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
av1_alloc_txb_buf(AV1_COMP * cpi)26 void av1_alloc_txb_buf(AV1_COMP *cpi) {
27 AV1_COMMON *cm = &cpi->common;
28 CoeffBufferPool *coeff_buf_pool = &cpi->coeff_buffer_pool;
29 const int num_sb_rows =
30 CEIL_POWER_OF_TWO(cm->mi_params.mi_rows, cm->seq_params->mib_size_log2);
31 const int num_sb_cols =
32 CEIL_POWER_OF_TWO(cm->mi_params.mi_cols, cm->seq_params->mib_size_log2);
33 const int size = num_sb_rows * num_sb_cols;
34 const int num_planes = av1_num_planes(cm);
35 const int subsampling_x = cm->seq_params->subsampling_x;
36 const int subsampling_y = cm->seq_params->subsampling_y;
37 const int luma_max_sb_square =
38 1 << num_pels_log2_lookup[cm->seq_params->sb_size];
39 const int chroma_max_sb_square =
40 luma_max_sb_square >> (subsampling_x + subsampling_y);
41 const int num_tcoeffs =
42 size * (luma_max_sb_square + (num_planes - 1) * chroma_max_sb_square);
43 const int txb_unit_size = TX_SIZE_W_MIN * TX_SIZE_H_MIN;
44
45 av1_free_txb_buf(cpi);
46 // TODO(jingning): This should be further reduced.
47 CHECK_MEM_ERROR(cm, cpi->coeff_buffer_base,
48 aom_malloc(sizeof(*cpi->coeff_buffer_base) * size));
49 CHECK_MEM_ERROR(
50 cm, coeff_buf_pool->tcoeff,
51 aom_memalign(32, sizeof(*coeff_buf_pool->tcoeff) * num_tcoeffs));
52 CHECK_MEM_ERROR(
53 cm, coeff_buf_pool->eobs,
54 aom_malloc(sizeof(*coeff_buf_pool->eobs) * num_tcoeffs / txb_unit_size));
55 CHECK_MEM_ERROR(cm, coeff_buf_pool->entropy_ctx,
56 aom_malloc(sizeof(*coeff_buf_pool->entropy_ctx) *
57 num_tcoeffs / txb_unit_size));
58
59 tran_low_t *tcoeff_ptr = coeff_buf_pool->tcoeff;
60 uint16_t *eob_ptr = coeff_buf_pool->eobs;
61 uint8_t *entropy_ctx_ptr = coeff_buf_pool->entropy_ctx;
62 for (int i = 0; i < size; i++) {
63 for (int plane = 0; plane < num_planes; plane++) {
64 const int max_sb_square =
65 (plane == AOM_PLANE_Y) ? luma_max_sb_square : chroma_max_sb_square;
66 cpi->coeff_buffer_base[i].tcoeff[plane] = tcoeff_ptr;
67 cpi->coeff_buffer_base[i].eobs[plane] = eob_ptr;
68 cpi->coeff_buffer_base[i].entropy_ctx[plane] = entropy_ctx_ptr;
69 tcoeff_ptr += max_sb_square;
70 eob_ptr += max_sb_square / txb_unit_size;
71 entropy_ctx_ptr += max_sb_square / txb_unit_size;
72 }
73 }
74 }
75
av1_free_txb_buf(AV1_COMP * cpi)76 void av1_free_txb_buf(AV1_COMP *cpi) {
77 CoeffBufferPool *coeff_buf_pool = &cpi->coeff_buffer_pool;
78 aom_free(cpi->coeff_buffer_base);
79 cpi->coeff_buffer_base = NULL;
80 aom_free(coeff_buf_pool->tcoeff);
81 coeff_buf_pool->tcoeff = NULL;
82 aom_free(coeff_buf_pool->eobs);
83 coeff_buf_pool->eobs = NULL;
84 aom_free(coeff_buf_pool->entropy_ctx);
85 coeff_buf_pool->entropy_ctx = NULL;
86 }
87
write_golomb(aom_writer * w,int level)88 static void write_golomb(aom_writer *w, int level) {
89 int x = level + 1;
90 int i = x;
91 int length = 0;
92
93 while (i) {
94 i >>= 1;
95 ++length;
96 }
97 assert(length > 0);
98
99 for (i = 0; i < length - 1; ++i) aom_write_bit(w, 0);
100
101 for (i = length - 1; i >= 0; --i) aom_write_bit(w, (x >> i) & 0x01);
102 }
103
104 static const int8_t eob_to_pos_small[33] = {
105 0, 1, 2, // 0-2
106 3, 3, // 3-4
107 4, 4, 4, 4, // 5-8
108 5, 5, 5, 5, 5, 5, 5, 5, // 9-16
109 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 6 // 17-32
110 };
111
112 static const int8_t eob_to_pos_large[17] = {
113 6, // place holder
114 7, // 33-64
115 8, 8, // 65-128
116 9, 9, 9, 9, // 129-256
117 10, 10, 10, 10, 10, 10, 10, 10, // 257-512
118 11 // 513-
119 };
120
av1_get_eob_pos_token(const int eob,int * const extra)121 int av1_get_eob_pos_token(const int eob, int *const extra) {
122 int t;
123
124 if (eob < 33) {
125 t = eob_to_pos_small[eob];
126 } else {
127 const int e = AOMMIN((eob - 1) >> 5, 16);
128 t = eob_to_pos_large[e];
129 }
130
131 *extra = eob - av1_eob_group_start[t];
132
133 return t;
134 }
135
136 #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)137 void av1_update_eob_context(int cdf_idx, int eob, TX_SIZE tx_size,
138 TX_CLASS tx_class, PLANE_TYPE plane,
139 FRAME_CONTEXT *ec_ctx, FRAME_COUNTS *counts,
140 uint8_t allow_update_cdf) {
141 #else
142 void av1_update_eob_context(int eob, TX_SIZE tx_size, TX_CLASS tx_class,
