xref: /aosp_15_r20/external/libaom/aom_dsp/x86/highbd_quantize_intrin_sse2.c (revision 77c1e3ccc04c968bd2bc212e87364f250e820521)
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 
14 #include "aom_dsp/aom_dsp_common.h"
15 #include "aom_mem/aom_mem.h"
16 #include "aom_ports/mem.h"
17 #include "config/aom_dsp_rtcd.h"
18 
aom_highbd_quantize_b_sse2(const tran_low_t * coeff_ptr,intptr_t count,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)19 void aom_highbd_quantize_b_sse2(const tran_low_t *coeff_ptr, intptr_t count,
20                                 const int16_t *zbin_ptr,
21                                 const int16_t *round_ptr,
22                                 const int16_t *quant_ptr,
23                                 const int16_t *quant_shift_ptr,
24                                 tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
25                                 const int16_t *dequant_ptr, uint16_t *eob_ptr,
26                                 const int16_t *scan, const int16_t *iscan) {
27   int i, j, non_zero_regs = (int)count / 4, eob_i = -1;
28   __m128i zbins[2];
29   __m128i nzbins[2];
30 
31   zbins[0] = _mm_set_epi32((int)zbin_ptr[1], (int)zbin_ptr[1], (int)zbin_ptr[1],
32                            (int)zbin_ptr[0]);
33   zbins[1] = _mm_set1_epi32((int)zbin_ptr[1]);
34 
35   nzbins[0] = _mm_setzero_si128();
36   nzbins[1] = _mm_setzero_si128();
37   nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
38   nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
39 
40   (void)scan;
41 
42   memset(qcoeff_ptr, 0, count * sizeof(*qcoeff_ptr));
43   memset(dqcoeff_ptr, 0, count * sizeof(*dqcoeff_ptr));
44 
45   // Pre-scan pass
46   for (i = ((int)count / 4) - 1; i >= 0; i--) {
47     __m128i coeffs, cmp1, cmp2;
48     int test;
49     coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
50     cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
51     cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
52     cmp1 = _mm_and_si128(cmp1, cmp2);
53     test = _mm_movemask_epi8(cmp1);
54     if (test == 0xffff)
55       non_zero_regs--;
56     else
57       break;
58   }
59 
60   // Quantization pass:
61   for (i = 0; i < non_zero_regs; i++) {
62     __m128i coeffs, coeffs_sign, tmp1, tmp2;
63     int test;
64     int abs_coeff[4];
65     int coeff_sign[4];
66 
67     coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
68     coeffs_sign = _mm_srai_epi32(coeffs, 31);
69     coeffs = _mm_sub_epi32(_mm_xor_si128(coeffs, coeffs_sign), coeffs_sign);
70     tmp1 = _mm_cmpgt_epi32(coeffs, zbins[i != 0]);
71     tmp2 = _mm_cmpeq_epi32(coeffs, zbins[i != 0]);
72     tmp1 = _mm_or_si128(tmp1, tmp2);
73     test = _mm_movemask_epi8(tmp1);
74     _mm_storeu_si128((__m128i *)abs_coeff, coeffs);
75     _mm_storeu_si128((__m128i *)coeff_sign, coeffs_sign);
76 
77     for (j = 0; j < 4; j++) {
78       if (test & (1 << (4 * j))) {
79         int k = 4 * i + j;
80         const int64_t tmp3 = abs_coeff[j] + round_ptr[k != 0];
81         const int64_t tmp4 = ((tmp3 * quant_ptr[k != 0]) >> 16) + tmp3;
82         const uint32_t abs_qcoeff =
83             (uint32_t)((tmp4 * quant_shift_ptr[k != 0]) >> 16);
84         qcoeff_ptr[k] =
85             (int)(abs_qcoeff ^ (uint32_t)coeff_sign[j]) - coeff_sign[j];
86         dqcoeff_ptr[k] = qcoeff_ptr[k] * dequant_ptr[k != 0];
87         if (abs_qcoeff) eob_i = iscan[k] > eob_i ? iscan[k] : eob_i;
88       }
89     }
90   }
91   *eob_ptr = eob_i + 1;
92 }
93 
aom_highbd_quantize_b_32x32_sse2(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)94 void aom_highbd_quantize_b_32x32_sse2(
95     const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
96     const int16_t *round_ptr, const int16_t *quant_ptr,
97     const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
98     tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
99     const int16_t *scan, const int16_t *iscan) {
100   __m128i zbins[2];
101   __m128i nzbins[2];
102   int idx = 0;
103   int idx_arr[1024];
104   int i, eob = -1;
105   const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 1);
106   const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 1);
107   (void)scan;
108   zbins[0] = _mm_set_epi32(zbin1_tmp, zbin1_tmp, zbin1_tmp, zbin0_tmp);
109   zbins[1] = _mm_set1_epi32(zbin1_tmp);
110 
111   nzbins[0] = _mm_setzero_si128();
112   nzbins[1] = _mm_setzero_si128();
113   nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
114   nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
115 
116   memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
117   memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
118 
119   // Pre-scan pass
120   for (i = 0; i < n_coeffs / 4; i++) {
121     __m128i coeffs, cmp1, cmp2;
122     int test;
123     coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
124     cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
125     cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
126     cmp1 = _mm_and_si128(cmp1, cmp2);
127     test = _mm_movemask_epi8(cmp1);
128     if (!(test & 0xf)) idx_arr[idx++] = i * 4;
129     if (!(test & 0xf0)) idx_arr[idx++] = i * 4 + 1;
130     if (!(test & 0xf00)) idx_arr[idx++] = i * 4 + 2;
131     if (!(test & 0xf000)) idx_arr[idx++] = i * 4 + 3;
132   }
133 
134   // Quantization pass: only process the coefficients selected in
135   // pre-scan pass. Note: idx can be zero.
