1 /*
2 * Copyright (c) 2017, 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/encoder/encodetxb.h"
13
14 #include "aom_ports/mem.h"
15 #include "av1/common/blockd.h"
16 #include "av1/common/idct.h"
17 #include "av1/common/pred_common.h"
18 #include "av1/common/scan.h"
19 #include "av1/encoder/bitstream.h"
20 #include "av1/encoder/cost.h"
21 #include "av1/encoder/encodeframe.h"
22 #include "av1/encoder/hash.h"
23 #include "av1/encoder/rdopt.h"
24 #include "av1/encoder/tokenize.h"
25
26 #if CONFIG_HTB_TRELLIS
27 static int hbt_needs_init = 1;
28 static CRC32C crc_calculator;
29 static const int HBT_EOB = 16; // also the length in opt_qcoeff
30 static const int HBT_TABLE_SIZE = 65536; // 16 bit: holds 65536 'arrays'
31 static const int HBT_ARRAY_LENGTH = 256; // 8 bit: 256 entries
32 // If removed in hbt_create_hashes or increased beyond int8_t, widen deltas type
33 static const int HBT_KICKOUT = 3;
34
35 typedef struct OptTxbQcoeff {
36 // Use larger type if larger/no kickout value is used in hbt_create_hashes
37 int8_t deltas[16];
38 uint32_t hbt_qc_hash;
39 uint32_t hbt_ctx_hash;
40 int init;
41 int rate_cost;
42 } OptTxbQcoeff;
43
44 OptTxbQcoeff *hbt_hash_table;
45 #endif // CONFIG_HTB_TRELLIS
46
47 typedef struct LevelDownStats {
48 int update;
49 tran_low_t low_qc;
50 tran_low_t low_dqc;
51 int64_t dist0;
52 int rate;
53 int rate_low;
54 int64_t dist;
55 int64_t dist_low;
56 int64_t rd;
57 int64_t rd_low;
58 int64_t nz_rd;
59 int64_t rd_diff;
60 int cost_diff;
61 int64_t dist_diff;
62 int new_eob;
63 } LevelDownStats;
64
get_dqv(const int16_t * dequant,int coeff_idx,const qm_val_t * iqmatrix)65 static INLINE int get_dqv(const int16_t *dequant, int coeff_idx,
66 const qm_val_t *iqmatrix) {
67 int dqv = dequant[!!coeff_idx];
68 if (iqmatrix != NULL)
69 dqv =
70 ((iqmatrix[coeff_idx] * dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
71 return dqv;
72 }
73
av1_alloc_txb_buf(AV1_COMP * cpi)74 void av1_alloc_txb_buf(AV1_COMP *cpi) {
75 AV1_COMMON *cm = &cpi->common;
76 int size = ((cm->mi_params.mi_rows >> cm->seq_params.mib_size_log2) + 1) *
77 ((cm->mi_params.mi_cols >> cm->seq_params.mib_size_log2) + 1);
78
79 av1_free_txb_buf(cpi);
80 // TODO(jingning): This should be further reduced.
81 CHECK_MEM_ERROR(cm, cpi->coeff_buffer_base,
82 aom_memalign(32, sizeof(*cpi->coeff_buffer_base) * size));
83 }
84
av1_free_txb_buf(AV1_COMP * cpi)85 void av1_free_txb_buf(AV1_COMP *cpi) { aom_free(cpi->coeff_buffer_base); }
86
write_golomb(aom_writer * w,int level)87 static void write_golomb(aom_writer *w, int level) {
88 int x = level + 1;
89 int i = x;
90 int length = 0;
91
92 while (i) {
93 i >>= 1;
94 ++length;
95 }
96 assert(length > 0);
97
98 for (i = 0; i < length - 1; ++i) aom_write_bit(w, 0);
99
100 for (i = length - 1; i >= 0; --i) aom_write_bit(w, (x >> i) & 0x01);
101 }
102
get_lower_coeff(tran_low_t qc)103 static INLINE tran_low_t get_lower_coeff(tran_low_t qc) {
104 if (qc == 0) {
105 return 0;
106 }
107 return qc > 0 ? qc - 1 : qc + 1;
108 }
109
qcoeff_to_dqcoeff(tran_low_t qc,int coeff_idx,int dqv,int shift,const qm_val_t * iqmatrix)110 static INLINE tran_low_t qcoeff_to_dqcoeff(tran_low_t qc, int coeff_idx,
111 int dqv, int shift,
112 const qm_val_t *iqmatrix) {
113 int sign = qc < 0 ? -1 : 1;
114 if (iqmatrix != NULL)
115 dqv =
116 ((iqmatrix[coeff_idx] * dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
117 return sign * ((abs(qc) * dqv) >> shift);
118 }
119
get_coeff_dist(tran_low_t tcoeff,tran_low_t dqcoeff,int shift)120 static INLINE int64_t get_coeff_dist(tran_low_t tcoeff, tran_low_t dqcoeff,
121 int shift) {
122 const int64_t diff = (tcoeff - dqcoeff) * (1 << shift);
123 const int64_t error = diff * diff;
124 return error;
125 }
126
127 static const int8_t eob_to_pos_small[33] = {
128 0, 1, 2, // 0-2
129 3, 3, // 3-4
130 4, 4, 4, 4, // 5-8
131 5, 5, 5, 5, 5, 5, 5, 5, // 9-16
132 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6 // 17-32
133 };
134
135 static const int8_t eob_to_pos_large[17] = {
136 6, // place holder
137 7, // 33-64
138 8, 8, // 65-128
139 9, 9, 9, 9, // 129-256
140 10, 10, 10, 10, 10, 10, 10, 10, // 257-512
141 11 // 513-
142 };
143
get_eob_pos_token(const int eob,int * const extra)144 static INLINE int get_eob_pos_token(const int eob, int *const extra) {
145 int t;
146
147 if (eob < 33) {
148 t = eob_to_pos_small[eob];
149 } else {
150 const int e = AOMMIN((eob - 1) >> 5, 16);
151 t = eob_to_pos_large[e];
152 }
153
154 *extra = eob - av1_eob_group_start[t];
155
156 return t;
157 }
158
159 #if CONFIG_ENTROPY_STATS
av1_update_eob_context(int cdf_idx,int eob,TX_SIZE tx_size,TX_CLASS tx_class,PLANE_TYPE plane,FRAME_CONTEXT * ec_ctx,FRAME_COUNTS * counts,uint8_t allow_update_cdf)160 void av1_update_eob_context(int cdf_idx, int eob, TX_SIZE tx_size,
161 TX_CLASS tx_class, PLANE_TYPE plane,
162 FRAME_CONTEXT *ec_ctx, FRAME_COUNTS *counts,
163 uint8_t allow_update_cdf) {
164 #else
165 void av1_update_eob_context(int eob, TX_SIZE tx_size, TX_CLASS tx_class,
166 PLANE_TYPE plane, FRAME_CONTEXT *ec_ctx,
167 uint8_t allow_update_cdf) {
168 #endif
169 int eob_extra;
170 const int eob_pt = get_eob_pos_token(eob, &eob_extra);
171 TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
172
173 const int eob_multi_size = txsize_log2_minus4[tx_size];
174 const int eob_multi_ctx = (tx_class == TX_CLASS_2D) ? 0 : 1;
175
176 switch (eob_multi_size) {
177 case 0:
178 #if CONFIG_ENTROPY_STATS
179 ++counts->eob_multi16[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
180 #endif
181 if (allow_update_cdf)
182 update_cdf(ec_ctx->eob_flag_cdf16[plane][eob_multi_ctx], eob_pt - 1, 5);
183 break;
184 case 1:
185 #if CONFIG_ENTROPY_STATS
186 ++counts->eob_multi32[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
187 #endif
188 if (allow_update_cdf)
189 update_cdf(ec_ctx->eob_flag_cdf32[plane][eob_multi_ctx], eob_pt - 1, 6);
190 break;
191 case 2:
192 #if CONFIG_ENTROPY_STATS
193 ++counts->eob_multi64[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
194 #endif
195 if (allow_update_cdf)
196 update_cdf(ec_ctx->eob_flag_cdf64[plane][eob_multi_ctx], eob_pt - 1, 7);
197 break;
198 case 3:
199 #if CONFIG_ENTROPY_STATS
200 ++counts->eob_multi128[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
201 #endif
202 if (allow_update_cdf) {
203 update_cdf(ec_ctx->eob_flag_cdf128[plane][eob_multi_ctx], eob_pt - 1,
204 8);
205 }
206 break;
207 case 4:
208 #if CONFIG_ENTROPY_STATS
209 ++counts->eob_multi256[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
210 #endif
211 if (allow_update_cdf) {
212 update_cdf(ec_ctx->eob_flag_cdf256[plane][eob_multi_ctx], eob_pt - 1,
213 9);
214 }
215 break;
216 case 5:
217 #if CONFIG_ENTROPY_STATS
218 ++counts->eob_multi512[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
219 #endif
220 if (allow_update_cdf) {
221 update_cdf(ec_ctx->eob_flag_cdf512[plane][eob_multi_ctx], eob_pt - 1,
222 10);
223 }
224 break;
225 case 6:
226 default:
227 #if CONFIG_ENTROPY_STATS
228 ++counts->eob_multi1024[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
229 #endif
230 if (allow_update_cdf) {
231 update_cdf(ec_ctx->eob_flag_cdf1024[plane][eob_multi_ctx], eob_pt - 1,
232 11);
233 }
234 break;
235 }
236
237 if (av1_eob_offset_bits[eob_pt] > 0) {
238 int eob_ctx = eob_pt - 3;
239 int eob_shift = av1_eob_offset_bits[eob_pt] - 1;
240 int bit = (eob_extra & (1 << eob_shift)) ? 1 : 0;
241 #if CONFIG_ENTROPY_STATS
242 counts->eob_extra[cdf_idx][txs_ctx][plane][eob_pt][bit]++;
243 #endif // CONFIG_ENTROPY_STATS
244 if (allow_update_cdf)
245 update_cdf(ec_ctx->eob_extra_cdf[txs_ctx][plane][eob_ctx], bit, 2);
246 }
247 }
248
249 static int get_eob_cost(int eob, const LV_MAP_EOB_COST *txb_eob_costs,
250 const LV_MAP_COEFF_COST *txb_costs, TX_CLASS tx_class) {
251 int eob_extra;
252 const int eob_pt = get_eob_pos_token(eob, &eob_extra);
253 int eob_cost = 0;
254 const int eob_multi_ctx = (tx_class == TX_CLASS_2D) ? 0 : 1;
255 eob_cost = txb_eob_costs->eob_cost[eob_multi_ctx][eob_pt - 1];
256
257 if (av1_eob_offset_bits[eob_pt] > 0) {
258 const int eob_ctx = eob_pt - 3;
259 const int eob_shift = av1_eob_offset_bits[eob_pt] - 1;
260 const int bit = (eob_extra & (1 << eob_shift)) ? 1 : 0;
261 eob_cost += txb_costs->eob_extra_cost[eob_ctx][bit];
262 const int offset_bits = av1_eob_offset_bits[eob_pt];
263 if (offset_bits > 1) eob_cost += av1_cost_literal(offset_bits - 1);
264 }
265 return eob_cost;
266 }
267
268 static INLINE int get_sign_bit_cost(tran_low_t qc, int coeff_idx,
269 const int (*dc_sign_cost)[2],
270 int dc_sign_ctx) {
271 if (coeff_idx == 0) {
272 const int sign = (qc < 0) ? 1 : 0;
273 return dc_sign_cost[dc_sign_ctx][sign];
274 }
275 return av1_cost_literal(1);
276 }
277
278 static const int golomb_bits_cost[32] = {
279 0, 512, 512 * 3, 512 * 3, 512 * 5, 512 * 5, 512 * 5, 512 * 5,
280 512 * 7, 512 * 7, 512 * 7, 512 * 7, 512 * 7, 512 * 7, 512 * 7, 512 * 7,
281 512 * 9, 512 * 9, 512 * 9, 512 * 9, 512 * 9, 512 * 9, 512 * 9, 512 * 9,
282 512 * 9, 512 * 9, 512 * 9, 512 * 9, 512 * 9, 512 * 9, 512 * 9, 512 * 9
283 };
284 static const int golomb_cost_diff[32] = {
285 0, 512, 512 * 2, 0, 512 * 2, 0, 0, 0, 512 * 2, 0, 0, 0, 0, 0, 0, 0,
286 512 * 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0
287 };
288
289 static INLINE int get_golomb_cost(int abs_qc) {
290 if (abs_qc >= 1 + NUM_BASE_LEVELS + COEFF_BASE_RANGE) {
291 const int r = abs_qc - COEFF_BASE_RANGE - NUM_BASE_LEVELS;
292 const int length = get_msb(r) + 1;
293 return av1_cost_literal(2 * length - 1);
294 }
295 return 0;
296 }
297
298 static INLINE int get_br_cost_with_diff(tran_low_t level, const int *coeff_lps,
299 int *diff) {
300 const int base_range = AOMMIN(level - 1 - NUM_BASE_LEVELS, COEFF_BASE_RANGE);
301 int golomb_bits = 0;
302 if (level <= COEFF_BASE_RANGE + 1 + NUM_BASE_LEVELS)
303 *diff += coeff_lps[base_range + COEFF_BASE_RANGE + 1];
304
305 if (level >= COEFF_BASE_RANGE + 1 + NUM_BASE_LEVELS) {
306 int r = level - COEFF_BASE_RANGE - NUM_BASE_LEVELS;
307 if (r < 32) {
308 golomb_bits = golomb_bits_cost[r];
309 *diff += golomb_cost_diff[r];
310 } else {
311 golomb_bits = get_golomb_cost(level);
312 *diff += (r & (r - 1)) == 0 ? 1024 : 0;
313 }
314 }
315
316 return coeff_lps[base_range] + golomb_bits;
317 }
318
319 static INLINE int get_br_cost(tran_low_t level, const int *coeff_lps) {
320 const int base_range = AOMMIN(level - 1 - NUM_BASE_LEVELS, COEFF_BASE_RANGE);
321 return coeff_lps[base_range] + get_golomb_cost(level);
322 }
323
324 static int get_coeff_cost(const tran_low_t qc, const int scan_idx,
325 const int is_eob, const TxbInfo *const txb_info,
326 const LV_MAP_COEFF_COST *const txb_costs,
