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 <immintrin.h>
13
14 #include "config/aom_dsp_rtcd.h"
15
16 #include "aom/aom_integer.h"
17
init_one_qp(const __m128i * p,__m256i * qp)18 static inline void init_one_qp(const __m128i *p, __m256i *qp) {
19 const __m128i sign = _mm_srai_epi16(*p, 15);
20 const __m128i dc = _mm_unpacklo_epi16(*p, sign);
21 const __m128i ac = _mm_unpackhi_epi16(*p, sign);
22 *qp = _mm256_insertf128_si256(_mm256_castsi128_si256(dc), ac, 1);
23 }
24
update_qp(__m256i * qp)25 static inline void update_qp(__m256i *qp) {
26 int i;
27 for (i = 0; i < 5; ++i) {
28 qp[i] = _mm256_permute2x128_si256(qp[i], qp[i], 0x11);
29 }
30 }
31
init_qp(const int16_t * zbin_ptr,const int16_t * round_ptr,const int16_t * quant_ptr,const int16_t * dequant_ptr,const int16_t * quant_shift_ptr,__m256i * qp,int log_scale)32 static inline void init_qp(const int16_t *zbin_ptr, const int16_t *round_ptr,
33 const int16_t *quant_ptr, const int16_t *dequant_ptr,
34 const int16_t *quant_shift_ptr, __m256i *qp,
35 int log_scale) {
36 const __m128i zbin = _mm_loadu_si128((const __m128i *)zbin_ptr);
37 const __m128i round = _mm_loadu_si128((const __m128i *)round_ptr);
38 const __m128i quant = _mm_loadu_si128((const __m128i *)quant_ptr);
39 const __m128i dequant = _mm_loadu_si128((const __m128i *)dequant_ptr);
40 const __m128i quant_shift = _mm_loadu_si128((const __m128i *)quant_shift_ptr);
41 init_one_qp(&zbin, &qp[0]);
42 init_one_qp(&round, &qp[1]);
43 init_one_qp(&quant, &qp[2]);
44 init_one_qp(&dequant, &qp[3]);
45 init_one_qp(&quant_shift, &qp[4]);
46 if (log_scale > 0) {
47 const __m256i rnd = _mm256_set1_epi32((int16_t)(1 << (log_scale - 1)));
48 qp[0] = _mm256_add_epi32(qp[0], rnd);
49 qp[0] = _mm256_srai_epi32(qp[0], log_scale);
50
51 qp[1] = _mm256_add_epi32(qp[1], rnd);
52 qp[1] = _mm256_srai_epi32(qp[1], log_scale);
53 }
54 // Subtracting 1 here eliminates a _mm256_cmpeq_epi32() instruction when
55 // calculating the zbin mask.
56 qp[0] = _mm256_sub_epi32(qp[0], _mm256_set1_epi32(1));
57 }
58
59 // Note:
60 // *x is vector multiplied by *y which is 16 int32_t parallel multiplication
61 // and right shift 16. The output, 16 int32_t is save in *p.
