Line data Source code
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 "./av1_rtcd.h"
13 : #include "./aom_config.h"
14 : #include "./aom_dsp_rtcd.h"
15 :
16 : #include "aom_dsp/bitwriter.h"
17 : #include "aom_dsp/quantize.h"
18 : #include "aom_mem/aom_mem.h"
19 : #include "aom_ports/mem.h"
20 :
21 : #include "av1/common/idct.h"
22 : #include "av1/common/reconinter.h"
23 : #include "av1/common/reconintra.h"
24 : #include "av1/common/scan.h"
25 :
26 : #include "av1/encoder/av1_quantize.h"
27 : #include "av1/encoder/encodemb.h"
28 : #if CONFIG_LV_MAP
29 : #include "av1/encoder/encodetxb.h"
30 : #endif
31 : #include "av1/encoder/hybrid_fwd_txfm.h"
32 : #include "av1/encoder/rd.h"
33 : #include "av1/encoder/tokenize.h"
34 :
35 : #if CONFIG_PVQ
36 : #include "av1/encoder/encint.h"
37 : #include "av1/common/partition.h"
38 : #include "av1/encoder/pvq_encoder.h"
39 : #endif
40 :
41 : #if CONFIG_CFL
42 : #include "av1/common/cfl.h"
43 : #endif
44 :
45 : // Check if one needs to use c version subtraction.
46 0 : static int check_subtract_block_size(int w, int h) { return w < 4 || h < 4; }
47 :
48 0 : static void subtract_block(const MACROBLOCKD *xd, int rows, int cols,
49 : int16_t *diff, ptrdiff_t diff_stride,
50 : const uint8_t *src8, ptrdiff_t src_stride,
51 : const uint8_t *pred8, ptrdiff_t pred_stride) {
52 : #if !CONFIG_HIGHBITDEPTH
53 : (void)xd;
54 : #endif
55 :
56 0 : if (check_subtract_block_size(rows, cols)) {
57 : #if CONFIG_HIGHBITDEPTH
58 0 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
59 0 : aom_highbd_subtract_block_c(rows, cols, diff, diff_stride, src8,
60 : src_stride, pred8, pred_stride, xd->bd);
61 0 : return;
62 : }
63 : #endif // CONFIG_HIGHBITDEPTH
64 0 : aom_subtract_block_c(rows, cols, diff, diff_stride, src8, src_stride, pred8,
65 : pred_stride);
66 :
67 0 : return;
68 : }
69 :
70 : #if CONFIG_HIGHBITDEPTH
71 0 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
72 0 : aom_highbd_subtract_block(rows, cols, diff, diff_stride, src8, src_stride,
73 : pred8, pred_stride, xd->bd);
74 0 : return;
75 : }
76 : #endif // CONFIG_HIGHBITDEPTH
77 0 : aom_subtract_block(rows, cols, diff, diff_stride, src8, src_stride, pred8,
78 : pred_stride);
79 : }
80 :
81 0 : void av1_subtract_txb(MACROBLOCK *x, int plane, BLOCK_SIZE plane_bsize,
82 : int blk_col, int blk_row, TX_SIZE tx_size) {
83 0 : MACROBLOCKD *const xd = &x->e_mbd;
84 0 : struct macroblock_plane *const p = &x->plane[plane];
85 0 : const struct macroblockd_plane *const pd = &x->e_mbd.plane[plane];
86 0 : const int diff_stride = block_size_wide[plane_bsize];
87 0 : const int src_stride = p->src.stride;
88 0 : const int dst_stride = pd->dst.stride;
89 0 : const int tx1d_width = tx_size_wide[tx_size];
90 0 : const int tx1d_height = tx_size_high[tx_size];
91 0 : uint8_t *dst =
92 0 : &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
93 0 : uint8_t *src =
94 0 : &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]];
95 0 : int16_t *src_diff =
96 0 : &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
97 0 : subtract_block(xd, tx1d_height, tx1d_width, src_diff, diff_stride, src,
98 : src_stride, dst, dst_stride);
99 0 : }
100 :
101 0 : void av1_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
102 0 : struct macroblock_plane *const p = &x->plane[plane];
103 0 : const struct macroblockd_plane *const pd = &x->e_mbd.plane[plane];
104 0 : const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
105 0 : const int bw = block_size_wide[plane_bsize];
106 0 : const int bh = block_size_high[plane_bsize];
107 0 : const MACROBLOCKD *xd = &x->e_mbd;
108 :
109 0 : subtract_block(xd, bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
110 0 : pd->dst.buf, pd->dst.stride);
111 0 : }
112 :
113 : // These numbers are empirically obtained.
114 : static const int plane_rd_mult[REF_TYPES][PLANE_TYPES] = {
115 : #if CONFIG_EC_ADAPT
116 : { 10, 7 }, { 8, 5 },
117 : #else
118 : { 10, 6 }, { 8, 6 },
119 : #endif
120 : };
121 :
122 : #define UPDATE_RD_COST() \
123 : { \
124 : rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0); \
125 : rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1); \
126 : }
127 :
128 0 : static INLINE unsigned int get_token_bit_costs(
129 : unsigned int token_costs[2][COEFF_CONTEXTS][ENTROPY_TOKENS], int skip_eob,
130 : int ctx, int token) {
131 : (void)skip_eob;
132 0 : return token_costs[token == ZERO_TOKEN || token == EOB_TOKEN][ctx][token];
133 : }
134 :
135 : #if !CONFIG_LV_MAP
136 : #define USE_GREEDY_OPTIMIZE_B 0
137 :
138 : #if USE_GREEDY_OPTIMIZE_B
139 :
140 : typedef struct av1_token_state_greedy {
141 : int16_t token;
142 : tran_low_t qc;
143 : tran_low_t dqc;
144 : } av1_token_state_greedy;
145 :
146 : static int optimize_b_greedy(const AV1_COMMON *cm, MACROBLOCK *mb, int plane,
147 : int block, TX_SIZE tx_size, int ctx) {
148 : MACROBLOCKD *const xd = &mb->e_mbd;
149 : struct macroblock_plane *const p = &mb->plane[plane];
150 : struct macroblockd_plane *const pd = &xd->plane[plane];
151 : const int ref = is_inter_block(&xd->mi[0]->mbmi);
152 : av1_token_state_greedy tokens[MAX_TX_SQUARE + 1][2];
153 : uint8_t token_cache[MAX_TX_SQUARE];
154 : const tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
155 : tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
156 : tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
157 : const int eob = p->eobs[block];
158 : const PLANE_TYPE plane_type = pd->plane_type;
159 : const int16_t *const dequant_ptr = pd->dequant;
160 : const uint8_t *const band_translate = get_band_translate(tx_size);
161 : TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
162 : const SCAN_ORDER *const scan_order =
163 : get_scan(cm, tx_size, tx_type, is_inter_block(&xd->mi[0]->mbmi));
164 : const int16_t *const scan = scan_order->scan;
165 : const int16_t *const nb = scan_order->neighbors;
166 : int dqv;
167 : const int shift = av1_get_tx_scale(tx_size);
168 : #if CONFIG_AOM_QM
169 : int seg_id = xd->mi[0]->mbmi.segment_id;
170 : const qm_val_t *iqmatrix = pd->seg_iqmatrix[seg_id][!ref][tx_size];
171 : #endif
172 : #if CONFIG_NEW_QUANT
173 : int dq = get_dq_profile_from_ctx(mb->qindex, ctx, ref, plane_type);
174 : const dequant_val_type_nuq *dequant_val = pd->dequant_val_nuq[dq];
175 : #endif // CONFIG_NEW_QUANT
176 : int sz = 0;
177 : const int64_t rddiv = mb->rddiv;
178 : int64_t rd_cost0, rd_cost1;
179 : int16_t t0, t1;
180 : int i, final_eob;
181 : const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
182 : unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
183 : mb->token_costs[txsize_sqr_map[tx_size]][plane_type][ref];
184 : const int default_eob = tx_size_2d[tx_size];
185 :
186 : assert(mb->qindex > 0);
187 :
188 : assert((!plane_type && !plane) || (plane_type && plane));
189 : assert(eob <= default_eob);
190 :
191 : int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][plane_type]) >> 1;
192 :
193 : int64_t rate0, rate1;
194 : for (i = 0; i < eob; i++) {
195 : const int rc = scan[i];
196 : int x = qcoeff[rc];
197 : t0 = av1_get_token(x);
198 :
199 : tokens[i][0].qc = x;
200 : tokens[i][0].token = t0;
201 : tokens[i][0].dqc = dqcoeff[rc];
202 :
203 : token_cache[rc] = av1_pt_energy_class[t0];
204 : }
205 : tokens[eob][0].token = EOB_TOKEN;
206 : tokens[eob][0].qc = 0;
207 : tokens[eob][0].dqc = 0;
208 : tokens[eob][1] = tokens[eob][0];
209 :
210 : unsigned int(*token_costs_ptr)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
211 : token_costs;
212 :
213 : final_eob = 0;
214 :
215 : int64_t eob_cost0, eob_cost1;
216 :
217 : const int ctx0 = ctx;
218 : /* Record the r-d cost */
219 : int64_t accu_rate = 0;
220 : int64_t accu_error = 0;
221 :
222 : rate0 = get_token_bit_costs(*(token_costs_ptr + band_translate[0]), 0, ctx0,
223 : EOB_TOKEN);
224 : int64_t best_block_rd_cost = RDCOST(rdmult, rddiv, rate0, accu_error);
225 :
226 : // int64_t best_block_rd_cost_all0 = best_block_rd_cost;
227 :
228 : int x_prev = 1;
229 :
230 : for (i = 0; i < eob; i++) {
231 : const int rc = scan[i];
232 : int x = qcoeff[rc];
233 : sz = -(x < 0);
234 :
235 : int band_cur = band_translate[i];
236 : int ctx_cur = (i == 0) ? ctx : get_coef_context(nb, token_cache, i);
237 : int token_tree_sel_cur = (x_prev == 0);
238 :
239 : if (x == 0) {
240 : // no need to search when x == 0
241 : rate0 =
242 : get_token_bit_costs(*(token_costs_ptr + band_cur), token_tree_sel_cur,
243 : ctx_cur, tokens[i][0].token);
244 : accu_rate += rate0;
245 : x_prev = 0;
246 : // accu_error does not change when x==0
247 : } else {
248 : /* Computing distortion
249 : */
250 : // compute the distortion for the first candidate
251 : // and the distortion for quantizing to 0.