327 const int coeff_ctx, const TX_CLASS tx_class) {
328 const TXB_CTX *const txb_ctx = txb_info->txb_ctx;
329 const int is_nz = (qc != 0);
330 const tran_low_t abs_qc = abs(qc);
331 int cost = 0;
332 const int16_t *const scan = txb_info->scan_order->scan;
333 const int pos = scan[scan_idx];
334
335 if (is_eob) {
336 cost += txb_costs->base_eob_cost[coeff_ctx][AOMMIN(abs_qc, 3) - 1];
337 } else {
338 cost += txb_costs->base_cost[coeff_ctx][AOMMIN(abs_qc, 3)];
339 }
340 if (is_nz) {
341 cost += get_sign_bit_cost(qc, scan_idx, txb_costs->dc_sign_cost,
342 txb_ctx->dc_sign_ctx);
343
344 if (abs_qc > NUM_BASE_LEVELS) {
345 const int ctx =
346 get_br_ctx(txb_info->levels, pos, txb_info->bwl, tx_class);
347 cost += get_br_cost(abs_qc, txb_costs->lps_cost[ctx]);
348 }
349 }
350 return cost;
351 }
352
353 static INLINE int get_nz_map_ctx(const uint8_t *const levels,
354 const int coeff_idx, const int bwl,
355 const int height, const int scan_idx,
356 const int is_eob, const TX_SIZE tx_size,
357 const TX_CLASS tx_class) {
358 if (is_eob) {
359 if (scan_idx == 0) return 0;
360 if (scan_idx <= (height << bwl) / 8) return 1;
361 if (scan_idx <= (height << bwl) / 4) return 2;
362 return 3;
363 }
364 const int stats =
365 get_nz_mag(levels + get_padded_idx(coeff_idx, bwl), bwl, tx_class);
366 return get_nz_map_ctx_from_stats(stats, coeff_idx, bwl, tx_size, tx_class);
367 }
368
369 static void get_dist_cost_stats(LevelDownStats *const stats, const int scan_idx,
370 const int is_eob,
371 const LV_MAP_COEFF_COST *const txb_costs,
372 const TxbInfo *const txb_info,
373 const TX_CLASS tx_class) {
374 const int16_t *const scan = txb_info->scan_order->scan;
375 const int coeff_idx = scan[scan_idx];
376 const tran_low_t qc = txb_info->qcoeff[coeff_idx];
377 const uint8_t *const levels = txb_info->levels;
378 stats->new_eob = -1;
379 stats->update = 0;
380 stats->rd_low = 0;
381 stats->rd = 0;
382 stats->nz_rd = 0;
383 stats->dist_low = 0;
384 stats->rate_low = 0;
385 stats->low_qc = 0;
386
387 const tran_low_t tqc = txb_info->tcoeff[coeff_idx];
388 const int dqv = txb_info->dequant[coeff_idx != 0];
389 const int coeff_ctx =
390 get_nz_map_ctx(levels, coeff_idx, txb_info->bwl, txb_info->height,
391 scan_idx, is_eob, txb_info->tx_size, tx_class);
392 const int qc_cost = get_coeff_cost(qc, scan_idx, is_eob, txb_info, txb_costs,
393 coeff_ctx, tx_class);
394 assert(qc != 0);
395 const tran_low_t dqc = qcoeff_to_dqcoeff(qc, coeff_idx, dqv, txb_info->shift,
396 txb_info->iqmatrix);
397 const int64_t dqc_dist = get_coeff_dist(tqc, dqc, txb_info->shift);
398
399 // distortion difference when coefficient is quantized to 0
400 const tran_low_t dqc0 =
401 qcoeff_to_dqcoeff(0, coeff_idx, dqv, txb_info->shift, txb_info->iqmatrix);
402
403 stats->dist0 = get_coeff_dist(tqc, dqc0, txb_info->shift);
404 stats->dist = dqc_dist - stats->dist0;
405 stats->rate = qc_cost;
406
407 stats->rd = RDCOST(txb_info->rdmult, stats->rate, stats->dist);
408
409 stats->low_qc = get_lower_coeff(qc);
410
411 if (is_eob && stats->low_qc == 0) {
412 stats->rd_low = stats->rd; // disable selection of low_qc in this case.
413 } else {
414 if (stats->low_qc == 0) {
415 stats->dist_low = 0;
416 } else {
417 stats->low_dqc = qcoeff_to_dqcoeff(stats->low_qc, coeff_idx, dqv,
418 txb_info->shift, txb_info->iqmatrix);
419 const int64_t low_dqc_dist =
420 get_coeff_dist(tqc, stats->low_dqc, txb_info->shift);
421 stats->dist_low = low_dqc_dist - stats->dist0;
422 }
423 const int low_qc_cost =
424 get_coeff_cost(stats->low_qc, scan_idx, is_eob, txb_info, txb_costs,
425 coeff_ctx, tx_class);
426 stats->rate_low = low_qc_cost;
427 stats->rd_low = RDCOST(txb_info->rdmult, stats->rate_low, stats->dist_low);
428 }
429 }
430
431 static void get_dist_cost_stats_with_eob(
432 LevelDownStats *const stats, const int scan_idx,
433 const LV_MAP_COEFF_COST *const txb_costs, const TxbInfo *const txb_info,
434 const TX_CLASS tx_class) {
435 const int is_eob = 0;
436 get_dist_cost_stats(stats, scan_idx, is_eob, txb_costs, txb_info, tx_class);
437
438 const int16_t *const scan = txb_info->scan_order->scan;
439 const int coeff_idx = scan[scan_idx];
440 const tran_low_t qc = txb_info->qcoeff[coeff_idx];
441 const int coeff_ctx_temp = get_nz_map_ctx(
442 txb_info->levels, coeff_idx, txb_info->bwl, txb_info->height, scan_idx, 1,
443 txb_info->tx_size, tx_class);
444 const int qc_eob_cost = get_coeff_cost(qc, scan_idx, 1, txb_info, txb_costs,
445 coeff_ctx_temp, tx_class);
446 int64_t rd_eob = RDCOST(txb_info->rdmult, qc_eob_cost, stats->dist);
447 if (stats->low_qc != 0) {
448 const int low_qc_eob_cost =
449 get_coeff_cost(stats->low_qc, scan_idx, 1, txb_info, txb_costs,
450 coeff_ctx_temp, tx_class);
451 int64_t rd_eob_low =
452 RDCOST(txb_info->rdmult, low_qc_eob_cost, stats->dist_low);
453 rd_eob = (rd_eob > rd_eob_low) ? rd_eob_low : rd_eob;
454 }
455
456 stats->nz_rd = AOMMIN(stats->rd_low, stats->rd) - rd_eob;
457 }
458
459 static INLINE void update_qcoeff(const int coeff_idx, const tran_low_t qc,
460 const TxbInfo *const txb_info) {
461 txb_info->qcoeff[coeff_idx] = qc;
462 txb_info->levels[get_padded_idx(coeff_idx, txb_info->bwl)] =
463 (uint8_t)clamp(abs(qc), 0, INT8_MAX);
464 }
465
466 static INLINE void update_coeff(const int coeff_idx, const tran_low_t qc,
467 const TxbInfo *const txb_info) {
468 update_qcoeff(coeff_idx, qc, txb_info);
469 const int dqv = txb_info->dequant[coeff_idx != 0];
470 txb_info->dqcoeff[coeff_idx] = qcoeff_to_dqcoeff(
471 qc, coeff_idx, dqv, txb_info->shift, txb_info->iqmatrix);
472 }
473
474 void av1_txb_init_levels_c(const tran_low_t *const coeff, const int width,
475 const int height, uint8_t *const levels) {
476 const int stride = width + TX_PAD_HOR;
477 uint8_t *ls = levels;
478
479 memset(levels + stride * height, 0,
480 sizeof(*levels) * (TX_PAD_BOTTOM * stride + TX_PAD_END));
481
482 for (int i = 0; i < height; i++) {
483 for (int j = 0; j < width; j++) {
484 *ls++ = (uint8_t)clamp(abs(coeff[i * width + j]), 0, INT8_MAX);
485 }
486 for (int j = 0; j < TX_PAD_HOR; j++) {
487 *ls++ = 0;
488 }
489 }
490 }
491
492 void av1_get_nz_map_contexts_c(const uint8_t *const levels,
493 const int16_t *const scan, const uint16_t eob,
494 const TX_SIZE tx_size, const TX_CLASS tx_class,
495 int8_t *const coeff_contexts) {
496 const int bwl = get_txb_bwl(tx_size);
497 const int height = get_txb_high(tx_size);
498 for (int i = 0; i < eob; ++i) {
499 const int pos = scan[i];
500 coeff_contexts[pos] = get_nz_map_ctx(levels, pos, bwl, height, i,
501 i == eob - 1, tx_size, tx_class);
502 }
503 }
504
505 void av1_write_coeffs_txb(const AV1_COMMON *const cm, MACROBLOCK *const x,
506 aom_writer *w, int blk_row, int blk_col, int plane,
507 int block, TX_SIZE tx_size) {
508 MACROBLOCKD *xd = &x->e_mbd;
509 const CB_COEFF_BUFFER *cb_coef_buff = x->cb_coef_buff;
510 const int txb_offset =
511 x->mbmi_ext_frame->cb_offset / (TX_SIZE_W_MIN * TX_SIZE_H_MIN);
512 const uint16_t *eob_txb = cb_coef_buff->eobs[plane] + txb_offset;
513 const uint16_t eob = eob_txb[block];
514 const uint8_t *entropy_ctx = cb_coef_buff->entropy_ctx[plane] + txb_offset;
515 const int txb_skip_ctx = entropy_ctx[block] & TXB_SKIP_CTX_MASK;
516 const TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
517 FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
518 aom_write_symbol(w, eob == 0, ec_ctx->txb_skip_cdf[txs_ctx][txb_skip_ctx], 2);
519 if (eob == 0) return;
520
521 const PLANE_TYPE plane_type = get_plane_type(plane);
522 const TX_TYPE tx_type =
523 av1_get_tx_type(xd, plane_type, blk_row, blk_col, tx_size,
524 cm->features.reduced_tx_set_used);
525 // Only y plane's tx_type is transmitted
526 if (plane == 0) {
527 av1_write_tx_type(cm, xd, tx_type, tx_size, w);
528 }
529
530 int eob_extra;
531 const int eob_pt = get_eob_pos_token(eob, &eob_extra);
532 const int eob_multi_size = txsize_log2_minus4[tx_size];
533 const TX_CLASS tx_class = tx_type_to_class[tx_type];
534 const int eob_multi_ctx = (tx_class == TX_CLASS_2D) ? 0 : 1;
535 switch (eob_multi_size) {
536 case 0:
537 aom_write_symbol(w, eob_pt - 1,
538 ec_ctx->eob_flag_cdf16[plane_type][eob_multi_ctx], 5);
539 break;
540 case 1:
541 aom_write_symbol(w, eob_pt - 1,
542 ec_ctx->eob_flag_cdf32[plane_type][eob_multi_ctx], 6);
543 break;
544 case 2:
545 aom_write_symbol(w, eob_pt - 1,
546 ec_ctx->eob_flag_cdf64[plane_type][eob_multi_ctx], 7);
547 break;
548 case 3:
549 aom_write_symbol(w, eob_pt - 1,
550 ec_ctx->eob_flag_cdf128[plane_type][eob_multi_ctx], 8);
551 break;
552 case 4:
553 aom_write_symbol(w, eob_pt - 1,
554 ec_ctx->eob_flag_cdf256[plane_type][eob_multi_ctx], 9);
555 break;
556 case 5:
557 aom_write_symbol(w, eob_pt - 1,
558 ec_ctx->eob_flag_cdf512[plane_type][eob_multi_ctx], 10);
559 break;
560 default:
561 aom_write_symbol(w, eob_pt - 1,
562 ec_ctx->eob_flag_cdf1024[plane_type][eob_multi_ctx], 11);
563 break;
564 }
565
566 const int eob_offset_bits = av1_eob_offset_bits[eob_pt];
567 if (eob_offset_bits > 0) {
568 const int eob_ctx = eob_pt - 3;
569 int eob_shift = eob_offset_bits - 1;
570 int bit = (eob_extra & (1 << eob_shift)) ? 1 : 0;
571 aom_write_symbol(w, bit,
572 ec_ctx->eob_extra_cdf[txs_ctx][plane_type][eob_ctx], 2);
573 for (int i = 1; i < eob_offset_bits; i++) {
574 eob_shift = eob_offset_bits - 1 - i;
575 bit = (eob_extra & (1 << eob_shift)) ? 1 : 0;
576 aom_write_bit(w, bit);
577 }
578 }
579
580 const int width = get_txb_wide(tx_size);
581 const int height = get_txb_high(tx_size);
582 uint8_t levels_buf[TX_PAD_2D];
583 uint8_t *const levels = set_levels(levels_buf, width);
584 const tran_low_t *tcoeff_txb =
585 cb_coef_buff->tcoeff[plane] + x->mbmi_ext_frame->cb_offset;
586 const tran_low_t *tcoeff = tcoeff_txb + BLOCK_OFFSET(block);
587 av1_txb_init_levels(tcoeff, width, height, levels);
588 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
589 const int16_t *const scan = scan_order->scan;
590 DECLARE_ALIGNED(16, int8_t, coeff_contexts[MAX_TX_SQUARE]);
591 av1_get_nz_map_contexts(levels, scan, eob, tx_size, tx_class, coeff_contexts);
592
593 const int bwl = get_txb_bwl(tx_size);
594 for (int c = eob - 1; c >= 0; --c) {
595 const int pos = scan[c];
596 const int coeff_ctx = coeff_contexts[pos];
597 const tran_low_t v = tcoeff[pos];
598 const tran_low_t level = abs(v);
599
600 if (c == eob - 1) {
601 aom_write_symbol(
602 w, AOMMIN(level, 3) - 1,
603 ec_ctx->coeff_base_eob_cdf[txs_ctx][plane_type][coeff_ctx], 3);
604 } else {
605 aom_write_symbol(w, AOMMIN(level, 3),
606 ec_ctx->coeff_base_cdf[txs_ctx][plane_type][coeff_ctx],
607 4);
608 }
609 if (level > NUM_BASE_LEVELS) {
610 // level is above 1.
