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 <stdint.h>
15
16 #include "aom_ports/mem.h"
17 #include "av1/common/blockd.h"
18 #include "av1/common/idct.h"
19 #include "av1/common/pred_common.h"
20 #include "av1/common/scan.h"
21 #include "av1/encoder/bitstream.h"
22 #include "av1/encoder/cost.h"
23 #include "av1/encoder/encodeframe.h"
24 #include "av1/encoder/hash.h"
25 #include "av1/encoder/rdopt.h"
26 #include "av1/encoder/tokenize.h"
27
av1_alloc_txb_buf(AV1_COMP * cpi)28 void av1_alloc_txb_buf(AV1_COMP *cpi) {
29 AV1_COMMON *cm = &cpi->common;
30 CoeffBufferPool *coeff_buf_pool = &cpi->coeff_buffer_pool;
31 const int num_sb_rows =
32 CEIL_POWER_OF_TWO(cm->mi_params.mi_rows, cm->seq_params->mib_size_log2);
33 const int num_sb_cols =
34 CEIL_POWER_OF_TWO(cm->mi_params.mi_cols, cm->seq_params->mib_size_log2);
35 const int size = num_sb_rows * num_sb_cols;
36 const int num_planes = av1_num_planes(cm);
37 const int subsampling_x = cm->seq_params->subsampling_x;
38 const int subsampling_y = cm->seq_params->subsampling_y;
39 const int luma_max_sb_square =
40 1 << num_pels_log2_lookup[cm->seq_params->sb_size];
41 const int chroma_max_sb_square =
42 luma_max_sb_square >> (subsampling_x + subsampling_y);
43 const int total_max_sb_square =
44 (luma_max_sb_square + (num_planes - 1) * chroma_max_sb_square);
45 if ((size_t)size > SIZE_MAX / (size_t)total_max_sb_square) {
46 aom_internal_error(cm->error, AOM_CODEC_ERROR,
47 "A multiplication would overflow size_t");
48 }
49 const size_t num_tcoeffs = (size_t)size * (size_t)total_max_sb_square;
50 const int txb_unit_size = TX_SIZE_W_MIN * TX_SIZE_H_MIN;
51
52 av1_free_txb_buf(cpi);
53 // TODO(jingning): This should be further reduced.
54 CHECK_MEM_ERROR(cm, cpi->coeff_buffer_base,
55 aom_malloc(sizeof(*cpi->coeff_buffer_base) * size));
56 if (sizeof(*coeff_buf_pool->tcoeff) > SIZE_MAX / num_tcoeffs) {
57 aom_internal_error(cm->error, AOM_CODEC_ERROR,
58 "A multiplication would overflow size_t");
59 }
60 CHECK_MEM_ERROR(
61 cm, coeff_buf_pool->tcoeff,
62 aom_memalign(32, sizeof(*coeff_buf_pool->tcoeff) * num_tcoeffs));
63 if (sizeof(*coeff_buf_pool->eobs) > SIZE_MAX / num_tcoeffs) {
64 aom_internal_error(cm->error, AOM_CODEC_ERROR,
65 "A multiplication would overflow size_t");
66 }
67 CHECK_MEM_ERROR(
68 cm, coeff_buf_pool->eobs,
69 aom_malloc(sizeof(*coeff_buf_pool->eobs) * num_tcoeffs / txb_unit_size));
70 if (sizeof(*coeff_buf_pool->entropy_ctx) > SIZE_MAX / num_tcoeffs) {
71 aom_internal_error(cm->error, AOM_CODEC_ERROR,
72 "A multiplication would overflow size_t");
73 }
74 CHECK_MEM_ERROR(cm, coeff_buf_pool->entropy_ctx,
75 aom_malloc(sizeof(*coeff_buf_pool->entropy_ctx) *
76 num_tcoeffs / txb_unit_size));
77
78 tran_low_t *tcoeff_ptr = coeff_buf_pool->tcoeff;
79 uint16_t *eob_ptr = coeff_buf_pool->eobs;
80 uint8_t *entropy_ctx_ptr = coeff_buf_pool->entropy_ctx;
81 for (int i = 0; i < size; i++) {
82 for (int plane = 0; plane < num_planes; plane++) {
83 const int max_sb_square =
84 (plane == AOM_PLANE_Y) ? luma_max_sb_square : chroma_max_sb_square;
85 cpi->coeff_buffer_base[i].tcoeff[plane] = tcoeff_ptr;
86 cpi->coeff_buffer_base[i].eobs[plane] = eob_ptr;
87 cpi->coeff_buffer_base[i].entropy_ctx[plane] = entropy_ctx_ptr;
88 tcoeff_ptr += max_sb_square;
89 eob_ptr += max_sb_square / txb_unit_size;
90 entropy_ctx_ptr += max_sb_square / txb_unit_size;
91 }
92 }
93 }
94
av1_free_txb_buf(AV1_COMP * cpi)95 void av1_free_txb_buf(AV1_COMP *cpi) {
96 CoeffBufferPool *coeff_buf_pool = &cpi->coeff_buffer_pool;
97 aom_free(cpi->coeff_buffer_base);
98 cpi->coeff_buffer_base = NULL;
99 aom_free(coeff_buf_pool->tcoeff);
100 coeff_buf_pool->tcoeff = NULL;
101 aom_free(coeff_buf_pool->eobs);
102 coeff_buf_pool->eobs = NULL;
103 aom_free(coeff_buf_pool->entropy_ctx);
104 coeff_buf_pool->entropy_ctx = NULL;
105 }
106
write_golomb(aom_writer * w,int level)107 static void write_golomb(aom_writer *w, int level) {
108 int x = level + 1;
109 int i = x;
110 int length = 0;
111
112 while (i) {
113 i >>= 1;
114 ++length;
115 }
116 assert(length > 0);
117
118 for (i = 0; i < length - 1; ++i) aom_write_bit(w, 0);
119
120 for (i = length - 1; i >= 0; --i) aom_write_bit(w, (x >> i) & 0x01);
121 }
122
123 static const int8_t eob_to_pos_small[33] = {
124 0, 1, 2, // 0-2
125 3, 3, // 3-4
126 4, 4, 4, 4, // 5-8
127 5, 5, 5, 5, 5, 5, 5, 5, // 9-16
128 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6 // 17-32
129 };
130
131 static const int8_t eob_to_pos_large[17] = {
132 6, // place holder
133 7, // 33-64
134 8, 8, // 65-128