252 : int dx0 = (-coeff[rc]) * (1 << shift);
253 : #if CONFIG_HIGHBITDEPTH
254 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
255 : dx0 >>= xd->bd - 8;
256 : }
257 : #endif
258 : int64_t d0 = (int64_t)dx0 * dx0;
259 :
260 : int x_a = x - 2 * sz - 1;
261 : int64_t d2, d2_a;
262 :
263 : int dx;
264 :
265 : #if CONFIG_AOM_QM
266 : int iwt = iqmatrix[rc];
267 : dqv = dequant_ptr[rc != 0];
268 : dqv = ((iwt * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
269 : #else
270 : dqv = dequant_ptr[rc != 0];
271 : #endif
272 :
273 : dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
274 : #if CONFIG_HIGHBITDEPTH
275 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
276 : dx >>= xd->bd - 8;
277 : }
278 : #endif // CONFIG_HIGHBITDEPTH
279 : d2 = (int64_t)dx * dx;
280 :
281 : /* compute the distortion for the second candidate
282 : * x_a = x - 2 * sz + 1;
283 : */
284 : if (x_a != 0) {
285 : #if CONFIG_NEW_QUANT
286 : dx = av1_dequant_coeff_nuq(x, dqv, dequant_val[band_translate[i]]) -
287 : (coeff[rc] << shift);
288 : #if CONFIG_HIGHBITDEPTH
289 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
290 : dx >>= xd->bd - 8;
291 : }
292 : #endif // CONFIG_HIGHBITDEPTH
293 : #else // CONFIG_NEW_QUANT
294 : #if CONFIG_HIGHBITDEPTH
295 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
296 : dx -= ((dqv >> (xd->bd - 8)) + sz) ^ sz;
297 : } else {
298 : dx -= (dqv + sz) ^ sz;
299 : }
300 : #else
301 : dx -= (dqv + sz) ^ sz;
302 : #endif // CONFIG_HIGHBITDEPTH
303 : #endif // CONFIG_NEW_QUANT
304 : d2_a = (int64_t)dx * dx;
305 : } else {
306 : d2_a = d0;
307 : }
308 : /* Computing rates and r-d cost
309 : */
310 :
311 : int best_x, best_eob_x;
312 : int64_t base_bits, next_bits0, next_bits1;
313 : int64_t next_eob_bits0, next_eob_bits1;
314 :
315 : // rate cost of x
316 : base_bits = av1_get_token_cost(x, &t0, cat6_bits);
317 : rate0 = base_bits + get_token_bit_costs(*(token_costs_ptr + band_cur),
318 : token_tree_sel_cur, ctx_cur, t0);
319 :
320 : base_bits = av1_get_token_cost(x_a, &t1, cat6_bits);
321 : rate1 = base_bits + get_token_bit_costs(*(token_costs_ptr + band_cur),
322 : token_tree_sel_cur, ctx_cur, t1);
323 :
324 : next_bits0 = 0;
325 : next_bits1 = 0;
326 : next_eob_bits0 = 0;
327 : next_eob_bits1 = 0;
328 :
329 : if (i < default_eob - 1) {
330 : int ctx_next, token_tree_sel_next;
331 : int band_next = band_translate[i + 1];
332 :
333 : token_cache[rc] = av1_pt_energy_class[t0];
334 : ctx_next = get_coef_context(nb, token_cache, i + 1);
335 : token_tree_sel_next = (x == 0);
336 :
337 : next_bits0 = get_token_bit_costs(*(token_costs_ptr + band_next),
338 : token_tree_sel_next, ctx_next,
339 : tokens[i + 1][0].token);
340 : next_eob_bits0 =
341 : get_token_bit_costs(*(token_costs_ptr + band_next),
342 : token_tree_sel_next, ctx_next, EOB_TOKEN);
343 :
344 : token_cache[rc] = av1_pt_energy_class[t1];
345 : ctx_next = get_coef_context(nb, token_cache, i + 1);
346 : token_tree_sel_next = (x_a == 0);
347 :
348 : next_bits1 = get_token_bit_costs(*(token_costs_ptr + band_next),
349 : token_tree_sel_next, ctx_next,
350 : tokens[i + 1][0].token);
351 :
352 : if (x_a != 0) {
353 : next_eob_bits1 =
354 : get_token_bit_costs(*(token_costs_ptr + band_next),
355 : token_tree_sel_next, ctx_next, EOB_TOKEN);
356 : }
357 : }
358 :
359 : rd_cost0 = RDCOST(rdmult, rddiv, (rate0 + next_bits0), d2);
360 : rd_cost1 = RDCOST(rdmult, rddiv, (rate1 + next_bits1), d2_a);
361 :
362 : best_x = (rd_cost1 < rd_cost0);
363 :
364 : eob_cost0 = RDCOST(rdmult, rddiv, (accu_rate + rate0 + next_eob_bits0),
365 : (accu_error + d2 - d0));
366 : eob_cost1 = eob_cost0;
367 : if (x_a != 0) {
368 : eob_cost1 = RDCOST(rdmult, rddiv, (accu_rate + rate1 + next_eob_bits1),
369 : (accu_error + d2_a - d0));
370 : best_eob_x = (eob_cost1 < eob_cost0);
371 : } else {
372 : best_eob_x = 0;
373 : }
374 :
375 : int dqc, dqc_a = 0;
376 :
377 : dqc = dqcoeff[rc];
378 : if (best_x + best_eob_x) {
379 : if (x_a != 0) {
380 : #if CONFIG_NEW_QUANT
381 : dqc_a = av1_dequant_abscoeff_nuq(abs(x_a), dqv,
382 : dequant_val[band_translate[i]]);
383 : dqc_a = shift ? ROUND_POWER_OF_TWO(dqc_a, shift) : dqc_a;
384 : if (sz) dqc_a = -dqc_a;
385 : #else
386 : if (x_a < 0)
387 : dqc_a = -((-x_a * dqv) >> shift);
388 : else
389 : dqc_a = (x_a * dqv) >> shift;
390 : #endif // CONFIG_NEW_QUANT
391 : } else {
392 : dqc_a = 0;
393 : } // if (x_a != 0)
394 : }
395 :
396 : // record the better quantized value
397 : if (best_x) {
398 : qcoeff[rc] = x_a;
399 : dqcoeff[rc] = dqc_a;
400 :
401 : accu_rate += rate1;
402 : accu_error += d2_a - d0;
403 : assert(d2_a <= d0);
404 :
405 : token_cache[rc] = av1_pt_energy_class[t1];
406 : } else {
407 : accu_rate += rate0;
408 : accu_error += d2 - d0;
409 : assert(d2 <= d0);
410 :
411 : token_cache[rc] = av1_pt_energy_class[t0];
412 : }
413 :
414 : x_prev = qcoeff[rc];
415 :
416 : // determine whether to move the eob position to i+1
417 : int64_t best_eob_cost_i = eob_cost0;
418 :
419 : tokens[i][1].token = t0;
420 : tokens[i][1].qc = x;
421 : tokens[i][1].dqc = dqc;
422 :
423 : if ((x_a != 0) && (best_eob_x)) {
424 : best_eob_cost_i = eob_cost1;
425 :
426 : tokens[i][1].token = t1;
427 : tokens[i][1].qc = x_a;
428 : tokens[i][1].dqc = dqc_a;
429 : }
430 :
431 : if (best_eob_cost_i < best_block_rd_cost) {
432 : best_block_rd_cost = best_eob_cost_i;
433 : final_eob = i + 1;
434 : }
435 : } // if (x==0)
436 : } // for (i)
437 :
438 : assert(final_eob <= eob);
439 : if (final_eob > 0) {
440 : assert(tokens[final_eob - 1][1].qc != 0);
441 : i = final_eob - 1;
442 : int rc = scan[i];
443 : qcoeff[rc] = tokens[i][1].qc;
444 : dqcoeff[rc] = tokens[i][1].dqc;
445 : }
446 :
447 : for (i = final_eob; i < eob; i++) {
448 : int rc = scan[i];
449 : qcoeff[rc] = 0;
450 : dqcoeff[rc] = 0;
451 : }
452 :
453 : mb->plane[plane].eobs[block] = final_eob;
454 : return final_eob;
455 : }
456 :
457 : #else // USE_GREEDY_OPTIMIZE_B
458 :
459 : typedef struct av1_token_state_org {
460 : int64_t error;
461 : int rate;
462 : int16_t next;
463 : int16_t token;
464 : tran_low_t qc;
465 : tran_low_t dqc;
466 : uint8_t best_index;
467 : } av1_token_state_org;
468 :
469 0 : static int optimize_b_org(const AV1_COMMON *cm, MACROBLOCK *mb, int plane,
470 : int block, TX_SIZE tx_size, int ctx) {
471 0 : MACROBLOCKD *const xd = &mb->e_mbd;
472 0 : struct macroblock_plane *const p = &mb->plane[plane];
473 0 : struct macroblockd_plane *const pd = &xd->plane[plane];
474 0 : const int ref = is_inter_block(&xd->mi[0]->mbmi);
475 : av1_token_state_org tokens[MAX_TX_SQUARE + 1][2];
476 : uint8_t token_cache[MAX_TX_SQUARE];
477 0 : const tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
478 0 : tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
479 0 : tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
480 0 : const int eob = p->eobs[block];
481 0 : const PLANE_TYPE plane_type = pd->plane_type;
482 0 : const int default_eob = tx_size_2d[tx_size];
483 0 : const int16_t *const dequant_ptr = pd->dequant;
484 0 : const uint8_t *const band_translate = get_band_translate(tx_size);
485 0 : TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
486 0 : const SCAN_ORDER *const scan_order =
487 0 : get_scan(cm, tx_size, tx_type, is_inter_block(&xd->mi[0]->mbmi));
488 0 : const int16_t *const scan = scan_order->scan;
489 0 : const int16_t *const nb = scan_order->neighbors;
490 : int dqv;
491 0 : const int shift = av1_get_tx_scale(tx_size);
492 : #if CONFIG_AOM_QM
493 : int seg_id = xd->mi[0]->mbmi.segment_id;
494 : const qm_val_t *iqmatrix = pd->seg_iqmatrix[seg_id][!ref][tx_size];
495 : #endif
496 : #if CONFIG_NEW_QUANT
497 : int dq = get_dq_profile_from_ctx(mb->qindex, ctx, ref, plane_type);
498 : const dequant_val_type_nuq *dequant_val = pd->dequant_val_nuq[dq];
499 : #endif // CONFIG_NEW_QUANT
500 0 : int next = eob, sz = 0;
501 0 : const int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][plane_type]) >> 1;
502 0 : const int64_t rddiv = mb->rddiv;
503 : int64_t rd_cost0, rd_cost1;
504 : int rate0, rate1;
505 : int64_t error0, error1;
506 : int16_t t0, t1;
507 0 : int best, band = (eob < default_eob) ? band_translate[eob]
508 0 : : band_translate[eob - 1];
509 : int pt, i, final_eob;
510 0 : const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
511 0 : unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
512 0 : mb->token_costs[txsize_sqr_map[tx_size]][plane_type][ref];
513 0 : const uint16_t *band_counts = &band_count_table[tx_size][band];
514 0 : uint16_t band_left = eob - band_cum_count_table[tx_size][band] + 1;
515 0 : int shortcut = 0;
516 0 : int next_shortcut = 0;
517 :
518 : #if CONFIG_EXT_DELTA_Q
519 0 : const int qindex = cm->seg.enabled
520 0 : ? av1_get_qindex(&cm->seg, xd->mi[0]->mbmi.segment_id,
521 : cm->base_qindex)
522 0 : : cm->base_qindex;
523 0 : assert(qindex > 0);
524 : (void)qindex;
525 : #else
526 : assert(mb->qindex > 0);
527 : #endif
528 :
529 0 : token_costs += band;
530 :
531 0 : assert((!plane_type && !plane) || (plane_type && plane));
532 0 : assert(eob <= default_eob);
533 :
534 : /* Now set up a Viterbi trellis to evaluate alternative roundings. */
535 : /* Initialize the sentinel node of the trellis. */
536 0 : tokens[eob][0].rate = 0;
537 0 : tokens[eob][0].error = 0;
538 0 : tokens[eob][0].next = default_eob;
539 0 : tokens[eob][0].token = EOB_TOKEN;
540 0 : tokens[eob][0].qc = 0;
541 0 : tokens[eob][1] = tokens[eob][0];
542 :
543 0 : for (i = 0; i < eob; i++) {
544 0 : const int rc = scan[i];
545 0 : tokens[i][0].rate = av1_get_token_cost(qcoeff[rc], &t0, cat6_bits);
546 0 : tokens[i][0].token = t0;
547 0 : token_cache[rc] = av1_pt_energy_class[t0];
548 : }
549 :
550 0 : for (i = eob; i-- > 0;) {
551 : int base_bits, dx;
552 : int64_t d2;
553 0 : const int rc = scan[i];
554 0 : int x = qcoeff[rc];
555 : #if CONFIG_AOM_QM
556 : int iwt = iqmatrix[rc];
557 : dqv = dequant_ptr[rc != 0];
558 : dqv = ((iwt * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
559 : #else
560 0 : dqv = dequant_ptr[rc != 0];
561 : #endif
562 0 : next_shortcut = shortcut;
563 :
564 : /* Only add a trellis state for non-zero coefficients. */
565 0 : if (UNLIKELY(x)) {
566 0 : error0 = tokens[next][0].error;
567 0 : error1 = tokens[next][1].error;
568 : /* Evaluate the first possibility for this state. */
569 0 : rate0 = tokens[next][0].rate;
570 0 : rate1 = tokens[next][1].rate;
571 :
572 0 : if (next_shortcut) {
573 : /* Consider both possible successor states. */
574 0 : if (next < default_eob) {
575 0 : pt = get_coef_context(nb, token_cache, i + 1);
576 0 : rate0 +=
577 0 : get_token_bit_costs(*token_costs, 0, pt, tokens[next][0].token);
578 0 : rate1 +=
579 0 : get_token_bit_costs(*token_costs, 0, pt, tokens[next][1].token);
580 : }
581 0 : UPDATE_RD_COST();
582 : /* And pick the best. */
583 0 : best = rd_cost1 < rd_cost0;
584 : } else {
585 0 : if (next < default_eob) {
586 0 : pt = get_coef_context(nb, token_cache, i + 1);
587 0 : rate0 +=
588 0 : get_token_bit_costs(*token_costs, 0, pt, tokens[next][0].token);
589 : }
590 0 : best = 0;
591 : }
592 :
593 0 : dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
594 : #if CONFIG_HIGHBITDEPTH
595 0 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
596 0 : dx >>= xd->bd - 8;
597 : }
598 : #endif // CONFIG_HIGHBITDEPTH
599 0 : d2 = (int64_t)dx * dx;
600 0 : tokens[i][0].rate += (best ? rate1 : rate0);
601 0 : tokens[i][0].error = d2 + (best ? error1 : error0);
602 0 : tokens[i][0].next = next;
603 0 : tokens[i][0].qc = x;
604 0 : tokens[i][0].dqc = dqcoeff[rc];
605 0 : tokens[i][0].best_index = best;
606 :
607 : /* Evaluate the second possibility for this state. */
608 0 : rate0 = tokens[next][0].rate;
609 0 : rate1 = tokens[next][1].rate;
610 :
611 : // The threshold of 3 is empirically obtained.
