Coverage Report

Created: 2025-06-13 07:07

/src/aom/aom_dsp/x86/intrapred_ssse3.c
Line
Count
Source (jump to first uncovered line)
1
/*
2
 * Copyright (c) 2017, Alliance for Open Media. All rights reserved.
3
 *
4
 * This source code is subject to the terms of the BSD 2 Clause License and
5
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6
 * was not distributed with this source code in the LICENSE file, you can
7
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8
 * Media Patent License 1.0 was not distributed with this source code in the
9
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10
 */
11
12
#include <tmmintrin.h>
13
14
#include "config/aom_dsp_rtcd.h"
15
16
#include "aom_dsp/intrapred_common.h"
17
18
// -----------------------------------------------------------------------------
19
// PAETH_PRED
20
21
// Return 8 16-bit pixels in one row
22
static inline __m128i paeth_8x1_pred(const __m128i *left, const __m128i *top,
23
7.53M
                                     const __m128i *topleft) {
24
7.53M
  const __m128i base = _mm_sub_epi16(_mm_add_epi16(*top, *left), *topleft);
25
26
7.53M
  __m128i pl = _mm_abs_epi16(_mm_sub_epi16(base, *left));
27
7.53M
  __m128i pt = _mm_abs_epi16(_mm_sub_epi16(base, *top));
28
7.53M
  __m128i ptl = _mm_abs_epi16(_mm_sub_epi16(base, *topleft));
29
30
7.53M
  __m128i mask1 = _mm_cmpgt_epi16(pl, pt);
31
7.53M
  mask1 = _mm_or_si128(mask1, _mm_cmpgt_epi16(pl, ptl));
32
7.53M
  __m128i mask2 = _mm_cmpgt_epi16(pt, ptl);
33
34
7.53M
  pl = _mm_andnot_si128(mask1, *left);
35
36
7.53M
  ptl = _mm_and_si128(mask2, *topleft);
37
7.53M
  pt = _mm_andnot_si128(mask2, *top);
38
7.53M
  pt = _mm_or_si128(pt, ptl);
39
7.53M
  pt = _mm_and_si128(mask1, pt);
40
41
7.53M
  return _mm_or_si128(pl, pt);
42
7.53M
}
43
44
void aom_paeth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t stride,
45
268k
                                   const uint8_t *above, const uint8_t *left) {
46
268k
  __m128i l = _mm_loadl_epi64((const __m128i *)left);
47
268k
  const __m128i t = _mm_loadl_epi64((const __m128i *)above);
48
268k
  const __m128i zero = _mm_setzero_si128();
49
268k
  const __m128i t16 = _mm_unpacklo_epi8(t, zero);
50
268k
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
51
268k
  __m128i rep = _mm_set1_epi16((short)0x8000);
52
268k
  const __m128i one = _mm_set1_epi16(1);
53
54
268k
  int i;
55
1.34M
  for (i = 0; i < 4; ++i) {
56
1.07M
    const __m128i l16 = _mm_shuffle_epi8(l, rep);
57
1.07M
    const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16);
58
59
1.07M
    *(int *)dst = _mm_cvtsi128_si32(_mm_packus_epi16(row, row));
60
1.07M
    dst += stride;
61
1.07M
    rep = _mm_add_epi16(rep, one);
62
1.07M
  }
63
268k
}
64
65
void aom_paeth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t stride,
66
58.3k
                                   const uint8_t *above, const uint8_t *left) {
67
58.3k
  __m128i l = _mm_loadl_epi64((const __m128i *)left);
68
58.3k
  const __m128i t = _mm_loadl_epi64((const __m128i *)above);
69
58.3k
  const __m128i zero = _mm_setzero_si128();
70
58.3k
  const __m128i t16 = _mm_unpacklo_epi8(t, zero);
71
58.3k
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
72
58.3k
  __m128i rep = _mm_set1_epi16((short)0x8000);
73
58.3k
  const __m128i one = _mm_set1_epi16(1);
74
75
58.3k
  int i;
76
525k
  for (i = 0; i < 8; ++i) {
77
467k
    const __m128i l16 = _mm_shuffle_epi8(l, rep);
78
467k
    const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16);
79
80
467k
    *(int *)dst = _mm_cvtsi128_si32(_mm_packus_epi16(row, row));
81
467k
    dst += stride;
82
467k
    rep = _mm_add_epi16(rep, one);
83
467k
  }
84
58.3k
}
85
86
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
87
void aom_paeth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t stride,
88
62.1k
                                    const uint8_t *above, const uint8_t *left) {
89
62.1k
  __m128i l = _mm_load_si128((const __m128i *)left);
90
62.1k
  const __m128i t = _mm_cvtsi32_si128(((const int *)above)[0]);
91
62.1k
  const __m128i zero = _mm_setzero_si128();
92
62.1k
  const __m128i t16 = _mm_unpacklo_epi8(t, zero);
93
62.1k
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
94
62.1k
  __m128i rep = _mm_set1_epi16((short)0x8000);
95
62.1k
  const __m128i one = _mm_set1_epi16(1);
96
97
1.05M
  for (int i = 0; i < 16; ++i) {
98
993k
    const __m128i l16 = _mm_shuffle_epi8(l, rep);
99
993k
    const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16);
100
101
993k
    *(int *)dst = _mm_cvtsi128_si32(_mm_packus_epi16(row, row));
102
993k
    dst += stride;
103
993k
    rep = _mm_add_epi16(rep, one);
104
993k
  }
105
62.1k
}
106
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
107
108
void aom_paeth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t stride,
109
85.2k
                                   const uint8_t *above, const uint8_t *left) {
110
85.2k
  __m128i l = _mm_loadl_epi64((const __m128i *)left);
111
85.2k
  const __m128i t = _mm_loadl_epi64((const __m128i *)above);
112
85.2k
  const __m128i zero = _mm_setzero_si128();
113
85.2k
  const __m128i t16 = _mm_unpacklo_epi8(t, zero);
114
85.2k
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
115
85.2k
  __m128i rep = _mm_set1_epi16((short)0x8000);
116
85.2k
  const __m128i one = _mm_set1_epi16(1);
117
118
85.2k
  int i;
119
426k
  for (i = 0; i < 4; ++i) {
120
341k
    const __m128i l16 = _mm_shuffle_epi8(l, rep);
121
341k
    const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16);
122
123
341k
    _mm_storel_epi64((__m128i *)dst, _mm_packus_epi16(row, row));
124
341k
    dst += stride;
125
341k
    rep = _mm_add_epi16(rep, one);
126
341k
  }
127
85.2k
}
128
129
void aom_paeth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t stride,
130
176k
                                   const uint8_t *above, const uint8_t *left) {
131
176k
  __m128i l = _mm_loadl_epi64((const __m128i *)left);
132
176k
  const __m128i t = _mm_loadl_epi64((const __m128i *)above);
133
176k
  const __m128i zero = _mm_setzero_si128();
134
176k
  const __m128i t16 = _mm_unpacklo_epi8(t, zero);
135
176k
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
136
176k
  __m128i rep = _mm_set1_epi16((short)0x8000);
137
176k
  const __m128i one = _mm_set1_epi16(1);
138
139
176k
  int i;
140
1.58M
  for (i = 0; i < 8; ++i) {
141
1.41M
    const __m128i l16 = _mm_shuffle_epi8(l, rep);
142
1.41M
    const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16);
143
144
1.41M
    _mm_storel_epi64((__m128i *)dst, _mm_packus_epi16(row, row));
145
1.41M
    dst += stride;
146
1.41M
    rep = _mm_add_epi16(rep, one);
147
1.41M
  }
148
176k
}
149
150
void aom_paeth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t stride,
151
47.2k
                                    const uint8_t *above, const uint8_t *left) {
152
47.2k
  __m128i l = _mm_load_si128((const __m128i *)left);
153
47.2k
  const __m128i t = _mm_loadl_epi64((const __m128i *)above);
154
47.2k
  const __m128i zero = _mm_setzero_si128();
155
47.2k
  const __m128i t16 = _mm_unpacklo_epi8(t, zero);
156
47.2k
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
157
47.2k
  __m128i rep = _mm_set1_epi16((short)0x8000);
158
47.2k
  const __m128i one = _mm_set1_epi16(1);
159
160
47.2k
  int i;
161
802k
  for (i = 0; i < 16; ++i) {
162
755k
    const __m128i l16 = _mm_shuffle_epi8(l, rep);
163
755k
    const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16);
164
165
755k
    _mm_storel_epi64((__m128i *)dst, _mm_packus_epi16(row, row));
166
755k
    dst += stride;
167
755k
    rep = _mm_add_epi16(rep, one);
168
755k
  }
169
47.2k
}
170
171
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
172
void aom_paeth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t stride,
173
31.2k
                                    const uint8_t *above, const uint8_t *left) {
174
31.2k
  const __m128i t = _mm_loadl_epi64((const __m128i *)above);
175
31.2k
  const __m128i zero = _mm_setzero_si128();
176
31.2k
  const __m128i t16 = _mm_unpacklo_epi8(t, zero);
177
31.2k
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
178
31.2k
  const __m128i one = _mm_set1_epi16(1);
179
180
93.8k
  for (int j = 0; j < 2; ++j) {
181
62.5k
    const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16));
182
62.5k
    __m128i rep = _mm_set1_epi16((short)0x8000);
183
1.06M
    for (int i = 0; i < 16; ++i) {
184
1.00M
      const __m128i l16 = _mm_shuffle_epi8(l, rep);
185
1.00M
      const __m128i row = paeth_8x1_pred(&l16, &t16, &tl16);
186
187
1.00M
      _mm_storel_epi64((__m128i *)dst, _mm_packus_epi16(row, row));
188
1.00M
      dst += stride;
189
1.00M
      rep = _mm_add_epi16(rep, one);
190
1.00M
    }
191
62.5k
  }
192
31.2k
}
193
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
194
195
// Return 16 8-bit pixels in one row
196
static inline __m128i paeth_16x1_pred(const __m128i *left, const __m128i *top0,
197
                                      const __m128i *top1,
198
745k
                                      const __m128i *topleft) {
199
745k
  const __m128i p0 = paeth_8x1_pred(left, top0, topleft);
200
745k
  const __m128i p1 = paeth_8x1_pred(left, top1, topleft);
201
745k
  return _mm_packus_epi16(p0, p1);
202
745k
}
203
204
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
205
void aom_paeth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t stride,
206
67.5k
                                    const uint8_t *above, const uint8_t *left) {
207
67.5k
  __m128i l = _mm_cvtsi32_si128(((const int *)left)[0]);
208
67.5k
  const __m128i t = _mm_load_si128((const __m128i *)above);
209
67.5k
  const __m128i zero = _mm_setzero_si128();
210
67.5k
  const __m128i top0 = _mm_unpacklo_epi8(t, zero);
211
67.5k
  const __m128i top1 = _mm_unpackhi_epi8(t, zero);
212
67.5k
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
213
67.5k
  __m128i rep = _mm_set1_epi16((short)0x8000);
214
67.5k
  const __m128i one = _mm_set1_epi16(1);
215
216
337k
  for (int i = 0; i < 4; ++i) {
217
270k
    const __m128i l16 = _mm_shuffle_epi8(l, rep);
218
270k
    const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16);
219
220
270k
    _mm_store_si128((__m128i *)dst, row);
221
270k
    dst += stride;
222
270k
    rep = _mm_add_epi16(rep, one);
223
270k
  }
224
67.5k
}
225
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
226
227
void aom_paeth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t stride,
228
0
                                    const uint8_t *above, const uint8_t *left) {
229
0
  __m128i l = _mm_loadl_epi64((const __m128i *)left);
230
0
  const __m128i t = _mm_load_si128((const __m128i *)above);
231
0
  const __m128i zero = _mm_setzero_si128();
232
0
  const __m128i top0 = _mm_unpacklo_epi8(t, zero);
233
0
  const __m128i top1 = _mm_unpackhi_epi8(t, zero);
234
0
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
235
0
  __m128i rep = _mm_set1_epi16((short)0x8000);
236
0
  const __m128i one = _mm_set1_epi16(1);
237
238
0
  int i;
239
0
  for (i = 0; i < 8; ++i) {
240
0
    const __m128i l16 = _mm_shuffle_epi8(l, rep);
241
0
    const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16);
242
243
0
    _mm_store_si128((__m128i *)dst, row);
244
0
    dst += stride;
245
0
    rep = _mm_add_epi16(rep, one);
246
0
  }
247
0
}
248
249
void aom_paeth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride,
250
                                     const uint8_t *above,
251
0
                                     const uint8_t *left) {
252
0
  __m128i l = _mm_load_si128((const __m128i *)left);
253
0
  const __m128i t = _mm_load_si128((const __m128i *)above);
254
0
  const __m128i zero = _mm_setzero_si128();
255
0
  const __m128i top0 = _mm_unpacklo_epi8(t, zero);
256
0
  const __m128i top1 = _mm_unpackhi_epi8(t, zero);
257
0
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
258
0
  __m128i rep = _mm_set1_epi16((short)0x8000);
259
0
  const __m128i one = _mm_set1_epi16(1);
260
261
0
  int i;
262
0
  for (i = 0; i < 16; ++i) {
263
0
    const __m128i l16 = _mm_shuffle_epi8(l, rep);
264
0
    const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16);
265
266
0
    _mm_store_si128((__m128i *)dst, row);
267
0
    dst += stride;
268
0
    rep = _mm_add_epi16(rep, one);
269
0
  }
270
0
}
271
272
void aom_paeth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t stride,
273
                                     const uint8_t *above,
274
0
                                     const uint8_t *left) {
275
0
  __m128i l = _mm_load_si128((const __m128i *)left);
276
0
  const __m128i t = _mm_load_si128((const __m128i *)above);
277
0
  const __m128i zero = _mm_setzero_si128();
278
0
  const __m128i top0 = _mm_unpacklo_epi8(t, zero);
279
0
  const __m128i top1 = _mm_unpackhi_epi8(t, zero);
280
0
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
281
0
  __m128i rep = _mm_set1_epi16((short)0x8000);
282
0
  const __m128i one = _mm_set1_epi16(1);
283
0
  __m128i l16;
284
285
0
  int i;
286
0
  for (i = 0; i < 16; ++i) {
287
0
    l16 = _mm_shuffle_epi8(l, rep);
288
0
    const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16);
289
290
0
    _mm_store_si128((__m128i *)dst, row);
291
0
    dst += stride;
292
0
    rep = _mm_add_epi16(rep, one);
293
0
  }
294
295
0
  l = _mm_load_si128((const __m128i *)(left + 16));
296
0
  rep = _mm_set1_epi16((short)0x8000);
297
0
  for (i = 0; i < 16; ++i) {
298
0
    l16 = _mm_shuffle_epi8(l, rep);
299
0
    const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16);
300
301
0
    _mm_store_si128((__m128i *)dst, row);
302
0
    dst += stride;
303
0
    rep = _mm_add_epi16(rep, one);
304
0
  }
305
0
}
306
307
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
308
void aom_paeth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t stride,
309
                                     const uint8_t *above,
310
0
                                     const uint8_t *left) {
311
0
  const __m128i t = _mm_load_si128((const __m128i *)above);
312
0
  const __m128i zero = _mm_setzero_si128();
313
0
  const __m128i top0 = _mm_unpacklo_epi8(t, zero);
314
0
  const __m128i top1 = _mm_unpackhi_epi8(t, zero);
315
0
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
316
0
  const __m128i one = _mm_set1_epi16(1);
317
318
0
  for (int j = 0; j < 4; ++j) {
319
0
    const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16));
320
0
    __m128i rep = _mm_set1_epi16((short)0x8000);
321
0
    for (int i = 0; i < 16; ++i) {
322
0
      const __m128i l16 = _mm_shuffle_epi8(l, rep);
323
0
      const __m128i row = paeth_16x1_pred(&l16, &top0, &top1, &tl16);
324
0
      _mm_store_si128((__m128i *)dst, row);
325
0
      dst += stride;
326
0
      rep = _mm_add_epi16(rep, one);
327
0
    }
328
0
  }
329
0
}
330
331
void aom_paeth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t stride,
332
29.6k
                                    const uint8_t *above, const uint8_t *left) {
333
29.6k
  const __m128i a = _mm_load_si128((const __m128i *)above);
334
29.6k
  const __m128i b = _mm_load_si128((const __m128i *)(above + 16));
335
29.6k
  const __m128i zero = _mm_setzero_si128();
336
29.6k
  const __m128i al = _mm_unpacklo_epi8(a, zero);
337
29.6k
  const __m128i ah = _mm_unpackhi_epi8(a, zero);
338
29.6k
  const __m128i bl = _mm_unpacklo_epi8(b, zero);
339
29.6k
  const __m128i bh = _mm_unpackhi_epi8(b, zero);
340
341
29.6k
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
342
29.6k
  __m128i rep = _mm_set1_epi16((short)0x8000);
343
29.6k
  const __m128i one = _mm_set1_epi16(1);
344
29.6k
  const __m128i l = _mm_loadl_epi64((const __m128i *)left);
345
29.6k
  __m128i l16;
346
347
267k
  for (int i = 0; i < 8; ++i) {
348
237k
    l16 = _mm_shuffle_epi8(l, rep);
349
237k
    const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16);
350
237k
    const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
351
352
237k
    _mm_store_si128((__m128i *)dst, r32l);
353
237k
    _mm_store_si128((__m128i *)(dst + 16), r32h);
354
237k
    dst += stride;
355
237k
    rep = _mm_add_epi16(rep, one);
356
237k
  }
357
29.6k
}
358
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
359
360
void aom_paeth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t stride,
361
                                     const uint8_t *above,
362
0
                                     const uint8_t *left) {
363
0
  const __m128i a = _mm_load_si128((const __m128i *)above);
364
0
  const __m128i b = _mm_load_si128((const __m128i *)(above + 16));
365
0
  const __m128i zero = _mm_setzero_si128();
366
0
  const __m128i al = _mm_unpacklo_epi8(a, zero);
367
0
  const __m128i ah = _mm_unpackhi_epi8(a, zero);
368
0
  const __m128i bl = _mm_unpacklo_epi8(b, zero);
369
0
  const __m128i bh = _mm_unpackhi_epi8(b, zero);
370
371
0
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
372
0
  __m128i rep = _mm_set1_epi16((short)0x8000);
373
0
  const __m128i one = _mm_set1_epi16(1);
374
0
  __m128i l = _mm_load_si128((const __m128i *)left);
375
0
  __m128i l16;
376
377
0
  int i;
378
0
  for (i = 0; i < 16; ++i) {
379
0
    l16 = _mm_shuffle_epi8(l, rep);
380
0
    const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16);
381
0
    const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
382
383
0
    _mm_store_si128((__m128i *)dst, r32l);
384
0
    _mm_store_si128((__m128i *)(dst + 16), r32h);
385
0
    dst += stride;
386
0
    rep = _mm_add_epi16(rep, one);
387
0
  }
388
0
}
389
390
