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
|
/*
* Copyright (c) 2017, Alliance for Open Media. All rights reserved
*
* This source code is subject to the terms of the BSD 2 Clause License and
* the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
* was not distributed with this source code in the LICENSE file, you can
* obtain it at www.aomedia.org/license/software. If the Alliance for Open
* Media Patent License 1.0 was not distributed with this source code in the
* PATENTS file, you can obtain it at www.aomedia.org/license/patent.
*/
#include <immintrin.h>
#include "config/av1_rtcd.h"
#include "aom/aom_integer.h"
#include "aom_dsp/aom_dsp_common.h"
static INLINE void read_coeff(const tran_low_t *coeff, __m256i *c) {
if (sizeof(tran_low_t) == 4) {
const __m256i x0 = _mm256_loadu_si256((const __m256i *)coeff);
const __m256i x1 = _mm256_loadu_si256((const __m256i *)coeff + 1);
*c = _mm256_packs_epi32(x0, x1);
*c = _mm256_permute4x64_epi64(*c, 0xD8);
} else {
*c = _mm256_loadu_si256((const __m256i *)coeff);
}
}
static INLINE void write_zero(tran_low_t *qcoeff) {
const __m256i zero = _mm256_setzero_si256();
if (sizeof(tran_low_t) == 4) {
_mm256_storeu_si256((__m256i *)qcoeff, zero);
_mm256_storeu_si256((__m256i *)qcoeff + 1, zero);
} else {
_mm256_storeu_si256((__m256i *)qcoeff, zero);
}
}
static INLINE void init_one_qp(const __m128i *p, __m256i *qp) {
const __m128i ac = _mm_unpackhi_epi64(*p, *p);
*qp = _mm256_insertf128_si256(_mm256_castsi128_si256(*p), ac, 1);
}
static INLINE void init_qp(const int16_t *round_ptr, const int16_t *quant_ptr,
const int16_t *dequant_ptr, int log_scale,
__m256i *thr, __m256i *qp) {
__m128i round = _mm_loadu_si128((const __m128i *)round_ptr);
const __m128i quant = _mm_loadu_si128((const __m128i *)quant_ptr);
const __m128i dequant = _mm_loadu_si128((const __m128i *)dequant_ptr);
if (log_scale > 0) {
const __m128i rnd = _mm_set1_epi16((int16_t)1 << (log_scale - 1));
round = _mm_add_epi16(round, rnd);
round = _mm_srai_epi16(round, log_scale);
}
init_one_qp(&round, &qp[0]);
init_one_qp(&quant, &qp[1]);
if (log_scale == 1) {
qp[1] = _mm256_slli_epi16(qp[1], log_scale);
}
init_one_qp(&dequant, &qp[2]);
*thr = _mm256_srai_epi16(qp[2], 1 + log_scale);
}
static INLINE void update_qp(int log_scale, __m256i *thr, __m256i *qp) {
qp[0] = _mm256_permute2x128_si256(qp[0], qp[0], 0x11);
qp[1] = _mm256_permute2x128_si256(qp[1], qp[1], 0x11);
qp[2] = _mm256_permute2x128_si256(qp[2], qp[2], 0x11);
*thr = _mm256_srai_epi16(qp[2], 1 + log_scale);
}
#define store_quan(q, addr) \
do { \
__m256i sign_bits = _mm256_srai_epi16(q, 15); \
__m256i y0 = _mm256_unpacklo_epi16(q, sign_bits); \
__m256i y1 = _mm256_unpackhi_epi16(q, sign_bits); \
__m256i x0 = _mm256_permute2x128_si256(y0, y1, 0x20); \
__m256i x1 = _mm256_permute2x128_si256(y0, y1, 0x31); \
_mm256_storeu_si256((__m256i *)addr, x0); \
_mm256_storeu_si256((__m256i *)addr + 1, x1); \
} while (0)
#define store_two_quan(q, addr1, dq, addr2) \
do { \
if (sizeof(tran_low_t) == 4) { \
store_quan(q, addr1); \
store_quan(dq, addr2); \
} else { \
_mm256_storeu_si256((__m256i *)addr1, q); \
_mm256_storeu_si256((__m256i *)addr2, dq); \
} \
} while (0)
static INLINE uint16_t quant_gather_eob(__m256i eob) {
const __m128i eob_lo = _mm256_castsi256_si128(eob);
const __m128i eob_hi = _mm256_extractf128_si256(eob, 1);
