xref: /aosp_15_r20/external/libaom/aom_dsp/x86/adaptive_quantize_avx2.c (revision 77c1e3ccc04c968bd2bc212e87364f250e820521)
1 /*
2  * Copyright (c) 2019, 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/quantize.h"
16 #include "aom_dsp/x86/quantize_x86.h"
17 
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)18 static inline void load_b_values_avx2(const int16_t *zbin_ptr, __m256i *zbin,
19                                       const int16_t *round_ptr, __m256i *round,
20                                       const int16_t *quant_ptr, __m256i *quant,
21                                       const int16_t *dequant_ptr,
22                                       __m256i *dequant,
23                                       const int16_t *shift_ptr,
24                                       __m256i *shift) {
25   *zbin = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)zbin_ptr));
26   *zbin = _mm256_permute4x64_epi64(*zbin, 0x54);
27   *zbin = _mm256_sub_epi16(*zbin, _mm256_set1_epi16(1));
28   *round = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)round_ptr));
29   *round = _mm256_permute4x64_epi64(*round, 0x54);
30   *quant = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)quant_ptr));
31   *quant = _mm256_permute4x64_epi64(*quant, 0x54);
32   *dequant =
33       _mm256_castsi128_si256(_mm_load_si128((const __m128i *)dequant_ptr));
34   *dequant = _mm256_permute4x64_epi64(*dequant, 0x54);
35   *shift = _mm256_castsi128_si256(_mm_load_si128((const __m128i *)shift_ptr));
36   *shift = _mm256_permute4x64_epi64(*shift, 0x54);
37 }
38 
load_coefficients_avx2(const tran_low_t * coeff_ptr)39 static inline __m256i load_coefficients_avx2(const tran_low_t *coeff_ptr) {
40   const __m256i coeff1 = _mm256_load_si256((__m256i *)(coeff_ptr));
41   const __m256i coeff2 = _mm256_load_si256((__m256i *)(coeff_ptr + 8));
42   return _mm256_packs_epi32(coeff1, coeff2);
43 }
44 
update_mask1_avx2(__m256i * cmp_mask,const int16_t * iscan_ptr,int * is_found,__m256i * mask)45 static inline void update_mask1_avx2(__m256i *cmp_mask,
46                                      const int16_t *iscan_ptr, int *is_found,
47                                      __m256i *mask) {
48   __m256i temp_mask = _mm256_setzero_si256();
49   if (_mm256_movemask_epi8(*cmp_mask)) {
50     __m256i iscan = _mm256_loadu_si256((const __m256i *)(iscan_ptr));
51     temp_mask = _mm256_and_si256(*cmp_mask, iscan);
52     *is_found = 1;
53   }
54   *mask = _mm256_max_epi16(temp_mask, *mask);
55 }
56 
update_mask0_avx2(__m256i * qcoeff,__m256i * threshold,const int16_t * iscan_ptr,int * is_found,__m256i * mask)57 static inline void update_mask0_avx2(__m256i *qcoeff, __m256i *threshold,
58                                      const int16_t *iscan_ptr, int *is_found,
59                                      __m256i *mask) {
60   __m256i zero = _mm256_setzero_si256();
61   __m256i coeff[2], cmp_mask0, cmp_mask1;
62   coeff[0] = _mm256_unpacklo_epi16(*qcoeff, zero);
63   coeff[1] = _mm256_unpackhi_epi16(*qcoeff, zero);
64   coeff[0] = _mm256_slli_epi32(coeff[0], AOM_QM_BITS);
65   cmp_mask0 = _mm256_cmpgt_epi32(coeff[0], threshold[0]);
66   coeff[1] = _mm256_slli_epi32(coeff[1], AOM_QM_BITS);
67   cmp_mask1 = _mm256_cmpgt_epi32(coeff[1], threshold[1]);
68   cmp_mask0 =
69       _mm256_permute4x64_epi64(_mm256_packs_epi32(cmp_mask0, cmp_mask1), 0xd8);
70   update_mask1_avx2(&cmp_mask0, iscan_ptr, is_found, mask);
71 }
72 
calculate_qcoeff_avx2(__m256i * coeff,const __m256i * round,const __m256i * quant,const __m256i * shift)73 static inline void calculate_qcoeff_avx2(__m256i *coeff, const __m256i *round,
74                                          const __m256i *quant,
75                                          const __m256i *shift) {
76   __m256i tmp, qcoeff;
77   qcoeff = _mm256_adds_epi16(*coeff, *round);
78   tmp = _mm256_mulhi_epi16(qcoeff, *quant);
79   qcoeff = _mm256_add_epi16(tmp, qcoeff);
80   *coeff = _mm256_mulhi_epi16(qcoeff, *shift);