mm256_mul_shift_epi32(const __m256i * x,const __m256i * y)62 static inline __m256i mm256_mul_shift_epi32(const __m256i *x,
63 const __m256i *y) {
64 __m256i prod_lo = _mm256_mul_epi32(*x, *y);
65 __m256i prod_hi = _mm256_srli_epi64(*x, 32);
66 const __m256i mult_hi = _mm256_srli_epi64(*y, 32);
67 prod_hi = _mm256_mul_epi32(prod_hi, mult_hi);
68
69 prod_lo = _mm256_srli_epi64(prod_lo, 16);
70 const __m256i mask = _mm256_set_epi32(0, -1, 0, -1, 0, -1, 0, -1);
71 prod_lo = _mm256_and_si256(prod_lo, mask);
72 prod_hi = _mm256_srli_epi64(prod_hi, 16);
73
74 prod_hi = _mm256_slli_epi64(prod_hi, 32);
75 return _mm256_or_si256(prod_lo, prod_hi);
76 }
77
get_max_lane_eob(const int16_t * iscan_ptr,__m256i eobmax,__m256i nz_mask)78 static AOM_FORCE_INLINE __m256i get_max_lane_eob(const int16_t *iscan_ptr,
79 __m256i eobmax,
80 __m256i nz_mask) {
81 const __m256i packed_nz_mask = _mm256_packs_epi32(nz_mask, nz_mask);
82 const __m256i packed_nz_mask_perm =
83 _mm256_permute4x64_epi64(packed_nz_mask, 0xD8);
84 const __m256i iscan =
85 _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)iscan_ptr));
86 const __m256i iscan_plus1 = _mm256_sub_epi16(iscan, packed_nz_mask_perm);
87 const __m256i nz_iscan = _mm256_and_si256(iscan_plus1, packed_nz_mask_perm);
88 return _mm256_max_epi16(eobmax, nz_iscan);
89 }
90
91 // Get the max eob from the lower 128 bits.
get_max_eob(__m256i eob)92 static AOM_FORCE_INLINE uint16_t get_max_eob(__m256i eob) {
93 __m256i eob_s;
94 eob_s = _mm256_shuffle_epi32(eob, 0xe);
95 eob = _mm256_max_epi16(eob, eob_s);
96 eob_s = _mm256_shufflelo_epi16(eob, 0xe);
97 eob = _mm256_max_epi16(eob, eob_s);
98 eob_s = _mm256_shufflelo_epi16(eob, 1);
99 eob = _mm256_max_epi16(eob, eob_s);
100 return (uint16_t)_mm256_extract_epi16(eob, 0);
101 }
102
mm256_mul_shift_epi32_logscale(const __m256i * x,const __m256i * y,int log_scale)103 static AOM_FORCE_INLINE __m256i mm256_mul_shift_epi32_logscale(const __m256i *x,
104 const __m256i *y,
105 int log_scale) {
106 __m256i prod_lo = _mm256_mul_epi32(*x, *y);
107 __m256i prod_hi = _mm256_srli_epi64(*x, 32);
108 const __m256i mult_hi = _mm256_srli_epi64(*y, 32);
109 prod_hi = _mm256_mul_epi32(prod_hi, mult_hi);
110 prod_lo = _mm256_srli_epi64(prod_lo, 16 - log_scale);
111 const __m256i mask = _mm256_set_epi32(0, -1, 0, -1, 0, -1, 0, -1);
112 prod_lo = _mm256_and_si256(prod_lo, mask);
113 prod_hi = _mm256_srli_epi64(prod_hi, 16 - log_scale);
114 prod_hi = _mm256_slli_epi64(prod_hi, 32);
115 return _mm256_or_si256(prod_lo, prod_hi);
116 }
117
quantize_logscale(const __m256i * qp,const tran_low_t * coeff_ptr,const int16_t * iscan_ptr,tran_low_t * qcoeff,tran_low_t * dqcoeff,__m256i * eob,int log_scale)118 static AOM_FORCE_INLINE void quantize_logscale(
119 const __m256i *qp, const tran_low_t *coeff_ptr, const int16_t *iscan_ptr,
120 tran_low_t *qcoeff, tran_low_t *dqcoeff, __m256i *eob, int log_scale) {
121 const __m256i coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr);
122 const __m256i abs_coeff = _mm256_abs_epi32(coeff);
