2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
15 #include "vpx/vpx_encoder.h"
16 #include "vpx_mem/vpx_mem.h"
17 #include "vpx_ports/mem_ops.h"
19 #include "vp9/common/vp9_entropy.h"
20 #include "vp9/common/vp9_entropymode.h"
21 #include "vp9/common/vp9_entropymv.h"
22 #include "vp9/common/vp9_mvref_common.h"
23 #include "vp9/common/vp9_pragmas.h"
24 #include "vp9/common/vp9_pred_common.h"
25 #include "vp9/common/vp9_seg_common.h"
26 #include "vp9/common/vp9_systemdependent.h"
27 #include "vp9/common/vp9_tile_common.h"
29 #include "vp9/encoder/vp9_cost.h"
30 #include "vp9/encoder/vp9_bitstream.h"
31 #include "vp9/encoder/vp9_encodemv.h"
32 #include "vp9/encoder/vp9_mcomp.h"
33 #include "vp9/encoder/vp9_segmentation.h"
34 #include "vp9/encoder/vp9_subexp.h"
35 #include "vp9/encoder/vp9_tokenize.h"
36 #include "vp9/encoder/vp9_write_bit_buffer.h"
38 static struct vp9_token intra_mode_encodings[INTRA_MODES];
39 static struct vp9_token switchable_interp_encodings[SWITCHABLE_FILTERS];
40 static struct vp9_token partition_encodings[PARTITION_TYPES];
41 static struct vp9_token inter_mode_encodings[INTER_MODES];
43 void vp9_entropy_mode_init() {
44 vp9_tokens_from_tree(intra_mode_encodings, vp9_intra_mode_tree);
45 vp9_tokens_from_tree(switchable_interp_encodings, vp9_switchable_interp_tree);
46 vp9_tokens_from_tree(partition_encodings, vp9_partition_tree);
47 vp9_tokens_from_tree(inter_mode_encodings, vp9_inter_mode_tree);
50 static void write_intra_mode(vp9_writer *w, PREDICTION_MODE mode,
51 const vp9_prob *probs) {
52 vp9_write_token(w, vp9_intra_mode_tree, probs, &intra_mode_encodings[mode]);
55 static void write_inter_mode(vp9_writer *w, PREDICTION_MODE mode,
56 const vp9_prob *probs) {
57 assert(is_inter_mode(mode));
58 vp9_write_token(w, vp9_inter_mode_tree, probs,
59 &inter_mode_encodings[INTER_OFFSET(mode)]);
62 static void encode_unsigned_max(struct vp9_write_bit_buffer *wb,
64 vp9_wb_write_literal(wb, data, get_unsigned_bits(max));
67 static void prob_diff_update(const vp9_tree_index *tree,
68 vp9_prob probs[/*n - 1*/],
69 const unsigned int counts[/*n - 1*/],
70 int n, vp9_writer *w) {
72 unsigned int branch_ct[32][2];
74 // Assuming max number of probabilities <= 32
77 vp9_tree_probs_from_distribution(tree, branch_ct, counts);
78 for (i = 0; i < n - 1; ++i)
79 vp9_cond_prob_diff_update(w, &probs[i], branch_ct[i]);
82 static void write_selected_tx_size(const VP9_COMP *cpi,
83 TX_SIZE tx_size, BLOCK_SIZE bsize,
85 const TX_SIZE max_tx_size = max_txsize_lookup[bsize];
86 const MACROBLOCKD *const xd = &cpi->mb.e_mbd;
87 const vp9_prob *const tx_probs = get_tx_probs2(max_tx_size, xd,
88 &cpi->common.fc.tx_probs);
89 vp9_write(w, tx_size != TX_4X4, tx_probs[0]);
90 if (tx_size != TX_4X4 && max_tx_size >= TX_16X16) {
91 vp9_write(w, tx_size != TX_8X8, tx_probs[1]);
92 if (tx_size != TX_8X8 && max_tx_size >= TX_32X32)
93 vp9_write(w, tx_size != TX_16X16, tx_probs[2]);
97 static int write_skip(const VP9_COMP *cpi, int segment_id, const MODE_INFO *mi,
99 const MACROBLOCKD *const xd = &cpi->mb.e_mbd;
100 if (vp9_segfeature_active(&cpi->common.seg, segment_id, SEG_LVL_SKIP)) {
103 const int skip = mi->mbmi.skip;
104 vp9_write(w, skip, vp9_get_skip_prob(&cpi->common, xd));
109 static void update_skip_probs(VP9_COMMON *cm, vp9_writer *w) {
112 for (k = 0; k < SKIP_CONTEXTS; ++k)
113 vp9_cond_prob_diff_update(w, &cm->fc.skip_probs[k], cm->counts.skip[k]);
116 static void update_switchable_interp_probs(VP9_COMMON *cm, vp9_writer *w) {
118 for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
119 prob_diff_update(vp9_switchable_interp_tree,
120 cm->fc.switchable_interp_prob[j],
121 cm->counts.switchable_interp[j], SWITCHABLE_FILTERS, w);
124 static void pack_mb_tokens(vp9_writer *w,
125 TOKENEXTRA **tp, const TOKENEXTRA *stop) {
128 while (p < stop && p->token != EOSB_TOKEN) {
129 const int t = p->token;
130 const struct vp9_token *const a = &vp9_coef_encodings[t];
131 const vp9_extra_bit *const b = &vp9_extra_bits[t];
136 /* skip one or two nodes */
137 if (p->skip_eob_node) {
138 n -= p->skip_eob_node;
139 i = 2 * p->skip_eob_node;
142 // TODO(jbb): expanding this can lead to big gains. It allows
143 // much better branch prediction and would enable us to avoid numerous
144 // lookups and compares.