611 const int base_range = level - 1 - NUM_BASE_LEVELS;
612 const int br_ctx = get_br_ctx(levels, pos, bwl, tx_class);
613 aom_cdf_prob *cdf =
614 ec_ctx->coeff_br_cdf[AOMMIN(txs_ctx, TX_32X32)][plane_type][br_ctx];
615 for (int idx = 0; idx < COEFF_BASE_RANGE; idx += BR_CDF_SIZE - 1) {
616 const int k = AOMMIN(base_range - idx, BR_CDF_SIZE - 1);
617 aom_write_symbol(w, k, cdf, BR_CDF_SIZE);
618 if (k < BR_CDF_SIZE - 1) break;
619 }
620 }
621 }
622
623 // Loop to code all signs in the transform block,
624 // starting with the sign of DC (if applicable)
625 for (int c = 0; c < eob; ++c) {
626 const tran_low_t v = tcoeff[scan[c]];
627 const tran_low_t level = abs(v);
628 const int sign = (v < 0) ? 1 : 0;
629 if (level) {
630 if (c == 0) {
631 const int dc_sign_ctx =
632 (entropy_ctx[block] >> DC_SIGN_CTX_SHIFT) & DC_SIGN_CTX_MASK;
633 aom_write_symbol(w, sign, ec_ctx->dc_sign_cdf[plane_type][dc_sign_ctx],
634 2);
635 } else {
636 aom_write_bit(w, sign);
637 }
638 if (level > COEFF_BASE_RANGE + NUM_BASE_LEVELS)
639 write_golomb(w, level - COEFF_BASE_RANGE - 1 - NUM_BASE_LEVELS);
640 }
641 }
642 }
643
644 typedef struct encode_txb_args {
645 const AV1_COMMON *cm;
646 MACROBLOCK *x;
647 aom_writer *w;
648 } ENCODE_TXB_ARGS;
649
650 void av1_write_coeffs_mb(const AV1_COMMON *const cm, MACROBLOCK *x,
651 aom_writer *w, BLOCK_SIZE bsize) {
652 MACROBLOCKD *xd = &x->e_mbd;
653 const int num_planes = av1_num_planes(cm);
654 int block[MAX_MB_PLANE] = { 0 };
655 int row, col;
656 assert(bsize == get_plane_block_size(bsize, xd->plane[0].subsampling_x,
657 xd->plane[0].subsampling_y));
658 const int max_blocks_wide = max_block_wide(xd, bsize, 0);
659 const int max_blocks_high = max_block_high(xd, bsize, 0);
660 const BLOCK_SIZE max_unit_bsize = BLOCK_64X64;
661 int mu_blocks_wide = mi_size_wide[max_unit_bsize];
662 int mu_blocks_high = mi_size_high[max_unit_bsize];
663 mu_blocks_wide = AOMMIN(max_blocks_wide, mu_blocks_wide);
664 mu_blocks_high = AOMMIN(max_blocks_high, mu_blocks_high);
665
666 for (row = 0; row < max_blocks_high; row += mu_blocks_high) {
667 for (col = 0; col < max_blocks_wide; col += mu_blocks_wide) {
668 for (int plane = 0; plane < num_planes; ++plane) {
669 if (plane && !xd->is_chroma_ref) break;
670 const TX_SIZE tx_size = av1_get_tx_size(plane, xd);
671 const int stepr = tx_size_high_unit[tx_size];
672 const int stepc = tx_size_wide_unit[tx_size];
673 const int step = stepr * stepc;
674 const struct macroblockd_plane *const pd = &xd->plane[plane];
675 const int unit_height = ROUND_POWER_OF_TWO(
676 AOMMIN(mu_blocks_high + row, max_blocks_high), pd->subsampling_y);
677 const int unit_width = ROUND_POWER_OF_TWO(
678 AOMMIN(mu_blocks_wide + col, max_blocks_wide), pd->subsampling_x);
679 for (int blk_row = row >> pd->subsampling_y; blk_row < unit_height;
680 blk_row += stepr) {
681 for (int blk_col = col >> pd->subsampling_x; blk_col < unit_width;
682 blk_col += stepc) {
683 av1_write_coeffs_txb(cm, x, w, blk_row, blk_col, plane,
684 block[plane], tx_size);
685 block[plane] += step;
686 }
687 }
688 }
689 }
690 }
691 }
692
693 // TODO(angiebird): use this function whenever it's possible
694 static int get_tx_type_cost(const MACROBLOCK *x, const MACROBLOCKD *xd,
695 int plane, TX_SIZE tx_size, TX_TYPE tx_type,
696 int reduced_tx_set_used) {
697 if (plane > 0) return 0;
698
699 const TX_SIZE square_tx_size = txsize_sqr_map[tx_size];
700
701 const MB_MODE_INFO *mbmi = xd->mi[0];
702 const int is_inter = is_inter_block(mbmi);
703 if (get_ext_tx_types(tx_size, is_inter, reduced_tx_set_used) > 1 &&
704 !xd->lossless[xd->mi[0]->segment_id]) {
705 const int ext_tx_set =
706 get_ext_tx_set(tx_size, is_inter, reduced_tx_set_used);
707 if (is_inter) {
708 if (ext_tx_set > 0)
709 return x->inter_tx_type_costs[ext_tx_set][square_tx_size][tx_type];
710 } else {
711 if (ext_tx_set > 0) {
712 PREDICTION_MODE intra_dir;
713 if (mbmi->filter_intra_mode_info.use_filter_intra)
714 intra_dir = fimode_to_intradir[mbmi->filter_intra_mode_info
715 .filter_intra_mode];
716 else
717 intra_dir = mbmi->mode;
718 return x->intra_tx_type_costs[ext_tx_set][square_tx_size][intra_dir]
719 [tx_type];
720 }
721 }
722 }
723 return 0;
724 }
725
726 static INLINE void update_coeff_eob_fast(int *eob, int shift,
727 const int16_t *dequant_ptr,
728 const int16_t *scan,
729 const tran_low_t *coeff_ptr,
730 tran_low_t *qcoeff_ptr,
731 tran_low_t *dqcoeff_ptr) {
732 // TODO(sarahparker) make this work for aomqm
733 int eob_out = *eob;
734 int zbin[2] = { dequant_ptr[0] + ROUND_POWER_OF_TWO(dequant_ptr[0] * 70, 7),
735 dequant_ptr[1] + ROUND_POWER_OF_TWO(dequant_ptr[1] * 70, 7) };
736
737 for (int i = *eob - 1; i >= 0; i--) {
738 const int rc = scan[i];
739 const int qcoeff = qcoeff_ptr[rc];
740 const int coeff = coeff_ptr[rc];
741 const int coeff_sign = AOMSIGN(coeff);
742 int64_t abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
743
744 if (((abs_coeff << (1 + shift)) < zbin[rc != 0]) || (qcoeff == 0)) {
745 eob_out--;
746 qcoeff_ptr[rc] = 0;
747 dqcoeff_ptr[rc] = 0;
748 } else {
749 break;
750 }
751 }
752
753 *eob = eob_out;
754 }
755
756 static AOM_FORCE_INLINE int warehouse_efficients_txb(
757 const MACROBLOCK *x, const int plane, const int block,
758 const TX_SIZE tx_size, const TXB_CTX *const txb_ctx,
759 const struct macroblock_plane *p, const int eob,
760 const PLANE_TYPE plane_type, const LV_MAP_COEFF_COST *const coeff_costs,
761 const MACROBLOCKD *const xd, const TX_TYPE tx_type, const TX_CLASS tx_class,
762 int reduced_tx_set_used) {
763 const tran_low_t *const qcoeff = p->qcoeff + BLOCK_OFFSET(block);
764 const int txb_skip_ctx = txb_ctx->txb_skip_ctx;
765 const int bwl = get_txb_bwl(tx_size);
766 const int width = get_txb_wide(tx_size);
767 const int height = get_txb_high(tx_size);
768 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
769 const int16_t *const scan = scan_order->scan;
770 uint8_t levels_buf[TX_PAD_2D];
771 uint8_t *const levels = set_levels(levels_buf, width);
772 DECLARE_ALIGNED(16, int8_t, coeff_contexts[MAX_TX_SQUARE]);
773 const int eob_multi_size = txsize_log2_minus4[tx_size];
774 const LV_MAP_EOB_COST *const eob_costs =
775 &x->eob_costs[eob_multi_size][plane_type];
776 int cost = coeff_costs->txb_skip_cost[txb_skip_ctx][0];
777
778 av1_txb_init_levels(qcoeff, width, height, levels);
779
780 cost += get_tx_type_cost(x, xd, plane, tx_size, tx_type, reduced_tx_set_used);
781
782 cost += get_eob_cost(eob, eob_costs, coeff_costs, tx_class);
783
784 av1_get_nz_map_contexts(levels, scan, eob, tx_size, tx_class, coeff_contexts);
785
786 const int(*lps_cost)[COEFF_BASE_RANGE + 1 + COEFF_BASE_RANGE + 1] =
787 coeff_costs->lps_cost;
788 int c = eob - 1;
789 {
790 const int pos = scan[c];
791 const tran_low_t v = qcoeff[pos];
792 const int sign = AOMSIGN(v);
793 const int level = (v ^ sign) - sign;
794 const int coeff_ctx = coeff_contexts[pos];
795 cost += coeff_costs->base_eob_cost[coeff_ctx][AOMMIN(level, 3) - 1];
796
797 if (v) {
798 // sign bit cost
799 if (level > NUM_BASE_LEVELS) {
800 const int ctx = get_br_ctx_eob(pos, bwl, tx_class);
801 cost += get_br_cost(level, lps_cost[ctx]);
802 }
803 if (c) {
804 cost += av1_cost_literal(1);
805 } else {
806 const int sign01 = (sign ^ sign) - sign;
807 const int dc_sign_ctx = txb_ctx->dc_sign_ctx;
808 cost += coeff_costs->dc_sign_cost[dc_sign_ctx][sign01];
809 return cost;
810 }
811 }
812 }
813 const int(*base_cost)[8] = coeff_costs->base_cost;
814 for (c = eob - 2; c >= 1; --c) {
815 const int pos = scan[c];
816 const int coeff_ctx = coeff_contexts[pos];
817 const tran_low_t v = qcoeff[pos];
818 const int level = abs(v);
819 cost += base_cost[coeff_ctx][AOMMIN(level, 3)];
820 if (v) {
821 // sign bit cost
822 cost += av1_cost_literal(1);
823 if (level > NUM_BASE_LEVELS) {
824 const int ctx = get_br_ctx(levels, pos, bwl, tx_class);
825 cost += get_br_cost(level, lps_cost[ctx]);
826 }
827 }
828 }
829 // c == 0 after previous loop
830 {
831 const int pos = scan[c];
832 const tran_low_t v = qcoeff[pos];
833 const int coeff_ctx = coeff_contexts[pos];
834 const int sign = AOMSIGN(v);
835 const int level = (v ^ sign) - sign;
836 cost += base_cost[coeff_ctx][AOMMIN(level, 3)];
837
838 if (v) {
839 // sign bit cost
840 const int sign01 = (sign ^ sign) - sign;
841 const int dc_sign_ctx = txb_ctx->dc_sign_ctx;
842 cost += coeff_costs->dc_sign_cost[dc_sign_ctx][sign01];
843 if (level > NUM_BASE_LEVELS) {
844 const int ctx = get_br_ctx(levels, pos, bwl, tx_class);
845 cost += get_br_cost(level, lps_cost[ctx]);
846 }
847 }
848 }
849 return cost;
850 }
851
852 static AOM_FORCE_INLINE int warehouse_efficients_txb_laplacian(
853 const MACROBLOCK *x, const int plane, const int block,
854 const TX_SIZE tx_size, const TXB_CTX *const txb_ctx, const int eob,
855 const PLANE_TYPE plane_type, const LV_MAP_COEFF_COST *const coeff_costs,
856 const MACROBLOCKD *const xd, const TX_TYPE tx_type, const TX_CLASS tx_class,
857 int reduced_tx_set_used) {
858 const int txb_skip_ctx = txb_ctx->txb_skip_ctx;
859
860 const int eob_multi_size = txsize_log2_minus4[tx_size];
861 const LV_MAP_EOB_COST *const eob_costs =
862 &x->eob_costs[eob_multi_size][plane_type];
863 int cost = coeff_costs->txb_skip_cost[txb_skip_ctx][0];
864
865 cost += get_tx_type_cost(x, xd, plane, tx_size, tx_type, reduced_tx_set_used);
866
867 cost += get_eob_cost(eob, eob_costs, coeff_costs, tx_class);
868
869 cost += av1_cost_coeffs_txb_estimate(x, plane, block, tx_size, tx_type);
870 return cost;
871 }
872
873 // Look up table of individual cost of coefficient by its quantization level.