143 PLANE_TYPE plane, FRAME_CONTEXT *ec_ctx,
144 uint8_t allow_update_cdf) {
145 #endif
146 int eob_extra;
147 const int eob_pt = av1_get_eob_pos_token(eob, &eob_extra);
148 TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
149
150 const int eob_multi_size = txsize_log2_minus4[tx_size];
151 const int eob_multi_ctx = (tx_class == TX_CLASS_2D) ? 0 : 1;
152
153 switch (eob_multi_size) {
154 case 0:
155 #if CONFIG_ENTROPY_STATS
156 ++counts->eob_multi16[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
157 #endif
158 if (allow_update_cdf)
159 update_cdf(ec_ctx->eob_flag_cdf16[plane][eob_multi_ctx], eob_pt - 1, 5);
160 break;
161 case 1:
162 #if CONFIG_ENTROPY_STATS
163 ++counts->eob_multi32[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
164 #endif
165 if (allow_update_cdf)
166 update_cdf(ec_ctx->eob_flag_cdf32[plane][eob_multi_ctx], eob_pt - 1, 6);
167 break;
168 case 2:
169 #if CONFIG_ENTROPY_STATS
170 ++counts->eob_multi64[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
171 #endif
172 if (allow_update_cdf)
173 update_cdf(ec_ctx->eob_flag_cdf64[plane][eob_multi_ctx], eob_pt - 1, 7);
174 break;
175 case 3:
176 #if CONFIG_ENTROPY_STATS
177 ++counts->eob_multi128[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
178 #endif
179 if (allow_update_cdf) {
180 update_cdf(ec_ctx->eob_flag_cdf128[plane][eob_multi_ctx], eob_pt - 1,
181 8);
182 }
183 break;
184 case 4:
185 #if CONFIG_ENTROPY_STATS
186 ++counts->eob_multi256[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
187 #endif
188 if (allow_update_cdf) {
189 update_cdf(ec_ctx->eob_flag_cdf256[plane][eob_multi_ctx], eob_pt - 1,
190 9);
191 }
192 break;
193 case 5:
194 #if CONFIG_ENTROPY_STATS
195 ++counts->eob_multi512[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
196 #endif
197 if (allow_update_cdf) {
198 update_cdf(ec_ctx->eob_flag_cdf512[plane][eob_multi_ctx], eob_pt - 1,
199 10);
200 }
201 break;
202 case 6:
203 default:
204 #if CONFIG_ENTROPY_STATS
205 ++counts->eob_multi1024[cdf_idx][plane][eob_multi_ctx][eob_pt - 1];
206 #endif
207 if (allow_update_cdf) {
208 update_cdf(ec_ctx->eob_flag_cdf1024[plane][eob_multi_ctx], eob_pt - 1,
209 11);
210 }
211 break;
212 }
213
214 if (av1_eob_offset_bits[eob_pt] > 0) {
215 int eob_ctx = eob_pt - 3;
216 int eob_shift = av1_eob_offset_bits[eob_pt] - 1;
217 int bit = (eob_extra & (1 << eob_shift)) ? 1 : 0;
218 #if CONFIG_ENTROPY_STATS
219 counts->eob_extra[cdf_idx][txs_ctx][plane][eob_pt][bit]++;
220 #endif // CONFIG_ENTROPY_STATS
221 if (allow_update_cdf)
222 update_cdf(ec_ctx->eob_extra_cdf[txs_ctx][plane][eob_ctx], bit, 2);
223 }
224 }
225
226 static INLINE int get_nz_map_ctx(const uint8_t *const levels,
227 const int coeff_idx, const int bhl,
228 const int width, const int scan_idx,
229 const int is_eob, const TX_SIZE tx_size,
230 const TX_CLASS tx_class) {
231 if (is_eob) {
232 if (scan_idx == 0) return 0;
233 if (scan_idx <= (width << bhl) / 8) return 1;
234 if (scan_idx <= (width << bhl) / 4) return 2;
235 return 3;
236 }
237 const int stats =
238 get_nz_mag(levels + get_padded_idx(coeff_idx, bhl), bhl, tx_class);
239 return get_nz_map_ctx_from_stats(stats, coeff_idx, bhl, tx_size, tx_class);
240 }
241
242 void av1_txb_init_levels_c(const tran_low_t *const coeff, const int width,
243 const int height, uint8_t *const levels) {
244 const int stride = height + TX_PAD_HOR;
245 uint8_t *ls = levels;
246
247 memset(levels + stride * width, 0,
248 sizeof(*levels) * (TX_PAD_BOTTOM * stride + TX_PAD_END));
249
250 for (int i = 0; i < width; i++) {
251 for (int j = 0; j < height; j++) {
252 *ls++ = (uint8_t)clamp(abs(coeff[i * height + j]), 0, INT8_MAX);
253 }
254 for (int j = 0; j < TX_PAD_HOR; j++) {
255 *ls++ = 0;
256 }
257 }
258 }
259
260 void av1_get_nz_map_contexts_c(const uint8_t *const levels,
261 const int16_t *const scan, const uint16_t eob,
262 const TX_SIZE tx_size, const TX_CLASS tx_class,
263 int8_t *const coeff_contexts) {
264 const int bhl = get_txb_bhl(tx_size);
265 const int width = get_txb_wide(tx_size);
266 for (int i = 0; i < eob; ++i) {
267 const int pos = scan[i];
268 coeff_contexts[pos] = get_nz_map_ctx(levels, pos, bhl, width, i,