612 0 : if (UNLIKELY(abs(x) > 3)) {
613 0 : shortcut = 0;
614 : } else {
615 : #if CONFIG_NEW_QUANT
616 : shortcut = ((av1_dequant_abscoeff_nuq(abs(x), dqv,
617 : dequant_val[band_translate[i]]) >
618 : (abs(coeff[rc]) << shift)) &&
619 : (av1_dequant_abscoeff_nuq(abs(x) - 1, dqv,
620 : dequant_val[band_translate[i]]) <
621 : (abs(coeff[rc]) << shift)));
622 : #else // CONFIG_NEW_QUANT
623 : #if CONFIG_AOM_QM
624 : if ((abs(x) * dequant_ptr[rc != 0] * iwt >
625 : ((abs(coeff[rc]) << shift) << AOM_QM_BITS)) &&
626 : (abs(x) * dequant_ptr[rc != 0] * iwt <
627 : (((abs(coeff[rc]) << shift) + dequant_ptr[rc != 0])
628 : << AOM_QM_BITS)))
629 : #else
630 0 : if ((abs(x) * dequant_ptr[rc != 0] > (abs(coeff[rc]) << shift)) &&
631 0 : (abs(x) * dequant_ptr[rc != 0] <
632 0 : (abs(coeff[rc]) << shift) + dequant_ptr[rc != 0]))
633 : #endif // CONFIG_AOM_QM
634 0 : shortcut = 1;
635 : else
636 0 : shortcut = 0;
637 : #endif // CONFIG_NEW_QUANT
638 : }
639 :
640 0 : if (shortcut) {
641 0 : sz = -(x < 0);
642 0 : x -= 2 * sz + 1;
643 : } else {
644 0 : tokens[i][1] = tokens[i][0];
645 0 : next = i;
646 :
647 0 : if (UNLIKELY(!(--band_left))) {
648 0 : --band_counts;
649 0 : band_left = *band_counts;
650 0 : --token_costs;
651 : }
652 0 : continue;
653 : }
654 :
655 : /* Consider both possible successor states. */
656 0 : if (!x) {
657 : /* If we reduced this coefficient to zero, check to see if
658 : * we need to move the EOB back here.
659 : */
660 0 : t0 = tokens[next][0].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
661 0 : t1 = tokens[next][1].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
662 0 : base_bits = 0;
663 : } else {
664 0 : base_bits = av1_get_token_cost(x, &t0, cat6_bits);
665 0 : t1 = t0;
666 : }
667 :
668 0 : if (next_shortcut) {
669 0 : if (LIKELY(next < default_eob)) {
670 0 : if (t0 != EOB_TOKEN) {
671 0 : token_cache[rc] = av1_pt_energy_class[t0];
672 0 : pt = get_coef_context(nb, token_cache, i + 1);
673 0 : rate0 += get_token_bit_costs(*token_costs, !x, pt,
674 0 : tokens[next][0].token);
675 : }
676 0 : if (t1 != EOB_TOKEN) {
677 0 : token_cache[rc] = av1_pt_energy_class[t1];
678 0 : pt = get_coef_context(nb, token_cache, i + 1);
679 0 : rate1 += get_token_bit_costs(*token_costs, !x, pt,
680 0 : tokens[next][1].token);
681 : }
682 : }
683 :
684 0 : UPDATE_RD_COST();
685 : /* And pick the best. */
686 0 : best = rd_cost1 < rd_cost0;
687 : } else {
688 : // The two states in next stage are identical.
689 0 : if (next < default_eob && t0 != EOB_TOKEN) {
690 0 : token_cache[rc] = av1_pt_energy_class[t0];
691 0 : pt = get_coef_context(nb, token_cache, i + 1);
692 0 : rate0 +=
693 0 : get_token_bit_costs(*token_costs, !x, pt, tokens[next][0].token);
694 : }
695 0 : best = 0;
696 : }
697 :
698 : #if CONFIG_NEW_QUANT
699 : dx = av1_dequant_coeff_nuq(x, dqv, dequant_val[band_translate[i]]) -
700 : (coeff[rc] << shift);
701 : #if CONFIG_HIGHBITDEPTH
702 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
703 : dx >>= xd->bd - 8;
704 : }
705 : #endif // CONFIG_HIGHBITDEPTH
706 : #else // CONFIG_NEW_QUANT
707 : #if CONFIG_HIGHBITDEPTH
708 0 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
709 0 : dx -= ((dqv >> (xd->bd - 8)) + sz) ^ sz;
710 : } else {
711 0 : dx -= (dqv + sz) ^ sz;
712 : }
713 : #else
714 : dx -= (dqv + sz) ^ sz;
715 : #endif // CONFIG_HIGHBITDEPTH
716 : #endif // CONFIG_NEW_QUANT
717 0 : d2 = (int64_t)dx * dx;
718 :
719 0 : tokens[i][1].rate = base_bits + (best ? rate1 : rate0);
720 0 : tokens[i][1].error = d2 + (best ? error1 : error0);
721 0 : tokens[i][1].next = next;
722 0 : tokens[i][1].token = best ? t1 : t0;
723 0 : tokens[i][1].qc = x;
724 :
725 0 : if (x) {
726 : #if CONFIG_NEW_QUANT
727 : tokens[i][1].dqc = av1_dequant_abscoeff_nuq(
728 : abs(x), dqv, dequant_val[band_translate[i]]);
729 : tokens[i][1].dqc = shift ? ROUND_POWER_OF_TWO(tokens[i][1].dqc, shift)
730 : : tokens[i][1].dqc;
731 : if (sz) tokens[i][1].dqc = -tokens[i][1].dqc;
732 : #else
733 0 : if (x < 0)
734 0 : tokens[i][1].dqc = -((-x * dqv) >> shift);
735 : else
736 0 : tokens[i][1].dqc = (x * dqv) >> shift;
737 : #endif // CONFIG_NEW_QUANT
738 : } else {
739 0 : tokens[i][1].dqc = 0;
740 : }
741 :
742 0 : tokens[i][1].best_index = best;
743 : /* Finally, make this the new head of the trellis. */
744 0 : next = i;
745 : } else {
746 : /* There's no choice to make for a zero coefficient, so we don't
747 : * add a new trellis node, but we do need to update the costs.
748 : */
749 0 : t0 = tokens[next][0].token;
750 0 : t1 = tokens[next][1].token;
751 0 : pt = get_coef_context(nb, token_cache, i + 1);
752 : /* Update the cost of each path if we're past the EOB token. */
753 0 : if (t0 != EOB_TOKEN) {
754 0 : tokens[next][0].rate += get_token_bit_costs(*token_costs, 1, pt, t0);
755 0 : tokens[next][0].token = ZERO_TOKEN;
756 : }
757 0 : if (t1 != EOB_TOKEN) {
758 0 : tokens[next][1].rate += get_token_bit_costs(*token_costs, 1, pt, t1);
759 0 : tokens[next][1].token = ZERO_TOKEN;
760 : }
761 0 : tokens[i][0].best_index = tokens[i][1].best_index = 0;
762 0 : shortcut = (tokens[next][0].rate != tokens[next][1].rate);
763 : /* Don't update next, because we didn't add a new node. */
764 : }
765 :
766 0 : if (UNLIKELY(!(--band_left))) {
767 0 : --band_counts;
768 0 : band_left = *band_counts;
769 0 : --token_costs;
770 : }
771 : }
772 :
773 : /* Now pick the best path through the whole trellis. */
774 0 : rate0 = tokens[next][0].rate;
775 0 : rate1 = tokens[next][1].rate;
776 0 : error0 = tokens[next][0].error;
777 0 : error1 = tokens[next][1].error;
778 0 : t0 = tokens[next][0].token;
779 0 : t1 = tokens[next][1].token;
780 0 : rate0 += get_token_bit_costs(*token_costs, 0, ctx, t0);
781 0 : rate1 += get_token_bit_costs(*token_costs, 0, ctx, t1);
782 0 : UPDATE_RD_COST();
783 0 : best = rd_cost1 < rd_cost0;
784 :
785 0 : final_eob = -1;
786 :
787 0 : for (i = next; i < eob; i = next) {
788 0 : const int x = tokens[i][best].qc;
789 0 : const int rc = scan[i];
790 0 : if (x) final_eob = i;
791 0 : qcoeff[rc] = x;
792 0 : dqcoeff[rc] = tokens[i][best].dqc;
793 :
794 0 : next = tokens[i][best].next;
795 0 : best = tokens[i][best].best_index;
796 : }
797 0 : final_eob++;
798 :
799 0 : mb->plane[plane].eobs[block] = final_eob;
800 0 : assert(final_eob <= default_eob);
801 0 : return final_eob;
802 : }
803 :
804 : #endif // USE_GREEDY_OPTIMIZE_B
805 : #endif // !CONFIG_LV_MAP
806 :
807 0 : int av1_optimize_b(const AV1_COMMON *cm, MACROBLOCK *mb, int plane, int block,
808 : BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
809 : const ENTROPY_CONTEXT *a, const ENTROPY_CONTEXT *l) {
810 0 : MACROBLOCKD *const xd = &mb->e_mbd;
811 0 : struct macroblock_plane *const p = &mb->plane[plane];
812 0 : const int eob = p->eobs[block];
813 0 : assert((mb->qindex == 0) ^ (xd->lossless[xd->mi[0]->mbmi.segment_id] == 0));
814 0 : if (eob == 0) return eob;
815 0 : if (xd->lossless[xd->mi[0]->mbmi.segment_id]) return eob;
816 : #if CONFIG_PVQ
817 : (void)cm;
818 : (void)tx_size;
819 : (void)a;
820 : (void)l;
821 : return eob;
822 : #endif
823 :
824 : #if !CONFIG_LV_MAP
825 : (void)plane_bsize;
826 : #if CONFIG_VAR_TX
827 0 : int ctx = get_entropy_context(tx_size, a, l);
828 : #else
829 : int ctx = combine_entropy_contexts(*a, *l);
830 : #endif
831 :
832 : #if USE_GREEDY_OPTIMIZE_B
833 : return optimize_b_greedy(cm, mb, plane, block, tx_size, ctx);
834 : #else // USE_GREEDY_OPTIMIZE_B
835 0 : return optimize_b_org(cm, mb, plane, block, tx_size, ctx);
836 : #endif // USE_GREEDY_OPTIMIZE_B
837 : #else // !CONFIG_LV_MAP
838 : TXB_CTX txb_ctx;
839 : get_txb_ctx(plane_bsize, tx_size, plane, a, l, &txb_ctx);
840 : return av1_optimize_txb(cm, mb, plane, block, tx_size, &txb_ctx);
841 : #endif // !CONFIG_LV_MAP
842 : }
843 :
844 : #if !CONFIG_PVQ
845 : #if CONFIG_HIGHBITDEPTH
846 : typedef enum QUANT_FUNC {
847 : QUANT_FUNC_LOWBD = 0,
848 : QUANT_FUNC_HIGHBD = 1,
849 : QUANT_FUNC_TYPES = 2
850 : } QUANT_FUNC;
851 :
852 : static AV1_QUANT_FACADE
853 : quant_func_list[AV1_XFORM_QUANT_TYPES][QUANT_FUNC_TYPES] = {
854 : #if !CONFIG_NEW_QUANT
855 : { av1_quantize_fp_facade, av1_highbd_quantize_fp_facade },
856 : { av1_quantize_b_facade, av1_highbd_quantize_b_facade },
857 : { av1_quantize_dc_facade, av1_highbd_quantize_dc_facade },
858 : #else // !CONFIG_NEW_QUANT
859 : { av1_quantize_fp_nuq_facade, av1_highbd_quantize_fp_nuq_facade },
860 : { av1_quantize_b_nuq_facade, av1_highbd_quantize_b_nuq_facade },
861 : { av1_quantize_dc_nuq_facade, av1_highbd_quantize_dc_nuq_facade },
862 : #endif // !CONFIG_NEW_QUANT
863 : { NULL, NULL }
864 : };
865 :
866 : #else
867 :
868 : typedef enum QUANT_FUNC {
869 : QUANT_FUNC_LOWBD = 0,
870 : QUANT_FUNC_TYPES = 1
871 : } QUANT_FUNC;
872 :
873 : static AV1_QUANT_FACADE quant_func_list[AV1_XFORM_QUANT_TYPES]
874 : [QUANT_FUNC_TYPES] = {
875 : #if !CONFIG_NEW_QUANT
876 : { av1_quantize_fp_facade },
877 : { av1_quantize_b_facade },
878 : { av1_quantize_dc_facade },
879 : #else // !CONFIG_NEW_QUANT
880 : { av1_quantize_fp_nuq_facade },
881 : { av1_quantize_b_nuq_facade },
882 : { av1_quantize_dc_nuq_facade },
883 : #endif // !CONFIG_NEW_QUANT
884 : { NULL }
885 : };
886 : #endif // CONFIG_HIGHBITDEPTH
887 : #endif // CONFIG_PVQ
888 :
889 0 : void av1_xform_quant(const AV1_COMMON *cm, MACROBLOCK *x, int plane, int block,
890 : int blk_row, int blk_col, BLOCK_SIZE plane_bsize,
891 : TX_SIZE tx_size, int ctx,
892 : AV1_XFORM_QUANT xform_quant_idx) {
893 0 : MACROBLOCKD *const xd = &x->e_mbd;
894 0 : MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
895 : #if !(CONFIG_PVQ || CONFIG_DAALA_DIST)
896 0 : const struct macroblock_plane *const p = &x->plane[plane];
897 0 : const struct macroblockd_plane *const pd = &xd->plane[plane];
898 : #else
899 : struct macroblock_plane *const p = &x->plane[plane];
900 : struct macroblockd_plane *const pd = &xd->plane[plane];
901 : #endif
902 0 : PLANE_TYPE plane_type = get_plane_type(plane);
903 0 : TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