void aom_paeth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride,
391
                                     const uint8_t *above,
392
0
                                     const uint8_t *left) {
393
0
  const __m128i a = _mm_load_si128((const __m128i *)above);
394
0
  const __m128i b = _mm_load_si128((const __m128i *)(above + 16));
395
0
  const __m128i zero = _mm_setzero_si128();
396
0
  const __m128i al = _mm_unpacklo_epi8(a, zero);
397
0
  const __m128i ah = _mm_unpackhi_epi8(a, zero);
398
0
  const __m128i bl = _mm_unpacklo_epi8(b, zero);
399
0
  const __m128i bh = _mm_unpackhi_epi8(b, zero);
400
401
0
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
402
0
  __m128i rep = _mm_set1_epi16((short)0x8000);
403
0
  const __m128i one = _mm_set1_epi16(1);
404
0
  __m128i l = _mm_load_si128((const __m128i *)left);
405
0
  __m128i l16;
406
407
0
  int i;
408
0
  for (i = 0; i < 16; ++i) {
409
0
    l16 = _mm_shuffle_epi8(l, rep);
410
0
    const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16);
411
0
    const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
412
413
0
    _mm_store_si128((__m128i *)dst, r32l);
414
0
    _mm_store_si128((__m128i *)(dst + 16), r32h);
415
0
    dst += stride;
416
0
    rep = _mm_add_epi16(rep, one);
417
0
  }
418
419
0
  rep = _mm_set1_epi16((short)0x8000);
420
0
  l = _mm_load_si128((const __m128i *)(left + 16));
421
0
  for (i = 0; i < 16; ++i) {
422
0
    l16 = _mm_shuffle_epi8(l, rep);
423
0
    const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16);
424
0
    const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
425
426
0
    _mm_store_si128((__m128i *)dst, r32l);
427
0
    _mm_store_si128((__m128i *)(dst + 16), r32h);
428
0
    dst += stride;
429
0
    rep = _mm_add_epi16(rep, one);
430
0
  }
431
0
}
432
433
void aom_paeth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t stride,
434
                                     const uint8_t *above,
435
0
                                     const uint8_t *left) {
436
0
  const __m128i a = _mm_load_si128((const __m128i *)above);
437
0
  const __m128i b = _mm_load_si128((const __m128i *)(above + 16));
438
0
  const __m128i zero = _mm_setzero_si128();
439
0
  const __m128i al = _mm_unpacklo_epi8(a, zero);
440
0
  const __m128i ah = _mm_unpackhi_epi8(a, zero);
441
0
  const __m128i bl = _mm_unpacklo_epi8(b, zero);
442
0
  const __m128i bh = _mm_unpackhi_epi8(b, zero);
443
444
0
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
445
0
  const __m128i one = _mm_set1_epi16(1);
446
0
  __m128i l16;
447
448
0
  int i, j;
449
0
  for (j = 0; j < 4; ++j) {
450
0
    const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16));
451
0
    __m128i rep = _mm_set1_epi16((short)0x8000);
452
0
    for (i = 0; i < 16; ++i) {
453
0
      l16 = _mm_shuffle_epi8(l, rep);
454
0
      const __m128i r32l = paeth_16x1_pred(&l16, &al, &ah, &tl16);
455
0
      const __m128i r32h = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
456
457
0
      _mm_store_si128((__m128i *)dst, r32l);
458
0
      _mm_store_si128((__m128i *)(dst + 16), r32h);
459
0
      dst += stride;
460
0
      rep = _mm_add_epi16(rep, one);
461
0
    }
462
0
  }
463
0
}
464
465
void aom_paeth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t stride,
466
                                     const uint8_t *above,
467
0
                                     const uint8_t *left) {
468
0
  const __m128i a = _mm_load_si128((const __m128i *)above);
469
0
  const __m128i b = _mm_load_si128((const __m128i *)(above + 16));
470
0
  const __m128i c = _mm_load_si128((const __m128i *)(above + 32));
471
0
  const __m128i d = _mm_load_si128((const __m128i *)(above + 48));
472
0
  const __m128i zero = _mm_setzero_si128();
473
0
  const __m128i al = _mm_unpacklo_epi8(a, zero);
474
0
  const __m128i ah = _mm_unpackhi_epi8(a, zero);
475
0
  const __m128i bl = _mm_unpacklo_epi8(b, zero);
476
0
  const __m128i bh = _mm_unpackhi_epi8(b, zero);
477
0
  const __m128i cl = _mm_unpacklo_epi8(c, zero);
478
0
  const __m128i ch = _mm_unpackhi_epi8(c, zero);
479
0
  const __m128i dl = _mm_unpacklo_epi8(d, zero);
480
0
  const __m128i dh = _mm_unpackhi_epi8(d, zero);
481
482
0
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
483
0
  const __m128i one = _mm_set1_epi16(1);
484
0
  __m128i l16;
485
486
0
  int i, j;
487
0
  for (j = 0; j < 2; ++j) {
488
0
    const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16));
489
0
    __m128i rep = _mm_set1_epi16((short)0x8000);
490
0
    for (i = 0; i < 16; ++i) {
491
0
      l16 = _mm_shuffle_epi8(l, rep);
492
0
      const __m128i r0 = paeth_16x1_pred(&l16, &al, &ah, &tl16);
493
0
      const __m128i r1 = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
494
0
      const __m128i r2 = paeth_16x1_pred(&l16, &cl, &ch, &tl16);
495
0
      const __m128i r3 = paeth_16x1_pred(&l16, &dl, &dh, &tl16);
496
497
0
      _mm_store_si128((__m128i *)dst, r0);
498
0
      _mm_store_si128((__m128i *)(dst + 16), r1);
499
0
      _mm_store_si128((__m128i *)(dst + 32), r2);
500
0
      _mm_store_si128((__m128i *)(dst + 48), r3);
501
0
      dst += stride;
502
0
      rep = _mm_add_epi16(rep, one);
503
0
    }
504
0
  }
505
0
}
506
507
void aom_paeth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t stride,
508
                                     const uint8_t *above,
509
0
                                     const uint8_t *left) {
510
0
  const __m128i a = _mm_load_si128((const __m128i *)above);
511
0
  const __m128i b = _mm_load_si128((const __m128i *)(above + 16));
512
0
  const __m128i c = _mm_load_si128((const __m128i *)(above + 32));
513
0
  const __m128i d = _mm_load_si128((const __m128i *)(above + 48));
514
0
  const __m128i zero = _mm_setzero_si128();
515
0
  const __m128i al = _mm_unpacklo_epi8(a, zero);
516
0
  const __m128i ah = _mm_unpackhi_epi8(a, zero);
517
0
  const __m128i bl = _mm_unpacklo_epi8(b, zero);
518
0
  const __m128i bh = _mm_unpackhi_epi8(b, zero);
519
0
  const __m128i cl = _mm_unpacklo_epi8(c, zero);
520
0
  const __m128i ch = _mm_unpackhi_epi8(c, zero);
521
0
  const __m128i dl = _mm_unpacklo_epi8(d, zero);
522
0
  const __m128i dh = _mm_unpackhi_epi8(d, zero);
523
524
0
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
525
0
  const __m128i one = _mm_set1_epi16(1);
526
0
  __m128i l16;
527
528
0
  int i, j;
529
0
  for (j = 0; j < 4; ++j) {
530
0
    const __m128i l = _mm_load_si128((const __m128i *)(left + j * 16));
531
0
    __m128i rep = _mm_set1_epi16((short)0x8000);
532
0
    for (i = 0; i < 16; ++i) {
533
0
      l16 = _mm_shuffle_epi8(l, rep);
534
0
      const __m128i r0 = paeth_16x1_pred(&l16, &al, &ah, &tl16);
535
0
      const __m128i r1 = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
536
0
      const __m128i r2 = paeth_16x1_pred(&l16, &cl, &ch, &tl16);
537
0
      const __m128i r3 = paeth_16x1_pred(&l16, &dl, &dh, &tl16);
538
539
0
      _mm_store_si128((__m128i *)dst, r0);
540
0
      _mm_store_si128((__m128i *)(dst + 16), r1);
541
0
      _mm_store_si128((__m128i *)(dst + 32), r2);
542
0
      _mm_store_si128((__m128i *)(dst + 48), r3);
543
0
      dst += stride;
544
0
      rep = _mm_add_epi16(rep, one);
545
0
    }
546
0
  }
547
0
}
548
549
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
550
void aom_paeth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t stride,
551
                                     const uint8_t *above,
552
0
                                     const uint8_t *left) {
553
0
  const __m128i a = _mm_load_si128((const __m128i *)above);
554
0
  const __m128i b = _mm_load_si128((const __m128i *)(above + 16));
555
0
  const __m128i c = _mm_load_si128((const __m128i *)(above + 32));
556
0
  const __m128i d = _mm_load_si128((const __m128i *)(above + 48));
557
0
  const __m128i zero = _mm_setzero_si128();
558
0
  const __m128i al = _mm_unpacklo_epi8(a, zero);
559
0
  const __m128i ah = _mm_unpackhi_epi8(a, zero);
560
0
  const __m128i bl = _mm_unpacklo_epi8(b, zero);
561
0
  const __m128i bh = _mm_unpackhi_epi8(b, zero);
562
0
  const __m128i cl = _mm_unpacklo_epi8(c, zero);
563
0
  const __m128i ch = _mm_unpackhi_epi8(c, zero);
564
0
  const __m128i dl = _mm_unpacklo_epi8(d, zero);
565
0
  const __m128i dh = _mm_unpackhi_epi8(d, zero);
566
567
0
  const __m128i tl16 = _mm_set1_epi16((int16_t)above[-1]);
568
0
  const __m128i one = _mm_set1_epi16(1);
569
0
  __m128i l16;
570
571
0
  int i;
572
0
  const __m128i l = _mm_load_si128((const __m128i *)left);
573
0
  __m128i rep = _mm_set1_epi16((short)0x8000);
574
0
  for (i = 0; i < 16; ++i) {
575
0
    l16 = _mm_shuffle_epi8(l, rep);
576
0
    const __m128i r0 = paeth_16x1_pred(&l16, &al, &ah, &tl16);
577
0
    const __m128i r1 = paeth_16x1_pred(&l16, &bl, &bh, &tl16);
578
0
    const __m128i r2 = paeth_16x1_pred(&l16, &cl, &ch, &tl16);
579
0
    const __m128i r3 = paeth_16x1_pred(&l16, &dl, &dh, &tl16);
580
581
0
    _mm_store_si128((__m128i *)dst, r0);
582
0
    _mm_store_si128((__m128i *)(dst + 16), r1);
583
0
    _mm_store_si128((__m128i *)(dst + 32), r2);
584
0
    _mm_store_si128((__m128i *)(dst + 48), r3);
585
0
    dst += stride;
586
0
    rep = _mm_add_epi16(rep, one);
587
0
  }
588
0
}
589
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
590
591
// -----------------------------------------------------------------------------
592
// SMOOTH_PRED
593
594
// pixels[0]: above and below_pred interleave vector
595
// pixels[1]: left vector
596
// pixels[2]: right_pred vector
597
static inline void load_pixel_w4(const uint8_t *above, const uint8_t *left,
598
296k
                                 int height, __m128i *pixels) {
599
296k
  __m128i d = _mm_cvtsi32_si128(((const int *)above)[0]);
600
296k
  if (height == 4)
601
201k
    pixels[1] = _mm_cvtsi32_si128(((const int *)left)[0]);
602
95.0k
  else if (height == 8)
603
62.6k
    pixels[1] = _mm_loadl_epi64(((const __m128i *)left));
604
32.4k
  else
605
32.4k
    pixels[1] = _mm_loadu_si128(((const __m128i *)left));
606
607
296k
  pixels[2] = _mm_set1_epi16((int16_t)above[3]);
608
609
296k
  const __m128i bp = _mm_set1_epi16((int16_t)left[height - 1]);
610
296k
  const __m128i zero = _mm_setzero_si128();
611
296k
  d = _mm_unpacklo_epi8(d, zero);
612
296k
  pixels[0] = _mm_unpacklo_epi16(d, bp);
613
296k
}
614
615
// weight_h[0]: weight_h vector
616
// weight_h[1]: scale - weight_h vector
617
// weight_h[2]: same as [0], second half for height = 16 only
618
// weight_h[3]: same as [1], second half for height = 16 only
619
// weight_w[0]: weights_w and scale - weights_w interleave vector
620
static inline void load_weight_w4(int height, __m128i *weight_h,
621
296k
                                  __m128i *weight_w) {
622
296k
  const __m128i zero = _mm_setzero_si128();
623
296k
  const __m128i d = _mm_set1_epi16((int16_t)(1 << SMOOTH_WEIGHT_LOG2_SCALE));
624
296k
  const __m128i t = _mm_cvtsi32_si128(((const int *)smooth_weights)[0]);
625
296k
  weight_h[0] = _mm_unpacklo_epi8(t, zero);
626
296k
  weight_h[1] = _mm_sub_epi16(d, weight_h[0]);
627
296k
  weight_w[0] = _mm_unpacklo_epi16(weight_h[0], weight_h[1]);
628
629
296k
  if (height == 8) {
630
62.6k
    const __m128i weight = _mm_loadl_epi64((const __m128i *)&smooth_weights[4]);
631
62.6k
    weight_h[0] = _mm_unpacklo_epi8(weight, zero);
632
62.6k
    weight_h[1] = _mm_sub_epi16(d, weight_h[0]);
633
233k
  } else if (height == 16) {
634
32.4k
    const __m128i weight =
635
32.4k
        _mm_loadu_si128((const __m128i *)&smooth_weights[12]);
636
32.4k
    weight_h[0] = _mm_unpacklo_epi8(weight, zero);
637
32.4k
    weight_h[1] = _mm_sub_epi16(d, weight_h[0]);
638
32.4k
    weight_h[2] = _mm_unpackhi_epi8(weight, zero);
639
32.4k
    weight_h[3] = _mm_sub_epi16(d, weight_h[2]);
640
32.4k
  }
641
296k
}
642
643
static inline void smooth_pred_4xh(const __m128i *pixel, const __m128i *wh,
644
                                   const __m128i *ww, int h, uint8_t *dst,
645
328k
                                   ptrdiff_t stride, int second_half) {
646
328k
  const __m128i round = _mm_set1_epi32((1 << SMOOTH_WEIGHT_LOG2_SCALE));
647
328k
  const __m128i one = _mm_set1_epi16(1);
648
328k
  const __m128i inc = _mm_set1_epi16(0x202);
649
328k
  const __m128i gat = _mm_set1_epi32(0xc080400);
650
328k
  __m128i rep = second_half ? _mm_set1_epi16((short)0x8008)
651
328k
                            : _mm_set1_epi16((short)0x8000);
652
328k
  __m128i d = _mm_set1_epi16(0x100);
653
654
2.15M
  for (int i = 0; i < h; ++i) {
655
1.82M
    const __m128i wg_wg = _mm_shuffle_epi8(wh[0], d);
656
1.82M
    const __m128i sc_sc = _mm_shuffle_epi8(wh[1], d);
657
1.82M
    const __m128i wh_sc = _mm_unpacklo_epi16(wg_wg, sc_sc);
658
1.82M
    __m128i s = _mm_madd_epi16(pixel[0], wh_sc);
659
660
1.82M
    __m128i b = _mm_shuffle_epi8(pixel[1], rep);
661
1.82M
    b = _mm_unpacklo_epi16(b, pixel[2]);
662
1.82M
    __m128i sum = _mm_madd_epi16(b, ww[0]);
663
664
1.82M
    sum = _mm_add_epi32(s, sum);
665
1.82M
    sum = _mm_add_epi32(sum, round);
666
1.82M
    sum = _mm_srai_epi32(sum, 1 + SMOOTH_WEIGHT_LOG2_SCALE);
667
668
1.82M
    sum = _mm_shuffle_epi8(sum, gat);
669
1.82M
    *(int *)dst = _mm_cvtsi128_si32(sum);
670
1.82M
    dst += stride;
671
672
1.82M
    rep = _mm_add_epi16(rep, one);
673
1.82M
    d = _mm_add_epi16(d, inc);
674
1.82M
  }
675
328k
}
676
677
void aom_smooth_predictor_4x4_ssse3(uint8_t *dst, ptrdiff_t stride,
678
201k
                                    const uint8_t *above, const uint8_t *left) {
679
201k
  __m128i pixels[3];
680
201k
  load_pixel_w4(above, left, 4, pixels);
681
682
201k
  __m128i wh[4], ww[2];
683
201k
  load_weight_w4(4, wh, ww);
684
685
201k
  smooth_pred_4xh(pixels, wh, ww, 4, dst, stride, 0);
686
201k
}
687
688
void aom_smooth_predictor_4x8_ssse3(uint8_t *dst, ptrdiff_t stride,
689
62.6k
                                    const uint8_t *above, const uint8_t *left) {
690
62.6k
  __m128i pixels[3];
691
62.6k
  load_pixel_w4(above, left, 8, pixels);
692
693
62.6k
  __m128i wh[4], ww[2];
694
62.6k
  load_weight_w4(8, wh, ww);
695
696
62.6k
  smooth_pred_4xh(pixels, wh, ww, 8, dst, stride, 0);
697
62.6k
}
698
699
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
700
void aom_smooth_predictor_4x16_ssse3(uint8_t *dst, ptrdiff_t stride,
701
                                     const uint8_t *above,
702
32.4k
                                     const uint8_t *left) {
703
32.4k
  __m128i pixels[3];
704
32.4k
  load_pixel_w4(above, left, 16, pixels);
705
706
32.4k
  __m128i wh[4], ww[2];
707
32.4k
  load_weight_w4(16, wh, ww);
708
709
32.4k
  smooth_pred_4xh(pixels, wh, ww, 8, dst, stride, 0);
710
32.4k
  dst += stride << 3;
711
32.4k
  smooth_pred_4xh(pixels, &wh[2], ww, 8, dst, stride, 1);
712
32.4k
}
713
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
714
715
// pixels[0]: above and below_pred interleave vector, first half
716
// pixels[1]: above and below_pred interleave vector, second half
717
// pixels[2]: left vector
718
// pixels[3]: right_pred vector
719
// pixels[4]: above and below_pred interleave vector, first half
720
// pixels[5]: above and below_pred interleave vector, second half
721
// pixels[6]: left vector + 16
722
// pixels[7]: right_pred vector
723
static inline void load_pixel_w8(const uint8_t *above, const uint8_t *left,
724
347k
                                 int height, __m128i *pixels) {
725
347k
  const __m128i zero = _mm_setzero_si128();
726
347k
  const __m128i bp = _mm_set1_epi16((int16_t)left[height - 1]);
727
347k
  __m128i d = _mm_loadl_epi64((const __m128i *)above);
728
347k
  d = _mm_unpacklo_epi8(d, zero);
729
347k
  pixels[0] = _mm_unpacklo_epi16(d, bp);
730
347k
  pixels[1] = _mm_unpackhi_epi16(d, bp);
731
732
347k
  pixels[3] = _mm_set1_epi16((int16_t)above[7]);
733
734
347k
  if (height == 4) {
735
91.3k
    pixels[2] = _mm_cvtsi32_si128(((const int *)left)[0]);
736
255k
  } else if (height == 8) {
737
186k
    pixels[2] = _mm_loadl_epi64((const __m128i *)left);
738
186k
  } else if (height == 16) {
739
50.7k
    pixels[2] = _mm_load_si128((const __m128i *)left);
740
50.7k
  } else {
741
18.6k
    pixels[2] = _mm_load_si128((const __m128i *)left);
742
18.6k
    pixels[4] = pixels[0];
743
18.6k
    pixels[5] = pixels[1];
744
18.6k
    pixels[6] = _mm_load_si128((const __m128i *)(left + 16));
745
18.6k
    pixels[7] = pixels[3];
746
18.6k
  }
747
347k
}
748
749
// weight_h[0]: weight_h vector
750
// weight_h[1]: scale - weight_h vector
751
// weight_h[2]: same as [0], offset 8
752
// weight_h[3]: same as [1], offset 8
753
// weight_h[4]: same as [0], offset 16
754
// weight_h[5]: same as [1], offset 16
755
// weight_h[6]: same as [0], offset 24
756
// weight_h[7]: same as [1], offset 24
757
// weight_w[0]: weights_w and scale - weights_w interleave vector, first half
758
// weight_w[1]: weights_w and scale - weights_w interleave vector, second half
759
static inline void load_weight_w8(int height, __m128i *weight_h,
760
347k
                                  __m128i *weight_w) {
761
347k
  const __m128i zero = _mm_setzero_si128();
762
347k
  const int we_offset = height < 8 ? 0 : 4;
763
347k
  __m128i we = _mm_loadu_si128((const __m128i *)&smooth_weights[we_offset]);
764
347k
  weight_h[0] = _mm_unpacklo_epi8(we, zero);
765
347k
  const __m128i d = _mm_set1_epi16((int16_t)(1 << SMOOTH_WEIGHT_LOG2_SCALE));
766
347k
  weight_h[1] = _mm_sub_epi16(d, weight_h[0]);
767
768
347k
  if (height == 4) {
769
91.3k
    we = _mm_srli_si128(we, 4);
770
91.3k
    __m128i tmp1 = _mm_unpacklo_epi8(we, zero);
771