146 // If we have a token that's in the constrained set, the coefficient tree
147 // is split into two treed writes. The first treed write takes care of the
148 // unconstrained nodes. The second treed write takes care of the
149 // constrained nodes.
150 if (t >= TWO_TOKEN && t < EOB_TOKEN) {
151 int len = UNCONSTRAINED_NODES - p->skip_eob_node;
152 int bits = v >> (n - len);
153 vp9_write_tree(w, vp9_coef_tree, p->context_tree, bits, len, i);
154 vp9_write_tree(w, vp9_coef_con_tree,
155 vp9_pareto8_full[p->context_tree[PIVOT_NODE] - 1],
158 vp9_write_tree(w, vp9_coef_tree, p->context_tree, v, n, i);
162 const int e = p->extra, l = b->len;
165 const unsigned char *pb = b->prob;
167 int n = l; /* number of bits in v, assumed nonzero */
171 const int bb = (v >> --n) & 1;
172 vp9_write(w, bb, pb[i >> 1]);
177 vp9_write_bit(w, e & 1);
182 *tp = p + (p->token == EOSB_TOKEN);
185 static void write_segment_id(vp9_writer *w, const struct segmentation *seg,
187 if (seg->enabled && seg->update_map)
188 vp9_write_tree(w, vp9_segment_tree, seg->tree_probs, segment_id, 3, 0);
191 // This function encodes the reference frame
192 static void write_ref_frames(const VP9_COMP *cpi, vp9_writer *w) {
193 const VP9_COMMON *const cm = &cpi->common;
194 const MACROBLOCKD *const xd = &cpi->mb.e_mbd;
195 const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
196 const int is_compound = has_second_ref(mbmi);
197 const int segment_id = mbmi->segment_id;
199 // If segment level coding of this signal is disabled...
200 // or the segment allows multiple reference frame options
201 if (vp9_segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
202 assert(!is_compound);
203 assert(mbmi->ref_frame[0] ==
204 vp9_get_segdata(&cm->seg, segment_id, SEG_LVL_REF_FRAME));
206 // does the feature use compound prediction or not
207 // (if not specified at the frame/segment level)
208 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
209 vp9_write(w, is_compound, vp9_get_reference_mode_prob(cm, xd));
211 assert(!is_compound == (cm->reference_mode == SINGLE_REFERENCE));
215 vp9_write(w, mbmi->ref_frame[0] == GOLDEN_FRAME,
216 vp9_get_pred_prob_comp_ref_p(cm, xd));
218 const int bit0 = mbmi->ref_frame[0] != LAST_FRAME;
219 vp9_write(w, bit0, vp9_get_pred_prob_single_ref_p1(cm, xd));
221 const int bit1 = mbmi->ref_frame[0] != GOLDEN_FRAME;
222 vp9_write(w, bit1, vp9_get_pred_prob_single_ref_p2(cm, xd));
228 static void pack_inter_mode_mvs(VP9_COMP *cpi, const MODE_INFO *mi,
230 VP9_COMMON *const cm = &cpi->common;
231 const nmv_context *nmvc = &cm->fc.nmvc;
232 const MACROBLOCK *const x = &cpi->mb;
233 const MACROBLOCKD *const xd = &x->e_mbd;
234 const struct segmentation *const seg = &cm->seg;
235 const MB_MODE_INFO *const mbmi = &mi->mbmi;
236 const PREDICTION_MODE mode = mbmi->mode;
237 const int segment_id = mbmi->segment_id;
238 const BLOCK_SIZE bsize = mbmi->sb_type;
239 const int allow_hp = cm->allow_high_precision_mv;
240 const int is_inter = is_inter_block(mbmi);
241 const int is_compound = has_second_ref(mbmi);
244 if (seg->update_map) {
245 if (seg->temporal_update) {
246 const int pred_flag = mbmi->seg_id_predicted;
247 vp9_prob pred_prob = vp9_get_pred_prob_seg_id(seg, xd);
248 vp9_write(w, pred_flag, pred_prob);
250 write_segment_id(w, seg, segment_id);
252 write_segment_id(w, seg, segment_id);
256 skip = write_skip(cpi, segment_id, mi, w);
258 if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
