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 static int hbt_needs_init = 1;
27 static CRC32C crc_calculator;
28 static const int HBT_EOB = 16; // also the length in opt_qcoeff
29 static const int HBT_TABLE_SIZE = 65536; // 16 bit: holds 65536 'arrays'
30 static const int HBT_ARRAY_LENGTH = 256; // 8 bit: 256 entries
31 // If removed in hbt_create_hashes or increased beyond int8_t, widen deltas type
32 static const int HBT_KICKOUT = 3;
33
34 typedef struct OptTxbQcoeff {
35 // Use larger type if larger/no kickout value is used in hbt_create_hashes
36 int8_t deltas[16];
37 uint32_t hbt_qc_hash;
38 uint32_t hbt_ctx_hash;
39 int init;
40 int rate_cost;
41 } OptTxbQcoeff;
42
43 OptTxbQcoeff *hbt_hash_table;
44
45 typedef struct LevelDownStats {
46 int update;
47 tran_low_t low_qc;
48 tran_low_t low_dqc;
49 int64_t dist0;
50 int rate;
51 int rate_low;
52 int64_t dist;
53 int64_t dist_low;
54 int64_t rd;
55 int64_t rd_low;
56 int64_t nz_rd;
57 int64_t rd_diff;
58 int cost_diff;
59 int64_t dist_diff;
60 int new_eob;
61 } LevelDownStats;
62
av1_alloc_txb_buf(AV1_COMP * cpi)63 void av1_alloc_txb_buf(AV1_COMP *cpi) {
64 AV1_COMMON *cm = &cpi->common;
65 int size = ((cm->mi_rows >> cm->seq_params.mib_size_log2) + 1) *
66 ((cm->mi_cols >> cm->seq_params.mib_size_log2) + 1);
67
68 av1_free_txb_buf(cpi);
69 // TODO(jingning): This should be further reduced.
70 CHECK_MEM_ERROR(cm, cpi->coeff_buffer_base,
71 aom_memalign(32, sizeof(*cpi->coeff_buffer_base) * size));
72 }
73
av1_free_txb_buf(AV1_COMP * cpi)74 void av1_free_txb_buf(AV1_COMP *cpi) { aom_free(cpi->coeff_buffer_base); }
75
av1_set_coeff_buffer(const AV1_COMP * const cpi,MACROBLOCK * const x,int mi_row,int mi_col)76 void av1_set_coeff_buffer(const AV1_COMP *const cpi, MACROBLOCK *const x,
77 int mi_row, int mi_col) {
78 const AV1_COMMON *const cm = &cpi->common;
79 int mib_size_log2 = cm->seq_params.mib_size_log2;
80 int stride = (cm->mi_cols >> mib_size_log2) + 1;
81 int offset = (mi_row >> mib_size_log2) * stride + (mi_col >> mib_size_log2);
82 x->mbmi_ext->cb_coef_buff = &cpi->coeff_buffer_base[offset];
83 x->mbmi_ext->cb_offset = x->cb_offset;
84 assert(x->cb_offset < (1 << num_pels_log2_lookup[cm->seq_params.sb_size]));
85 }
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 - k_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 (k_eob_offset_bits[eob_pt] > 0) {
238 int eob_ctx = eob_pt - 3;
239 int eob_shift = k_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 (k_eob_offset_bits[eob_pt] > 0) {
258 const int eob_ctx = eob_pt - 3;
259 const int eob_shift = k_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 = k_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, MACROBLOCKD *xd,
506 aom_writer *w, int blk_row, int blk_col, int plane,
507 TX_SIZE tx_size, const tran_low_t *tcoeff,
508 uint16_t eob, TXB_CTX *txb_ctx) {
509 const TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
510 FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
511 aom_write_symbol(w, eob == 0,
512 ec_ctx->txb_skip_cdf[txs_ctx][txb_ctx->txb_skip_ctx], 2);
513 if (eob == 0) return;
514 const PLANE_TYPE plane_type = get_plane_type(plane);
515 const TX_TYPE tx_type = av1_get_tx_type(plane_type, xd, blk_row, blk_col,
516 tx_size, cm->reduced_tx_set_used);
517 const TX_CLASS tx_class = tx_type_to_class[tx_type];
518 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
519 const int16_t *const scan = scan_order->scan;
520 int c;
521 const int bwl = get_txb_bwl(tx_size);
522 const int width = get_txb_wide(tx_size);
523 const int height = get_txb_high(tx_size);
524
525 uint8_t levels_buf[TX_PAD_2D];
526 uint8_t *const levels = set_levels(levels_buf, width);
527 DECLARE_ALIGNED(16, int8_t, coeff_contexts[MAX_TX_SQUARE]);
528 av1_txb_init_levels(tcoeff, width, height, levels);
529
530 av1_write_tx_type(cm, xd, blk_row, blk_col, plane, tx_size, w);
531
532 int eob_extra;
533 const int eob_pt = get_eob_pos_token(eob, &eob_extra);
534 const int eob_multi_size = txsize_log2_minus4[tx_size];
535 const int eob_multi_ctx = (tx_class == TX_CLASS_2D) ? 0 : 1;
536 switch (eob_multi_size) {
537 case 0:
538 aom_write_symbol(w, eob_pt - 1,
539 ec_ctx->eob_flag_cdf16[plane_type][eob_multi_ctx], 5);
540 break;
541 case 1:
542 aom_write_symbol(w, eob_pt - 1,
543 ec_ctx->eob_flag_cdf32[plane_type][eob_multi_ctx], 6);
544 break;
545 case 2:
546 aom_write_symbol(w, eob_pt - 1,
547 ec_ctx->eob_flag_cdf64[plane_type][eob_multi_ctx], 7);
548 break;
549 case 3:
550 aom_write_symbol(w, eob_pt - 1,
551 ec_ctx->eob_flag_cdf128[plane_type][eob_multi_ctx], 8);
552 break;
553 case 4:
554 aom_write_symbol(w, eob_pt - 1,
555 ec_ctx->eob_flag_cdf256[plane_type][eob_multi_ctx], 9);
556 break;
557 case 5:
558 aom_write_symbol(w, eob_pt - 1,
559 ec_ctx->eob_flag_cdf512[plane_type][eob_multi_ctx], 10);
560 break;
561 default:
562 aom_write_symbol(w, eob_pt - 1,
563 ec_ctx->eob_flag_cdf1024[plane_type][eob_multi_ctx], 11);
564 break;
565 }
566
567 const int eob_offset_bits = k_eob_offset_bits[eob_pt];
568 if (eob_offset_bits > 0) {
569 const int eob_ctx = eob_pt - 3;
570 int eob_shift = eob_offset_bits - 1;
571 int bit = (eob_extra & (1 << eob_shift)) ? 1 : 0;
572 aom_write_symbol(w, bit,
573 ec_ctx->eob_extra_cdf[txs_ctx][plane_type][eob_ctx], 2);
574 for (int i = 1; i < eob_offset_bits; i++) {
575 eob_shift = eob_offset_bits - 1 - i;
576 bit = (eob_extra & (1 << eob_shift)) ? 1 : 0;
577 aom_write_bit(w, bit);
578 }
579 }
580
581 av1_get_nz_map_contexts(levels, scan, eob, tx_size, tx_class, coeff_contexts);
582
583 for (c = eob - 1; c >= 0; --c) {
584 const int pos = scan[c];
585 const int coeff_ctx = coeff_contexts[pos];
586 const tran_low_t v = tcoeff[pos];
587 const tran_low_t level = abs(v);
588
589 if (c == eob - 1) {
590 aom_write_symbol(
591 w, AOMMIN(level, 3) - 1,
592 ec_ctx->coeff_base_eob_cdf[txs_ctx][plane_type][coeff_ctx], 3);
593 } else {
594 aom_write_symbol(w, AOMMIN(level, 3),
595 ec_ctx->coeff_base_cdf[txs_ctx][plane_type][coeff_ctx],
596 4);
597 }
598 if (level > NUM_BASE_LEVELS) {
599 // level is above 1.