269 i == eob - 1, tx_size, tx_class);
270 }
271 }
272
273 void av1_write_coeffs_txb(const AV1_COMMON *const cm, MACROBLOCK *const x,
274 aom_writer *w, int blk_row, int blk_col, int plane,
275 int block, TX_SIZE tx_size) {
276 MACROBLOCKD *xd = &x->e_mbd;
277 const CB_COEFF_BUFFER *cb_coef_buff = x->cb_coef_buff;
278 const PLANE_TYPE plane_type = get_plane_type(plane);
279 const int txb_offset = x->mbmi_ext_frame->cb_offset[plane_type] /
280 (TX_SIZE_W_MIN * TX_SIZE_H_MIN);
281 const uint16_t *eob_txb = cb_coef_buff->eobs[plane] + txb_offset;
282 const uint16_t eob = eob_txb[block];
283 const uint8_t *entropy_ctx = cb_coef_buff->entropy_ctx[plane] + txb_offset;
284 const int txb_skip_ctx = entropy_ctx[block] & TXB_SKIP_CTX_MASK;
285 const TX_SIZE txs_ctx = get_txsize_entropy_ctx(tx_size);
286 FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
287 aom_write_symbol(w, eob == 0, ec_ctx->txb_skip_cdf[txs_ctx][txb_skip_ctx], 2);
288 if (eob == 0) return;
289
290 const TX_TYPE tx_type =
291 av1_get_tx_type(xd, plane_type, blk_row, blk_col, tx_size,
292 cm->features.reduced_tx_set_used);
293 // Only y plane's tx_type is transmitted
294 if (plane == 0) {
295 av1_write_tx_type(cm, xd, tx_type, tx_size, w);
296 }
297
298 int eob_extra;
299 const int eob_pt = av1_get_eob_pos_token(eob, &eob_extra);
300 const int eob_multi_size = txsize_log2_minus4[tx_size];
301 const TX_CLASS tx_class = tx_type_to_class[tx_type];
302 const int eob_multi_ctx = (tx_class == TX_CLASS_2D) ? 0 : 1;
303 switch (eob_multi_size) {
304 case 0:
305 aom_write_symbol(w, eob_pt - 1,
306 ec_ctx->eob_flag_cdf16[plane_type][eob_multi_ctx], 5);
307 break;
308 case 1:
309 aom_write_symbol(w, eob_pt - 1,
310 ec_ctx->eob_flag_cdf32[plane_type][eob_multi_ctx], 6);
311 break;
312 case 2:
313 aom_write_symbol(w, eob_pt - 1,
314 ec_ctx->eob_flag_cdf64[plane_type][eob_multi_ctx], 7);
315 break;
316 case 3:
317 aom_write_symbol(w, eob_pt - 1,
318 ec_ctx->eob_flag_cdf128[plane_type][eob_multi_ctx], 8);
319 break;
320 case 4:
321 aom_write_symbol(w, eob_pt - 1,
322 ec_ctx->eob_flag_cdf256[plane_type][eob_multi_ctx], 9);
323 break;
324 case 5:
325 aom_write_symbol(w, eob_pt - 1,
326 ec_ctx->eob_flag_cdf512[plane_type][eob_multi_ctx], 10);
327 break;
328 default:
329 aom_write_symbol(w, eob_pt - 1,
330 ec_ctx->eob_flag_cdf1024[plane_type][eob_multi_ctx], 11);
331 break;
332 }
333
334 const int eob_offset_bits = av1_eob_offset_bits[eob_pt];
335 if (eob_offset_bits > 0) {
336 const int eob_ctx = eob_pt - 3;
337 int eob_shift = eob_offset_bits - 1;
338 int bit = (eob_extra & (1 << eob_shift)) ? 1 : 0;
339 aom_write_symbol(w, bit,
340 ec_ctx->eob_extra_cdf[txs_ctx][plane_type][eob_ctx], 2);
341 for (int i = 1; i < eob_offset_bits; i++) {
342 eob_shift = eob_offset_bits - 1 - i;
343 bit = (eob_extra & (1 << eob_shift)) ? 1 : 0;
344 aom_write_bit(w, bit);
345 }
346 }
347
348 const int width = get_txb_wide(tx_size);
349 const int height = get_txb_high(tx_size);
350 uint8_t levels_buf[TX_PAD_2D];
351 uint8_t *const levels = set_levels(levels_buf, height);
352 const tran_low_t *tcoeff_txb =
353 cb_coef_buff->tcoeff[plane] + x->mbmi_ext_frame->cb_offset[plane_type];
354 const tran_low_t *tcoeff = tcoeff_txb + BLOCK_OFFSET(block);
355 av1_txb_init_levels(tcoeff, width, height, levels);
356 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
357 const int16_t *const scan = scan_order->scan;
358 DECLARE_ALIGNED(16, int8_t, coeff_contexts[MAX_TX_SQUARE]);
359 av1_get_nz_map_contexts(levels, scan, eob, tx_size, tx_class, coeff_contexts);
360
361 const int bhl = get_txb_bhl(tx_size);
362 for (int c = eob - 1; c >= 0; --c) {
363 const int pos = scan[c];
364 const int coeff_ctx = coeff_contexts[pos];
365 const tran_low_t v = tcoeff[pos];
366 const tran_low_t level = abs(v);
367
368 if (c == eob - 1) {
369 aom_write_symbol(
370 w, AOMMIN(level, 3) - 1,
371 ec_ctx->coeff_base_eob_cdf[txs_ctx][plane_type][coeff_ctx], 3);
372 } else {
373 aom_write_symbol(w, AOMMIN(level, 3),
374 ec_ctx->coeff_base_cdf[txs_ctx][plane_type][coeff_ctx],
375 4);
376 }
377 if (level > NUM_BASE_LEVELS) {
378 // level is above 1.