259 vp9_write(w, is_inter, vp9_get_intra_inter_prob(cm, xd));
261 if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT &&
263 (skip || vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP)))) {
264 write_selected_tx_size(cpi, mbmi->tx_size, bsize, w);
268 if (bsize >= BLOCK_8X8) {
269 write_intra_mode(w, mode, cm->fc.y_mode_prob[size_group_lookup[bsize]]);
272 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
273 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
274 for (idy = 0; idy < 2; idy += num_4x4_h) {
275 for (idx = 0; idx < 2; idx += num_4x4_w) {
276 const PREDICTION_MODE b_mode = mi->bmi[idy * 2 + idx].as_mode;
277 write_intra_mode(w, b_mode, cm->fc.y_mode_prob[0]);
281 write_intra_mode(w, mbmi->uv_mode, cm->fc.uv_mode_prob[mode]);
283 const int mode_ctx = mbmi->mode_context[mbmi->ref_frame[0]];
284 const vp9_prob *const inter_probs = cm->fc.inter_mode_probs[mode_ctx];
285 write_ref_frames(cpi, w);
287 // If segment skip is not enabled code the mode.
288 if (!vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP)) {
289 if (bsize >= BLOCK_8X8) {
290 write_inter_mode(w, mode, inter_probs);
291 ++cm->counts.inter_mode[mode_ctx][INTER_OFFSET(mode)];
295 if (cm->interp_filter == SWITCHABLE) {
296 const int ctx = vp9_get_pred_context_switchable_interp(xd);
297 vp9_write_token(w, vp9_switchable_interp_tree,
298 cm->fc.switchable_interp_prob[ctx],
299 &switchable_interp_encodings[mbmi->interp_filter]);
301 assert(mbmi->interp_filter == cm->interp_filter);
304 if (bsize < BLOCK_8X8) {
305 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
306 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
308 for (idy = 0; idy < 2; idy += num_4x4_h) {
309 for (idx = 0; idx < 2; idx += num_4x4_w) {
310 const int j = idy * 2 + idx;
311 const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
312 write_inter_mode(w, b_mode, inter_probs);
313 ++cm->counts.inter_mode[mode_ctx][INTER_OFFSET(b_mode)];
314 if (b_mode == NEWMV) {
315 for (ref = 0; ref < 1 + is_compound; ++ref)
316 vp9_encode_mv(cpi, w, &mi->bmi[j].as_mv[ref].as_mv,
317 &mbmi->ref_mvs[mbmi->ref_frame[ref]][0].as_mv,
324 for (ref = 0; ref < 1 + is_compound; ++ref)
325 vp9_encode_mv(cpi, w, &mbmi->mv[ref].as_mv,
326 &mbmi->ref_mvs[mbmi->ref_frame[ref]][0].as_mv, nmvc,
333 static void write_mb_modes_kf(const VP9_COMP *cpi, MODE_INFO **mi_8x8,
335 const VP9_COMMON *const cm = &cpi->common;
336 const MACROBLOCKD *const xd = &cpi->mb.e_mbd;
337 const struct segmentation *const seg = &cm->seg;
338 const MODE_INFO *const mi = mi_8x8[0];
339 const MODE_INFO *const above_mi = mi_8x8[-xd->mi_stride];
340 const MODE_INFO *const left_mi = xd->left_available ? mi_8x8[-1] : NULL;
341 const MB_MODE_INFO *const mbmi = &mi->mbmi;
342 const BLOCK_SIZE bsize = mbmi->sb_type;
345 write_segment_id(w, seg, mbmi->segment_id);
347 write_skip(cpi, mbmi->segment_id, mi, w);
349 if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT)
350 write_selected_tx_size(cpi, mbmi->tx_size, bsize, w);
352 if (bsize >= BLOCK_8X8) {
353 write_intra_mode(w, mbmi->mode, get_y_mode_probs(mi, above_mi, left_mi, 0));
355 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
356 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
359 for (idy = 0; idy < 2; idy += num_4x4_h) {
360 for (idx = 0; idx < 2; idx += num_4x4_w) {
361 const int block = idy * 2 + idx;
362 write_intra_mode(w, mi->bmi[block].as_mode,
363 get_y_mode_probs(mi, above_mi, left_mi, block));
368 write_intra_mode(w, mbmi->uv_mode, vp9_kf_uv_mode_prob[mbmi->mode]);