874 // determined based on Laplacian distribution conditioned on estimated context
875 static const int costLUT[15] = { -1143, 53, 545, 825, 1031,
876 1209, 1393, 1577, 1763, 1947,
877 2132, 2317, 2501, 2686, 2871 };
878 static const int const_term = (1 << AV1_PROB_COST_SHIFT);
879 static const int loge_par = ((14427 << AV1_PROB_COST_SHIFT) + 5000) / 10000;
880 int av1_cost_coeffs_txb_estimate(const MACROBLOCK *x, const int plane,
881 const int block, const TX_SIZE tx_size,
882 const TX_TYPE tx_type) {
883 assert(plane == 0);
884
885 int cost = 0;
886 const struct macroblock_plane *p = &x->plane[plane];
887 const SCAN_ORDER *scan_order = get_scan(tx_size, tx_type);
888 const int16_t *scan = scan_order->scan;
889 tran_low_t *qcoeff = p->qcoeff + BLOCK_OFFSET(block);
890
891 int eob = p->eobs[block];
892
893 // coeffs
894 int c = eob - 1;
895 // eob
896 {
897 const int pos = scan[c];
898 const tran_low_t v = abs(qcoeff[pos]) - 1;
899 cost += (v << (AV1_PROB_COST_SHIFT + 2));
900 }
901 // other coeffs
902 for (c = eob - 2; c >= 0; c--) {
903 const int pos = scan[c];
904 const tran_low_t v = abs(qcoeff[pos]);
905 const int idx = AOMMIN(v, 14);
906
907 cost += costLUT[idx];
908 }
909
910 // const_term does not contain DC, and log(e) does not contain eob, so both
911 // (eob-1)
912 cost += (const_term + loge_par) * (eob - 1);
913
914 return cost;
915 }
916
917 int av1_cost_coeffs_txb(const MACROBLOCK *x, const int plane, const int block,
918 const TX_SIZE tx_size, const TX_TYPE tx_type,
919 const TXB_CTX *const txb_ctx, int reduced_tx_set_used) {
920 const struct macroblock_plane *p = &x->plane[plane];
921 const int eob = p->eobs[block];
922 const TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
923 const PLANE_TYPE plane_type = get_plane_type(plane);
924 const LV_MAP_COEFF_COST *const coeff_costs =
925 &x->coeff_costs[txs_ctx][plane_type];
926 if (eob == 0) {
927 return coeff_costs->txb_skip_cost[txb_ctx->txb_skip_ctx][1];
928 }
929
930 const MACROBLOCKD *const xd = &x->e_mbd;
931 const TX_CLASS tx_class = tx_type_to_class[tx_type];
932
933 return warehouse_efficients_txb(x, plane, block, tx_size, txb_ctx, p, eob,
934 plane_type, coeff_costs, xd, tx_type,
935 tx_class, reduced_tx_set_used);
936 }
937
938 int av1_cost_coeffs_txb_laplacian(const MACROBLOCK *x, const int plane,
939 const int block, const TX_SIZE tx_size,
940 const TX_TYPE tx_type,
941 const TXB_CTX *const txb_ctx,
942 const int reduced_tx_set_used,
943 const int adjust_eob) {
944 const struct macroblock_plane *p = &x->plane[plane];
945 int eob = p->eobs[block];
946
947 if (adjust_eob) {
948 const SCAN_ORDER *scan_order = get_scan(tx_size, tx_type);
949 const int16_t *scan = scan_order->scan;
950 tran_low_t *tcoeff = p->coeff + BLOCK_OFFSET(block);
951 tran_low_t *qcoeff = p->qcoeff + BLOCK_OFFSET(block);
952 const MACROBLOCKD *xd = &x->e_mbd;
953 const struct macroblockd_plane *const pd = &xd->plane[plane];
954 tran_low_t *dqcoeff = pd->dqcoeff + BLOCK_OFFSET(block);
955 update_coeff_eob_fast(&eob, av1_get_tx_scale(tx_size), p->dequant_QTX, scan,
956 tcoeff, qcoeff, dqcoeff);
957 p->eobs[block] = eob;
958 }
959
960 const TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
961 const PLANE_TYPE plane_type = get_plane_type(plane);
962 const LV_MAP_COEFF_COST *const coeff_costs =
963 &x->coeff_costs[txs_ctx][plane_type];
964 if (eob == 0) {
965 return coeff_costs->txb_skip_cost[txb_ctx->txb_skip_ctx][1];
966 }
967
968 const MACROBLOCKD *const xd = &x->e_mbd;
969 const TX_CLASS tx_class = tx_type_to_class[tx_type];
970
971 return warehouse_efficients_txb_laplacian(
972 x, plane, block, tx_size, txb_ctx, eob, plane_type, coeff_costs, xd,
973 tx_type, tx_class, reduced_tx_set_used);
974 }
975
976 static int optimize_txb(TxbInfo *txb_info, const LV_MAP_COEFF_COST *txb_costs,
977 const LV_MAP_EOB_COST *txb_eob_costs, int *rate_cost) {
978 int update = 0;
979 if (txb_info->eob == 0) return update;
980 const int16_t *const scan = txb_info->scan_order->scan;
981 // forward optimize the nz_map`
982 const int init_eob = txb_info->eob;
983 const TX_CLASS tx_class = tx_type_to_class[txb_info->tx_type];
984 const int eob_cost =
985 get_eob_cost(init_eob, txb_eob_costs, txb_costs, tx_class);
986
987 // backward optimize the level-k map
988 int accu_rate = eob_cost;
989 int64_t accu_dist = 0;
990 int64_t prev_eob_rd_cost = INT64_MAX;
991 int64_t cur_eob_rd_cost = 0;
992
993 {
994 const int si = init_eob - 1;
995 const int coeff_idx = scan[si];
996 LevelDownStats stats;
997 get_dist_cost_stats(&stats, si, si == init_eob - 1, txb_costs, txb_info,
998 tx_class);
999 if ((stats.rd_low < stats.rd) && (stats.low_qc != 0)) {
1000 update = 1;
1001 update_coeff(coeff_idx, stats.low_qc, txb_info);
1002 accu_rate += stats.rate_low;
1003 accu_dist += stats.dist_low;
1004 } else {
1005 accu_rate += stats.rate;
1006 accu_dist += stats.dist;
1007 }
1008 }
1009
1010 int si = init_eob - 2;
1011 int8_t has_nz_tail = 0;
1012 // eob is not fixed
1013 for (; si >= 0 && has_nz_tail < 2; --si) {
1014 assert(si != init_eob - 1);
1015 const int coeff_idx = scan[si];
1016 tran_low_t qc = txb_info->qcoeff[coeff_idx];
1017
1018 if (qc == 0) {
1019 const int coeff_ctx =
1020 get_lower_levels_ctx(txb_info->levels, coeff_idx, txb_info->bwl,
1021 txb_info->tx_size, tx_class);
1022 accu_rate += txb_costs->base_cost[coeff_ctx][0];
1023 } else {
1024 LevelDownStats stats;
1025 get_dist_cost_stats_with_eob(&stats, si, txb_costs, txb_info, tx_class);
1026 // check if it is better to make this the last significant coefficient
1027 int cur_eob_rate =
1028 get_eob_cost(si + 1, txb_eob_costs, txb_costs, tx_class);
1029 cur_eob_rd_cost = RDCOST(txb_info->rdmult, cur_eob_rate, 0);
1030 prev_eob_rd_cost =
1031 RDCOST(txb_info->rdmult, accu_rate, accu_dist) + stats.nz_rd;
1032 if (cur_eob_rd_cost <= prev_eob_rd_cost) {
1033 update = 1;
1034 for (int j = si + 1; j < txb_info->eob; j++) {
1035 const int coeff_pos_j = scan[j];
1036 update_coeff(coeff_pos_j, 0, txb_info);
1037 }
1038 txb_info->eob = si + 1;
1039
1040 // rerun cost calculation due to change of eob
1041 accu_rate = cur_eob_rate;
1042 accu_dist = 0;
1043 get_dist_cost_stats(&stats, si, 1, txb_costs, txb_info, tx_class);
1044 if ((stats.rd_low < stats.rd) && (stats.low_qc != 0)) {
1045 update = 1;
1046 update_coeff(coeff_idx, stats.low_qc, txb_info);
1047 accu_rate += stats.rate_low;
1048 accu_dist += stats.dist_low;
1049 } else {
1050 accu_rate += stats.rate;
1051 accu_dist += stats.dist;
1052 }
1053
1054 // reset non zero tail when new eob is found
1055 has_nz_tail = 0;
1056 } else {
1057 int bUpdCoeff = 0;
1058 if (stats.rd_low < stats.rd) {
1059 if ((si < txb_info->eob - 1)) {
1060 bUpdCoeff = 1;
1061 update = 1;
1062 }
1063 } else {
1064 ++has_nz_tail;
1065 }
1066
1067 if (bUpdCoeff) {
1068 update_coeff(coeff_idx, stats.low_qc, txb_info);
1069 accu_rate += stats.rate_low;
1070 accu_dist += stats.dist_low;
1071 } else {
1072 accu_rate += stats.rate;
1073 accu_dist += stats.dist;
1074 }
1075 }
1076 }
1077 } // for (si)
1078
1079 // eob is fixed
1080 for (; si >= 0; --si) {
1081 assert(si != init_eob - 1);
1082 const int coeff_idx = scan[si];
1083 tran_low_t qc = txb_info->qcoeff[coeff_idx];
1084
1085 if (qc == 0) {
1086 const int coeff_ctx =
1087 get_lower_levels_ctx(txb_info->levels, coeff_idx, txb_info->bwl,
1088 txb_info->tx_size, tx_class);
1089 accu_rate += txb_costs->base_cost[coeff_ctx][0];
1090 } else {
1091 LevelDownStats stats;
1092 get_dist_cost_stats(&stats, si, 0, txb_costs, txb_info, tx_class);
1093
1094 int bUpdCoeff = 0;
1095 if (stats.rd_low < stats.rd) {
1096 if ((si < txb_info->eob - 1)) {
1097 bUpdCoeff = 1;
1098 update = 1;
1099 }
1100 }
1101 if (bUpdCoeff) {
1102 update_coeff(coeff_idx, stats.low_qc, txb_info);
1103 accu_rate += stats.rate_low;
1104 accu_dist += stats.dist_low;
1105 } else {
1106 accu_rate += stats.rate;
1107 accu_dist += stats.dist;
1108 }
1109 }
1110 } // for (si)
1111
1112 int non_zero_blk_rate =
1113 txb_costs->txb_skip_cost[txb_info->txb_ctx->txb_skip_ctx][0];
1114 prev_eob_rd_cost =
1115 RDCOST(txb_info->rdmult, accu_rate + non_zero_blk_rate, accu_dist);
1116
1117 int zero_blk_rate =
1118 txb_costs->txb_skip_cost[txb_info->txb_ctx->txb_skip_ctx][1];
1119 int64_t zero_blk_rd_cost = RDCOST(txb_info->rdmult, zero_blk_rate, 0);
1120 if (zero_blk_rd_cost <= prev_eob_rd_cost) {
1121 update = 1;
1122 for (int j = 0; j < txb_info->eob; j++) {
1123 const int coeff_pos_j = scan[j];
1124 update_coeff(coeff_pos_j, 0, txb_info);
1125 }
1126 txb_info->eob = 0;
1127 }
1128
1129 // record total rate cost
1130 *rate_cost = zero_blk_rd_cost <= prev_eob_rd_cost
1131 ? zero_blk_rate
1132 : accu_rate + non_zero_blk_rate;
1133
1134 if (txb_info->eob > 0) {
1135 *rate_cost += txb_info->tx_type_cost;
1136 }
1137
1138 return update;
1139 }
1140
1141 #if CONFIG_HTB_TRELLIS
1142 static void hbt_init() {
1143 hbt_hash_table =
1144 aom_malloc(sizeof(OptTxbQcoeff) * HBT_TABLE_SIZE * HBT_ARRAY_LENGTH);
1145 memset(hbt_hash_table, 0,
1146 sizeof(OptTxbQcoeff) * HBT_TABLE_SIZE * HBT_ARRAY_LENGTH);
1147 av1_crc32c_calculator_init(&crc_calculator); // 31 bit: qc & ctx
1148
1149 hbt_needs_init = 0;
1150 }
1151
1152 void hbt_destroy() { aom_free(hbt_hash_table); }
1153
1154 static int hbt_hash_miss(uint32_t hbt_ctx_hash, uint32_t hbt_qc_hash,
1155 TxbInfo *txb_info, const LV_MAP_COEFF_COST *txb_costs,
1156 const LV_MAP_EOB_COST *txb_eob_costs,
1157 const struct macroblock_plane *p, int block,
1158 int fast_mode, int *rate_cost) {
1159 (void)fast_mode;
1160 const int16_t *scan = txb_info->scan_order->scan;
1161 int prev_eob = txb_info->eob;
1162 assert(HBT_EOB <= 16); // Lengthen array if allowing longer eob.
1163 int32_t prev_coeff[16];
1164 for (int i = 0; i < prev_eob; i++) {
1165 prev_coeff[i] = txb_info->qcoeff[scan[i]];
1166 }
1167 for (int i = prev_eob; i < HBT_EOB; i++) {
1168 prev_coeff[i] = 0; // For compiler piece of mind.
