1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
|
/*
* Copyright (c) 2016, Alliance for Open Media. All rights reserved
*
* This source code is subject to the terms of the BSD 2 Clause License and
* the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
* was not distributed with this source code in the LICENSE file, you can
* obtain it at www.aomedia.org/license/software. If the Alliance for Open
* Media Patent License 1.0 was not distributed with this source code in the
* PATENTS file, you can obtain it at www.aomedia.org/license/patent.
*/
#include <smmintrin.h>
#include "config/av1_rtcd.h"
#include "av1/common/warped_motion.h"
static const uint8_t warp_highbd_arrange_bytes[16] = {
0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15
};
static INLINE void horizontal_filter(__m128i src, __m128i src2, __m128i *tmp,
int sx, int alpha, int k,
const int offset_bits_horiz,
const int reduce_bits_horiz) {
// Filter even-index pixels
const __m128i tmp_0 = _mm_loadu_si128(
(__m128i *)(warped_filter + ((sx + 0 * alpha) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_2 = _mm_loadu_si128(
(__m128i *)(warped_filter + ((sx + 2 * alpha) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_4 = _mm_loadu_si128(
(__m128i *)(warped_filter + ((sx + 4 * alpha) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_6 = _mm_loadu_si128(
(__m128i *)(warped_filter + ((sx + 6 * alpha) >> WARPEDDIFF_PREC_BITS)));
// coeffs 0 1 0 1 2 3 2 3 for pixels 0, 2
const __m128i tmp_8 = _mm_unpacklo_epi32(tmp_0, tmp_2);
// coeffs 0 1 0 1 2 3 2 3 for pixels 4, 6
const __m128i tmp_10 = _mm_unpacklo_epi32(tmp_4, tmp_6);
// coeffs 4 5 4 5 6 7 6 7 for pixels 0, 2
const __m128i tmp_12 = _mm_unpackhi_epi32(tmp_0, tmp_2);
// coeffs 4 5 4 5 6 7 6 7 for pixels 4, 6
const __m128i tmp_14 = _mm_unpackhi_epi32(tmp_4, tmp_6);
// coeffs 0 1 0 1 0 1 0 1 for pixels 0, 2, 4, 6
const __m128i coeff_0 = _mm_unpacklo_epi64(tmp_8, tmp_10);
// coeffs 2 3 2 3 2 3 2 3 for pixels 0, 2, 4, 6
const __m128i coeff_2 = _mm_unpackhi_epi64(tmp_8, tmp_10);
// coeffs 4 5 4 5 4 5 4 5 for pixels 0, 2, 4, 6
const __m128i coeff_4 = _mm_unpacklo_epi64(tmp_12, tmp_14);
// coeffs 6 7 6 7 6 7 6 7 for pixels 0, 2, 4, 6
const __m128i coeff_6 = _mm_unpackhi_epi64(tmp_12, tmp_14);
const __m128i round_const = _mm_set1_epi32((1 << offset_bits_horiz) +
((1 << reduce_bits_horiz) >> 1));
// Calculate filtered results
const __m128i res_0 = _mm_madd_epi16(src, coeff_0);
const __m128i res_2 = _mm_madd_epi16(_mm_alignr_epi8(src2, src, 4), coeff_2);
const __m128i res_4 = _mm_madd_epi16(_mm_alignr_epi8(src2, src, 8), coeff_4);
const __m128i res_6 = _mm_madd_epi16(_mm_alignr_epi8(src2, src, 12), coeff_6);
__m128i res_even =
_mm_add_epi32(_mm_add_epi32(res_0, res_4), _mm_add_epi32(res_2, res_6));
res_even = _mm_sra_epi32(_mm_add_epi32(res_even, round_const),
_mm_cvtsi32_si128(reduce_bits_horiz));
// Filter odd-index pixels
const __m128i tmp_1 = _mm_loadu_si128(
(__m128i *)(warped_filter + ((sx + 1 * alpha) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_3 = _mm_loadu_si128(
(__m128i *)(warped_filter + ((sx + 3 * alpha) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_5 = _mm_loadu_si128(
(__m128i *)(warped_filter + ((sx + 5 * alpha) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_7 = _mm_loadu_si128(
(__m128i *)(warped_filter + ((sx + 7 * alpha) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_9 = _mm_unpacklo_epi32(tmp_1, tmp_3);
const __m128i tmp_11 = _mm_unpacklo_epi32(tmp_5, tmp_7);
const __m128i tmp_13 = _mm_unpackhi_epi32(tmp_1, tmp_3);
const __m128i tmp_15 = _mm_unpackhi_epi32(tmp_5, tmp_7);
const __m128i coeff_1 = _mm_unpacklo_epi64(tmp_9, tmp_11);
const __m128i coeff_3 = _mm_unpackhi_epi64(tmp_9, tmp_11);
const __m128i coeff_5 = _mm_unpacklo_epi64(tmp_13, tmp_15);
const __m128i coeff_7 = _mm_unpackhi_epi64(tmp_13, tmp_15);
const __m128i res_1 = _mm_madd_epi16(_mm_alignr_epi8(src2, src, 2), coeff_1);
const __m128i res_3 = _mm_madd_epi16(_mm_alignr_epi8(src2, src, 6), coeff_3);
const __m128i res_5 = _mm_madd_epi16(_mm_alignr_epi8(src2, src, 10), coeff_5);
const __m128i res_7 = _mm_madd_epi16(_mm_alignr_epi8(src2, src, 14), coeff_7);
__m128i res_odd =
_mm_add_epi32(_mm_add_epi32(res_1, res_5), _mm_add_epi32(res_3, res_7));
res_odd = _mm_sra_epi32(_mm_add_epi32(res_odd, round_const),
_mm_cvtsi32_si128(reduce_bits_horiz));
// Combine results into one register.