379 const int base_range = level - 1 - NUM_BASE_LEVELS;
380 const int br_ctx = get_br_ctx(levels, pos, bhl, tx_class);
381 aom_cdf_prob *cdf =
382 ec_ctx->coeff_br_cdf[AOMMIN(txs_ctx, TX_32X32)][plane_type][br_ctx];
383 for (int idx = 0; idx < COEFF_BASE_RANGE; idx += BR_CDF_SIZE - 1) {
384 const int k = AOMMIN(base_range - idx, BR_CDF_SIZE - 1);
385 aom_write_symbol(w, k, cdf, BR_CDF_SIZE);
386 if (k < BR_CDF_SIZE - 1) break;
387 }
388 }
389 }
390
391 // Loop to code all signs in the transform block,
392 // starting with the sign of DC (if applicable)
393 for (int c = 0; c < eob; ++c) {
394 const tran_low_t v = tcoeff[scan[c]];
395 const tran_low_t level = abs(v);
396 const int sign = (v < 0) ? 1 : 0;
397 if (level) {
398 if (c == 0) {
399 const int dc_sign_ctx =
400 (entropy_ctx[block] >> DC_SIGN_CTX_SHIFT) & DC_SIGN_CTX_MASK;
401 aom_write_symbol(w, sign, ec_ctx->dc_sign_cdf[plane_type][dc_sign_ctx],
402 2);
403 } else {
404 aom_write_bit(w, sign);
405 }
406 if (level > COEFF_BASE_RANGE + NUM_BASE_LEVELS)
407 write_golomb(w, level - COEFF_BASE_RANGE - 1 - NUM_BASE_LEVELS);
408 }
409 }
410 }
411
412 void av1_write_intra_coeffs_mb(const AV1_COMMON *const cm, MACROBLOCK *x,
413 aom_writer *w, BLOCK_SIZE bsize) {
414 MACROBLOCKD *xd = &x->e_mbd;
415 const int num_planes = av1_num_planes(cm);
416 int block[MAX_MB_PLANE] = { 0 };
417 int row, col;
418 assert(bsize == get_plane_block_size(bsize, xd->plane[0].subsampling_x,
419 xd->plane[0].subsampling_y));
420 const int max_blocks_wide = max_block_wide(xd, bsize, 0);
421 const int max_blocks_high = max_block_high(xd, bsize, 0);
422 const BLOCK_SIZE max_unit_bsize = BLOCK_64X64;
423 int mu_blocks_wide = mi_size_wide[max_unit_bsize];
424 int mu_blocks_high = mi_size_high[max_unit_bsize];
425 mu_blocks_wide = AOMMIN(max_blocks_wide, mu_blocks_wide);
426 mu_blocks_high = AOMMIN(max_blocks_high, mu_blocks_high);
427
428 for (row = 0; row < max_blocks_high; row += mu_blocks_high) {
429 for (col = 0; col < max_blocks_wide; col += mu_blocks_wide) {
430 for (int plane = 0; plane < num_planes; ++plane) {
431 if (plane && !xd->is_chroma_ref) break;
432 const TX_SIZE tx_size = av1_get_tx_size(plane, xd);
433 const int stepr = tx_size_high_unit[tx_size];
434 const int stepc = tx_size_wide_unit[tx_size];
435 const int step = stepr * stepc;
436 const struct macroblockd_plane *const pd = &xd->plane[plane];
437 const int unit_height = ROUND_POWER_OF_TWO(
438 AOMMIN(mu_blocks_high + row, max_blocks_high), pd->subsampling_y);
439 const int unit_width = ROUND_POWER_OF_TWO(
440 AOMMIN(mu_blocks_wide + col, max_blocks_wide), pd->subsampling_x);
441 for (int blk_row = row >> pd->subsampling_y; blk_row < unit_height;
442 blk_row += stepr) {
443 for (int blk_col = col >> pd->subsampling_x; blk_col < unit_width;
444 blk_col += stepc) {
445 av1_write_coeffs_txb(cm, x, w, blk_row, blk_col, plane,
446 block[plane], tx_size);
447 block[plane] += step;
448 }
449 }
450 }
451 }
452 }
453 }
454
455 uint8_t av1_get_txb_entropy_context(const tran_low_t *qcoeff,
456 const SCAN_ORDER *scan_order, int eob) {
457 const int16_t *const scan = scan_order->scan;
458 int cul_level = 0;
459 int c;
460
461 if (eob == 0) return 0;
462 for (c = 0; c < eob; ++c) {
463 cul_level += abs(qcoeff[scan[c]]);
464 if (cul_level > COEFF_CONTEXT_MASK) break;
465 }
466
467 cul_level = AOMMIN(COEFF_CONTEXT_MASK, cul_level);
468 set_dc_sign(&cul_level, qcoeff[0]);
469
470 return (uint8_t)cul_level;
471 }
472
473 static void update_tx_type_count(const AV1_COMP *cpi, const AV1_COMMON *cm,
474 MACROBLOCKD *xd, int blk_row, int blk_col,
475 int plane, TX_SIZE tx_size,
476 FRAME_COUNTS *counts,
477 uint8_t allow_update_cdf) {
478 MB_MODE_INFO *mbmi = xd->mi[0];
479 int is_inter = is_inter_block(mbmi);
480 const int reduced_tx_set_used = cm->features.reduced_tx_set_used;
481 FRAME_CONTEXT *fc = xd->tile_ctx;
482 #if !CONFIG_ENTROPY_STATS
483 (void)counts;
484 #endif // !CONFIG_ENTROPY_STATS
485
486 // Only y plane's tx_type is updated
487 if (plane > 0) return;
488 const TX_TYPE tx_type = av1_get_tx_type(xd, PLANE_TYPE_Y, blk_row, blk_col,
489 tx_size, reduced_tx_set_used);
490 if (is_inter) {
491 if (cpi->oxcf.txfm_cfg.use_inter_dct_only) {
492 assert(tx_type == DCT_DCT);
493 }
494 } else {
495 if (cpi->oxcf.txfm_cfg.use_intra_dct_only) {
496 assert(tx_type == DCT_DCT);
497 } else if (cpi->oxcf.txfm_cfg.use_intra_default_tx_only) {
498 const TX_TYPE default_type = get_default_tx_type(
499 PLANE_TYPE_Y, xd, tx_size, cpi->use_screen_content_tools);
500 (void)default_type;
501 // TODO(kyslov): We don't always respect use_intra_default_tx_only flag in
502 // NonRD and REALTIME case. Specifically we ignore it in hybrid inta mode
503 // search, when picking up intra mode in nonRD inter mode search and in RD
504 // REALTIME mode when we limit TX type usage.