136   for (i = 0; i < idx; i++) {
137     const int rc = idx_arr[i];
138     const int coeff = coeff_ptr[rc];
139     const int coeff_sign = AOMSIGN(coeff);
140     const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
141     const int64_t tmp1 = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
142     const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
143     const uint32_t abs_qcoeff =
144         (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
145     qcoeff_ptr[rc] = (int)(abs_qcoeff ^ (uint32_t)coeff_sign) - coeff_sign;
146     dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
147     if (abs_qcoeff) eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob;
148   }
149   *eob_ptr = eob + 1;
150 }
151 
aom_highbd_quantize_b_64x64_sse2(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)152 void aom_highbd_quantize_b_64x64_sse2(
153     const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
154     const int16_t *round_ptr, const int16_t *quant_ptr,
155     const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
156     tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
157     const int16_t *scan, const int16_t *iscan) {
158   __m128i zbins[2];
159   __m128i nzbins[2];
160   int idx = 0;
161   int idx_arr[1024];
162   int i, eob = -1;
163   const int zbin0_tmp = ROUND_POWER_OF_TWO(zbin_ptr[0], 2);
164   const int zbin1_tmp = ROUND_POWER_OF_TWO(zbin_ptr[1], 2);
165   (void)scan;
166   zbins[0] = _mm_set_epi32(zbin1_tmp, zbin1_tmp, zbin1_tmp, zbin0_tmp);
167   zbins[1] = _mm_set1_epi32(zbin1_tmp);
168 
169   nzbins[0] = _mm_setzero_si128();
170   nzbins[1] = _mm_setzero_si128();
171   nzbins[0] = _mm_sub_epi32(nzbins[0], zbins[0]);
172   nzbins[1] = _mm_sub_epi32(nzbins[1], zbins[1]);
173 
174   memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
175   memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
176 
177   // Pre-scan pass
178   for (i = 0; i < n_coeffs / 4; i++) {
179     __m128i coeffs, cmp1, cmp2;
180     int test;
181     coeffs = _mm_load_si128((const __m128i *)(coeff_ptr + i * 4));
182     cmp1 = _mm_cmplt_epi32(coeffs, zbins[i != 0]);
183     cmp2 = _mm_cmpgt_epi32(coeffs, nzbins[i != 0]);
184     cmp1 = _mm_and_si128(cmp1, cmp2);
185     test = _mm_movemask_epi8(cmp1);
186     if (!(test & 0xf)) idx_arr[idx++] = i * 4;
187     if (!(test & 0xf0)) idx_arr[idx++] = i * 4 + 1;
188     if (!(test & 0xf00)) idx_arr[idx++] = i * 4 + 2;
189     if (!(test & 0xf000)) idx_arr[idx++] = i * 4 + 3;
190   }
191 
192   // Quantization pass: only process the coefficients selected in
193   // pre-scan pass. Note: idx can be zero.
194   for (i = 0; i < idx; i++) {
195     const int rc = idx_arr[i];
196     const int coeff = coeff_ptr[rc];
197     const int coeff_sign = AOMSIGN(coeff);
198     const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
199     const int64_t tmp1 = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 2);
200     const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
201     const uint32_t abs_qcoeff =
202         (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 14);
203     qcoeff_ptr[rc] = (int)(abs_qcoeff ^ (uint32_t)coeff_sign) - coeff_sign;
204     dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 4;
205     if (abs_qcoeff) eob = iscan[idx_arr[i]] > eob ? iscan[idx_arr[i]] : eob;
206   }
207   *eob_ptr = eob + 1;
208 }
209