91.3k
    __m128i tmp2 = _mm_sub_epi16(d, tmp1);
772
91.3k
    weight_w[0] = _mm_unpacklo_epi16(tmp1, tmp2);
773
91.3k
    weight_w[1] = _mm_unpackhi_epi16(tmp1, tmp2);
774
255k
  } else {
775
255k
    weight_w[0] = _mm_unpacklo_epi16(weight_h[0], weight_h[1]);
776
255k
    weight_w[1] = _mm_unpackhi_epi16(weight_h[0], weight_h[1]);
777
255k
  }
778
779
347k
  if (height == 16) {
780
50.7k
    we = _mm_loadu_si128((const __m128i *)&smooth_weights[12]);
781
50.7k
    weight_h[0] = _mm_unpacklo_epi8(we, zero);
782
50.7k
    weight_h[1] = _mm_sub_epi16(d, weight_h[0]);
783
50.7k
    weight_h[2] = _mm_unpackhi_epi8(we, zero);
784
50.7k
    weight_h[3] = _mm_sub_epi16(d, weight_h[2]);
785
296k
  } else if (height == 32) {
786
18.6k
    const __m128i weight_lo =
787
18.6k
        _mm_loadu_si128((const __m128i *)&smooth_weights[28]);
788
18.6k
    weight_h[0] = _mm_unpacklo_epi8(weight_lo, zero);
789
18.6k
    weight_h[1] = _mm_sub_epi16(d, weight_h[0]);
790
18.6k
    weight_h[2] = _mm_unpackhi_epi8(weight_lo, zero);
791
18.6k
    weight_h[3] = _mm_sub_epi16(d, weight_h[2]);
792
18.6k
    const __m128i weight_hi =
793
18.6k
        _mm_loadu_si128((const __m128i *)&smooth_weights[28 + 16]);
794
18.6k
    weight_h[4] = _mm_unpacklo_epi8(weight_hi, zero);
795
18.6k
    weight_h[5] = _mm_sub_epi16(d, weight_h[4]);
796
18.6k
    weight_h[6] = _mm_unpackhi_epi8(weight_hi, zero);
797
18.6k
    weight_h[7] = _mm_sub_epi16(d, weight_h[6]);
798
18.6k
  }
799
347k
}
800
801
static inline void smooth_pred_8xh(const __m128i *pixels, const __m128i *wh,
802
                                   const __m128i *ww, int h, uint8_t *dst,
803
453k
                                   ptrdiff_t stride, int second_half) {
804
453k
  const __m128i round = _mm_set1_epi32((1 << SMOOTH_WEIGHT_LOG2_SCALE));
805
453k
  const __m128i one = _mm_set1_epi16(1);
806
453k
  const __m128i inc = _mm_set1_epi16(0x202);
807
453k
  const __m128i gat = _mm_set_epi32(0, 0, 0xe0c0a08, 0x6040200);
808
809
453k
  __m128i rep = second_half ? _mm_set1_epi16((short)0x8008)
810
453k
                            : _mm_set1_epi16((short)0x8000);
811
453k
  __m128i d = _mm_set1_epi16(0x100);
812
813
453k
  int i;
814
3.71M
  for (i = 0; i < h; ++i) {
815
3.26M
    const __m128i wg_wg = _mm_shuffle_epi8(wh[0], d);
816
3.26M
    const __m128i sc_sc = _mm_shuffle_epi8(wh[1], d);
817
3.26M
    const __m128i wh_sc = _mm_unpacklo_epi16(wg_wg, sc_sc);
818
3.26M
    __m128i s0 = _mm_madd_epi16(pixels[0], wh_sc);
819
3.26M
    __m128i s1 = _mm_madd_epi16(pixels[1], wh_sc);
820
821
3.26M
    __m128i b = _mm_shuffle_epi8(pixels[2], rep);
822
3.26M
    b = _mm_unpacklo_epi16(b, pixels[3]);
823
3.26M
    __m128i sum0 = _mm_madd_epi16(b, ww[0]);
824
3.26M
    __m128i sum1 = _mm_madd_epi16(b, ww[1]);
825
826
3.26M
    s0 = _mm_add_epi32(s0, sum0);
827
3.26M
    s0 = _mm_add_epi32(s0, round);
828
3.26M
    s0 = _mm_srai_epi32(s0, 1 + SMOOTH_WEIGHT_LOG2_SCALE);
829
830
3.26M
    s1 = _mm_add_epi32(s1, sum1);
831
3.26M
    s1 = _mm_add_epi32(s1, round);
832
3.26M
    s1 = _mm_srai_epi32(s1, 1 + SMOOTH_WEIGHT_LOG2_SCALE);
833
834
3.26M
    sum0 = _mm_packus_epi16(s0, s1);
835
3.26M
    sum0 = _mm_shuffle_epi8(sum0, gat);
836
3.26M
    _mm_storel_epi64((__m128i *)dst, sum0);
837
3.26M
    dst += stride;
838
839
3.26M
    rep = _mm_add_epi16(rep, one);
840
3.26M
    d = _mm_add_epi16(d, inc);
841
3.26M
  }
842
453k
}
843
844
void aom_smooth_predictor_8x4_ssse3(uint8_t *dst, ptrdiff_t stride,
845
91.3k
                                    const uint8_t *above, const uint8_t *left) {
846
91.3k
  __m128i pixels[4];
847
91.3k
  load_pixel_w8(above, left, 4, pixels);
848
849
91.3k
  __m128i wh[4], ww[2];
850
91.3k
  load_weight_w8(4, wh, ww);
851
852
91.3k
  smooth_pred_8xh(pixels, wh, ww, 4, dst, stride, 0);
853
91.3k
}
854
855
void aom_smooth_predictor_8x8_ssse3(uint8_t *dst, ptrdiff_t stride,
856
186k
                                    const uint8_t *above, const uint8_t *left) {
857
186k
  __m128i pixels[4];
858
186k
  load_pixel_w8(above, left, 8, pixels);
859
860
186k
  __m128i wh[4], ww[2];
861
186k
  load_weight_w8(8, wh, ww);
862
863
186k
  smooth_pred_8xh(pixels, wh, ww, 8, dst, stride, 0);
864
186k
}
865
866
void aom_smooth_predictor_8x16_ssse3(uint8_t *dst, ptrdiff_t stride,
867
                                     const uint8_t *above,
868
50.7k
                                     const uint8_t *left) {
869
50.7k
  __m128i pixels[4];
870
50.7k
  load_pixel_w8(above, left, 16, pixels);
871
872
50.7k
  __m128i wh[4], ww[2];
873
50.7k
  load_weight_w8(16, wh, ww);
874
875
50.7k
  smooth_pred_8xh(pixels, wh, ww, 8, dst, stride, 0);
876
50.7k
  dst += stride << 3;
877
50.7k
  smooth_pred_8xh(pixels, &wh[2], ww, 8, dst, stride, 1);
878
50.7k
}
879
880
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
881
void aom_smooth_predictor_8x32_ssse3(uint8_t *dst, ptrdiff_t stride,
882
                                     const uint8_t *above,
883
18.6k
                                     const uint8_t *left) {
884
18.6k
  __m128i pixels[8];
885
18.6k
  load_pixel_w8(above, left, 32, pixels);
886
887
18.6k
  __m128i wh[8], ww[2];
888
18.6k
  load_weight_w8(32, wh, ww);
889
890
18.6k
  smooth_pred_8xh(&pixels[0], wh, ww, 8, dst, stride, 0);
891
18.6k
  dst += stride << 3;
892
18.6k
  smooth_pred_8xh(&pixels[0], &wh[2], ww, 8, dst, stride, 1);
893
18.6k
  dst += stride << 3;
894
18.6k
  smooth_pred_8xh(&pixels[4], &wh[4], ww, 8, dst, stride, 0);
895
18.6k
  dst += stride << 3;
896
18.6k
  smooth_pred_8xh(&pixels[4], &wh[6], ww, 8, dst, stride, 1);
897
18.6k
}
898
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
899
900
// TODO(slavarnway): Visual Studio only supports restrict when /std:c11
901
// (available in 2019+) or greater is specified; __restrict can be used in that
902
// case. This should be moved to rtcd and used consistently between the
903
// function declarations and definitions to avoid warnings in Visual Studio
904
// when defining LIBAOM_RESTRICT to restrict or __restrict.
905
#if defined(_MSC_VER)
906
#define LIBAOM_RESTRICT
907
#else
908
#define LIBAOM_RESTRICT restrict
909
#endif
910
911
491k
static AOM_FORCE_INLINE __m128i Load4(const void *src) {
912
  // With new compilers such as clang 8.0.0 we can use the new _mm_loadu_si32
913
  // intrinsic. Both _mm_loadu_si32(src) and the code here are compiled into a
914
  // movss instruction.
915
  //
916
  // Until compiler support of _mm_loadu_si32 is widespread, use of
917
  // _mm_loadu_si32 is banned.
918
491k
  int val;
919
491k
  memcpy(&val, src, sizeof(val));
920
491k
  return _mm_cvtsi32_si128(val);
921
491k
}
922
923
88.3M
static AOM_FORCE_INLINE __m128i LoadLo8(const void *a) {
924
88.3M
  return _mm_loadl_epi64((const __m128i *)(a));
925
88.3M
}
926
927
815k
static AOM_FORCE_INLINE __m128i LoadUnaligned16(const void *a) {
928
815k
  return _mm_loadu_si128((const __m128i *)(a));
929
815k
}
930
931
1.14M
static AOM_FORCE_INLINE void Store4(void *dst, const __m128i x) {
932
1.14M
  const int val = _mm_cvtsi128_si32(x);
933
1.14M
  memcpy(dst, &val, sizeof(val));
934
1.14M
}
935
936
45.5M
static AOM_FORCE_INLINE void StoreLo8(void *a, const __m128i v) {
937
45.5M
  _mm_storel_epi64((__m128i *)(a), v);
938
45.5M
}
939
940
13.6M
static AOM_FORCE_INLINE void StoreUnaligned16(void *a, const __m128i v) {
941
13.6M
  _mm_storeu_si128((__m128i *)(a), v);
942
13.6M
}
943
944
89.7M
static AOM_FORCE_INLINE __m128i cvtepu8_epi16(__m128i x) {
945
89.7M
  return _mm_unpacklo_epi8((x), _mm_setzero_si128());
946
89.7M
}
947
948
292k
static AOM_FORCE_INLINE __m128i cvtepu8_epi32(__m128i x) {
949
292k
  const __m128i tmp = _mm_unpacklo_epi8((x), _mm_setzero_si128());
950
292k
  return _mm_unpacklo_epi16(tmp, _mm_setzero_si128());
951
292k
}
952
953
43.7M
static AOM_FORCE_INLINE __m128i cvtepu16_epi32(__m128i x) {
954
43.7M
  return _mm_unpacklo_epi16((x), _mm_setzero_si128());
955
43.7M
}
956
957
static void smooth_predictor_wxh(uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
958
                                 const uint8_t *LIBAOM_RESTRICT top_row,
959
                                 const uint8_t *LIBAOM_RESTRICT left_column,
960
505k
                                 int width, int height) {
961
505k
  const uint8_t *const sm_weights_h = smooth_weights + height - 4;
962
505k
  const uint8_t *const sm_weights_w = smooth_weights + width - 4;
963
505k
  const __m128i zero = _mm_setzero_si128();
964
505k
  const __m128i scale_value = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
965
505k
  const __m128i bottom_left = _mm_cvtsi32_si128(left_column[height - 1]);
966
505k
  const __m128i top_right = _mm_set1_epi16(top_row[width - 1]);
967
505k
  const __m128i round = _mm_set1_epi32(1 << SMOOTH_WEIGHT_LOG2_SCALE);
968
10.8M
  for (int y = 0; y < height; ++y) {
969
10.3M
    const __m128i weights_y = _mm_cvtsi32_si128(sm_weights_h[y]);
970
10.3M
    const __m128i left_y = _mm_cvtsi32_si128(left_column[y]);
971
10.3M
    const __m128i scale_m_weights_y = _mm_sub_epi16(scale_value, weights_y);
972
10.3M
    __m128i scaled_bottom_left =
973
10.3M
        _mm_mullo_epi16(scale_m_weights_y, bottom_left);
974
10.3M
    const __m128i weight_left_y =
975
10.3M
        _mm_shuffle_epi32(_mm_unpacklo_epi16(weights_y, left_y), 0);
976
10.3M
    scaled_bottom_left = _mm_add_epi32(scaled_bottom_left, round);
977
10.3M
    scaled_bottom_left = _mm_shuffle_epi32(scaled_bottom_left, 0);
978
54.1M
    for (int x = 0; x < width; x += 8) {
979
43.7M
      const __m128i top_x = LoadLo8(top_row + x);
980
43.7M
      const __m128i weights_x = LoadLo8(sm_weights_w + x);
981
43.7M
      const __m128i top_weights_x = _mm_unpacklo_epi8(top_x, weights_x);
982
43.7M
      const __m128i top_weights_x_lo = cvtepu8_epi16(top_weights_x);
983
43.7M
      const __m128i top_weights_x_hi = _mm_unpackhi_epi8(top_weights_x, zero);
984
985
      // Here opposite weights and pixels are multiplied, where the order of
986
      // interleaving is indicated in the names.
987
43.7M
      __m128i pred_lo = _mm_madd_epi16(top_weights_x_lo, weight_left_y);
988
43.7M
      __m128i pred_hi = _mm_madd_epi16(top_weights_x_hi, weight_left_y);
989
990
      // |scaled_bottom_left| is always scaled by the same weight each row, so
991
      // we only derive |scaled_top_right| values here.
992
43.7M
      const __m128i inverted_weights_x =
993
43.7M
          _mm_sub_epi16(scale_value, cvtepu8_epi16(weights_x));
994
43.7M
      const __m128i scaled_top_right =
995
43.7M
          _mm_mullo_epi16(inverted_weights_x, top_right);
996
43.7M
      const __m128i scaled_top_right_lo = cvtepu16_epi32(scaled_top_right);
997
43.7M
      const __m128i scaled_top_right_hi =
998
43.7M
          _mm_unpackhi_epi16(scaled_top_right, zero);
999
43.7M
      pred_lo = _mm_add_epi32(pred_lo, scaled_bottom_left);
1000
43.7M
      pred_hi = _mm_add_epi32(pred_hi, scaled_bottom_left);
1001
43.7M
      pred_lo = _mm_add_epi32(pred_lo, scaled_top_right_lo);
1002
43.7M
      pred_hi = _mm_add_epi32(pred_hi, scaled_top_right_hi);
1003
1004
      // The round value for RightShiftWithRounding was added with
1005
      // |scaled_bottom_left|.
1006
43.7M
      pred_lo = _mm_srli_epi32(pred_lo, (1 + SMOOTH_WEIGHT_LOG2_SCALE));
1007
43.7M
      pred_hi = _mm_srli_epi32(pred_hi, (1 + SMOOTH_WEIGHT_LOG2_SCALE));
1008
43.7M
      const __m128i pred = _mm_packus_epi16(pred_lo, pred_hi);
1009
43.7M
      StoreLo8(dst + x, _mm_packus_epi16(pred, pred));
1010
43.7M
    }
1011
10.3M
    dst += stride;
1012
10.3M
  }
1013
505k
}
1014
1015
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1016
void aom_smooth_predictor_16x4_ssse3(uint8_t *dst, ptrdiff_t stride,
1017
                                     const uint8_t *above,
1018
62.4k
                                     const uint8_t *left) {
1019
62.4k
  smooth_predictor_wxh(dst, stride, above, left, 16, 4);
1020
62.4k
}
1021
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1022
1023
void aom_smooth_predictor_16x8_ssse3(uint8_t *dst, ptrdiff_t stride,
1024
                                     const uint8_t *above,
1025
66.8k
                                     const uint8_t *left) {
1026
66.8k
  smooth_predictor_wxh(dst, stride, above, left, 16, 8);
1027
66.8k
}
1028
1029
void aom_smooth_predictor_16x16_ssse3(uint8_t *dst, ptrdiff_t stride,
1030
                                      const uint8_t *above,
1031
110k
                                      const uint8_t *left) {
1032
110k
  smooth_predictor_wxh(dst, stride, above, left, 16, 16);
1033
110k
}
1034
1035
void aom_smooth_predictor_16x32_ssse3(uint8_t *dst, ptrdiff_t stride,
1036
                                      const uint8_t *above,
1037
28.1k
                                      const uint8_t *left) {
1038
28.1k
  smooth_predictor_wxh(dst, stride, above, left, 16, 32);
1039
28.1k
}
1040
1041
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1042
void aom_smooth_predictor_16x64_ssse3(uint8_t *dst, ptrdiff_t stride,
1043
                                      const uint8_t *above,
1044
5.19k
                                      const uint8_t *left) {
1045
5.19k
  smooth_predictor_wxh(dst, stride, above, left, 16, 64);
1046
5.19k
}
1047
1048
void aom_smooth_predictor_32x8_ssse3(uint8_t *dst, ptrdiff_t stride,
1049
                                     const uint8_t *above,
1050
43.9k
                                     const uint8_t *left) {
1051
43.9k
  smooth_predictor_wxh(dst, stride, above, left, 32, 8);
1052
43.9k
}
1053
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1054
1055
void aom_smooth_predictor_32x16_ssse3(uint8_t *dst, ptrdiff_t stride,
1056
                                      const uint8_t *above,
1057
28.1k
                                      const uint8_t *left) {
1058
28.1k
  smooth_predictor_wxh(dst, stride, above, left, 32, 16);
1059
28.1k
}
1060
1061
void aom_smooth_predictor_32x32_ssse3(uint8_t *dst, ptrdiff_t stride,
1062
                                      const uint8_t *above,
1063
97.1k
                                      const uint8_t *left) {
1064
97.1k
  smooth_predictor_wxh(dst, stride, above, left, 32, 32);
1065
97.1k
}
1066
1067
void aom_smooth_predictor_32x64_ssse3(uint8_t *dst, ptrdiff_t stride,
1068
                                      const uint8_t *above,
1069
2.88k
                                      const uint8_t *left) {
1070
2.88k
  smooth_predictor_wxh(dst, stride, above, left, 32, 64);
1071
2.88k
}
1072
1073
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1074
void aom_smooth_predictor_64x16_ssse3(uint8_t *dst, ptrdiff_t stride,
1075
                                      const uint8_t *above,
1076
25.1k
                                      const uint8_t *left) {
1077
25.1k
  smooth_predictor_wxh(dst, stride, above, left, 64, 16);
1078
25.1k
}
1079
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1080
1081
void aom_smooth_predictor_64x32_ssse3(uint8_t *dst, ptrdiff_t stride,
1082
                                      const uint8_t *above,
1083
5.39k
                                      const uint8_t *left) {
1084
5.39k
  smooth_predictor_wxh(dst, stride, above, left, 64, 32);
1085
5.39k
}
1086
1087
void aom_smooth_predictor_64x64_ssse3(uint8_t *dst, ptrdiff_t stride,
1088
                                      const uint8_t *above,
1089
29.7k
                                      const uint8_t *left) {
1090
29.7k
  smooth_predictor_wxh(dst, stride, above, left, 64, 64);
1091
29.7k
}
1092
1093
// -----------------------------------------------------------------------------
1094
// Smooth horizontal/vertical helper functions.
1095
1096
// For Horizontal, pixels1 and pixels2 are the same repeated value. For
1097
// Vertical, weights1 and weights2 are the same, and scaled_corner1 and
1098
// scaled_corner2 are the same.
1099
static AOM_FORCE_INLINE void write_smooth_directional_sum16(
1100
    uint8_t *LIBAOM_RESTRICT dst, const __m128i pixels1, const __m128i pixels2,
1101
    const __m128i weights1, const __m128i weights2,
1102
    const __m128i scaled_corner1, const __m128i scaled_corner2,
1103
13.6M
    const __m128i round) {
1104
13.6M
  const __m128i weighted_px1 = _mm_mullo_epi16(pixels1, weights1);
1105
13.6M
  const __m128i weighted_px2 = _mm_mullo_epi16(pixels2, weights2);
1106
13.6M
  const __m128i pred_sum1 = _mm_add_epi16(scaled_corner1, weighted_px1);
1107
13.6M
  const __m128i pred_sum2 = _mm_add_epi16(scaled_corner2, weighted_px2);
1108
  // Equivalent to RightShiftWithRounding(pred[x][y], 8).
1109
13.6M
  const __m128i pred1 = _mm_srli_epi16(_mm_add_epi16(pred_sum1, round), 8);
1110
13.6M
  const __m128i pred2 = _mm_srli_epi16(_mm_add_epi16(pred_sum2, round), 8);
1111
13.6M
  StoreUnaligned16(dst, _mm_packus_epi16(pred1, pred2));
1112
13.6M
}
1113
1114
static AOM_FORCE_INLINE __m128i smooth_directional_sum8(
1115
1.79M
    const __m128i pixels, const __m128i weights, const __m128i scaled_corner) {
1116
1.79M
  const __m128i weighted_px = _mm_mullo_epi16(pixels, weights);
1117
1.79M
  return _mm_add_epi16(scaled_corner, weighted_px);
1118
1.79M
}
1119
1120
static AOM_FORCE_INLINE void write_smooth_directional_sum8(
1121
    uint8_t *LIBAOM_RESTRICT dst, const __m128i *pixels, const __m128i *weights,
1122
1.79M
    const __m128i *scaled_corner, const __m128i *round) {
1123
1.79M
  const __m128i pred_sum =
1124
1.79M
      smooth_directional_sum8(*pixels, *weights, *scaled_corner);
1125
  // Equivalent to RightShiftWithRounding(pred[x][y], 8).