// We store the columns in the order 0, 2, 4, 6, 1, 3, 5, 7
// as this order helps with the vertical filter.
tmp[k + 7] = _mm_packs_epi32(res_even, res_odd);
}
void av1_highbd_warp_affine_sse4_1(const int32_t *mat, const uint16_t *ref,
int width, int height, int stride,
uint16_t *pred, int p_col, int p_row,
int p_width, int p_height, int p_stride,
int subsampling_x, int subsampling_y, int bd,
ConvolveParams *conv_params, int16_t alpha,
int16_t beta, int16_t gamma, int16_t delta) {
__m128i tmp[15];
int i, j, k;
const int reduce_bits_horiz =
conv_params->round_0 +
AOMMAX(bd + FILTER_BITS - conv_params->round_0 - 14, 0);
const int reduce_bits_vert = conv_params->is_compound
? conv_params->round_1
: 2 * FILTER_BITS - reduce_bits_horiz;
const int offset_bits_horiz = bd + FILTER_BITS - 1;
assert(IMPLIES(conv_params->is_compound, conv_params->dst != NULL));
assert(!(bd == 12 && reduce_bits_horiz < 5));
assert(IMPLIES(conv_params->do_average, conv_params->is_compound));
const int offset_bits_vert = bd + 2 * FILTER_BITS - reduce_bits_horiz;
const __m128i clip_pixel =
_mm_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
const __m128i reduce_bits_vert_shift = _mm_cvtsi32_si128(reduce_bits_vert);
const __m128i reduce_bits_vert_const =
_mm_set1_epi32(((1 << reduce_bits_vert) >> 1));
const __m128i res_add_const = _mm_set1_epi32(1 << offset_bits_vert);
const int round_bits =
2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
const __m128i res_sub_const =
_mm_set1_epi32(-(1 << (offset_bits - conv_params->round_1)) -
(1 << (offset_bits - conv_params->round_1 - 1)));
__m128i round_bits_shift = _mm_cvtsi32_si128(round_bits);
__m128i round_bits_const = _mm_set1_epi32(((1 << round_bits) >> 1));
const int w0 = conv_params->fwd_offset;
const int w1 = conv_params->bck_offset;
const __m128i wt0 = _mm_set1_epi32(w0);
const __m128i wt1 = _mm_set1_epi32(w1);
/* Note: For this code to work, the left/right frame borders need to be
extended by at least 13 pixels each. By the time we get here, other
code will have set up this border, but we allow an explicit check
for debugging purposes.
