xref: /aosp_15_r20/external/libaom/aom_dsp/x86/quantize_avx2.c (revision 77c1e3ccc04c968bd2bc212e87364f250e820521)
1 /*
2  * Copyright (c) 2022, 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 #include "config/aom_dsp_rtcd.h"
14 #include "aom/aom_integer.h"
15 #include "aom_dsp/x86/quantize_x86.h"
16 
load_b_values_avx2(const int16_t * zbin_ptr,__m256i * zbin,const int16_t * round_ptr,__m256i * round,const int16_t * quant_ptr,__m256i * quant,const int16_t * dequant_ptr,__m256i * dequant,const int16_t * shift_ptr,__m256i * shift,int log_scale)17 static inline void load_b_values_avx2(const int16_t *zbin_ptr, __m256i *zbin,
18                                       const int16_t *round_ptr, __m256i *round,
19                                       const int16_t *quant_ptr, __m256i *quant,
20                                       const int16_t *dequant_ptr,
21                                       __m256i *dequant,
22                                       const int16_t *shift_ptr, __m256i *shift,
23                                       int log_scale) {
24   *zbin = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)zbin_ptr));
25   *zbin = _mm256_permute4x64_epi64(*zbin, 0x54);
26   if (log_scale > 0) {
27     const __m256i rnd = _mm256_set1_epi16((int16_t)(1 << (log_scale - 1)));
28     *zbin = _mm256_add_epi16(*zbin, rnd);
29     *zbin = _mm256_srai_epi16(*zbin, log_scale);
30   }
31   // Subtracting 1 here eliminates a _mm256_cmpeq_epi16() instruction when
32   // calculating the zbin mask. (See quantize_b_logscale{0,1,2}_16)
33   *zbin = _mm256_sub_epi16(*zbin, _mm256_set1_epi16(1));
34 
35   *round = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)round_ptr));
36   *round = _mm256_permute4x64_epi64(*round, 0x54);
37   if (log_scale > 0) {
38     const __m256i rnd = _mm256_set1_epi16((int16_t)(1 << (log_scale - 1)));
39     *round = _mm256_add_epi16(*round, rnd);
40     *round = _mm256_srai_epi16(*round, log_scale);
41   }
42 
43   *quant = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)quant_ptr));
44   *quant = _mm256_permute4x64_epi64(*quant, 0x54);
45   *dequant =
46       _mm256_castsi128_si256(_mm_load_si128((const __m128i *)dequant_ptr));
47   *dequant = _mm256_permute4x64_epi64(*dequant, 0x54);
48   *shift = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)shift_ptr));
49   *shift = _mm256_permute4x64_epi64(*shift, 0x54);
50 }
51 
load_coefficients_avx2(const tran_low_t * coeff_ptr)52 static inline __m256i load_coefficients_avx2(const tran_low_t *coeff_ptr) {
53   const __m256i coeff1 = _mm256_load_si256((__m256i *)coeff_ptr);
54   const __m256i coeff2 = _mm256_load_si256((__m256i *)(coeff_ptr + 8));
55   return _mm256_packs_epi32(coeff1, coeff2);
56 }
57 
store_coefficients_avx2(__m256i coeff_vals,tran_low_t * coeff_ptr)58 static inline void store_coefficients_avx2(__m256i coeff_vals,
59                                            tran_low_t *coeff_ptr) {
60   __m256i coeff_sign = _mm256_srai_epi16(coeff_vals, 15);
61   __m256i coeff_vals_lo = _mm256_unpacklo_epi16(coeff_vals, coeff_sign);
62   __m256i coeff_vals_hi = _mm256_unpackhi_epi16(coeff_vals, coeff_sign);
63   _mm256_store_si256((__m256i *)coeff_ptr, coeff_vals_lo);
64   _mm256_store_si256((__m256i *)(coeff_ptr + 8), coeff_vals_hi);
65 }
66 
quantize_b_logscale0_16(const tran_low_t * coeff_ptr,tran_low_t * qcoeff_ptr,tran_low_t * dqcoeff_ptr,__m256i * v_quant,__m256i * v_dequant,__m256i * v_round,__m256i * v_zbin,__m256i * v_quant_shift)67 static AOM_FORCE_INLINE __m256i quantize_b_logscale0_16(
68     const tran_low_t *coeff_ptr, tran_low_t *qcoeff_ptr,
69     tran_low_t *dqcoeff_ptr, __m256i *v_quant, __m256i *v_dequant,
70     __m256i *v_round, __m256i *v_zbin, __m256i *v_quant_shift) {