371 static void write_modes_b(VP9_COMP *cpi, const TileInfo *const tile,
372 vp9_writer *w, TOKENEXTRA **tok, TOKENEXTRA *tok_end,
373 int mi_row, int mi_col) {
374 VP9_COMMON *const cm = &cpi->common;
375 MACROBLOCKD *const xd = &cpi->mb.e_mbd;
378 xd->mi = cm->mi_grid_visible + (mi_row * cm->mi_stride + mi_col);
381 set_mi_row_col(xd, tile,
382 mi_row, num_8x8_blocks_high_lookup[m->mbmi.sb_type],
383 mi_col, num_8x8_blocks_wide_lookup[m->mbmi.sb_type],
384 cm->mi_rows, cm->mi_cols);
385 if (frame_is_intra_only(cm)) {
386 write_mb_modes_kf(cpi, xd->mi, w);
388 pack_inter_mode_mvs(cpi, m, w);
391 assert(*tok < tok_end);
392 pack_mb_tokens(w, tok, tok_end);
395 static void write_partition(VP9_COMMON *cm, MACROBLOCKD *xd,
396 int hbs, int mi_row, int mi_col,
397 PARTITION_TYPE p, BLOCK_SIZE bsize, vp9_writer *w) {
398 const int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
399 const vp9_prob *const probs = get_partition_probs(cm, ctx);
400 const int has_rows = (mi_row + hbs) < cm->mi_rows;
401 const int has_cols = (mi_col + hbs) < cm->mi_cols;
403 if (has_rows && has_cols) {
404 vp9_write_token(w, vp9_partition_tree, probs, &partition_encodings[p]);
405 } else if (!has_rows && has_cols) {
406 assert(p == PARTITION_SPLIT || p == PARTITION_HORZ);
407 vp9_write(w, p == PARTITION_SPLIT, probs[1]);
408 } else if (has_rows && !has_cols) {
409 assert(p == PARTITION_SPLIT || p == PARTITION_VERT);
410 vp9_write(w, p == PARTITION_SPLIT, probs[2]);
412 assert(p == PARTITION_SPLIT);
416 static void write_modes_sb(VP9_COMP *cpi,
417 const TileInfo *const tile,
418 vp9_writer *w, TOKENEXTRA **tok, TOKENEXTRA *tok_end,
419 int mi_row, int mi_col, BLOCK_SIZE bsize) {
420 VP9_COMMON *const cm = &cpi->common;
421 MACROBLOCKD *const xd = &cpi->mb.e_mbd;
423 const int bsl = b_width_log2(bsize);
424 const int bs = (1 << bsl) / 4;
425 PARTITION_TYPE partition;
427 MODE_INFO *m = cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col];
429 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
432 partition = partition_lookup[bsl][m->mbmi.sb_type];
433 write_partition(cm, xd, bs, mi_row, mi_col, partition, bsize, w);
434 subsize = get_subsize(bsize, partition);
435 if (subsize < BLOCK_8X8) {
436 write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
440 write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
443 write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
444 if (mi_row + bs < cm->mi_rows)
445 write_modes_b(cpi, tile, w, tok, tok_end, mi_row + bs, mi_col);
448 write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
449 if (mi_col + bs < cm->mi_cols)
450 write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col + bs);
452 case PARTITION_SPLIT:
453 write_modes_sb(cpi, tile, w, tok, tok_end, mi_row, mi_col, subsize);
454 write_modes_sb(cpi, tile, w, tok, tok_end, mi_row, mi_col + bs,
456 write_modes_sb(cpi, tile, w, tok, tok_end, mi_row + bs, mi_col,
458 write_modes_sb(cpi, tile, w, tok, tok_end, mi_row + bs, mi_col + bs,
466 // update partition context
467 if (bsize >= BLOCK_8X8 &&
468 (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
469 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
472 static void write_modes(VP9_COMP *cpi,
473 const TileInfo *const tile,
474 vp9_writer *w, TOKENEXTRA **tok, TOKENEXTRA *tok_end) {
477 for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
478 mi_row += MI_BLOCK_SIZE) {
479 vp9_zero(cpi->mb.e_mbd.left_seg_context);
