1 /*
2  * Copyright (c) 2016, 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 <emmintrin.h>
13 #include <xmmintrin.h>
14 
15 #include "./aom_dsp_rtcd.h"
16 #include "aom/aom_integer.h"
17 
load_coefficients(const tran_low_t * coeff_ptr)18 static INLINE __m128i load_coefficients(const tran_low_t *coeff_ptr) {
19   if (sizeof(tran_low_t) == 4) {
20     return _mm_setr_epi16((int16_t)coeff_ptr[0], (int16_t)coeff_ptr[1],
21                           (int16_t)coeff_ptr[2], (int16_t)coeff_ptr[3],
22                           (int16_t)coeff_ptr[4], (int16_t)coeff_ptr[5],
23                           (int16_t)coeff_ptr[6], (int16_t)coeff_ptr[7]);
24   } else {
25     return _mm_load_si128((const __m128i *)coeff_ptr);
26   }
27 }
28 
store_coefficients(__m128i coeff_vals,tran_low_t * coeff_ptr)29 static INLINE void store_coefficients(__m128i coeff_vals,
30                                       tran_low_t *coeff_ptr) {
31   if (sizeof(tran_low_t) == 4) {
32     __m128i one = _mm_set1_epi16(1);
33     __m128i coeff_vals_hi = _mm_mulhi_epi16(coeff_vals, one);
34     __m128i coeff_vals_lo = _mm_mullo_epi16(coeff_vals, one);
35     __m128i coeff_vals_1 = _mm_unpacklo_epi16(coeff_vals_lo, coeff_vals_hi);
36     __m128i coeff_vals_2 = _mm_unpackhi_epi16(coeff_vals_lo, coeff_vals_hi);
37     _mm_store_si128((__m128i *)(coeff_ptr), coeff_vals_1);
38     _mm_store_si128((__m128i *)(coeff_ptr + 4), coeff_vals_2);
39   } else {
40     _mm_store_si128((__m128i *)(coeff_ptr), coeff_vals);
41   }
42 }
43 
aom_quantize_b_sse2(const tran_low_t * coeff_ptr,intptr_t n_coeffs,int skip_block,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)44 void aom_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
45                          int skip_block, const int16_t *zbin_ptr,
46                          const int16_t *round_ptr, const int16_t *quant_ptr,
47                          const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
48                          tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
49                          uint16_t *eob_ptr, const int16_t *scan_ptr,
50                          const int16_t *iscan_ptr) {
51   __m128i zero;
52   (void)scan_ptr;
53 
54   coeff_ptr += n_coeffs;
55   iscan_ptr += n_coeffs;
56   qcoeff_ptr += n_coeffs;
57   dqcoeff_ptr += n_coeffs;
58   n_coeffs = -n_coeffs;
59   zero = _mm_setzero_si128();
60   if (!skip_block) {
61     __m128i eob;
62     __m128i zbin;
63     __m128i round, quant, dequant, shift;
64     {
65       __m128i coeff0, coeff1;
66 
67       // Setup global values
68       {
69         __m128i pw_1;
70         zbin = _mm_load_si128((const __m128i *)zbin_ptr);
71         round = _mm_load_si128((const __m128i *)round_ptr);
72         quant = _mm_load_si128((const __m128i *)quant_ptr);
73         pw_1 = _mm_set1_epi16(1);
74         zbin = _mm_sub_epi16(zbin, pw_1);
75         dequant = _mm_load_si128((const __m128i *)dequant_ptr);
76         shift = _mm_load_si128((const __m128i *)quant_shift_ptr);
77       }
78 
79       {
80         __m128i coeff0_sign, coeff1_sign;
81         __m128i qcoeff0, qcoeff1;
82         __m128i qtmp0, qtmp1;
83         __m128i cmp_mask0, cmp_mask1;
84         // Do DC and first 15 AC
85         coeff0 = load_coefficients(coeff_ptr + n_coeffs);
86         coeff1 = load_coefficients(coeff_ptr + n_coeffs + 8);
87 
88         // Poor man's sign extract
89         coeff0_sign = _mm_srai_epi16(coeff0, 15);
90         coeff1_sign = _mm_srai_epi16(coeff1, 15);
91         qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
92         qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
93         qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
94         qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
95 
96         cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
97         zbin = _mm_unpackhi_epi64(zbin, zbin);  // Switch DC to AC
98         cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
99         qcoeff0 = _mm_adds_epi16(qcoeff0, round);
100         round = _mm_unpackhi_epi64(round, round);
101         qcoeff1 = _mm_adds_epi16(qcoeff1, round);
102         qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
103         quant = _mm_unpackhi_epi64(quant, quant);
104         qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
105         qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
106         qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
107         qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
108         shift = _mm_unpackhi_epi64(shift, shift);
109         qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
110 
111         // Reinsert signs
112         qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
113         qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
114         qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
115         qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
116 
117         // Mask out zbin threshold coeffs
118         qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
119         qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
120 
121         store_coefficients(qcoeff0, qcoeff_ptr + n_coeffs);
122         store_coefficients(qcoeff1, qcoeff_ptr + n_coeffs + 8);
123 
124         coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
125         dequant = _mm_unpackhi_epi64(dequant, dequant);