__m128i eob_s = _mm_max_epi16(eob_lo, eob_hi);
eob_s = _mm_subs_epu16(_mm_set1_epi16(INT16_MAX), eob_s);
eob_s = _mm_minpos_epu16(eob_s);
return INT16_MAX - _mm_extract_epi16(eob_s, 0);
}
static INLINE void quantize(const __m256i *thr, const __m256i *qp, __m256i *c,
const int16_t *iscan_ptr, tran_low_t *qcoeff,
tran_low_t *dqcoeff, __m256i *eob) {
const __m256i abs_coeff = _mm256_abs_epi16(*c);
__m256i mask = _mm256_cmpgt_epi16(abs_coeff, *thr);
mask = _mm256_or_si256(mask, _mm256_cmpeq_epi16(abs_coeff, *thr));
const int nzflag = _mm256_movemask_epi8(mask);
if (nzflag) {
__m256i q = _mm256_adds_epi16(abs_coeff, qp[0]);
q = _mm256_mulhi_epi16(q, qp[1]);
q = _mm256_sign_epi16(q, *c);
const __m256i dq = _mm256_mullo_epi16(q, qp[2]);
store_two_quan(q, qcoeff, dq, dqcoeff);
const __m256i zero = _mm256_setzero_si256();
const __m256i iscan = _mm256_loadu_si256((const __m256i *)iscan_ptr);
const __m256i zero_coeff = _mm256_cmpeq_epi16(dq, zero);
const __m256i nzero_coeff = _mm256_cmpeq_epi16(zero_coeff, zero);
__m256i cur_eob = _mm256_sub_epi16(iscan, nzero_coeff);
cur_eob = _mm256_and_si256(cur_eob, nzero_coeff);
*eob = _mm256_max_epi16(*eob, cur_eob);
} else {
write_zero(qcoeff);
write_zero(dqcoeff);
}
}
void av1_quantize_fp_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
const int16_t *zbin_ptr, const int16_t *round_ptr,
const int16_t *quant_ptr,
const int16_t *quant_shift_ptr,
tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
const int16_t *dequant_ptr, uint16_t *eob_ptr,
const int16_t *scan_ptr, const int16_t *iscan_ptr) {
(void)scan_ptr;
(void)zbin_ptr;
(void)quant_shift_ptr;
const unsigned int step = 16;
__m256i qp[3];
__m256i coeff, thr;
const int log_scale = 0;
init_qp(round_ptr, quant_ptr, dequant_ptr, log_scale, &thr, qp);
read_coeff(coeff_ptr, &coeff);
__m256i eob = _mm256_setzero_si256();
quantize(&thr, qp, &coeff, iscan_ptr, qcoeff_ptr, dqcoeff_ptr, &eob);
coeff_ptr += step;
qcoeff_ptr += step;
dqcoeff_ptr += step;
iscan_ptr += step;
n_coeffs -= step;
update_qp(log_scale, &thr, qp);
while (n_coeffs > 0) {
read_coeff(coeff_ptr, &coeff);
quantize(&thr, qp, &coeff, iscan_ptr, qcoeff_ptr, dqcoeff_ptr, &eob);
coeff_ptr += step;
qcoeff_ptr += step;
dqcoeff_ptr += step;
iscan_ptr += step;
n_coeffs -= step;
}
*eob_ptr = quant_gather_eob(eob);
}
static INLINE void quantize_32x32(const __m256i *thr, const __m256i *qp,
__m256i *c, const int16_t *iscan_ptr,
tran_low_t *qcoeff, tran_low_t *dqcoeff,
__m256i *eob) {
const __m256i abs_coeff = _mm256_abs_epi16(*c);
__m256i mask = _mm256_cmpgt_epi16(abs_coeff, *thr);
mask = _mm256_or_si256(mask, _mm256_cmpeq_epi16(abs_coeff, *thr));
const int nzflag = _mm256_movemask_epi8(mask);
if (nzflag) {
__m256i q = _mm256_adds_epi16(abs_coeff, qp[0]);
q = _mm256_mulhi_epu16(q, qp[1]);
__m256i dq = _mm256_mullo_epi16(q, qp[2]);
dq = _mm256_srli_epi16(dq, 1);
q = _mm256_sign_epi16(q, *c);
dq = _mm256_sign_epi16(dq, *c);
store_two_quan(q, qcoeff, dq, dqcoeff);
const __m256i zero = _mm256_setzero_si256();
const __m256i iscan = _mm256_loadu_si256((const __m256i *)iscan_ptr);
const __m256i zero_coeff = _mm256_cmpeq_epi16(dq, zero);
const __m256i nzero_coeff = _mm256_cmpeq_epi16(zero_coeff, zero);
__m256i cur_eob = _mm256_sub_epi16(iscan, nzero_coeff);
cur_eob = _mm256_and_si256(cur_eob, nzero_coeff);
*eob = _mm256_max_epi16(*eob, cur_eob);
} else {
write_zero(qcoeff);
write_zero(dqcoeff);
}
}