600 const int base_range = level - 1 - NUM_BASE_LEVELS;
601 const int br_ctx = get_br_ctx(levels, pos, bwl, tx_class);
602 aom_cdf_prob *cdf =
603 ec_ctx->coeff_br_cdf[AOMMIN(txs_ctx, TX_32X32)][plane_type][br_ctx];
604 for (int idx = 0; idx < COEFF_BASE_RANGE; idx += BR_CDF_SIZE - 1) {
605 const int k = AOMMIN(base_range - idx, BR_CDF_SIZE - 1);
606 aom_write_symbol(w, k, cdf, BR_CDF_SIZE);
607 if (k < BR_CDF_SIZE - 1) break;
608 }
609 }
610 }
611
612 // Loop to code all signs in the transform block,
613 // starting with the sign of DC (if applicable)
614 for (c = 0; c < eob; ++c) {
615 const tran_low_t v = tcoeff[scan[c]];
616 const tran_low_t level = abs(v);
617 const int sign = (v < 0) ? 1 : 0;
618 if (level) {
619 if (c == 0) {
620 aom_write_symbol(
621 w, sign, ec_ctx->dc_sign_cdf[plane_type][txb_ctx->dc_sign_ctx], 2);
622 } else {
623 aom_write_bit(w, sign);
624 }
625 if (level > COEFF_BASE_RANGE + NUM_BASE_LEVELS)
626 write_golomb(w, level - COEFF_BASE_RANGE - 1 - NUM_BASE_LEVELS);
627 }
628 }
629 }
630
631 typedef struct encode_txb_args {
632 const AV1_COMMON *cm;
633 MACROBLOCK *x;
634 aom_writer *w;
635 } ENCODE_TXB_ARGS;
636
637 static void write_coeffs_txb_wrap(const AV1_COMMON *cm, MACROBLOCK *x,
638 aom_writer *w, int plane, int block,
639 int blk_row, int blk_col, TX_SIZE tx_size) {
640 MACROBLOCKD *xd = &x->e_mbd;
641 const int txb_offset =
642 x->mbmi_ext->cb_offset / (TX_SIZE_W_MIN * TX_SIZE_H_MIN);
643 tran_low_t *tcoeff_txb =
644 x->mbmi_ext->cb_coef_buff->tcoeff[plane] + x->mbmi_ext->cb_offset;
645 uint16_t *eob_txb = x->mbmi_ext->cb_coef_buff->eobs[plane] + txb_offset;
646 uint8_t *txb_skip_ctx_txb =
647 x->mbmi_ext->cb_coef_buff->txb_skip_ctx[plane] + txb_offset;
648 int *dc_sign_ctx_txb =
649 x->mbmi_ext->cb_coef_buff->dc_sign_ctx[plane] + txb_offset;
650 tran_low_t *tcoeff = BLOCK_OFFSET(tcoeff_txb, block);
651 uint16_t eob = eob_txb[block];
652 TXB_CTX txb_ctx = { txb_skip_ctx_txb[block], dc_sign_ctx_txb[block] };
653 av1_write_coeffs_txb(cm, xd, w, blk_row, blk_col, plane, tx_size, tcoeff, eob,
654 &txb_ctx);
655 }
656
657 void av1_write_coeffs_mb(const AV1_COMMON *const cm, MACROBLOCK *x, int mi_row,
658 int mi_col, aom_writer *w, BLOCK_SIZE bsize) {
659 MACROBLOCKD *xd = &x->e_mbd;
660 const int num_planes = av1_num_planes(cm);
661 int block[MAX_MB_PLANE] = { 0 };
662 int row, col;
663 assert(bsize == get_plane_block_size(bsize, xd->plane[0].subsampling_x,
664 xd->plane[0].subsampling_y));
665 const int max_blocks_wide = max_block_wide(xd, bsize, 0);
666 const int max_blocks_high = max_block_high(xd, bsize, 0);
667 const BLOCK_SIZE max_unit_bsize = BLOCK_64X64;
668 int mu_blocks_wide = block_size_wide[max_unit_bsize] >> tx_size_wide_log2[0];
669 int mu_blocks_high = block_size_high[max_unit_bsize] >> tx_size_high_log2[0];
670 mu_blocks_wide = AOMMIN(max_blocks_wide, mu_blocks_wide);
671 mu_blocks_high = AOMMIN(max_blocks_high, mu_blocks_high);
672
673 for (row = 0; row < max_blocks_high; row += mu_blocks_high) {
674 for (col = 0; col < max_blocks_wide; col += mu_blocks_wide) {
675 for (int plane = 0; plane < num_planes; ++plane) {
676 const struct macroblockd_plane *const pd = &xd->plane[plane];
677 if (!is_chroma_reference(mi_row, mi_col, bsize, pd->subsampling_x,
678 pd->subsampling_y))
679 continue;
680 const TX_SIZE tx_size = av1_get_tx_size(plane, xd);
681 const int stepr = tx_size_high_unit[tx_size];
682 const int stepc = tx_size_wide_unit[tx_size];
683 const int step = stepr * stepc;
684
685 const int unit_height = ROUND_POWER_OF_TWO(
686 AOMMIN(mu_blocks_high + row, max_blocks_high), pd->subsampling_y);
687 const int unit_width = ROUND_POWER_OF_TWO(
688 AOMMIN(mu_blocks_wide + col, max_blocks_wide), pd->subsampling_x);
689 for (int blk_row = row >> pd->subsampling_y; blk_row < unit_height;
690 blk_row += stepr) {
691 for (int blk_col = col >> pd->subsampling_x; blk_col < unit_width;
692 blk_col += stepc) {
693 write_coeffs_txb_wrap(cm, x, w, plane, block[plane], blk_row,
694 blk_col, tx_size);
695 block[plane] += step;
696 }
697 }
698 }
699 }
700 }
701 }
702
703 // TODO(angiebird): use this function whenever it's possible
704 static int get_tx_type_cost(const AV1_COMMON *cm, const MACROBLOCK *x,
705 const MACROBLOCKD *xd, int plane, TX_SIZE tx_size,
706 TX_TYPE tx_type) {
707 if (plane > 0) return 0;
708
709 const TX_SIZE square_tx_size = txsize_sqr_map[tx_size];
710
711 const MB_MODE_INFO *mbmi = xd->mi[0];
712 const int is_inter = is_inter_block(mbmi);
713 if (get_ext_tx_types(tx_size, is_inter, cm->reduced_tx_set_used) > 1 &&
714 !xd->lossless[xd->mi[0]->segment_id]) {
715 const int ext_tx_set =
716 get_ext_tx_set(tx_size, is_inter, cm->reduced_tx_set_used);
717 if (is_inter) {
718 if (ext_tx_set > 0)
719 return x->inter_tx_type_costs[ext_tx_set][square_tx_size][tx_type];
720 } else {
721 if (ext_tx_set > 0) {
722 PREDICTION_MODE intra_dir;
723 if (mbmi->filter_intra_mode_info.use_filter_intra)
724 intra_dir = fimode_to_intradir[mbmi->filter_intra_mode_info
725 .filter_intra_mode];
726 else
727 intra_dir = mbmi->mode;
728 return x->intra_tx_type_costs[ext_tx_set][square_tx_size][intra_dir]
729 [tx_type];
730 }
731 }
732 }
733 return 0;
734 }
735
736 static AOM_FORCE_INLINE int warehouse_efficients_txb(
737 const AV1_COMMON *const cm, const MACROBLOCK *x, const int plane,
738 const int block, const TX_SIZE tx_size, const TXB_CTX *const txb_ctx,
739 const struct macroblock_plane *p, const int eob,
740 const PLANE_TYPE plane_type, const LV_MAP_COEFF_COST *const coeff_costs,
741 const MACROBLOCKD *const xd, const TX_TYPE tx_type,
742 const TX_CLASS tx_class) {
743 const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
744 const int txb_skip_ctx = txb_ctx->txb_skip_ctx;
745 const int bwl = get_txb_bwl(tx_size);
746 const int width = get_txb_wide(tx_size);
747 const int height = get_txb_high(tx_size);
748 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
749 const int16_t *const scan = scan_order->scan;
750 uint8_t levels_buf[TX_PAD_2D];
751 uint8_t *const levels = set_levels(levels_buf, width);
752 DECLARE_ALIGNED(16, int8_t, coeff_contexts[MAX_TX_SQUARE]);
753 const int eob_multi_size = txsize_log2_minus4[tx_size];
754 const LV_MAP_EOB_COST *const eob_costs =
755 &x->eob_costs[eob_multi_size][plane_type];
756 int cost = coeff_costs->txb_skip_cost[txb_skip_ctx][0];
757
758 av1_txb_init_levels(qcoeff, width, height, levels);
759
760 cost += get_tx_type_cost(cm, x, xd, plane, tx_size, tx_type);
761
762 cost += get_eob_cost(eob, eob_costs, coeff_costs, tx_class);
763
764 av1_get_nz_map_contexts(levels, scan, eob, tx_size, tx_class, coeff_contexts);
765
766 const int(*lps_cost)[COEFF_BASE_RANGE + 1 + COEFF_BASE_RANGE + 1] =
767 coeff_costs->lps_cost;
768 int c = eob - 1;
769 {
770 const int pos = scan[c];
771 const tran_low_t v = qcoeff[pos];
772 const int sign = v >> 31;
773 const int level = (v ^ sign) - sign;
774 const int coeff_ctx = coeff_contexts[pos];
775 cost += coeff_costs->base_eob_cost[coeff_ctx][AOMMIN(level, 3) - 1];
776
777 if (v) {
778 // sign bit cost
779 if (level > NUM_BASE_LEVELS) {
780 const int ctx = get_br_ctx_eob(pos, bwl, tx_class);
781 cost += get_br_cost(level, lps_cost[ctx]);
782 }
783 if (c) {
784 cost += av1_cost_literal(1);
785 } else {
786 const int sign01 = (sign ^ sign) - sign;
787 const int dc_sign_ctx = txb_ctx->dc_sign_ctx;
788 cost += coeff_costs->dc_sign_cost[dc_sign_ctx][sign01];
789 return cost;
790 }
791 }
792 }
793 const int(*base_cost)[8] = coeff_costs->base_cost;
794 for (c = eob - 2; c >= 1; --c) {
795 const int pos = scan[c];
796 const int coeff_ctx = coeff_contexts[pos];
797 const tran_low_t v = qcoeff[pos];
798 const int level = abs(v);
799 const int cost0 = base_cost[coeff_ctx][AOMMIN(level, 3)];
800 if (v) {
801 // sign bit cost
802 cost += av1_cost_literal(1);
803 if (level > NUM_BASE_LEVELS) {
804 const int ctx = get_br_ctx(levels, pos, bwl, tx_class);
805 cost += get_br_cost(level, lps_cost[ctx]);
806 }
807 }
808 cost += cost0;
809 }
810 if (c == 0) {
811 const int pos = scan[c];
812 const tran_low_t v = qcoeff[pos];
813 const int coeff_ctx = coeff_contexts[pos];
814 const int sign = v >> 31;
815 const int level = (v ^ sign) - sign;
816 cost += base_cost[coeff_ctx][AOMMIN(level, 3)];
817
818 if (v) {
819 // sign bit cost
820 const int sign01 = (sign ^ sign) - sign;
821 const int dc_sign_ctx = txb_ctx->dc_sign_ctx;
822 cost += coeff_costs->dc_sign_cost[dc_sign_ctx][sign01];
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 return cost;
830 }
831
832 int av1_cost_coeffs_txb(const AV1_COMMON *const cm, const MACROBLOCK *x,
833 const int plane, const int block, const TX_SIZE tx_size,
834 const TX_TYPE tx_type, const TXB_CTX *const txb_ctx) {
835 const struct macroblock_plane *p = &x->plane[plane];
836 const int eob = p->eobs[block];
837 const TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
838 const PLANE_TYPE plane_type = get_plane_type(plane);
839 const LV_MAP_COEFF_COST *const coeff_costs =
840 &x->coeff_costs[txs_ctx][plane_type];
841 if (eob == 0) {
842 return coeff_costs->txb_skip_cost[txb_ctx->txb_skip_ctx][1];