505 // We need to fix txfm cfg for these cases. Meanwhile relieving the
506 // assert.
507 assert(tx_type == default_type || cpi->sf.rt_sf.use_nonrd_pick_mode ||
508 cpi->oxcf.mode == REALTIME);
509 }
510 }
511
512 if (get_ext_tx_types(tx_size, is_inter, reduced_tx_set_used) > 1 &&
513 cm->quant_params.base_qindex > 0 && !mbmi->skip_txfm &&
514 !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
515 const int eset = get_ext_tx_set(tx_size, is_inter, reduced_tx_set_used);
516 if (eset > 0) {
517 const TxSetType tx_set_type =
518 av1_get_ext_tx_set_type(tx_size, is_inter, reduced_tx_set_used);
519 if (is_inter) {
520 if (allow_update_cdf) {
521 update_cdf(fc->inter_ext_tx_cdf[eset][txsize_sqr_map[tx_size]],
522 av1_ext_tx_ind[tx_set_type][tx_type],
523 av1_num_ext_tx_set[tx_set_type]);
524 }
525 #if CONFIG_ENTROPY_STATS
526 ++counts->inter_ext_tx[eset][txsize_sqr_map[tx_size]]
527 [av1_ext_tx_ind[tx_set_type][tx_type]];
528 #endif // CONFIG_ENTROPY_STATS
529 } else {
530 PREDICTION_MODE intra_dir;
531 if (mbmi->filter_intra_mode_info.use_filter_intra)
532 intra_dir = fimode_to_intradir[mbmi->filter_intra_mode_info
533 .filter_intra_mode];
534 else
535 intra_dir = mbmi->mode;
536 #if CONFIG_ENTROPY_STATS
537 ++counts->intra_ext_tx[eset][txsize_sqr_map[tx_size]][intra_dir]
538 [av1_ext_tx_ind[tx_set_type][tx_type]];
539 #endif // CONFIG_ENTROPY_STATS
540 if (allow_update_cdf) {
541 update_cdf(
542 fc->intra_ext_tx_cdf[eset][txsize_sqr_map[tx_size]][intra_dir],
543 av1_ext_tx_ind[tx_set_type][tx_type],
544 av1_num_ext_tx_set[tx_set_type]);
545 }
546 }
547 }
548 }
549 }
550
551 void av1_update_and_record_txb_context(int plane, int block, int blk_row,
552 int blk_col, BLOCK_SIZE plane_bsize,
553 TX_SIZE tx_size, void *arg) {
554 struct tokenize_b_args *const args = arg;
555 const AV1_COMP *cpi = args->cpi;
556 const AV1_COMMON *cm = &cpi->common;
557 ThreadData *const td = args->td;
558 MACROBLOCK *const x = &td->mb;
559 MACROBLOCKD *const xd = &x->e_mbd;
560 struct macroblock_plane *p = &x->plane[plane];
561 struct macroblockd_plane *pd = &xd->plane[plane];
562 const int eob = p->eobs[block];
563 const int block_offset = BLOCK_OFFSET(block);
564 tran_low_t *qcoeff = p->qcoeff + block_offset;
565 const PLANE_TYPE plane_type = pd->plane_type;
566 const TX_TYPE tx_type =
567 av1_get_tx_type(xd, plane_type, blk_row, blk_col, tx_size,
568 cm->features.reduced_tx_set_used);
569 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
570 tran_low_t *tcoeff;
571 assert(args->dry_run != DRY_RUN_COSTCOEFFS);
572 if (args->dry_run == OUTPUT_ENABLED) {
573 MB_MODE_INFO *mbmi = xd->mi[0];
574 TXB_CTX txb_ctx;
575 get_txb_ctx(plane_bsize, tx_size, plane,
576 pd->above_entropy_context + blk_col,
577 pd->left_entropy_context + blk_row, &txb_ctx);
578 const int bhl = get_txb_bhl(tx_size);
579 const int width = get_txb_wide(tx_size);
580 const int height = get_txb_high(tx_size);
581 const uint8_t allow_update_cdf = args->allow_update_cdf;
582 const TX_SIZE txsize_ctx = get_txsize_entropy_ctx(tx_size);
583 FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
584 #if CONFIG_ENTROPY_STATS
585 int cdf_idx = cm->coef_cdf_category;
586 ++td->counts->txb_skip[cdf_idx][txsize_ctx][txb_ctx.txb_skip_ctx][eob == 0];
587 #endif // CONFIG_ENTROPY_STATS
588 if (allow_update_cdf) {
589 update_cdf(ec_ctx->txb_skip_cdf[txsize_ctx][txb_ctx.txb_skip_ctx],
590 eob == 0, 2);
591 }
592
593 CB_COEFF_BUFFER *cb_coef_buff = x->cb_coef_buff;
594 const int txb_offset = x->mbmi_ext_frame->cb_offset[plane_type] /
595 (TX_SIZE_W_MIN * TX_SIZE_H_MIN);
596 uint16_t *eob_txb = cb_coef_buff->eobs[plane] + txb_offset;
597 uint8_t *const entropy_ctx = cb_coef_buff->entropy_ctx[plane] + txb_offset;