1126
1.79M
  const __m128i pred = _mm_srli_epi16(_mm_add_epi16(pred_sum, *round), 8);
1127
1.79M
  StoreLo8(dst, _mm_packus_epi16(pred, pred));
1128
1.79M
}
1129
1130
// -----------------------------------------------------------------------------
1131
// SMOOTH_V_PRED
1132
1133
static AOM_FORCE_INLINE void load_smooth_vertical_pixels4(
1134
    const uint8_t *LIBAOM_RESTRICT above, const uint8_t *LIBAOM_RESTRICT left,
1135
62.4k
    const int height, __m128i *pixels) {
1136
62.4k
  __m128i top = Load4(above);
1137
62.4k
  const __m128i bottom_left = _mm_set1_epi16(left[height - 1]);
1138
62.4k
  top = cvtepu8_epi16(top);
1139
62.4k
  pixels[0] = _mm_unpacklo_epi16(top, bottom_left);
1140
62.4k
}
1141
1142
// |weight_array| alternates weight vectors from the table with their inverted
1143
// (256-w) counterparts. This is precomputed by the compiler when the weights
1144
// table is visible to this module. Removing this visibility can cut speed by up
1145
// to half in both 4xH and 8xH transforms.
1146
static AOM_FORCE_INLINE void load_smooth_vertical_weights4(
1147
    const uint8_t *LIBAOM_RESTRICT weight_array, const int height,
1148
62.4k
    __m128i *weights) {
1149
62.4k
  const __m128i inverter = _mm_set1_epi16(256);
1150
1151
62.4k
  if (height == 4) {
1152
39.3k
    const __m128i weight = Load4(weight_array);
1153
39.3k
    weights[0] = cvtepu8_epi16(weight);
1154
39.3k
    weights[1] = _mm_sub_epi16(inverter, weights[0]);
1155
39.3k
  } else if (height == 8) {
1156
13.4k
    const __m128i weight = LoadLo8(weight_array + 4);
1157
13.4k
    weights[0] = cvtepu8_epi16(weight);
1158
13.4k
    weights[1] = _mm_sub_epi16(inverter, weights[0]);
1159
13.4k
  } else {
1160
9.59k
    const __m128i weight = LoadUnaligned16(weight_array + 12);
1161
9.59k
    const __m128i zero = _mm_setzero_si128();
1162
9.59k
    weights[0] = cvtepu8_epi16(weight);
1163
9.59k
    weights[1] = _mm_sub_epi16(inverter, weights[0]);
1164
9.59k
    weights[2] = _mm_unpackhi_epi8(weight, zero);
1165
9.59k
    weights[3] = _mm_sub_epi16(inverter, weights[2]);
1166
9.59k
  }
1167
62.4k
}
1168
1169
static AOM_FORCE_INLINE void write_smooth_vertical4xh(
1170
    const __m128i *pixel, const __m128i *weight, const int height,
1171
72.0k
    uint8_t *LIBAOM_RESTRICT dst, const ptrdiff_t stride) {
1172
72.0k
  const __m128i pred_round = _mm_set1_epi32(128);
1173
72.0k
  const __m128i mask_increment = _mm_set1_epi16(0x0202);
1174
72.0k
  const __m128i cvtepu8_epi32 = _mm_set1_epi32(0xC080400);
1175
72.0k
  __m128i y_select = _mm_set1_epi16(0x0100);
1176
1177
490k
  for (int y = 0; y < height; ++y) {
1178
418k
    const __m128i weight_y = _mm_shuffle_epi8(weight[0], y_select);
1179
418k
    const __m128i inverted_weight_y = _mm_shuffle_epi8(weight[1], y_select);
1180
418k
    const __m128i alternate_weights =
1181
418k
        _mm_unpacklo_epi16(weight_y, inverted_weight_y);
1182
    // Here the pixel vector is top_row[0], corner, top_row[1], corner, ...
1183
    // The madd instruction yields four results of the form:
1184
    // (top_row[x] * weight[y] + corner * inverted_weight[y])
1185
418k
    __m128i sum = _mm_madd_epi16(pixel[0], alternate_weights);
1186
418k
    sum = _mm_add_epi32(sum, pred_round);
1187
418k
    sum = _mm_srai_epi32(sum, 8);
1188
418k
    sum = _mm_shuffle_epi8(sum, cvtepu8_epi32);
1189
418k
    Store4(dst, sum);
1190
418k
    dst += stride;
1191
418k
    y_select = _mm_add_epi16(y_select, mask_increment);
1192
418k
  }
1193
72.0k
}
1194
1195
void aom_smooth_v_predictor_4x4_ssse3(
1196
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1197
    const uint8_t *LIBAOM_RESTRICT top_row,
1198
39.3k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1199
39.3k
  __m128i pixels;
1200
39.3k
  load_smooth_vertical_pixels4(top_row, left_column, 4, &pixels);
1201
1202
39.3k
  __m128i weights[2];
1203
39.3k
  load_smooth_vertical_weights4(smooth_weights, 4, weights);
1204
1205
39.3k
  write_smooth_vertical4xh(&pixels, weights, 4, dst, stride);
1206
39.3k
}
1207
1208
void aom_smooth_v_predictor_4x8_ssse3(
1209
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1210
    const uint8_t *LIBAOM_RESTRICT top_row,
1211
13.4k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1212
13.4k
  __m128i pixels;
1213
13.4k
  load_smooth_vertical_pixels4(top_row, left_column, 8, &pixels);
1214
1215
13.4k
  __m128i weights[2];
1216
13.4k
  load_smooth_vertical_weights4(smooth_weights, 8, weights);
1217
1218
13.4k
  write_smooth_vertical4xh(&pixels, weights, 8, dst, stride);
1219
13.4k
}
1220
1221
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1222
void aom_smooth_v_predictor_4x16_ssse3(
1223
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1224
    const uint8_t *LIBAOM_RESTRICT top_row,
1225
9.59k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1226
9.59k
  __m128i pixels;
1227
9.59k
  load_smooth_vertical_pixels4(top_row, left_column, 16, &pixels);
1228
1229
9.59k
  __m128i weights[4];
1230
9.59k
  load_smooth_vertical_weights4(smooth_weights, 16, weights);
1231
1232
9.59k
  write_smooth_vertical4xh(&pixels, weights, 8, dst, stride);
1233
9.59k
  dst += stride << 3;
1234
9.59k
  write_smooth_vertical4xh(&pixels, &weights[2], 8, dst, stride);
1235
9.59k
}
1236
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1237
1238
void aom_smooth_v_predictor_8x4_ssse3(
1239
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1240
    const uint8_t *LIBAOM_RESTRICT top_row,
1241
19.6k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1242
19.6k
  const __m128i bottom_left = _mm_set1_epi16(left_column[3]);
1243
19.6k
  const __m128i weights = cvtepu8_epi16(Load4(smooth_weights));
1244
19.6k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
1245
19.6k
  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
1246
19.6k
  const __m128i scaled_bottom_left =
1247
19.6k
      _mm_mullo_epi16(inverted_weights, bottom_left);
1248
19.6k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
1249
19.6k
  __m128i y_select = _mm_set1_epi32(0x01000100);
1250
19.6k
  const __m128i top = cvtepu8_epi16(LoadLo8(top_row));
1251
19.6k
  __m128i weights_y = _mm_shuffle_epi8(weights, y_select);
1252
19.6k
  __m128i scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
1253
19.6k
  write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
1254
19.6k
                                &round);
1255
19.6k
  dst += stride;
1256
19.6k
  y_select = _mm_set1_epi32(0x03020302);
1257
19.6k
  weights_y = _mm_shuffle_epi8(weights, y_select);
1258
19.6k
  scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
1259
19.6k
  write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
1260
19.6k
                                &round);
1261
19.6k
  dst += stride;
1262
19.6k
  y_select = _mm_set1_epi32(0x05040504);
1263
19.6k
  weights_y = _mm_shuffle_epi8(weights, y_select);
1264
19.6k
  scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
1265
19.6k
  write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
1266
19.6k
                                &round);
1267
19.6k
  dst += stride;
1268
19.6k
  y_select = _mm_set1_epi32(0x07060706);
1269
19.6k
  weights_y = _mm_shuffle_epi8(weights, y_select);
1270
19.6k
  scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
1271
19.6k
  write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
1272
19.6k
                                &round);
1273
19.6k
}
1274
1275
void aom_smooth_v_predictor_8x8_ssse3(
1276
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1277
    const uint8_t *LIBAOM_RESTRICT top_row,
1278
42.5k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1279
42.5k
  const __m128i bottom_left = _mm_set1_epi16(left_column[7]);
1280
42.5k
  const __m128i weights = cvtepu8_epi16(LoadLo8(smooth_weights + 4));
1281
42.5k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
1282
42.5k
  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
1283
42.5k
  const __m128i scaled_bottom_left =
1284
42.5k
      _mm_mullo_epi16(inverted_weights, bottom_left);
1285
42.5k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
1286
42.5k
  const __m128i top = cvtepu8_epi16(LoadLo8(top_row));
1287
383k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1288
340k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1289
340k
    const __m128i weights_y = _mm_shuffle_epi8(weights, y_select);
1290
340k
    const __m128i scaled_bottom_left_y =
1291
340k
        _mm_shuffle_epi8(scaled_bottom_left, y_select);
1292
340k
    write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
1293
340k
                                  &round);
1294
340k
    dst += stride;
1295
340k
  }
1296
42.5k
}
1297
1298
void aom_smooth_v_predictor_8x16_ssse3(
1299
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1300
    const uint8_t *LIBAOM_RESTRICT top_row,
1301
13.2k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1302
13.2k
  const __m128i bottom_left = _mm_set1_epi16(left_column[15]);
1303
13.2k
  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
1304
1305
13.2k
  const __m128i weights1 = cvtepu8_epi16(weights);
1306
13.2k
  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights, 8));
1307
13.2k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
1308
13.2k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
1309
13.2k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
1310
13.2k
  const __m128i scaled_bottom_left1 =
1311
13.2k
      _mm_mullo_epi16(inverted_weights1, bottom_left);
1312
13.2k
  const __m128i scaled_bottom_left2 =
1313
13.2k
      _mm_mullo_epi16(inverted_weights2, bottom_left);
1314
13.2k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
1315
13.2k
  const __m128i top = cvtepu8_epi16(LoadLo8(top_row));
1316
119k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1317
106k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1318
106k
    const __m128i weights_y = _mm_shuffle_epi8(weights1, y_select);
1319
106k
    const __m128i scaled_bottom_left_y =
1320
106k
        _mm_shuffle_epi8(scaled_bottom_left1, y_select);
1321
106k
    write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
1322
106k
                                  &round);
1323
106k
    dst += stride;
1324
106k
  }
1325
119k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1326
106k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1327
106k
    const __m128i weights_y = _mm_shuffle_epi8(weights2, y_select);
1328
106k
    const __m128i scaled_bottom_left_y =
1329
106k
        _mm_shuffle_epi8(scaled_bottom_left2, y_select);
1330
106k
    write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
1331
106k
                                  &round);
1332
106k
    dst += stride;
1333
106k
  }
1334
13.2k
}
1335
1336
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1337
void aom_smooth_v_predictor_8x32_ssse3(
1338
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1339
    const uint8_t *LIBAOM_RESTRICT top_row,
1340
4.73k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1341
4.73k
  const __m128i zero = _mm_setzero_si128();
1342
4.73k
  const __m128i bottom_left = _mm_set1_epi16(left_column[31]);
1343
4.73k
  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
1344
4.73k
  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
1345
4.73k
  const __m128i weights1 = cvtepu8_epi16(weights_lo);
1346
4.73k
  const __m128i weights2 = _mm_unpackhi_epi8(weights_lo, zero);
1347
4.73k
  const __m128i weights3 = cvtepu8_epi16(weights_hi);
1348
4.73k
  const __m128i weights4 = _mm_unpackhi_epi8(weights_hi, zero);
1349
4.73k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
1350
4.73k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
1351
4.73k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
1352
4.73k
  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
1353
4.73k
  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
1354
4.73k
  const __m128i scaled_bottom_left1 =
1355
4.73k
      _mm_mullo_epi16(inverted_weights1, bottom_left);
1356
4.73k
  const __m128i scaled_bottom_left2 =
1357
4.73k
      _mm_mullo_epi16(inverted_weights2, bottom_left);
1358
4.73k
  const __m128i scaled_bottom_left3 =
1359
4.73k
      _mm_mullo_epi16(inverted_weights3, bottom_left);
1360
4.73k
  const __m128i scaled_bottom_left4 =
1361
4.73k
      _mm_mullo_epi16(inverted_weights4, bottom_left);
1362
4.73k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
1363
4.73k
  const __m128i top = cvtepu8_epi16(LoadLo8(top_row));
1364
42.6k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1365
37.8k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1366
37.8k
    const __m128i weights_y = _mm_shuffle_epi8(weights1, y_select);
1367
37.8k
    const __m128i scaled_bottom_left_y =
1368
37.8k
        _mm_shuffle_epi8(scaled_bottom_left1, y_select);
1369
37.8k
    write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
1370
37.8k
                                  &round);
1371
37.8k
    dst += stride;
1372
37.8k
  }
1373
42.6k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1374
37.8k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1375
37.8k
    const __m128i weights_y = _mm_shuffle_epi8(weights2, y_select);
1376
37.8k
    const __m128i scaled_bottom_left_y =
1377
37.8k
        _mm_shuffle_epi8(scaled_bottom_left2, y_select);
1378
37.8k
    write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
1379
37.8k
                                  &round);
1380
37.8k
    dst += stride;
1381
37.8k
  }
1382
42.6k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1383
37.8k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1384
37.8k
    const __m128i weights_y = _mm_shuffle_epi8(weights3, y_select);
1385
37.8k
    const __m128i scaled_bottom_left_y =
1386
37.8k
        _mm_shuffle_epi8(scaled_bottom_left3, y_select);
1387
37.8k
    write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
1388
37.8k
                                  &round);
1389
37.8k
    dst += stride;
1390
37.8k
  }
1391
42.6k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1392
37.8k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1393
37.8k
    const __m128i weights_y = _mm_shuffle_epi8(weights4, y_select);
1394
37.8k
    const __m128i scaled_bottom_left_y =
1395
37.8k
        _mm_shuffle_epi8(scaled_bottom_left4, y_select);
1396
37.8k
    write_smooth_directional_sum8(dst, &top, &weights_y, &scaled_bottom_left_y,
1397
37.8k
                                  &round);
1398
37.8k
    dst += stride;
1399
37.8k
  }
1400
4.73k
}
1401
1402
void aom_smooth_v_predictor_16x4_ssse3(
1403
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1404
    const uint8_t *LIBAOM_RESTRICT top_row,
1405
18.3k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1406
18.3k
  const __m128i bottom_left = _mm_set1_epi16(left_column[3]);
1407
18.3k
  const __m128i weights = cvtepu8_epi16(Load4(smooth_weights));
1408
18.3k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
1409
18.3k
  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
1410
18.3k
  const __m128i scaled_bottom_left =
1411
18.3k
      _mm_mullo_epi16(inverted_weights, bottom_left);
1412
18.3k
  const __m128i round = _mm_set1_epi16(128);
1413
18.3k
  const __m128i top = LoadUnaligned16(top_row);
1414
18.3k
  const __m128i top_lo = cvtepu8_epi16(top);
1415
18.3k
  const __m128i top_hi = cvtepu8_epi16(_mm_srli_si128(top, 8));
1416
1417
18.3k
  __m128i y_select = _mm_set1_epi32(0x01000100);
1418
18.3k
  __m128i weights_y = _mm_shuffle_epi8(weights, y_select);
1419
18.3k
  __m128i scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
1420
18.3k
  write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
1421
18.3k
                                 scaled_bottom_left_y, scaled_bottom_left_y,
1422
18.3k
                                 round);
1423
18.3k
  dst += stride;
1424
18.3k
  y_select = _mm_set1_epi32(0x03020302);
1425
18.3k
  weights_y = _mm_shuffle_epi8(weights, y_select);
1426
18.3k
  scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
1427
18.3k
  write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
1428
18.3k
                                 scaled_bottom_left_y, scaled_bottom_left_y,
1429
18.3k
                                 round);
1430
18.3k
  dst += stride;
1431
18.3k
  y_select = _mm_set1_epi32(0x05040504);
1432
18.3k
  weights_y = _mm_shuffle_epi8(weights, y_select);
1433
18.3k
  scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
1434
18.3k
  write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
1435
18.3k
                                 scaled_bottom_left_y, scaled_bottom_left_y,
1436
18.3k
                                 round);
1437
18.3k
  dst += stride;
1438
18.3k
  y_select = _mm_set1_epi32(0x07060706);
1439
18.3k
  weights_y = _mm_shuffle_epi8(weights, y_select);
1440
18.3k
  scaled_bottom_left_y = _mm_shuffle_epi8(scaled_bottom_left, y_select);
1441
18.3k
  write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
1442
18.3k
                                 scaled_bottom_left_y, scaled_bottom_left_y,
1443
18.3k
                                 round);
1444
18.3k
}
1445
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1446
1447
void aom_smooth_v_predictor_16x8_ssse3(
1448
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1449
    const uint8_t *LIBAOM_RESTRICT top_row,
1450
14.0k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1451
14.0k
  const __m128i bottom_left = _mm_set1_epi16(left_column[7]);
1452
14.0k
  const __m128i weights = cvtepu8_epi16(LoadLo8(smooth_weights + 4));
1453
14.0k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
1454
14.0k
  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
1455
14.0k
  const __m128i scaled_bottom_left =
1456
14.0k
      _mm_mullo_epi16(inverted_weights, bottom_left);
1457
14.0k
  const __m128i round = _mm_set1_epi16(128);
1458
14.0k
  const __m128i top = LoadUnaligned16(top_row);
1459
14.0k
  const __m128i top_lo = cvtepu8_epi16(top);
1460
14.0k
  const __m128i top_hi = cvtepu8_epi16(_mm_srli_si128(top, 8));
1461
126k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1462
112k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1463
112k
    const __m128i weights_y = _mm_shuffle_epi8(weights, y_select);
1464
112k
    const __m128i scaled_bottom_left_y =
1465
112k
        _mm_shuffle_epi8(scaled_bottom_left, y_select);
1466
112k
    write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
1467
112k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1468
112k
                                   round);
1469
112k
    dst += stride;
1470
112k
  }
1471
14.0k
}
1472
1473
void aom_smooth_v_predictor_16x16_ssse3(
1474
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1475
    const uint8_t *LIBAOM_RESTRICT top_row,
1476
33.2k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1477
33.2k
  const __m128i bottom_left = _mm_set1_epi16(left_column[15]);
1478
33.2k
  const __m128i zero = _mm_setzero_si128();
1479
33.2k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
1480
33.2k
  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
1481
33.2k
  const __m128i weights_lo = cvtepu8_epi16(weights);
1482
33.2k
  const __m128i weights_hi = _mm_unpackhi_epi8(weights, zero);
1483
33.2k
  const __m128i inverted_weights_lo = _mm_sub_epi16(scale, weights_lo);
1484
33.2k
  const __m128i inverted_weights_hi = _mm_sub_epi16(scale, weights_hi);
1485
33.2k
  const __m128i scaled_bottom_left_lo =
1486
33.2k
      _mm_mullo_epi16(inverted_weights_lo, bottom_left);
1487
33.2k
  const __m128i scaled_bottom_left_hi =
1488
33.2k
      _mm_mullo_epi16(inverted_weights_hi, bottom_left);
1489
33.2k
  const __m128i round = _mm_set1_epi16(128);
1490
1491
33.2k
  const __m128i top = LoadUnaligned16(top_row);
1492
33.2k
  const __m128i top_lo = cvtepu8_epi16(top);
1493
33.2k
  const __m128i top_hi = _mm_unpackhi_epi8(top, zero);
1494
299k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1495
266k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1496
266k
    const __m128i weights_y = _mm_shuffle_epi8(weights_lo, y_select);
1497
266k
    const __m128i scaled_bottom_left_y =
1498
266k
        _mm_shuffle_epi8(scaled_bottom_left_lo, y_select);
1499
266k
    write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
1500
266k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1501
266k
                                   round);
1502
266k
    dst += stride;
1503
266k
  }
1504
299k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1505
266k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1506
266k
    const __m128i weights_y = _mm_shuffle_epi8(weights_hi, y_select);
1507
266k
    const __m128i scaled_bottom_left_y =
1508
266k
        _mm_shuffle_epi8(scaled_bottom_left_hi, y_select);
1509
266k
    write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
1510
266k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1511
266k
                                   round);
1512
266k
    dst += stride;
1513
266k
  }
1514
33.2k
}
1515
1516
void aom_smooth_v_predictor_16x32_ssse3(
1517
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1518
    const uint8_t *LIBAOM_RESTRICT top_row,
1519
9.09k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1520
9.09k
  const __m128i bottom_left = _mm_set1_epi16(left_column[31]);
1521
9.09k
  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
1522
9.09k
  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
1523
9.09k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
1524
9.09k
  const __m128i zero = _mm_setzero_si128();
1525
9.09k
  const __m128i weights1 = cvtepu8_epi16(weights_lo);
1526
9.09k
  const __m128i weights2 = _mm_unpackhi_epi8(weights_lo, zero);
1527
9.09k
  const __m128i weights3 = cvtepu8_epi16(weights_hi);
1528
9.09k
  const __m128i weights4 = _mm_unpackhi_epi8(weights_hi, zero);
1529
9.09k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
1530