480 for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
481 mi_col += MI_BLOCK_SIZE)
482 write_modes_sb(cpi, tile, w, tok, tok_end, mi_row, mi_col,
487 static void build_tree_distribution(VP9_COMP *cpi, TX_SIZE tx_size,
488 vp9_coeff_stats *coef_branch_ct) {
489 vp9_coeff_probs_model *coef_probs = cpi->frame_coef_probs[tx_size];
490 vp9_coeff_count *coef_counts = cpi->coef_counts[tx_size];
491 unsigned int (*eob_branch_ct)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS] =
492 cpi->common.counts.eob_branch[tx_size];
495 for (i = 0; i < PLANE_TYPES; ++i) {
496 for (j = 0; j < REF_TYPES; ++j) {
497 for (k = 0; k < COEF_BANDS; ++k) {
498 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
499 vp9_tree_probs_from_distribution(vp9_coef_tree,
500 coef_branch_ct[i][j][k][l],
501 coef_counts[i][j][k][l]);
502 coef_branch_ct[i][j][k][l][0][1] = eob_branch_ct[i][j][k][l] -
503 coef_branch_ct[i][j][k][l][0][0];
504 for (m = 0; m < UNCONSTRAINED_NODES; ++m)
505 coef_probs[i][j][k][l][m] = get_binary_prob(
506 coef_branch_ct[i][j][k][l][m][0],
507 coef_branch_ct[i][j][k][l][m][1]);
514 static void update_coef_probs_common(vp9_writer* const bc, VP9_COMP *cpi,
516 vp9_coeff_stats *frame_branch_ct) {
517 vp9_coeff_probs_model *new_frame_coef_probs = cpi->frame_coef_probs[tx_size];
518 vp9_coeff_probs_model *old_frame_coef_probs =
519 cpi->common.fc.coef_probs[tx_size];
520 const vp9_prob upd = DIFF_UPDATE_PROB;
521 const int entropy_nodes_update = UNCONSTRAINED_NODES;
523 switch (cpi->sf.use_fast_coef_updates) {
525 /* dry run to see if there is any udpate at all needed */
527 int update[2] = {0, 0};
528 for (i = 0; i < PLANE_TYPES; ++i) {
529 for (j = 0; j < REF_TYPES; ++j) {
530 for (k = 0; k < COEF_BANDS; ++k) {
531 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
532 for (t = 0; t < entropy_nodes_update; ++t) {
533 vp9_prob newp = new_frame_coef_probs[i][j][k][l][t];
534 const vp9_prob oldp = old_frame_coef_probs[i][j][k][l][t];
538 s = vp9_prob_diff_update_savings_search_model(
539 frame_branch_ct[i][j][k][l][0],
540 old_frame_coef_probs[i][j][k][l], &newp, upd);
542 s = vp9_prob_diff_update_savings_search(
543 frame_branch_ct[i][j][k][l][t], oldp, &newp, upd);
544 if (s > 0 && newp != oldp)
547 savings += s - (int)(vp9_cost_zero(upd));
549 savings -= (int)(vp9_cost_zero(upd));
557 // printf("Update %d %d, savings %d\n", update[0], update[1], savings);
558 /* Is coef updated at all */
559 if (update[1] == 0 || savings < 0) {
560 vp9_write_bit(bc, 0);
563 vp9_write_bit(bc, 1);
564 for (i = 0; i < PLANE_TYPES; ++i) {
565 for (j = 0; j < REF_TYPES; ++j) {
566 for (k = 0; k < COEF_BANDS; ++k) {
567 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
568 // calc probs and branch cts for this frame only
569 for (t = 0; t < entropy_nodes_update; ++t) {
570 vp9_prob newp = new_frame_coef_probs[i][j][k][l][t];
571 vp9_prob *oldp = old_frame_coef_probs[i][j][k][l] + t;
572 const vp9_prob upd = DIFF_UPDATE_PROB;
576 s = vp9_prob_diff_update_savings_search_model(
577 frame_branch_ct[i][j][k][l][0],
578 old_frame_coef_probs[i][j][k][l], &newp, upd);
580 s = vp9_prob_diff_update_savings_search(
581 frame_branch_ct[i][j][k][l][t],
583 if (s > 0 && newp != *oldp)
585 vp9_write(bc, u, upd);
587 /* send/use new probability */
588 vp9_write_prob_diff_update(bc, newp, *oldp);
600 case ONE_LOOP_REDUCED: {
601 const int prev_coef_contexts_to_update =
602 cpi->sf.use_fast_coef_updates == ONE_LOOP_REDUCED ?