598 entropy_ctx[block] = txb_ctx.txb_skip_ctx;
599 eob_txb[block] = eob;
600
601 if (eob == 0) {
602 av1_set_entropy_contexts(xd, pd, plane, plane_bsize, tx_size, 0, blk_col,
603 blk_row);
604 return;
605 }
606 const int segment_id = mbmi->segment_id;
607 const int seg_eob = av1_get_tx_eob(&cpi->common.seg, segment_id, tx_size);
608 tran_low_t *tcoeff_txb =
609 cb_coef_buff->tcoeff[plane] + x->mbmi_ext_frame->cb_offset[plane_type];
610 tcoeff = tcoeff_txb + block_offset;
611 memcpy(tcoeff, qcoeff, sizeof(*tcoeff) * seg_eob);
612
613 uint8_t levels_buf[TX_PAD_2D];
614 uint8_t *const levels = set_levels(levels_buf, height);
615 av1_txb_init_levels(tcoeff, width, height, levels);
616 update_tx_type_count(cpi, cm, xd, blk_row, blk_col, plane, tx_size,
617 td->counts, allow_update_cdf);
618
619 const TX_CLASS tx_class = tx_type_to_class[tx_type];
620 const int16_t *const scan = scan_order->scan;
621
622 // record tx type usage
623 td->rd_counts.tx_type_used[tx_size][tx_type]++;
624
625 #if CONFIG_ENTROPY_STATS
626 av1_update_eob_context(cdf_idx, eob, tx_size, tx_class, plane_type, ec_ctx,
627 td->counts, allow_update_cdf);
628 #else
629 av1_update_eob_context(eob, tx_size, tx_class, plane_type, ec_ctx,
630 allow_update_cdf);
631 #endif
632
633 DECLARE_ALIGNED(16, int8_t, coeff_contexts[MAX_TX_SQUARE]);
634 av1_get_nz_map_contexts(levels, scan, eob, tx_size, tx_class,
635 coeff_contexts);
636
637 for (int c = eob - 1; c >= 0; --c) {
638 const int pos = scan[c];
639 const int coeff_ctx = coeff_contexts[pos];
640 const tran_low_t v = qcoeff[pos];
641 const tran_low_t level = abs(v);
642 /* abs_sum_level is needed to decide the job scheduling order of
643 * pack bitstream multi-threading. This data is not needed if
644 * multi-threading is disabled. */
645 if (cpi->mt_info.pack_bs_mt_enabled) td->abs_sum_level += level;
646
647 if (allow_update_cdf) {
648 if (c == eob - 1) {
649 assert(coeff_ctx < 4);
650 update_cdf(
651 ec_ctx->coeff_base_eob_cdf[txsize_ctx][plane_type][coeff_ctx],
652 AOMMIN(level, 3) - 1, 3);
653 } else {
654 update_cdf(ec_ctx->coeff_base_cdf[txsize_ctx][plane_type][coeff_ctx],
655 AOMMIN(level, 3), 4);
656 }
657 }
658 if (c == eob - 1) {
659 assert(coeff_ctx < 4);
660 #if CONFIG_ENTROPY_STATS
661 ++td->counts->coeff_base_eob_multi[cdf_idx][txsize_ctx][plane_type]
662 [coeff_ctx][AOMMIN(level, 3) - 1];
663 } else {
664 ++td->counts->coeff_base_multi[cdf_idx][txsize_ctx][plane_type]
665 [coeff_ctx][AOMMIN(level, 3)];
666 #endif
667 }
668 if (level > NUM_BASE_LEVELS) {
669 const int base_range = level - 1 - NUM_BASE_LEVELS;
670 const int br_ctx = get_br_ctx(levels, pos, bhl, tx_class);
671 for (int idx = 0; idx < COEFF_BASE_RANGE; idx += BR_CDF_SIZE - 1) {
672 const int k = AOMMIN(base_range - idx, BR_CDF_SIZE - 1);
673 if (allow_update_cdf) {
674 update_cdf(ec_ctx->coeff_br_cdf[AOMMIN(txsize_ctx, TX_32X32)]
675 [plane_type][br_ctx],
676 k, BR_CDF_SIZE);
677 }
678 for (int lps = 0; lps < BR_CDF_SIZE - 1; lps++) {
679 #if CONFIG_ENTROPY_STATS
680 ++td->counts->coeff_lps[AOMMIN(txsize_ctx, TX_32X32)][plane_type]
681 [lps][br_ctx][lps == k];
682 #endif // CONFIG_ENTROPY_STATS
683 if (lps == k) break;
684 }
685 #if CONFIG_ENTROPY_STATS
686 ++td->counts->coeff_lps_multi[cdf_idx][AOMMIN(txsize_ctx, TX_32X32)]
687 [plane_type][br_ctx][k];
688 #endif
689 if (k < BR_CDF_SIZE - 1) break;
690 }
691 }
692 }
693 // Update the context needed to code the DC sign (if applicable)
694 if (tcoeff[0] != 0) {
695 const int dc_sign = (tcoeff[0] < 0) ? 1 : 0;
696 const int dc_sign_ctx = txb_ctx.dc_sign_ctx;
697 #if CONFIG_ENTROPY_STATS
698 ++td->counts->dc_sign[plane_type][dc_sign_ctx][dc_sign];
699 #endif // CONFIG_ENTROPY_STATS
