1 /*
2 * Copyright (c) 2010 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 #include <math.h>
13 #include "./vpx_dsp_rtcd.h"
14 #include "vpx_mem/vpx_mem.h"
15 #include "vpx_ports/mem.h"
16
17 #include "vp9/common/vp9_quant_common.h"
18 #include "vp9/common/vp9_seg_common.h"
19
20 #include "vp9/encoder/vp9_encoder.h"
21 #include "vp9/encoder/vp9_quantize.h"
22 #include "vp9/encoder/vp9_rd.h"
23
vp9_quantize_fp_c(const tran_low_t * coeff_ptr,intptr_t n_coeffs,int skip_block,const int16_t * round_ptr,const int16_t * quant_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)24 void vp9_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
25 int skip_block, const int16_t *round_ptr,
26 const int16_t *quant_ptr, tran_low_t *qcoeff_ptr,
27 tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr,
28 uint16_t *eob_ptr, const int16_t *scan,
29 const int16_t *iscan) {
30 int i, eob = -1;
31 (void)iscan;
32 (void)skip_block;
33 assert(!skip_block);
34
35 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
36 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
37
38 // Quantization pass: All coefficients with index >= zero_flag are
39 // skippable. Note: zero_flag can be zero.
40 for (i = 0; i < n_coeffs; i++) {
41 const int rc = scan[i];
42 const int coeff = coeff_ptr[rc];
43 const int coeff_sign = (coeff >> 31);
44 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
45
46 int tmp = clamp(abs_coeff + round_ptr[rc != 0], INT16_MIN, INT16_MAX);
47 tmp = (tmp * quant_ptr[rc != 0]) >> 16;
48
49 qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
50 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
51
52 if (tmp) eob = i;
53 }
54 *eob_ptr = eob + 1;
55 }
56
57 #if CONFIG_VP9_HIGHBITDEPTH
vp9_highbd_quantize_fp_c(const tran_low_t * coeff_ptr,intptr_t n_coeffs,int skip_block,const int16_t * round_ptr,const int16_t * quant_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)58 void vp9_highbd_quantize_fp_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
59 int skip_block, const int16_t *round_ptr,
60 const int16_t *quant_ptr, tran_low_t *qcoeff_ptr,
61 tran_low_t *dqcoeff_ptr,
62 const int16_t *dequant_ptr, uint16_t *eob_ptr,
63 const int16_t *scan, const int16_t *iscan) {
64 int i;
65 int eob = -1;
66
67 (void)iscan;
68 (void)skip_block;
69 assert(!skip_block);
70
71 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
72 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
73
74 // Quantization pass: All coefficients with index >= zero_flag are
75 // skippable. Note: zero_flag can be zero.
76 for (i = 0; i < n_coeffs; i++) {
77 const int rc = scan[i];
78 const int coeff = coeff_ptr[rc];
79 const int coeff_sign = (coeff >> 31);
80 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
81 const int64_t tmp = abs_coeff + round_ptr[rc != 0];
82 const int abs_qcoeff = (int)((tmp * quant_ptr[rc != 0]) >> 16);
83 qcoeff_ptr[rc] = (tran_low_t)(abs_qcoeff ^ coeff_sign) - coeff_sign;
84 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0];
85 if (abs_qcoeff) eob = i;
86 }
87 *eob_ptr = eob + 1;
88 }
89 #endif
90
91 // TODO(jingning) Refactor this file and combine functions with similar
92 // operations.