71   const __m256i v_coeff = load_coefficients_avx2(coeff_ptr);
72   const __m256i v_abs_coeff = _mm256_abs_epi16(v_coeff);
73   const __m256i v_zbin_mask = _mm256_cmpgt_epi16(v_abs_coeff, *v_zbin);
74 
75   if (_mm256_movemask_epi8(v_zbin_mask) == 0) {
76     _mm256_store_si256((__m256i *)qcoeff_ptr, _mm256_setzero_si256());
77     _mm256_store_si256((__m256i *)dqcoeff_ptr, _mm256_setzero_si256());
78     _mm256_store_si256((__m256i *)(qcoeff_ptr + 8), _mm256_setzero_si256());
79     _mm256_store_si256((__m256i *)(dqcoeff_ptr + 8), _mm256_setzero_si256());
80     return _mm256_setzero_si256();
81   }
82 
83   // tmp = v_zbin_mask ? (int64_t)abs_coeff + log_scaled_round : 0
84   const __m256i v_tmp_rnd =
85       _mm256_and_si256(_mm256_adds_epi16(v_abs_coeff, *v_round), v_zbin_mask);
86   //  tmp32 = (int)(((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
87   //                 quant_shift_ptr[rc != 0]) >>
88   //                (16 - log_scale + AOM_QM_BITS));
89   const __m256i v_tmp32_a = _mm256_mulhi_epi16(v_tmp_rnd, *v_quant);
90   const __m256i v_tmp32_b = _mm256_add_epi16(v_tmp32_a, v_tmp_rnd);
91   const __m256i v_tmp32 = _mm256_mulhi_epi16(v_tmp32_b, *v_quant_shift);
92   const __m256i v_nz_mask = _mm256_cmpgt_epi16(v_tmp32, _mm256_setzero_si256());
93   const __m256i v_qcoeff = _mm256_sign_epi16(v_tmp32, v_coeff);
94   const __m256i v_dqcoeff = _mm256_mullo_epi16(v_qcoeff, *v_dequant);
95   store_coefficients_avx2(v_qcoeff, qcoeff_ptr);
96   store_coefficients_avx2(v_dqcoeff, dqcoeff_ptr);
97   return v_nz_mask;
98 }
99 
get_max_lane_eob(const int16_t * iscan,__m256i v_eobmax,__m256i v_mask)100 static inline __m256i get_max_lane_eob(const int16_t *iscan, __m256i v_eobmax,
101                                        __m256i v_mask) {
102   const __m256i v_iscan = _mm256_loadu_si256((const __m256i *)iscan);
103   const __m256i v_iscan_perm = _mm256_permute4x64_epi64(v_iscan, 0xD8);
104   const __m256i v_iscan_plus1 = _mm256_sub_epi16(v_iscan_perm, v_mask);
105   const __m256i v_nz_iscan = _mm256_and_si256(v_iscan_plus1, v_mask);
106   return _mm256_max_epi16(v_eobmax, v_nz_iscan);
107 }
108 
accumulate_eob256(__m256i eob256)109 static inline int16_t accumulate_eob256(__m256i eob256) {
110   const __m128i eob_lo = _mm256_castsi256_si128(eob256);
111   const __m128i eob_hi = _mm256_extractf128_si256(eob256, 1);
112   __m128i eob = _mm_max_epi16(eob_lo, eob_hi);
113   __m128i eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
114   eob = _mm_max_epi16(eob, eob_shuffled);
115   eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
116   eob = _mm_max_epi16(eob, eob_shuffled);
117   eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
118   eob = _mm_max_epi16(eob, eob_shuffled);
119   return _mm_extract_epi16(eob, 1);
120 }
121 
aom_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)122 void aom_quantize_b_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
123                          const int16_t *zbin_ptr, const int16_t *round_ptr,
124                          const int16_t *quant_ptr,
125                          const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
126                          tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
127                          uint16_t *eob_ptr, const int16_t *scan,
128                          const int16_t *iscan) {
129   (void)scan;
130   __m256i v_zbin, v_round, v_quant, v_dequant, v_quant_shift;
131   __m256i v_eobmax = _mm256_setzero_si256();
132 
133   load_b_values_avx2(zbin_ptr, &v_zbin, round_ptr, &v_round, quant_ptr,
134                      &v_quant, dequant_ptr, &v_dequant, quant_shift_ptr,
135                      &v_quant_shift, 0);
136 
137   // Do DC and first 15 AC.