135 9, 9, 9, 9, // 129-256
136 10, 10, 10, 10, 10, 10, 10, 10, // 257-512
137 11 // 513-
138 };
139
av1_get_eob_pos_token(const int eob,int * const extra)140 int av1_get_eob_pos_token(const int eob, int *const extra) {
141 int t;
142
143 if (eob < 33) {
144 t = eob_to_pos_small[eob];
145 } else {
146 const int e = AOMMIN((eob - 1) >> 5, 16);
147 t = eob_to_pos_large[e];
148 }
149
150 *extra = eob - av1_eob_group_start[t];
151
152 return t;
153 }
154
155 #if CONFIG_ENTROPY_STATS
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)156 static void update_eob_context(int cdf_idx, int eob, TX_SIZE tx_size,
157 TX_CLASS tx_class, PLANE_TYPE plane,
158 FRAME_CONTEXT *ec_ctx, FRAME_COUNTS *counts,
159 uint8_t allow_update_cdf) {
160 #else
161 static void update_eob_context(int eob, TX_SIZE tx_size, TX_CLASS tx_class,
162 PLANE_TYPE plane, FRAME_CONTEXT *ec_ctx,
163 uint8_t allow_update_cdf) {
164 #endif
165 int eob_extra;
166 const int eob_pt = av1_get_eob_pos_token(eob, &eob_extra);
167 TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
168
169 const int eob_multi_size = txsize_log2_minus4[tx_size];
170 const int eob_multi_ctx = (tx_class == TX_CLASS_2D) ? 0 : 1;
171
172 switch (eob_multi_size) {
173 case 0:
174 #if CONFIG_ENTROPY_STATS
175 ++counts->eob_multi16[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
176 #endif
177 if (allow_update_cdf)
178 update_cdf(ec_ctx->eob_flag_cdf16[plane][eob_multi_ctx], eob_pt - 1, 5);
179 break;
180 case 1:
181 #if CONFIG_ENTROPY_STATS
182 ++counts->eob_multi32[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
183 #endif
184 if (allow_update_cdf)
185 update_cdf(ec_ctx->eob_flag_cdf32[plane][eob_multi_ctx], eob_pt - 1, 6);
186 break;
187 case 2:
188 #if CONFIG_ENTROPY_STATS
189 ++counts->eob_multi64[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
190 #endif
191 if (allow_update_cdf)
192 update_cdf(ec_ctx->eob_flag_cdf64[plane][eob_multi_ctx], eob_pt - 1, 7);
193 break;
194 case 3:
195 #if CONFIG_ENTROPY_STATS
196 ++counts->eob_multi128[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
197 #endif
198 if (allow_update_cdf) {
199 update_cdf(ec_ctx->eob_flag_cdf128[plane][eob_multi_ctx], eob_pt - 1,
200 8);
201 }
202 break;
203 case 4:
204 #if CONFIG_ENTROPY_STATS
205 ++counts->eob_multi256[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
206 #endif
207 if (allow_update_cdf) {
208 update_cdf(ec_ctx->eob_flag_cdf256[plane][eob_multi_ctx], eob_pt - 1,
209 9);
210 }
211 break;
212 case 5:
213 #if CONFIG_ENTROPY_STATS
214 ++counts->eob_multi512[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
215 #endif
216 if (allow_update_cdf) {
217 update_cdf(ec_ctx->eob_flag_cdf512[plane][eob_multi_ctx], eob_pt - 1,
218 10);
219 }
220 break;
221 case 6:
222 default:
223 #if CONFIG_ENTROPY_STATS
224 ++counts->eob_multi1024[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
225 #endif
226 if (allow_update_cdf) {
227 update_cdf(ec_ctx->eob_flag_cdf1024[plane][eob_multi_ctx], eob_pt - 1,
228 11);
229 }
230 break;
231 }
232
233 if (av1_eob_offset_bits[eob_pt] > 0) {
234 int eob_ctx = eob_pt - 3;
235 int eob_shift = av1_eob_offset_bits[eob_pt] - 1;
236 int bit = (eob_extra & (1 << eob_shift)) ? 1 : 0;
237 #if CONFIG_ENTROPY_STATS
238 counts->eob_extra[cdf_idx][txs_ctx][plane][eob_pt][bit]++;
239 #endif // CONFIG_ENTROPY_STATS
240 if (allow_update_cdf)
241 update_cdf(ec_ctx->eob_extra_cdf[txs_ctx][plane][eob_ctx], bit, 2);
242 }
243 }
244
245 static inline int get_nz_map_ctx(const uint8_t *const levels,
246 const int coeff_idx, const int bhl,
247 const int width, const int scan_idx,
248 const int is_eob, const TX_SIZE tx_size,
249 const TX_CLASS tx_class) {
250 if (is_eob) {
251 if (scan_idx == 0) return 0;
252 if (scan_idx <= (width << bhl) / 8) return 1;
253 if (scan_idx <= (width << bhl) / 4) return 2;
254 return 3;
255 }
256 const int stats =
257 get_nz_mag(levels + get_padded_idx(coeff_idx, bhl), bhl, tx_class);
258 return get_nz_map_ctx_from_stats(stats, coeff_idx, bhl, tx_size, tx_class);
259 }
260
261 void av1_txb_init_levels_c(const tran_low_t *const coeff, const int width,
262 const int height, uint8_t *const levels) {
263 const int stride = height + TX_PAD_HOR;
264 uint8_t *ls = levels;
265
266 memset(levels + stride * width, 0,
267 sizeof(*levels) * (TX_PAD_BOTTOM * stride + TX_PAD_END));
268
269 for (int i = 0; i < width; i++) {
270 for (int j = 0; j < height; j++) {
271 *ls++ = (uint8_t)clamp(abs(coeff[i * height + j]), 0, INT8_MAX);
272 }
273 for (int j = 0; j < TX_PAD_HOR; j++) {
274 *ls++ = 0;
275 }
276 }
277 }