9.09k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
1531
9.09k
  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
1532
9.09k
  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
1533
9.09k
  const __m128i scaled_bottom_left1 =
1534
9.09k
      _mm_mullo_epi16(inverted_weights1, bottom_left);
1535
9.09k
  const __m128i scaled_bottom_left2 =
1536
9.09k
      _mm_mullo_epi16(inverted_weights2, bottom_left);
1537
9.09k
  const __m128i scaled_bottom_left3 =
1538
9.09k
      _mm_mullo_epi16(inverted_weights3, bottom_left);
1539
9.09k
  const __m128i scaled_bottom_left4 =
1540
9.09k
      _mm_mullo_epi16(inverted_weights4, bottom_left);
1541
9.09k
  const __m128i round = _mm_set1_epi16(128);
1542
1543
9.09k
  const __m128i top = LoadUnaligned16(top_row);
1544
9.09k
  const __m128i top_lo = cvtepu8_epi16(top);
1545
9.09k
  const __m128i top_hi = _mm_unpackhi_epi8(top, zero);
1546
81.8k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1547
72.7k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1548
72.7k
    const __m128i weights_y = _mm_shuffle_epi8(weights1, y_select);
1549
72.7k
    const __m128i scaled_bottom_left_y =
1550
72.7k
        _mm_shuffle_epi8(scaled_bottom_left1, y_select);
1551
72.7k
    write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
1552
72.7k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1553
72.7k
                                   round);
1554
72.7k
    dst += stride;
1555
72.7k
  }
1556
81.8k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1557
72.7k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1558
72.7k
    const __m128i weights_y = _mm_shuffle_epi8(weights2, y_select);
1559
72.7k
    const __m128i scaled_bottom_left_y =
1560
72.7k
        _mm_shuffle_epi8(scaled_bottom_left2, y_select);
1561
72.7k
    write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
1562
72.7k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1563
72.7k
                                   round);
1564
72.7k
    dst += stride;
1565
72.7k
  }
1566
81.8k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1567
72.7k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1568
72.7k
    const __m128i weights_y = _mm_shuffle_epi8(weights3, y_select);
1569
72.7k
    const __m128i scaled_bottom_left_y =
1570
72.7k
        _mm_shuffle_epi8(scaled_bottom_left3, y_select);
1571
72.7k
    write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
1572
72.7k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1573
72.7k
                                   round);
1574
72.7k
    dst += stride;
1575
72.7k
  }
1576
81.8k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1577
72.7k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1578
72.7k
    const __m128i weights_y = _mm_shuffle_epi8(weights4, y_select);
1579
72.7k
    const __m128i scaled_bottom_left_y =
1580
72.7k
        _mm_shuffle_epi8(scaled_bottom_left4, y_select);
1581
72.7k
    write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
1582
72.7k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1583
72.7k
                                   round);
1584
72.7k
    dst += stride;
1585
72.7k
  }
1586
9.09k
}
1587
1588
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1589
void aom_smooth_v_predictor_16x64_ssse3(
1590
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1591
    const uint8_t *LIBAOM_RESTRICT top_row,
1592
1.77k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1593
1.77k
  const __m128i bottom_left = _mm_set1_epi16(left_column[63]);
1594
1.77k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
1595
1.77k
  const __m128i round = _mm_set1_epi16(128);
1596
1.77k
  const __m128i zero = _mm_setzero_si128();
1597
1.77k
  const __m128i top = LoadUnaligned16(top_row);
1598
1.77k
  const __m128i top_lo = cvtepu8_epi16(top);
1599
1.77k
  const __m128i top_hi = _mm_unpackhi_epi8(top, zero);
1600
1.77k
  const uint8_t *weights_base_ptr = smooth_weights + 60;
1601
8.86k
  for (int left_offset = 0; left_offset < 64; left_offset += 16) {
1602
7.09k
    const __m128i weights = LoadUnaligned16(weights_base_ptr + left_offset);
1603
7.09k
    const __m128i weights_lo = cvtepu8_epi16(weights);
1604
7.09k
    const __m128i weights_hi = _mm_unpackhi_epi8(weights, zero);
1605
7.09k
    const __m128i inverted_weights_lo = _mm_sub_epi16(scale, weights_lo);
1606
7.09k
    const __m128i inverted_weights_hi = _mm_sub_epi16(scale, weights_hi);
1607
7.09k
    const __m128i scaled_bottom_left_lo =
1608
7.09k
        _mm_mullo_epi16(inverted_weights_lo, bottom_left);
1609
7.09k
    const __m128i scaled_bottom_left_hi =
1610
7.09k
        _mm_mullo_epi16(inverted_weights_hi, bottom_left);
1611
1612
63.8k
    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1613
56.7k
      const __m128i y_select = _mm_set1_epi32(y_mask);
1614
56.7k
      const __m128i weights_y = _mm_shuffle_epi8(weights_lo, y_select);
1615
56.7k
      const __m128i scaled_bottom_left_y =
1616
56.7k
          _mm_shuffle_epi8(scaled_bottom_left_lo, y_select);
1617
56.7k
      write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
1618
56.7k
                                     scaled_bottom_left_y, scaled_bottom_left_y,
1619
56.7k
                                     round);
1620
56.7k
      dst += stride;
1621
56.7k
    }
1622
63.8k
    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1623
56.7k
      const __m128i y_select = _mm_set1_epi32(y_mask);
1624
56.7k
      const __m128i weights_y = _mm_shuffle_epi8(weights_hi, y_select);
1625
56.7k
      const __m128i scaled_bottom_left_y =
1626
56.7k
          _mm_shuffle_epi8(scaled_bottom_left_hi, y_select);
1627
56.7k
      write_smooth_directional_sum16(dst, top_lo, top_hi, weights_y, weights_y,
1628
56.7k
                                     scaled_bottom_left_y, scaled_bottom_left_y,
1629
56.7k
                                     round);
1630
56.7k
      dst += stride;
1631
56.7k
    }
1632
7.09k
  }
1633
1.77k
}
1634
1635
void aom_smooth_v_predictor_32x8_ssse3(
1636
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1637
    const uint8_t *LIBAOM_RESTRICT top_row,
1638
18.8k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1639
18.8k
  const __m128i zero = _mm_setzero_si128();
1640
18.8k
  const __m128i bottom_left = _mm_set1_epi16(left_column[7]);
1641
18.8k
  const __m128i top_lo = LoadUnaligned16(top_row);
1642
18.8k
  const __m128i top_hi = LoadUnaligned16(top_row + 16);
1643
18.8k
  const __m128i top1 = cvtepu8_epi16(top_lo);
1644
18.8k
  const __m128i top2 = _mm_unpackhi_epi8(top_lo, zero);
1645
18.8k
  const __m128i top3 = cvtepu8_epi16(top_hi);
1646
18.8k
  const __m128i top4 = _mm_unpackhi_epi8(top_hi, zero);
1647
18.8k
  __m128i scale = _mm_set1_epi16(256);
1648
18.8k
  const __m128i weights = cvtepu8_epi16(LoadLo8(smooth_weights + 4));
1649
18.8k
  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
1650
18.8k
  const __m128i scaled_bottom_left =
1651
18.8k
      _mm_mullo_epi16(inverted_weights, bottom_left);
1652
18.8k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
1653
169k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1654
150k
    __m128i y_select = _mm_set1_epi32(y_mask);
1655
150k
    const __m128i weights_y = _mm_shuffle_epi8(weights, y_select);
1656
150k
    const __m128i scaled_bottom_left_y =
1657
150k
        _mm_shuffle_epi8(scaled_bottom_left, y_select);
1658
150k
    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
1659
150k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1660
150k
                                   round);
1661
150k
    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
1662
150k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1663
150k
                                   round);
1664
150k
    dst += stride;
1665
150k
  }
1666
18.8k
}
1667
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1668
1669
void aom_smooth_v_predictor_32x16_ssse3(
1670
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1671
    const uint8_t *LIBAOM_RESTRICT top_row,
1672
9.13k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1673
9.13k
  const __m128i zero = _mm_setzero_si128();
1674
9.13k
  const __m128i bottom_left = _mm_set1_epi16(left_column[15]);
1675
9.13k
  const __m128i top_lo = LoadUnaligned16(top_row);
1676
9.13k
  const __m128i top_hi = LoadUnaligned16(top_row + 16);
1677
9.13k
  const __m128i top1 = cvtepu8_epi16(top_lo);
1678
9.13k
  const __m128i top2 = _mm_unpackhi_epi8(top_lo, zero);
1679
9.13k
  const __m128i top3 = cvtepu8_epi16(top_hi);
1680
9.13k
  const __m128i top4 = _mm_unpackhi_epi8(top_hi, zero);
1681
9.13k
  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
1682
9.13k
  const __m128i weights1 = cvtepu8_epi16(weights);
1683
9.13k
  const __m128i weights2 = _mm_unpackhi_epi8(weights, zero);
1684
9.13k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
1685
9.13k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
1686
9.13k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
1687
9.13k
  const __m128i scaled_bottom_left1 =
1688
9.13k
      _mm_mullo_epi16(inverted_weights1, bottom_left);
1689
9.13k
  const __m128i scaled_bottom_left2 =
1690
9.13k
      _mm_mullo_epi16(inverted_weights2, bottom_left);
1691
9.13k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
1692
82.1k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1693
73.0k
    __m128i y_select = _mm_set1_epi32(y_mask);
1694
73.0k
    const __m128i weights_y = _mm_shuffle_epi8(weights1, y_select);
1695
73.0k
    const __m128i scaled_bottom_left_y =
1696
73.0k
        _mm_shuffle_epi8(scaled_bottom_left1, y_select);
1697
73.0k
    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
1698
73.0k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1699
73.0k
                                   round);
1700
73.0k
    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
1701
73.0k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1702
73.0k
                                   round);
1703
73.0k
    dst += stride;
1704
73.0k
  }
1705
82.1k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1706
73.0k
    __m128i y_select = _mm_set1_epi32(y_mask);
1707
73.0k
    const __m128i weights_y = _mm_shuffle_epi8(weights2, y_select);
1708
73.0k
    const __m128i scaled_bottom_left_y =
1709
73.0k
        _mm_shuffle_epi8(scaled_bottom_left2, y_select);
1710
73.0k
    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
1711
73.0k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1712
73.0k
                                   round);
1713
73.0k
    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
1714
73.0k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1715
73.0k
                                   round);
1716
73.0k
    dst += stride;
1717
73.0k
  }
1718
9.13k
}
1719
1720
void aom_smooth_v_predictor_32x32_ssse3(
1721
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1722
    const uint8_t *LIBAOM_RESTRICT top_row,
1723
39.3k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1724
39.3k
  const __m128i bottom_left = _mm_set1_epi16(left_column[31]);
1725
39.3k
  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
1726
39.3k
  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
1727
39.3k
  const __m128i zero = _mm_setzero_si128();
1728
39.3k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
1729
39.3k
  const __m128i top_lo = LoadUnaligned16(top_row);
1730
39.3k
  const __m128i top_hi = LoadUnaligned16(top_row + 16);
1731
39.3k
  const __m128i top1 = cvtepu8_epi16(top_lo);
1732
39.3k
  const __m128i top2 = _mm_unpackhi_epi8(top_lo, zero);
1733
39.3k
  const __m128i top3 = cvtepu8_epi16(top_hi);
1734
39.3k
  const __m128i top4 = _mm_unpackhi_epi8(top_hi, zero);
1735
39.3k
  const __m128i weights1 = cvtepu8_epi16(weights_lo);
1736
39.3k
  const __m128i weights2 = _mm_unpackhi_epi8(weights_lo, zero);
1737
39.3k
  const __m128i weights3 = cvtepu8_epi16(weights_hi);
1738
39.3k
  const __m128i weights4 = _mm_unpackhi_epi8(weights_hi, zero);
1739
39.3k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
1740
39.3k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
1741
39.3k
  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
1742
39.3k
  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
1743
39.3k
  const __m128i scaled_bottom_left1 =
1744
39.3k
      _mm_mullo_epi16(inverted_weights1, bottom_left);
1745
39.3k
  const __m128i scaled_bottom_left2 =
1746
39.3k
      _mm_mullo_epi16(inverted_weights2, bottom_left);
1747
39.3k
  const __m128i scaled_bottom_left3 =
1748
39.3k
      _mm_mullo_epi16(inverted_weights3, bottom_left);
1749
39.3k
  const __m128i scaled_bottom_left4 =
1750
39.3k
      _mm_mullo_epi16(inverted_weights4, bottom_left);
1751
39.3k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
1752
353k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1753
314k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1754
314k
    const __m128i weights_y = _mm_shuffle_epi8(weights1, y_select);
1755
314k
    const __m128i scaled_bottom_left_y =
1756
314k
        _mm_shuffle_epi8(scaled_bottom_left1, y_select);
1757
314k
    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
1758
314k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1759
314k
                                   round);
1760
314k
    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
1761
314k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1762
314k
                                   round);
1763
314k
    dst += stride;
1764
314k
  }
1765
353k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1766
314k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1767
314k
    const __m128i weights_y = _mm_shuffle_epi8(weights2, y_select);
1768
314k
    const __m128i scaled_bottom_left_y =
1769
314k
        _mm_shuffle_epi8(scaled_bottom_left2, y_select);
1770
314k
    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
1771
314k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1772
314k
                                   round);
1773
314k
    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
1774
314k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1775
314k
                                   round);
1776
314k
    dst += stride;
1777
314k
  }
1778
353k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1779
314k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1780
314k
    const __m128i weights_y = _mm_shuffle_epi8(weights3, y_select);
1781
314k
    const __m128i scaled_bottom_left_y =
1782
314k
        _mm_shuffle_epi8(scaled_bottom_left3, y_select);
1783
314k
    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
1784
314k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1785
314k
                                   round);
1786
314k
    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
1787
314k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1788
314k
                                   round);
1789
314k
    dst += stride;
1790
314k
  }
1791
353k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1792
314k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1793
314k
    const __m128i weights_y = _mm_shuffle_epi8(weights4, y_select);
1794
314k
    const __m128i scaled_bottom_left_y =
1795
314k
        _mm_shuffle_epi8(scaled_bottom_left4, y_select);
1796
314k
    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
1797
314k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1798
314k
                                   round);
1799
314k
    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
1800
314k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1801
314k
                                   round);
1802
314k
    dst += stride;
1803
314k
  }
1804
39.3k
}
1805
1806
void aom_smooth_v_predictor_32x64_ssse3(
1807
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1808
    const uint8_t *LIBAOM_RESTRICT top_row,
1809
741
    const uint8_t *LIBAOM_RESTRICT left_column) {
1810
741
  const __m128i zero = _mm_setzero_si128();
1811
741
  const __m128i bottom_left = _mm_set1_epi16(left_column[63]);
1812
741
  const __m128i top_lo = LoadUnaligned16(top_row);
1813
741
  const __m128i top_hi = LoadUnaligned16(top_row + 16);
1814
741
  const __m128i top1 = cvtepu8_epi16(top_lo);
1815
741
  const __m128i top2 = _mm_unpackhi_epi8(top_lo, zero);
1816
741
  const __m128i top3 = cvtepu8_epi16(top_hi);
1817
741
  const __m128i top4 = _mm_unpackhi_epi8(top_hi, zero);
1818
741
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
1819
741
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
1820
741
  const uint8_t *weights_base_ptr = smooth_weights + 60;
1821
3.70k
  for (int left_offset = 0; left_offset < 64; left_offset += 16) {
1822
2.96k
    const __m128i weights = LoadUnaligned16(weights_base_ptr + left_offset);
1823
2.96k
    const __m128i weights_lo = cvtepu8_epi16(weights);
1824
2.96k
    const __m128i weights_hi = _mm_unpackhi_epi8(weights, zero);
1825
2.96k
    const __m128i inverted_weights_lo = _mm_sub_epi16(scale, weights_lo);
1826
2.96k
    const __m128i inverted_weights_hi = _mm_sub_epi16(scale, weights_hi);
1827
2.96k
    const __m128i scaled_bottom_left_lo =
1828
2.96k
        _mm_mullo_epi16(inverted_weights_lo, bottom_left);
1829
2.96k
    const __m128i scaled_bottom_left_hi =
1830
2.96k
        _mm_mullo_epi16(inverted_weights_hi, bottom_left);
1831
1832
26.6k
    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1833
23.7k
      const __m128i y_select = _mm_set1_epi32(y_mask);
1834
23.7k
      const __m128i weights_y = _mm_shuffle_epi8(weights_lo, y_select);
1835
23.7k
      const __m128i scaled_bottom_left_y =
1836
23.7k
          _mm_shuffle_epi8(scaled_bottom_left_lo, y_select);
1837
23.7k
      write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
1838
23.7k
                                     scaled_bottom_left_y, scaled_bottom_left_y,
1839
23.7k
                                     round);
1840
23.7k
      write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
1841
23.7k
                                     scaled_bottom_left_y, scaled_bottom_left_y,
1842
23.7k
                                     round);
1843
23.7k
      dst += stride;
1844
23.7k
    }
1845
26.6k
    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1846
23.7k
      const __m128i y_select = _mm_set1_epi32(y_mask);
1847
23.7k
      const __m128i weights_y = _mm_shuffle_epi8(weights_hi, y_select);
1848
23.7k
      const __m128i scaled_bottom_left_y =
1849
23.7k
          _mm_shuffle_epi8(scaled_bottom_left_hi, y_select);
1850
23.7k
      write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
1851
23.7k
                                     scaled_bottom_left_y, scaled_bottom_left_y,
1852
23.7k
                                     round);
1853
23.7k
      write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
1854
23.7k
                                     scaled_bottom_left_y, scaled_bottom_left_y,
1855
23.7k
                                     round);
1856
23.7k
      dst += stride;
1857
23.7k
    }
1858
2.96k
  }
1859
741
}
1860
1861
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1862
void aom_smooth_v_predictor_64x16_ssse3(
1863
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1864
    const uint8_t *LIBAOM_RESTRICT top_row,
1865
13.5k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1866
13.5k
  const __m128i bottom_left = _mm_set1_epi16(left_column[15]);
1867
13.5k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
1868
13.5k
  const __m128i zero = _mm_setzero_si128();
1869
13.5k
  const __m128i top_lolo = LoadUnaligned16(top_row);
1870
13.5k
  const __m128i top_lohi = LoadUnaligned16(top_row + 16);
1871
13.5k
  const __m128i top1 = cvtepu8_epi16(top_lolo);
1872
13.5k
  const __m128i top2 = _mm_unpackhi_epi8(top_lolo, zero);
1873
13.5k
  const __m128i top3 = cvtepu8_epi16(top_lohi);
1874
13.5k
  const __m128i top4 = _mm_unpackhi_epi8(top_lohi, zero);
1875
1876
13.5k
  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
1877
13.5k
  const __m128i weights1 = cvtepu8_epi16(weights);
1878
13.5k
  const __m128i weights2 = _mm_unpackhi_epi8(weights, zero);
1879
13.5k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
1880
13.5k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
1881
13.5k
  const __m128i top_hilo = LoadUnaligned16(top_row + 32);
1882
13.5k
  const __m128i top_hihi = LoadUnaligned16(top_row + 48);
1883
13.5k
  const __m128i top5 = cvtepu8_epi16(top_hilo);
1884
13.5k
  const __m128i top6 = _mm_unpackhi_epi8(top_hilo, zero);
1885
13.5k
  const __m128i top7 = cvtepu8_epi16(top_hihi);
1886
13.5k
  const __m128i top8 = _mm_unpackhi_epi8(top_hihi, zero);
1887
13.5k
  const __m128i scaled_bottom_left1 =
1888
13.5k
      _mm_mullo_epi16(inverted_weights1, bottom_left);
1889
13.5k
  const __m128i scaled_bottom_left2 =
1890
13.5k
      _mm_mullo_epi16(inverted_weights2, bottom_left);
1891
13.5k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
1892
121k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1893
108k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1894
108k
    const __m128i weights_y = _mm_shuffle_epi8(weights1, y_select);
1895
108k
    const __m128i scaled_bottom_left_y =
1896
108k
        _mm_shuffle_epi8(scaled_bottom_left1, y_select);
1897
108k
    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
1898
108k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1899
108k
                                   round);
1900
108k
    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
1901
108k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1902
108k
                                   round);
1903
108k
    write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
1904
108k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1905
108k
                                   round);
1906
108k
    write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