603 COEFF_CONTEXTS >> 1 : COEFF_CONTEXTS;
604 const int coef_band_to_update =
605 cpi->sf.use_fast_coef_updates == ONE_LOOP_REDUCED ?
606 COEF_BANDS >> 1 : COEF_BANDS;
608 int noupdates_before_first = 0;
609 for (i = 0; i < PLANE_TYPES; ++i) {
610 for (j = 0; j < REF_TYPES; ++j) {
611 for (k = 0; k < COEF_BANDS; ++k) {
612 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
613 // calc probs and branch cts for this frame only
614 for (t = 0; t < entropy_nodes_update; ++t) {
615 vp9_prob newp = new_frame_coef_probs[i][j][k][l][t];
616 vp9_prob *oldp = old_frame_coef_probs[i][j][k][l] + t;
619 if (l >= prev_coef_contexts_to_update ||
620 k >= coef_band_to_update) {
624 s = vp9_prob_diff_update_savings_search_model(
625 frame_branch_ct[i][j][k][l][0],
626 old_frame_coef_probs[i][j][k][l], &newp, upd);
628 s = vp9_prob_diff_update_savings_search(
629 frame_branch_ct[i][j][k][l][t],
631 if (s > 0 && newp != *oldp)
635 if (u == 0 && updates == 0) {
636 noupdates_before_first++;
639 if (u == 1 && updates == 1) {
642 vp9_write_bit(bc, 1);
643 for (v = 0; v < noupdates_before_first; ++v)
644 vp9_write(bc, 0, upd);
646 vp9_write(bc, u, upd);
648 /* send/use new probability */
649 vp9_write_prob_diff_update(bc, newp, *oldp);
658 vp9_write_bit(bc, 0); // no updates
668 static void update_coef_probs(VP9_COMP *cpi, vp9_writer* w) {
669 const TX_MODE tx_mode = cpi->common.tx_mode;
670 const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
672 vp9_coeff_stats frame_branch_ct[TX_SIZES][PLANE_TYPES];
674 vp9_clear_system_state();
676 for (tx_size = TX_4X4; tx_size <= TX_32X32; ++tx_size)
677 build_tree_distribution(cpi, tx_size, frame_branch_ct[tx_size]);
679 for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size)
680 update_coef_probs_common(w, cpi, tx_size, frame_branch_ct[tx_size]);
683 static void encode_loopfilter(struct loopfilter *lf,
684 struct vp9_write_bit_buffer *wb) {
687 // Encode the loop filter level and type
688 vp9_wb_write_literal(wb, lf->filter_level, 6);
689 vp9_wb_write_literal(wb, lf->sharpness_level, 3);
691 // Write out loop filter deltas applied at the MB level based on mode or
692 // ref frame (if they are enabled).