*/
/*for (i = 0; i < height; ++i) {
for (j = 0; j < 13; ++j) {
assert(ref[i * stride - 13 + j] == ref[i * stride]);
assert(ref[i * stride + width + j] == ref[i * stride + (width - 1)]);
}
}*/
for (i = 0; i < p_height; i += 8) {
for (j = 0; j < p_width; j += 8) {
const int32_t src_x = (p_col + j + 4) << subsampling_x;
const int32_t src_y = (p_row + i + 4) << subsampling_y;
const int32_t dst_x = mat[2] * src_x + mat[3] * src_y + mat[0];
const int32_t dst_y = mat[4] * src_x + mat[5] * src_y + mat[1];
const int32_t x4 = dst_x >> subsampling_x;
const int32_t y4 = dst_y >> subsampling_y;
int32_t ix4 = x4 >> WARPEDMODEL_PREC_BITS;
int32_t sx4 = x4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
int32_t iy4 = y4 >> WARPEDMODEL_PREC_BITS;
int32_t sy4 = y4 & ((1 << WARPEDMODEL_PREC_BITS) - 1);
// Add in all the constant terms, including rounding and offset
sx4 += alpha * (-4) + beta * (-4) + (1 << (WARPEDDIFF_PREC_BITS - 1)) +
(WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
sy4 += gamma * (-4) + delta * (-4) + (1 << (WARPEDDIFF_PREC_BITS - 1)) +
(WARPEDPIXEL_PREC_SHIFTS << WARPEDDIFF_PREC_BITS);
sx4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1);
sy4 &= ~((1 << WARP_PARAM_REDUCE_BITS) - 1);
// Horizontal filter
// If the block is aligned such that, after clamping, every sample
// would be taken from the leftmost/rightmost column, then we can
// skip the expensive horizontal filter.
if (ix4 <= -7) {
for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
int iy = iy4 + k;
if (iy < 0)
iy = 0;
else if (iy > height - 1)
iy = height - 1;
tmp[k + 7] = _mm_set1_epi16(
(1 << (bd + FILTER_BITS - reduce_bits_horiz - 1)) +
ref[iy * stride] * (1 << (FILTER_BITS - reduce_bits_horiz)));
}
} else if (ix4 >= width + 6) {
for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
int iy = iy4 + k;
if (iy < 0)
iy = 0;
else if (iy > height - 1)
iy = height - 1;
tmp[k + 7] =
_mm_set1_epi16((1 << (bd + FILTER_BITS - reduce_bits_horiz - 1)) +
ref[iy * stride + (width - 1)] *
(1 << (FILTER_BITS - reduce_bits_horiz)));
}
} else if (((ix4 - 7) < 0) || ((ix4 + 9) > width)) {
const int out_of_boundary_left = -(ix4 - 6);
const int out_of_boundary_right = (ix4 + 8) - width;
for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
int iy = iy4 + k;
if (iy < 0)
iy = 0;
else if (iy > height - 1)
iy = height - 1;
int sx = sx4 + beta * (k + 4);
// Load source pixels
const __m128i src =
_mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
const __m128i src2 =
_mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 + 1));
const __m128i src_01 = _mm_shuffle_epi8(
src, _mm_loadu_si128((__m128i *)warp_highbd_arrange_bytes));
const __m128i src2_01 = _mm_shuffle_epi8(
src2, _mm_loadu_si128((__m128i *)warp_highbd_arrange_bytes));
__m128i src_lo = _mm_unpacklo_epi64(src_01, src2_01);
__m128i src_hi = _mm_unpackhi_epi64(src_01, src2_01);
if (out_of_boundary_left >= 0) {
const __m128i shuffle_reg_left =
_mm_loadu_si128((__m128i *)warp_pad_left[out_of_boundary_left]);
src_lo = _mm_shuffle_epi8(src_lo, shuffle_reg_left);
src_hi = _mm_shuffle_epi8(src_hi, shuffle_reg_left);
}
if (out_of_boundary_right >= 0) {
const __m128i shuffle_reg_right = _mm_loadu_si128(
(__m128i *)warp_pad_right[out_of_boundary_right]);
src_lo = _mm_shuffle_epi8(src_lo, shuffle_reg_right);
src_hi = _mm_shuffle_epi8(src_hi, shuffle_reg_right);
}
const __m128i src_padded = _mm_unpacklo_epi8(src_lo, src_hi);
const __m128i src2_padded = _mm_unpackhi_epi8(src_lo, src_hi);
horizontal_filter(src_padded, src2_padded, tmp, sx, alpha, k,