904 0 : const int is_inter = is_inter_block(mbmi);
905 0 : const SCAN_ORDER *const scan_order = get_scan(cm, tx_size, tx_type, is_inter);
906 0 : tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
907 0 : tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
908 0 : tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
909 0 : uint16_t *const eob = &p->eobs[block];
910 0 : const int diff_stride = block_size_wide[plane_bsize];
911 : #if CONFIG_AOM_QM
912 : int seg_id = mbmi->segment_id;
913 : const qm_val_t *qmatrix = pd->seg_qmatrix[seg_id][!is_inter][tx_size];
914 : const qm_val_t *iqmatrix = pd->seg_iqmatrix[seg_id][!is_inter][tx_size];
915 : #endif
916 :
917 : FWD_TXFM_PARAM fwd_txfm_param;
918 :
919 : #if CONFIG_PVQ || CONFIG_DAALA_DIST
920 : uint8_t *dst;
921 : int16_t *pred;
922 : const int dst_stride = pd->dst.stride;
923 : int tx_blk_size;
924 : int i, j;
925 : #endif
926 :
927 : #if !CONFIG_PVQ
928 0 : const int tx2d_size = tx_size_2d[tx_size];
929 : QUANT_PARAM qparam;
930 : const int16_t *src_diff;
931 :
932 0 : src_diff =
933 0 : &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
934 0 : qparam.log_scale = av1_get_tx_scale(tx_size);
935 : #if CONFIG_NEW_QUANT
936 : qparam.tx_size = tx_size;
937 : qparam.dq = get_dq_profile_from_ctx(x->qindex, ctx, is_inter, plane_type);
938 : #endif // CONFIG_NEW_QUANT
939 : #if CONFIG_AOM_QM
940 : qparam.qmatrix = qmatrix;
941 : qparam.iqmatrix = iqmatrix;
942 : #endif // CONFIG_AOM_QM
943 : #else
944 : tran_low_t *ref_coeff = BLOCK_OFFSET(pd->pvq_ref_coeff, block);
945 : int skip = 1;
946 : PVQ_INFO *pvq_info = NULL;
947 : uint8_t *src;
948 : int16_t *src_int16;
949 : const int src_stride = p->src.stride;
950 :
951 : (void)ctx;
952 : (void)scan_order;
953 : (void)qcoeff;
954 :
955 : if (x->pvq_coded) {
956 : assert(block < MAX_PVQ_BLOCKS_IN_SB);
957 : pvq_info = &x->pvq[block][plane];
958 : }
959 : src = &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]];
960 : src_int16 =
961 : &p->src_int16[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
962 :
963 : // transform block size in pixels
964 : tx_blk_size = tx_size_wide[tx_size];
965 : #if CONFIG_HIGHBITDEPTH
966 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
967 : for (j = 0; j < tx_blk_size; j++)
968 : for (i = 0; i < tx_blk_size; i++)
969 : src_int16[diff_stride * j + i] =
970 : CONVERT_TO_SHORTPTR(src)[src_stride * j + i];
971 : } else {
972 : #endif // CONFIG_HIGHBITDEPTH
973 : for (j = 0; j < tx_blk_size; j++)
974 : for (i = 0; i < tx_blk_size; i++)
975 : src_int16[diff_stride * j + i] = src[src_stride * j + i];
976 : #if CONFIG_HIGHBITDEPTH
977 : }
978 : #endif // CONFIG_HIGHBITDEPTH
979 : #endif
980 :
981 : #if CONFIG_PVQ || CONFIG_DAALA_DIST
982 : dst = &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
983 : pred = &pd->pred[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
984 :
985 : // transform block size in pixels
986 : tx_blk_size = tx_size_wide[tx_size];
987 :
988 : // copy uint8 orig and predicted block to int16 buffer
989 : // in order to use existing VP10 transform functions
990 : #if CONFIG_HIGHBITDEPTH
991 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
992 : for (j = 0; j < tx_blk_size; j++)
993 : for (i = 0; i < tx_blk_size; i++)
994 : pred[diff_stride * j + i] =
995 : CONVERT_TO_SHORTPTR(dst)[dst_stride * j + i];
996 : } else {
997 : #endif // CONFIG_HIGHBITDEPTH
998 : for (j = 0; j < tx_blk_size; j++)
999 : for (i = 0; i < tx_blk_size; i++)
1000 : pred[diff_stride * j + i] = dst[dst_stride * j + i];
1001 : #if CONFIG_HIGHBITDEPTH
1002 : }
1003 : #endif // CONFIG_HIGHBITDEPTH
1004 : #endif
1005 :
1006 : (void)ctx;
1007 :
1008 0 : fwd_txfm_param.tx_type = tx_type;
1009 0 : fwd_txfm_param.tx_size = tx_size;
1010 0 : fwd_txfm_param.lossless = xd->lossless[mbmi->segment_id];
1011 :
1012 : #if !CONFIG_PVQ
1013 : #if CONFIG_HIGHBITDEPTH
1014 0 : fwd_txfm_param.bd = xd->bd;
1015 0 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1016 0 : av1_highbd_fwd_txfm(src_diff, coeff, diff_stride, &fwd_txfm_param);
1017 0 : if (xform_quant_idx != AV1_XFORM_QUANT_SKIP_QUANT) {
1018 0 : if (LIKELY(!x->skip_block)) {
1019 0 : quant_func_list[xform_quant_idx][QUANT_FUNC_HIGHBD](
1020 : coeff, tx2d_size, p, qcoeff, pd, dqcoeff, eob, scan_order, &qparam);
1021 : } else {
1022 0 : av1_quantize_skip(tx2d_size, qcoeff, dqcoeff, eob);
1023 : }
1024 : }
1025 : #if CONFIG_LV_MAP
1026 : p->txb_entropy_ctx[block] =
1027 : (uint8_t)av1_get_txb_entropy_context(qcoeff, scan_order, *eob);
1028 : #endif // CONFIG_LV_MAP
1029 0 : return;
1030 : }
1031 : #endif // CONFIG_HIGHBITDEPTH
1032 0 : av1_fwd_txfm(src_diff, coeff, diff_stride, &fwd_txfm_param);
1033 0 : if (xform_quant_idx != AV1_XFORM_QUANT_SKIP_QUANT) {
1034 0 : if (LIKELY(!x->skip_block)) {
1035 0 : quant_func_list[xform_quant_idx][QUANT_FUNC_LOWBD](
1036 : coeff, tx2d_size, p, qcoeff, pd, dqcoeff, eob, scan_order, &qparam);
1037 : } else {
1038 0 : av1_quantize_skip(tx2d_size, qcoeff, dqcoeff, eob);
1039 : }
1040 : }
1041 : #if CONFIG_LV_MAP
1042 : p->txb_entropy_ctx[block] =
1043 : (uint8_t)av1_get_txb_entropy_context(qcoeff, scan_order, *eob);
1044 : #endif // CONFIG_LV_MAP
1045 : #else // #if !CONFIG_PVQ
1046 : (void)xform_quant_idx;
1047 : #if CONFIG_HIGHBITDEPTH
1048 : fwd_txfm_param.bd = xd->bd;
1049 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1050 : av1_highbd_fwd_txfm(src_int16, coeff, diff_stride, &fwd_txfm_param);
1051 : av1_highbd_fwd_txfm(pred, ref_coeff, diff_stride, &fwd_txfm_param);
1052 : } else {
1053 : #endif
1054 : av1_fwd_txfm(src_int16, coeff, diff_stride, &fwd_txfm_param);
1055 : av1_fwd_txfm(pred, ref_coeff, diff_stride, &fwd_txfm_param);
1056 : #if CONFIG_HIGHBITDEPTH
1057 : }
1058 : #endif
1059 :
1060 : // PVQ for inter mode block
1061 : if (!x->skip_block) {
1062 : PVQ_SKIP_TYPE ac_dc_coded =
1063 : av1_pvq_encode_helper(x,
1064 : coeff, // target original vector
1065 : ref_coeff, // reference vector
1066 : dqcoeff, // de-quantized vector
1067 : eob, // End of Block marker
1068 : pd->dequant, // aom's quantizers
1069 : plane, // image plane
1070 : tx_size, // block size in log_2 - 2
1071 : tx_type,
1072 : &x->rate, // rate measured
1073 : x->pvq_speed,
1074 : pvq_info); // PVQ info for a block
1075 : skip = ac_dc_coded == PVQ_SKIP;
1076 : }
1077 : x->pvq_skip[plane] = skip;
1078 :
1079 : if (!skip) mbmi->skip = 0;
1080 : #endif // #if !CONFIG_PVQ
1081 : }
1082 :
1083 0 : static void encode_block(int plane, int block, int blk_row, int blk_col,
1084 : BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
1085 0 : struct encode_b_args *const args = arg;
1086 0 : AV1_COMMON *cm = args->cm;
1087 0 : MACROBLOCK *const x = args->x;
1088 0 : MACROBLOCKD *const xd = &x->e_mbd;
1089 : int ctx;
1090 0 : struct macroblock_plane *const p = &x->plane[plane];
1091 0 : struct macroblockd_plane *const pd = &xd->plane[plane];
1092 0 : tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
1093 : uint8_t *dst;
1094 : #if !CONFIG_PVQ
1095 : ENTROPY_CONTEXT *a, *l;
1096 : #endif
1097 : #if CONFIG_VAR_TX
1098 0 : int bw = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
1099 : #endif
1100 0 : dst = &pd->dst
1101 0 : .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];
1102 :
1103 : #if !CONFIG_PVQ
1104 0 : a = &args->ta[blk_col];
1105 0 : l = &args->tl[blk_row];
1106 : #if CONFIG_VAR_TX
1107 0 : ctx = get_entropy_context(tx_size, a, l);
1108 : #else
1109 : ctx = combine_entropy_contexts(*a, *l);
1110 : #endif
1111 : #else
1112 : ctx = 0;
1113 : #endif // CONFIG_PVQ
1114 :
1115 : #if CONFIG_VAR_TX
1116 : // Assert not magic number (uninitialized).
1117 0 : assert(x->blk_skip[plane][blk_row * bw + blk_col] != 234);
1118 :
1119 0 : if (x->blk_skip[plane][blk_row * bw + blk_col] == 0) {
1120 : #else
1121 : {
1122 : #endif
1123 0 : av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
1124 : ctx, AV1_XFORM_QUANT_FP);
1125 : }
1126 : #if CONFIG_VAR_TX
1127 : else {
1128 0 : p->eobs[block] = 0;
1129 : }
1130 : #endif
1131 :
1132 : #if !CONFIG_PVQ
1133 0 : av1_optimize_b(cm, x, plane, block, plane_bsize, tx_size, a, l);
1134 :
1135 0 : av1_set_txb_context(x, plane, block, tx_size, a, l);
1136 :
1137 0 : if (p->eobs[block]) *(args->skip) = 0;
1138 :
1139 0 : if (p->eobs[block] == 0) return;
1140 : #else
1141 : (void)ctx;
1142 : if (!x->pvq_skip[plane]) *(args->skip) = 0;
1143 :
1144 : if (x->pvq_skip[plane]) return;
1145 : #endif
1146 0 : TX_TYPE tx_type = get_tx_type(pd->plane_type, xd, block, tx_size);
1147 0 : av1_inverse_transform_block(xd, dqcoeff, tx_type, tx_size, dst,
1148 0 : pd->dst.stride, p->eobs[block]);
1149 : }
1150 :
1151 : #if CONFIG_VAR_TX
1152 0 : static void encode_block_inter(int plane, int block, int blk_row, int blk_col,
1153 : BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
1154 : void *arg) {
1155 0 : struct encode_b_args *const args = arg;
1156 0 : MACROBLOCK *const x = args->x;
1157 0 : MACROBLOCKD *const xd = &x->e_mbd;
1158 0 : MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1159 0 : const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
1160 0 : const struct macroblockd_plane *const pd = &xd->plane[plane];
1161 0 : const int tx_row = blk_row >> (1 - pd->subsampling_y);
1162 0 : const int tx_col = blk_col >> (1 - pd->subsampling_x);
1163 : TX_SIZE plane_tx_size;
1164 0 : const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
1165 0 : const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
1166 :
1167 0 : if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
1168 :
1169 0 : plane_tx_size =
1170 0 : plane ? uv_txsize_lookup[bsize][mbmi->inter_tx_size[tx_row][tx_col]][0][0]
1171 0 : : mbmi->inter_tx_size[tx_row][tx_col];
1172 :
1173 0 : if (tx_size == plane_tx_size) {
1174 0 : encode_block(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg);
1175 : } else {
1176 0 : assert(tx_size < TX_SIZES_ALL);
1177 0 : const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
1178 0 : assert(sub_txs < tx_size);
1179 : // This is the square transform block partition entry point.