843 }
844
845 const MACROBLOCKD *const xd = &x->e_mbd;
846 const TX_CLASS tx_class = tx_type_to_class[tx_type];
847
848 #define WAREHOUSE_EFFICIENTS_TXB_CASE(tx_class_literal) \
849 case tx_class_literal: \
850 return warehouse_efficients_txb(cm, x, plane, block, tx_size, txb_ctx, p, \
851 eob, plane_type, coeff_costs, xd, tx_type, \
852 tx_class_literal);
853 switch (tx_class) {
854 WAREHOUSE_EFFICIENTS_TXB_CASE(TX_CLASS_2D);
855 WAREHOUSE_EFFICIENTS_TXB_CASE(TX_CLASS_HORIZ);
856 WAREHOUSE_EFFICIENTS_TXB_CASE(TX_CLASS_VERT);
857 #undef WAREHOUSE_EFFICIENTS_TXB_CASE
858 default: assert(false); return 0;
859 }
860 }
861
862 static int optimize_txb(TxbInfo *txb_info, const LV_MAP_COEFF_COST *txb_costs,
863 const LV_MAP_EOB_COST *txb_eob_costs, int *rate_cost) {
864 int update = 0;
865 if (txb_info->eob == 0) return update;
866 const int16_t *const scan = txb_info->scan_order->scan;
867 // forward optimize the nz_map`
868 const int init_eob = txb_info->eob;
869 const TX_CLASS tx_class = tx_type_to_class[txb_info->tx_type];
870 const int eob_cost =
871 get_eob_cost(init_eob, txb_eob_costs, txb_costs, tx_class);
872
873 // backward optimize the level-k map
874 int accu_rate = eob_cost;
875 int64_t accu_dist = 0;
876 int64_t prev_eob_rd_cost = INT64_MAX;
877 int64_t cur_eob_rd_cost = 0;
878
879 {
880 const int si = init_eob - 1;
881 const int coeff_idx = scan[si];
882 LevelDownStats stats;
883 get_dist_cost_stats(&stats, si, si == init_eob - 1, txb_costs, txb_info,
884 tx_class);
885 if ((stats.rd_low < stats.rd) && (stats.low_qc != 0)) {
886 update = 1;
887 update_coeff(coeff_idx, stats.low_qc, txb_info);
888 accu_rate += stats.rate_low;
889 accu_dist += stats.dist_low;
890 } else {
891 accu_rate += stats.rate;
892 accu_dist += stats.dist;
893 }
894 }
895
896 int si = init_eob - 2;
897 int8_t has_nz_tail = 0;
898 // eob is not fixed
899 for (; si >= 0 && has_nz_tail < 2; --si) {
900 assert(si != init_eob - 1);
901 const int coeff_idx = scan[si];
902 tran_low_t qc = txb_info->qcoeff[coeff_idx];
903
904 if (qc == 0) {
905 const int coeff_ctx =
906 get_lower_levels_ctx(txb_info->levels, coeff_idx, txb_info->bwl,
907 txb_info->tx_size, tx_class);
908 accu_rate += txb_costs->base_cost[coeff_ctx][0];
909 } else {
910 LevelDownStats stats;
911 get_dist_cost_stats_with_eob(&stats, si, txb_costs, txb_info, tx_class);
912 // check if it is better to make this the last significant coefficient
913 int cur_eob_rate =
914 get_eob_cost(si + 1, txb_eob_costs, txb_costs, tx_class);
915 cur_eob_rd_cost = RDCOST(txb_info->rdmult, cur_eob_rate, 0);
916 prev_eob_rd_cost =
917 RDCOST(txb_info->rdmult, accu_rate, accu_dist) + stats.nz_rd;
918 if (cur_eob_rd_cost <= prev_eob_rd_cost) {
919 update = 1;
920 for (int j = si + 1; j < txb_info->eob; j++) {
921 const int coeff_pos_j = scan[j];
922 update_coeff(coeff_pos_j, 0, txb_info);
923 }
924 txb_info->eob = si + 1;
925
926 // rerun cost calculation due to change of eob
927 accu_rate = cur_eob_rate;
928 accu_dist = 0;
929 get_dist_cost_stats(&stats, si, 1, txb_costs, txb_info, tx_class);
930 if ((stats.rd_low < stats.rd) && (stats.low_qc != 0)) {
931 update = 1;
932 update_coeff(coeff_idx, stats.low_qc, txb_info);
933 accu_rate += stats.rate_low;
934 accu_dist += stats.dist_low;
935 } else {
936 accu_rate += stats.rate;
937 accu_dist += stats.dist;
938 }
939
940 // reset non zero tail when new eob is found
941 has_nz_tail = 0;
942 } else {
943 int bUpdCoeff = 0;
944 if (stats.rd_low < stats.rd) {
945 if ((si < txb_info->eob - 1)) {
946 bUpdCoeff = 1;
947 update = 1;
948 }
949 } else {
950 ++has_nz_tail;
951 }
952
953 if (bUpdCoeff) {
954 update_coeff(coeff_idx, stats.low_qc, txb_info);
955 accu_rate += stats.rate_low;
956 accu_dist += stats.dist_low;
957 } else {
958 accu_rate += stats.rate;
959 accu_dist += stats.dist;
960 }
961 }
962 }
963 } // for (si)
964
965 // eob is fixed
966 for (; si >= 0; --si) {
967 assert(si != init_eob - 1);
968 const int coeff_idx = scan[si];
969 tran_low_t qc = txb_info->qcoeff[coeff_idx];
970
971 if (qc == 0) {
972 const int coeff_ctx =
973 get_lower_levels_ctx(txb_info->levels, coeff_idx, txb_info->bwl,
974 txb_info->tx_size, tx_class);
975 accu_rate += txb_costs->base_cost[coeff_ctx][0];
976 } else {
977 LevelDownStats stats;
978 get_dist_cost_stats(&stats, si, 0, txb_costs, txb_info, tx_class);
979
980 int bUpdCoeff = 0;
981 if (stats.rd_low < stats.rd) {
982 if ((si < txb_info->eob - 1)) {
983 bUpdCoeff = 1;
984 update = 1;
985 }
986 }
987 if (bUpdCoeff) {
988 update_coeff(coeff_idx, stats.low_qc, txb_info);
989 accu_rate += stats.rate_low;
990 accu_dist += stats.dist_low;
991 } else {
992 accu_rate += stats.rate;
993 accu_dist += stats.dist;
994 }
995 }
996 } // for (si)
997
998 int non_zero_blk_rate =
999 txb_costs->txb_skip_cost[txb_info->txb_ctx->txb_skip_ctx][0];
1000 prev_eob_rd_cost =
1001 RDCOST(txb_info->rdmult, accu_rate + non_zero_blk_rate, accu_dist);
1002
1003 int zero_blk_rate =
1004 txb_costs->txb_skip_cost[txb_info->txb_ctx->txb_skip_ctx][1];
1005 int64_t zero_blk_rd_cost = RDCOST(txb_info->rdmult, zero_blk_rate, 0);
1006 if (zero_blk_rd_cost <= prev_eob_rd_cost) {
1007 update = 1;
1008 for (int j = 0; j < txb_info->eob; j++) {
1009 const int coeff_pos_j = scan[j];
1010 update_coeff(coeff_pos_j, 0, txb_info);
1011 }
1012 txb_info->eob = 0;
1013 }
1014
1015 // record total rate cost
1016 *rate_cost = zero_blk_rd_cost <= prev_eob_rd_cost
1017 ? zero_blk_rate
1018 : accu_rate + non_zero_blk_rate;
1019
1020 if (txb_info->eob > 0) {
1021 *rate_cost += txb_info->tx_type_cost;
1022 }
1023
1024 return update;
1025 }
1026
1027 static void hbt_init() {
1028 hbt_hash_table =
1029 aom_malloc(sizeof(OptTxbQcoeff) * HBT_TABLE_SIZE * HBT_ARRAY_LENGTH);
1030 memset(hbt_hash_table, 0,
1031 sizeof(OptTxbQcoeff) * HBT_TABLE_SIZE * HBT_ARRAY_LENGTH);
1032 av1_crc32c_calculator_init(&crc_calculator); // 31 bit: qc & ctx
1033
1034 hbt_needs_init = 0;
1035 }
1036
1037 void hbt_destroy() { aom_free(hbt_hash_table); }
1038
1039 static int hbt_hash_miss(uint32_t hbt_ctx_hash, uint32_t hbt_qc_hash,
1040 TxbInfo *txb_info, const LV_MAP_COEFF_COST *txb_costs,
1041 const LV_MAP_EOB_COST *txb_eob_costs,
1042 const struct macroblock_plane *p, int block,
1043 int fast_mode, int *rate_cost) {
1044 (void)fast_mode;
1045 const int16_t *scan = txb_info->scan_order->scan;
1046 int prev_eob = txb_info->eob;
1047 assert(HBT_EOB <= 16); // Lengthen array if allowing longer eob.
1048 int32_t prev_coeff[16];
1049 for (int i = 0; i < prev_eob; i++) {
1050 prev_coeff[i] = txb_info->qcoeff[scan[i]];
1051 }
1052 for (int i = prev_eob; i < HBT_EOB; i++) {
1053 prev_coeff[i] = 0; // For compiler piece of mind.
1054 }
1055
1056 av1_txb_init_levels(txb_info->qcoeff, txb_info->width, txb_info->height,
1057 txb_info->levels);
1058
1059 const int update =
1060 optimize_txb(txb_info, txb_costs, txb_eob_costs, rate_cost);
1061
1062 // Overwrite old entry
1063 uint16_t hbt_table_index = hbt_ctx_hash % HBT_TABLE_SIZE;
1064 uint16_t hbt_array_index = hbt_qc_hash % HBT_ARRAY_LENGTH;
1065 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1066 .rate_cost = *rate_cost;
1067 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index].init = 1;
1068 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1069 .hbt_qc_hash = hbt_qc_hash;
1070 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1071 .hbt_ctx_hash = hbt_ctx_hash;
1072 assert(prev_eob >= txb_info->eob); // eob can't get longer
1073 for (int i = 0; i < txb_info->eob; i++) {
1074 // Record how coeff changed. Convention: towards zero is negative.
1075 if (txb_info->qcoeff[scan[i]] > 0)
1076 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1077 .deltas[i] = txb_info->qcoeff[scan[i]] - prev_coeff[i];
1078 else
1079 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1080 .deltas[i] = prev_coeff[i] - txb_info->qcoeff[scan[i]];
1081 }
1082 for (int i = txb_info->eob; i < prev_eob; i++) {
1083 // If eob got shorter, record that all after it changed to zero.
1084 if (prev_coeff[i] > 0)
1085 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1086 .deltas[i] = -prev_coeff[i];
1087 else
1088 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1089 .deltas[i] = prev_coeff[i];
1090 }
1091 for (int i = prev_eob; i < HBT_EOB; i++) {
1092 // Record 'no change' after optimized coefficients run out.
1093 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1094 .deltas[i] = 0;
1095 }
1096
1097 if (update) {
1098 p->eobs[block] = txb_info->eob;
1099 p->txb_entropy_ctx[block] = av1_get_txb_entropy_context(
1100 txb_info->qcoeff, txb_info->scan_order, txb_info->eob);
1101 }
1102 return txb_info->eob;
1103 }
1104
1105 static int hbt_hash_hit(uint32_t hbt_table_index, int hbt_array_index,
1106 TxbInfo *txb_info, const struct macroblock_plane *p,
1107 int block, int *rate_cost) {
1108 const int16_t *scan = txb_info->scan_order->scan;
1109 int new_eob = 0;
1110 int update = 0;
1111
1112 for (int i = 0; i < txb_info->eob; i++) {
1113 // Delta convention is negatives go towards zero, so only apply those ones.