693 vp9_wb_write_bit(wb, lf->mode_ref_delta_enabled);
695 if (lf->mode_ref_delta_enabled) {
696 vp9_wb_write_bit(wb, lf->mode_ref_delta_update);
697 if (lf->mode_ref_delta_update) {
698 for (i = 0; i < MAX_REF_LF_DELTAS; i++) {
699 const int delta = lf->ref_deltas[i];
700 const int changed = delta != lf->last_ref_deltas[i];
701 vp9_wb_write_bit(wb, changed);
703 lf->last_ref_deltas[i] = delta;
704 vp9_wb_write_literal(wb, abs(delta) & 0x3F, 6);
705 vp9_wb_write_bit(wb, delta < 0);
709 for (i = 0; i < MAX_MODE_LF_DELTAS; i++) {
710 const int delta = lf->mode_deltas[i];
711 const int changed = delta != lf->last_mode_deltas[i];
712 vp9_wb_write_bit(wb, changed);
714 lf->last_mode_deltas[i] = delta;
715 vp9_wb_write_literal(wb, abs(delta) & 0x3F, 6);
716 vp9_wb_write_bit(wb, delta < 0);
723 static void write_delta_q(struct vp9_write_bit_buffer *wb, int delta_q) {
725 vp9_wb_write_bit(wb, 1);
726 vp9_wb_write_literal(wb, abs(delta_q), 4);
727 vp9_wb_write_bit(wb, delta_q < 0);
729 vp9_wb_write_bit(wb, 0);
733 static void encode_quantization(VP9_COMMON *cm,
734 struct vp9_write_bit_buffer *wb) {
735 vp9_wb_write_literal(wb, cm->base_qindex, QINDEX_BITS);
736 write_delta_q(wb, cm->y_dc_delta_q);
737 write_delta_q(wb, cm->uv_dc_delta_q);
738 write_delta_q(wb, cm->uv_ac_delta_q);
742 static void encode_segmentation(VP9_COMP *cpi,
743 struct vp9_write_bit_buffer *wb) {
746 struct segmentation *seg = &cpi->common.seg;
748 vp9_wb_write_bit(wb, seg->enabled);
753 vp9_wb_write_bit(wb, seg->update_map);
754 if (seg->update_map) {
755 // Select the coding strategy (temporal or spatial)
756 vp9_choose_segmap_coding_method(cpi);
757 // Write out probabilities used to decode unpredicted macro-block segments
758 for (i = 0; i < SEG_TREE_PROBS; i++) {
759 const int prob = seg->tree_probs[i];
760 const int update = prob != MAX_PROB;
761 vp9_wb_write_bit(wb, update);
763 vp9_wb_write_literal(wb, prob, 8);
766 // Write out the chosen coding method.
767 vp9_wb_write_bit(wb, seg->temporal_update);
768 if (seg->temporal_update) {
769 for (i = 0; i < PREDICTION_PROBS; i++) {
770 const int prob = seg->pred_probs[i];
771 const int update = prob != MAX_PROB;
772 vp9_wb_write_bit(wb, update);
774 vp9_wb_write_literal(wb, prob, 8);
780 vp9_wb_write_bit(wb, seg->update_data);
781 if (seg->update_data) {
782 vp9_wb_write_bit(wb, seg->abs_delta);
784 for (i = 0; i < MAX_SEGMENTS; i++) {
785 for (j = 0; j < SEG_LVL_MAX; j++) {
786 const int active = vp9_segfeature_active(seg, i, j);
787 vp9_wb_write_bit(wb, active);
789 const int data = vp9_get_segdata(seg, i, j);
790 const int data_max = vp9_seg_feature_data_max(j);
792 if (vp9_is_segfeature_signed(j)) {
793 encode_unsigned_max(wb, abs(data), data_max);
794 vp9_wb_write_bit(wb, data < 0);
796 encode_unsigned_max(wb, data, data_max);
805 static void encode_txfm_probs(VP9_COMMON *cm, vp9_writer *w) {
807 vp9_write_literal(w, MIN(cm->tx_mode, ALLOW_32X32), 2);
808 if (cm->tx_mode >= ALLOW_32X32)
809 vp9_write_bit(w, cm->tx_mode == TX_MODE_SELECT);
812 if (cm->tx_mode == TX_MODE_SELECT) {
814 unsigned int ct_8x8p[TX_SIZES - 3][2];
815 unsigned int ct_16x16p[TX_SIZES - 2][2];
816 unsigned int ct_32x32p[TX_SIZES - 1][2];
819 for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
820 tx_counts_to_branch_counts_8x8(cm->counts.tx.p8x8[i], ct_8x8p);
821 for (j = 0; j < TX_SIZES - 3; j++)
822 vp9_cond_prob_diff_update(w, &cm->fc.tx_probs.p8x8[i][j], ct_8x8p[j]);
825 for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
826 tx_counts_to_branch_counts_16x16(cm->counts.tx.p16x16[i], ct_16x16p);
827 for (j = 0; j < TX_SIZES - 2; j++)
828 vp9_cond_prob_diff_update(w, &cm->fc.tx_probs.p16x16[i][j],
832 for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
833 tx_counts_to_branch_counts_32x32(cm->counts.tx.p32x32[i], ct_32x32p);
834 for (j = 0; j < TX_SIZES - 1; j++)