1907
108k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1908
108k
                                   round);
1909
108k
    dst += stride;
1910
108k
  }
1911
121k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1912
108k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1913
108k
    const __m128i weights_y = _mm_shuffle_epi8(weights2, y_select);
1914
108k
    const __m128i scaled_bottom_left_y =
1915
108k
        _mm_shuffle_epi8(scaled_bottom_left2, y_select);
1916
108k
    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
1917
108k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1918
108k
                                   round);
1919
108k
    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
1920
108k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1921
108k
                                   round);
1922
108k
    write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
1923
108k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1924
108k
                                   round);
1925
108k
    write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
1926
108k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1927
108k
                                   round);
1928
108k
    dst += stride;
1929
108k
  }
1930
13.5k
}
1931
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
1932
1933
void aom_smooth_v_predictor_64x32_ssse3(
1934
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
1935
    const uint8_t *LIBAOM_RESTRICT top_row,
1936
2.41k
    const uint8_t *LIBAOM_RESTRICT left_column) {
1937
2.41k
  const __m128i zero = _mm_setzero_si128();
1938
2.41k
  const __m128i bottom_left = _mm_set1_epi16(left_column[31]);
1939
2.41k
  const __m128i top_lolo = LoadUnaligned16(top_row);
1940
2.41k
  const __m128i top_lohi = LoadUnaligned16(top_row + 16);
1941
2.41k
  const __m128i top1 = cvtepu8_epi16(top_lolo);
1942
2.41k
  const __m128i top2 = _mm_unpackhi_epi8(top_lolo, zero);
1943
2.41k
  const __m128i top3 = cvtepu8_epi16(top_lohi);
1944
2.41k
  const __m128i top4 = _mm_unpackhi_epi8(top_lohi, zero);
1945
2.41k
  const __m128i top_hilo = LoadUnaligned16(top_row + 32);
1946
2.41k
  const __m128i top_hihi = LoadUnaligned16(top_row + 48);
1947
2.41k
  const __m128i top5 = cvtepu8_epi16(top_hilo);
1948
2.41k
  const __m128i top6 = _mm_unpackhi_epi8(top_hilo, zero);
1949
2.41k
  const __m128i top7 = cvtepu8_epi16(top_hihi);
1950
2.41k
  const __m128i top8 = _mm_unpackhi_epi8(top_hihi, zero);
1951
2.41k
  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
1952
2.41k
  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
1953
2.41k
  const __m128i weights1 = cvtepu8_epi16(weights_lo);
1954
2.41k
  const __m128i weights2 = _mm_unpackhi_epi8(weights_lo, zero);
1955
2.41k
  const __m128i weights3 = cvtepu8_epi16(weights_hi);
1956
2.41k
  const __m128i weights4 = _mm_unpackhi_epi8(weights_hi, zero);
1957
2.41k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
1958
2.41k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
1959
2.41k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
1960
2.41k
  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
1961
2.41k
  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
1962
2.41k
  const __m128i scaled_bottom_left1 =
1963
2.41k
      _mm_mullo_epi16(inverted_weights1, bottom_left);
1964
2.41k
  const __m128i scaled_bottom_left2 =
1965
2.41k
      _mm_mullo_epi16(inverted_weights2, bottom_left);
1966
2.41k
  const __m128i scaled_bottom_left3 =
1967
2.41k
      _mm_mullo_epi16(inverted_weights3, bottom_left);
1968
2.41k
  const __m128i scaled_bottom_left4 =
1969
2.41k
      _mm_mullo_epi16(inverted_weights4, bottom_left);
1970
2.41k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
1971
1972
21.7k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1973
19.3k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1974
19.3k
    const __m128i weights_y = _mm_shuffle_epi8(weights1, y_select);
1975
19.3k
    const __m128i scaled_bottom_left_y =
1976
19.3k
        _mm_shuffle_epi8(scaled_bottom_left1, y_select);
1977
19.3k
    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
1978
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1979
19.3k
                                   round);
1980
19.3k
    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
1981
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1982
19.3k
                                   round);
1983
19.3k
    write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
1984
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1985
19.3k
                                   round);
1986
19.3k
    write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
1987
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1988
19.3k
                                   round);
1989
19.3k
    dst += stride;
1990
19.3k
  }
1991
21.7k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
1992
19.3k
    const __m128i y_select = _mm_set1_epi32(y_mask);
1993
19.3k
    const __m128i weights_y = _mm_shuffle_epi8(weights2, y_select);
1994
19.3k
    const __m128i scaled_bottom_left_y =
1995
19.3k
        _mm_shuffle_epi8(scaled_bottom_left2, y_select);
1996
19.3k
    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
1997
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
1998
19.3k
                                   round);
1999
19.3k
    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
2000
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
2001
19.3k
                                   round);
2002
19.3k
    write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
2003
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
2004
19.3k
                                   round);
2005
19.3k
    write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
2006
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
2007
19.3k
                                   round);
2008
19.3k
    dst += stride;
2009
19.3k
  }
2010
21.7k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2011
19.3k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2012
19.3k
    const __m128i weights_y = _mm_shuffle_epi8(weights3, y_select);
2013
19.3k
    const __m128i scaled_bottom_left_y =
2014
19.3k
        _mm_shuffle_epi8(scaled_bottom_left3, y_select);
2015
19.3k
    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
2016
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
2017
19.3k
                                   round);
2018
19.3k
    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
2019
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
2020
19.3k
                                   round);
2021
19.3k
    write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
2022
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
2023
19.3k
                                   round);
2024
19.3k
    write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
2025
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
2026
19.3k
                                   round);
2027
19.3k
    dst += stride;
2028
19.3k
  }
2029
21.7k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2030
19.3k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2031
19.3k
    const __m128i weights_y = _mm_shuffle_epi8(weights4, y_select);
2032
19.3k
    const __m128i scaled_bottom_left_y =
2033
19.3k
        _mm_shuffle_epi8(scaled_bottom_left4, y_select);
2034
19.3k
    write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
2035
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
2036
19.3k
                                   round);
2037
19.3k
    write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
2038
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
2039
19.3k
                                   round);
2040
19.3k
    write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
2041
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
2042
19.3k
                                   round);
2043
19.3k
    write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
2044
19.3k
                                   scaled_bottom_left_y, scaled_bottom_left_y,
2045
19.3k
                                   round);
2046
19.3k
    dst += stride;
2047
19.3k
  }
2048
2.41k
}
2049
2050
void aom_smooth_v_predictor_64x64_ssse3(
2051
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2052
    const uint8_t *LIBAOM_RESTRICT top_row,
2053
5.79k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2054
5.79k
  const __m128i zero = _mm_setzero_si128();
2055
5.79k
  const __m128i bottom_left = _mm_set1_epi16(left_column[63]);
2056
5.79k
  const __m128i top_lolo = LoadUnaligned16(top_row);
2057
5.79k
  const __m128i top_lohi = LoadUnaligned16(top_row + 16);
2058
5.79k
  const __m128i top1 = cvtepu8_epi16(top_lolo);
2059
5.79k
  const __m128i top2 = _mm_unpackhi_epi8(top_lolo, zero);
2060
5.79k
  const __m128i top3 = cvtepu8_epi16(top_lohi);
2061
5.79k
  const __m128i top4 = _mm_unpackhi_epi8(top_lohi, zero);
2062
5.79k
  const __m128i top_hilo = LoadUnaligned16(top_row + 32);
2063
5.79k
  const __m128i top_hihi = LoadUnaligned16(top_row + 48);
2064
5.79k
  const __m128i top5 = cvtepu8_epi16(top_hilo);
2065
5.79k
  const __m128i top6 = _mm_unpackhi_epi8(top_hilo, zero);
2066
5.79k
  const __m128i top7 = cvtepu8_epi16(top_hihi);
2067
5.79k
  const __m128i top8 = _mm_unpackhi_epi8(top_hihi, zero);
2068
5.79k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2069
5.79k
  const __m128i round = _mm_set1_epi16(128);
2070
5.79k
  const uint8_t *weights_base_ptr = smooth_weights + 60;
2071
28.9k
  for (int left_offset = 0; left_offset < 64; left_offset += 16) {
2072
23.1k
    const __m128i weights = LoadUnaligned16(weights_base_ptr + left_offset);
2073
23.1k
    const __m128i weights_lo = cvtepu8_epi16(weights);
2074
23.1k
    const __m128i weights_hi = _mm_unpackhi_epi8(weights, zero);
2075
23.1k
    const __m128i inverted_weights_lo = _mm_sub_epi16(scale, weights_lo);
2076
23.1k
    const __m128i inverted_weights_hi = _mm_sub_epi16(scale, weights_hi);
2077
23.1k
    const __m128i scaled_bottom_left_lo =
2078
23.1k
        _mm_mullo_epi16(inverted_weights_lo, bottom_left);
2079
23.1k
    const __m128i scaled_bottom_left_hi =
2080
23.1k
        _mm_mullo_epi16(inverted_weights_hi, bottom_left);
2081
208k
    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2082
185k
      const __m128i y_select = _mm_set1_epi32(y_mask);
2083
185k
      const __m128i weights_y = _mm_shuffle_epi8(weights_lo, y_select);
2084
185k
      const __m128i scaled_bottom_left_y =
2085
185k
          _mm_shuffle_epi8(scaled_bottom_left_lo, y_select);
2086
185k
      write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
2087
185k
                                     scaled_bottom_left_y, scaled_bottom_left_y,
2088
185k
                                     round);
2089
185k
      write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
2090
185k
                                     scaled_bottom_left_y, scaled_bottom_left_y,
2091
185k
                                     round);
2092
185k
      write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
2093
185k
                                     scaled_bottom_left_y, scaled_bottom_left_y,
2094
185k
                                     round);
2095
185k
      write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
2096
185k
                                     scaled_bottom_left_y, scaled_bottom_left_y,
2097
185k
                                     round);
2098
185k
      dst += stride;
2099
185k
    }
2100
208k
    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2101
185k
      const __m128i y_select = _mm_set1_epi32(y_mask);
2102
185k
      const __m128i weights_y = _mm_shuffle_epi8(weights_hi, y_select);
2103
185k
      const __m128i scaled_bottom_left_y =
2104
185k
          _mm_shuffle_epi8(scaled_bottom_left_hi, y_select);
2105
185k
      write_smooth_directional_sum16(dst, top1, top2, weights_y, weights_y,
2106
185k
                                     scaled_bottom_left_y, scaled_bottom_left_y,
2107
185k
                                     round);
2108
185k
      write_smooth_directional_sum16(dst + 16, top3, top4, weights_y, weights_y,
2109
185k
                                     scaled_bottom_left_y, scaled_bottom_left_y,
2110
185k
                                     round);
2111
185k
      write_smooth_directional_sum16(dst + 32, top5, top6, weights_y, weights_y,
2112
185k
                                     scaled_bottom_left_y, scaled_bottom_left_y,
2113
185k
                                     round);
2114
185k
      write_smooth_directional_sum16(dst + 48, top7, top8, weights_y, weights_y,
2115
185k
                                     scaled_bottom_left_y, scaled_bottom_left_y,
2116
185k
                                     round);
2117
185k
      dst += stride;
2118
185k
    }
2119
23.1k
  }
2120
5.79k
}
2121
2122
// -----------------------------------------------------------------------------
2123
// SMOOTH_H_PRED
2124
static AOM_FORCE_INLINE void write_smooth_horizontal_sum4(
2125
    uint8_t *LIBAOM_RESTRICT dst, const __m128i *left_y, const __m128i *weights,
2126
723k
    const __m128i *scaled_top_right, const __m128i *round) {
2127
723k
  const __m128i weighted_left_y = _mm_mullo_epi16(*left_y, *weights);
2128
723k
  const __m128i pred_sum = _mm_add_epi32(*scaled_top_right, weighted_left_y);
2129
  // Equivalent to RightShiftWithRounding(pred[x][y], 8).
2130
723k
  const __m128i pred = _mm_srli_epi32(_mm_add_epi32(pred_sum, *round), 8);
2131
723k
  const __m128i cvtepi32_epi8 = _mm_set1_epi32(0x0C080400);
2132
723k
  Store4(dst, _mm_shuffle_epi8(pred, cvtepi32_epi8));
2133
723k
}
2134
2135
void aom_smooth_h_predictor_4x4_ssse3(
2136
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2137
    const uint8_t *LIBAOM_RESTRICT top_row,
2138
74.8k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2139
74.8k
  const __m128i top_right = _mm_set1_epi32(top_row[3]);
2140
74.8k
  const __m128i left = cvtepu8_epi32(Load4(left_column));
2141
74.8k
  const __m128i weights = cvtepu8_epi32(Load4(smooth_weights));
2142
74.8k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2143
74.8k
  const __m128i inverted_weights = _mm_sub_epi32(scale, weights);
2144
74.8k
  const __m128i scaled_top_right = _mm_mullo_epi16(inverted_weights, top_right);
2145
74.8k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2146
74.8k
  __m128i left_y = _mm_shuffle_epi32(left, 0);
2147
74.8k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2148
74.8k
                               &round);
2149
74.8k
  dst += stride;
2150
74.8k
  left_y = _mm_shuffle_epi32(left, 0x55);
2151
74.8k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2152
74.8k
                               &round);
2153
74.8k
  dst += stride;
2154
74.8k
  left_y = _mm_shuffle_epi32(left, 0xaa);
2155
74.8k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2156
74.8k
                               &round);
2157
74.8k
  dst += stride;
2158
74.8k
  left_y = _mm_shuffle_epi32(left, 0xff);
2159
74.8k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2160
74.8k
                               &round);
2161
74.8k
}
2162
2163
void aom_smooth_h_predictor_4x8_ssse3(
2164
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2165
    const uint8_t *LIBAOM_RESTRICT top_row,
2166
21.1k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2167
21.1k
  const __m128i top_right = _mm_set1_epi32(top_row[3]);
2168
21.1k
  const __m128i weights = cvtepu8_epi32(Load4(smooth_weights));
2169
21.1k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2170
21.1k
  const __m128i inverted_weights = _mm_sub_epi32(scale, weights);
2171
21.1k
  const __m128i scaled_top_right = _mm_mullo_epi16(inverted_weights, top_right);
2172
21.1k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2173
21.1k
  __m128i left = cvtepu8_epi32(Load4(left_column));
2174
21.1k
  __m128i left_y = _mm_shuffle_epi32(left, 0);
2175
21.1k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2176
21.1k
                               &round);
2177
21.1k
  dst += stride;
2178
21.1k
  left_y = _mm_shuffle_epi32(left, 0x55);
2179
21.1k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2180
21.1k
                               &round);
2181
21.1k
  dst += stride;
2182
21.1k
  left_y = _mm_shuffle_epi32(left, 0xaa);
2183
21.1k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2184
21.1k
                               &round);
2185
21.1k
  dst += stride;
2186
21.1k
  left_y = _mm_shuffle_epi32(left, 0xff);
2187
21.1k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2188
21.1k
                               &round);
2189
21.1k
  dst += stride;
2190
2191
21.1k
  left = cvtepu8_epi32(Load4(left_column + 4));
2192
21.1k
  left_y = _mm_shuffle_epi32(left, 0);
2193
21.1k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2194
21.1k
                               &round);
2195
21.1k
  dst += stride;
2196
21.1k
  left_y = _mm_shuffle_epi32(left, 0x55);
2197
21.1k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2198
21.1k
                               &round);
2199
21.1k
  dst += stride;
2200
21.1k
  left_y = _mm_shuffle_epi32(left, 0xaa);
2201
21.1k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2202
21.1k
                               &round);
2203
21.1k
  dst += stride;
2204
21.1k
  left_y = _mm_shuffle_epi32(left, 0xff);
2205
21.1k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2206
21.1k
                               &round);
2207
21.1k
}
2208
2209
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
2210
void aom_smooth_h_predictor_4x16_ssse3(
2211
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2212
    const uint8_t *LIBAOM_RESTRICT top_row,
2213
15.9k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2214
15.9k
  const __m128i top_right = _mm_set1_epi32(top_row[3]);
2215
15.9k
  const __m128i weights = cvtepu8_epi32(Load4(smooth_weights));
2216
15.9k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2217
15.9k
  const __m128i inverted_weights = _mm_sub_epi32(scale, weights);
2218
15.9k
  const __m128i scaled_top_right = _mm_mullo_epi16(inverted_weights, top_right);
2219
15.9k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2220
15.9k
  __m128i left = cvtepu8_epi32(Load4(left_column));
2221
15.9k
  __m128i left_y = _mm_shuffle_epi32(left, 0);
2222
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2223
15.9k
                               &round);
2224
15.9k
  dst += stride;
2225
15.9k
  left_y = _mm_shuffle_epi32(left, 0x55);
2226
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2227
15.9k
                               &round);
2228
15.9k
  dst += stride;
2229
15.9k
  left_y = _mm_shuffle_epi32(left, 0xaa);
2230
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2231
15.9k
                               &round);
2232
15.9k
  dst += stride;
2233
15.9k
  left_y = _mm_shuffle_epi32(left, 0xff);
2234
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2235
15.9k
                               &round);
2236
15.9k
  dst += stride;
2237
2238
15.9k
  left = cvtepu8_epi32(Load4(left_column + 4));
2239
15.9k
  left_y = _mm_shuffle_epi32(left, 0);
2240
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2241
15.9k
                               &round);
2242
15.9k
  dst += stride;
2243
15.9k
  left_y = _mm_shuffle_epi32(left, 0x55);
2244
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2245
15.9k
                               &round);
2246
15.9k
  dst += stride;
2247
15.9k
  left_y = _mm_shuffle_epi32(left, 0xaa);
2248
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2249
15.9k
                               &round);
2250
15.9k
  dst += stride;
2251
15.9k
  left_y = _mm_shuffle_epi32(left, 0xff);
2252
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2253
15.9k
                               &round);
2254
15.9k
  dst += stride;
2255
2256
15.9k
  left = cvtepu8_epi32(Load4(left_column + 8));
2257
15.9k
  left_y = _mm_shuffle_epi32(left, 0);
2258
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2259
15.9k
                               &round);
2260
15.9k
  dst += stride;
2261
15.9k
  left_y = _mm_shuffle_epi32(left, 0x55);
2262
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2263
15.9k
                               &round);
2264
15.9k
  dst += stride;
2265
15.9k
  left_y = _mm_shuffle_epi32(left, 0xaa);
2266
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2267
15.9k
                               &round);
2268
15.9k
  dst += stride;
2269
15.9k
  left_y = _mm_shuffle_epi32(left, 0xff);
2270
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2271
15.9k
                               &round);
2272
15.9k
  dst += stride;
2273
2274
15.9k
  left = cvtepu8_epi32(Load4(left_column + 12));
2275
15.9k
  left_y = _mm_shuffle_epi32(left, 0);
2276
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2277
15.9k
                               &round);
2278
15.9k
  dst += stride;
2279
15.9k
  left_y = _mm_shuffle_epi32(left, 0x55);
2280
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2281
15.9k
                               &round);
2282
15.9k
  dst += stride;
2283
15.9k
  left_y = _mm_shuffle_epi32(left, 0xaa);
2284
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2285
15.9k
                               &round);
2286
15.9k
  dst += stride;
2287
15.9k
  left_y = _mm_shuffle_epi32(left, 0xff);
2288
15.9k
  write_smooth_horizontal_sum4(dst, &left_y, &weights, &scaled_top_right,
2289
15.9k
                               &round);
2290
15.9k
}
2291
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
2292
2293
// For SMOOTH_H, |pixels| is the repeated left value for the row. For SMOOTH_V,
2294
// |pixels| is a segment of the top row or the whole top row, and |weights| is
2295
// repeated.