1114 if (hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1115 .deltas[i] < 0) {
1116 if (txb_info->qcoeff[scan[i]] > 0)
1117 txb_info->qcoeff[scan[i]] +=
1118 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1119 .deltas[i];
1120 else
1121 txb_info->qcoeff[scan[i]] -=
1122 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1123 .deltas[i];
1124
1125 update = 1;
1126 update_coeff(scan[i], txb_info->qcoeff[scan[i]], txb_info);
1127 }
1128 if (txb_info->qcoeff[scan[i]]) new_eob = i + 1;
1129 }
1130
1131 // Rate_cost can be calculated here instead (av1_cost_coeffs_txb), but
1132 // it is expensive and gives little benefit as long as qc_hash is high bit
1133 *rate_cost =
1134 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1135 .rate_cost;
1136
1137 if (update) {
1138 txb_info->eob = new_eob;
1139 p->eobs[block] = txb_info->eob;
1140 p->txb_entropy_ctx[block] = av1_get_txb_entropy_context(
1141 txb_info->qcoeff, txb_info->scan_order, txb_info->eob);
1142 }
1143
1144 return txb_info->eob;
1145 }
1146
1147 static int hbt_search_match(uint32_t hbt_ctx_hash, uint32_t hbt_qc_hash,
1148 TxbInfo *txb_info,
1149 const LV_MAP_COEFF_COST *txb_costs,
1150 const LV_MAP_EOB_COST *txb_eob_costs,
1151 const struct macroblock_plane *p, int block,
1152 int fast_mode, int *rate_cost) {
1153 // Check for qcoeff match
1154 int hbt_array_index = hbt_qc_hash % HBT_ARRAY_LENGTH;
1155 int hbt_table_index = hbt_ctx_hash % HBT_TABLE_SIZE;
1156
1157 if (hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1158 .hbt_qc_hash == hbt_qc_hash &&
1159 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1160 .hbt_ctx_hash == hbt_ctx_hash &&
1161 hbt_hash_table[hbt_table_index * HBT_ARRAY_LENGTH + hbt_array_index]
1162 .init) {
1163 return hbt_hash_hit(hbt_table_index, hbt_array_index, txb_info, p, block,
1164 rate_cost);
1165 } else {
1166 return hbt_hash_miss(hbt_ctx_hash, hbt_qc_hash, txb_info, txb_costs,
1167 txb_eob_costs, p, block, fast_mode, rate_cost);
1168 }
1169 }
1170
1171 static int hbt_create_hashes(TxbInfo *txb_info,
1172 const LV_MAP_COEFF_COST *txb_costs,
1173 const LV_MAP_EOB_COST *txb_eob_costs,
1174 const struct macroblock_plane *p, int block,
1175 int fast_mode, int *rate_cost) {
1176 // Initialize hash table if needed.
1177 if (hbt_needs_init) {
1178 hbt_init();
1179 }
1180
1181 //// Hash creation
1182 uint8_t txb_hash_data[256]; // Asserts below to ensure enough space.
1183 const int16_t *scan = txb_info->scan_order->scan;
1184 uint8_t chunk = 0;
1185 int hash_data_index = 0;
1186
1187 // Make qc_hash.
1188 int packing_index = 0; // needed for packing.
1189 for (int i = 0; i < txb_info->eob; i++) {
1190 tran_low_t prechunk = txb_info->qcoeff[scan[i]];
1191
1192 // Softening: Improves speed. Aligns with signed deltas.
1193 if (prechunk < 0) prechunk *= -1;
1194
1195 // Early kick out: Don't apply feature if there are large coeffs:
1196 // If this kickout value is removed or raised beyond int8_t,
1197 // widen deltas type in OptTxbQcoeff struct.
1198 assert((int8_t)HBT_KICKOUT == HBT_KICKOUT); // If not, widen types.
1199 if (prechunk > HBT_KICKOUT) {
1200 av1_txb_init_levels(txb_info->qcoeff, txb_info->width, txb_info->height,
1201 txb_info->levels);
1202
1203 const int update =
1204 optimize_txb(txb_info, txb_costs, txb_eob_costs, rate_cost);
1205
1206 if (update) {
1207 p->eobs[block] = txb_info->eob;
1208 p->txb_entropy_ctx[block] = av1_get_txb_entropy_context(
1209 txb_info->qcoeff, txb_info->scan_order, txb_info->eob);
1210 }
1211 return txb_info->eob;
1212 }
1213
1214 // Since coeffs are 0 to 3, only 2 bits are needed: pack into bytes
1215 if (packing_index == 0) txb_hash_data[hash_data_index] = 0;
1216 chunk = prechunk << packing_index;
1217 packing_index += 2;
1218 txb_hash_data[hash_data_index] |= chunk;
1219
1220 // Full byte:
1221 if (packing_index == 8) {
1222 packing_index = 0;
1223 hash_data_index++;
1224 }
1225 }
1226 // Needed when packing_index != 0, to include final byte.
1227 hash_data_index++;
1228 assert(hash_data_index <= 64);
1229 // 31 bit qc_hash: index to array
1230 uint32_t hbt_qc_hash =
1231 av1_get_crc32c_value(&crc_calculator, txb_hash_data, hash_data_index);
1232
1233 // Make ctx_hash.
1234 hash_data_index = 0;
1235 tran_low_t prechunk;
1236
1237 for (int i = 0; i < txb_info->eob; i++) {
1238 // Save as magnitudes towards or away from zero.
1239 if (txb_info->tcoeff[scan[i]] >= 0)
1240 prechunk = txb_info->tcoeff[scan[i]] - txb_info->dqcoeff[scan[i]];
1241 else
1242 prechunk = txb_info->dqcoeff[scan[i]] - txb_info->tcoeff[scan[i]];
1243
1244 chunk = prechunk & 0xff;
1245 txb_hash_data[hash_data_index++] = chunk;
1246 }
1247
1248 // Extra ctx data:
1249 // Include dequants.
1250 txb_hash_data[hash_data_index++] = txb_info->dequant[0] & 0xff;
1251 txb_hash_data[hash_data_index++] = txb_info->dequant[1] & 0xff;
1252 chunk = txb_info->txb_ctx->txb_skip_ctx & 0xff;
1253 txb_hash_data[hash_data_index++] = chunk;
1254 chunk = txb_info->txb_ctx->dc_sign_ctx & 0xff;
1255 txb_hash_data[hash_data_index++] = chunk;
1256 // eob
1257 chunk = txb_info->eob & 0xff;
1258 txb_hash_data[hash_data_index++] = chunk;
1259 // rdmult (int64)
1260 chunk = txb_info->rdmult & 0xff;
1261 txb_hash_data[hash_data_index++] = chunk;
1262 // tx_type
1263 chunk = txb_info->tx_type & 0xff;
1264 txb_hash_data[hash_data_index++] = chunk;
1265 // base_eob_cost
1266 for (int i = 1; i < 3; i++) { // i = 0 are softened away
1267 for (int j = 0; j < SIG_COEF_CONTEXTS_EOB; j++) {
1268 chunk = (txb_costs->base_eob_cost[j][i] & 0xff00) >> 8;
1269 txb_hash_data[hash_data_index++] = chunk;
1270 }
1271 }
1272 // eob_cost
1273 for (int i = 0; i < 11; i++) {
1274 for (int j = 0; j < 2; j++) {
1275 chunk = (txb_eob_costs->eob_cost[j][i] & 0xff00) >> 8;
1276 txb_hash_data[hash_data_index++] = chunk;
1277 }
1278 }
1279 // dc_sign_cost
1280 for (int i = 0; i < 2; i++) {
1281 for (int j = 0; j < DC_SIGN_CONTEXTS; j++) {
1282 chunk = (txb_costs->dc_sign_cost[j][i] & 0xff00) >> 8;
1283 txb_hash_data[hash_data_index++] = chunk;
1284 }
1285 }
1286
1287 assert(hash_data_index <= 256);
1288 // 31 bit ctx_hash: used to index table
1289 uint32_t hbt_ctx_hash =
1290 av1_get_crc32c_value(&crc_calculator, txb_hash_data, hash_data_index);
1291 //// End hash creation
1292
1293 return hbt_search_match(hbt_ctx_hash, hbt_qc_hash, txb_info, txb_costs,
1294 txb_eob_costs, p, block, fast_mode, rate_cost);
1295 }
1296
1297 static AOM_FORCE_INLINE int get_two_coeff_cost_simple(
1298 int ci, tran_low_t abs_qc, int coeff_ctx,
1299 const LV_MAP_COEFF_COST *txb_costs, int bwl, TX_CLASS tx_class,
1300 const uint8_t *levels, int *cost_low) {
1301 // this simple version assumes the coeff's scan_idx is not DC (scan_idx != 0)
1302 // and not the last (scan_idx != eob - 1)
1303 assert(ci > 0);
1304 int cost = txb_costs->base_cost[coeff_ctx][AOMMIN(abs_qc, 3)];
1305 int diff = 0;
1306 if (abs_qc <= 3) diff = txb_costs->base_cost[coeff_ctx][abs_qc + 4];
1307 if (abs_qc) {
1308 cost += av1_cost_literal(1);
1309 if (abs_qc > NUM_BASE_LEVELS) {
1310 const int br_ctx = get_br_ctx(levels, ci, bwl, tx_class);
1311 int brcost_diff = 0;
1312 cost += get_br_cost_with_diff(abs_qc, txb_costs->lps_cost[br_ctx],
1313 &brcost_diff);
1314 diff += brcost_diff;
1315 }
1316 }
1317 *cost_low = cost - diff;
1318
1319 return cost;
1320 }
1321
1322 static INLINE int get_coeff_cost_eob(int ci, tran_low_t abs_qc, int sign,
1323 int coeff_ctx, int dc_sign_ctx,
1324 const LV_MAP_COEFF_COST *txb_costs,
1325 int bwl, TX_CLASS tx_class) {
1326 int cost = 0;
1327 cost += txb_costs->base_eob_cost[coeff_ctx][AOMMIN(abs_qc, 3) - 1];
1328 if (abs_qc != 0) {
1329 if (ci == 0) {
1330 cost += txb_costs->dc_sign_cost[dc_sign_ctx][sign];