278
279 void av1_get_nz_map_contexts_c(const uint8_t *const levels,
280 const int16_t *const scan, const uint16_t eob,
281 const TX_SIZE tx_size, const TX_CLASS tx_class,
282 int8_t *const coeff_contexts) {
283 const int bhl = get_txb_bhl(tx_size);
284 const int width = get_txb_wide(tx_size);
285 for (int i = 0; i < eob; ++i) {
286 const int pos = scan[i];
287 coeff_contexts[pos] = get_nz_map_ctx(levels, pos, bhl, width, i,
288 i == eob - 1, tx_size, tx_class);
289 }
290 }
291
292 void av1_write_coeffs_txb(const AV1_COMMON *const cm, MACROBLOCK *const x,
293 aom_writer *w, int blk_row, int blk_col, int plane,
294 int block, TX_SIZE tx_size) {
295 MACROBLOCKD *xd = &x->e_mbd;
296 const CB_COEFF_BUFFER *cb_coef_buff = x->cb_coef_buff;
297 const PLANE_TYPE plane_type = get_plane_type(plane);
298 const int txb_offset = x->mbmi_ext_frame->cb_offset[plane_type] /
299 (TX_SIZE_W_MIN * TX_SIZE_H_MIN);
300 const uint16_t *eob_txb = cb_coef_buff->eobs[plane] + txb_offset;
301 const uint16_t eob = eob_txb[block];
302 const uint8_t *entropy_ctx = cb_coef_buff->entropy_ctx[plane] + txb_offset;
303 const int txb_skip_ctx = entropy_ctx[block] & TXB_SKIP_CTX_MASK;
304 const TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
305 FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
306 aom_write_symbol(w, eob == 0, ec_ctx->txb_skip_cdf[txs_ctx][txb_skip_ctx], 2);
307 if (eob == 0) return;
308
309 const TX_TYPE tx_type =
310 av1_get_tx_type(xd, plane_type, blk_row, blk_col, tx_size,
311 cm->features.reduced_tx_set_used);
312 // Only y plane's tx_type is transmitted
313 if (plane == 0) {
314 av1_write_tx_type(cm, xd, tx_type, tx_size, w);
315 }
316
317 int eob_extra;
318 const int eob_pt = av1_get_eob_pos_token(eob, &eob_extra);
319 const int eob_multi_size = txsize_log2_minus4[tx_size];
320 const TX_CLASS tx_class = tx_type_to_class[tx_type];
321 const int eob_multi_ctx = (tx_class == TX_CLASS_2D) ? 0 : 1;
322 switch (eob_multi_size) {
323 case 0:
324 aom_write_symbol(w, eob_pt - 1,
325 ec_ctx->eob_flag_cdf16[plane_type][eob_multi_ctx], 5);
326 break;
327 case 1:
328 aom_write_symbol(w, eob_pt - 1,
329 ec_ctx->eob_flag_cdf32[plane_type][eob_multi_ctx], 6);
330 break;
331 case 2:
332 aom_write_symbol(w, eob_pt - 1,
333 ec_ctx->eob_flag_cdf64[plane_type][eob_multi_ctx], 7);
334 break;
335 case 3:
336 aom_write_symbol(w, eob_pt - 1,
337 ec_ctx->eob_flag_cdf128[plane_type][eob_multi_ctx], 8);
338 break;
339 case 4:
340 aom_write_symbol(w, eob_pt - 1,
341 ec_ctx->eob_flag_cdf256[plane_type][eob_multi_ctx], 9);
342 break;
343 case 5:
344 aom_write_symbol(w, eob_pt - 1,
345 ec_ctx->eob_flag_cdf512[plane_type][eob_multi_ctx], 10);
346 break;
347 default:
348 aom_write_symbol(w, eob_pt - 1,
349 ec_ctx->eob_flag_cdf1024[plane_type][eob_multi_ctx], 11);
350 break;
351 }
352
353 const int eob_offset_bits = av1_eob_offset_bits[eob_pt];
354 if (eob_offset_bits > 0) {
355 const int eob_ctx = eob_pt - 3;
356 int eob_shift = eob_offset_bits - 1;
357 int bit = (eob_extra & (1 << eob_shift)) ? 1 : 0;
358 aom_write_symbol(w, bit,
359 ec_ctx->eob_extra_cdf[txs_ctx][plane_type][eob_ctx], 2);
360 for (int i = 1; i < eob_offset_bits; i++) {
361 eob_shift = eob_offset_bits - 1 - i;
362 bit = (eob_extra & (1 << eob_shift)) ? 1 : 0;
363 aom_write_bit(w, bit);
364 }
365 }
366
367 const int width = get_txb_wide(tx_size);
368 const int height = get_txb_high(tx_size);
369 uint8_t levels_buf[TX_PAD_2D];
370 uint8_t *const levels = set_levels(levels_buf, height);
371 const tran_low_t *tcoeff_txb =
372 cb_coef_buff->tcoeff[plane] + x->mbmi_ext_frame->cb_offset[plane_type];
373 const tran_low_t *tcoeff = tcoeff_txb + BLOCK_OFFSET(block);
374 av1_txb_init_levels(tcoeff, width, height, levels);
375 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
376 const int16_t *const scan = scan_order->scan;
377 DECLARE_ALIGNED(16, int8_t, coeff_contexts[MAX_TX_SQUARE]);
378 av1_get_nz_map_contexts(levels, scan, eob, tx_size, tx_class, coeff_contexts);
379
380 const int bhl = get_txb_bhl(tx_size);
381 for (int c = eob - 1; c >= 0; --c) {
382 const int pos = scan[c];
383 const int coeff_ctx = coeff_contexts[pos];
384 const tran_low_t v = tcoeff[pos];
385 const tran_low_t level = abs(v);
386
387 if (c == eob - 1) {
388 aom_write_symbol(
389 w, AOMMIN(level, 3) - 1,
390 ec_ctx->coeff_base_eob_cdf[txs_ctx][plane_type][coeff_ctx], 3);
391 } else {
392 aom_write_symbol(w, AOMMIN(level, 3),
393 ec_ctx->coeff_base_cdf[txs_ctx][plane_type][coeff_ctx],
394 4);
395 }
396 if (level > NUM_BASE_LEVELS) {
397 // level is above 1.