1180 0 : int bsl = tx_size_wide_unit[sub_txs];
1181 : int i;
1182 0 : assert(bsl > 0);
1183 :
1184 0 : for (i = 0; i < 4; ++i) {
1185 0 : const int offsetr = blk_row + ((i >> 1) * bsl);
1186 0 : const int offsetc = blk_col + ((i & 0x01) * bsl);
1187 0 : int step = tx_size_wide_unit[sub_txs] * tx_size_high_unit[sub_txs];
1188 :
1189 0 : if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide) continue;
1190 :
1191 0 : encode_block_inter(plane, block, offsetr, offsetc, plane_bsize, sub_txs,
1192 : arg);
1193 0 : block += step;
1194 : }
1195 : }
1196 : }
1197 : #endif
1198 :
1199 : typedef struct encode_block_pass1_args {
1200 : AV1_COMMON *cm;
1201 : MACROBLOCK *x;
1202 : } encode_block_pass1_args;
1203 :
1204 0 : static void encode_block_pass1(int plane, int block, int blk_row, int blk_col,
1205 : BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
1206 : void *arg) {
1207 0 : encode_block_pass1_args *args = (encode_block_pass1_args *)arg;
1208 0 : AV1_COMMON *cm = args->cm;
1209 0 : MACROBLOCK *const x = args->x;
1210 0 : MACROBLOCKD *const xd = &x->e_mbd;
1211 0 : struct macroblock_plane *const p = &x->plane[plane];
1212 0 : struct macroblockd_plane *const pd = &xd->plane[plane];
1213 0 : tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
1214 : uint8_t *dst;
1215 0 : int ctx = 0;
1216 0 : dst = &pd->dst
1217 0 : .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];
1218 :
1219 0 : av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
1220 : ctx, AV1_XFORM_QUANT_B);
1221 : #if !CONFIG_PVQ
1222 0 : if (p->eobs[block] > 0) {
1223 : #else
1224 : if (!x->pvq_skip[plane]) {
1225 : {
1226 : int tx_blk_size;
1227 : int i, j;
1228 : // transform block size in pixels
1229 : tx_blk_size = tx_size_wide[tx_size];
1230 :
1231 : // Since av1 does not have separate function which does inverse transform
1232 : // but av1_inv_txfm_add_*x*() also does addition of predicted image to
1233 : // inverse transformed image,
1234 : // pass blank dummy image to av1_inv_txfm_add_*x*(), i.e. set dst as zeros
1235 : #if CONFIG_HIGHBITDEPTH
1236 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1237 : for (j = 0; j < tx_blk_size; j++)
1238 : for (i = 0; i < tx_blk_size; i++)
1239 : CONVERT_TO_SHORTPTR(dst)[j * pd->dst.stride + i] = 0;
1240 : } else {
1241 : #endif // CONFIG_HIGHBITDEPTH
1242 : for (j = 0; j < tx_blk_size; j++)
1243 : for (i = 0; i < tx_blk_size; i++) dst[j * pd->dst.stride + i] = 0;
1244 : #if CONFIG_HIGHBITDEPTH
1245 : }
1246 : #endif // CONFIG_HIGHBITDEPTH
1247 : }
1248 : #endif // !CONFIG_PVQ
1249 : #if CONFIG_HIGHBITDEPTH
1250 0 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1251 0 : if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {
1252 0 : av1_highbd_iwht4x4_add(dqcoeff, dst, pd->dst.stride, p->eobs[block],
1253 : xd->bd);
1254 : } else {
1255 0 : av1_highbd_idct4x4_add(dqcoeff, dst, pd->dst.stride, p->eobs[block],
1256 : xd->bd);
1257 : }
1258 0 : return;
1259 : }
1260 : #endif // CONFIG_HIGHBITDEPTH
1261 0 : if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {
1262 0 : av1_iwht4x4_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
1263 : } else {
1264 0 : av1_idct4x4_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
1265 : }
1266 : }
1267 : }
1268 :
1269 0 : void av1_encode_sby_pass1(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize) {
1270 0 : encode_block_pass1_args args = { cm, x };
1271 0 : av1_subtract_plane(x, bsize, 0);
1272 0 : av1_foreach_transformed_block_in_plane(&x->e_mbd, bsize, 0,
1273 : encode_block_pass1, &args);
1274 0 : }
1275 :
1276 0 : void av1_encode_sb(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize, int mi_row,
1277 : int mi_col) {
1278 0 : MACROBLOCKD *const xd = &x->e_mbd;
1279 : struct optimize_ctx ctx;
1280 0 : MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
1281 0 : struct encode_b_args arg = { cm, x, &ctx, &mbmi->skip, NULL, NULL, 1 };
1282 : int plane;
1283 :
1284 0 : mbmi->skip = 1;
1285 :
1286 0 : if (x->skip) return;
1287 :
1288 0 : for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
1289 : #if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
1290 0 : const int subsampling_x = xd->plane[plane].subsampling_x;
1291 0 : const int subsampling_y = xd->plane[plane].subsampling_y;
1292 :
1293 0 : if (!is_chroma_reference(mi_row, mi_col, bsize, subsampling_x,
1294 : subsampling_y))
1295 0 : continue;
1296 :
1297 0 : bsize = scale_chroma_bsize(bsize, subsampling_x, subsampling_y);
1298 : #else
1299 : (void)mi_row;
1300 : (void)mi_col;
1301 : #endif
1302 :
1303 : #if CONFIG_VAR_TX
1304 : // TODO(jingning): Clean this up.
1305 0 : const struct macroblockd_plane *const pd = &xd->plane[plane];
1306 0 : const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
1307 0 : const int mi_width = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
1308 0 : const int mi_height = block_size_high[plane_bsize] >> tx_size_wide_log2[0];
1309 0 : const TX_SIZE max_tx_size = get_vartx_max_txsize(mbmi, plane_bsize);
1310 0 : const BLOCK_SIZE txb_size = txsize_to_bsize[max_tx_size];
1311 0 : const int bw = block_size_wide[txb_size] >> tx_size_wide_log2[0];
1312 0 : const int bh = block_size_high[txb_size] >> tx_size_wide_log2[0];
1313 : int idx, idy;
1314 0 : int block = 0;
1315 0 : int step = tx_size_wide_unit[max_tx_size] * tx_size_high_unit[max_tx_size];
1316 0 : av1_get_entropy_contexts(bsize, 0, pd, ctx.ta[plane], ctx.tl[plane]);
1317 : #else
1318 : const struct macroblockd_plane *const pd = &xd->plane[plane];
1319 : const TX_SIZE tx_size = get_tx_size(plane, xd);
1320 : av1_get_entropy_contexts(bsize, tx_size, pd, ctx.ta[plane], ctx.tl[plane]);
1321 : #endif
1322 :
1323 : #if !CONFIG_PVQ
1324 0 : av1_subtract_plane(x, bsize, plane);
1325 : #endif
1326 0 : arg.ta = ctx.ta[plane];
1327 0 : arg.tl = ctx.tl[plane];
1328 :
1329 : #if CONFIG_VAR_TX
1330 0 : for (idy = 0; idy < mi_height; idy += bh) {
1331 0 : for (idx = 0; idx < mi_width; idx += bw) {
1332 0 : encode_block_inter(plane, block, idy, idx, plane_bsize, max_tx_size,
1333 : &arg);
1334 0 : block += step;
1335 : }
1336 : }
1337 : #else
1338 : av1_foreach_transformed_block_in_plane(xd, bsize, plane, encode_block,
1339 : &arg);
1340 : #endif
1341 : }
1342 : }
1343 :
1344 : #if CONFIG_SUPERTX
1345 : void av1_encode_sb_supertx(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize) {
1346 : MACROBLOCKD *const xd = &x->e_mbd;
1347 : struct optimize_ctx ctx;
1348 : MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
1349 : struct encode_b_args arg = { cm, x, &ctx, &mbmi->skip, NULL, NULL, 1 };
1350 : int plane;
1351 :
1352 : mbmi->skip = 1;
1353 : if (x->skip) return;
1354 :
1355 : for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
1356 : const struct macroblockd_plane *const pd = &xd->plane[plane];
1357 : #if CONFIG_VAR_TX
1358 : const TX_SIZE tx_size = TX_4X4;
1359 : #else
1360 : const TX_SIZE tx_size = get_tx_size(plane, xd);
1361 : #endif
1362 : av1_subtract_plane(x, bsize, plane);
1363 : av1_get_entropy_contexts(bsize, tx_size, pd, ctx.ta[plane], ctx.tl[plane]);
1364 : arg.ta = ctx.ta[plane];
1365 : arg.tl = ctx.tl[plane];
1366 : av1_foreach_transformed_block_in_plane(xd, bsize, plane, encode_block,
1367 : &arg);
1368 : }
1369 : }
1370 : #endif // CONFIG_SUPERTX
1371 :
1372 : #if !CONFIG_PVQ
1373 0 : void av1_set_txb_context(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
1374 : ENTROPY_CONTEXT *a, ENTROPY_CONTEXT *l) {
1375 : (void)tx_size;
1376 0 : struct macroblock_plane *p = &x->plane[plane];
1377 :
1378 : #if !CONFIG_LV_MAP
1379 0 : *a = *l = p->eobs[block] > 0;
1380 : #else // !CONFIG_LV_MAP
1381 : *a = *l = p->txb_entropy_ctx[block];
1382 : #endif // !CONFIG_LV_MAP
1383 :
1384 : #if CONFIG_VAR_TX || CONFIG_LV_MAP
1385 : int i;
1386 0 : for (i = 0; i < tx_size_wide_unit[tx_size]; ++i) a[i] = a[0];
1387 :
1388 0 : for (i = 0; i < tx_size_high_unit[tx_size]; ++i) l[i] = l[0];
1389 : #endif
1390 0 : }
1391 : #endif
1392 :
1393 0 : static void encode_block_intra_and_set_context(int plane, int block,
1394 : int blk_row, int blk_col,
1395 : BLOCK_SIZE plane_bsize,
1396 : TX_SIZE tx_size, void *arg) {
1397 0 : av1_encode_block_intra(plane, block, blk_row, blk_col, plane_bsize, tx_size,
1398 : arg);
1399 : #if !CONFIG_PVQ
1400 0 : struct encode_b_args *const args = arg;
1401 0 : MACROBLOCK *x = args->x;
1402 0 : ENTROPY_CONTEXT *a = &args->ta[blk_col];
1403 0 : ENTROPY_CONTEXT *l = &args->tl[blk_row];
1404 0 : av1_set_txb_context(x, plane, block, tx_size, a, l);
1405 : #endif
1406 0 : }
1407 :
1408 : #if CONFIG_DPCM_INTRA
1409 : static int get_eob(const tran_low_t *qcoeff, intptr_t n_coeffs,
1410 : const int16_t *scan) {
1411 : int eob = -1;
1412 : for (int i = (int)n_coeffs - 1; i >= 0; i--) {
1413 : const int rc = scan[i];
1414 : if (qcoeff[rc]) {
1415 : eob = i;
1416 : break;
1417 : }
1418 : }
1419 : return eob + 1;
1420 : }
1421 :
1422 : static void quantize_scaler(int coeff, int16_t zbin, int16_t round_value,
1423 : int16_t quant, int16_t quant_shift, int16_t dequant,
1424 : int log_scale, tran_low_t *const qcoeff,
1425 : tran_low_t *const dqcoeff) {
1426 : zbin = ROUND_POWER_OF_TWO(zbin, log_scale);
1427 : round_value = ROUND_POWER_OF_TWO(round_value, log_scale);
1428 : const int coeff_sign = (coeff >> 31);
1429 : const int abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
1430 : if (abs_coeff >= zbin) {
1431 : int tmp = clamp(abs_coeff + round_value, INT16_MIN, INT16_MAX);
1432 : tmp = ((((tmp * quant) >> 16) + tmp) * quant_shift) >> (16 - log_scale);
1433 : *qcoeff = (tmp ^ coeff_sign) - coeff_sign;
1434 : *dqcoeff = (*qcoeff * dequant) / (1 << log_scale);
1435 : }
1436 : }
1437 :
1438 : typedef void (*dpcm_fwd_tx_func)(const int16_t *input, int stride,
1439 : TX_TYPE_1D tx_type, tran_low_t *output);
1440 :
1441 : static dpcm_fwd_tx_func get_dpcm_fwd_tx_func(int tx_length) {
1442 : switch (tx_length) {
1443 : case 4: return av1_dpcm_ft4_c;
1444 : case 8: return av1_dpcm_ft8_c;
1445 : case 16: return av1_dpcm_ft16_c;
1446 : case 32:
1447 : return av1_dpcm_ft32_c;
1448 : // TODO(huisu): add support for TX_64X64.