vp9_quantize_fp_32x32_c(const tran_low_t * coeff_ptr,intptr_t n_coeffs,int skip_block,const int16_t * round_ptr,const int16_t * quant_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)93 void vp9_quantize_fp_32x32_c(const tran_low_t *coeff_ptr, intptr_t n_coeffs,
94 int skip_block, const int16_t *round_ptr,
95 const int16_t *quant_ptr, tran_low_t *qcoeff_ptr,
96 tran_low_t *dqcoeff_ptr,
97 const int16_t *dequant_ptr, uint16_t *eob_ptr,
98 const int16_t *scan, const int16_t *iscan) {
99 int i, eob = -1;
100 (void)iscan;
101 (void)skip_block;
102 assert(!skip_block);
103
104 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
105 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
106
107 for (i = 0; i < n_coeffs; i++) {
108 const int rc = scan[i];
109 const int coeff = coeff_ptr[rc];
110 const int coeff_sign = (coeff >> 31);
111 int tmp = 0;
112 int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
113
114 if (abs_coeff >= (dequant_ptr[rc != 0] >> 2)) {
115 abs_coeff += ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
116 abs_coeff = clamp(abs_coeff, INT16_MIN, INT16_MAX);
117 tmp = (abs_coeff * quant_ptr[rc != 0]) >> 15;
118 qcoeff_ptr[rc] = (tmp ^ coeff_sign) - coeff_sign;
119 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
120 }
121
122 if (tmp) eob = i;
123 }
124 *eob_ptr = eob + 1;
125 }
126
127 #if CONFIG_VP9_HIGHBITDEPTH
vp9_highbd_quantize_fp_32x32_c(const tran_low_t * coeff_ptr,intptr_t n_coeffs,int skip_block,const int16_t * round_ptr,const int16_t * quant_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)128 void vp9_highbd_quantize_fp_32x32_c(
129 const tran_low_t *coeff_ptr, intptr_t n_coeffs, int skip_block,
130 const int16_t *round_ptr, const int16_t *quant_ptr, tran_low_t *qcoeff_ptr,
131 tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
132 const int16_t *scan, const int16_t *iscan) {
133 int i, eob = -1;
134
135 (void)iscan;
136 (void)skip_block;
137 assert(!skip_block);
138
139 memset(qcoeff_ptr, 0, n_coeffs * sizeof(*qcoeff_ptr));
140 memset(dqcoeff_ptr, 0, n_coeffs * sizeof(*dqcoeff_ptr));
141
142 for (i = 0; i < n_coeffs; i++) {
143 int abs_qcoeff = 0;
144 const int rc = scan[i];
145 const int coeff = coeff_ptr[rc];
146 const int coeff_sign = (coeff >> 31);
147 const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
148
149 if (abs_coeff >= (dequant_ptr[rc != 0] >> 2)) {
150 const int64_t tmp = abs_coeff + ROUND_POWER_OF_TWO(round_ptr[rc != 0], 1);
151 abs_qcoeff = (int)((tmp * quant_ptr[rc != 0]) >> 15);
152 qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
153 dqcoeff_ptr[rc] = qcoeff_ptr[rc] * dequant_ptr[rc != 0] / 2;
154 }
155
156 if (abs_qcoeff) eob = i;
157 }
158 *eob_ptr = eob + 1;
159 }
160 #endif
161
vp9_regular_quantize_b_4x4(MACROBLOCK * x,int plane,int block,const int16_t * scan,const int16_t * iscan)162 void vp9_regular_quantize_b_4x4(MACROBLOCK *x, int plane, int block,
163 const int16_t *scan, const int16_t *iscan) {
164 MACROBLOCKD *const xd = &x->e_mbd;
165 struct macroblock_plane *p = &x->plane[plane];
166 struct macroblockd_plane *pd = &xd->plane[plane];
167 tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block),
168 *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
169 const int n_coeffs = 4 * 4;
170
171 if (x->skip_block) {
172 memset(qcoeff, 0, n_coeffs * sizeof(*qcoeff));
173 memset(dqcoeff, 0, n_coeffs * sizeof(*dqcoeff));
174 return;
175 }
176
177 #if CONFIG_VP9_HIGHBITDEPTH
178 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