398 const int base_range = level - 1 - NUM_BASE_LEVELS;
399 const int br_ctx = get_br_ctx(levels, pos, bhl, tx_class);
400 aom_cdf_prob *cdf =
401 ec_ctx->coeff_br_cdf[AOMMIN(txs_ctx, TX_32X32)][plane_type][br_ctx];
402 for (int idx = 0; idx < COEFF_BASE_RANGE; idx += BR_CDF_SIZE - 1) {
403 const int k = AOMMIN(base_range - idx, BR_CDF_SIZE - 1);
404 aom_write_symbol(w, k, cdf, BR_CDF_SIZE);
405 if (k < BR_CDF_SIZE - 1) break;
406 }
407 }
408 }
409
410 // Loop to code all signs in the transform block,
411 // starting with the sign of DC (if applicable)
412 for (int c = 0; c < eob; ++c) {
413 const tran_low_t v = tcoeff[scan[c]];
414 const tran_low_t level = abs(v);
415 const int sign = (v < 0) ? 1 : 0;
416 if (level) {
417 if (c == 0) {
418 const int dc_sign_ctx =
419 (entropy_ctx[block] >> DC_SIGN_CTX_SHIFT) & DC_SIGN_CTX_MASK;
420 aom_write_symbol(w, sign, ec_ctx->dc_sign_cdf[plane_type][dc_sign_ctx],
421 2);
422 } else {
423 aom_write_bit(w, sign);
424 }
425 if (level > COEFF_BASE_RANGE + NUM_BASE_LEVELS)
426 write_golomb(w, level - COEFF_BASE_RANGE - 1 - NUM_BASE_LEVELS);
427 }
428 }
429 }
430
431 void av1_write_intra_coeffs_mb(const AV1_COMMON *const cm, MACROBLOCK *x,
432 aom_writer *w, BLOCK_SIZE bsize) {
433 MACROBLOCKD *xd = &x->e_mbd;
434 const int num_planes = av1_num_planes(cm);
435 int block[MAX_MB_PLANE] = { 0 };
436 int row, col;
437 assert(bsize == get_plane_block_size(bsize, xd->plane[0].subsampling_x,
438 xd->plane[0].subsampling_y));
439 const int max_blocks_wide = max_block_wide(xd, bsize, 0);
440 const int max_blocks_high = max_block_high(xd, bsize, 0);
441 const BLOCK_SIZE max_unit_bsize = BLOCK_64X64;
442 int mu_blocks_wide = mi_size_wide[max_unit_bsize];
443 int mu_blocks_high = mi_size_high[max_unit_bsize];
444 mu_blocks_wide = AOMMIN(max_blocks_wide, mu_blocks_wide);
445 mu_blocks_high = AOMMIN(max_blocks_high, mu_blocks_high);
446
447 for (row = 0; row < max_blocks_high; row += mu_blocks_high) {
448 for (col = 0; col < max_blocks_wide; col += mu_blocks_wide) {
449 for (int plane = 0; plane < num_planes; ++plane) {
450 if (plane && !xd->is_chroma_ref) break;
451 const TX_SIZE tx_size = av1_get_tx_size(plane, xd);
452 const int stepr = tx_size_high_unit[tx_size];
453 const int stepc = tx_size_wide_unit[tx_size];
454 const int step = stepr * stepc;
455 const struct macroblockd_plane *const pd = &xd->plane[plane];
456 const int unit_height = ROUND_POWER_OF_TWO(
457 AOMMIN(mu_blocks_high + row, max_blocks_high), pd->subsampling_y);
458 const int unit_width = ROUND_POWER_OF_TWO(
459 AOMMIN(mu_blocks_wide + col, max_blocks_wide), pd->subsampling_x);
460 for (int blk_row = row >> pd->subsampling_y; blk_row < unit_height;
461 blk_row += stepr) {
462 for (int blk_col = col >> pd->subsampling_x; blk_col < unit_width;
463 blk_col += stepc) {
464 av1_write_coeffs_txb(cm, x, w, blk_row, blk_col, plane,
465 block[plane], tx_size);
466 block[plane] += step;
467 }
468 }
469 }
470 }
471 }
472 }
473
474 uint8_t av1_get_txb_entropy_context(const tran_low_t *qcoeff,
475 const SCAN_ORDER *scan_order, int eob) {
476 const int16_t *const scan = scan_order->scan;
477 int cul_level = 0;
478 int c;
479
480 if (eob == 0) return 0;
481 for (c = 0; c < eob; ++c) {
482 cul_level += abs(qcoeff[scan[c]]);
483 if (cul_level > COEFF_CONTEXT_MASK) break;
484 }
485
486 cul_level = AOMMIN(COEFF_CONTEXT_MASK, cul_level);
487 set_dc_sign(&cul_level, qcoeff[0]);
488
489 return (uint8_t)cul_level;
490 }
491
492 static void update_tx_type_count(const AV1_COMP *cpi, const AV1_COMMON *cm,
493 MACROBLOCKD *xd, int blk_row, int blk_col,
494 int plane, TX_SIZE tx_size,
495 FRAME_COUNTS *counts,
496 uint8_t allow_update_cdf) {
497 MB_MODE_INFO *mbmi = xd->mi[0];
498 int is_inter = is_inter_block(mbmi);
499 const int reduced_tx_set_used = cm->features.reduced_tx_set_used;
500 FRAME_CONTEXT *fc = xd->tile_ctx;
501 #if !CONFIG_ENTROPY_STATS
502 (void)counts;