1169 }
1170
1171 av1_txb_init_levels(txb_info->qcoeff, txb_info->width, txb_info->height,
1172 txb_info->levels);
1173
1174 const int update =
1175 optimize_txb(txb_info, txb_costs, txb_eob_costs, rate_cost);
1176
1177 // Overwrite old entry
1178 uint16_t hbt_table_index = hbt_ctx_hash % HBT_TABLE_SIZE;
1179 uint16_t hbt_array_index = hbt_qc_hash % HBT_ARRAY_LENGTH;
1180 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1181 .rate_cost = *rate_cost;
1182 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index].init = 1;
1183 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1184 .hbt_qc_hash = hbt_qc_hash;
1185 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1186 .hbt_ctx_hash = hbt_ctx_hash;
1187 assert(prev_eob >= txb_info->eob); // eob can't get longer
1188 for (int i = 0; i < txb_info->eob; i++) {
1189 // Record how coeff changed. Convention: towards zero is negative.
1190 if (txb_info->qcoeff[scan[i]] > 0)
1191 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1192 .deltas[i] = txb_info->qcoeff[scan[i]] - prev_coeff[i];
1193 else
1194 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1195 .deltas[i] = prev_coeff[i] - txb_info->qcoeff[scan[i]];
1196 }
1197 for (int i = txb_info->eob; i < prev_eob; i++) {
1198 // If eob got shorter, record that all after it changed to zero.
1199 if (prev_coeff[i] > 0)
1200 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1201 .deltas[i] = -prev_coeff[i];
1202 else
1203 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1204 .deltas[i] = prev_coeff[i];
1205 }
1206 for (int i = prev_eob; i < HBT_EOB; i++) {
1207 // Record 'no change' after optimized coefficients run out.
1208 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1209 .deltas[i] = 0;
1210 }
1211
1212 if (update) {
1213 p->eobs[block] = txb_info->eob;
1214 p->txb_entropy_ctx[block] = av1_get_txb_entropy_context(
1215 txb_info->qcoeff, txb_info->scan_order, txb_info->eob);
1216 }
1217 return txb_info->eob;
1218 }
1219
1220 static int hbt_hash_hit(uint32_t hbt_table_index, int hbt_array_index,
1221 TxbInfo *txb_info, const struct macroblock_plane *p,
1222 int block, int *rate_cost) {
1223 const int16_t *scan = txb_info->scan_order->scan;
1224 int new_eob = 0;
1225 int update = 0;
1226
1227 for (int i = 0; i < txb_info->eob; i++) {
1228 // Delta convention is negatives go towards zero, so only apply those ones.
1229 if (hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1230 .deltas[i] < 0) {
1231 if (txb_info->qcoeff[scan[i]] > 0)
1232 txb_info->qcoeff[scan[i]] +=
1233 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1234 .deltas[i];
1235 else
1236 txb_info->qcoeff[scan[i]] -=
1237 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1238 .deltas[i];
1239
1240 update = 1;
1241 update_coeff(scan[i], txb_info->qcoeff[scan[i]], txb_info);
1242 }
1243 if (txb_info->qcoeff[scan[i]]) new_eob = i + 1;
1244 }
1245
1246 // Rate_cost can be calculated here instead (av1_cost_coeffs_txb), but
1247 // it is expensive and gives little benefit as long as qc_hash is high bit
1248 *rate_cost =
1249 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1250 .rate_cost;
1251
1252 if (update) {
1253 txb_info->eob = new_eob;
1254 p->eobs[block] = txb_info->eob;
1255 p->txb_entropy_ctx[block] = av1_get_txb_entropy_context(
1256 txb_info->qcoeff, txb_info->scan_order, txb_info->eob);
1257 }
1258
1259 return txb_info->eob;
1260 }
1261
1262 static int hbt_search_match(uint32_t hbt_ctx_hash, uint32_t hbt_qc_hash,
1263 TxbInfo *txb_info,
1264 const LV_MAP_COEFF_COST *txb_costs,
1265 const LV_MAP_EOB_COST *txb_eob_costs,
1266 const struct macroblock_plane *p, int block,
1267 int fast_mode, int *rate_cost) {
1268 // Check for qcoeff match
1269 int hbt_array_index = hbt_qc_hash % HBT_ARRAY_LENGTH;
1270 int hbt_table_index = hbt_ctx_hash % HBT_TABLE_SIZE;
1271
1272 if (hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1273 .hbt_qc_hash == hbt_qc_hash &&
1274 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1275 .hbt_ctx_hash == hbt_ctx_hash &&
1276 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1277 .init) {
1278 return hbt_hash_hit(hbt_table_index, hbt_array_index, txb_info, p, block,
1279 rate_cost);
1280 } else {
1281 return hbt_hash_miss(hbt_ctx_hash, hbt_qc_hash, txb_info, txb_costs,
1282 txb_eob_costs, p, block, fast_mode, rate_cost);
1283 }
1284 }
1285
1286 static int hbt_create_hashes(TxbInfo *txb_info,
1287 const LV_MAP_COEFF_COST *txb_costs,
1288 const LV_MAP_EOB_COST *txb_eob_costs,
1289 const struct macroblock_plane *p, int block,
1290 int fast_mode, int *rate_cost) {
1291 // Initialize hash table if needed.
1292 if (hbt_needs_init) {
1293 hbt_init();
1294 }
1295
1296 //// Hash creation
1297 uint8_t txb_hash_data[256]; // Asserts below to ensure enough space.
1298 const int16_t *scan = txb_info->scan_order->scan;
1299 uint8_t chunk = 0;
1300 int hash_data_index = 0;
1301
1302 // Make qc_hash.
1303 int packing_index = 0; // needed for packing.
1304 for (int i = 0; i < txb_info->eob; i++) {
1305 tran_low_t prechunk = txb_info->qcoeff[scan[i]];
1306
1307 // Softening: Improves speed. Aligns with signed deltas.
1308 if (prechunk < 0) prechunk *= -1;
1309
1310 // Early kick out: Don't apply feature if there are large coeffs:
1311 // If this kickout value is removed or raised beyond int8_t,
1312 // widen deltas type in OptTxbQcoeff struct.
1313 assert((int8_t)HBT_KICKOUT == HBT_KICKOUT); // If not, widen types.
1314 if (prechunk > HBT_KICKOUT) {
1315 av1_txb_init_levels(txb_info->qcoeff, txb_info->width, txb_info->height,
1316 txb_info->levels);
1317
1318 const int update =
1319 optimize_txb(txb_info, txb_costs, txb_eob_costs, rate_cost);
1320
1321 if (update) {
1322 p->eobs[block] = txb_info->eob;
1323 p->txb_entropy_ctx[block] = av1_get_txb_entropy_context(
1324 txb_info->qcoeff, txb_info->scan_order, txb_info->eob);
1325 }
1326 return txb_info->eob;
1327 }
1328
1329 // Since coeffs are 0 to 3, only 2 bits are needed: pack into bytes
1330 if (packing_index == 0) txb_hash_data[hash_data_index] = 0;
1331 chunk = prechunk << packing_index;
1332 packing_index += 2;
1333 txb_hash_data[hash_data_index] |= chunk;
1334
1335 // Full byte:
1336 if (packing_index == 8) {
1337 packing_index = 0;
1338 hash_data_index++;
1339 }
1340 }
1341 // Needed when packing_index != 0, to include final byte.
1342 hash_data_index++;
1343 assert(hash_data_index <= 64);
1344 // 31 bit qc_hash: index to array
1345 uint32_t hbt_qc_hash =
1346 av1_get_crc32c_value(&crc_calculator, txb_hash_data, hash_data_index);
1347
1348 // Make ctx_hash.
1349 hash_data_index = 0;
1350 tran_low_t prechunk;
1351
1352 for (int i = 0; i < txb_info->eob; i++) {
1353 // Save as magnitudes towards or away from zero.
1354 if (txb_info->tcoeff[scan[i]] >= 0)
1355 prechunk = txb_info->tcoeff[scan[i]] - txb_info->dqcoeff[scan[i]];
1356 else
1357 prechunk = txb_info->dqcoeff[scan[i]] - txb_info->tcoeff[scan[i]];
1358
1359 chunk = prechunk & 0xff;
1360 txb_hash_data[hash_data_index++] = chunk;
1361 }
1362
1363 // Extra ctx data:
1364 // Include dequants.
1365 txb_hash_data[hash_data_index++] = txb_info->dequant[0] & 0xff;
1366 txb_hash_data[hash_data_index++] = txb_info->dequant[1] & 0xff;
1367 chunk = txb_info->txb_ctx->txb_skip_ctx & 0xff;
1368 txb_hash_data[hash_data_index++] = chunk;
1369 chunk = txb_info->txb_ctx->dc_sign_ctx & 0xff;
1370 txb_hash_data[hash_data_index++] = chunk;
1371 // eob
1372 chunk = txb_info->eob & 0xff;
1373 txb_hash_data[hash_data_index++] = chunk;
1374 // rdmult (int64)
1375 chunk = txb_info->rdmult & 0xff;
1376 txb_hash_data[hash_data_index++] = chunk;
1377 // tx_type
1378 chunk = txb_info->tx_type & 0xff;
1379 txb_hash_data[hash_data_index++] = chunk;
1380 // base_eob_cost
1381 for (int i = 1; i < 3; i++) { // i = 0 are softened away
1382 for (int j = 0; j < SIG_COEF_CONTEXTS_EOB; j++) {
1383 chunk = (txb_costs->base_eob_cost[j][i] & 0xff00) >> 8;
1384 txb_hash_data[hash_data_index++] = chunk;
1385 }
1386 }
1387 // eob_cost
1388 for (int i = 0; i < 11; i++) {
1389 for (int j = 0; j < 2; j++) {
1390 chunk = (txb_eob_costs->eob_cost[j][i] & 0xff00) >> 8;
1391 txb_hash_data[hash_data_index++] = chunk;
1392 }
1393 }
1394 // dc_sign_cost
1395 for (int i = 0; i < 2; i++) {
1396 for (int j = 0; j < DC_SIGN_CONTEXTS; j++) {
1397 chunk = (txb_costs->dc_sign_cost[j][i] & 0xff00) >> 8;
1398 txb_hash_data[hash_data_index++] = chunk;
1399 }
1400 }
1401
1402 assert(hash_data_index <= 256);
1403 // 31 bit ctx_hash: used to index table
1404 uint32_t hbt_ctx_hash =
1405 av1_get_crc32c_value(&crc_calculator, txb_hash_data, hash_data_index);
1406 //// End hash creation
1407
1408 return hbt_search_match(hbt_ctx_hash, hbt_qc_hash, txb_info, txb_costs,
1409 txb_eob_costs, p, block, fast_mode, rate_cost);
1410 }
1411 #endif // CONFIG_HTB_TRELLIS
1412
1413 static AOM_FORCE_INLINE int get_two_coeff_cost_simple(
1414 int ci, tran_low_t abs_qc, int coeff_ctx,
1415 const LV_MAP_COEFF_COST *txb_costs, int bwl, TX_CLASS tx_class,
1416 const uint8_t *levels, int *cost_low) {
1417 // this simple version assumes the coeff's scan_idx is not DC (scan_idx != 0)
1418 // and not the last (scan_idx != eob - 1)
1419 assert(ci > 0);
1420 int cost = txb_costs->base_cost[coeff_ctx][AOMMIN(abs_qc, 3)];
1421 int diff = 0;
1422 if (abs_qc <= 3) diff = txb_costs->base_cost[coeff_ctx][abs_qc + 4];
1423 if (abs_qc) {
1424 cost += av1_cost_literal(1);
1425 if (abs_qc > NUM_BASE_LEVELS) {
1426 const int br_ctx = get_br_ctx(levels, ci, bwl, tx_class);
1427 int brcost_diff = 0;
1428 cost += get_br_cost_with_diff(abs_qc, txb_costs->lps_cost[br_ctx],
1429 &brcost_diff);
1430 diff += brcost_diff;
1431 }
1432 }
1433 *cost_low = cost - diff;
1434
1435 return cost;
1436 }
1437
1438 static INLINE int get_coeff_cost_eob(int ci, tran_low_t abs_qc, int sign,
1439 int coeff_ctx, int dc_sign_ctx,
1440 const LV_MAP_COEFF_COST *txb_costs,
1441 int bwl, TX_CLASS tx_class) {
1442 int cost = 0;
1443 cost += txb_costs->base_eob_cost[coeff_ctx][AOMMIN(abs_qc, 3) - 1];
1444 if (abs_qc != 0) {
1445 if (ci == 0) {
1446 cost += txb_costs->dc_sign_cost[dc_sign_ctx][sign];
1447 } else {
1448 cost += av1_cost_literal(1);
1449 }
1450 if (abs_qc > NUM_BASE_LEVELS) {
1451 int br_ctx;
1452 br_ctx = get_br_ctx_eob(ci, bwl, tx_class);
1453 cost += get_br_cost(abs_qc, txb_costs->lps_cost[br_ctx]);
1454 }
1455 }
1456 return cost;
1457 }
1458
1459 static INLINE int get_coeff_cost_general(int is_last, int ci, tran_low_t abs_qc,
1460 int sign, int coeff_ctx,
1461 int dc_sign_ctx,
1462 const LV_MAP_COEFF_COST *txb_costs,
1463 int bwl, TX_CLASS tx_class,
1464 const uint8_t *levels) {
1465 int cost = 0;
1466 if (is_last) {
1467 cost += txb_costs->base_eob_cost[coeff_ctx][AOMMIN(abs_qc, 3) - 1];
1468 } else {