179 vpx_highbd_quantize_b(BLOCK_OFFSET(p->coeff, block), n_coeffs,
180 x->skip_block, p->zbin, p->round, p->quant,
181 p->quant_shift, qcoeff, dqcoeff, pd->dequant,
182 &p->eobs[block], scan, iscan);
183 return;
184 }
185 #endif
186 vpx_quantize_b(BLOCK_OFFSET(p->coeff, block), n_coeffs, x->skip_block,
187 p->zbin, p->round, p->quant, p->quant_shift, qcoeff, dqcoeff,
188 pd->dequant, &p->eobs[block], scan, iscan);
189 }
190
invert_quant(int16_t * quant,int16_t * shift,int d)191 static void invert_quant(int16_t *quant, int16_t *shift, int d) {
192 unsigned t;
193 int l, m;
194 t = d;
195 for (l = 0; t > 1; l++) t >>= 1;
196 m = 1 + (1 << (16 + l)) / d;
197 *quant = (int16_t)(m - (1 << 16));
198 *shift = 1 << (16 - l);
199 }
200
get_qzbin_factor(int q,vpx_bit_depth_t bit_depth)201 static int get_qzbin_factor(int q, vpx_bit_depth_t bit_depth) {
202 const int quant = vp9_dc_quant(q, 0, bit_depth);
203 #if CONFIG_VP9_HIGHBITDEPTH
204 switch (bit_depth) {
205 case VPX_BITS_8: return q == 0 ? 64 : (quant < 148 ? 84 : 80);
206 case VPX_BITS_10: return q == 0 ? 64 : (quant < 592 ? 84 : 80);
207 default:
208 assert(bit_depth == VPX_BITS_12);
209 return q == 0 ? 64 : (quant < 2368 ? 84 : 80);
210 }
211 #else
212 (void)bit_depth;
213 return q == 0 ? 64 : (quant < 148 ? 84 : 80);
214 #endif
215 }
216
vp9_init_quantizer(VP9_COMP * cpi)217 void vp9_init_quantizer(VP9_COMP *cpi) {
218 VP9_COMMON *const cm = &cpi->common;
219 QUANTS *const quants = &cpi->quants;
220 int i, q, quant;
221
222 for (q = 0; q < QINDEX_RANGE; q++) {
223 int qzbin_factor = get_qzbin_factor(q, cm->bit_depth);
224 int qrounding_factor = q == 0 ? 64 : 48;
225 const int sharpness_adjustment = 16 * (7 - cpi->oxcf.sharpness) / 7;
226
227 if (cpi->oxcf.sharpness > 0 && q > 0) {
228 qzbin_factor = 64 + sharpness_adjustment;
229 qrounding_factor = 64 - sharpness_adjustment;
230 }
231
232 for (i = 0; i < 2; ++i) {
233 int qrounding_factor_fp = i == 0 ? 48 : 42;
234 if (q == 0) qrounding_factor_fp = 64;
235 if (cpi->oxcf.sharpness > 0)
236 qrounding_factor_fp = 64 - sharpness_adjustment;
237 // y
238 quant = i == 0 ? vp9_dc_quant(q, cm->y_dc_delta_q, cm->bit_depth)
239 : vp9_ac_quant(q, 0, cm->bit_depth);
240 invert_quant(&quants->y_quant[q][i], &quants->y_quant_shift[q][i], quant);
241 quants->y_quant_fp[q][i] = (1 << 16) / quant;
242 quants->y_round_fp[q][i] = (qrounding_factor_fp * quant) >> 7;
243 quants->y_zbin[q][i] = ROUND_POWER_OF_TWO(qzbin_factor * quant, 7);
244 quants->y_round[q][i] = (qrounding_factor * quant) >> 7;
245 cpi->y_dequant[q][i] = quant;
246
247 // uv
248 quant = i == 0 ? vp9_dc_quant(q, cm->uv_dc_delta_q, cm->bit_depth)
249 : vp9_ac_quant(q, cm->uv_ac_delta_q, cm->bit_depth);
250 invert_quant(&quants->uv_quant[q][i], &quants->uv_quant_shift[q][i],
251 quant);
252 quants->uv_quant_fp[q][i] = (1 << 16) / quant;
253 quants->uv_round_fp[q][i] = (qrounding_factor_fp * quant) >> 7;
254 quants->uv_zbin[q][i] = ROUND_POWER_OF_TWO(qzbin_factor * quant, 7);
255 quants->uv_round[q][i] = (qrounding_factor * quant) >> 7;
256 cpi->uv_dequant[q][i] = quant;
257 }
258
259 for (i = 2; i < 8; i++) {
260 quants->y_quant[q][i] = quants->y_quant[q][1];
261 quants->y_quant_fp[q][i] = quants->y_quant_fp[q][1];
262 quants->y_round_fp[q][i] = quants->y_round_fp[q][1];
263 quants->y_quant_shift[q][i] = quants->y_quant_shift[q][1];
264 quants->y_zbin[q][i] = quants->y_zbin[q][1];
265 quants->y_round[q][i] = quants->y_round[q][1];
266 cpi->y_dequant[q][i] = cpi->y_dequant[q][1];
267