138   __m256i v_nz_mask =
139       quantize_b_logscale0_16(coeff_ptr, qcoeff_ptr, dqcoeff_ptr, &v_quant,
140                               &v_dequant, &v_round, &v_zbin, &v_quant_shift);
141 
142   v_eobmax = get_max_lane_eob(iscan, v_eobmax, v_nz_mask);
143 
144   v_round = _mm256_unpackhi_epi64(v_round, v_round);
145   v_quant = _mm256_unpackhi_epi64(v_quant, v_quant);
146   v_dequant = _mm256_unpackhi_epi64(v_dequant, v_dequant);
147   v_quant_shift = _mm256_unpackhi_epi64(v_quant_shift, v_quant_shift);
148   v_zbin = _mm256_unpackhi_epi64(v_zbin, v_zbin);
149 
150   for (intptr_t count = n_coeffs - 16; count > 0; count -= 16) {
151     coeff_ptr += 16;
152     qcoeff_ptr += 16;
153     dqcoeff_ptr += 16;
154     iscan += 16;
155     v_nz_mask =
156         quantize_b_logscale0_16(coeff_ptr, qcoeff_ptr, dqcoeff_ptr, &v_quant,
157                                 &v_dequant, &v_round, &v_zbin, &v_quant_shift);
158 
159     v_eobmax = get_max_lane_eob(iscan, v_eobmax, v_nz_mask);
160   }
161 
162   *eob_ptr = accumulate_eob256(v_eobmax);
163 }
164 
quantize_b_logscale_16(const tran_low_t * coeff_ptr,tran_low_t * qcoeff_ptr,tran_low_t * dqcoeff_ptr,__m256i * v_quant,__m256i * v_dequant,__m256i * v_round,__m256i * v_zbin,__m256i * v_quant_shift,int log_scale)165 static AOM_FORCE_INLINE __m256i quantize_b_logscale_16(
166     const tran_low_t *coeff_ptr, tran_low_t *qcoeff_ptr,
167     tran_low_t *dqcoeff_ptr, __m256i *v_quant, __m256i *v_dequant,
168     __m256i *v_round, __m256i *v_zbin, __m256i *v_quant_shift, int log_scale) {
169   const __m256i v_coeff = load_coefficients_avx2(coeff_ptr);
170   const __m256i v_abs_coeff = _mm256_abs_epi16(v_coeff);
171   const __m256i v_zbin_mask = _mm256_cmpgt_epi16(v_abs_coeff, *v_zbin);
172 
173   if (_mm256_movemask_epi8(v_zbin_mask) == 0) {
174     _mm256_store_si256((__m256i *)qcoeff_ptr, _mm256_setzero_si256());
175     _mm256_store_si256((__m256i *)dqcoeff_ptr, _mm256_setzero_si256());
176     _mm256_store_si256((__m256i *)(qcoeff_ptr + 8), _mm256_setzero_si256());
177     _mm256_store_si256((__m256i *)(dqcoeff_ptr + 8), _mm256_setzero_si256());
178     return _mm256_setzero_si256();
179   }
180 
181   // tmp = v_zbin_mask ? (int64_t)abs_coeff + log_scaled_round : 0
182   const __m256i v_tmp_rnd =
183       _mm256_and_si256(_mm256_adds_epi16(v_abs_coeff, *v_round), v_zbin_mask);
184   //  tmp32 = (int)(((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
185   //                 quant_shift_ptr[rc != 0]) >>
186   //                (16 - log_scale + AOM_QM_BITS));
187   const __m256i v_tmp32_a = _mm256_mulhi_epi16(v_tmp_rnd, *v_quant);
188   const __m256i v_tmp32_b = _mm256_add_epi16(v_tmp32_a, v_tmp_rnd);
189   const __m256i v_tmp32_hi = _mm256_slli_epi16(
190       _mm256_mulhi_epi16(v_tmp32_b, *v_quant_shift), log_scale);
191   const __m256i v_tmp32_lo = _mm256_srli_epi16(
192       _mm256_mullo_epi16(v_tmp32_b, *v_quant_shift), 16 - log_scale);
193   const __m256i v_tmp32 = _mm256_or_si256(v_tmp32_hi, v_tmp32_lo);
194   const __m256i v_dqcoeff_hi = _mm256_slli_epi16(
195       _mm256_mulhi_epi16(v_tmp32, *v_dequant), 16 - log_scale);
196   const __m256i v_dqcoeff_lo =
197       _mm256_srli_epi16(_mm256_mullo_epi16(v_tmp32, *v_dequant), log_scale);
198   const __m256i v_dqcoeff =
199       _mm256_sign_epi16(_mm256_or_si256(v_dqcoeff_hi, v_dqcoeff_lo), v_coeff);
200   const __m256i v_qcoeff = _mm256_sign_epi16(v_tmp32, v_coeff);
201   const __m256i v_nz_mask = _mm256_cmpgt_epi16(v_tmp32, _mm256_setzero_si256());
202   store_coefficients_avx2(v_qcoeff, qcoeff_ptr);
203   store_coefficients_avx2(v_dqcoeff, dqcoeff_ptr);
204   return v_nz_mask;
205 }
206 