1469 cost += txb_costs->base_cost[coeff_ctx][AOMMIN(abs_qc, 3)];
1470 }
1471 if (abs_qc != 0) {
1472 if (ci == 0) {
1473 cost += txb_costs->dc_sign_cost[dc_sign_ctx][sign];
1474 } else {
1475 cost += av1_cost_literal(1);
1476 }
1477 if (abs_qc > NUM_BASE_LEVELS) {
1478 int br_ctx;
1479 if (is_last)
1480 br_ctx = get_br_ctx_eob(ci, bwl, tx_class);
1481 else
1482 br_ctx = get_br_ctx(levels, ci, bwl, tx_class);
1483 cost += get_br_cost(abs_qc, txb_costs->lps_cost[br_ctx]);
1484 }
1485 }
1486 return cost;
1487 }
1488
1489 static INLINE void get_qc_dqc_low(tran_low_t abs_qc, int sign, int dqv,
1490 int shift, tran_low_t *qc_low,
1491 tran_low_t *dqc_low) {
1492 tran_low_t abs_qc_low = abs_qc - 1;
1493 *qc_low = (-sign ^ abs_qc_low) + sign;
1494 assert((sign ? -abs_qc_low : abs_qc_low) == *qc_low);
1495 tran_low_t abs_dqc_low = (abs_qc_low * dqv) >> shift;
1496 *dqc_low = (-sign ^ abs_dqc_low) + sign;
1497 assert((sign ? -abs_dqc_low : abs_dqc_low) == *dqc_low);
1498 }
1499
1500 static INLINE void update_coeff_general(
1501 int *accu_rate, int64_t *accu_dist, int si, int eob, TX_SIZE tx_size,
1502 TX_CLASS tx_class, int bwl, int height, int64_t rdmult, int shift,
1503 int dc_sign_ctx, const int16_t *dequant, const int16_t *scan,
1504 const LV_MAP_COEFF_COST *txb_costs, const tran_low_t *tcoeff,
1505 tran_low_t *qcoeff, tran_low_t *dqcoeff, uint8_t *levels,
1506 const qm_val_t *iqmatrix) {
1507 const int dqv = get_dqv(dequant, scan[si], iqmatrix);
1508 const int ci = scan[si];
1509 const tran_low_t qc = qcoeff[ci];
1510 const int is_last = si == (eob - 1);
1511 const int coeff_ctx = get_lower_levels_ctx_general(
1512 is_last, si, bwl, height, levels, ci, tx_size, tx_class);
1513 if (qc == 0) {
1514 *accu_rate += txb_costs->base_cost[coeff_ctx][0];
1515 } else {
1516 const int sign = (qc < 0) ? 1 : 0;
1517 const tran_low_t abs_qc = abs(qc);
1518 const tran_low_t tqc = tcoeff[ci];
1519 const tran_low_t dqc = dqcoeff[ci];
1520 const int64_t dist = get_coeff_dist(tqc, dqc, shift);
1521 const int64_t dist0 = get_coeff_dist(tqc, 0, shift);
1522 const int rate =
1523 get_coeff_cost_general(is_last, ci, abs_qc, sign, coeff_ctx,
1524 dc_sign_ctx, txb_costs, bwl, tx_class, levels);
1525 const int64_t rd = RDCOST(rdmult, rate, dist);
1526
1527 tran_low_t qc_low, dqc_low;
1528 tran_low_t abs_qc_low;
1529 int64_t dist_low, rd_low;
1530 int rate_low;
1531 if (abs_qc == 1) {
1532 abs_qc_low = qc_low = dqc_low = 0;
1533 dist_low = dist0;
1534 rate_low = txb_costs->base_cost[coeff_ctx][0];
1535 } else {
1536 get_qc_dqc_low(abs_qc, sign, dqv, shift, &qc_low, &dqc_low);
1537 abs_qc_low = abs_qc - 1;
1538 dist_low = get_coeff_dist(tqc, dqc_low, shift);
1539 rate_low =
1540 get_coeff_cost_general(is_last, ci, abs_qc_low, sign, coeff_ctx,
1541 dc_sign_ctx, txb_costs, bwl, tx_class, levels);
1542 }
1543
1544 rd_low = RDCOST(rdmult, rate_low, dist_low);
1545 if (rd_low < rd) {
1546 qcoeff[ci] = qc_low;
1547 dqcoeff[ci] = dqc_low;
1548 levels[get_padded_idx(ci, bwl)] = AOMMIN(abs_qc_low, INT8_MAX);
1549 *accu_rate += rate_low;
1550 *accu_dist += dist_low - dist0;
1551 } else {
1552 *accu_rate += rate;
1553 *accu_dist += dist - dist0;
1554 }
1555 }
1556 }
1557
1558 static AOM_FORCE_INLINE void update_coeff_simple(
1559 int *accu_rate, int si, int eob, TX_SIZE tx_size, TX_CLASS tx_class,
1560 int bwl, int64_t rdmult, int shift, const int16_t *dequant,
1561 const int16_t *scan, const LV_MAP_COEFF_COST *txb_costs,
1562 const tran_low_t *tcoeff, tran_low_t *qcoeff, tran_low_t *dqcoeff,
1563 uint8_t *levels, const qm_val_t *iqmatrix) {
1564 const int dqv = get_dqv(dequant, scan[si], iqmatrix);
1565 (void)eob;
1566 // this simple version assumes the coeff's scan_idx is not DC (scan_idx != 0)
1567 // and not the last (scan_idx != eob - 1)
1568 assert(si != eob - 1);
1569 assert(si > 0);
1570 const int ci = scan[si];
1571 const tran_low_t qc = qcoeff[ci];
1572 const int coeff_ctx =
1573 get_lower_levels_ctx(levels, ci, bwl, tx_size, tx_class);
1574 if (qc == 0) {
1575 *accu_rate += txb_costs->base_cost[coeff_ctx][0];
1576 } else {
1577 const tran_low_t abs_qc = abs(qc);
1578 const tran_low_t abs_tqc = abs(tcoeff[ci]);
1579 const tran_low_t abs_dqc = abs(dqcoeff[ci]);
1580 int rate_low = 0;
1581 const int rate = get_two_coeff_cost_simple(
1582 ci, abs_qc, coeff_ctx, txb_costs, bwl, tx_class, levels, &rate_low);
1583 if (abs_dqc < abs_tqc) {
1584 *accu_rate += rate;
1585 return;
1586 }
1587
1588 const int64_t dist = get_coeff_dist(abs_tqc, abs_dqc, shift);
1589 const int64_t rd = RDCOST(rdmult, rate, dist);
1590
1591 const tran_low_t abs_qc_low = abs_qc - 1;
1592 const tran_low_t abs_dqc_low = (abs_qc_low * dqv) >> shift;
1593 const int64_t dist_low = get_coeff_dist(abs_tqc, abs_dqc_low, shift);
1594 const int64_t rd_low = RDCOST(rdmult, rate_low, dist_low);
1595
1596 if (rd_low < rd) {
1597 const int sign = (qc < 0) ? 1 : 0;
1598 qcoeff[ci] = (-sign ^ abs_qc_low) + sign;
1599 dqcoeff[ci] = (-sign ^ abs_dqc_low) + sign;
1600 levels[get_padded_idx(ci, bwl)] = AOMMIN(abs_qc_low, INT8_MAX);
1601 *accu_rate += rate_low;
1602 } else {
1603 *accu_rate += rate;
1604 }
1605 }
1606 }
1607
1608 static AOM_FORCE_INLINE void update_coeff_eob(
1609 int *accu_rate, int64_t *accu_dist, int *eob, int *nz_num, int *nz_ci,
1610 int si, TX_SIZE tx_size, TX_CLASS tx_class, int bwl, int height,
1611 int dc_sign_ctx, int64_t rdmult, int shift, const int16_t *dequant,
1612 const int16_t *scan, const LV_MAP_EOB_COST *txb_eob_costs,
1613 const LV_MAP_COEFF_COST *txb_costs, const tran_low_t *tcoeff,
1614 tran_low_t *qcoeff, tran_low_t *dqcoeff, uint8_t *levels, int sharpness,
1615 const qm_val_t *iqmatrix) {
1616 const int dqv = get_dqv(dequant, scan[si], iqmatrix);
1617 assert(si != *eob - 1);
1618 const int ci = scan[si];
1619 const tran_low_t qc = qcoeff[ci];
1620 const int coeff_ctx =
1621 get_lower_levels_ctx(levels, ci, bwl, tx_size, tx_class);
1622 if (qc == 0) {
1623 *accu_rate += txb_costs->base_cost[coeff_ctx][0];
1624 } else {
1625 int lower_level = 0;
1626 const tran_low_t abs_qc = abs(qc);
1627 const tran_low_t tqc = tcoeff[ci];
1628 const tran_low_t dqc = dqcoeff[ci];
1629 const int sign = (qc < 0) ? 1 : 0;
1630 const int64_t dist0 = get_coeff_dist(tqc, 0, shift);
1631 int64_t dist = get_coeff_dist(tqc, dqc, shift) - dist0;
1632 int rate =
1633 get_coeff_cost_general(0, ci, abs_qc, sign, coeff_ctx, dc_sign_ctx,
1634 txb_costs, bwl, tx_class, levels);
1635 int64_t rd = RDCOST(rdmult, *accu_rate + rate, *accu_dist + dist);
1636
1637 tran_low_t qc_low, dqc_low;
1638 tran_low_t abs_qc_low;
1639 int64_t dist_low, rd_low;
1640 int rate_low;
1641 if (abs_qc == 1) {
1642 abs_qc_low = 0;
1643 dqc_low = qc_low = 0;
1644 dist_low = 0;
1645 rate_low = txb_costs->base_cost[coeff_ctx][0];
1646 rd_low = RDCOST(rdmult, *accu_rate + rate_low, *accu_dist);
1647 } else {
1648 get_qc_dqc_low(abs_qc, sign, dqv, shift, &qc_low, &dqc_low);
1649 abs_qc_low = abs_qc - 1;
1650 dist_low = get_coeff_dist(tqc, dqc_low, shift) - dist0;
1651 rate_low =
1652 get_coeff_cost_general(0, ci, abs_qc_low, sign, coeff_ctx,
1653 dc_sign_ctx, txb_costs, bwl, tx_class, levels);
1654 rd_low = RDCOST(rdmult, *accu_rate + rate_low, *accu_dist + dist_low);
1655 }
1656
1657 int lower_level_new_eob = 0;
1658 const int new_eob = si + 1;
1659 const int coeff_ctx_new_eob = get_lower_levels_ctx_eob(bwl, height, si);
1660 const int new_eob_cost =
1661 get_eob_cost(new_eob, txb_eob_costs, txb_costs, tx_class);
1662 int rate_coeff_eob =
1663 new_eob_cost + get_coeff_cost_eob(ci, abs_qc, sign, coeff_ctx_new_eob,
1664 dc_sign_ctx, txb_costs, bwl,
1665 tx_class);
1666 int64_t dist_new_eob = dist;
1667 int64_t rd_new_eob = RDCOST(rdmult, rate_coeff_eob, dist_new_eob);
1668
1669 if (abs_qc_low > 0) {
1670 const int rate_coeff_eob_low =
1671 new_eob_cost + get_coeff_cost_eob(ci, abs_qc_low, sign,
1672 coeff_ctx_new_eob, dc_sign_ctx,
1673 txb_costs, bwl, tx_class);
1674 const int64_t dist_new_eob_low = dist_low;
1675 const int64_t rd_new_eob_low =
1676 RDCOST(rdmult, rate_coeff_eob_low, dist_new_eob_low);
1677 if (rd_new_eob_low < rd_new_eob) {
1678 lower_level_new_eob = 1;
1679 rd_new_eob = rd_new_eob_low;
1680 rate_coeff_eob = rate_coeff_eob_low;
1681 dist_new_eob = dist_new_eob_low;
1682 }
1683 }
1684
1685 if (rd_low < rd) {
1686 lower_level = 1;
1687 rd = rd_low;
1688 rate = rate_low;
1689 dist = dist_low;
1690 }
1691
1692 if (sharpness == 0 && rd_new_eob < rd) {
1693 for (int ni = 0; ni < *nz_num; ++ni) {
1694 int last_ci = nz_ci[ni];
1695 levels[get_padded_idx(last_ci, bwl)] = 0;
1696 qcoeff[last_ci] = 0;
1697 dqcoeff[last_ci] = 0;
1698 }
1699 *eob = new_eob;
1700 *nz_num = 0;
1701 *accu_rate = rate_coeff_eob;
1702 *accu_dist = dist_new_eob;
1703 lower_level = lower_level_new_eob;
1704 } else {
1705 *accu_rate += rate;
1706 *accu_dist += dist;
1707 }
1708
1709 if (lower_level) {
1710 qcoeff[ci] = qc_low;
1711 dqcoeff[ci] = dqc_low;
1712 levels[get_padded_idx(ci, bwl)] = AOMMIN(abs_qc_low, INT8_MAX);
1713 }
1714 if (qcoeff[ci]) {
1715 nz_ci[*nz_num] = ci;
1716 ++*nz_num;
1717 }
1718 }
1719 }
1720
1721 static INLINE void update_skip(int *accu_rate, int64_t accu_dist, int *eob,
1722 int nz_num, int *nz_ci, int64_t rdmult,
1723 int skip_cost, int non_skip_cost,
1724 tran_low_t *qcoeff, tran_low_t *dqcoeff,
1725 int sharpness) {
1726 const int64_t rd = RDCOST(rdmult, *accu_rate + non_skip_cost, accu_dist);
1727 const int64_t rd_new_eob = RDCOST(rdmult, skip_cost, 0);
1728 if (sharpness == 0 && rd_new_eob < rd) {
1729 for (int i = 0; i < nz_num; ++i) {
1730 const int ci = nz_ci[i];
1731 qcoeff[ci] = 0;
1732 dqcoeff[ci] = 0;
1733 // no need to set up levels because this is the last step
1734 // levels[get_padded_idx(ci, bwl)] = 0;
1735 }
1736 *accu_rate = 0;
1737 *eob = 0;
1738 }
1739 }
1740
1741 int av1_optimize_txb_new(const struct AV1_COMP *cpi, MACROBLOCK *x, int plane,
1742 int block, TX_SIZE tx_size, TX_TYPE tx_type,
1743 const TXB_CTX *const txb_ctx, int *rate_cost,
1744 int sharpness, int fast_mode) {
1745 MACROBLOCKD *xd = &x->e_mbd;
1746 struct macroblockd_plane *pd = &xd->plane[plane];
1747 const struct macroblock_plane *p = &x->plane[plane];
1748 const SCAN_ORDER *scan_order = get_scan(tx_size, tx_type);
1749 const int16_t *scan = scan_order->scan;
1750 const int shift = av1_get_tx_scale(tx_size);
1751 int eob = p->eobs[block];
1752 const int16_t *dequant = p->dequant_QTX;
1753 const qm_val_t *iqmatrix =
1754 av1_get_iqmatrix(&cpi->common.quant_params, xd, plane, tx_size, tx_type);
1755 const int block_offset = BLOCK_OFFSET(block);
1756 tran_low_t *qcoeff = p->qcoeff + block_offset;
1757 tran_low_t *dqcoeff = pd->dqcoeff + block_offset;
1758 const tran_low_t *tcoeff = p->coeff + block_offset;
1759
1760 // This function is not called if eob = 0.