503 #endif // !CONFIG_ENTROPY_STATS
504
505 // Only y plane's tx_type is updated
506 if (plane > 0) return;
507 const TX_TYPE tx_type = av1_get_tx_type(xd, PLANE_TYPE_Y, blk_row, blk_col,
508 tx_size, reduced_tx_set_used);
509 if (is_inter) {
510 if (cpi->oxcf.txfm_cfg.use_inter_dct_only) {
511 assert(tx_type == DCT_DCT);
512 }
513 } else {
514 if (cpi->oxcf.txfm_cfg.use_intra_dct_only) {
515 assert(tx_type == DCT_DCT);
516 } else if (cpi->oxcf.txfm_cfg.use_intra_default_tx_only) {
517 const TX_TYPE default_type = get_default_tx_type(
518 PLANE_TYPE_Y, xd, tx_size, cpi->use_screen_content_tools);
519 (void)default_type;
520 // TODO(kyslov): We don't always respect use_intra_default_tx_only flag in
521 // NonRD and REALTIME case. Specifically we ignore it in hybrid inta mode
522 // search, when picking up intra mode in nonRD inter mode search and in RD
523 // REALTIME mode when we limit TX type usage.
524 // We need to fix txfm cfg for these cases. Meanwhile relieving the
525 // assert.
526 assert(tx_type == default_type || cpi->sf.rt_sf.use_nonrd_pick_mode ||
527 cpi->oxcf.mode == REALTIME);
528 }
529 }
530
531 if (get_ext_tx_types(tx_size, is_inter, reduced_tx_set_used) > 1 &&
532 cm->quant_params.base_qindex > 0 && !mbmi->skip_txfm &&
533 !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
534 const int eset = get_ext_tx_set(tx_size, is_inter, reduced_tx_set_used);
535 if (eset > 0) {
536 const TxSetType tx_set_type =
537 av1_get_ext_tx_set_type(tx_size, is_inter, reduced_tx_set_used);
538 if (is_inter) {
539 if (allow_update_cdf) {
540 update_cdf(fc->inter_ext_tx_cdf[eset][txsize_sqr_map[tx_size]],
541 av1_ext_tx_ind[tx_set_type][tx_type],
542 av1_num_ext_tx_set[tx_set_type]);
543 }
544 #if CONFIG_ENTROPY_STATS
545 ++counts->inter_ext_tx[eset][txsize_sqr_map[tx_size]]
546 [av1_ext_tx_ind[tx_set_type][tx_type]];
547 #endif // CONFIG_ENTROPY_STATS
548 } else {
549 PREDICTION_MODE intra_dir;
550 if (mbmi->filter_intra_mode_info.use_filter_intra)
551 intra_dir = fimode_to_intradir[mbmi->filter_intra_mode_info
552 .filter_intra_mode];
553 else
554 intra_dir = mbmi->mode;
555 #if CONFIG_ENTROPY_STATS
556 ++counts->intra_ext_tx[eset][txsize_sqr_map[tx_size]][intra_dir]
557 [av1_ext_tx_ind[tx_set_type][tx_type]];
558 #endif // CONFIG_ENTROPY_STATS
559 if (allow_update_cdf) {
560 update_cdf(
561 fc->intra_ext_tx_cdf[eset][txsize_sqr_map[tx_size]][intra_dir],
562 av1_ext_tx_ind[tx_set_type][tx_type],
563 av1_num_ext_tx_set[tx_set_type]);
564 }
565 }
566 }
567 }
568 }
569
570 void av1_update_and_record_txb_context(int plane, int block, int blk_row,
571 int blk_col, BLOCK_SIZE plane_bsize,
572 TX_SIZE tx_size, void *arg) {
573 struct tokenize_b_args *const args = arg;
574 const AV1_COMP *cpi = args->cpi;
575 const AV1_COMMON *cm = &cpi->common;
576 ThreadData *const td = args->td;
577 MACROBLOCK *const x = &td->mb;
578 MACROBLOCKD *const xd = &x->e_mbd;
579 struct macroblock_plane *p = &x->plane[plane];
580 struct macroblockd_plane *pd = &xd->plane[plane];
581 const int eob = p->eobs[block];
582 const int block_offset = BLOCK_OFFSET(block);
583 tran_low_t *qcoeff = p->qcoeff + block_offset;
584 const PLANE_TYPE plane_type = pd->plane_type;
585 const TX_TYPE tx_type =
586 av1_get_tx_type(xd, plane_type, blk_row, blk_col, tx_size,
587 cm->features.reduced_tx_set_used);
588 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
589 tran_low_t *tcoeff;
590 assert(args->dry_run != DRY_RUN_COSTCOEFFS);
591 if (args->dry_run == OUTPUT_ENABLED) {
592 MB_MODE_INFO *mbmi = xd->mi[0];
593 TXB_CTX txb_ctx;
594 get_txb_ctx(plane_bsize, tx_size, plane,
595 pd->above_entropy_context + blk_col,
596 pd->left_entropy_context + blk_row, &txb_ctx);
597 const int bhl = get_txb_bhl(tx_size);
598 const int width = get_txb_wide(tx_size);
599 const int height = get_txb_high(tx_size);
600 const uint8_t allow_update_cdf = args->allow_update_cdf;
601 const TX_SIZE txsize_ctx = get_txsize_entropy_ctx(tx_size);