offset_bits_horiz, reduce_bits_horiz);
}
} else {
for (k = -7; k < AOMMIN(8, p_height - i); ++k) {
int iy = iy4 + k;
if (iy < 0)
iy = 0;
else if (iy > height - 1)
iy = height - 1;
int sx = sx4 + beta * (k + 4);
// Load source pixels
const __m128i src =
_mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 - 7));
const __m128i src2 =
_mm_loadu_si128((__m128i *)(ref + iy * stride + ix4 + 1));
horizontal_filter(src, src2, tmp, sx, alpha, k, offset_bits_horiz,
reduce_bits_horiz);
}
}
// Vertical filter
for (k = -4; k < AOMMIN(4, p_height - i - 4); ++k) {
int sy = sy4 + delta * (k + 4);
// Load from tmp and rearrange pairs of consecutive rows into the
// column order 0 0 2 2 4 4 6 6; 1 1 3 3 5 5 7 7
const __m128i *src = tmp + (k + 4);
const __m128i src_0 = _mm_unpacklo_epi16(src[0], src[1]);
const __m128i src_2 = _mm_unpacklo_epi16(src[2], src[3]);
const __m128i src_4 = _mm_unpacklo_epi16(src[4], src[5]);
const __m128i src_6 = _mm_unpacklo_epi16(src[6], src[7]);
// Filter even-index pixels
const __m128i tmp_0 = _mm_loadu_si128(
(__m128i *)(warped_filter +
((sy + 0 * gamma) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_2 = _mm_loadu_si128(
(__m128i *)(warped_filter +
((sy + 2 * gamma) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_4 = _mm_loadu_si128(
(__m128i *)(warped_filter +
((sy + 4 * gamma) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_6 = _mm_loadu_si128(
(__m128i *)(warped_filter +
((sy + 6 * gamma) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_8 = _mm_unpacklo_epi32(tmp_0, tmp_2);
const __m128i tmp_10 = _mm_unpacklo_epi32(tmp_4, tmp_6);
const __m128i tmp_12 = _mm_unpackhi_epi32(tmp_0, tmp_2);
const __m128i tmp_14 = _mm_unpackhi_epi32(tmp_4, tmp_6);
const __m128i coeff_0 = _mm_unpacklo_epi64(tmp_8, tmp_10);
const __m128i coeff_2 = _mm_unpackhi_epi64(tmp_8, tmp_10);
const __m128i coeff_4 = _mm_unpacklo_epi64(tmp_12, tmp_14);
const __m128i coeff_6 = _mm_unpackhi_epi64(tmp_12, tmp_14);
const __m128i res_0 = _mm_madd_epi16(src_0, coeff_0);
const __m128i res_2 = _mm_madd_epi16(src_2, coeff_2);
const __m128i res_4 = _mm_madd_epi16(src_4, coeff_4);
const __m128i res_6 = _mm_madd_epi16(src_6, coeff_6);
const __m128i res_even = _mm_add_epi32(_mm_add_epi32(res_0, res_2),
_mm_add_epi32(res_4, res_6));
// Filter odd-index pixels
const __m128i src_1 = _mm_unpackhi_epi16(src[0], src[1]);
const __m128i src_3 = _mm_unpackhi_epi16(src[2], src[3]);
const __m128i src_5 = _mm_unpackhi_epi16(src[4], src[5]);
const __m128i src_7 = _mm_unpackhi_epi16(src[6], src[7]);
const __m128i tmp_1 = _mm_loadu_si128(
(__m128i *)(warped_filter +
((sy + 1 * gamma) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_3 = _mm_loadu_si128(
(__m128i *)(warped_filter +
((sy + 3 * gamma) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_5 = _mm_loadu_si128(
(__m128i *)(warped_filter +
((sy + 5 * gamma) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_7 = _mm_loadu_si128(
(__m128i *)(warped_filter +
((sy + 7 * gamma) >> WARPEDDIFF_PREC_BITS)));
const __m128i tmp_9 = _mm_unpacklo_epi32(tmp_1, tmp_3);
const __m128i tmp_11 = _mm_unpacklo_epi32(tmp_5, tmp_7);
const __m128i tmp_13 = _mm_unpackhi_epi32(tmp_1, tmp_3);
const __m128i tmp_15 = _mm_unpackhi_epi32(tmp_5, tmp_7);
const __m128i coeff_1 = _mm_unpacklo_epi64(tmp_9, tmp_11);
const __m128i coeff_3 = _mm_unpackhi_epi64(tmp_9, tmp_11);
const __m128i coeff_5 = _mm_unpacklo_epi64(tmp_13, tmp_15);
const __m128i coeff_7 = _mm_unpackhi_epi64(tmp_13, tmp_15);
const __m128i res_1 = _mm_madd_epi16(src_1, coeff_1);