81 }
82 
calculate_dqcoeff_avx2(__m256i qcoeff,__m256i dequant)83 static inline __m256i calculate_dqcoeff_avx2(__m256i qcoeff, __m256i dequant) {
84   return _mm256_mullo_epi16(qcoeff, dequant);
85 }
86 
store_coefficients_avx2(__m256i coeff_vals,tran_low_t * coeff_ptr)87 static inline void store_coefficients_avx2(__m256i coeff_vals,
88                                            tran_low_t *coeff_ptr) {
89   __m256i coeff_sign = _mm256_srai_epi16(coeff_vals, 15);
90   __m256i coeff_vals_lo = _mm256_unpacklo_epi16(coeff_vals, coeff_sign);
91   __m256i coeff_vals_hi = _mm256_unpackhi_epi16(coeff_vals, coeff_sign);
92   _mm256_store_si256((__m256i *)(coeff_ptr), coeff_vals_lo);
93   _mm256_store_si256((__m256i *)(coeff_ptr + 8), coeff_vals_hi);
94 }
95 
aom_quantize_b_adaptive_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)96 void aom_quantize_b_adaptive_avx2(
97     const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
98     const int16_t *round_ptr, const int16_t *quant_ptr,
99     const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
100     tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
101     const int16_t *scan, const int16_t *iscan) {
102   int index = 16;
103   int non_zero_count = 0;
104   int non_zero_count_prescan_add_zero = 0;
105   int is_found0 = 0, is_found1 = 0;
106   int eob = -1;
107   const __m256i zero = _mm256_setzero_si256();
108   __m256i zbin, round, quant, dequant, shift;
109   __m256i coeff, qcoeff;
110   __m256i cmp_mask, mask0 = zero, mask1 = zero;
111   __m128i temp_mask0, temp_mask1;
112   int prescan_add[2];
113   int thresh[2];
114   const qm_val_t wt = (1 << AOM_QM_BITS);
115   for (int i = 0; i < 2; ++i) {
116     prescan_add[i] = ROUND_POWER_OF_TWO(dequant_ptr[i] * EOB_FACTOR, 7);
117     thresh[i] = (zbin_ptr[i] * wt + prescan_add[i]) - 1;
118   }
119   __m256i threshold[2];
120   threshold[0] = _mm256_set1_epi32(thresh[0]);
121   threshold[1] = _mm256_set1_epi32(thresh[1]);
122   threshold[0] = _mm256_blend_epi32(threshold[0], threshold[1], 0xfe);
123 
124 #if SKIP_EOB_FACTOR_ADJUST
125   int first = -1;
126 #endif
127 
128   // Setup global values.
129   load_b_values_avx2(zbin_ptr, &zbin, round_ptr, &round, quant_ptr, &quant,
130                      dequant_ptr, &dequant, quant_shift_ptr, &shift);
131 
132   // Do DC and first 15 AC.
133   coeff = load_coefficients_avx2(coeff_ptr);
134   qcoeff = _mm256_abs_epi16(coeff);
135   update_mask0_avx2(&qcoeff, threshold, iscan, &is_found0, &mask0);
136   __m256i temp0 = _mm256_cmpgt_epi16(qcoeff, zbin);
137   zbin = _mm256_unpackhi_epi64(zbin, zbin);
138   cmp_mask = _mm256_permute4x64_epi64(temp0, 0xd8);
139   update_mask1_avx2(&cmp_mask, iscan, &is_found1, &mask1);
140   threshold[0] = threshold[1];
141   if (_mm256_movemask_epi8(cmp_mask) == 0) {
142     _mm256_store_si256((__m256i *)(qcoeff_ptr), zero);
143     _mm256_store_si256((__m256i *)(qcoeff_ptr + 8), zero);
144     _mm256_store_si256((__m256i *)(dqcoeff_ptr), zero);
145     _mm256_store_si256((__m256i *)(dqcoeff_ptr + 8), zero);
146     round = _mm256_unpackhi_epi64(round, round);
147     quant = _mm256_unpackhi_epi64(quant, quant);
148     shift = _mm256_unpackhi_epi64(shift, shift);
149     dequant = _mm256_unpackhi_epi64(dequant, dequant);
150   } else {
151     calculate_qcoeff_avx2(&qcoeff, &round, &quant, &shift);
152     round = _mm256_unpackhi_epi64(round, round);
153     quant = _mm256_unpackhi_epi64(quant, quant);
154     shift = _mm256_unpackhi_epi64(shift, shift);
155     // Reinsert signs
156     qcoeff = _mm256_sign_epi16(qcoeff, coeff);
157     // Mask out zbin threshold coeffs
158     qcoeff = _mm256_and_si256(qcoeff, temp0);
159     store_coefficients_avx2(qcoeff, qcoeff_ptr);
160     coeff = calculate_dqcoeff_avx2(qcoeff, dequant);
161     dequant = _mm256_unpackhi_epi64(dequant, dequant);
162     store_coefficients_avx2(coeff, dqcoeff_ptr);
163   }
164 
165   // AC only loop.