1331 } else {
1332 cost += av1_cost_literal(1);
1333 }
1334 if (abs_qc > NUM_BASE_LEVELS) {
1335 int br_ctx;
1336 br_ctx = get_br_ctx_eob(ci, bwl, tx_class);
1337 cost += get_br_cost(abs_qc, txb_costs->lps_cost[br_ctx]);
1338 }
1339 }
1340 return cost;
1341 }
1342
1343 static INLINE int get_coeff_cost_general(int is_last, int ci, tran_low_t abs_qc,
1344 int sign, int coeff_ctx,
1345 int dc_sign_ctx,
1346 const LV_MAP_COEFF_COST *txb_costs,
1347 int bwl, TX_CLASS tx_class,
1348 const uint8_t *levels) {
1349 int cost = 0;
1350 if (is_last) {
1351 cost += txb_costs->base_eob_cost[coeff_ctx][AOMMIN(abs_qc, 3) - 1];
1352 } else {
1353 cost += txb_costs->base_cost[coeff_ctx][AOMMIN(abs_qc, 3)];
1354 }
1355 if (abs_qc != 0) {
1356 if (ci == 0) {
1357 cost += txb_costs->dc_sign_cost[dc_sign_ctx][sign];
1358 } else {
1359 cost += av1_cost_literal(1);
1360 }
1361 if (abs_qc > NUM_BASE_LEVELS) {
1362 int br_ctx;
1363 if (is_last)
1364 br_ctx = get_br_ctx_eob(ci, bwl, tx_class);
1365 else
1366 br_ctx = get_br_ctx(levels, ci, bwl, tx_class);
1367 cost += get_br_cost(abs_qc, txb_costs->lps_cost[br_ctx]);
1368 }
1369 }
1370 return cost;
1371 }
1372
1373 static INLINE void get_qc_dqc_low(tran_low_t abs_qc, int sign, int dqv,
1374 int shift, tran_low_t *qc_low,
1375 tran_low_t *dqc_low) {
1376 tran_low_t abs_qc_low = abs_qc - 1;
1377 *qc_low = (-sign ^ abs_qc_low) + sign;
1378 assert((sign ? -abs_qc_low : abs_qc_low) == *qc_low);
1379 tran_low_t abs_dqc_low = (abs_qc_low * dqv) >> shift;
1380 *dqc_low = (-sign ^ abs_dqc_low) + sign;
1381 assert((sign ? -abs_dqc_low : abs_dqc_low) == *dqc_low);
1382 }
1383
1384 static INLINE void update_coeff_general(
1385 int *accu_rate, int64_t *accu_dist, int si, int eob, TX_SIZE tx_size,
1386 TX_CLASS tx_class, int bwl, int height, int64_t rdmult, int shift,
1387 int dc_sign_ctx, const int16_t *dequant, const int16_t *scan,
1388 const LV_MAP_COEFF_COST *txb_costs, const tran_low_t *tcoeff,
1389 tran_low_t *qcoeff, tran_low_t *dqcoeff, uint8_t *levels) {
1390 const int dqv = dequant[si != 0];
1391 const int ci = scan[si];
1392 const tran_low_t qc = qcoeff[ci];
1393 const int is_last = si == (eob - 1);
1394 const int coeff_ctx = get_lower_levels_ctx_general(
1395 is_last, si, bwl, height, levels, ci, tx_size, tx_class);
1396 if (qc == 0) {
1397 *accu_rate += txb_costs->base_cost[coeff_ctx][0];
1398 } else {
1399 const int sign = (qc < 0) ? 1 : 0;
1400 const tran_low_t abs_qc = abs(qc);
1401 const tran_low_t tqc = tcoeff[ci];
1402 const tran_low_t dqc = dqcoeff[ci];
1403 const int64_t dist = get_coeff_dist(tqc, dqc, shift);
1404 const int64_t dist0 = get_coeff_dist(tqc, 0, shift);
1405 const int rate =
1406 get_coeff_cost_general(is_last, ci, abs_qc, sign, coeff_ctx,
1407 dc_sign_ctx, txb_costs, bwl, tx_class, levels);
1408 const int64_t rd = RDCOST(rdmult, rate, dist);
1409
1410 tran_low_t qc_low, dqc_low;
1411 tran_low_t abs_qc_low;
1412 int64_t dist_low, rd_low;
1413 int rate_low;
1414 if (abs_qc == 1) {
1415 abs_qc_low = qc_low = dqc_low = 0;
1416 dist_low = dist0;
1417 rate_low = txb_costs->base_cost[coeff_ctx][0];
1418 } else {
1419 get_qc_dqc_low(abs_qc, sign, dqv, shift, &qc_low, &dqc_low);
1420 abs_qc_low = abs_qc - 1;
1421 dist_low = get_coeff_dist(tqc, dqc_low, shift);
1422 rate_low =
1423 get_coeff_cost_general(is_last, ci, abs_qc_low, sign, coeff_ctx,
1424 dc_sign_ctx, txb_costs, bwl, tx_class, levels);
1425 }
1426
1427 rd_low = RDCOST(rdmult, rate_low, dist_low);
1428 if (rd_low < rd) {
1429 qcoeff[ci] = qc_low;
1430 dqcoeff[ci] = dqc_low;
1431 levels[get_padded_idx(ci, bwl)] = AOMMIN(abs_qc_low, INT8_MAX);
1432 *accu_rate += rate_low;
1433 *accu_dist += dist_low - dist0;
1434 } else {
1435 *accu_rate += rate;
1436 *accu_dist += dist - dist0;
1437 }
1438 }
1439 }
1440
1441 static AOM_FORCE_INLINE void update_coeff_simple(
1442 int *accu_rate, int si, int eob, TX_SIZE tx_size, TX_CLASS tx_class,
1443 int bwl, int64_t rdmult, int shift, const int16_t *dequant,
1444 const int16_t *scan, const LV_MAP_COEFF_COST *txb_costs,
1445 const tran_low_t *tcoeff, tran_low_t *qcoeff, tran_low_t *dqcoeff,
1446 uint8_t *levels) {
1447 const int dqv = dequant[1];
1448 (void)eob;
1449 // this simple version assumes the coeff's scan_idx is not DC (scan_idx != 0)
1450 // and not the last (scan_idx != eob - 1)
1451 assert(si != eob - 1);
1452 assert(si > 0);
1453 const int ci = scan[si];
1454 const tran_low_t qc = qcoeff[ci];
1455 const int coeff_ctx =
1456 get_lower_levels_ctx(levels, ci, bwl, tx_size, tx_class);
1457 if (qc == 0) {
1458 *accu_rate += txb_costs->base_cost[coeff_ctx][0];
1459 } else {
1460 const tran_low_t abs_qc = abs(qc);
1461 const tran_low_t abs_tqc = abs(tcoeff[ci]);
1462 const tran_low_t abs_dqc = abs(dqcoeff[ci]);
1463 int rate_low = 0;
1464 const int rate = get_two_coeff_cost_simple(
1465 ci, abs_qc, coeff_ctx, txb_costs, bwl, tx_class, levels, &rate_low);
1466 if (abs_dqc < abs_tqc) {
1467 *accu_rate += rate;
1468 return;
1469 }
1470
1471 const int64_t dist = get_coeff_dist(abs_tqc, abs_dqc, shift);
1472 const int64_t rd = RDCOST(rdmult, rate, dist);
1473
1474 const tran_low_t abs_qc_low = abs_qc - 1;
1475 const tran_low_t abs_dqc_low = (abs_qc_low * dqv) >> shift;
1476 const int64_t dist_low = get_coeff_dist(abs_tqc, abs_dqc_low, shift);
1477 const int64_t rd_low = RDCOST(rdmult, rate_low, dist_low);
1478
1479 if (rd_low < rd) {
1480 const int sign = (qc < 0) ? 1 : 0;
1481 qcoeff[ci] = (-sign ^ abs_qc_low) + sign;
1482 dqcoeff[ci] = (-sign ^ abs_dqc_low) + sign;
1483 levels[get_padded_idx(ci, bwl)] = AOMMIN(abs_qc_low, INT8_MAX);
1484 *accu_rate += rate_low;
1485 } else {
1486 *accu_rate += rate;
1487 }
1488 }
1489 }
1490
1491 static INLINE void update_coeff_eob_fast(int *eob, int shift,
1492 const int16_t *dequant_ptr,
1493 const int16_t *scan,
1494 const tran_low_t *coeff_ptr,
1495 tran_low_t *qcoeff_ptr,
1496 tran_low_t *dqcoeff_ptr) {
1497 // TODO(sarahparker) make this work for aomqm
1498 int eob_out = *eob;
1499 int zbin[2] = { dequant_ptr[0] + ROUND_POWER_OF_TWO(dequant_ptr[0] * 70, 7),
1500 dequant_ptr[1] + ROUND_POWER_OF_TWO(dequant_ptr[1] * 70, 7) };
1501
1502 for (int i = *eob - 1; i >= 0; i--) {
1503 const int rc = scan[i];
1504 const int qcoeff = qcoeff_ptr[rc];
1505 const int coeff = coeff_ptr[rc];
1506 const int coeff_sign = (coeff >> 31);
1507 int64_t abs_coeff = (coeff ^ coeff_sign) - coeff_sign;
1508
1509 if (((abs_coeff << (1 + shift)) < zbin[rc != 0]) || (qcoeff == 0)) {
1510 eob_out--;
1511 qcoeff_ptr[rc] = 0;
1512 dqcoeff_ptr[rc] = 0;
1513 } else {
1514 break;
1515 }
1516 }
1517
1518 *eob = eob_out;
1519 }
1520
1521 static AOM_FORCE_INLINE void update_coeff_eob(
1522 int *accu_rate, int64_t *accu_dist, int *eob, int *nz_num, int *nz_ci,
1523 int si, TX_SIZE tx_size, TX_CLASS tx_class, int bwl, int height,
1524 int dc_sign_ctx, int64_t rdmult, int shift, const int16_t *dequant,
1525 const int16_t *scan, const LV_MAP_EOB_COST *txb_eob_costs,
1526 const LV_MAP_COEFF_COST *txb_costs, const tran_low_t *tcoeff,
1527 tran_low_t *qcoeff, tran_low_t *dqcoeff, uint8_t *levels, int sharpness) {
1528 const int dqv = dequant[si != 0];
1529 assert(si != *eob - 1);
1530 const int ci = scan[si];
1531 const tran_low_t qc = qcoeff[ci];
1532 const int coeff_ctx =
1533 get_lower_levels_ctx(levels, ci, bwl, tx_size, tx_class);
1534 if (qc == 0) {
1535 *accu_rate += txb_costs->base_cost[coeff_ctx][0];
1536 } else {
1537 int lower_level = 0;
1538 const tran_low_t abs_qc = abs(qc);
1539 const tran_low_t tqc = tcoeff[ci];
1540 const tran_low_t dqc = dqcoeff[ci];
1541 const int sign = (qc < 0) ? 1 : 0;
1542 const int64_t dist0 = get_coeff_dist(tqc, 0, shift);
1543 int64_t dist = get_coeff_dist(tqc, dqc, shift) - dist0;