700 if (allow_update_cdf)
701 update_cdf(ec_ctx->dc_sign_cdf[plane_type][dc_sign_ctx], dc_sign, 2);
702 entropy_ctx[block] |= dc_sign_ctx << DC_SIGN_CTX_SHIFT;
703 }
704 } else {
705 tcoeff = qcoeff;
706 }
707 const uint8_t cul_level =
708 av1_get_txb_entropy_context(tcoeff, scan_order, eob);
709 av1_set_entropy_contexts(xd, pd, plane, plane_bsize, tx_size, cul_level,
710 blk_col, blk_row);
711 }
712
713 void av1_record_txb_context(int plane, int block, int blk_row, int blk_col,
714 BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
715 void *arg) {
716 struct tokenize_b_args *const args = arg;
717 const AV1_COMP *cpi = args->cpi;
718 const AV1_COMMON *cm = &cpi->common;
719 ThreadData *const td = args->td;
720 MACROBLOCK *const x = &td->mb;
721 MACROBLOCKD *const xd = &x->e_mbd;
722 struct macroblock_plane *p = &x->plane[plane];
723 struct macroblockd_plane *pd = &xd->plane[plane];
724 const int eob = p->eobs[block];
725 const int block_offset = BLOCK_OFFSET(block);
726 tran_low_t *qcoeff = p->qcoeff + block_offset;
727 const PLANE_TYPE plane_type = pd->plane_type;
728 const TX_TYPE tx_type =
729 av1_get_tx_type(xd, plane_type, blk_row, blk_col, tx_size,
730 cm->features.reduced_tx_set_used);
731 const SCAN_ORDER *const scan_order = get_scan(tx_size, tx_type);
732 tran_low_t *tcoeff;
733 assert(args->dry_run != DRY_RUN_COSTCOEFFS);
734 if (args->dry_run == OUTPUT_ENABLED) {
735 MB_MODE_INFO *mbmi = xd->mi[0];
736 TXB_CTX txb_ctx;
737 get_txb_ctx(plane_bsize, tx_size, plane,
738 pd->above_entropy_context + blk_col,
739 pd->left_entropy_context + blk_row, &txb_ctx);
740 #if CONFIG_ENTROPY_STATS
741 const TX_SIZE txsize_ctx = get_txsize_entropy_ctx(tx_size);
742 const int bhl = get_txb_bhl(tx_size);
743 const int width = get_txb_wide(tx_size);
744 const int height = get_txb_high(tx_size);
745 int cdf_idx = cm->coef_cdf_category;
746 ++td->counts->txb_skip[cdf_idx][txsize_ctx][txb_ctx.txb_skip_ctx][eob == 0];
747 #endif // CONFIG_ENTROPY_STATS
748
749 CB_COEFF_BUFFER *cb_coef_buff = x->cb_coef_buff;
750 const int txb_offset = x->mbmi_ext_frame->cb_offset[plane_type] /
751 (TX_SIZE_W_MIN * TX_SIZE_H_MIN);
752 uint16_t *eob_txb = cb_coef_buff->eobs[plane] + txb_offset;
753 uint8_t *const entropy_ctx = cb_coef_buff->entropy_ctx[plane] + txb_offset;
754 entropy_ctx[block] = txb_ctx.txb_skip_ctx;
755 eob_txb[block] = eob;
756
757 if (eob == 0) {
758 av1_set_entropy_contexts(xd, pd, plane, plane_bsize, tx_size, 0, blk_col,
759 blk_row);
760 return;
761 }
762 const int segment_id = mbmi->segment_id;
763 const int seg_eob = av1_get_tx_eob(&cpi->common.seg, segment_id, tx_size);
764 tran_low_t *tcoeff_txb =
765 cb_coef_buff->tcoeff[plane] + x->mbmi_ext_frame->cb_offset[plane_type];
766 tcoeff = tcoeff_txb + block_offset;
767 memcpy(tcoeff, qcoeff, sizeof(*tcoeff) * seg_eob);
768
769 #if CONFIG_ENTROPY_STATS
770 uint8_t levels_buf[TX_PAD_2D];
771 uint8_t *const levels = set_levels(levels_buf, height);
772 av1_txb_init_levels(tcoeff, width, height, levels);
773 update_tx_type_count(cpi, cm, xd, blk_row, blk_col, plane, tx_size,
774 td->counts, 0 /*allow_update_cdf*/);
775
776 const TX_CLASS tx_class = tx_type_to_class[tx_type];
777 const bool do_coeff_scan = true;
778 #else
779 const bool do_coeff_scan = cpi->mt_info.pack_bs_mt_enabled;
780 #endif
781 const int16_t *const scan = scan_order->scan;
782
783 // record tx type usage
784 td->rd_counts.tx_type_used[tx_size][tx_type]++;
785
786 #if CONFIG_ENTROPY_STATS
787 FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
788 av1_update_eob_context(cdf_idx, eob, tx_size, tx_class, plane_type, ec_ctx,
789 td->counts, 0 /*allow_update_cdf*/);
790
791 DECLARE_ALIGNED(16, int8_t, coeff_contexts[MAX_TX_SQUARE]);