1761 assert(eob > 0);
1762
1763 if (fast_mode) {
1764 update_coeff_eob_fast(&eob, shift, dequant, scan, tcoeff, qcoeff, dqcoeff);
1765 p->eobs[block] = eob;
1766 if (eob == 0) {
1767 *rate_cost = av1_cost_skip_txb(x, txb_ctx, plane, tx_size);
1768 return eob;
1769 }
1770 }
1771
1772 const AV1_COMMON *cm = &cpi->common;
1773 const PLANE_TYPE plane_type = get_plane_type(plane);
1774 const TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
1775 const TX_CLASS tx_class = tx_type_to_class[tx_type];
1776 const MB_MODE_INFO *mbmi = xd->mi[0];
1777 const int bwl = get_txb_bwl(tx_size);
1778 const int width = get_txb_wide(tx_size);
1779 const int height = get_txb_high(tx_size);
1780 assert(width == (1 << bwl));
1781 const int is_inter = is_inter_block(mbmi);
1782 const LV_MAP_COEFF_COST *txb_costs = &x->coeff_costs[txs_ctx][plane_type];
1783 const int eob_multi_size = txsize_log2_minus4[tx_size];
1784 const LV_MAP_EOB_COST *txb_eob_costs =
1785 &x->eob_costs[eob_multi_size][plane_type];
1786
1787 const int rshift =
1788 (sharpness +
1789 (cpi->oxcf.aq_mode == VARIANCE_AQ && mbmi->segment_id < 4
1790 ? 7 - mbmi->segment_id
1791 : 2) +
1792 (cpi->oxcf.aq_mode != VARIANCE_AQ &&
1793 cpi->oxcf.deltaq_mode == DELTA_Q_PERCEPTUAL &&
1794 cm->delta_q_info.delta_q_present_flag && x->sb_energy_level < 0
1795 ? (3 - x->sb_energy_level)
1796 : 0));
1797 const int64_t rdmult =
1798 (((int64_t)x->rdmult *
1799 (plane_rd_mult[is_inter][plane_type] << (2 * (xd->bd - 8)))) +
1800 2) >>
1801 rshift;
1802
1803 uint8_t levels_buf[TX_PAD_2D];
1804 uint8_t *const levels = set_levels(levels_buf, width);
1805
1806 if (eob > 1) av1_txb_init_levels(qcoeff, width, height, levels);
1807
1808 // TODO(angirbird): check iqmatrix
1809
1810 const int non_skip_cost = txb_costs->txb_skip_cost[txb_ctx->txb_skip_ctx][0];
1811 const int skip_cost = txb_costs->txb_skip_cost[txb_ctx->txb_skip_ctx][1];
1812 const int eob_cost = get_eob_cost(eob, txb_eob_costs, txb_costs, tx_class);
1813 int accu_rate = eob_cost;
1814 int64_t accu_dist = 0;
1815 int si = eob - 1;
1816 const int ci = scan[si];
1817 const tran_low_t qc = qcoeff[ci];
1818 const tran_low_t abs_qc = abs(qc);
1819 const int sign = qc < 0;
1820 const int max_nz_num = 2;
1821 int nz_num = 1;
1822 int nz_ci[3] = { ci, 0, 0 };
1823 if (abs_qc >= 2) {
1824 update_coeff_general(&accu_rate, &accu_dist, si, eob, tx_size, tx_class,
1825 bwl, height, rdmult, shift, txb_ctx->dc_sign_ctx,
1826 dequant, scan, txb_costs, tcoeff, qcoeff, dqcoeff,
1827 levels, iqmatrix);
1828 --si;
1829 } else {
1830 assert(abs_qc == 1);
1831 const int coeff_ctx = get_lower_levels_ctx_eob(bwl, height, si);
1832 accu_rate +=
1833 get_coeff_cost_eob(ci, abs_qc, sign, coeff_ctx, txb_ctx->dc_sign_ctx,
1834 txb_costs, bwl, tx_class);
1835 const tran_low_t tqc = tcoeff[ci];
1836 const tran_low_t dqc = dqcoeff[ci];
1837 const int64_t dist = get_coeff_dist(tqc, dqc, shift);
1838 const int64_t dist0 = get_coeff_dist(tqc, 0, shift);
1839 accu_dist += dist - dist0;
1840 --si;
1841 }
1842
1843 #define UPDATE_COEFF_EOB_CASE(tx_class_literal) \
1844 case tx_class_literal: \
1845 for (; si >= 0 && nz_num <= max_nz_num && !fast_mode; --si) { \
1846 update_coeff_eob(&accu_rate, &accu_dist, &eob, &nz_num, nz_ci, si, \
1847 tx_size, tx_class_literal, bwl, height, \
1848 txb_ctx->dc_sign_ctx, rdmult, shift, dequant, scan, \
1849 txb_eob_costs, txb_costs, tcoeff, qcoeff, dqcoeff, \
1850 levels, sharpness, iqmatrix); \
1851 } \
1852 break;
1853 switch (tx_class) {
1854 UPDATE_COEFF_EOB_CASE(TX_CLASS_2D);
1855 UPDATE_COEFF_EOB_CASE(TX_CLASS_HORIZ);
1856 UPDATE_COEFF_EOB_CASE(TX_CLASS_VERT);
1857 #undef UPDATE_COEFF_EOB_CASE
1858 default: assert(false);
1859 }
1860
1861 if (si == -1 && nz_num <= max_nz_num) {
1862 update_skip(&accu_rate, accu_dist, &eob, nz_num, nz_ci, rdmult, skip_cost,
1863 non_skip_cost, qcoeff, dqcoeff, sharpness);
1864 }
1865
1866 #define UPDATE_COEFF_SIMPLE_CASE(tx_class_literal) \
1867 case tx_class_literal: \
1868 for (; si >= 1; --si) { \
1869 update_coeff_simple(&accu_rate, si, eob, tx_size, tx_class_literal, bwl, \
1870 rdmult, shift, dequant, scan, txb_costs, tcoeff, \
1871 qcoeff, dqcoeff, levels, iqmatrix); \
1872 } \
1873 break;
1874 switch (tx_class) {
1875 UPDATE_COEFF_SIMPLE_CASE(TX_CLASS_2D);
1876 UPDATE_COEFF_SIMPLE_CASE(TX_CLASS_HORIZ);
1877 UPDATE_COEFF_SIMPLE_CASE(TX_CLASS_VERT);
1878 #undef UPDATE_COEFF_SIMPLE_CASE
1879 default: assert(false);
1880 }
1881
1882 // DC position
1883 if (si == 0) {
1884 // no need to update accu_dist because it's not used after this point
1885 int64_t dummy_dist = 0;
1886 update_coeff_general(&accu_rate, &dummy_dist, si, eob, tx_size, tx_class,
1887 bwl, height, rdmult, shift, txb_ctx->dc_sign_ctx,
1888 dequant, scan, txb_costs, tcoeff, qcoeff, dqcoeff,
1889 levels, iqmatrix);
1890 }
1891
1892 const int tx_type_cost = get_tx_type_cost(x, xd, plane, tx_size, tx_type,
1893 cm->features.reduced_tx_set_used);
1894 if (eob == 0)
1895 accu_rate += skip_cost;
1896 else
1897 accu_rate += non_skip_cost + tx_type_cost;
1898
1899 p->eobs[block] = eob;
1900 p->txb_entropy_ctx[block] =
1901 av1_get_txb_entropy_context(qcoeff, scan_order, p->eobs[block]);
1902
1903 *rate_cost = accu_rate;
1904 return eob;
1905 }
1906
1907 // This function is deprecated, but we keep it here because hash trellis
1908 // is not integrated with av1_optimize_txb_new yet
1909 int av1_optimize_txb(const struct AV1_COMP *cpi, MACROBLOCK *x, int plane,
1910 int blk_row, int blk_col, int block, TX_SIZE tx_size,
1911 TXB_CTX *txb_ctx, int fast_mode, int *rate_cost) {
1912 const AV1_COMMON *cm = &cpi->common;
1913 const int reduced_tx_set_used = cm->features.reduced_tx_set_used;
1914 MACROBLOCKD *const xd = &x->e_mbd;
1915 const PLANE_TYPE plane_type = get_plane_type(plane);
1916 const TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
1917 const TX_TYPE tx_type = av1_get_tx_type(xd, plane_type, blk_row, blk_col,
1918 tx_size, reduced_tx_set_used);
1919 const MB_MODE_INFO *mbmi = xd->mi[0];
1920 const struct macroblock_plane *p = &x->plane[plane];
1921 struct macroblockd_plane *pd = &xd->plane[plane];
1922 const int eob = p->eobs[block];
1923 const int block_offset = BLOCK_OFFSET(block);
1924 tran_low_t *qcoeff = p->qcoeff + block_offset;
1925 tran_low_t *dqcoeff = pd->dqcoeff + block_offset;
1926 const tran_low_t *tcoeff = p->coeff + block_offset;
1927 const int16_t *dequant = p->dequant_QTX;
1928 const int seg_eob = av1_get_max_eob(tx_size);
1929 const int bwl = get_txb_bwl(tx_size);
1930 const int width = get_txb_wide(tx_size);
1931 const int height = get_txb_high(tx_size);
1932 const int is_inter = is_inter_block(mbmi);
1933 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
1934 const LV_MAP_COEFF_COST *txb_costs = &x->coeff_costs[txs_ctx][plane_type];
1935 const int eob_multi_size = txsize_log2_minus4[tx_size];
1936 const LV_MAP_EOB_COST txb_eob_costs =
1937 x->eob_costs[eob_multi_size][plane_type];
1938
1939 const int shift = av1_get_tx_scale(tx_size);
1940 const int64_t rdmult =
1941 (((int64_t)x->rdmult * plane_rd_mult[is_inter][plane_type]
1942 << (2 * (xd->bd - 8))) +
1943 2) >>
1944 2;
1945 uint8_t levels_buf[TX_PAD_2D];
1946 uint8_t *const levels = set_levels(levels_buf, width);
1947 const qm_val_t *iqmatrix =
1948 av1_get_iqmatrix(&cpi->common.quant_params, xd, plane, tx_size, tx_type);
1949 assert(width == (1 << bwl));
1950 const int tx_type_cost =
1951 get_tx_type_cost(x, xd, plane, tx_size, tx_type, reduced_tx_set_used);
1952 TxbInfo txb_info = {
1953 qcoeff, levels, dqcoeff, tcoeff, dequant, shift, tx_size,
1954 txs_ctx, tx_type, bwl, width, height, eob, seg_eob,
1955 scan_order, txb_ctx, rdmult, iqmatrix, tx_type_cost,
1956 };
1957
1958 #if CONFIG_HTB_TRELLIS
1959 // Hash based trellis (hbt) speed feature: avoid expensive optimize_txb calls
1960 // by storing the coefficient deltas in a hash table.