166   while (index < n_coeffs) {
167     coeff = load_coefficients_avx2(coeff_ptr + index);
168     qcoeff = _mm256_abs_epi16(coeff);
169     update_mask0_avx2(&qcoeff, threshold, iscan + index, &is_found0, &mask0);
170     temp0 = _mm256_cmpgt_epi16(qcoeff, zbin);
171     cmp_mask = _mm256_permute4x64_epi64(temp0, 0xd8);
172     update_mask1_avx2(&cmp_mask, iscan + index, &is_found1, &mask1);
173     if (_mm256_movemask_epi8(cmp_mask) == 0) {
174       _mm256_store_si256((__m256i *)(qcoeff_ptr + index), zero);
175       _mm256_store_si256((__m256i *)(qcoeff_ptr + index + 8), zero);
176       _mm256_store_si256((__m256i *)(dqcoeff_ptr + index), zero);
177       _mm256_store_si256((__m256i *)(dqcoeff_ptr + index + 8), zero);
178       index += 16;
179       continue;
180     }
181     calculate_qcoeff_avx2(&qcoeff, &round, &quant, &shift);
182     qcoeff = _mm256_sign_epi16(qcoeff, coeff);
183     qcoeff = _mm256_and_si256(qcoeff, temp0);
184     store_coefficients_avx2(qcoeff, qcoeff_ptr + index);
185     coeff = calculate_dqcoeff_avx2(qcoeff, dequant);
186     store_coefficients_avx2(coeff, dqcoeff_ptr + index);
187     index += 16;
188   }
189   if (is_found0) {
190     temp_mask0 = _mm_max_epi16(_mm256_castsi256_si128(mask0),
191                                _mm256_extracti128_si256(mask0, 1));
192     non_zero_count = calculate_non_zero_count(temp_mask0);
193   }
194   if (is_found1) {
195     temp_mask1 = _mm_max_epi16(_mm256_castsi256_si128(mask1),
196                                _mm256_extracti128_si256(mask1, 1));
197     non_zero_count_prescan_add_zero = calculate_non_zero_count(temp_mask1);
198   }
199 
200   for (int i = non_zero_count_prescan_add_zero - 1; i >= non_zero_count; i--) {
201     const int rc = scan[i];
202     qcoeff_ptr[rc] = 0;
203     dqcoeff_ptr[rc] = 0;
204   }
205 
206   for (int i = non_zero_count - 1; i >= 0; i--) {
207     const int rc = scan[i];
208     if (qcoeff_ptr[rc]) {
209       eob = i;
210       break;
211     }
212   }
213 
214   *eob_ptr = eob + 1;
215 #if SKIP_EOB_FACTOR_ADJUST
216   // TODO(Aniket): Experiment the following loop with intrinsic by combining
217   // with the quantization loop above
218   for (int i = 0; i < non_zero_count; i++) {
219     const int rc = scan[i];
220     const int qcoeff0 = qcoeff_ptr[rc];
221     if (qcoeff0) {
222       first = i;
223       break;
224     }
225   }
226   if ((*eob_ptr - 1) >= 0 && first == (*eob_ptr - 1)) {
227     const int rc = scan[(*eob_ptr - 1)];
228     if (qcoeff_ptr[rc] == 1 || qcoeff_ptr[rc] == -1) {
229       const int coeff0 = coeff_ptr[rc] * wt;
230       const int coeff_sign = AOMSIGN(coeff0);
231       const int abs_coeff = (coeff0 ^ coeff_sign) - coeff_sign;
232       const int factor = EOB_FACTOR + SKIP_EOB_FACTOR_ADJUST;
233       const int prescan_add_val =
234           ROUND_POWER_OF_TWO(dequant_ptr[rc != 0] * factor, 7);
235       if (abs_coeff <
236           (zbin_ptr[rc != 0] * (1 << AOM_QM_BITS) + prescan_add_val)) {
237         qcoeff_ptr[rc] = 0;
238         dqcoeff_ptr[rc] = 0;
239         *eob_ptr = 0;
240       }
241     }
242   }
243 #endif
244 }
245