268 quants->uv_quant[q][i] = quants->uv_quant[q][1];
269 quants->uv_quant_fp[q][i] = quants->uv_quant_fp[q][1];
270 quants->uv_round_fp[q][i] = quants->uv_round_fp[q][1];
271 quants->uv_quant_shift[q][i] = quants->uv_quant_shift[q][1];
272 quants->uv_zbin[q][i] = quants->uv_zbin[q][1];
273 quants->uv_round[q][i] = quants->uv_round[q][1];
274 cpi->uv_dequant[q][i] = cpi->uv_dequant[q][1];
275 }
276 }
277 }
278
vp9_init_plane_quantizers(VP9_COMP * cpi,MACROBLOCK * x)279 void vp9_init_plane_quantizers(VP9_COMP *cpi, MACROBLOCK *x) {
280 const VP9_COMMON *const cm = &cpi->common;
281 MACROBLOCKD *const xd = &x->e_mbd;
282 QUANTS *const quants = &cpi->quants;
283 const int segment_id = xd->mi[0]->segment_id;
284 const int qindex = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
285 const int rdmult = vp9_compute_rd_mult(cpi, qindex + cm->y_dc_delta_q);
286 int i;
287
288 // Y
289 x->plane[0].quant = quants->y_quant[qindex];
290 x->plane[0].quant_fp = quants->y_quant_fp[qindex];
291 x->plane[0].round_fp = quants->y_round_fp[qindex];
292 x->plane[0].quant_shift = quants->y_quant_shift[qindex];
293 x->plane[0].zbin = quants->y_zbin[qindex];
294 x->plane[0].round = quants->y_round[qindex];
295 xd->plane[0].dequant = cpi->y_dequant[qindex];
296
297 x->plane[0].quant_thred[0] = x->plane[0].zbin[0] * x->plane[0].zbin[0];
298 x->plane[0].quant_thred[1] = x->plane[0].zbin[1] * x->plane[0].zbin[1];
299
300 // UV
301 for (i = 1; i < 3; i++) {
302 x->plane[i].quant = quants->uv_quant[qindex];
303 x->plane[i].quant_fp = quants->uv_quant_fp[qindex];
304 x->plane[i].round_fp = quants->uv_round_fp[qindex];
305 x->plane[i].quant_shift = quants->uv_quant_shift[qindex];
306 x->plane[i].zbin = quants->uv_zbin[qindex];
307 x->plane[i].round = quants->uv_round[qindex];
308 xd->plane[i].dequant = cpi->uv_dequant[qindex];
309
310 x->plane[i].quant_thred[0] = x->plane[i].zbin[0] * x->plane[i].zbin[0];
311 x->plane[i].quant_thred[1] = x->plane[i].zbin[1] * x->plane[i].zbin[1];
312 }
313
314 x->skip_block = segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP);
315 x->q_index = qindex;
316
317 set_error_per_bit(x, rdmult);
318
319 vp9_initialize_me_consts(cpi, x, x->q_index);
320 }
321
vp9_frame_init_quantizer(VP9_COMP * cpi)322 void vp9_frame_init_quantizer(VP9_COMP *cpi) {
323 vp9_init_plane_quantizers(cpi, &cpi->td.mb);
324 }
325
vp9_set_quantizer(VP9_COMMON * cm,int q)326 void vp9_set_quantizer(VP9_COMMON *cm, int q) {
327 // quantizer has to be reinitialized with vp9_init_quantizer() if any
328 // delta_q changes.
329 cm->base_qindex = q;
330 cm->y_dc_delta_q = 0;
331 cm->uv_dc_delta_q = 0;
332 cm->uv_ac_delta_q = 0;
333 }
334
335 // Table that converts 0-63 Q-range values passed in outside to the Qindex
336 // range used internally.
337 static const int quantizer_to_qindex[] = {
338 0, 4, 8, 12, 16, 20, 24, 28, 32, 36, 40, 44, 48,
339 52, 56, 60, 64, 68, 72, 76, 80, 84, 88, 92, 96, 100,
340 104, 108, 112, 116, 120, 124, 128, 132, 136, 140, 144, 148, 152,
341 156, 160, 164, 168, 172, 176, 180, 184, 188, 192, 196, 200, 204,
342 208, 212, 216, 220, 224, 228, 232, 236, 240, 244, 249, 255,
343 };
344
vp9_quantizer_to_qindex(int quantizer)345 int vp9_quantizer_to_qindex(int quantizer) {
346 return quantizer_to_qindex[quantizer];
347 }
348
vp9_qindex_to_quantizer(int qindex)349 int vp9_qindex_to_quantizer(int qindex) {
350 int quantizer;
351
352 for (quantizer = 0; quantizer < 64; ++quantizer)
353 if (quantizer_to_qindex[quantizer] >= qindex) return quantizer;
354
355 return 63;
356 }
357