1544 int rate =
1545 get_coeff_cost_general(0, ci, abs_qc, sign, coeff_ctx, dc_sign_ctx,
1546 txb_costs, bwl, tx_class, levels);
1547 int64_t rd = RDCOST(rdmult, *accu_rate + rate, *accu_dist + dist);
1548
1549 tran_low_t qc_low, dqc_low;
1550 tran_low_t abs_qc_low;
1551 int64_t dist_low, rd_low;
1552 int rate_low;
1553 if (abs_qc == 1) {
1554 abs_qc_low = 0;
1555 dqc_low = qc_low = 0;
1556 dist_low = 0;
1557 rate_low = txb_costs->base_cost[coeff_ctx][0];
1558 rd_low = RDCOST(rdmult, *accu_rate + rate_low, *accu_dist);
1559 } else {
1560 get_qc_dqc_low(abs_qc, sign, dqv, shift, &qc_low, &dqc_low);
1561 abs_qc_low = abs_qc - 1;
1562 dist_low = get_coeff_dist(tqc, dqc_low, shift) - dist0;
1563 rate_low =
1564 get_coeff_cost_general(0, ci, abs_qc_low, sign, coeff_ctx,
1565 dc_sign_ctx, txb_costs, bwl, tx_class, levels);
1566 rd_low = RDCOST(rdmult, *accu_rate + rate_low, *accu_dist + dist_low);
1567 }
1568
1569 int lower_level_new_eob = 0;
1570 const int new_eob = si + 1;
1571 const int coeff_ctx_new_eob = get_lower_levels_ctx_eob(bwl, height, si);
1572 const int new_eob_cost =
1573 get_eob_cost(new_eob, txb_eob_costs, txb_costs, tx_class);
1574 int rate_coeff_eob =
1575 new_eob_cost + get_coeff_cost_eob(ci, abs_qc, sign, coeff_ctx_new_eob,
1576 dc_sign_ctx, txb_costs, bwl,
1577 tx_class);
1578 int64_t dist_new_eob = dist;
1579 int64_t rd_new_eob = RDCOST(rdmult, rate_coeff_eob, dist_new_eob);
1580
1581 if (abs_qc_low > 0) {
1582 const int rate_coeff_eob_low =
1583 new_eob_cost + get_coeff_cost_eob(ci, abs_qc_low, sign,
1584 coeff_ctx_new_eob, dc_sign_ctx,
1585 txb_costs, bwl, tx_class);
1586 const int64_t dist_new_eob_low = dist_low;
1587 const int64_t rd_new_eob_low =
1588 RDCOST(rdmult, rate_coeff_eob_low, dist_new_eob_low);
1589 if (rd_new_eob_low < rd_new_eob) {
1590 lower_level_new_eob = 1;
1591 rd_new_eob = rd_new_eob_low;
1592 rate_coeff_eob = rate_coeff_eob_low;
1593 dist_new_eob = dist_new_eob_low;
1594 }
1595 }
1596
1597 if (rd_low < rd) {
1598 lower_level = 1;
1599 rd = rd_low;
1600 rate = rate_low;
1601 dist = dist_low;
1602 }
1603
1604 if (sharpness == 0 && rd_new_eob < rd) {
1605 for (int ni = 0; ni < *nz_num; ++ni) {
1606 int last_ci = nz_ci[ni];
1607 levels[get_padded_idx(last_ci, bwl)] = 0;
1608 qcoeff[last_ci] = 0;
1609 dqcoeff[last_ci] = 0;
1610 }
1611 *eob = new_eob;
1612 *nz_num = 0;
1613 *accu_rate = rate_coeff_eob;
1614 *accu_dist = dist_new_eob;
1615 lower_level = lower_level_new_eob;
1616 } else {
1617 *accu_rate += rate;
1618 *accu_dist += dist;
1619 }
1620
1621 if (lower_level) {
1622 qcoeff[ci] = qc_low;
1623 dqcoeff[ci] = dqc_low;
1624 levels[get_padded_idx(ci, bwl)] = AOMMIN(abs_qc_low, INT8_MAX);
1625 }
1626 if (qcoeff[ci]) {
1627 nz_ci[*nz_num] = ci;
1628 ++*nz_num;
1629 }
1630 }
1631 }
1632
1633 static INLINE void update_skip(int *accu_rate, int64_t accu_dist, int *eob,
1634 int nz_num, int *nz_ci, int64_t rdmult,
1635 int skip_cost, int non_skip_cost,
1636 tran_low_t *qcoeff, tran_low_t *dqcoeff,
1637 int sharpness) {
1638 const int64_t rd = RDCOST(rdmult, *accu_rate + non_skip_cost, accu_dist);
1639 const int64_t rd_new_eob = RDCOST(rdmult, skip_cost, 0);
1640 if (sharpness == 0 && rd_new_eob < rd) {
1641 for (int i = 0; i < nz_num; ++i) {
1642 const int ci = nz_ci[i];
1643 qcoeff[ci] = 0;
1644 dqcoeff[ci] = 0;
1645 // no need to set up levels because this is the last step
1646 // levels[get_padded_idx(ci, bwl)] = 0;
1647 }
1648 *accu_rate = 0;
1649 *eob = 0;
1650 }
1651 }
1652
1653 int av1_optimize_txb_new(const struct AV1_COMP *cpi, MACROBLOCK *x, int plane,
1654 int block, TX_SIZE tx_size, TX_TYPE tx_type,
1655 const TXB_CTX *const txb_ctx, int *rate_cost,
1656 int sharpness, int fast_mode) {
1657 MACROBLOCKD *xd = &x->e_mbd;
1658 struct macroblockd_plane *pd = &xd->plane[plane];
1659 const struct macroblock_plane *p = &x->plane[plane];
1660 const SCAN_ORDER *scan_order = get_scan(tx_size, tx_type);
1661 const int16_t *scan = scan_order->scan;
1662 const int shift = av1_get_tx_scale(tx_size);
1663 int eob = p->eobs[block];
1664 const int16_t *dequant = p->dequant_QTX;
1665 tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
1666 tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
1667 const tran_low_t *tcoeff = BLOCK_OFFSET(p->coeff, block);
1668
1669 if (fast_mode) {
1670 update_coeff_eob_fast(&eob, shift, dequant, scan, tcoeff, qcoeff, dqcoeff);
1671 p->eobs[block] = eob;
1672 if (eob == 0) {
1673 *rate_cost = av1_cost_skip_txb(x, txb_ctx, plane, tx_size);
1674 return eob;
1675 }
1676 }
1677
1678 const AV1_COMMON *cm = &cpi->common;
1679 const PLANE_TYPE plane_type = get_plane_type(plane);
1680 const TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
1681 const TX_CLASS tx_class = tx_type_to_class[tx_type];
1682 const MB_MODE_INFO *mbmi = xd->mi[0];
1683 const int bwl = get_txb_bwl(tx_size);
1684 const int width = get_txb_wide(tx_size);
1685 const int height = get_txb_high(tx_size);
1686 assert(width == (1 << bwl));
1687 const int is_inter = is_inter_block(mbmi);
1688 const LV_MAP_COEFF_COST *txb_costs = &x->coeff_costs[txs_ctx][plane_type];
1689 const int eob_multi_size = txsize_log2_minus4[tx_size];
1690 const LV_MAP_EOB_COST *txb_eob_costs =
1691 &x->eob_costs[eob_multi_size][plane_type];
1692
1693 const int rshift =
1694 (sharpness +
1695 (cpi->oxcf.aq_mode == VARIANCE_AQ && mbmi->segment_id < 4
1696 ? 7 - mbmi->segment_id
1697 : 2) +
1698 (cpi->oxcf.aq_mode != VARIANCE_AQ &&
1699 cpi->oxcf.deltaq_mode > NO_DELTA_Q && x->sb_energy_level < 0
1700 ? (3 - x->sb_energy_level)
1701 : 0));
1702 const int64_t rdmult =
1703 (((int64_t)x->rdmult *
1704 (plane_rd_mult[is_inter][plane_type] << (2 * (xd->bd - 8)))) +
1705 2) >>
1706 rshift;
1707
1708 uint8_t levels_buf[TX_PAD_2D];
1709 uint8_t *const levels = set_levels(levels_buf, width);
1710
1711 if (eob > 1) av1_txb_init_levels(qcoeff, width, height, levels);
1712
1713 // TODO(angirbird): check iqmatrix
1714
1715 const int non_skip_cost = txb_costs->txb_skip_cost[txb_ctx->txb_skip_ctx][0];
1716 const int skip_cost = txb_costs->txb_skip_cost[txb_ctx->txb_skip_ctx][1];
1717 const int eob_cost = get_eob_cost(eob, txb_eob_costs, txb_costs, tx_class);
1718 int accu_rate = eob_cost;
1719 int64_t accu_dist = 0;
1720 int si = eob - 1;
1721 const int ci = scan[si];
1722 const tran_low_t qc = qcoeff[ci];
1723 const tran_low_t abs_qc = abs(qc);
1724 const int sign = qc < 0;
1725 const int max_nz_num = 2;
1726 int nz_num = 1;
1727 int nz_ci[3] = { ci, 0, 0 };
1728 if (abs_qc >= 2) {
1729 update_coeff_general(&accu_rate, &accu_dist, si, eob, tx_size, tx_class,
1730 bwl, height, rdmult, shift, txb_ctx->dc_sign_ctx,
1731 dequant, scan, txb_costs, tcoeff, qcoeff, dqcoeff,
1732 levels);
1733 --si;
1734 } else {
1735 assert(abs_qc == 1);
1736 const int coeff_ctx = get_lower_levels_ctx_eob(bwl, height, si);
1737 accu_rate +=
1738 get_coeff_cost_eob(ci, abs_qc, sign, coeff_ctx, txb_ctx->dc_sign_ctx,
1739 txb_costs, bwl, tx_class);
1740 const tran_low_t tqc = tcoeff[ci];
1741 const tran_low_t dqc = dqcoeff[ci];
1742 const int64_t dist = get_coeff_dist(tqc, dqc, shift);
1743 const int64_t dist0 = get_coeff_dist(tqc, 0, shift);
1744 accu_dist += dist - dist0;
1745 --si;
1746 }
1747
1748 #define UPDATE_COEFF_EOB_CASE(tx_class_literal) \
1749 case tx_class_literal: \
1750 for (; si >= 0 && nz_num <= max_nz_num && !fast_mode; --si) { \
1751 update_coeff_eob(&accu_rate, &accu_dist, &eob, &nz_num, nz_ci, si, \
1752 tx_size, tx_class_literal, bwl, height, \
1753 txb_ctx->dc_sign_ctx, rdmult, shift, dequant, scan, \
1754 txb_eob_costs, txb_costs, tcoeff, qcoeff, dqcoeff, \
1755 levels, sharpness); \
1756 } \
1757 break;
1758 switch (tx_class) {