2296
void aom_smooth_h_predictor_8x4_ssse3(
2297
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2298
    const uint8_t *LIBAOM_RESTRICT top_row,
2299
34.9k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2300
34.9k
  const __m128i top_right = _mm_set1_epi16(top_row[7]);
2301
34.9k
  const __m128i left = cvtepu8_epi16(Load4(left_column));
2302
34.9k
  const __m128i weights = cvtepu8_epi16(LoadLo8(smooth_weights + 4));
2303
34.9k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2304
34.9k
  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
2305
34.9k
  const __m128i scaled_top_right = _mm_mullo_epi16(inverted_weights, top_right);
2306
34.9k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2307
34.9k
  __m128i y_select = _mm_set1_epi32(0x01000100);
2308
34.9k
  __m128i left_y = _mm_shuffle_epi8(left, y_select);
2309
34.9k
  write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
2310
34.9k
                                &round);
2311
34.9k
  dst += stride;
2312
34.9k
  y_select = _mm_set1_epi32(0x03020302);
2313
34.9k
  left_y = _mm_shuffle_epi8(left, y_select);
2314
34.9k
  write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
2315
34.9k
                                &round);
2316
34.9k
  dst += stride;
2317
34.9k
  y_select = _mm_set1_epi32(0x05040504);
2318
34.9k
  left_y = _mm_shuffle_epi8(left, y_select);
2319
34.9k
  write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
2320
34.9k
                                &round);
2321
34.9k
  dst += stride;
2322
34.9k
  y_select = _mm_set1_epi32(0x07060706);
2323
34.9k
  left_y = _mm_shuffle_epi8(left, y_select);
2324
34.9k
  write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
2325
34.9k
                                &round);
2326
34.9k
}
2327
2328
void aom_smooth_h_predictor_8x8_ssse3(
2329
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2330
    const uint8_t *LIBAOM_RESTRICT top_row,
2331
49.1k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2332
49.1k
  const __m128i top_right = _mm_set1_epi16(top_row[7]);
2333
49.1k
  const __m128i left = cvtepu8_epi16(LoadLo8(left_column));
2334
49.1k
  const __m128i weights = cvtepu8_epi16(LoadLo8(smooth_weights + 4));
2335
49.1k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2336
49.1k
  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
2337
49.1k
  const __m128i scaled_top_right = _mm_mullo_epi16(inverted_weights, top_right);
2338
49.1k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2339
442k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2340
393k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2341
393k
    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2342
393k
    write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
2343
393k
                                  &round);
2344
393k
    dst += stride;
2345
393k
  }
2346
49.1k
}
2347
2348
void aom_smooth_h_predictor_8x16_ssse3(
2349
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2350
    const uint8_t *LIBAOM_RESTRICT top_row,
2351
17.8k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2352
17.8k
  const __m128i top_right = _mm_set1_epi16(top_row[7]);
2353
17.8k
  const __m128i weights = cvtepu8_epi16(LoadLo8(smooth_weights + 4));
2354
17.8k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2355
17.8k
  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
2356
17.8k
  const __m128i scaled_top_right = _mm_mullo_epi16(inverted_weights, top_right);
2357
17.8k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2358
17.8k
  __m128i left = cvtepu8_epi16(LoadLo8(left_column));
2359
160k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2360
142k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2361
142k
    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2362
142k
    write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
2363
142k
                                  &round);
2364
142k
    dst += stride;
2365
142k
  }
2366
17.8k
  left = cvtepu8_epi16(LoadLo8(left_column + 8));
2367
160k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2368
142k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2369
142k
    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2370
142k
    write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
2371
142k
                                  &round);
2372
142k
    dst += stride;
2373
142k
  }
2374
17.8k
}
2375
2376
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
2377
void aom_smooth_h_predictor_8x32_ssse3(
2378
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2379
    const uint8_t *LIBAOM_RESTRICT top_row,
2380
6.02k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2381
6.02k
  const __m128i top_right = _mm_set1_epi16(top_row[7]);
2382
6.02k
  const __m128i weights = cvtepu8_epi16(LoadLo8(smooth_weights + 4));
2383
6.02k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2384
6.02k
  const __m128i inverted_weights = _mm_sub_epi16(scale, weights);
2385
6.02k
  const __m128i scaled_top_right = _mm_mullo_epi16(inverted_weights, top_right);
2386
6.02k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2387
6.02k
  __m128i left = cvtepu8_epi16(LoadLo8(left_column));
2388
54.1k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2389
48.1k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2390
48.1k
    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2391
48.1k
    write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
2392
48.1k
                                  &round);
2393
48.1k
    dst += stride;
2394
48.1k
  }
2395
6.02k
  left = cvtepu8_epi16(LoadLo8(left_column + 8));
2396
54.1k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2397
48.1k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2398
48.1k
    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2399
48.1k
    write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
2400
48.1k
                                  &round);
2401
48.1k
    dst += stride;
2402
48.1k
  }
2403
6.02k
  left = cvtepu8_epi16(LoadLo8(left_column + 16));
2404
54.1k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2405
48.1k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2406
48.1k
    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2407
48.1k
    write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
2408
48.1k
                                  &round);
2409
48.1k
    dst += stride;
2410
48.1k
  }
2411
6.02k
  left = cvtepu8_epi16(LoadLo8(left_column + 24));
2412
54.1k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2413
48.1k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2414
48.1k
    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2415
48.1k
    write_smooth_directional_sum8(dst, &left_y, &weights, &scaled_top_right,
2416
48.1k
                                  &round);
2417
48.1k
    dst += stride;
2418
48.1k
  }
2419
6.02k
}
2420
2421
void aom_smooth_h_predictor_16x4_ssse3(
2422
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2423
    const uint8_t *LIBAOM_RESTRICT top_row,
2424
23.9k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2425
23.9k
  const __m128i top_right = _mm_set1_epi16(top_row[15]);
2426
23.9k
  const __m128i left = cvtepu8_epi16(Load4(left_column));
2427
23.9k
  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
2428
23.9k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2429
23.9k
  const __m128i weights1 = cvtepu8_epi16(weights);
2430
23.9k
  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights, 8));
2431
23.9k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
2432
23.9k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
2433
23.9k
  const __m128i scaled_top_right1 =
2434
23.9k
      _mm_mullo_epi16(inverted_weights1, top_right);
2435
23.9k
  const __m128i scaled_top_right2 =
2436
23.9k
      _mm_mullo_epi16(inverted_weights2, top_right);
2437
23.9k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2438
23.9k
  __m128i y_mask = _mm_set1_epi32(0x01000100);
2439
23.9k
  __m128i left_y = _mm_shuffle_epi8(left, y_mask);
2440
23.9k
  write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2441
23.9k
                                 scaled_top_right1, scaled_top_right2, round);
2442
23.9k
  dst += stride;
2443
23.9k
  y_mask = _mm_set1_epi32(0x03020302);
2444
23.9k
  left_y = _mm_shuffle_epi8(left, y_mask);
2445
23.9k
  write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2446
23.9k
                                 scaled_top_right1, scaled_top_right2, round);
2447
23.9k
  dst += stride;
2448
23.9k
  y_mask = _mm_set1_epi32(0x05040504);
2449
23.9k
  left_y = _mm_shuffle_epi8(left, y_mask);
2450
23.9k
  write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2451
23.9k
                                 scaled_top_right1, scaled_top_right2, round);
2452
23.9k
  dst += stride;
2453
23.9k
  y_mask = _mm_set1_epi32(0x07060706);
2454
23.9k
  left_y = _mm_shuffle_epi8(left, y_mask);
2455
23.9k
  write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2456
23.9k
                                 scaled_top_right1, scaled_top_right2, round);
2457
23.9k
}
2458
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
2459
2460
void aom_smooth_h_predictor_16x8_ssse3(
2461
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2462
    const uint8_t *LIBAOM_RESTRICT top_row,
2463
22.2k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2464
22.2k
  const __m128i top_right = _mm_set1_epi16(top_row[15]);
2465
22.2k
  const __m128i left = cvtepu8_epi16(LoadLo8(left_column));
2466
22.2k
  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
2467
22.2k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2468
22.2k
  const __m128i weights1 = cvtepu8_epi16(weights);
2469
22.2k
  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights, 8));
2470
22.2k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
2471
22.2k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
2472
22.2k
  const __m128i scaled_top_right1 =
2473
22.2k
      _mm_mullo_epi16(inverted_weights1, top_right);
2474
22.2k
  const __m128i scaled_top_right2 =
2475
22.2k
      _mm_mullo_epi16(inverted_weights2, top_right);
2476
22.2k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2477
200k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2478
178k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2479
178k
    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2480
178k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2481
178k
                                   scaled_top_right1, scaled_top_right2, round);
2482
178k
    dst += stride;
2483
178k
  }
2484
22.2k
}
2485
2486
void aom_smooth_h_predictor_16x16_ssse3(
2487
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2488
    const uint8_t *LIBAOM_RESTRICT top_row,
2489
42.1k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2490
42.1k
  const __m128i top_right = _mm_set1_epi16(top_row[15]);
2491
42.1k
  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
2492
42.1k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2493
42.1k
  const __m128i weights1 = cvtepu8_epi16(weights);
2494
42.1k
  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights, 8));
2495
42.1k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
2496
42.1k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
2497
42.1k
  const __m128i scaled_top_right1 =
2498
42.1k
      _mm_mullo_epi16(inverted_weights1, top_right);
2499
42.1k
  const __m128i scaled_top_right2 =
2500
42.1k
      _mm_mullo_epi16(inverted_weights2, top_right);
2501
42.1k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2502
42.1k
  __m128i left = cvtepu8_epi16(LoadLo8(left_column));
2503
379k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2504
337k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2505
337k
    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2506
337k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2507
337k
                                   scaled_top_right1, scaled_top_right2, round);
2508
337k
    dst += stride;
2509
337k
  }
2510
42.1k
  left = cvtepu8_epi16(LoadLo8(left_column + 8));
2511
379k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2512
337k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2513
337k
    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2514
337k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2515
337k
                                   scaled_top_right1, scaled_top_right2, round);
2516
337k
    dst += stride;
2517
337k
  }
2518
42.1k
}
2519
2520
void aom_smooth_h_predictor_16x32_ssse3(
2521
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2522
    const uint8_t *LIBAOM_RESTRICT top_row,
2523
12.3k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2524
12.3k
  const __m128i top_right = _mm_set1_epi16(top_row[15]);
2525
12.3k
  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
2526
12.3k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2527
12.3k
  const __m128i weights1 = cvtepu8_epi16(weights);
2528
12.3k
  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights, 8));
2529
12.3k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
2530
12.3k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
2531
12.3k
  const __m128i scaled_top_right1 =
2532
12.3k
      _mm_mullo_epi16(inverted_weights1, top_right);
2533
12.3k
  const __m128i scaled_top_right2 =
2534
12.3k
      _mm_mullo_epi16(inverted_weights2, top_right);
2535
12.3k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2536
12.3k
  __m128i left = cvtepu8_epi16(LoadLo8(left_column));
2537
111k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2538
98.8k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2539
98.8k
    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2540
98.8k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2541
98.8k
                                   scaled_top_right1, scaled_top_right2, round);
2542
98.8k
    dst += stride;
2543
98.8k
  }
2544
12.3k
  left = cvtepu8_epi16(LoadLo8(left_column + 8));
2545
111k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2546
98.8k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2547
98.8k
    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2548
98.8k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2549
98.8k
                                   scaled_top_right1, scaled_top_right2, round);
2550
98.8k
    dst += stride;
2551
98.8k
  }
2552
12.3k
  left = cvtepu8_epi16(LoadLo8(left_column + 16));
2553
111k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2554
98.8k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2555
98.8k
    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2556
98.8k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2557
98.8k
                                   scaled_top_right1, scaled_top_right2, round);
2558
98.8k
    dst += stride;
2559
98.8k
  }
2560
12.3k
  left = cvtepu8_epi16(LoadLo8(left_column + 24));
2561
111k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2562
98.8k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2563
98.8k
    const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2564
98.8k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2565
98.8k
                                   scaled_top_right1, scaled_top_right2, round);
2566
98.8k
    dst += stride;
2567
98.8k
  }
2568
12.3k
}
2569
2570
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
2571
void aom_smooth_h_predictor_16x64_ssse3(
2572
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2573
    const uint8_t *LIBAOM_RESTRICT top_row,
2574
3.03k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2575
3.03k
  const __m128i top_right = _mm_set1_epi16(top_row[15]);
2576
3.03k
  const __m128i weights = LoadUnaligned16(smooth_weights + 12);
2577
3.03k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2578
3.03k
  const __m128i weights1 = cvtepu8_epi16(weights);
2579
3.03k
  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights, 8));
2580
3.03k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
2581
3.03k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
2582
3.03k
  const __m128i scaled_top_right1 =
2583
3.03k
      _mm_mullo_epi16(inverted_weights1, top_right);
2584
3.03k
  const __m128i scaled_top_right2 =
2585
3.03k
      _mm_mullo_epi16(inverted_weights2, top_right);
2586
3.03k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2587
27.3k
  for (int left_offset = 0; left_offset < 64; left_offset += 8) {
2588
24.3k
    const __m128i left = cvtepu8_epi16(LoadLo8(left_column + left_offset));
2589
218k
    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2590
194k
      const __m128i y_select = _mm_set1_epi32(y_mask);
2591
194k
      const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2592
194k
      write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2593
194k
                                     scaled_top_right1, scaled_top_right2,
2594
194k
                                     round);
2595
194k
      dst += stride;
2596
194k
    }
2597
24.3k
  }
2598
3.03k
}
2599
2600
void aom_smooth_h_predictor_32x8_ssse3(
2601
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2602
    const uint8_t *LIBAOM_RESTRICT top_row,
2603
18.1k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2604
18.1k
  const __m128i top_right = _mm_set1_epi16(top_row[31]);
2605
18.1k
  const __m128i left = cvtepu8_epi16(LoadLo8(left_column));
2606
18.1k
  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
2607
18.1k
  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
2608
18.1k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2609
18.1k
  const __m128i weights1 = cvtepu8_epi16(weights_lo);
2610
18.1k
  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights_lo, 8));
2611
18.1k
  const __m128i weights3 = cvtepu8_epi16(weights_hi);
2612
18.1k
  const __m128i weights4 = cvtepu8_epi16(_mm_srli_si128(weights_hi, 8));
2613
18.1k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
2614
18.1k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
2615
18.1k
  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
2616
18.1k
  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
2617
18.1k
  const __m128i scaled_top_right1 =
2618
18.1k
      _mm_mullo_epi16(inverted_weights1, top_right);
2619
18.1k
  const __m128i scaled_top_right2 =
2620
18.1k
      _mm_mullo_epi16(inverted_weights2, top_right);
2621
18.1k
  const __m128i scaled_top_right3 =
2622
18.1k
      _mm_mullo_epi16(inverted_weights3, top_right);
2623
18.1k
  const __m128i scaled_top_right4 =
2624
18.1k
      _mm_mullo_epi16(inverted_weights4, top_right);
2625
18.1k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2626
163k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2627
145k
    __m128i y_select = _mm_set1_epi32(y_mask);
2628
145k
    __m128i left_y = _mm_shuffle_epi8(left, y_select);
2629
145k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2630
145k
                                   scaled_top_right1, scaled_top_right2, round);
2631
145k
    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
2632
145k
                                   scaled_top_right3, scaled_top_right4, round);
2633
145k
    dst += stride;
2634
145k
  }
2635
18.1k
}
2636
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
2637
2638
void aom_smooth_h_predictor_32x16_ssse3(
2639
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2640
    const uint8_t *LIBAOM_RESTRICT top_row,
2641
10.5k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2642
10.5k
  const __m128i top_right = _mm_set1_epi16(top_row[31]);
2643
10.5k
  const __m128i left1 = cvtepu8_epi16(LoadLo8(left_column));
2644
10.5k
  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
2645
10.5k
  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
2646
10.5k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2647
10.5k
  const __m128i weights1 = cvtepu8_epi16(weights_lo);
2648
10.5k
  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights_lo, 8));
2649
10.5k
  const __m128i weights3 = cvtepu8_epi16(weights_hi);
2650
10.5k
  const __m128i weights4 = cvtepu8_epi16(_mm_srli_si128(weights_hi, 8));
2651
10.5k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
2652
10.5k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
2653
10.5k
  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
2654
10.5k
  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
2655
10.5k
  const __m128i scaled_top_right1 =
2656
10.5k
      _mm_mullo_epi16(inverted_weights1, top_right);
2657
10.5k
  const __m128i scaled_top_right2 =
2658
10.5k
      _mm_mullo_epi16(inverted_weights2, top_right);
2659
10.5k
  const __m128i scaled_top_right3 =
2660
10.5k
      _mm_mullo_epi16(inverted_weights3, top_right);
2661
10.5k
  const __m128i scaled_top_right4 =
2662
10.5k
      _mm_mullo_epi16(inverted_weights4, top_right);
2663
10.5k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2664
95.3k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2665
84.7k
    __m128i y_select = _mm_set1_epi32(y_mask);
2666
84.7k
    __m128i left_y = _mm_shuffle_epi8(left1, y_select);
2667
84.7k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2668
84.7k
                                   scaled_top_right1, scaled_top_right2, round);
2669
84.7k
    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
2670
84.7k
                                   scaled_top_right3, scaled_top_right4, round);
2671
84.7k
    dst += stride;
2672
84.7k
  }
2673
10.5k
  const __m128i left2 =
2674
10.5k
      cvtepu8_epi16(LoadLo8((const uint8_t *)left_column + 8));
2675
95.3k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2676
84.7k
    __m128i y_select = _mm_set1_epi32(y_mask);
2677
84.7k
    __m128i left_y = _mm_shuffle_epi8(left2, y_select);
2678
84.7k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2679
84.7k
                                   scaled_top_right1, scaled_top_right2, round);
2680
84.7k
    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
2681
84.7k
                                   scaled_top_right3, scaled_top_right4, round);
2682
84.7k
    dst += stride;
2683
84.7k
  }
2684
10.5k
}
2685
2686
void aom_smooth_h_predictor_32x32_ssse3(
2687
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2688
    const uint8_t *LIBAOM_RESTRICT top_row,
2689
41.4k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2690
41.4k