void av1_quantize_fp_32x32_avx2(
const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
const int16_t *round_ptr, const int16_t *quant_ptr,
const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
const int16_t *scan_ptr, const int16_t *iscan_ptr) {
(void)scan_ptr;
(void)zbin_ptr;
(void)quant_shift_ptr;
const unsigned int step = 16;
__m256i qp[3];
__m256i coeff, thr;
const int log_scale = 1;
init_qp(round_ptr, quant_ptr, dequant_ptr, log_scale, &thr, qp);
read_coeff(coeff_ptr, &coeff);
__m256i eob = _mm256_setzero_si256();
quantize_32x32(&thr, qp, &coeff, iscan_ptr, qcoeff_ptr, dqcoeff_ptr, &eob);
coeff_ptr += step;
qcoeff_ptr += step;
dqcoeff_ptr += step;
iscan_ptr += step;
n_coeffs -= step;
update_qp(log_scale, &thr, qp);
while (n_coeffs > 0) {
read_coeff(coeff_ptr, &coeff);
quantize_32x32(&thr, qp, &coeff, iscan_ptr, qcoeff_ptr, dqcoeff_ptr, &eob);
coeff_ptr += step;
qcoeff_ptr += step;
dqcoeff_ptr += step;
iscan_ptr += step;
n_coeffs -= step;
}
*eob_ptr = quant_gather_eob(eob);
}
static INLINE void quantize_64x64(const __m256i *thr, const __m256i *qp,
__m256i *c, const int16_t *iscan_ptr,
tran_low_t *qcoeff, tran_low_t *dqcoeff,
__m256i *eob) {
const __m256i abs_coeff = _mm256_abs_epi16(*c);
__m256i mask = _mm256_cmpgt_epi16(abs_coeff, *thr);
mask = _mm256_or_si256(mask, _mm256_cmpeq_epi16(abs_coeff, *thr));
const int nzflag = _mm256_movemask_epi8(mask);
if (nzflag) {
__m256i q = _mm256_adds_epi16(abs_coeff, qp[0]);
__m256i qh = _mm256_mulhi_epi16(q, qp[1]);
__m256i ql = _mm256_mullo_epi16(q, qp[1]);
qh = _mm256_slli_epi16(qh, 2);
ql = _mm256_srli_epi16(ql, 14);
q = _mm256_or_si256(qh, ql);
const __m256i dqh = _mm256_slli_epi16(_mm256_mulhi_epi16(q, qp[2]), 14);
const __m256i dql = _mm256_srli_epi16(_mm256_mullo_epi16(q, qp[2]), 2);
__m256i dq = _mm256_or_si256(dqh, dql);
q = _mm256_sign_epi16(q, *c);
dq = _mm256_sign_epi16(dq, *c);
store_two_quan(q, qcoeff, dq, dqcoeff);
const __m256i zero = _mm256_setzero_si256();
const __m256i iscan = _mm256_loadu_si256((const __m256i *)iscan_ptr);
const __m256i zero_coeff = _mm256_cmpeq_epi16(dq, zero);
const __m256i nzero_coeff = _mm256_cmpeq_epi16(zero_coeff, zero);
__m256i cur_eob = _mm256_sub_epi16(iscan, nzero_coeff);
cur_eob = _mm256_and_si256(cur_eob, nzero_coeff);
*eob = _mm256_max_epi16(*eob, cur_eob);
} else {
write_zero(qcoeff);
write_zero(dqcoeff);
}
}
void av1_quantize_fp_64x64_avx2(
const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
const int16_t *round_ptr, const int16_t *quant_ptr,
const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
const int16_t *scan_ptr, const int16_t *iscan_ptr) {
(void)scan_ptr;
(void)zbin_ptr;
(void)quant_shift_ptr;
const unsigned int step = 16;
__m256i qp[3];
__m256i coeff, thr;
const int log_scale = 2;
init_qp(round_ptr, quant_ptr, dequant_ptr, log_scale, &thr, qp);
read_coeff(coeff_ptr, &coeff);
__m256i eob = _mm256_setzero_si256();
quantize_64x64(&thr, qp, &coeff, iscan_ptr, qcoeff_ptr, dqcoeff_ptr, &eob);
coeff_ptr += step;
qcoeff_ptr += step;
dqcoeff_ptr += step;
iscan_ptr += step;
n_coeffs -= step;
update_qp(log_scale, &thr, qp);
while (n_coeffs > 0) {
read_coeff(coeff_ptr, &coeff);
quantize_64x64(&thr, qp, &coeff, iscan_ptr, qcoeff_ptr, dqcoeff_ptr, &eob);
coeff_ptr += step;
qcoeff_ptr += step;
dqcoeff_ptr += step;
iscan_ptr += step;
n_coeffs -= step;
}
*eob_ptr = quant_gather_eob(eob);
}
|
