1 /*
2 * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include <assert.h>
12
13 #include "./vpx_dsp_rtcd.h"
14 #include "vpx_dsp/quantize.h"
15 #include "vpx_dsp/vpx_dsp_common.h"
16 #include "vpx_mem/vpx_mem.h"
17
vpx_quantize_dc(const tran_low_t * coeff_ptr,int n_coeffs,int skip_block,const int16_t * round_ptr,const int16_t quant,tran_low_t * qcoeff_ptr,tran_low_t * dqcoeff_ptr,const int16_t dequant,uint16_t * eob_ptr)18 void vpx_quantize_dc(const tran_low_t *coeff_ptr, int n_coeffs, int skip_block,
19 const int16_t *round_ptr, const int16_t quant,
20 tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
21 const int16_t dequant, uint16_t *eob_ptr) {
22 const int rc = 0;
23 const int coeff = coeff_ptr[rc];
24 const int coeff_sign = (coeff >> 31);
25 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
26 int tmp, eob = -1;
27
28 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
29 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
30
31 if (!skip_block) {
32 tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
33 tmp = (tmp * quant) >> 16;
34 qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
35 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant;
36 if (tmp) eob = 0;
37 }
38 *eob_ptr = eob + 1;
39 }
40
41 #if CONFIG_VP9_HIGHBITDEPTH
vpx_highbd_quantize_dc(const tran_low_t * coeff_ptr,int n_coeffs,int skip_block,const int16_t * round_ptr,const int16_t quant,tran_low_t * qcoeff_ptr,tran_low_t * dqcoeff_ptr,const int16_t dequant,uint16_t * eob_ptr)42 void vpx_highbd_quantize_dc(const tran_low_t *coeff_ptr, int n_coeffs,
43 int skip_block, const int16_t *round_ptr,
44 const int16_t quant, tran_low_t *qcoeff_ptr,
45 tran_low_t *dqcoeff_ptr, const int16_t dequant,
46 uint16_t *eob_ptr) {
47 int eob = -1;
48
49 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
50 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
51
52 if (!skip_block) {
53 const int coeff = coeff_ptr[0];
54 const int coeff_sign = (coeff >> 31);
55 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
56 const int64_t tmp = abs_coeff + round_ptr[0];
57 const int abs_qcoeff = (int)((tmp * quant) >> 16);
58 qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
59 dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant;
60 if (abs_qcoeff) eob = 0;
61 }
62 *eob_ptr = eob + 1;
63 }
64 #endif
65
vpx_quantize_dc_32x32(const tran_low_t * coeff_ptr,int skip_block,const int16_t * round_ptr,const int16_t quant,tran_low_t * qcoeff_ptr,tran_low_t * dqcoeff_ptr,const int16_t dequant,uint16_t * eob_ptr)66 void vpx_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
67 const int16_t *round_ptr, const int16_t quant,
68 tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
69 const int16_t dequant, uint16_t *eob_ptr) {
70 const int n_coeffs = 1024;
71 const int rc = 0;
72 const int coeff = coeff_ptr[rc];
73 const int coeff_sign = (coeff >> 31);
74 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
75 int tmp, eob = -1;
76
77 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
78 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
79
80 if (!skip_block) {
81 tmp = clamp(abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1),
82 INT16_MIN, INT16_MAX);
83 tmp = (tmp * quant) >> 15;
84 qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
85 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant / 2;
86 if (tmp) eob = 0;
87 }
88 *eob_ptr = eob + 1;
89 }
90
91 #if CONFIG_VP9_HIGHBITDEPTH
vpx_highbd_quantize_dc_32x32(const tran_low_t * coeff_ptr,int skip_block,const int16_t * round_ptr,const int16_t quant,tran_low_t * qcoeff_ptr,tran_low_t * dqcoeff_ptr,const int16_t dequant,uint16_t * eob_ptr)92 void vpx_highbd_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
93 const int16_t *round_ptr, const int16_t quant,
94 tran_low_t *qcoeff_ptr,
95 tran_low_t *dqcoeff_ptr,
96 const int16_t dequant, uint16_t *eob_ptr) {
97 const int n_coeffs = 1024;
98 int eob = -1;
99
100 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
101 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
102
103 if (!skip_block) {
104 const int coeff = coeff_ptr[0];
105 const int coeff_sign = (coeff >> 31);
106 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
107 const int64_t tmp = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[0], 1);
108 const int abs_qcoeff = (int)((tmp * quant) >> 15);
109 qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
110 dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant / 2;
111 if (abs_qcoeff) eob = 0;
112 }
113 *eob_ptr = eob + 1;
114 }
115 #endif
116
vpx_quantize_b_c(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,const int16_t * iscan)117 void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