1449 : default: assert(0); return NULL;
1450 : }
1451 : }
1452 :
1453 : static void process_block_dpcm_vert(TX_SIZE tx_size, TX_TYPE_1D tx_type_1d,
1454 : struct macroblockd_plane *const pd,
1455 : struct macroblock_plane *const p,
1456 : uint8_t *src, int src_stride, uint8_t *dst,
1457 : int dst_stride, int16_t *src_diff,
1458 : int diff_stride, tran_low_t *coeff,
1459 : tran_low_t *qcoeff, tran_low_t *dqcoeff) {
1460 : const int tx1d_width = tx_size_wide[tx_size];
1461 : dpcm_fwd_tx_func forward_tx = get_dpcm_fwd_tx_func(tx1d_width);
1462 : dpcm_inv_txfm_add_func inverse_tx =
1463 : av1_get_dpcm_inv_txfm_add_func(tx1d_width);
1464 : const int tx1d_height = tx_size_high[tx_size];
1465 : const int log_scale = av1_get_tx_scale(tx_size);
1466 : int q_idx = 0;
1467 : for (int r = 0; r < tx1d_height; ++r) {
1468 : // Update prediction.
1469 : if (r > 0) memcpy(dst, dst - dst_stride, tx1d_width * sizeof(dst[0]));
1470 : // Subtraction.
1471 : for (int c = 0; c < tx1d_width; ++c) src_diff[c] = src[c] - dst[c];
1472 : // Forward transform.
1473 : forward_tx(src_diff, 1, tx_type_1d, coeff);
1474 : // Quantization.
1475 : for (int c = 0; c < tx1d_width; ++c) {
1476 : quantize_scaler(coeff[c], p->zbin[q_idx], p->round[q_idx],
1477 : p->quant[q_idx], p->quant_shift[q_idx],
1478 : pd->dequant[q_idx], log_scale, &qcoeff[c], &dqcoeff[c]);
1479 : q_idx = 1;
1480 : }
1481 : // Inverse transform.
1482 : inverse_tx(dqcoeff, 1, tx_type_1d, dst);
1483 : // Move to the next row.
1484 : coeff += tx1d_width;
1485 : qcoeff += tx1d_width;
1486 : dqcoeff += tx1d_width;
1487 : src_diff += diff_stride;
1488 : dst += dst_stride;
1489 : src += src_stride;
1490 : }
1491 : }
1492 :
1493 : static void process_block_dpcm_horz(TX_SIZE tx_size, TX_TYPE_1D tx_type_1d,
1494 : struct macroblockd_plane *const pd,
1495 : struct macroblock_plane *const p,
1496 : uint8_t *src, int src_stride, uint8_t *dst,
1497 : int dst_stride, int16_t *src_diff,
1498 : int diff_stride, tran_low_t *coeff,
1499 : tran_low_t *qcoeff, tran_low_t *dqcoeff) {
1500 : const int tx1d_height = tx_size_high[tx_size];
1501 : dpcm_fwd_tx_func forward_tx = get_dpcm_fwd_tx_func(tx1d_height);
1502 : dpcm_inv_txfm_add_func inverse_tx =
1503 : av1_get_dpcm_inv_txfm_add_func(tx1d_height);
1504 : const int tx1d_width = tx_size_wide[tx_size];
1505 : const int log_scale = av1_get_tx_scale(tx_size);
1506 : int q_idx = 0;
1507 : for (int c = 0; c < tx1d_width; ++c) {
1508 : for (int r = 0; r < tx1d_height; ++r) {
1509 : // Update prediction.
1510 : if (c > 0) dst[r * dst_stride] = dst[r * dst_stride - 1];
1511 : // Subtraction.
1512 : src_diff[r * diff_stride] = src[r * src_stride] - dst[r * dst_stride];
1513 : }
1514 : // Forward transform.
1515 : tran_low_t tx_buff[64];
1516 : forward_tx(src_diff, diff_stride, tx_type_1d, tx_buff);
1517 : for (int r = 0; r < tx1d_height; ++r) coeff[r * tx1d_width] = tx_buff[r];
1518 : // Quantization.
1519 : for (int r = 0; r < tx1d_height; ++r) {
1520 : quantize_scaler(coeff[r * tx1d_width], p->zbin[q_idx], p->round[q_idx],
1521 : p->quant[q_idx], p->quant_shift[q_idx],
1522 : pd->dequant[q_idx], log_scale, &qcoeff[r * tx1d_width],
1523 : &dqcoeff[r * tx1d_width]);
1524 : q_idx = 1;
1525 : }
1526 : // Inverse transform.
1527 : for (int r = 0; r < tx1d_height; ++r) tx_buff[r] = dqcoeff[r * tx1d_width];
1528 : inverse_tx(tx_buff, dst_stride, tx_type_1d, dst);
1529 : // Move to the next column.
1530 : ++coeff, ++qcoeff, ++dqcoeff, ++src_diff, ++dst, ++src;
1531 : }
1532 : }
1533 :
1534 : #if CONFIG_HIGHBITDEPTH
1535 : static void hbd_process_block_dpcm_vert(
1536 : TX_SIZE tx_size, TX_TYPE_1D tx_type_1d, int bd,
1537 : struct macroblockd_plane *const pd, struct macroblock_plane *const p,
1538 : uint8_t *src8, int src_stride, uint8_t *dst8, int dst_stride,
1539 : int16_t *src_diff, int diff_stride, tran_low_t *coeff, tran_low_t *qcoeff,
1540 : tran_low_t *dqcoeff) {
1541 : const int tx1d_width = tx_size_wide[tx_size];
1542 : dpcm_fwd_tx_func forward_tx = get_dpcm_fwd_tx_func(tx1d_width);
1543 : hbd_dpcm_inv_txfm_add_func inverse_tx =
1544 : av1_get_hbd_dpcm_inv_txfm_add_func(tx1d_width);
1545 : uint16_t *src = CONVERT_TO_SHORTPTR(src8);
1546 : uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
1547 : const int tx1d_height = tx_size_high[tx_size];
1548 : const int log_scale = av1_get_tx_scale(tx_size);
1549 : int q_idx = 0;
1550 : for (int r = 0; r < tx1d_height; ++r) {
1551 : // Update prediction.
1552 : if (r > 0) memcpy(dst, dst - dst_stride, tx1d_width * sizeof(dst[0]));
1553 : // Subtraction.
1554 : for (int c = 0; c < tx1d_width; ++c) src_diff[c] = src[c] - dst[c];
1555 : // Forward transform.
1556 : forward_tx(src_diff, 1, tx_type_1d, coeff);
1557 : // Quantization.
1558 : for (int c = 0; c < tx1d_width; ++c) {
1559 : quantize_scaler(coeff[c], p->zbin[q_idx], p->round[q_idx],
1560 : p->quant[q_idx], p->quant_shift[q_idx],
1561 : pd->dequant[q_idx], log_scale, &qcoeff[c], &dqcoeff[c]);
1562 : q_idx = 1;
1563 : }
1564 : // Inverse transform.
1565 : inverse_tx(dqcoeff, 1, tx_type_1d, bd, dst);
1566 : // Move to the next row.
1567 : coeff += tx1d_width;
1568 : qcoeff += tx1d_width;
1569 : dqcoeff += tx1d_width;
1570 : src_diff += diff_stride;
1571 : dst += dst_stride;
1572 : src += src_stride;
1573 : }
1574 : }
1575 :
1576 : static void hbd_process_block_dpcm_horz(
1577 : TX_SIZE tx_size, TX_TYPE_1D tx_type_1d, int bd,
1578 : struct macroblockd_plane *const pd, struct macroblock_plane *const p,
1579 : uint8_t *src8, int src_stride, uint8_t *dst8, int dst_stride,
1580 : int16_t *src_diff, int diff_stride, tran_low_t *coeff, tran_low_t *qcoeff,
1581 : tran_low_t *dqcoeff) {
1582 : const int tx1d_height = tx_size_high[tx_size];
1583 : dpcm_fwd_tx_func forward_tx = get_dpcm_fwd_tx_func(tx1d_height);
1584 : hbd_dpcm_inv_txfm_add_func inverse_tx =
1585 : av1_get_hbd_dpcm_inv_txfm_add_func(tx1d_height);
1586 : uint16_t *src = CONVERT_TO_SHORTPTR(src8);
1587 : uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
1588 : const int tx1d_width = tx_size_wide[tx_size];
1589 : const int log_scale = av1_get_tx_scale(tx_size);
1590 : int q_idx = 0;
1591 : for (int c = 0; c < tx1d_width; ++c) {
1592 : for (int r = 0; r < tx1d_height; ++r) {
1593 : // Update prediction.
1594 : if (c > 0) dst[r * dst_stride] = dst[r * dst_stride - 1];
1595 : // Subtraction.
1596 : src_diff[r * diff_stride] = src[r * src_stride] - dst[r * dst_stride];
1597 : }
1598 : // Forward transform.
1599 : tran_low_t tx_buff[64];
1600 : forward_tx(src_diff, diff_stride, tx_type_1d, tx_buff);
1601 : for (int r = 0; r < tx1d_height; ++r) coeff[r * tx1d_width] = tx_buff[r];
1602 : // Quantization.
1603 : for (int r = 0; r < tx1d_height; ++r) {
1604 : quantize_scaler(coeff[r * tx1d_width], p->zbin[q_idx], p->round[q_idx],
1605 : p->quant[q_idx], p->quant_shift[q_idx],
1606 : pd->dequant[q_idx], log_scale, &qcoeff[r * tx1d_width],
1607 : &dqcoeff[r * tx1d_width]);
1608 : q_idx = 1;
1609 : }
1610 : // Inverse transform.
1611 : for (int r = 0; r < tx1d_height; ++r) tx_buff[r] = dqcoeff[r * tx1d_width];
1612 : inverse_tx(tx_buff, dst_stride, tx_type_1d, bd, dst);
1613 : // Move to the next column.