  const __m128i top_right = _mm_set1_epi16(top_row[31]);
2691
41.4k
  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
2692
41.4k
  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
2693
41.4k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2694
41.4k
  const __m128i weights1 = cvtepu8_epi16(weights_lo);
2695
41.4k
  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights_lo, 8));
2696
41.4k
  const __m128i weights3 = cvtepu8_epi16(weights_hi);
2697
41.4k
  const __m128i weights4 = cvtepu8_epi16(_mm_srli_si128(weights_hi, 8));
2698
41.4k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
2699
41.4k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
2700
41.4k
  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
2701
41.4k
  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
2702
41.4k
  const __m128i scaled_top_right1 =
2703
41.4k
      _mm_mullo_epi16(inverted_weights1, top_right);
2704
41.4k
  const __m128i scaled_top_right2 =
2705
41.4k
      _mm_mullo_epi16(inverted_weights2, top_right);
2706
41.4k
  const __m128i scaled_top_right3 =
2707
41.4k
      _mm_mullo_epi16(inverted_weights3, top_right);
2708
41.4k
  const __m128i scaled_top_right4 =
2709
41.4k
      _mm_mullo_epi16(inverted_weights4, top_right);
2710
41.4k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2711
41.4k
  __m128i left = cvtepu8_epi16(LoadLo8(left_column));
2712
373k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2713
331k
    __m128i y_select = _mm_set1_epi32(y_mask);
2714
331k
    __m128i left_y = _mm_shuffle_epi8(left, y_select);
2715
331k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2716
331k
                                   scaled_top_right1, scaled_top_right2, round);
2717
331k
    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
2718
331k
                                   scaled_top_right3, scaled_top_right4, round);
2719
331k
    dst += stride;
2720
331k
  }
2721
41.4k
  left = cvtepu8_epi16(LoadLo8(left_column + 8));
2722
373k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2723
331k
    __m128i y_select = _mm_set1_epi32(y_mask);
2724
331k
    __m128i left_y = _mm_shuffle_epi8(left, y_select);
2725
331k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2726
331k
                                   scaled_top_right1, scaled_top_right2, round);
2727
331k
    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
2728
331k
                                   scaled_top_right3, scaled_top_right4, round);
2729
331k
    dst += stride;
2730
331k
  }
2731
41.4k
  left = cvtepu8_epi16(LoadLo8(left_column + 16));
2732
373k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2733
331k
    __m128i y_select = _mm_set1_epi32(y_mask);
2734
331k
    __m128i left_y = _mm_shuffle_epi8(left, y_select);
2735
331k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2736
331k
                                   scaled_top_right1, scaled_top_right2, round);
2737
331k
    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
2738
331k
                                   scaled_top_right3, scaled_top_right4, round);
2739
331k
    dst += stride;
2740
331k
  }
2741
41.4k
  left = cvtepu8_epi16(LoadLo8(left_column + 24));
2742
373k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2743
331k
    __m128i y_select = _mm_set1_epi32(y_mask);
2744
331k
    __m128i left_y = _mm_shuffle_epi8(left, y_select);
2745
331k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2746
331k
                                   scaled_top_right1, scaled_top_right2, round);
2747
331k
    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
2748
331k
                                   scaled_top_right3, scaled_top_right4, round);
2749
331k
    dst += stride;
2750
331k
  }
2751
41.4k
}
2752
2753
void aom_smooth_h_predictor_32x64_ssse3(
2754
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2755
    const uint8_t *LIBAOM_RESTRICT top_row,
2756
815
    const uint8_t *LIBAOM_RESTRICT left_column) {
2757
815
  const __m128i top_right = _mm_set1_epi16(top_row[31]);
2758
815
  const __m128i weights_lo = LoadUnaligned16(smooth_weights + 28);
2759
815
  const __m128i weights_hi = LoadUnaligned16(smooth_weights + 44);
2760
815
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2761
815
  const __m128i weights1 = cvtepu8_epi16(weights_lo);
2762
815
  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights_lo, 8));
2763
815
  const __m128i weights3 = cvtepu8_epi16(weights_hi);
2764
815
  const __m128i weights4 = cvtepu8_epi16(_mm_srli_si128(weights_hi, 8));
2765
815
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
2766
815
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
2767
815
  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
2768
815
  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
2769
815
  const __m128i scaled_top_right1 =
2770
815
      _mm_mullo_epi16(inverted_weights1, top_right);
2771
815
  const __m128i scaled_top_right2 =
2772
815
      _mm_mullo_epi16(inverted_weights2, top_right);
2773
815
  const __m128i scaled_top_right3 =
2774
815
      _mm_mullo_epi16(inverted_weights3, top_right);
2775
815
  const __m128i scaled_top_right4 =
2776
815
      _mm_mullo_epi16(inverted_weights4, top_right);
2777
815
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2778
7.33k
  for (int left_offset = 0; left_offset < 64; left_offset += 8) {
2779
6.52k
    const __m128i left = cvtepu8_epi16(LoadLo8(left_column + left_offset));
2780
58.6k
    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2781
52.1k
      const __m128i y_select = _mm_set1_epi32(y_mask);
2782
52.1k
      const __m128i left_y = _mm_shuffle_epi8(left, y_select);
2783
52.1k
      write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2784
52.1k
                                     scaled_top_right1, scaled_top_right2,
2785
52.1k
                                     round);
2786
52.1k
      write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3,
2787
52.1k
                                     weights4, scaled_top_right3,
2788
52.1k
                                     scaled_top_right4, round);
2789
52.1k
      dst += stride;
2790
52.1k
    }
2791
6.52k
  }
2792
815
}
2793
2794
#if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
2795
void aom_smooth_h_predictor_64x16_ssse3(
2796
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2797
    const uint8_t *LIBAOM_RESTRICT top_row,
2798
5.70k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2799
5.70k
  const __m128i top_right = _mm_set1_epi16(top_row[63]);
2800
5.70k
  const __m128i left1 = cvtepu8_epi16(LoadLo8(left_column));
2801
5.70k
  const __m128i weights_lolo = LoadUnaligned16(smooth_weights + 60);
2802
5.70k
  const __m128i weights_lohi = LoadUnaligned16(smooth_weights + 76);
2803
5.70k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2804
5.70k
  const __m128i weights1 = cvtepu8_epi16(weights_lolo);
2805
5.70k
  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights_lolo, 8));
2806
5.70k
  const __m128i weights3 = cvtepu8_epi16(weights_lohi);
2807
5.70k
  const __m128i weights4 = cvtepu8_epi16(_mm_srli_si128(weights_lohi, 8));
2808
5.70k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
2809
5.70k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
2810
5.70k
  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
2811
5.70k
  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
2812
5.70k
  const __m128i scaled_top_right1 =
2813
5.70k
      _mm_mullo_epi16(inverted_weights1, top_right);
2814
5.70k
  const __m128i scaled_top_right2 =
2815
5.70k
      _mm_mullo_epi16(inverted_weights2, top_right);
2816
5.70k
  const __m128i scaled_top_right3 =
2817
5.70k
      _mm_mullo_epi16(inverted_weights3, top_right);
2818
5.70k
  const __m128i scaled_top_right4 =
2819
5.70k
      _mm_mullo_epi16(inverted_weights4, top_right);
2820
5.70k
  const __m128i weights_hilo = LoadUnaligned16(smooth_weights + 92);
2821
5.70k
  const __m128i weights_hihi = LoadUnaligned16(smooth_weights + 108);
2822
5.70k
  const __m128i weights5 = cvtepu8_epi16(weights_hilo);
2823
5.70k
  const __m128i weights6 = cvtepu8_epi16(_mm_srli_si128(weights_hilo, 8));
2824
5.70k
  const __m128i weights7 = cvtepu8_epi16(weights_hihi);
2825
5.70k
  const __m128i weights8 = cvtepu8_epi16(_mm_srli_si128(weights_hihi, 8));
2826
5.70k
  const __m128i inverted_weights5 = _mm_sub_epi16(scale, weights5);
2827
5.70k
  const __m128i inverted_weights6 = _mm_sub_epi16(scale, weights6);
2828
5.70k
  const __m128i inverted_weights7 = _mm_sub_epi16(scale, weights7);
2829
5.70k
  const __m128i inverted_weights8 = _mm_sub_epi16(scale, weights8);
2830
5.70k
  const __m128i scaled_top_right5 =
2831
5.70k
      _mm_mullo_epi16(inverted_weights5, top_right);
2832
5.70k
  const __m128i scaled_top_right6 =
2833
5.70k
      _mm_mullo_epi16(inverted_weights6, top_right);
2834
5.70k
  const __m128i scaled_top_right7 =
2835
5.70k
      _mm_mullo_epi16(inverted_weights7, top_right);
2836
5.70k
  const __m128i scaled_top_right8 =
2837
5.70k
      _mm_mullo_epi16(inverted_weights8, top_right);
2838
5.70k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2839
51.3k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2840
45.6k
    __m128i y_select = _mm_set1_epi32(y_mask);
2841
45.6k
    __m128i left_y = _mm_shuffle_epi8(left1, y_select);
2842
45.6k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2843
45.6k
                                   scaled_top_right1, scaled_top_right2, round);
2844
45.6k
    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
2845
45.6k
                                   scaled_top_right3, scaled_top_right4, round);
2846
45.6k
    write_smooth_directional_sum16(dst + 32, left_y, left_y, weights5, weights6,
2847
45.6k
                                   scaled_top_right5, scaled_top_right6, round);
2848
45.6k
    write_smooth_directional_sum16(dst + 48, left_y, left_y, weights7, weights8,
2849
45.6k
                                   scaled_top_right7, scaled_top_right8, round);
2850
45.6k
    dst += stride;
2851
45.6k
  }
2852
5.70k
  const __m128i left2 = cvtepu8_epi16(LoadLo8(left_column + 8));
2853
51.3k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2854
45.6k
    __m128i y_select = _mm_set1_epi32(y_mask);
2855
45.6k
    __m128i left_y = _mm_shuffle_epi8(left2, y_select);
2856
45.6k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2857
45.6k
                                   scaled_top_right1, scaled_top_right2, round);
2858
45.6k
    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
2859
45.6k
                                   scaled_top_right3, scaled_top_right4, round);
2860
45.6k
    write_smooth_directional_sum16(dst + 32, left_y, left_y, weights5, weights6,
2861
45.6k
                                   scaled_top_right5, scaled_top_right6, round);
2862
45.6k
    write_smooth_directional_sum16(dst + 48, left_y, left_y, weights7, weights8,
2863
45.6k
                                   scaled_top_right7, scaled_top_right8, round);
2864
45.6k
    dst += stride;
2865
45.6k
  }
2866
5.70k
}
2867
#endif  // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER
2868
2869
void aom_smooth_h_predictor_64x32_ssse3(
2870
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2871
    const uint8_t *LIBAOM_RESTRICT top_row,
2872
1.45k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2873
1.45k
  const __m128i top_right = _mm_set1_epi16(top_row[63]);
2874
1.45k
  const __m128i left1 = cvtepu8_epi16(LoadLo8(left_column));
2875
1.45k
  const __m128i weights_lolo = LoadUnaligned16(smooth_weights + 60);
2876
1.45k
  const __m128i weights_lohi = LoadUnaligned16(smooth_weights + 76);
2877
1.45k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2878
1.45k
  const __m128i weights1 = cvtepu8_epi16(weights_lolo);
2879
1.45k
  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights_lolo, 8));
2880
1.45k
  const __m128i weights3 = cvtepu8_epi16(weights_lohi);
2881
1.45k
  const __m128i weights4 = cvtepu8_epi16(_mm_srli_si128(weights_lohi, 8));
2882
1.45k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
2883
1.45k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
2884
1.45k
  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
2885
1.45k
  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
2886
1.45k
  const __m128i scaled_top_right1 =
2887
1.45k
      _mm_mullo_epi16(inverted_weights1, top_right);
2888
1.45k
  const __m128i scaled_top_right2 =
2889
1.45k
      _mm_mullo_epi16(inverted_weights2, top_right);
2890
1.45k
  const __m128i scaled_top_right3 =
2891
1.45k
      _mm_mullo_epi16(inverted_weights3, top_right);
2892
1.45k
  const __m128i scaled_top_right4 =
2893
1.45k
      _mm_mullo_epi16(inverted_weights4, top_right);
2894
1.45k
  const __m128i weights_hilo = LoadUnaligned16(smooth_weights + 92);
2895
1.45k
  const __m128i weights_hihi = LoadUnaligned16(smooth_weights + 108);
2896
1.45k
  const __m128i weights5 = cvtepu8_epi16(weights_hilo);
2897
1.45k
  const __m128i weights6 = cvtepu8_epi16(_mm_srli_si128(weights_hilo, 8));
2898
1.45k
  const __m128i weights7 = cvtepu8_epi16(weights_hihi);
2899
1.45k
  const __m128i weights8 = cvtepu8_epi16(_mm_srli_si128(weights_hihi, 8));
2900
1.45k
  const __m128i inverted_weights5 = _mm_sub_epi16(scale, weights5);
2901
1.45k
  const __m128i inverted_weights6 = _mm_sub_epi16(scale, weights6);
2902
1.45k
  const __m128i inverted_weights7 = _mm_sub_epi16(scale, weights7);
2903
1.45k
  const __m128i inverted_weights8 = _mm_sub_epi16(scale, weights8);
2904
1.45k
  const __m128i scaled_top_right5 =
2905
1.45k
      _mm_mullo_epi16(inverted_weights5, top_right);
2906
1.45k
  const __m128i scaled_top_right6 =
2907
1.45k
      _mm_mullo_epi16(inverted_weights6, top_right);
2908
1.45k
  const __m128i scaled_top_right7 =
2909
1.45k
      _mm_mullo_epi16(inverted_weights7, top_right);
2910
1.45k
  const __m128i scaled_top_right8 =
2911
1.45k
      _mm_mullo_epi16(inverted_weights8, top_right);
2912
1.45k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
2913
13.0k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2914
11.6k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2915
11.6k
    const __m128i left_y = _mm_shuffle_epi8(left1, y_select);
2916
11.6k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2917
11.6k
                                   scaled_top_right1, scaled_top_right2, round);
2918
11.6k
    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
2919
11.6k
                                   scaled_top_right3, scaled_top_right4, round);
2920
11.6k
    write_smooth_directional_sum16(dst + 32, left_y, left_y, weights5, weights6,
2921
11.6k
                                   scaled_top_right5, scaled_top_right6, round);
2922
11.6k
    write_smooth_directional_sum16(dst + 48, left_y, left_y, weights7, weights8,
2923
11.6k
                                   scaled_top_right7, scaled_top_right8, round);
2924
11.6k
    dst += stride;
2925
11.6k
  }
2926
1.45k
  const __m128i left2 = cvtepu8_epi16(LoadLo8(left_column + 8));
2927
13.0k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2928
11.6k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2929
11.6k
    const __m128i left_y = _mm_shuffle_epi8(left2, y_select);
2930
11.6k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2931
11.6k
                                   scaled_top_right1, scaled_top_right2, round);
2932
11.6k
    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
2933
11.6k
                                   scaled_top_right3, scaled_top_right4, round);
2934
11.6k
    write_smooth_directional_sum16(dst + 32, left_y, left_y, weights5, weights6,
2935
11.6k
                                   scaled_top_right5, scaled_top_right6, round);
2936
11.6k
    write_smooth_directional_sum16(dst + 48, left_y, left_y, weights7, weights8,
2937
11.6k
                                   scaled_top_right7, scaled_top_right8, round);
2938
11.6k
    dst += stride;
2939
11.6k
  }
2940
1.45k
  const __m128i left3 = cvtepu8_epi16(LoadLo8(left_column + 16));
2941
13.0k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2942
11.6k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2943
11.6k
    const __m128i left_y = _mm_shuffle_epi8(left3, y_select);
2944
11.6k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2945
11.6k
                                   scaled_top_right1, scaled_top_right2, round);
2946
11.6k
    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
2947
11.6k
                                   scaled_top_right3, scaled_top_right4, round);
2948
11.6k
    write_smooth_directional_sum16(dst + 32, left_y, left_y, weights5, weights6,
2949
11.6k
                                   scaled_top_right5, scaled_top_right6, round);
2950
11.6k
    write_smooth_directional_sum16(dst + 48, left_y, left_y, weights7, weights8,
2951
11.6k
                                   scaled_top_right7, scaled_top_right8, round);
2952
11.6k
    dst += stride;
2953
11.6k
  }
2954
1.45k
  const __m128i left4 = cvtepu8_epi16(LoadLo8(left_column + 24));
2955
13.0k
  for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
2956
11.6k
    const __m128i y_select = _mm_set1_epi32(y_mask);
2957
11.6k
    const __m128i left_y = _mm_shuffle_epi8(left4, y_select);
2958
11.6k
    write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
2959
11.6k
                                   scaled_top_right1, scaled_top_right2, round);
2960
11.6k
    write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3, weights4,
2961
11.6k
                                   scaled_top_right3, scaled_top_right4, round);
2962
11.6k
    write_smooth_directional_sum16(dst + 32, left_y, left_y, weights5, weights6,
2963
11.6k
                                   scaled_top_right5, scaled_top_right6, round);
2964
11.6k
    write_smooth_directional_sum16(dst + 48, left_y, left_y, weights7, weights8,
2965
11.6k
                                   scaled_top_right7, scaled_top_right8, round);
2966
11.6k
    dst += stride;
2967
11.6k
  }
2968
1.45k
}
2969
2970
void aom_smooth_h_predictor_64x64_ssse3(
2971
    uint8_t *LIBAOM_RESTRICT dst, ptrdiff_t stride,
2972
    const uint8_t *LIBAOM_RESTRICT top_row,
2973
4.67k
    const uint8_t *LIBAOM_RESTRICT left_column) {
2974
4.67k
  const __m128i top_right = _mm_set1_epi16(top_row[63]);
2975
4.67k
  const __m128i weights_lolo = LoadUnaligned16(smooth_weights + 60);
2976
4.67k
  const __m128i weights_lohi = LoadUnaligned16(smooth_weights + 76);
2977
4.67k
  const __m128i scale = _mm_set1_epi16(1 << SMOOTH_WEIGHT_LOG2_SCALE);
2978
4.67k
  const __m128i weights1 = cvtepu8_epi16(weights_lolo);
2979
4.67k
  const __m128i weights2 = cvtepu8_epi16(_mm_srli_si128(weights_lolo, 8));
2980
4.67k
  const __m128i weights3 = cvtepu8_epi16(weights_lohi);
2981
4.67k
  const __m128i weights4 = cvtepu8_epi16(_mm_srli_si128(weights_lohi, 8));
2982
4.67k
  const __m128i inverted_weights1 = _mm_sub_epi16(scale, weights1);
2983
4.67k
  const __m128i inverted_weights2 = _mm_sub_epi16(scale, weights2);
2984
4.67k
  const __m128i inverted_weights3 = _mm_sub_epi16(scale, weights3);
2985
4.67k
  const __m128i inverted_weights4 = _mm_sub_epi16(scale, weights4);
2986
4.67k
  const __m128i scaled_top_right1 =
2987
4.67k
      _mm_mullo_epi16(inverted_weights1, top_right);
2988
4.67k
  const __m128i scaled_top_right2 =
2989
4.67k
      _mm_mullo_epi16(inverted_weights2, top_right);
2990
4.67k
  const __m128i scaled_top_right3 =
2991
4.67k
      _mm_mullo_epi16(inverted_weights3, top_right);
2992
4.67k
  const __m128i scaled_top_right4 =
2993
4.67k
      _mm_mullo_epi16(inverted_weights4, top_right);
2994
4.67k
  const __m128i weights_hilo = LoadUnaligned16(smooth_weights + 92);
2995
4.67k
  const __m128i weights_hihi = LoadUnaligned16(smooth_weights + 108);
2996
4.67k
  const __m128i weights5 = cvtepu8_epi16(weights_hilo);
2997
4.67k
  const __m128i weights6 = cvtepu8_epi16(_mm_srli_si128(weights_hilo, 8));
2998
4.67k
  const __m128i weights7 = cvtepu8_epi16(weights_hihi);
2999
4.67k
  const __m128i weights8 = cvtepu8_epi16(_mm_srli_si128(weights_hihi, 8));
3000
4.67k
  const __m128i inverted_weights5 = _mm_sub_epi16(scale, weights5);
3001
4.67k
  const __m128i inverted_weights6 = _mm_sub_epi16(scale, weights6);
3002
4.67k
  const __m128i inverted_weights7 = _mm_sub_epi16(scale, weights7);
3003
4.67k
  const __m128i inverted_weights8 = _mm_sub_epi16(scale, weights8);
3004
4.67k
  const __m128i scaled_top_right5 =
3005
4.67k
      _mm_mullo_epi16(inverted_weights5, top_right);
3006
4.67k
  const __m128i scaled_top_right6 =
3007
4.67k
      _mm_mullo_epi16(inverted_weights6, top_right);
3008
4.67k
  const __m128i scaled_top_right7 =
3009
4.67k
      _mm_mullo_epi16(inverted_weights7, top_right);
3010
4.67k
  const __m128i scaled_top_right8 =
3011
4.67k
      _mm_mullo_epi16(inverted_weights8, top_right);
3012
4.67k
  const __m128i round = _mm_set1_epi16(1 << (SMOOTH_WEIGHT_LOG2_SCALE - 1));
3013
42.0k
  for (int left_offset = 0; left_offset < 64; left_offset += 8) {
3014
37.3k
    const __m128i left = cvtepu8_epi16(LoadLo8(left_column + left_offset));
3015
336k
    for (int y_mask = 0x01000100; y_mask < 0x0F0E0F0F; y_mask += 0x02020202) {
3016
299k
      const __m128i y_select = _mm_set1_epi32(y_mask);
3017
299k
      const __m128i left_y = _mm_shuffle_epi8(left, y_select);
3018
299k
      write_smooth_directional_sum16(dst, left_y, left_y, weights1, weights2,
3019
299k
                                     scaled_top_right1, scaled_top_right2,
3020
299k
                                     round);
3021
299k
      write_smooth_directional_sum16(dst + 16, left_y, left_y, weights3,
3022
299k
                                     weights4, scaled_top_right3,
3023
299k
                                     scaled_top_right4, round);
3024
299k
      write_smooth_directional_sum16(dst + 32, left_y, left_y, weights5,
3025
299k
                                     weights6, scaled_top_right5,
3026
299k
                                     scaled_top_right6, round);
3027
299k
      write_smooth_directional_sum16(dst + 48, left_y, left_y, weights7,
3028
299k
                                     weights8, scaled_top_right7,
3029
299k
                                     scaled_top_right8, round);
3030
299k
      dst += stride;
3031
299k
    }
3032
37.3k
  }
3033
4.67k
}