602 FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
603 #if CONFIG_ENTROPY_STATS
604 int cdf_idx = cm->coef_cdf_category;
605 ++td->counts->txb_skip[cdf_idx][txsize_ctx][txb_ctx.txb_skip_ctx][eob == 0];
606 #endif // CONFIG_ENTROPY_STATS
607 if (allow_update_cdf) {
608 update_cdf(ec_ctx->txb_skip_cdf[txsize_ctx][txb_ctx.txb_skip_ctx],
609 eob == 0, 2);
610 }
611
612 CB_COEFF_BUFFER *cb_coef_buff = x->cb_coef_buff;
613 const int txb_offset = x->mbmi_ext_frame->cb_offset[plane_type] /
614 (TX_SIZE_W_MIN * TX_SIZE_H_MIN);
615 uint16_t *eob_txb = cb_coef_buff->eobs[plane] + txb_offset;
616 uint8_t *const entropy_ctx = cb_coef_buff->entropy_ctx[plane] + txb_offset;
617 entropy_ctx[block] = txb_ctx.txb_skip_ctx;
618 eob_txb[block] = eob;
619
620 if (eob == 0) {
621 av1_set_entropy_contexts(xd, pd, plane, plane_bsize, tx_size, 0, blk_col,
622 blk_row);
623 return;
624 }
625 const int segment_id = mbmi->segment_id;
626 const int seg_eob = av1_get_tx_eob(&cpi->common.seg, segment_id, tx_size);
627 tran_low_t *tcoeff_txb =
628 cb_coef_buff->tcoeff[plane] + x->mbmi_ext_frame->cb_offset[plane_type];
629 tcoeff = tcoeff_txb + block_offset;
630 memcpy(tcoeff, qcoeff, sizeof(*tcoeff) * seg_eob);
631
632 uint8_t levels_buf[TX_PAD_2D];
633 uint8_t *const levels = set_levels(levels_buf, height);
634 av1_txb_init_levels(tcoeff, width, height, levels);
635 update_tx_type_count(cpi, cm, xd, blk_row, blk_col, plane, tx_size,
636 td->counts, allow_update_cdf);
637
638 const TX_CLASS tx_class = tx_type_to_class[tx_type];
639 const int16_t *const scan = scan_order->scan;
640
641 // record tx type usage
642 td->rd_counts.tx_type_used[tx_size][tx_type]++;
643
644 #if CONFIG_ENTROPY_STATS
645 update_eob_context(cdf_idx, eob, tx_size, tx_class, plane_type, ec_ctx,
646 td->counts, allow_update_cdf);
647 #else
648 update_eob_context(eob, tx_size, tx_class, plane_type, ec_ctx,
649 allow_update_cdf);
650 #endif
651
652 DECLARE_ALIGNED(16, int8_t, coeff_contexts[MAX_TX_SQUARE]);
653 av1_get_nz_map_contexts(levels, scan, eob, tx_size, tx_class,
654 coeff_contexts);
655
656 for (int c = eob - 1; c >= 0; --c) {
657 const int pos = scan[c];
658 const int coeff_ctx = coeff_contexts[pos];
659 const tran_low_t v = qcoeff[pos];
660 const tran_low_t level = abs(v);
661 /* abs_sum_level is needed to decide the job scheduling order of
662 * pack bitstream multi-threading. This data is not needed if
663 * multi-threading is disabled. */
664 if (cpi->mt_info.pack_bs_mt_enabled) td->abs_sum_level += level;
665
666 if (allow_update_cdf) {
667 if (c == eob - 1) {
668 assert(coeff_ctx < 4);
669 update_cdf(
670 ec_ctx->coeff_base_eob_cdf[txsize_ctx][plane_type][coeff_ctx],
671 AOMMIN(level, 3) - 1, 3);
672 } else {
673 update_cdf(ec_ctx->coeff_base_cdf[txsize_ctx][plane_type][coeff_ctx],
674 AOMMIN(level, 3), 4);
675 }
676 }
677 if (c == eob - 1) {
678 assert(coeff_ctx < 4);
679 #if CONFIG_ENTROPY_STATS
680 ++td->counts->coeff_base_eob_multi[cdf_idx][txsize_ctx][plane_type]
681 [coeff_ctx][AOMMIN(level, 3) - 1];
682 } else {
683 ++td->counts->coeff_base_multi[cdf_idx][txsize_ctx][plane_type]
684 [coeff_ctx][AOMMIN(level, 3)];
685 #endif
686 }
687 if (level > NUM_BASE_LEVELS) {
688 const int base_range = level - 1 - NUM_BASE_LEVELS;
689 const int br_ctx = get_br_ctx(levels, pos, bhl, tx_class);
690 for (int idx = 0; idx < COEFF_BASE_RANGE; idx += BR_CDF_SIZE - 1) {
691 const int k = AOMMIN(base_range - idx, BR_CDF_SIZE - 1);
692 if (allow_update_cdf) {
693 update_cdf(ec_ctx->coeff_br_cdf[AOMMIN(txsize_ctx, TX_32X32)]
694 [plane_type][br_ctx],
695 k, BR_CDF_SIZE);
696 }
697 for (int lps = 0; lps < BR_CDF_SIZE - 1; lps++) {
698 #if CONFIG_ENTROPY_STATS
699 ++td->counts->coeff_lps[AOMMIN(txsize_ctx, TX_32X32)][plane_type]
700 [lps][br_ctx][lps == k];
701 #endif // CONFIG_ENTROPY_STATS
702 if (lps == k) break;
703 }
704 #if CONFIG_ENTROPY_STATS