118 int skip_block, const int16_t *zbin_ptr,
119 const int16_t *round_ptr, const int16_t *quant_ptr,
120 const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
121 tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
122 uint16_t *eob_ptr, const int16_t *scan,
123 const int16_t *iscan) {
124 int i, non_zero_count = (int)n_coeffs, eob = -1;
125 const int zbins[2] = { zbin_ptr[0], zbin_ptr[1] };
126 const int nzbins[2] = { zbins[0] * -1, zbins[1] * -1 };
127 (void)iscan;
128 (void)skip_block;
129 assert(!skip_block);
130
131 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
132 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
133
134 // Pre-scan pass
135 for (i = (int)n_coeffs - 1; i >= 0; i--) {
136 const int rc = scan[i];
137 const int coeff = coeff_ptr[rc];
138
139 if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0])
140 non_zero_count--;
141 else
142 break;
143 }
144
145 // Quantization pass: All coefficients with index >= zero_flag are
146 // skippable. Note: zero_flag can be zero.
147 for (i = 0; i < non_zero_count; i++) {
148 const int rc = scan[i];
149 const int coeff = coeff_ptr[rc];
150 const int coeff_sign = (coeff >> 31);
151 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
152
153 if (abs_coeff >= zbins[rc != 0]) {
154 int tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
155 tmp = ((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
156 quant_shift_ptr[rc != 0]) >>
157 16; // quantization
158 qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
159 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
160
161 if (tmp) eob = i;
162 }
163 }
164 *eob_ptr = eob + 1;
165 }
166
167 #if CONFIG_VP9_HIGHBITDEPTH
vpx_highbd_quantize_b_c(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,const int16_t * iscan)168 void vpx_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
169 int skip_block, const int16_t *zbin_ptr,
170 const int16_t *round_ptr, const int16_t *quant_ptr,
171 const int16_t *quant_shift_ptr,
172 tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
173 const int16_t *dequant_ptr, uint16_t *eob_ptr,
174 const int16_t *scan, const int16_t *iscan) {
175 int i, non_zero_count = (int)n_coeffs, eob = -1;
176 const int zbins[2] = { zbin_ptr[0], zbin_ptr[1] };
177 const int nzbins[2] = { zbins[0] * -1, zbins[1] * -1 };
178 (void)iscan;
179 (void)skip_block;
180 assert(!skip_block);
181
182 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
183 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
184
185 // Pre-scan pass
186 for (i = (int)n_coeffs - 1; i >= 0; i--) {
187 const int rc = scan[i];
188 const int coeff = coeff_ptr[rc];
189
190 if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0])
191 non_zero_count--;
192 else
193 break;
194 }
195
196 // Quantization pass: All coefficients with index >= zero_flag are
197 // skippable. Note: zero_flag can be zero.
198 for (i = 0; i < non_zero_count; i++) {
199 const int rc = scan[i];
200 const int coeff = coeff_ptr[rc];
201 const int coeff_sign = (coeff >> 31);
202 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
203
204 if (abs_coeff >= zbins[rc != 0]) {
205 const int64_t tmp1 = abs_coeff + round_ptr[rc != 0];
206 const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
207 const int abs_qcoeff = (int)((tmp2 * quant_shift_ptr[rc != 0]) >> 16);
208 qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
209 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
210 if (abs_qcoeff) eob = i;
211 }
212 }
213 *eob_ptr = eob + 1;
214 }
215 #endif
216
vpx_quantize_b_32x32_c(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,const int16_t * iscan)217 void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
218 int skip_block, const int16_t *zbin_ptr,
219 const int16_t *round_ptr, const int16_t *quant_ptr,
220 const int16_t *quant_shift_ptr,
221 tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
222 const int16_t *dequant_ptr, uint16_t *eob_ptr,
223 const int16_t *scan, const int16_t *iscan) {
224 const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], 1),
225 ROUND_POWER_OF_TWO(zbin_ptr[1], 1) };
226 const int nzbins[2] = { zbins[0] * -1, zbins[1] * -1 };
227
228 int idx = 0;
229 int idx_arr[1024];
230 int i, eob = -1;
231 (void)iscan;
232 (void)skip_block;
233 assert(!skip_block);
234
235 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
236 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
237
238 // Pre-scan pass
239 for (i = 0; i < n_coeffs; i++) {
240 const int rc = scan[i];
241 const int coeff = coeff_ptr[rc];
242
243 // If the coefficient is out of the base ZBIN range, keep it for
244 // quantization.