1614 : ++coeff, ++qcoeff, ++dqcoeff, ++src_diff, ++dst, ++src;
1615 : }
1616 : }
1617 : #endif // CONFIG_HIGHBITDEPTH
1618 :
1619 : void av1_encode_block_intra_dpcm(const AV1_COMMON *cm, MACROBLOCK *x,
1620 : PREDICTION_MODE mode, int plane, int block,
1621 : int blk_row, int blk_col,
1622 : BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
1623 : TX_TYPE tx_type, ENTROPY_CONTEXT *ta,
1624 : ENTROPY_CONTEXT *tl, int8_t *skip) {
1625 : MACROBLOCKD *const xd = &x->e_mbd;
1626 : struct macroblock_plane *const p = &x->plane[plane];
1627 : struct macroblockd_plane *const pd = &xd->plane[plane];
1628 : tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
1629 : const int diff_stride = block_size_wide[plane_bsize];
1630 : const int src_stride = p->src.stride;
1631 : const int dst_stride = pd->dst.stride;
1632 : const int tx1d_width = tx_size_wide[tx_size];
1633 : const int tx1d_height = tx_size_high[tx_size];
1634 : const SCAN_ORDER *const scan_order = get_scan(cm, tx_size, tx_type, 0);
1635 : tran_low_t *coeff = BLOCK_OFFSET(p->coeff, block);
1636 : tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
1637 : uint8_t *dst =
1638 : &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
1639 : uint8_t *src =
1640 : &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]];
1641 : int16_t *src_diff =
1642 : &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
1643 : uint16_t *eob = &p->eobs[block];
1644 : *eob = 0;
1645 : memset(qcoeff, 0, tx1d_height * tx1d_width * sizeof(*qcoeff));
1646 : memset(dqcoeff, 0, tx1d_height * tx1d_width * sizeof(*dqcoeff));
1647 :
1648 : if (LIKELY(!x->skip_block)) {
1649 : TX_TYPE_1D tx_type_1d = DCT_1D;
1650 : switch (tx_type) {
1651 : case IDTX: tx_type_1d = IDTX_1D; break;
1652 : case V_DCT:
1653 : assert(mode == H_PRED);
1654 : tx_type_1d = DCT_1D;
1655 : break;
1656 : case H_DCT:
1657 : assert(mode == V_PRED);
1658 : tx_type_1d = DCT_1D;
1659 : break;
1660 : default: assert(0);
1661 : }
1662 : switch (mode) {
1663 : case V_PRED:
1664 : #if CONFIG_HIGHBITDEPTH
1665 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1666 : hbd_process_block_dpcm_vert(tx_size, tx_type_1d, xd->bd, pd, p, src,
1667 : src_stride, dst, dst_stride, src_diff,
1668 : diff_stride, coeff, qcoeff, dqcoeff);
1669 : } else {
1670 : #endif // CONFIG_HIGHBITDEPTH
1671 : process_block_dpcm_vert(tx_size, tx_type_1d, pd, p, src, src_stride,
1672 : dst, dst_stride, src_diff, diff_stride, coeff,
1673 : qcoeff, dqcoeff);
1674 : #if CONFIG_HIGHBITDEPTH
1675 : }
1676 : #endif // CONFIG_HIGHBITDEPTH
1677 : break;
1678 : case H_PRED:
1679 : #if CONFIG_HIGHBITDEPTH
1680 : if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1681 : hbd_process_block_dpcm_horz(tx_size, tx_type_1d, xd->bd, pd, p, src,
1682 : src_stride, dst, dst_stride, src_diff,
1683 : diff_stride, coeff, qcoeff, dqcoeff);
1684 : } else {
1685 : #endif // CONFIG_HIGHBITDEPTH
1686 : process_block_dpcm_horz(tx_size, tx_type_1d, pd, p, src, src_stride,
1687 : dst, dst_stride, src_diff, diff_stride, coeff,
1688 : qcoeff, dqcoeff);
1689 : #if CONFIG_HIGHBITDEPTH
1690 : }
1691 : #endif // CONFIG_HIGHBITDEPTH
1692 : break;
1693 : default: assert(0);
1694 : }
1695 : *eob = get_eob(qcoeff, tx1d_height * tx1d_width, scan_order->scan);
1696 : }
1697 :
1698 : ta[blk_col] = tl[blk_row] = *eob > 0;
1699 : if (*eob) *skip = 0;
1700 : }
1701 : #endif // CONFIG_DPCM_INTRA
1702 :
1703 0 : void av1_encode_block_intra(int plane, int block, int blk_row, int blk_col,
1704 : BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
1705 : void *arg) {
1706 0 : struct encode_b_args *const args = arg;
1707 0 : AV1_COMMON *cm = args->cm;
1708 0 : MACROBLOCK *const x = args->x;
1709 0 : MACROBLOCKD *const xd = &x->e_mbd;
1710 0 : struct macroblock_plane *const p = &x->plane[plane];
1711 0 : struct macroblockd_plane *const pd = &xd->plane[plane];
1712 0 : tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
1713 0 : PLANE_TYPE plane_type = get_plane_type(plane);
1714 0 : const TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
1715 0 : uint16_t *eob = &p->eobs[block];
1716 0 : const int dst_stride = pd->dst.stride;
1717 0 : uint8_t *dst =
1718 0 : &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
1719 : #if CONFIG_CFL
1720 :
1721 : #if CONFIG_EC_ADAPT
1722 : FRAME_CONTEXT *const ec_ctx = xd->tile_ctx;
1723 : #else
1724 : FRAME_CONTEXT *const ec_ctx = cm->fc;
1725 : #endif // CONFIG_EC_ADAPT
1726 :
1727 : av1_predict_intra_block_encoder_facade(x, ec_ctx, plane, block, blk_col,
1728 : blk_row, tx_size, plane_bsize);
1729 : #else
1730 0 : av1_predict_intra_block_facade(xd, plane, block, blk_col, blk_row, tx_size);
1731 : #endif
1732 :
1733 : #if CONFIG_DPCM_INTRA
1734 : const int block_raster_idx = av1_block_index_to_raster_order(tx_size, block);
1735 : const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1736 : const PREDICTION_MODE mode =
1737 : (plane == 0) ? get_y_mode(xd->mi[0], block_raster_idx) : mbmi->uv_mode;
1738 : if (av1_use_dpcm_intra(plane, mode, tx_type, mbmi)) {
1739 : av1_encode_block_intra_dpcm(cm, x, mode, plane, block, blk_row, blk_col,
1740 : plane_bsize, tx_size, tx_type, args->ta,
1741 : args->tl, args->skip);
1742 : return;
1743 : }
1744 : #endif // CONFIG_DPCM_INTRA
1745 :
1746 0 : av1_subtract_txb(x, plane, plane_bsize, blk_col, blk_row, tx_size);
1747 :
1748 0 : const ENTROPY_CONTEXT *a = &args->ta[blk_col];
1749 0 : const ENTROPY_CONTEXT *l = &args->tl[blk_row];
1750 0 : int ctx = combine_entropy_contexts(*a, *l);
1751 0 : if (args->enable_optimize_b) {
1752 0 : av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
1753 : ctx, AV1_XFORM_QUANT_FP);
1754 0 : av1_optimize_b(cm, x, plane, block, plane_bsize, tx_size, a, l);
1755 : } else {
1756 0 : av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
1757 : ctx, AV1_XFORM_QUANT_B);
1758 : }
1759 :
1760 : #if CONFIG_PVQ
1761 : // *(args->skip) == mbmi->skip
1762 : if (!x->pvq_skip[plane]) *(args->skip) = 0;
1763 :
1764 : if (x->pvq_skip[plane]) return;
1765 : #endif // CONFIG_PVQ
1766 0 : av1_inverse_transform_block(xd, dqcoeff, tx_type, tx_size, dst, dst_stride,
1767 0 : *eob);
1768 : #if !CONFIG_PVQ
1769 0 : if (*eob) *(args->skip) = 0;
1770 : #else
1771 : // Note : *(args->skip) == mbmi->skip
1772 : #endif
1773 : #if CONFIG_CFL
1774 : MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
1775 : if (plane == AOM_PLANE_Y && x->cfl_store_y) {
1776 : cfl_store(xd->cfl, dst, dst_stride, blk_row, blk_col, tx_size);
1777 : }
1778 :
1779 : if (mbmi->uv_mode == DC_PRED) {
1780 : // TODO(ltrudeau) find a cleaner way to detect last transform block
1781 : if (plane == AOM_PLANE_U) {
1782 : xd->cfl->num_tx_blk[CFL_PRED_U] =
1783 : (blk_row == 0 && blk_col == 0) ? 1
1784 : : xd->cfl->num_tx_blk[CFL_PRED_U] + 1;
1785 : }
1786 :
1787 : if (plane == AOM_PLANE_V) {
1788 : xd->cfl->num_tx_blk[CFL_PRED_V] =
1789 : (blk_row == 0 && blk_col == 0) ? 1
1790 : : xd->cfl->num_tx_blk[CFL_PRED_V] + 1;
1791 :
1792 : if (mbmi->skip &&
1793 : xd->cfl->num_tx_blk[CFL_PRED_U] == xd->cfl->num_tx_blk[CFL_PRED_V]) {
1794 : assert(plane_bsize != BLOCK_INVALID);
1795 : const int block_width = block_size_wide[plane_bsize];
1796 : const int block_height = block_size_high[plane_bsize];
1797 :
1798 : // if SKIP is chosen at the block level, and ind != 0, we must change
1799 : // the prediction
1800 : if (mbmi->cfl_alpha_idx != 0) {
1801 : const struct macroblockd_plane *const pd_cb = &xd->plane[AOM_PLANE_U];
1802 : uint8_t *const dst_cb = pd_cb->dst.buf;
1803 : const int dst_stride_cb = pd_cb->dst.stride;
1804 : uint8_t *const dst_cr = pd->dst.buf;
1805 : const int dst_stride_cr = pd->dst.stride;
1806 : for (int j = 0; j < block_height; j++) {
1807 : for (int i = 0; i < block_width; i++) {
1808 : dst_cb[dst_stride_cb * j + i] =
1809 : (uint8_t)(xd->cfl->dc_pred[CFL_PRED_U] + 0.5);
1810 : dst_cr[dst_stride_cr * j + i] =
1811 : (uint8_t)(xd->cfl->dc_pred[CFL_PRED_V] + 0.5);
1812 : }
1813 : }
1814 : mbmi->cfl_alpha_idx = 0;
1815 : mbmi->cfl_alpha_signs[CFL_PRED_U] = CFL_SIGN_POS;
1816 : mbmi->cfl_alpha_signs[CFL_PRED_V] = CFL_SIGN_POS;
1817 : }
1818 : }
1819 : }
1820 : }
1821 : #endif
1822 0 : }
1823 :
1824 : #if CONFIG_CFL
1825 : static int cfl_alpha_dist(const uint8_t *y_pix, int y_stride, double y_avg,
1826 : const uint8_t *src, int src_stride, int blk_width,
1827 : int blk_height, double dc_pred, double alpha,
1828 : int *dist_neg_out) {
1829 : const double dc_pred_bias = dc_pred + 0.5;
1830 : int dist = 0;
1831 : int diff;
1832 :
1833 : if (alpha == 0.0) {
1834 : const int dc_pred_i = (int)dc_pred_bias;
1835 : for (int j = 0; j < blk_height; j++) {
1836 : for (int i = 0; i < blk_width; i++) {
1837 : diff = src[i] - dc_pred_i;
1838 : dist += diff * diff;
1839 : }
1840 : src += src_stride;
1841 : }
1842 :
1843 : if (dist_neg_out) *dist_neg_out = dist;
1844 :
1845 : return dist;
1846 : }
1847 :
1848 : int dist_neg = 0;
1849 : for (int j = 0; j < blk_height; j++) {
1850 : for (int i = 0; i < blk_width; i++) {
1851 : const double scaled_luma = alpha * (y_pix[i] - y_avg);
1852 : const int uv = src[i];
1853 : diff = uv - (int)(scaled_luma + dc_pred_bias);
1854 : dist += diff * diff;
1855 : diff = uv + (int)(scaled_luma - dc_pred_bias);
1856 : dist_neg += diff * diff;
1857 : }
1858 : y_pix += y_stride;
1859 : src += src_stride;
1860 : }
1861 :
1862 : if (dist_neg_out) *dist_neg_out = dist_neg;
1863 :
1864 : return dist;
1865 : }
1866 :
1867 : static int cfl_compute_alpha_ind(MACROBLOCK *const x, const CFL_CTX *const cfl,
1868 : BLOCK_SIZE bsize,
1869 : CFL_SIGN_TYPE signs_out[CFL_SIGNS]) {
1870 : const struct macroblock_plane *const p_u = &x->plane[AOM_PLANE_U];
1871 : const struct macroblock_plane *const p_v = &x->plane[AOM_PLANE_V];
1872 : const uint8_t *const src_u = p_u->src.buf;
1873 : const uint8_t *const src_v = p_v->src.buf;
1874 : const int src_stride_u = p_u->src.stride;
1875 : const int src_stride_v = p_v->src.stride;
1876 : const int block_width = block_size_wide[bsize];
1877 : const int block_height = block_size_high[bsize];
1878 : const double dc_pred_u = cfl->dc_pred[CFL_PRED_U];
1879 : const double dc_pred_v = cfl->dc_pred[CFL_PRED_V];
1880 :
1881 : // Temporary pixel buffer used to store the CfL prediction when we compute the
1882 : // alpha index.