705 ++td->counts->coeff_lps_multi[cdf_idx][AOMMIN(txsize_ctx, TX_32X32)]
706 [plane_type][br_ctx][k];
707 #endif
708 if (k < BR_CDF_SIZE - 1) break;
709 }
710 }
711 }
712 // Update the context needed to code the DC sign (if applicable)
713 if (tcoeff[0] != 0) {
714 const int dc_sign = (tcoeff[0] < 0) ? 1 : 0;
715 const int dc_sign_ctx = txb_ctx.dc_sign_ctx;
716 #if CONFIG_ENTROPY_STATS
717 ++td->counts->dc_sign[plane_type][dc_sign_ctx][dc_sign];
718 #endif // CONFIG_ENTROPY_STATS
719 if (allow_update_cdf)
720 update_cdf(ec_ctx->dc_sign_cdf[plane_type][dc_sign_ctx], dc_sign, 2);
721 entropy_ctx[block] |= dc_sign_ctx << DC_SIGN_CTX_SHIFT;
722 }
723 } else {
724 tcoeff = qcoeff;
725 }
726 const uint8_t cul_level =
727 av1_get_txb_entropy_context(tcoeff, scan_order, eob);
728 av1_set_entropy_contexts(xd, pd, plane, plane_bsize, tx_size, cul_level,
729 blk_col, blk_row);
730 }
731
732 void av1_record_txb_context(int plane, int block, int blk_row, int blk_col,
733 BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
734 void *arg) {
735 struct tokenize_b_args *const args = arg;
736 const AV1_COMP *cpi = args->cpi;
737 const AV1_COMMON *cm = &cpi->common;
738 ThreadData *const td = args->td;
739 MACROBLOCK *const x = &td->mb;
740 MACROBLOCKD *const xd = &x->e_mbd;
741 struct macroblock_plane *p = &x->plane[plane];
742 struct macroblockd_plane *pd = &xd->plane[plane];
743 const int eob = p->eobs[block];
744 const int block_offset = BLOCK_OFFSET(block);
745 tran_low_t *qcoeff = p->qcoeff + block_offset;
746 const PLANE_TYPE plane_type = pd->plane_type;
747 const TX_TYPE tx_type =
748 av1_get_tx_type(xd, plane_type, blk_row, blk_col, tx_size,
749 cm->features.reduced_tx_set_used);
750 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
751 tran_low_t *tcoeff;
752 assert(args->dry_run != DRY_RUN_COSTCOEFFS);
753 if (args->dry_run == OUTPUT_ENABLED) {
754 MB_MODE_INFO *mbmi = xd->mi[0];
755 TXB_CTX txb_ctx;
756 get_txb_ctx(plane_bsize, tx_size, plane,
757 pd->above_entropy_context + blk_col,
758 pd->left_entropy_context + blk_row, &txb_ctx);
759 #if CONFIG_ENTROPY_STATS
760 const TX_SIZE txsize_ctx = get_txsize_entropy_ctx(tx_size);
761 const int bhl = get_txb_bhl(tx_size);
762 const int width = get_txb_wide(tx_size);
763 const int height = get_txb_high(tx_size);
764 int cdf_idx = cm->coef_cdf_category;
765 ++td->counts->txb_skip[cdf_idx][txsize_ctx][txb_ctx.txb_skip_ctx][eob == 0];
766 #endif // CONFIG_ENTROPY_STATS
767
768 CB_COEFF_BUFFER *cb_coef_buff = x->cb_coef_buff;
769 const int txb_offset = x->mbmi_ext_frame->cb_offset[plane_type] /
770 (TX_SIZE_W_MIN * TX_SIZE_H_MIN);
771 uint16_t *eob_txb = cb_coef_buff->eobs[plane] + txb_offset;
772 uint8_t *const entropy_ctx = cb_coef_buff->entropy_ctx[plane] + txb_offset;
773 entropy_ctx[block] = txb_ctx.txb_skip_ctx;
774 eob_txb[block] = eob;
775
776 if (eob == 0) {
777 av1_set_entropy_contexts(xd, pd, plane, plane_bsize, tx_size, 0, blk_col,
778 blk_row);
779 return;
780 }
781 const int segment_id = mbmi->segment_id;
782 const int seg_eob = av1_get_tx_eob(&cpi->common.seg, segment_id, tx_size);
783 tran_low_t *tcoeff_txb =
784 cb_coef_buff->tcoeff[plane] + x->mbmi_ext_frame->cb_offset[plane_type];
785 tcoeff = tcoeff_txb + block_offset;
786 memcpy(tcoeff, qcoeff, sizeof(*tcoeff) * seg_eob);
787
788 #if CONFIG_ENTROPY_STATS
789 uint8_t levels_buf[TX_PAD_2D];
790 uint8_t *const levels = set_levels(levels_buf, height);
791 av1_txb_init_levels(tcoeff, width, height, levels);
792 update_tx_type_count(cpi, cm, xd, blk_row, blk_col, plane, tx_size,
793 td->counts, 0 /*allow_update_cdf*/);
794
795 const TX_CLASS tx_class = tx_type_to_class[tx_type];
796 const bool do_coeff_scan = true;
797 #else
798 const bool do_coeff_scan = cpi->mt_info.pack_bs_mt_enabled;
799 #endif
800 const int16_t *const scan = scan_order->scan;
801
802 // record tx type usage
803 td->rd_counts.tx_type_used[tx_size][tx_type]++;
804
805 #if CONFIG_ENTROPY_STATS