792 av1_get_nz_map_contexts(levels, scan, eob, tx_size, tx_class,
793 coeff_contexts);
794 #endif
795
796 for (int c = eob - 1; (c >= 0) && do_coeff_scan; --c) {
797 const int pos = scan[c];
798 const tran_low_t v = qcoeff[pos];
799 const tran_low_t level = abs(v);
800 /* abs_sum_level is needed to decide the job scheduling order of
801 * pack bitstream multi-threading. This data is not needed if
802 * multi-threading is disabled. */
803 if (cpi->mt_info.pack_bs_mt_enabled) td->abs_sum_level += level;
804
805 #if CONFIG_ENTROPY_STATS
806 const int coeff_ctx = coeff_contexts[pos];
807 if (c == eob - 1) {
808 assert(coeff_ctx < 4);
809 ++td->counts->coeff_base_eob_multi[cdf_idx][txsize_ctx][plane_type]
810 [coeff_ctx][AOMMIN(level, 3) - 1];
811 } else {
812 ++td->counts->coeff_base_multi[cdf_idx][txsize_ctx][plane_type]
813 [coeff_ctx][AOMMIN(level, 3)];
814 }
815 if (level > NUM_BASE_LEVELS) {
816 const int base_range = level - 1 - NUM_BASE_LEVELS;
817 const int br_ctx = get_br_ctx(levels, pos, bhl, tx_class);
818 for (int idx = 0; idx < COEFF_BASE_RANGE; idx += BR_CDF_SIZE - 1) {
819 const int k = AOMMIN(base_range - idx, BR_CDF_SIZE - 1);
820 for (int lps = 0; lps < BR_CDF_SIZE - 1; lps++) {
821 ++td->counts->coeff_lps[AOMMIN(txsize_ctx, TX_32X32)][plane_type]
822 [lps][br_ctx][lps == k];
823 if (lps == k) break;
824 }
825 ++td->counts->coeff_lps_multi[cdf_idx][AOMMIN(txsize_ctx, TX_32X32)]
826 [plane_type][br_ctx][k];
827 if (k < BR_CDF_SIZE - 1) break;
828 }
829 }
830 #endif
831 }
832 // Update the context needed to code the DC sign (if applicable)
833 if (tcoeff[0] != 0) {
834 const int dc_sign_ctx = txb_ctx.dc_sign_ctx;
835 #if CONFIG_ENTROPY_STATS
836 const int dc_sign = (tcoeff[0] < 0) ? 1 : 0;
837 ++td->counts->dc_sign[plane_type][dc_sign_ctx][dc_sign];
838 #endif // CONFIG_ENTROPY_STATS
839 entropy_ctx[block] |= dc_sign_ctx << DC_SIGN_CTX_SHIFT;
840 }
841 } else {
842 tcoeff = qcoeff;
843 }
844 const uint8_t cul_level =
845 av1_get_txb_entropy_context(tcoeff, scan_order, eob);
846 av1_set_entropy_contexts(xd, pd, plane, plane_bsize, tx_size, cul_level,
847 blk_col, blk_row);
848 }
849
850 void av1_update_intra_mb_txb_context(const AV1_COMP *cpi, ThreadData *td,
851 RUN_TYPE dry_run, BLOCK_SIZE bsize,
852 uint8_t allow_update_cdf) {
853 const AV1_COMMON *const cm = &cpi->common;
854 const int num_planes = av1_num_planes(cm);
855 MACROBLOCK *const x = &td->mb;
856 MACROBLOCKD *const xd = &x->e_mbd;
857 MB_MODE_INFO *const mbmi = xd->mi[0];
858 struct tokenize_b_args arg = { cpi, td, 0, allow_update_cdf, dry_run };
859 if (mbmi->skip_txfm) {
860 av1_reset_entropy_context(xd, bsize, num_planes);
861 return;
862 }
863 const foreach_transformed_block_visitor visit =
864 allow_update_cdf ? av1_update_and_record_txb_context
865 : av1_record_txb_context;
866
867 for (int plane = 0; plane < num_planes; ++plane) {
868 if (plane && !xd->is_chroma_ref) break;
869 const struct macroblockd_plane *const pd = &xd->plane[plane];
870 const int ss_x = pd->subsampling_x;
871 const int ss_y = pd->subsampling_y;
872 const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, ss_x, ss_y);
873 av1_foreach_transformed_block_in_plane(xd, plane_bsize, plane, visit, &arg);
874 }
875 }
876
877 CB_COEFF_BUFFER *av1_get_cb_coeff_buffer(const struct AV1_COMP *cpi, int mi_row,
878 int mi_col) {
879 const AV1_COMMON *const cm = &cpi->common;
880 const int mib_size_log2 = cm->seq_params->mib_size_log2;
881 const int stride =
882 CEIL_POWER_OF_TWO(cm->mi_params.mi_cols, cm->seq_params->mib_size_log2);
883 const int offset =
884 (mi_row >> mib_size_log2) * stride + (mi_col >> mib_size_log2);
885 return cpi->coeff_buffer_base + offset;
886 }
887