VPX: removed step checks from neon convolve code
[aom.git] / vpx_dsp / quantize.c
blobe4e741a9089a896f078b73c17a3ecb49c5fa73e6
1 /*
2 * Copyright (c) 2015 The WebM project authors. All Rights Reserved.
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 */
11 #include "vpx_dsp/quantize.h"
12 #include "vpx_mem/vpx_mem.h"
14 void vpx_quantize_dc(const tran_low_t *coeff_ptr,
15 int n_coeffs, int skip_block,
16 const int16_t *round_ptr, const int16_t quant,
17 tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
18 const int16_t dequant_ptr, uint16_t *eob_ptr) {
19 const int rc = 0;
20 const int coeff = coeff_ptr[rc];
21 const int coeff_sign = (coeff >> 31);
22 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
23 int tmp, eob = -1;
25 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
26 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
28 if (!skip_block) {
29 tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
30 tmp = (tmp * quant) >> 16;
31 qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
32 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr;
33 if (tmp)
34 eob = 0;
36 *eob_ptr = eob + 1;
39 #if CONFIG_VP9_HIGHBITDEPTH
40 void vpx_highbd_quantize_dc(const tran_low_t *coeff_ptr,
41 int n_coeffs, int skip_block,
42 const int16_t *round_ptr, const int16_t quant,
43 tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
44 const int16_t dequant_ptr, uint16_t *eob_ptr) {
45 int eob = -1;
47 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
48 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
50 if (!skip_block) {
51 const int coeff = coeff_ptr[0];
52 const int coeff_sign = (coeff >> 31);
53 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
54 const int64_t tmp = abs_coeff + round_ptr[0];
55 const uint32_t abs_qcoeff = (uint32_t)((tmp * quant) >> 16);
56 qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
57 dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant_ptr;
58 if (abs_qcoeff)
59 eob = 0;
61 *eob_ptr = eob + 1;
63 #endif
65 void vpx_quantize_dc_32x32(const tran_low_t *coeff_ptr, int skip_block,
66 const int16_t *round_ptr, const int16_t quant,
67 tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
68 const int16_t dequant_ptr, uint16_t *eob_ptr) {
69 const int n_coeffs = 1024;
70 const int rc = 0;
71 const int coeff = coeff_ptr[rc];
72 const int coeff_sign = (coeff >> 31);
73 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
74 int tmp, eob = -1;
76 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
77 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
79 if (!skip_block) {
80 tmp = clamp(abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1),
81 INT16_MIN, INT16_MAX);
82 tmp = (tmp * quant) >> 15;
83 qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
84 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr / 2;
85 if (tmp)
86 eob = 0;
88 *eob_ptr = eob + 1;
91 #if CONFIG_VP9_HIGHBITDEPTH
92 void vpx_highbd_quantize_dc_32x32(const tran_low_t *coeff_ptr,
93 int skip_block,
94 const int16_t *round_ptr,
95 const int16_t quant,
96 tran_low_t *qcoeff_ptr,
97 tran_low_t *dqcoeff_ptr,
98 const int16_t dequant_ptr,
99 uint16_t *eob_ptr) {
100 const int n_coeffs = 1024;
101 int eob = -1;
103 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
104 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
106 if (!skip_block) {
107 const int coeff = coeff_ptr[0];
108 const int coeff_sign = (coeff >> 31);
109 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
110 const int64_t tmp = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[0], 1);
111 const uint32_t abs_qcoeff = (uint32_t)((tmp * quant) >> 15);
112 qcoeff_ptr[0] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
113 dqcoeff_ptr[0] = qcoeff_ptr[0] * dequant_ptr / 2;
114 if (abs_qcoeff)
115 eob = 0;
117 *eob_ptr = eob + 1;
119 #endif
121 void vpx_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
122 int skip_block,
123 const int16_t *zbin_ptr, const int16_t *round_ptr,
124 const int16_t *quant_ptr, const int16_t *quant_shift_ptr,
125 tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
126 const int16_t *dequant_ptr,
127 uint16_t *eob_ptr,
128 const int16_t *scan, const int16_t *iscan) {
129 int i, non_zero_count = (int)n_coeffs, eob = -1;
130 const int zbins[2] = {zbin_ptr[0], zbin_ptr[1]};
131 const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
132 (void)iscan;
134 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
135 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
137 if (!skip_block) {
138 // Pre-scan pass
139 for (i = (int)n_coeffs - 1; i >= 0; i--) {
140 const int rc = scan[i];
141 const int coeff = coeff_ptr[rc];
143 if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0])
144 non_zero_count--;
145 else
146 break;
149 // Quantization pass: All coefficients with index >= zero_flag are
150 // skippable. Note: zero_flag can be zero.
151 for (i = 0; i < non_zero_count; i++) {
152 const int rc = scan[i];
153 const int coeff = coeff_ptr[rc];
154 const int coeff_sign = (coeff >> 31);
155 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
157 if (abs_coeff >= zbins[rc != 0]) {
158 int tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
159 tmp = ((((tmp * quant_ptr[rc != 0]) >> 16) + tmp) *
160 quant_shift_ptr[rc != 0]) >> 16; // quantization
161 qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
162 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
164 if (tmp)
165 eob = i;
169 *eob_ptr = eob + 1;
172 #if CONFIG_VP9_HIGHBITDEPTH
173 void vpx_highbd_quantize_b_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
174 int skip_block, const int16_t *zbin_ptr,
175 const int16_t *round_ptr, const int16_t *quant_ptr,
176 const int16_t *quant_shift_ptr,
177 tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
178 const int16_t *dequant_ptr,
179 uint16_t *eob_ptr, const int16_t *scan,
180 const int16_t *iscan) {
181 int i, non_zero_count = (int)n_coeffs, eob = -1;
182 const int zbins[2] = {zbin_ptr[0], zbin_ptr[1]};
183 const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
184 (void)iscan;
186 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
187 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
189 if (!skip_block) {
190 // Pre-scan pass
191 for (i = (int)n_coeffs - 1; i >= 0; i--) {
192 const int rc = scan[i];
193 const int coeff = coeff_ptr[rc];
195 if (coeff < zbins[rc != 0] && coeff > nzbins[rc != 0])
196 non_zero_count--;
197 else
198 break;
201 // Quantization pass: All coefficients with index >= zero_flag are
202 // skippable. Note: zero_flag can be zero.