1759 UPDATE_COEFF_EOB_CASE(TX_CLASS_2D);
1760 UPDATE_COEFF_EOB_CASE(TX_CLASS_HORIZ);
1761 UPDATE_COEFF_EOB_CASE(TX_CLASS_VERT);
1762 #undef UPDATE_COEFF_EOB_CASE
1763 default: assert(false);
1764 }
1765
1766 if (si == -1 && nz_num <= max_nz_num) {
1767 update_skip(&accu_rate, accu_dist, &eob, nz_num, nz_ci, rdmult, skip_cost,
1768 non_skip_cost, qcoeff, dqcoeff, sharpness);
1769 }
1770
1771 #define UPDATE_COEFF_SIMPLE_CASE(tx_class_literal) \
1772 case tx_class_literal: \
1773 for (; si >= 1; --si) { \
1774 update_coeff_simple(&accu_rate, si, eob, tx_size, tx_class_literal, bwl, \
1775 rdmult, shift, dequant, scan, txb_costs, tcoeff, \
1776 qcoeff, dqcoeff, levels); \
1777 } \
1778 break;
1779 switch (tx_class) {
1780 UPDATE_COEFF_SIMPLE_CASE(TX_CLASS_2D);
1781 UPDATE_COEFF_SIMPLE_CASE(TX_CLASS_HORIZ);
1782 UPDATE_COEFF_SIMPLE_CASE(TX_CLASS_VERT);
1783 #undef UPDATE_COEFF_SIMPLE_CASE
1784 default: assert(false);
1785 }
1786
1787 // DC position
1788 if (si == 0) {
1789 // no need to update accu_dist because it's not used after this point
1790 int64_t dummy_dist = 0;
1791 update_coeff_general(&accu_rate, &dummy_dist, si, eob, tx_size, tx_class,
1792 bwl, height, rdmult, shift, txb_ctx->dc_sign_ctx,
1793 dequant, scan, txb_costs, tcoeff, qcoeff, dqcoeff,
1794 levels);
1795 }
1796
1797 const int tx_type_cost = get_tx_type_cost(cm, x, xd, plane, tx_size, tx_type);
1798 if (eob == 0)
1799 accu_rate += skip_cost;
1800 else
1801 accu_rate += non_skip_cost + tx_type_cost;
1802
1803 p->eobs[block] = eob;
1804 p->txb_entropy_ctx[block] =
1805 av1_get_txb_entropy_context(qcoeff, scan_order, p->eobs[block]);
1806
1807 *rate_cost = accu_rate;
1808 return eob;
1809 }
1810
1811 // This function is deprecated, but we keep it here because hash trellis
1812 // is not integrated with av1_optimize_txb_new yet
1813 int av1_optimize_txb(const struct AV1_COMP *cpi, MACROBLOCK *x, int plane,
1814 int blk_row, int blk_col, int block, TX_SIZE tx_size,
1815 TXB_CTX *txb_ctx, int fast_mode, int *rate_cost) {
1816 const AV1_COMMON *cm = &cpi->common;
1817 MACROBLOCKD *const xd = &x->e_mbd;
1818 const PLANE_TYPE plane_type = get_plane_type(plane);
1819 const TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
1820 const TX_TYPE tx_type = av1_get_tx_type(plane_type, xd, blk_row, blk_col,
1821 tx_size, cm->reduced_tx_set_used);
1822 const MB_MODE_INFO *mbmi = xd->mi[0];
1823 const struct macroblock_plane *p = &x->plane[plane];
1824 struct macroblockd_plane *pd = &xd->plane[plane];
1825 const int eob = p->eobs[block];
1826 tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
1827 tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
1828 const tran_low_t *tcoeff = BLOCK_OFFSET(p->coeff, block);
1829 const int16_t *dequant = p->dequant_QTX;
1830 const int seg_eob = av1_get_max_eob(tx_size);
1831 const int bwl = get_txb_bwl(tx_size);
1832 const int width = get_txb_wide(tx_size);
1833 const int height = get_txb_high(tx_size);
1834 const int is_inter = is_inter_block(mbmi);
1835 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
1836 const LV_MAP_COEFF_COST *txb_costs = &x->coeff_costs[txs_ctx][plane_type];
1837 const int eob_multi_size = txsize_log2_minus4[tx_size];
1838 const LV_MAP_EOB_COST txb_eob_costs =
1839 x->eob_costs[eob_multi_size][plane_type];
1840
1841 const int shift = av1_get_tx_scale(tx_size);
1842 const int64_t rdmult =
1843 (((int64_t)x->rdmult * plane_rd_mult[is_inter][plane_type]
1844 << (2 * (xd->bd - 8))) +
1845 2) >>
1846 2;
1847 uint8_t levels_buf[TX_PAD_2D];
1848 uint8_t *const levels = set_levels(levels_buf, width);
1849 const TX_SIZE qm_tx_size = av1_get_adjusted_tx_size(tx_size);
1850 const qm_val_t *iqmatrix =
1851 IS_2D_TRANSFORM(tx_type)
1852 ? pd->seg_iqmatrix[mbmi->segment_id][qm_tx_size]
1853 : cm->giqmatrix[NUM_QM_LEVELS - 1][0][qm_tx_size];
1854 assert(width == (1 << bwl));
1855 const int tx_type_cost = get_tx_type_cost(cm, x, xd, plane, tx_size, tx_type);
1856 TxbInfo txb_info = {
1857 qcoeff, levels, dqcoeff, tcoeff, dequant, shift, tx_size,
1858 txs_ctx, tx_type, bwl, width, height, eob, seg_eob,
1859 scan_order, txb_ctx, rdmult, iqmatrix, tx_type_cost,
1860 };
1861
1862 // Hash based trellis (hbt) speed feature: avoid expensive optimize_txb calls
1863 // by storing the coefficient deltas in a hash table.
1864 // Currently disabled in speedfeatures.c
1865 if (eob <= HBT_EOB && eob > 0 && cpi->sf.use_hash_based_trellis) {
1866 return hbt_create_hashes(&txb_info, txb_costs, &txb_eob_costs, p, block,
1867 fast_mode, rate_cost);
1868 }
1869
1870 av1_txb_init_levels(qcoeff, width, height, levels);
1871
1872 const int update =
1873 optimize_txb(&txb_info, txb_costs, &txb_eob_costs, rate_cost);
1874
1875 if (update) {
1876 p->eobs[block] = txb_info.eob;
1877 p->txb_entropy_ctx[block] =
1878 av1_get_txb_entropy_context(qcoeff, scan_order, txb_info.eob);
1879 }
1880 return txb_info.eob;
1881 }
1882
1883 int av1_get_txb_entropy_context(const tran_low_t *qcoeff,
1884 const SCAN_ORDER *scan_order, int eob) {
1885 const int16_t *const scan = scan_order->scan;
1886 int cul_level = 0;
1887 int c;
1888
1889 if (eob == 0) return 0;
1890 for (c = 0; c < eob; ++c) {
1891 cul_level += abs(qcoeff[scan[c]]);
1892 if (cul_level > COEFF_CONTEXT_MASK) break;
1893 }
1894
1895 cul_level = AOMMIN(COEFF_CONTEXT_MASK, cul_level);
1896 set_dc_sign(&cul_level, qcoeff[0]);
1897
1898 return cul_level;
1899 }
1900
1901 void av1_update_txb_context_b(int plane, int block, int blk_row, int blk_col,
1902 BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
1903 void *arg) {
1904 struct tokenize_b_args *const args = arg;
1905 const AV1_COMP *cpi = args->cpi;
1906 const AV1_COMMON *cm = &cpi->common;
1907 ThreadData *const td = args->td;
1908 MACROBLOCK *const x = &td->mb;
1909 MACROBLOCKD *const xd = &x->e_mbd;
1910 struct macroblock_plane *p = &x->plane[plane];
1911 struct macroblockd_plane *pd = &xd->plane[plane];
1912 const uint16_t eob = p->eobs[block];
1913 const tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
1914 const PLANE_TYPE plane_type = pd->plane_type;
1915 const TX_TYPE tx_type = av1_get_tx_type(plane_type, xd, blk_row, blk_col,
1916 tx_size, cm->reduced_tx_set_used);
1917 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
1918 const int cul_level = av1_get_txb_entropy_context(qcoeff, scan_order, eob);
1919 av1_set_contexts(xd, pd, plane, plane_bsize, tx_size, cul_level, blk_col,
1920 blk_row);
1921 }
1922
1923 static void update_tx_type_count(const AV1_COMMON *cm, MACROBLOCKD *xd,
1924 int blk_row, int blk_col, int plane,
1925 TX_SIZE tx_size, FRAME_COUNTS *counts,
1926 uint8_t allow_update_cdf) {
1927 MB_MODE_INFO *mbmi = xd->mi[0];
1928 int is_inter = is_inter_block(mbmi);
1929 FRAME_CONTEXT *fc = xd->tile_ctx;
1930 #if !CONFIG_ENTROPY_STATS
1931 (void)counts;
1932 #endif // !CONFIG_ENTROPY_STATS
1933
1934 // Only y plane's tx_type is updated
1935 if (plane > 0) return;
1936 TX_TYPE tx_type = av1_get_tx_type(PLANE_TYPE_Y, xd, blk_row, blk_col, tx_size,
1937 cm->reduced_tx_set_used);
1938 if (get_ext_tx_types(tx_size, is_inter, cm->reduced_tx_set_used) > 1 &&
1939 cm->base_qindex > 0 && !mbmi->skip &&
1940 !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
1941 const int eset = get_ext_tx_set(tx_size, is_inter, cm->reduced_tx_set_used);
1942 if (eset > 0) {
1943 const TxSetType tx_set_type =
1944 av1_get_ext_tx_set_type(tx_size, is_inter, cm->reduced_tx_set_used);
1945 if (is_inter) {
1946 if (allow_update_cdf) {
1947 update_cdf(fc->inter_ext_tx_cdf[eset][txsize_sqr_map[tx_size]],
1948 av1_ext_tx_ind[tx_set_type][tx_type],
1949 av1_num_ext_tx_set[tx_set_type]);
1950 }
1951 #if CONFIG_ENTROPY_STATS
1952 ++counts->inter_ext_tx[eset][txsize_sqr_map[tx_size]]