123 const __m256i zbin_mask = _mm256_cmpgt_epi32(abs_coeff, qp[0]);
124
125 if (UNLIKELY(_mm256_movemask_epi8(zbin_mask) == 0)) {
126 const __m256i zero = _mm256_setzero_si256();
127 _mm256_storeu_si256((__m256i *)qcoeff, zero);
128 _mm256_storeu_si256((__m256i *)dqcoeff, zero);
129 return;
130 }
131
132 const __m256i tmp_rnd =
133 _mm256_and_si256(_mm256_add_epi32(abs_coeff, qp[1]), zbin_mask);
134 // const int64_t tmp2 = ((tmpw * quant_ptr[rc != 0]) >> 16) + tmpw;
135 const __m256i tmp = mm256_mul_shift_epi32_logscale(&tmp_rnd, &qp[2], 0);
136 const __m256i tmp2 = _mm256_add_epi32(tmp, tmp_rnd);
137 // const int abs_qcoeff = (int)((tmp2 * quant_shift_ptr[rc != 0]) >>
138 // (16 - log_scale + AOM_QM_BITS));
139 const __m256i abs_q =
140 mm256_mul_shift_epi32_logscale(&tmp2, &qp[4], log_scale);
141 const __m256i abs_dq =
142 _mm256_srli_epi32(_mm256_mullo_epi32(abs_q, qp[3]), log_scale);
143 const __m256i nz_mask = _mm256_cmpgt_epi32(abs_q, _mm256_setzero_si256());
144 const __m256i q = _mm256_sign_epi32(abs_q, coeff);
145 const __m256i dq = _mm256_sign_epi32(abs_dq, coeff);
146
147 _mm256_storeu_si256((__m256i *)qcoeff, q);
148 _mm256_storeu_si256((__m256i *)dqcoeff, dq);
149
150 *eob = get_max_lane_eob(iscan_ptr, *eob, nz_mask);
151 }
152
quantize(const __m256i * qp,const tran_low_t * coeff_ptr,const int16_t * iscan_ptr,tran_low_t * qcoeff,tran_low_t * dqcoeff,__m256i * eob)153 static AOM_FORCE_INLINE void quantize(const __m256i *qp,
154 const tran_low_t *coeff_ptr,
155 const int16_t *iscan_ptr,
156 tran_low_t *qcoeff, tran_low_t *dqcoeff,
157 __m256i *eob) {
158 const __m256i coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr);
159 const __m256i abs_coeff = _mm256_abs_epi32(coeff);
160 const __m256i zbin_mask = _mm256_cmpgt_epi32(abs_coeff, qp[0]);
161
162 if (UNLIKELY(_mm256_movemask_epi8(zbin_mask) == 0)) {
163 const __m256i zero = _mm256_setzero_si256();
164 _mm256_storeu_si256((__m256i *)qcoeff, zero);
165 _mm256_storeu_si256((__m256i *)dqcoeff, zero);
166 return;
167 }
168
169 const __m256i tmp_rnd =
170 _mm256_and_si256(_mm256_add_epi32(abs_coeff, qp[1]), zbin_mask);
171 const __m256i tmp = mm256_mul_shift_epi32(&tmp_rnd, &qp[2]);
172 const __m256i tmp2 = _mm256_add_epi32(tmp, tmp_rnd);
173 const __m256i abs_q = mm256_mul_shift_epi32(&tmp2, &qp[4]);
174 const __m256i abs_dq = _mm256_mullo_epi32(abs_q, qp[3]);
175 const __m256i nz_mask = _mm256_cmpgt_epi32(abs_q, _mm256_setzero_si256());
176 const __m256i q = _mm256_sign_epi32(abs_q, coeff);
177 const __m256i dq = _mm256_sign_epi32(abs_dq, coeff);
178
179 _mm256_storeu_si256((__m256i *)qcoeff, q);
180 _mm256_storeu_si256((__m256i *)dqcoeff, dq);
181
182 *eob = get_max_lane_eob(iscan_ptr, *eob, nz_mask);
183 }
184
aom_highbd_quantize_b_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,const int16_t * iscan)185 void aom_highbd_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
186 const int16_t *zbin_ptr,
187 const int16_t *round_ptr,
188 const int16_t *quant_ptr,
189 const int16_t *quant_shift_ptr,
190 tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