const __m128i res_3 = _mm_madd_epi16(src_3, coeff_3);
const __m128i res_5 = _mm_madd_epi16(src_5, coeff_5);
const __m128i res_7 = _mm_madd_epi16(src_7, coeff_7);
const __m128i res_odd = _mm_add_epi32(_mm_add_epi32(res_1, res_3),
_mm_add_epi32(res_5, res_7));
// Rearrange pixels back into the order 0 ... 7
__m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
__m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
if (conv_params->is_compound) {
__m128i *const p =
(__m128i *)&conv_params
->dst[(i + k + 4) * conv_params->dst_stride + j];
res_lo = _mm_add_epi32(res_lo, res_add_const);
res_lo = _mm_sra_epi32(_mm_add_epi32(res_lo, reduce_bits_vert_const),
reduce_bits_vert_shift);
if (conv_params->do_average) {
__m128i *const dst16 = (__m128i *)&pred[(i + k + 4) * p_stride + j];
__m128i p_32 = _mm_cvtepu16_epi32(_mm_loadl_epi64(p));
if (conv_params->use_jnt_comp_avg) {
res_lo = _mm_add_epi32(_mm_mullo_epi32(p_32, wt0),
_mm_mullo_epi32(res_lo, wt1));
res_lo = _mm_srai_epi32(res_lo, DIST_PRECISION_BITS);
} else {
res_lo = _mm_srai_epi32(_mm_add_epi32(p_32, res_lo), 1);
}
__m128i res32_lo = _mm_add_epi32(res_lo, res_sub_const);
res32_lo = _mm_sra_epi32(_mm_add_epi32(res32_lo, round_bits_const),
round_bits_shift);
__m128i res16_lo = _mm_packus_epi32(res32_lo, res32_lo);
res16_lo = _mm_min_epi16(res16_lo, clip_pixel);
_mm_storel_epi64(dst16, res16_lo);
} else {
res_lo = _mm_packus_epi32(res_lo, res_lo);
_mm_storel_epi64(p, res_lo);
}
if (p_width > 4) {
__m128i *const p4 =
(__m128i *)&conv_params
->dst[(i + k + 4) * conv_params->dst_stride + j + 4];
res_hi = _mm_add_epi32(res_hi, res_add_const);
res_hi =
_mm_sra_epi32(_mm_add_epi32(res_hi, reduce_bits_vert_const),
reduce_bits_vert_shift);
if (conv_params->do_average) {
__m128i *const dst16_4 =
(__m128i *)&pred[(i + k + 4) * p_stride + j + 4];
__m128i p4_32 = _mm_cvtepu16_epi32(_mm_loadl_epi64(p4));
if (conv_params->use_jnt_comp_avg) {
res_hi = _mm_add_epi32(_mm_mullo_epi32(p4_32, wt0),
_mm_mullo_epi32(res_hi, wt1));
res_hi = _mm_srai_epi32(res_hi, DIST_PRECISION_BITS);
} else {
res_hi = _mm_srai_epi32(_mm_add_epi32(p4_32, res_hi), 1);
}
__m128i res32_hi = _mm_add_epi32(res_hi, res_sub_const);
res32_hi = _mm_sra_epi32(
_mm_add_epi32(res32_hi, round_bits_const), round_bits_shift);
__m128i res16_hi = _mm_packus_epi32(res32_hi, res32_hi);
res16_hi = _mm_min_epi16(res16_hi, clip_pixel);
_mm_storel_epi64(dst16_4, res16_hi);
} else {
res_hi = _mm_packus_epi32(res_hi, res_hi);
_mm_storel_epi64(p4, res_hi);
}
}
} else {
// Round and pack into 8 bits
const __m128i round_const =
_mm_set1_epi32(-(1 << (bd + reduce_bits_vert - 1)) +
((1 << reduce_bits_vert) >> 1));
const __m128i res_lo_round = _mm_srai_epi32(
_mm_add_epi32(res_lo, round_const), reduce_bits_vert);
const __m128i res_hi_round = _mm_srai_epi32(
_mm_add_epi32(res_hi, round_const), reduce_bits_vert);
__m128i res_16bit = _mm_packs_epi32(res_lo_round, res_hi_round);
// Clamp res_16bit to the range [0, 2^bd - 1]
const __m128i max_val = _mm_set1_epi16((1 << bd) - 1);
const __m128i zero = _mm_setzero_si128();
res_16bit = _mm_max_epi16(_mm_min_epi16(res_16bit, max_val), zero);
// Store, blending with 'pred' if needed
__m128i *const p = (__m128i *)&pred[(i + k + 4) * p_stride + j];
// Note: If we're outputting a 4x4 block, we need to be very careful
// to only output 4 pixels at this point, to avoid encode/decode
// mismatches when encoding with multiple threads.
if (p_width == 4) {
_mm_storel_epi64(p, res_16bit);
} else {
_mm_storeu_si128(p, res_16bit);
}
}
}
}
}
}
|