1883 : uint8_t tmp_pix[MAX_SB_SQUARE];
1884 : // Load CfL Prediction over the entire block
1885 : const double y_avg =
1886 : cfl_load(cfl, tmp_pix, MAX_SB_SIZE, 0, 0, block_width, block_height);
1887 :
1888 : int sse[CFL_PRED_PLANES][CFL_MAGS_SIZE];
1889 : sse[CFL_PRED_U][0] =
1890 : cfl_alpha_dist(tmp_pix, MAX_SB_SIZE, y_avg, src_u, src_stride_u,
1891 : block_width, block_height, dc_pred_u, 0, NULL);
1892 : sse[CFL_PRED_V][0] =
1893 : cfl_alpha_dist(tmp_pix, MAX_SB_SIZE, y_avg, src_v, src_stride_v,
1894 : block_width, block_height, dc_pred_v, 0, NULL);
1895 : for (int m = 1; m < CFL_MAGS_SIZE; m += 2) {
1896 : assert(cfl_alpha_mags[m + 1] == -cfl_alpha_mags[m]);
1897 : sse[CFL_PRED_U][m] = cfl_alpha_dist(
1898 : tmp_pix, MAX_SB_SIZE, y_avg, src_u, src_stride_u, block_width,
1899 : block_height, dc_pred_u, cfl_alpha_mags[m], &sse[CFL_PRED_U][m + 1]);
1900 : sse[CFL_PRED_V][m] = cfl_alpha_dist(
1901 : tmp_pix, MAX_SB_SIZE, y_avg, src_v, src_stride_v, block_width,
1902 : block_height, dc_pred_v, cfl_alpha_mags[m], &sse[CFL_PRED_V][m + 1]);
1903 : }
1904 :
1905 : int dist;
1906 : int64_t cost;
1907 : int64_t best_cost;
1908 :
1909 : // Compute least squares parameter of the entire block
1910 : // IMPORTANT: We assume that the first code is 0,0
1911 : int ind = 0;
1912 : signs_out[CFL_PRED_U] = CFL_SIGN_POS;
1913 : signs_out[CFL_PRED_V] = CFL_SIGN_POS;
1914 :
1915 : dist = sse[CFL_PRED_U][0] + sse[CFL_PRED_V][0];
1916 : dist *= 16;
1917 : best_cost = RDCOST(x->rdmult, x->rddiv, cfl->costs[0], dist);
1918 :
1919 : for (int c = 1; c < CFL_ALPHABET_SIZE; c++) {
1920 : const int idx_u = cfl_alpha_codes[c][CFL_PRED_U];
1921 : const int idx_v = cfl_alpha_codes[c][CFL_PRED_V];
1922 : for (CFL_SIGN_TYPE sign_u = idx_u == 0; sign_u < CFL_SIGNS; sign_u++) {
1923 : for (CFL_SIGN_TYPE sign_v = idx_v == 0; sign_v < CFL_SIGNS; sign_v++) {
1924 : dist = sse[CFL_PRED_U][idx_u + (sign_u == CFL_SIGN_NEG)] +
1925 : sse[CFL_PRED_V][idx_v + (sign_v == CFL_SIGN_NEG)];
1926 : dist *= 16;
1927 : cost = RDCOST(x->rdmult, x->rddiv, cfl->costs[c], dist);
1928 : if (cost < best_cost) {
1929 : best_cost = cost;
1930 : ind = c;
1931 : signs_out[CFL_PRED_U] = sign_u;
1932 : signs_out[CFL_PRED_V] = sign_v;
1933 : }
1934 : }
1935 : }
1936 : }
1937 :
1938 : return ind;
1939 : }
1940 :
1941 : static inline void cfl_update_costs(CFL_CTX *cfl, FRAME_CONTEXT *ec_ctx) {
1942 : assert(ec_ctx->cfl_alpha_cdf[CFL_ALPHABET_SIZE - 1] ==
1943 : AOM_ICDF(CDF_PROB_TOP));
1944 : const int prob_den = CDF_PROB_TOP;
1945 :
1946 : int prob_num = AOM_ICDF(ec_ctx->cfl_alpha_cdf[0]);
1947 : cfl->costs[0] = av1_cost_zero(get_prob(prob_num, prob_den));
1948 :
1949 : for (int c = 1; c < CFL_ALPHABET_SIZE; c++) {
1950 : int sign_bit_cost = (cfl_alpha_codes[c][CFL_PRED_U] != 0) +
1951 : (cfl_alpha_codes[c][CFL_PRED_V] != 0);
1952 : prob_num = AOM_ICDF(ec_ctx->cfl_alpha_cdf[c]) -
1953 : AOM_ICDF(ec_ctx->cfl_alpha_cdf[c - 1]);
1954 : cfl->costs[c] = av1_cost_zero(get_prob(prob_num, prob_den)) +
1955 : av1_cost_literal(sign_bit_cost);
1956 : }
1957 : }
1958 :
1959 : void av1_predict_intra_block_encoder_facade(MACROBLOCK *x,
1960 : FRAME_CONTEXT *ec_ctx, int plane,
1961 : int block_idx, int blk_col,
1962 : int blk_row, TX_SIZE tx_size,
1963 : BLOCK_SIZE plane_bsize) {
1964 : MACROBLOCKD *const xd = &x->e_mbd;
1965 : MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
1966 : if (plane != AOM_PLANE_Y && mbmi->uv_mode == DC_PRED) {
1967 : if (blk_col == 0 && blk_row == 0 && plane == AOM_PLANE_U) {
1968 : CFL_CTX *const cfl = xd->cfl;
1969 : cfl_update_costs(cfl, ec_ctx);
1970 : cfl_dc_pred(xd, plane_bsize, tx_size);
1971 : mbmi->cfl_alpha_idx =
1972 : cfl_compute_alpha_ind(x, cfl, plane_bsize, mbmi->cfl_alpha_signs);
1973 : }
1974 : }
1975 : av1_predict_intra_block_facade(xd, plane, block_idx, blk_col, blk_row,
1976 : tx_size);
1977 : }
1978 : #endif
1979 :
1980 0 : void av1_encode_intra_block_plane(AV1_COMMON *cm, MACROBLOCK *x,
1981 : BLOCK_SIZE bsize, int plane,
1982 : int enable_optimize_b, int mi_row,
1983 : int mi_col) {
1984 0 : const MACROBLOCKD *const xd = &x->e_mbd;
1985 0 : ENTROPY_CONTEXT ta[2 * MAX_MIB_SIZE] = { 0 };
1986 0 : ENTROPY_CONTEXT tl[2 * MAX_MIB_SIZE] = { 0 };
1987 :
1988 0 : struct encode_b_args arg = {
1989 0 : cm, x, NULL, &xd->mi[0]->mbmi.skip, ta, tl, enable_optimize_b
1990 : };
1991 :
1992 : #if CONFIG_CB4X4
1993 0 : if (!is_chroma_reference(mi_row, mi_col, bsize,
1994 : xd->plane[plane].subsampling_x,
1995 : xd->plane[plane].subsampling_y))
1996 0 : return;
1997 : #else
1998 : (void)mi_row;
1999 : (void)mi_col;
2000 : #endif
2001 :
2002 0 : if (enable_optimize_b) {
2003 0 : const struct macroblockd_plane *const pd = &xd->plane[plane];
2004 0 : const TX_SIZE tx_size = get_tx_size(plane, xd);
2005 0 : av1_get_entropy_contexts(bsize, tx_size, pd, ta, tl);
2006 : }
2007 0 : av1_foreach_transformed_block_in_plane(
2008 : xd, bsize, plane, encode_block_intra_and_set_context, &arg);
2009 : }
2010 :
2011 : #if CONFIG_PVQ
2012 : PVQ_SKIP_TYPE av1_pvq_encode_helper(MACROBLOCK *x, tran_low_t *const coeff,
2013 : tran_low_t *ref_coeff,
2014 : tran_low_t *const dqcoeff, uint16_t *eob,
2015 : const int16_t *quant, int plane,
2016 : int tx_size, TX_TYPE tx_type, int *rate,
2017 : int speed, PVQ_INFO *pvq_info) {
2018 : const int tx_blk_size = tx_size_wide[tx_size];
2019 : daala_enc_ctx *daala_enc = &x->daala_enc;
2020 : PVQ_SKIP_TYPE ac_dc_coded;
2021 : int coeff_shift = 3 - av1_get_tx_scale(tx_size);
2022 : int hbd_downshift = 0;
2023 : int rounding_mask;
2024 : int pvq_dc_quant;
2025 : int use_activity_masking = daala_enc->use_activity_masking;
2026 : int tell;
2027 : int has_dc_skip = 1;
2028 : int i;
2029 : int off = od_qm_offset(tx_size, plane ? 1 : 0);
2030 :
2031 : DECLARE_ALIGNED(16, tran_low_t, coeff_pvq[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
2032 : DECLARE_ALIGNED(16, tran_low_t, ref_coeff_pvq[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
2033 : DECLARE_ALIGNED(16, tran_low_t, dqcoeff_pvq[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
2034 :
2035 : DECLARE_ALIGNED(16, int32_t, in_int32[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
2036 : DECLARE_ALIGNED(16, int32_t, ref_int32[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
2037 : DECLARE_ALIGNED(16, int32_t, out_int32[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
2038 :
2039 : hbd_downshift = x->e_mbd.bd - 8;
2040 :
2041 : assert(OD_COEFF_SHIFT >= 4);
2042 : // DC quantizer for PVQ
2043 : if (use_activity_masking)
2044 : pvq_dc_quant =
2045 : OD_MAXI(1, (quant[0] << (OD_COEFF_SHIFT - 3) >> hbd_downshift) *
2046 : daala_enc->state
2047 : .pvq_qm_q4[plane][od_qm_get_index(tx_size, 0)] >>
2048 : 4);
2049 : else
2050 : pvq_dc_quant =
2051 : OD_MAXI(1, quant[0] << (OD_COEFF_SHIFT - 3) >> hbd_downshift);
2052 :
2053 : *eob = 0;
2054 :
2055 : #if !CONFIG_ANS
2056 : tell = od_ec_enc_tell_frac(&daala_enc->w.ec);
2057 : #else
2058 : #error "CONFIG_PVQ currently requires !CONFIG_ANS."
2059 : #endif
2060 :
2061 : // Change coefficient ordering for pvq encoding.
2062 : od_raster_to_coding_order(coeff_pvq, tx_blk_size, tx_type, coeff,
2063 : tx_blk_size);
2064 : od_raster_to_coding_order(ref_coeff_pvq, tx_blk_size, tx_type, ref_coeff,
2065 : tx_blk_size);
2066 :
2067 : // copy int16 inputs to int32
2068 : for (i = 0; i < tx_blk_size * tx_blk_size; i++) {
2069 : ref_int32[i] =
2070 : AOM_SIGNED_SHL(ref_coeff_pvq[i], OD_COEFF_SHIFT - coeff_shift) >>
2071 : hbd_downshift;
2072 : in_int32[i] = AOM_SIGNED_SHL(coeff_pvq[i], OD_COEFF_SHIFT - coeff_shift) >>
2073 : hbd_downshift;
2074 : }
2075 :
2076 : if (abs(in_int32[0] - ref_int32[0]) < pvq_dc_quant * 141 / 256) { /* 0.55 */
2077 : out_int32[0] = 0;
2078 : } else {
2079 : out_int32[0] = OD_DIV_R0(in_int32[0] - ref_int32[0], pvq_dc_quant);
2080 : }
2081 :
2082 : ac_dc_coded =
2083 : od_pvq_encode(daala_enc, ref_int32, in_int32, out_int32,
2084 : OD_MAXI(1, quant[0] << (OD_COEFF_SHIFT - 3) >>
2085 : hbd_downshift), // scale/quantizer
2086 : OD_MAXI(1, quant[1] << (OD_COEFF_SHIFT - 3) >>
2087 : hbd_downshift), // scale/quantizer
2088 : plane,
2089 : tx_size, OD_PVQ_BETA[use_activity_masking][plane][tx_size],
2090 : 0, // is_keyframe,
2091 : daala_enc->state.qm + off, daala_enc->state.qm_inv + off,
2092 : speed, // speed
2093 : pvq_info);
2094 :
2095 : // Encode residue of DC coeff, if required.
2096 : if (!has_dc_skip || out_int32[0]) {
2097 : generic_encode(&daala_enc->w, &daala_enc->state.adapt->model_dc[plane],
2098 : abs(out_int32[0]) - has_dc_skip,
2099 : &daala_enc->state.adapt->ex_dc[plane][tx_size][0], 2);
2100 : }
2101 : if (out_int32[0]) {
2102 : aom_write_bit(&daala_enc->w, out_int32[0] < 0);
2103 : }
2104 :
2105 : // need to save quantized residue of DC coeff
2106 : // so that final pvq bitstream writing can know whether DC is coded.
2107 : if (pvq_info) pvq_info->dq_dc_residue = out_int32[0];
2108 :
2109 : out_int32[0] = out_int32[0] * pvq_dc_quant;
2110 : out_int32[0] += ref_int32[0];
2111 :
2112 : // copy int32 result back to int16
2113 : assert(OD_COEFF_SHIFT > coeff_shift);
2114 : rounding_mask = (1 << (OD_COEFF_SHIFT - coeff_shift - 1)) - 1;
2115 : for (i = 0; i < tx_blk_size * tx_blk_size; i++) {
2116 : out_int32[i] = AOM_SIGNED_SHL(out_int32[i], hbd_downshift);
2117 : dqcoeff_pvq[i] = (out_int32[i] + (out_int32[i] < 0) + rounding_mask) >>
2118 : (OD_COEFF_SHIFT - coeff_shift);
2119 : }
2120 :
2121 : // Back to original coefficient order
2122 : od_coding_order_to_raster(dqcoeff, tx_blk_size, tx_type, dqcoeff_pvq,
2123 : tx_blk_size);
2124 :
2125 : *eob = tx_blk_size * tx_blk_size;
2126 :
2127 : #if !CONFIG_ANS
2128 : *rate = (od_ec_enc_tell_frac(&daala_enc->w.ec) - tell)
2129 : << (AV1_PROB_COST_SHIFT - OD_BITRES);
2130 : #else
2131 : #error "CONFIG_PVQ currently requires !CONFIG_ANS."
2132 : #endif
2133 : assert(*rate >= 0);
2134 :
2135 : return ac_dc_coded;
2136 : }
2137 :
2138 : void av1_store_pvq_enc_info(PVQ_INFO *pvq_info, int *qg, int *theta, int *k,
2139 : od_coeff *y, int nb_bands, const int *off,
2140 : int *size, int skip_rest, int skip_dir,
2141 : int bs) { // block size in log_2 -2
2142 : int i;
2143 : const int tx_blk_size = tx_size_wide[bs];
2144 :
2145 : for (i = 0; i < nb_bands; i++) {
2146 : pvq_info->qg[i] = qg[i];
2147 : pvq_info->theta[i] = theta[i];
2148 : pvq_info->k[i] = k[i];
2149 : pvq_info->off[i] = off[i];
2150 : pvq_info->size[i] = size[i];
2151 : }
2152 :
2153 : memcpy(pvq_info->y, y, tx_blk_size * tx_blk_size * sizeof(od_coeff));
2154 :
2155 : pvq_info->nb_bands = nb_bands;
2156 : pvq_info->skip_rest = skip_rest;
2157 : pvq_info->skip_dir = skip_dir;
2158 : pvq_info->bs = bs;
2159 : }
2160 : #endif
|