126         coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
127 
128         store_coefficients(coeff0, dqcoeff_ptr + n_coeffs);
129         store_coefficients(coeff1, dqcoeff_ptr + n_coeffs + 8);
130       }
131 
132       {
133         // Scan for eob
134         __m128i zero_coeff0, zero_coeff1;
135         __m128i nzero_coeff0, nzero_coeff1;
136         __m128i iscan0, iscan1;
137         __m128i eob1;
138         zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
139         zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
140         nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
141         nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
142         iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
143         iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
144         // Add one to convert from indices to counts
145         iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
146         iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
147         eob = _mm_and_si128(iscan0, nzero_coeff0);
148         eob1 = _mm_and_si128(iscan1, nzero_coeff1);
149         eob = _mm_max_epi16(eob, eob1);
150       }
151       n_coeffs += 8 * 2;
152     }
153 
154     // AC only loop
155     while (n_coeffs < 0) {
156       __m128i coeff0, coeff1;
157       {
158         __m128i coeff0_sign, coeff1_sign;
159         __m128i qcoeff0, qcoeff1;
160         __m128i qtmp0, qtmp1;
161         __m128i cmp_mask0, cmp_mask1;
162 
163         coeff0 = load_coefficients(coeff_ptr + n_coeffs);
164         coeff1 = load_coefficients(coeff_ptr + n_coeffs + 8);
165 
166         // Poor man's sign extract
167         coeff0_sign = _mm_srai_epi16(coeff0, 15);
168         coeff1_sign = _mm_srai_epi16(coeff1, 15);
169         qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
170         qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
171         qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
172         qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
173 
174         cmp_mask0 = _mm_cmpgt_epi16(qcoeff0, zbin);
175         cmp_mask1 = _mm_cmpgt_epi16(qcoeff1, zbin);
176         qcoeff0 = _mm_adds_epi16(qcoeff0, round);
177         qcoeff1 = _mm_adds_epi16(qcoeff1, round);
178         qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
179         qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);
180         qtmp0 = _mm_add_epi16(qtmp0, qcoeff0);
181         qtmp1 = _mm_add_epi16(qtmp1, qcoeff1);
182         qcoeff0 = _mm_mulhi_epi16(qtmp0, shift);
183         qcoeff1 = _mm_mulhi_epi16(qtmp1, shift);
184 
185         // Reinsert signs
186         qcoeff0 = _mm_xor_si128(qcoeff0, coeff0_sign);
187         qcoeff1 = _mm_xor_si128(qcoeff1, coeff1_sign);
188         qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
189         qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);
190 
191         // Mask out zbin threshold coeffs
192         qcoeff0 = _mm_and_si128(qcoeff0, cmp_mask0);
193         qcoeff1 = _mm_and_si128(qcoeff1, cmp_mask1);
194 
195         store_coefficients(qcoeff0, qcoeff_ptr + n_coeffs);
196         store_coefficients(qcoeff1, qcoeff_ptr + n_coeffs + 8);
197 
198         coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
199         coeff1 = _mm_mullo_epi16(qcoeff1, dequant);
200 
201         store_coefficients(coeff0, dqcoeff_ptr + n_coeffs);
202         store_coefficients(coeff1, dqcoeff_ptr + n_coeffs + 8);
203       }
204 
205       {
206         // Scan for eob
207         __m128i zero_coeff0, zero_coeff1;
208         __m128i nzero_coeff0, nzero_coeff1;
209         __m128i iscan0, iscan1;
210         __m128i eob0, eob1;
211         zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
212         zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
213         nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
214         nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
215         iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
216         iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
217         // Add one to convert from indices to counts
218         iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
219         iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
220         eob0 = _mm_and_si128(iscan0, nzero_coeff0);
221         eob1 = _mm_and_si128(iscan1, nzero_coeff1);
222         eob0 = _mm_max_epi16(eob0, eob1);
223         eob = _mm_max_epi16(eob, eob0);
224       }
225       n_coeffs += 8 * 2;
226     }
227 
228     // Accumulate EOB
229     {
230       __m128i eob_shuffled;
231       eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
232       eob = _mm_max_epi16(eob, eob_shuffled);
233       eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
234       eob = _mm_max_epi16(eob, eob_shuffled);
235       eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
236       eob = _mm_max_epi16(eob, eob_shuffled);
237       *eob_ptr = _mm_extract_epi16(eob, 1);
238     }
239   } else {
240     do {
241       store_coefficients(zero, dqcoeff_ptr + n_coeffs);
242       store_coefficients(zero, dqcoeff_ptr + n_coeffs + 8);
243       store_coefficients(zero, qcoeff_ptr + n_coeffs);
244       store_coefficients(zero, qcoeff_ptr + n_coeffs + 8);
245       n_coeffs += 8 * 2;
246     } while (n_coeffs < 0);
247     *eob_ptr = 0;
248   }
249 }
250