quantize_b_no_qmatrix_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 * iscan,int log_scale)207 static AOM_FORCE_INLINE void quantize_b_no_qmatrix_avx2(
208     const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
209     const int16_t *round_ptr, const int16_t *quant_ptr,
210     const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
211     tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
212     const int16_t *iscan, int log_scale) {
213   __m256i v_zbin, v_round, v_quant, v_dequant, v_quant_shift;
214   __m256i v_eobmax = _mm256_setzero_si256();
215 
216   load_b_values_avx2(zbin_ptr, &v_zbin, round_ptr, &v_round, quant_ptr,
217                      &v_quant, dequant_ptr, &v_dequant, quant_shift_ptr,
218                      &v_quant_shift, log_scale);
219 
220   // Do DC and first 15 AC.
221   __m256i v_nz_mask = quantize_b_logscale_16(
222       coeff_ptr, qcoeff_ptr, dqcoeff_ptr, &v_quant, &v_dequant, &v_round,
223       &v_zbin, &v_quant_shift, log_scale);
224 
225   v_eobmax = get_max_lane_eob(iscan, v_eobmax, v_nz_mask);
226 
227   v_round = _mm256_unpackhi_epi64(v_round, v_round);
228   v_quant = _mm256_unpackhi_epi64(v_quant, v_quant);
229   v_dequant = _mm256_unpackhi_epi64(v_dequant, v_dequant);
230   v_quant_shift = _mm256_unpackhi_epi64(v_quant_shift, v_quant_shift);
231   v_zbin = _mm256_unpackhi_epi64(v_zbin, v_zbin);
232 
233   for (intptr_t count = n_coeffs - 16; count > 0; count -= 16) {
234     coeff_ptr += 16;
235     qcoeff_ptr += 16;
236     dqcoeff_ptr += 16;
237     iscan += 16;
238     v_nz_mask = quantize_b_logscale_16(coeff_ptr, qcoeff_ptr, dqcoeff_ptr,
239                                        &v_quant, &v_dequant, &v_round, &v_zbin,
240                                        &v_quant_shift, log_scale);
241 
242     v_eobmax = get_max_lane_eob(iscan, v_eobmax, v_nz_mask);
243   }
244 
245   *eob_ptr = accumulate_eob256(v_eobmax);
246 }
247 
aom_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)248 void aom_quantize_b_32x32_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
249                                const int16_t *zbin_ptr,
250                                const int16_t *round_ptr,
251                                const int16_t *quant_ptr,
252                                const int16_t *quant_shift_ptr,
253                                tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
254                                const int16_t *dequant_ptr, uint16_t *eob_ptr,
255                                const int16_t *scan, const int16_t *iscan) {
256   (void)scan;
257   quantize_b_no_qmatrix_avx2(coeff_ptr, n_coeffs, zbin_ptr, round_ptr,
258                              quant_ptr, quant_shift_ptr, qcoeff_ptr,
259                              dqcoeff_ptr, dequant_ptr, eob_ptr, iscan, 1);
260 }
261 
aom_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)262 void aom_quantize_b_64x64_avx2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
263                                const int16_t *zbin_ptr,
264                                const int16_t *round_ptr,
265                                const int16_t *quant_ptr,
266                                const int16_t *quant_shift_ptr,
267                                tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
268                                const int16_t *dequant_ptr, uint16_t *eob_ptr,
269                                const int16_t *scan, const int16_t *iscan) {
270   (void)scan;
271   quantize_b_no_qmatrix_avx2(coeff_ptr, n_coeffs, zbin_ptr, round_ptr,
272                              quant_ptr, quant_shift_ptr, qcoeff_ptr,
273                              dqcoeff_ptr, dequant_ptr, eob_ptr, iscan, 2);
274 }
275