245 if (coeff >= zbins[rc != 0] || coeff <= nzbins[rc != 0]) idx_arr[idx++] = i;
246 }
247
248 // Quantization pass: only process the coefficients selected in
249 // pre-scan pass. Note: idx can be zero.
250 for (i = 0; i < idx; i++) {
251 const int rc = scan[idx_arr[i]];
252 const int coeff = coeff_ptr[rc];
253 const int coeff_sign = (coeff >> 31);
254 int tmp;
255 int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
256 abs_coeff += ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
257 abs_coeff = clamp(abs_coeff, INT16_MIN, INT16_MAX);
258 tmp = ((((abs_coeff * quant_ptr[rc != 0]) >> 16) + abs_coeff) *
259 quant_shift_ptr[rc != 0]) >>
260 15;
261
262 qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
263 #if (ARCH_X86 || ARCH_X86_64) && !CONFIG_VP9_HIGHBITDEPTH
264 // When tran_low_t is only 16 bits dqcoeff can outrange it. Rather than
265 // truncating with a cast, saturate the value. This is easier to implement
266 // on x86 and preserves the sign of the value.
267 dqcoeff_ptr[rc] =
268 clamp(qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2, INT16_MIN, INT16_MAX);
269 #else
270 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
271 #endif // ARCH_X86 && CONFIG_VP9_HIGHBITDEPTH
272
273 if (tmp) eob = idx_arr[i];
274 }
275 *eob_ptr = eob + 1;
276 }
277
278 #if CONFIG_VP9_HIGHBITDEPTH
vpx_highbd_quantize_b_32x32_c(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,const int16_t * iscan)279 void vpx_highbd_quantize_b_32x32_c(
280 const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block,
281 const int16_t *zbin_ptr, const int16_t *round_ptr, const int16_t *quant_ptr,
282 const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
283 tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
284 const int16_t *scan, const int16_t *iscan) {
285 const int zbins[2] = { ROUND_POWER_OF_TWO(zbin_ptr[0], 1),
286 ROUND_POWER_OF_TWO(zbin_ptr[1], 1) };
287 const int nzbins[2] = { zbins[0] * -1, zbins[1] * -1 };
288
289 int idx = 0;
290 int idx_arr[1024];
291 int i, eob = -1;
292 (void)iscan;
293 (void)skip_block;
294 assert(!skip_block);
295
296 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
297 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
298
299 // Pre-scan pass
300 for (i = 0; i < n_coeffs; i++) {
301 const int rc = scan[i];
302 const int coeff = coeff_ptr[rc];
303
304 // If the coefficient is out of the base ZBIN range, keep it for
305 // quantization.
306 if (coeff >= zbins[rc != 0] || coeff <= nzbins[rc != 0]) idx_arr[idx++] = i;
307 }
308
309 // Quantization pass: only process the coefficients selected in
310 // pre-scan pass. Note: idx can be zero.
311 for (i = 0; i < idx; i++) {
312 const int rc = scan[idx_arr[i]];
313 const int coeff = coeff_ptr[rc];
314 const int coeff_sign = (coeff >> 31);
315 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
316 const int64_t tmp1 = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
317 const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
318 const int abs_qcoeff = (int)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
319 qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
320 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
321 if (abs_qcoeff) eob = idx_arr[i];
322 }
323 *eob_ptr = eob + 1;
324 }
325 #endif
326