191 const int16_t *dequant_ptr, uint16_t *eob_ptr,
192 const int16_t *scan, const int16_t *iscan) {
193 (void)scan;
194 const int step = 8;
195
196 __m256i eob = _mm256_setzero_si256();
197 __m256i qp[5];
198
199 init_qp(zbin_ptr, round_ptr, quant_ptr, dequant_ptr, quant_shift_ptr, qp, 0);
200
201 quantize(qp, coeff_ptr, iscan, qcoeff_ptr, dqcoeff_ptr, &eob);
202
203 coeff_ptr += step;
204 qcoeff_ptr += step;
205 dqcoeff_ptr += step;
206 iscan += step;
207 n_coeffs -= step;
208
209 update_qp(qp);
210
211 while (n_coeffs > 0) {
212 quantize(qp, coeff_ptr, iscan, qcoeff_ptr, dqcoeff_ptr, &eob);
213
214 coeff_ptr += step;
215 qcoeff_ptr += step;
216 dqcoeff_ptr += step;
217 iscan += step;
218 n_coeffs -= step;
219 }
220
221 *eob_ptr = get_max_eob(eob);
222 }
223
aom_highbd_quantize_b_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,const int16_t * iscan)224 void aom_highbd_quantize_b_32x32_avx2(
225 const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
226 const int16_t *round_ptr, const int16_t *quant_ptr,
227 const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
228 tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
229 const int16_t *scan, const int16_t *iscan) {
230 (void)scan;
231 const unsigned int step = 8;
232
233 __m256i eob = _mm256_setzero_si256();
234 __m256i qp[5];
235 init_qp(zbin_ptr, round_ptr, quant_ptr, dequant_ptr, quant_shift_ptr, qp, 1);
236
237 quantize_logscale(qp, coeff_ptr, iscan, qcoeff_ptr, dqcoeff_ptr, &eob, 1);
238
239 coeff_ptr += step;
240 qcoeff_ptr += step;
241 dqcoeff_ptr += step;
242 iscan += step;
243 n_coeffs -= step;
244
245 update_qp(qp);
246
247 while (n_coeffs > 0) {
248 quantize_logscale(qp, coeff_ptr, iscan, qcoeff_ptr, dqcoeff_ptr, &eob, 1);
249
250 coeff_ptr += step;
251 qcoeff_ptr += step;
252 dqcoeff_ptr += step;
253 iscan += step;
254 n_coeffs -= step;
255 }
256
257 *eob_ptr = get_max_eob(eob);
258 }
259
aom_highbd_quantize_b_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,const int16_t * iscan)260 void aom_highbd_quantize_b_64x64_avx2(
261 const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
262 const int16_t *round_ptr, const int16_t *quant_ptr,
263 const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
264 tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
265 const int16_t *scan, const int16_t *iscan) {
266 (void)scan;
267 const int step = 8;
268
269 __m256i eob = _mm256_setzero_si256();
270 __m256i qp[5];
271 init_qp(zbin_ptr, round_ptr, quant_ptr, dequant_ptr, quant_shift_ptr, qp, 2);
272
273 quantize_logscale(qp, coeff_ptr, iscan, qcoeff_ptr, dqcoeff_ptr, &eob, 2);
274
275 coeff_ptr += step;
276 qcoeff_ptr += step;
277 dqcoeff_ptr += step;
278 iscan += step;
279 n_coeffs -= step;
280
281 update_qp(qp);
282
283 while (n_coeffs > 0) {
284 quantize_logscale(qp, coeff_ptr, iscan, qcoeff_ptr, dqcoeff_ptr, &eob, 2);
285
286 coeff_ptr += step;
287 qcoeff_ptr += step;
288 dqcoeff_ptr += step;
289 iscan += step;
290 n_coeffs -= step;
291 }
292
293 *eob_ptr = get_max_eob(eob);
294 }
295