203 for (i = 0; i < non_zero_count; i++) {
204 const int rc = scan[i];
205 const int coeff = coeff_ptr[rc];
206 const int coeff_sign = (coeff >> 31);
207 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
209 if (abs_coeff >= zbins[rc != 0]) {
210 const int64_t tmp1 = abs_coeff + round_ptr[rc != 0];
211 const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
212 const uint32_t abs_qcoeff =
213 (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 16);
214 qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
215 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
216 if (abs_qcoeff)
217 eob = i;
221 *eob_ptr = eob + 1;
223 #endif
225 void vpx_quantize_b_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
226 int skip_block,
227 const int16_t *zbin_ptr, const int16_t *round_ptr,
228 const int16_t *quant_ptr,
229 const int16_t *quant_shift_ptr,
230 tran_low_t *qcoeff_ptr, tran_low_t *dqcoeff_ptr,
231 const int16_t *dequant_ptr,
232 uint16_t *eob_ptr,
233 const int16_t *scan, const int16_t *iscan) {
234 const int zbins[2] = {ROUND_POWER_OF_TWO(zbin_ptr[0], 1),
235 ROUND_POWER_OF_TWO(zbin_ptr[1], 1)};
236 const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
238 int idx = 0;
239 int idx_arr[1024];
240 int i, eob = -1;
241 (void)iscan;
243 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
244 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
246 if (!skip_block) {
247 // Pre-scan pass
248 for (i = 0; i < n_coeffs; i++) {
249 const int rc = scan[i];
250 const int coeff = coeff_ptr[rc];
252 // If the coefficient is out of the base ZBIN range, keep it for
253 // quantization.
254 if (coeff >= zbins[rc != 0] || coeff <= nzbins[rc != 0])
255 idx_arr[idx++] = i;
258 // Quantization pass: only process the coefficients selected in
259 // pre-scan pass. Note: idx can be zero.
260 for (i = 0; i < idx; i++) {
261 const int rc = scan[idx_arr[i]];
262 const int coeff = coeff_ptr[rc];
263 const int coeff_sign = (coeff >> 31);
264 int tmp;
265 int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
266 abs_coeff += ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
267 abs_coeff = clamp(abs_coeff, INT16_MIN, INT16_MAX);
268 tmp = ((((abs_coeff * quant_ptr[rc != 0]) >> 16) + abs_coeff) *
269 quant_shift_ptr[rc != 0]) >> 15;
271 qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
272 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
274 if (tmp)
275 eob = idx_arr[i];
278 *eob_ptr = eob + 1;
281 #if CONFIG_VP9_HIGHBITDEPTH
282 void vpx_highbd_quantize_b_32x32_c(const tran_low_t *coeff_ptr,
283 intptr_t n_coeffs, int skip_block,
284 const int16_t *zbin_ptr,
285 const int16_t *round_ptr,
286 const int16_t *quant_ptr,
287 const int16_t *quant_shift_ptr,
288 tran_low_t *qcoeff_ptr,
289 tran_low_t *dqcoeff_ptr,
290 const int16_t *dequant_ptr,
291 uint16_t *eob_ptr,
292 const int16_t *scan, const int16_t *iscan) {
293 const int zbins[2] = {ROUND_POWER_OF_TWO(zbin_ptr[0], 1),
294 ROUND_POWER_OF_TWO(zbin_ptr[1], 1)};
295 const int nzbins[2] = {zbins[0] * -1, zbins[1] * -1};
297 int idx = 0;
298 int idx_arr[1024];
299 int i, eob = -1;
300 (void)iscan;
302 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
303 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
305 if (!skip_block) {
306 // Pre-scan pass
307 for (i = 0; i < n_coeffs; i++) {
308 const int rc = scan[i];
309 const int coeff = coeff_ptr[rc];
311 // If the coefficient is out of the base ZBIN range, keep it for
312 // quantization.
313 if (coeff >= zbins[rc != 0] || coeff <= nzbins[rc != 0])
314 idx_arr[idx++] = i;
317 // Quantization pass: only process the coefficients selected in
318 // pre-scan pass. Note: idx can be zero.
319 for (i = 0; i < idx; i++) {
320 const int rc = scan[idx_arr[i]];
321 const int coeff = coeff_ptr[rc];
322 const int coeff_sign = (coeff >> 31);
323 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
324 const int64_t tmp1 = abs_coeff
325 + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
326 const int64_t tmp2 = ((tmp1 * quant_ptr[rc != 0]) >> 16) + tmp1;
327 const uint32_t abs_qcoeff =
328 (uint32_t)((tmp2 * quant_shift_ptr[rc != 0]) >> 15);
329 qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
330 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
331 if (abs_qcoeff)
332 eob = idx_arr[i];
335 *eob_ptr = eob + 1;
337 #endif