806 FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
807 update_eob_context(cdf_idx, eob, tx_size, tx_class, plane_type, ec_ctx,
808 td->counts, 0 /*allow_update_cdf*/);
809
810 DECLARE_ALIGNED(16, int8_t, coeff_contexts[MAX_TX_SQUARE]);
811 av1_get_nz_map_contexts(levels, scan, eob, tx_size, tx_class,
812 coeff_contexts);
813 #endif
814
815 for (int c = eob - 1; (c >= 0) && do_coeff_scan; --c) {
816 const int pos = scan[c];
817 const tran_low_t v = qcoeff[pos];
818 const tran_low_t level = abs(v);
819 /* abs_sum_level is needed to decide the job scheduling order of
820 * pack bitstream multi-threading. This data is not needed if
821 * multi-threading is disabled. */
822 if (cpi->mt_info.pack_bs_mt_enabled) td->abs_sum_level += level;
823
824 #if CONFIG_ENTROPY_STATS
825 const int coeff_ctx = coeff_contexts[pos];
826 if (c == eob - 1) {
827 assert(coeff_ctx < 4);
828 ++td->counts->coeff_base_eob_multi[cdf_idx][txsize_ctx][plane_type]
829 [coeff_ctx][AOMMIN(level, 3) - 1];
830 } else {
831 ++td->counts->coeff_base_multi[cdf_idx][txsize_ctx][plane_type]
832 [coeff_ctx][AOMMIN(level, 3)];
833 }
834 if (level > NUM_BASE_LEVELS) {
835 const int base_range = level - 1 - NUM_BASE_LEVELS;
836 const int br_ctx = get_br_ctx(levels, pos, bhl, tx_class);
837 for (int idx = 0; idx < COEFF_BASE_RANGE; idx += BR_CDF_SIZE - 1) {
838 const int k = AOMMIN(base_range - idx, BR_CDF_SIZE - 1);
839 for (int lps = 0; lps < BR_CDF_SIZE - 1; lps++) {
840 ++td->counts->coeff_lps[AOMMIN(txsize_ctx, TX_32X32)][plane_type]
841 [lps][br_ctx][lps == k];
842 if (lps == k) break;
843 }
844 ++td->counts->coeff_lps_multi[cdf_idx][AOMMIN(txsize_ctx, TX_32X32)]
845 [plane_type][br_ctx][k];
846 if (k < BR_CDF_SIZE - 1) break;
847 }
848 }
849 #endif
850 }
851 // Update the context needed to code the DC sign (if applicable)
852 if (tcoeff[0] != 0) {
853 const int dc_sign_ctx = txb_ctx.dc_sign_ctx;
854 #if CONFIG_ENTROPY_STATS
855 const int dc_sign = (tcoeff[0] < 0) ? 1 : 0;
856 ++td->counts->dc_sign[plane_type][dc_sign_ctx][dc_sign];
857 #endif // CONFIG_ENTROPY_STATS
858 entropy_ctx[block] |= dc_sign_ctx << DC_SIGN_CTX_SHIFT;
859 }
860 } else {
861 tcoeff = qcoeff;
862 }
863 const uint8_t cul_level =
864 av1_get_txb_entropy_context(tcoeff, scan_order, eob);
865 av1_set_entropy_contexts(xd, pd, plane, plane_bsize, tx_size, cul_level,
866 blk_col, blk_row);
867 }
868
869 void av1_update_intra_mb_txb_context(const AV1_COMP *cpi, ThreadData *td,
870 RUN_TYPE dry_run, BLOCK_SIZE bsize,
871 uint8_t allow_update_cdf) {
872 const AV1_COMMON *const cm = &cpi->common;
873 const int num_planes = av1_num_planes(cm);
874 MACROBLOCK *const x = &td->mb;
875 MACROBLOCKD *const xd = &x->e_mbd;
876 MB_MODE_INFO *const mbmi = xd->mi[0];
877 struct tokenize_b_args arg = { cpi, td, 0, allow_update_cdf, dry_run };
878 if (mbmi->skip_txfm) {
879 av1_reset_entropy_context(xd, bsize, num_planes);
880 return;
881 }
882 const foreach_transformed_block_visitor visit =
883 allow_update_cdf ? av1_update_and_record_txb_context
884 : av1_record_txb_context;
885
886 for (int plane = 0; plane < num_planes; ++plane) {
887 if (plane && !xd->is_chroma_ref) break;
888 const struct macroblockd_plane *const pd = &xd->plane[plane];
889 const int ss_x = pd->subsampling_x;
890 const int ss_y = pd->subsampling_y;
891 const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, ss_x, ss_y);
892 av1_foreach_transformed_block_in_plane(xd, plane_bsize, plane, visit, &arg);
893 }
894 }
895
896 CB_COEFF_BUFFER *av1_get_cb_coeff_buffer(const struct AV1_COMP *cpi, int mi_row,
897 int mi_col) {
898 const AV1_COMMON *const cm = &cpi->common;
899 const int mib_size_log2 = cm->seq_params->mib_size_log2;
900 const int stride =
901 CEIL_POWER_OF_TWO(cm->mi_params.mi_cols, cm->seq_params->mib_size_log2);
902 const int offset =
903 (mi_row >> mib_size_log2) * stride + (mi_col >> mib_size_log2);
904 return cpi->coeff_buffer_base + offset;
905 }
906