1953 [av1_ext_tx_ind[tx_set_type][tx_type]];
1954 #endif // CONFIG_ENTROPY_STATS
1955 } else {
1956 PREDICTION_MODE intra_dir;
1957 if (mbmi->filter_intra_mode_info.use_filter_intra)
1958 intra_dir = fimode_to_intradir[mbmi->filter_intra_mode_info
1959 .filter_intra_mode];
1960 else
1961 intra_dir = mbmi->mode;
1962 #if CONFIG_ENTROPY_STATS
1963 ++counts->intra_ext_tx[eset][txsize_sqr_map[tx_size]][intra_dir]
1964 [av1_ext_tx_ind[tx_set_type][tx_type]];
1965 #endif // CONFIG_ENTROPY_STATS
1966 if (allow_update_cdf) {
1967 update_cdf(
1968 fc->intra_ext_tx_cdf[eset][txsize_sqr_map[tx_size]][intra_dir],
1969 av1_ext_tx_ind[tx_set_type][tx_type],
1970 av1_num_ext_tx_set[tx_set_type]);
1971 }
1972 }
1973 }
1974 }
1975 }
1976
1977 void av1_update_and_record_txb_context(int plane, int block, int blk_row,
1978 int blk_col, BLOCK_SIZE plane_bsize,
1979 TX_SIZE tx_size, void *arg) {
1980 struct tokenize_b_args *const args = arg;
1981 const AV1_COMP *cpi = args->cpi;
1982 const AV1_COMMON *cm = &cpi->common;
1983 ThreadData *const td = args->td;
1984 MACROBLOCK *const x = &td->mb;
1985 MACROBLOCKD *const xd = &x->e_mbd;
1986 struct macroblock_plane *p = &x->plane[plane];
1987 struct macroblockd_plane *pd = &xd->plane[plane];
1988 MB_MODE_INFO *mbmi = xd->mi[0];
1989 const int eob = p->eobs[block];
1990 TXB_CTX txb_ctx;
1991 get_txb_ctx(plane_bsize, tx_size, plane, pd->above_context + blk_col,
1992 pd->left_context + blk_row, &txb_ctx);
1993 const int bwl = get_txb_bwl(tx_size);
1994 const int width = get_txb_wide(tx_size);
1995 const int height = get_txb_high(tx_size);
1996 const uint8_t allow_update_cdf = args->allow_update_cdf;
1997 const TX_SIZE txsize_ctx = get_txsize_entropy_ctx(tx_size);
1998 FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
1999 #if CONFIG_ENTROPY_STATS
2000 int cdf_idx = cm->coef_cdf_category;
2001 #endif // CONFIG_ENTROPY_STATS
2002
2003 #if CONFIG_ENTROPY_STATS
2004 ++td->counts->txb_skip[cdf_idx][txsize_ctx][txb_ctx.txb_skip_ctx][eob == 0];
2005 #endif // CONFIG_ENTROPY_STATS
2006 if (allow_update_cdf) {
2007 update_cdf(ec_ctx->txb_skip_cdf[txsize_ctx][txb_ctx.txb_skip_ctx], eob == 0,
2008 2);
2009 }
2010
2011 const int txb_offset =
2012 x->mbmi_ext->cb_offset / (TX_SIZE_W_MIN * TX_SIZE_H_MIN);
2013 uint16_t *eob_txb = x->mbmi_ext->cb_coef_buff->eobs[plane] + txb_offset;
2014 uint8_t *txb_skip_ctx_txb =
2015 x->mbmi_ext->cb_coef_buff->txb_skip_ctx[plane] + txb_offset;
2016 txb_skip_ctx_txb[block] = txb_ctx.txb_skip_ctx;
2017 eob_txb[block] = eob;
2018
2019 if (eob == 0) {
2020 av1_set_contexts(xd, pd, plane, plane_bsize, tx_size, 0, blk_col, blk_row);
2021 return;
2022 }
2023
2024 tran_low_t *tcoeff_txb =
2025 x->mbmi_ext->cb_coef_buff->tcoeff[plane] + x->mbmi_ext->cb_offset;
2026 tran_low_t *tcoeff = BLOCK_OFFSET(tcoeff_txb, block);
2027 const int segment_id = mbmi->segment_id;
2028 const int seg_eob = av1_get_tx_eob(&cpi->common.seg, segment_id, tx_size);
2029 const tran_low_t *qcoeff = BLOCK_OFFSET(p->qcoeff, block);
2030 memcpy(tcoeff, qcoeff, sizeof(*tcoeff) * seg_eob);
2031
2032 uint8_t levels_buf[TX_PAD_2D];
2033 uint8_t *const levels = set_levels(levels_buf, width);
2034 av1_txb_init_levels(tcoeff, width, height, levels);
2035 update_tx_type_count(cm, xd, blk_row, blk_col, plane, tx_size, td->counts,
2036 allow_update_cdf);
2037
2038 const PLANE_TYPE plane_type = pd->plane_type;
2039 const TX_TYPE tx_type = av1_get_tx_type(plane_type, xd, blk_row, blk_col,
2040 tx_size, cm->reduced_tx_set_used);
2041 const TX_CLASS tx_class = tx_type_to_class[tx_type];
2042 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
2043 const int16_t *const scan = scan_order->scan;
2044 #if CONFIG_ENTROPY_STATS
2045 av1_update_eob_context(cdf_idx, eob, tx_size, tx_class, plane_type, ec_ctx,
2046 td->counts, allow_update_cdf);
2047 #else
2048 av1_update_eob_context(eob, tx_size, tx_class, plane_type, ec_ctx,
2049 allow_update_cdf);
2050 #endif
2051
2052 DECLARE_ALIGNED(16, int8_t, coeff_contexts[MAX_TX_SQUARE]);
2053 av1_get_nz_map_contexts(levels, scan, eob, tx_size, tx_class, coeff_contexts);
2054
2055 for (int c = eob - 1; c >= 0; --c) {
2056 const int pos = scan[c];
2057 const int coeff_ctx = coeff_contexts[pos];
2058 const tran_low_t v = qcoeff[pos];
2059 const tran_low_t level = abs(v);
2060
2061 if (allow_update_cdf) {
2062 if (c == eob - 1) {
2063 assert(coeff_ctx < 4);
2064 update_cdf(
2065 ec_ctx->coeff_base_eob_cdf[txsize_ctx][plane_type][coeff_ctx],
2066 AOMMIN(level, 3) - 1, 3);
2067 } else {
2068 update_cdf(ec_ctx->coeff_base_cdf[txsize_ctx][plane_type][coeff_ctx],
2069 AOMMIN(level, 3), 4);
2070 }
2071 }
2072 {
2073 if (c == eob - 1) {
2074 assert(coeff_ctx < 4);
2075 #if CONFIG_ENTROPY_STATS
2076 ++td->counts->coeff_base_eob_multi[cdf_idx][txsize_ctx][plane_type]
2077 [coeff_ctx][AOMMIN(level, 3) - 1];
2078 } else {
2079 ++td->counts->coeff_base_multi[cdf_idx][txsize_ctx][plane_type]
2080 [coeff_ctx][AOMMIN(level, 3)];
2081 #endif
2082 }
2083 }
2084 if (level > NUM_BASE_LEVELS) {
2085 const int base_range = level - 1 - NUM_BASE_LEVELS;
2086 const int br_ctx = get_br_ctx(levels, pos, bwl, tx_class);
2087 for (int idx = 0; idx < COEFF_BASE_RANGE; idx += BR_CDF_SIZE - 1) {
2088 const int k = AOMMIN(base_range - idx, BR_CDF_SIZE - 1);
2089 if (allow_update_cdf) {
2090 update_cdf(ec_ctx->coeff_br_cdf[AOMMIN(txsize_ctx, TX_32X32)]
2091 [plane_type][br_ctx],
2092 k, BR_CDF_SIZE);
2093 }
2094 for (int lps = 0; lps < BR_CDF_SIZE - 1; lps++) {
2095 #if CONFIG_ENTROPY_STATS
2096 ++td->counts->coeff_lps[AOMMIN(txsize_ctx, TX_32X32)][plane_type][lps]
2097 [br_ctx][lps == k];
2098 #endif // CONFIG_ENTROPY_STATS
2099 if (lps == k) break;
2100 }
2101 #if CONFIG_ENTROPY_STATS
2102 ++td->counts->coeff_lps_multi[cdf_idx][AOMMIN(txsize_ctx, TX_32X32)]
2103 [plane_type][br_ctx][k];
2104 #endif
2105 if (k < BR_CDF_SIZE - 1) break;
2106 }
2107 }
2108 }
2109
2110 // Update the context needed to code the DC sign (if applicable)
2111 if (tcoeff[0] != 0) {
2112 const int dc_sign = (tcoeff[0] < 0) ? 1 : 0;
2113 const int dc_sign_ctx = txb_ctx.dc_sign_ctx;
2114 #if CONFIG_ENTROPY_STATS
2115 ++td->counts->dc_sign[plane_type][dc_sign_ctx][dc_sign];
2116 #endif // CONFIG_ENTROPY_STATS
2117 if (allow_update_cdf)
2118 update_cdf(ec_ctx->dc_sign_cdf[plane_type][dc_sign_ctx], dc_sign, 2);
2119 int *dc_sign_ctx_txb =
2120 x->mbmi_ext->cb_coef_buff->dc_sign_ctx[plane] + txb_offset;
2121 dc_sign_ctx_txb[block] = dc_sign_ctx;
2122 }
2123
2124 const int cul_level = av1_get_txb_entropy_context(tcoeff, scan_order, eob);
2125 av1_set_contexts(xd, pd, plane, plane_bsize, tx_size, cul_level, blk_col,
2126 blk_row);
2127 }
2128
2129 void av1_update_txb_context(const AV1_COMP *cpi, ThreadData *td,
2130 RUN_TYPE dry_run, BLOCK_SIZE bsize, int *rate,
2131 int mi_row, int mi_col, uint8_t allow_update_cdf) {
2132 const AV1_COMMON *const cm = &cpi->common;
2133 const int num_planes = av1_num_planes(cm);
2134 MACROBLOCK *const x = &td->mb;
2135 MACROBLOCKD *const xd = &x->e_mbd;
2136 MB_MODE_INFO *const mbmi = xd->mi[0];
2137 struct tokenize_b_args arg = { cpi, td, NULL, 0, allow_update_cdf };
2138 (void)rate;
2139 (void)mi_row;
2140 (void)mi_col;
2141 if (mbmi->skip) {
2142 av1_reset_skip_context(xd, mi_row, mi_col, bsize, num_planes);
2143 return;
2144 }
2145
2146 if (!dry_run) {
2147 av1_foreach_transformed_block(xd, bsize, mi_row, mi_col,
2148 av1_update_and_record_txb_context, &arg,
2149 num_planes);
2150 } else if (dry_run == DRY_RUN_NORMAL) {
2151 av1_foreach_transformed_block(xd, bsize, mi_row, mi_col,
2152 av1_update_txb_context_b, &arg, num_planes);
2153 } else {
2154 printf("DRY_RUN_COSTCOEFFS is not supported yet\n");
2155 assert(0);
2156 }
2157 }
2158