1961 // Currently disabled in speedfeatures.c
1962 if (eob <= HBT_EOB && eob > 0 && cpi->sf.use_hash_based_trellis) {
1963 return hbt_create_hashes(&txb_info, txb_costs, &txb_eob_costs, p, block,
1964 fast_mode, rate_cost);
1965 }
1966 #else
1967 (void)fast_mode;
1968 #endif // CONFIG_HTB_TRELLIS
1969 av1_txb_init_levels(qcoeff, width, height, levels);
1970
1971 const int update =
1972 optimize_txb(&txb_info, txb_costs, &txb_eob_costs, rate_cost);
1973
1974 if (update) {
1975 p->eobs[block] = txb_info.eob;
1976 p->txb_entropy_ctx[block] =
1977 av1_get_txb_entropy_context(qcoeff, scan_order, txb_info.eob);
1978 }
1979 return txb_info.eob;
1980 }
1981
1982 int av1_get_txb_entropy_context(const tran_low_t *qcoeff,
1983 const SCAN_ORDER *scan_order, int eob) {
1984 const int16_t *const scan = scan_order->scan;
1985 int cul_level = 0;
1986 int c;
1987
1988 if (eob == 0) return 0;
1989 for (c = 0; c < eob; ++c) {
1990 cul_level += abs(qcoeff[scan[c]]);
1991 if (cul_level > COEFF_CONTEXT_MASK) break;
1992 }
1993
1994 cul_level = AOMMIN(COEFF_CONTEXT_MASK, cul_level);
1995 set_dc_sign(&cul_level, qcoeff[0]);
1996
1997 return cul_level;
1998 }
1999
2000 static void update_tx_type_count(const AV1_COMP *cpi, const AV1_COMMON *cm,
2001 MACROBLOCKD *xd, int blk_row, int blk_col,
2002 int plane, TX_SIZE tx_size,
2003 FRAME_COUNTS *counts,
2004 uint8_t allow_update_cdf) {
2005 MB_MODE_INFO *mbmi = xd->mi[0];
2006 int is_inter = is_inter_block(mbmi);
2007 const int reduced_tx_set_used = cm->features.reduced_tx_set_used;
2008 FRAME_CONTEXT *fc = xd->tile_ctx;
2009 #if !CONFIG_ENTROPY_STATS
2010 (void)counts;
2011 #endif // !CONFIG_ENTROPY_STATS
2012
2013 // Only y plane's tx_type is updated
2014 if (plane > 0) return;
2015 const TX_TYPE tx_type = av1_get_tx_type(xd, PLANE_TYPE_Y, blk_row, blk_col,
2016 tx_size, reduced_tx_set_used);
2017 if (is_inter) {
2018 if (cpi->oxcf.use_inter_dct_only) {
2019 assert(tx_type == DCT_DCT);
2020 }
2021 } else {
2022 if (cpi->oxcf.use_intra_dct_only) {
2023 assert(tx_type == DCT_DCT);
2024 } else if (cpi->oxcf.use_intra_default_tx_only) {
2025 const TX_TYPE default_type = get_default_tx_type(
2026 PLANE_TYPE_Y, xd, tx_size, cpi->is_screen_content_type);
2027 (void)default_type;
2028 assert(tx_type == default_type);
2029 }
2030 }
2031
2032 if (get_ext_tx_types(tx_size, is_inter, reduced_tx_set_used) > 1 &&
2033 cm->quant_params.base_qindex > 0 && !mbmi->skip &&
2034 !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
2035 const int eset = get_ext_tx_set(tx_size, is_inter, reduced_tx_set_used);
2036 if (eset > 0) {
2037 const TxSetType tx_set_type =
2038 av1_get_ext_tx_set_type(tx_size, is_inter, reduced_tx_set_used);
2039 if (is_inter) {
2040 if (allow_update_cdf) {
2041 update_cdf(fc->inter_ext_tx_cdf[eset][txsize_sqr_map[tx_size]],
2042 av1_ext_tx_ind[tx_set_type][tx_type],
2043 av1_num_ext_tx_set[tx_set_type]);
2044 }
2045 #if CONFIG_ENTROPY_STATS
2046 ++counts->inter_ext_tx[eset][txsize_sqr_map[tx_size]]
2047 [av1_ext_tx_ind[tx_set_type][tx_type]];
2048 #endif // CONFIG_ENTROPY_STATS
2049 } else {
2050 PREDICTION_MODE intra_dir;
2051 if (mbmi->filter_intra_mode_info.use_filter_intra)
2052 intra_dir = fimode_to_intradir[mbmi->filter_intra_mode_info
2053 .filter_intra_mode];
2054 else
2055 intra_dir = mbmi->mode;
2056 #if CONFIG_ENTROPY_STATS
2057 ++counts->intra_ext_tx[eset][txsize_sqr_map[tx_size]][intra_dir]
2058 [av1_ext_tx_ind[tx_set_type][tx_type]];
2059 #endif // CONFIG_ENTROPY_STATS
2060 if (allow_update_cdf) {
2061 update_cdf(
2062 fc->intra_ext_tx_cdf[eset][txsize_sqr_map[tx_size]][intra_dir],
2063 av1_ext_tx_ind[tx_set_type][tx_type],
2064 av1_num_ext_tx_set[tx_set_type]);
2065 }
2066 }
2067 }
2068 }
2069 }
2070
2071 void av1_update_and_record_txb_context(int plane, int block, int blk_row,
2072 int blk_col, BLOCK_SIZE plane_bsize,
2073 TX_SIZE tx_size, void *arg) {
2074 struct tokenize_b_args *const args = arg;
2075 const AV1_COMP *cpi = args->cpi;
2076 const AV1_COMMON *cm = &cpi->common;
2077 ThreadData *const td = args->td;
2078 MACROBLOCK *const x = &td->mb;
2079 MACROBLOCKD *const xd = &x->e_mbd;
2080 struct macroblock_plane *p = &x->plane[plane];
2081 struct macroblockd_plane *pd = &xd->plane[plane];
2082 const int eob = p->eobs[block];
2083 const int block_offset = BLOCK_OFFSET(block);
2084 tran_low_t *qcoeff = p->qcoeff + block_offset;
2085 const PLANE_TYPE plane_type = pd->plane_type;
2086 const TX_TYPE tx_type =
2087 av1_get_tx_type(xd, plane_type, blk_row, blk_col, tx_size,
2088 cm->features.reduced_tx_set_used);
2089 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
2090 tran_low_t *tcoeff;
2091 assert(args->dry_run != DRY_RUN_COSTCOEFFS);
2092 if (args->dry_run == OUTPUT_ENABLED) {
2093 MB_MODE_INFO *mbmi = xd->mi[0];
2094 TXB_CTX txb_ctx;
2095 get_txb_ctx(plane_bsize, tx_size, plane,
2096 pd->above_entropy_context + blk_col,
2097 pd->left_entropy_context + blk_row, &txb_ctx);
2098 const int bwl = get_txb_bwl(tx_size);
2099 const int width = get_txb_wide(tx_size);
2100 const int height = get_txb_high(tx_size);
2101 const uint8_t allow_update_cdf = args->allow_update_cdf;
2102 const TX_SIZE txsize_ctx = get_txsize_entropy_ctx(tx_size);
2103 FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
2104 #if CONFIG_ENTROPY_STATS
2105 int cdf_idx = cm->coef_cdf_category;
2106 ++td->counts->txb_skip[cdf_idx][txsize_ctx][txb_ctx.txb_skip_ctx][eob == 0];
2107 #endif // CONFIG_ENTROPY_STATS
2108 if (allow_update_cdf) {
2109 update_cdf(ec_ctx->txb_skip_cdf[txsize_ctx][txb_ctx.txb_skip_ctx],
2110 eob == 0, 2);
2111 }
2112
2113 CB_COEFF_BUFFER *cb_coef_buff = x->cb_coef_buff;
2114 const int txb_offset =
2115 x->mbmi_ext_frame->cb_offset / (TX_SIZE_W_MIN * TX_SIZE_H_MIN);
2116 uint16_t *eob_txb = cb_coef_buff->eobs[plane] + txb_offset;
2117 uint8_t *const entropy_ctx = cb_coef_buff->entropy_ctx[plane] + txb_offset;
2118 entropy_ctx[block] = txb_ctx.txb_skip_ctx;
2119 eob_txb[block] = eob;
2120
2121 if (eob == 0) {
2122 av1_set_entropy_contexts(xd, pd, plane, plane_bsize, tx_size, 0, blk_col,
2123 blk_row);
2124 return;
2125 }
2126 const int segment_id = mbmi->segment_id;
2127 const int seg_eob = av1_get_tx_eob(&cpi->common.seg, segment_id, tx_size);
2128 tran_low_t *tcoeff_txb =
2129 cb_coef_buff->tcoeff[plane] + x->mbmi_ext_frame->cb_offset;
2130 tcoeff = tcoeff_txb + block_offset;
2131 memcpy(tcoeff, qcoeff, sizeof(*tcoeff) * seg_eob);
2132
2133 uint8_t levels_buf[TX_PAD_2D];
2134 uint8_t *const levels = set_levels(levels_buf, width);
2135 av1_txb_init_levels(tcoeff, width, height, levels);
2136 update_tx_type_count(cpi, cm, xd, blk_row, blk_col, plane, tx_size,
2137 td->counts, allow_update_cdf);
2138
2139 const TX_CLASS tx_class = tx_type_to_class[tx_type];
2140 const int16_t *const scan = scan_order->scan;
2141
2142 // record tx type usage
2143 td->rd_counts.tx_type_used[tx_size][tx_type]++;
2144
2145 #if CONFIG_ENTROPY_STATS
2146 av1_update_eob_context(cdf_idx, eob, tx_size, tx_class, plane_type, ec_ctx,
2147 td->counts, allow_update_cdf);
2148 #else
2149 av1_update_eob_context(eob, tx_size, tx_class, plane_type, ec_ctx,
2150 allow_update_cdf);
2151 #endif
2152
2153 DECLARE_ALIGNED(16, int8_t, coeff_contexts[MAX_TX_SQUARE]);
2154 av1_get_nz_map_contexts(levels, scan, eob, tx_size, tx_class,
2155 coeff_contexts);
2156
2157 for (int c = eob - 1; c >= 0; --c) {
2158 const int pos = scan[c];
2159 const int coeff_ctx = coeff_contexts[pos];
2160 const tran_low_t v = qcoeff[pos];
2161 const tran_low_t level = abs(v);
2162
2163 if (allow_update_cdf) {
2164 if (c == eob - 1) {
2165 assert(coeff_ctx < 4);
2166 update_cdf(
2167 ec_ctx->coeff_base_eob_cdf[txsize_ctx][plane_type][coeff_ctx],
2168 AOMMIN(level, 3) - 1, 3);
2169 } else {
2170 update_cdf(ec_ctx->coeff_base_cdf[txsize_ctx][plane_type][coeff_ctx],
2171 AOMMIN(level, 3), 4);
2172 }
2173 }
2174 if (c == eob - 1) {
2175 assert(coeff_ctx < 4);
2176 #if CONFIG_ENTROPY_STATS
2177 ++td->counts->coeff_base_eob_multi[cdf_idx][txsize_ctx][plane_type]
2178 [coeff_ctx][AOMMIN(level, 3) - 1];
2179 } else {
2180 ++td->counts->coeff_base_multi[cdf_idx][txsize_ctx][plane_type]
2181 [coeff_ctx][AOMMIN(level, 3)];
2182 #endif
2183 }
2184 if (level > NUM_BASE_LEVELS) {
2185 const int base_range = level - 1 - NUM_BASE_LEVELS;
2186 const int br_ctx = get_br_ctx(levels, pos, bwl, tx_class);
2187 for (int idx = 0; idx < COEFF_BASE_RANGE; idx += BR_CDF_SIZE - 1) {
2188 const int k = AOMMIN(base_range - idx, BR_CDF_SIZE - 1);
2189 if (allow_update_cdf) {
2190 update_cdf(ec_ctx->coeff_br_cdf[AOMMIN(txsize_ctx, TX_32X32)]
2191 [plane_type][br_ctx],
2192 k, BR_CDF_SIZE);
2193 }
2194 for (int lps = 0; lps < BR_CDF_SIZE - 1; lps++) {
2195 #if CONFIG_ENTROPY_STATS
2196 ++td->counts->coeff_lps[AOMMIN(txsize_ctx, TX_32X32)][plane_type]
2197 [lps][br_ctx][lps == k];
2198 #endif // CONFIG_ENTROPY_STATS
2199 if (lps == k) break;
2200 }
2201 #if CONFIG_ENTROPY_STATS
2202 ++td->counts->coeff_lps_multi[cdf_idx][AOMMIN(txsize_ctx, TX_32X32)]
2203 [plane_type][br_ctx][k];
2204 #endif
2205 if (k < BR_CDF_SIZE - 1) break;
2206 }
2207 }
2208 }
2209 // Update the context needed to code the DC sign (if applicable)
2210 if (tcoeff[0] != 0) {
2211 const int dc_sign = (tcoeff[0] < 0) ? 1 : 0;
2212 const int dc_sign_ctx = txb_ctx.dc_sign_ctx;
2213 #if CONFIG_ENTROPY_STATS
2214 ++td->counts->dc_sign[plane_type][dc_sign_ctx][dc_sign];
2215 #endif // CONFIG_ENTROPY_STATS
2216 if (allow_update_cdf)
2217 update_cdf(ec_ctx->dc_sign_cdf[plane_type][dc_sign_ctx], dc_sign, 2);
2218 entropy_ctx[block] |= dc_sign_ctx << DC_SIGN_CTX_SHIFT;
2219 }
2220 } else {
2221 tcoeff = qcoeff;
2222 }
2223 const int cul_level = av1_get_txb_entropy_context(tcoeff, scan_order, eob);
2224 av1_set_entropy_contexts(xd, pd, plane, plane_bsize, tx_size, cul_level,
2225 blk_col, blk_row);
2226 }
2227
2228 void av1_update_txb_context(const AV1_COMP *cpi, ThreadData *td,
2229 RUN_TYPE dry_run, BLOCK_SIZE bsize,
2230 uint8_t allow_update_cdf) {
2231 const AV1_COMMON *const cm = &cpi->common;
2232 const int num_planes = av1_num_planes(cm);
2233 MACROBLOCK *const x = &td->mb;
2234 MACROBLOCKD *const xd = &x->e_mbd;
2235 MB_MODE_INFO *const mbmi = xd->mi[0];
2236 struct tokenize_b_args arg = { cpi, td, 0, allow_update_cdf, dry_run };
2237 if (mbmi->skip) {
2238 av1_reset_entropy_context(xd, bsize, num_planes);
2239 return;
2240 }
2241
2242 for (int plane = 0; plane < num_planes; ++plane) {
2243 if (plane && !xd->is_chroma_ref) break;
2244 const struct macroblockd_plane *const pd = &xd->plane[plane];
2245 const int ss_x = pd->subsampling_x;
2246 const int ss_y = pd->subsampling_y;
2247 const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, ss_x, ss_y);
2248 av1_foreach_transformed_block_in_plane(
2249 xd, plane_bsize, plane, av1_update_and_record_txb_context, &arg);
2250 }
2251 }
2252
2253 CB_COEFF_BUFFER *av1_get_cb_coeff_buffer(const struct AV1_COMP *cpi, int mi_row,
2254 int mi_col) {
2255 const AV1_COMMON *const cm = &cpi->common;
2256 const int mib_size_log2 = cm->seq_params.mib_size_log2;
2257 const int stride = (cm->mi_params.mi_cols >> mib_size_log2) + 1;
2258 const int offset =
2259 (mi_row >> mib_size_log2) * stride + (mi_col >> mib_size_log2);
2260 return cpi->coeff_buffer_base + offset;
2261 }
2262