835 vp9_cond_prob_diff_update(w, &cm->fc.tx_probs.p32x32[i][j],
841 static void write_interp_filter(INTERP_FILTER filter,
842 struct vp9_write_bit_buffer *wb) {
843 const int filter_to_literal[] = { 1, 0, 2, 3 };
845 vp9_wb_write_bit(wb, filter == SWITCHABLE);
846 if (filter != SWITCHABLE)
847 vp9_wb_write_literal(wb, filter_to_literal[filter], 2);
850 static void fix_interp_filter(VP9_COMMON *cm) {
851 if (cm->interp_filter == SWITCHABLE) {
852 // Check to see if only one of the filters is actually used
853 int count[SWITCHABLE_FILTERS];
855 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
857 for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
858 count[i] += cm->counts.switchable_interp[j][i];
862 // Only one filter is used. So set the filter at frame level
863 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
865 cm->interp_filter = i;
873 static void write_tile_info(VP9_COMMON *cm, struct vp9_write_bit_buffer *wb) {
874 int min_log2_tile_cols, max_log2_tile_cols, ones;
875 vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
878 ones = cm->log2_tile_cols - min_log2_tile_cols;
880 vp9_wb_write_bit(wb, 1);
882 if (cm->log2_tile_cols < max_log2_tile_cols)
883 vp9_wb_write_bit(wb, 0);
886 vp9_wb_write_bit(wb, cm->log2_tile_rows != 0);
887 if (cm->log2_tile_rows != 0)
888 vp9_wb_write_bit(wb, cm->log2_tile_rows != 1);
891 static int get_refresh_mask(VP9_COMP *cpi) {
892 // Should the GF or ARF be updated using the transmitted frame or buffer
893 #if CONFIG_MULTIPLE_ARF
894 if (!cpi->multi_arf_enabled && cpi->refresh_golden_frame &&
895 !cpi->refresh_alt_ref_frame) {
897 if (cpi->refresh_golden_frame && !cpi->refresh_alt_ref_frame &&
900 // Preserve the previously existing golden frame and update the frame in
901 // the alt ref slot instead. This is highly specific to the use of
902 // alt-ref as a forward reference, and this needs to be generalized as
903 // other uses are implemented (like RTC/temporal scaling)
905 // gld_fb_idx and alt_fb_idx need to be swapped for future frames, but
906 // that happens in vp9_encoder.c:update_reference_frames() so that it can
907 // be done outside of the recode loop.
908 return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
909 (cpi->refresh_golden_frame << cpi->alt_fb_idx);
911 int arf_idx = cpi->alt_fb_idx;
912 #if CONFIG_MULTIPLE_ARF
913 // Determine which ARF buffer to use to encode this ARF frame.
914 if (cpi->multi_arf_enabled) {
915 int sn = cpi->sequence_number;
916 arf_idx = (cpi->frame_coding_order[sn] < 0) ?
917 cpi->arf_buffer_idx[sn + 1] :
918 cpi->arf_buffer_idx[sn];
921 return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
922 (cpi->refresh_golden_frame << cpi->gld_fb_idx) |
923 (cpi->refresh_alt_ref_frame << arf_idx);
927 static size_t encode_tiles(VP9_COMP *cpi, uint8_t *data_ptr) {
928 VP9_COMMON *const cm = &cpi->common;
929 vp9_writer residual_bc;
931 int tile_row, tile_col;
932 TOKENEXTRA *tok[4][1 << 6], *tok_end;
933 size_t total_size = 0;
934 const int tile_cols = 1 << cm->log2_tile_cols;
935 const int tile_rows = 1 << cm->log2_tile_rows;
937 vpx_memset(cm->above_seg_context, 0, sizeof(*cm->above_seg_context) *
938 mi_cols_aligned_to_sb(cm->mi_cols));
940 tok[0][0] = cpi->tok;
941 for (tile_row = 0; tile_row < tile_rows; tile_row++) {
943 tok[tile_row][0] = tok[tile_row - 1][tile_cols - 1] +
944 cpi->tok_count[tile_row - 1][tile_cols - 1];
946 for (tile_col = 1; tile_col < tile_cols; tile_col++)
947 tok[tile_row][tile_col] = tok[tile_row][tile_col - 1] +
948 cpi->tok_count[tile_row][tile_col - 1];
951 for (tile_row = 0; tile_row < tile_rows; tile_row++) {
952 for (tile_col = 0; tile_col < tile_cols; tile_col++) {
955 vp9_tile_init(&tile, cm, tile_row, tile_col);
956 tok_end = tok[tile_row][tile_col] + cpi->tok_count[tile_row][tile_col];
958 if (tile_col < tile_cols - 1 || tile_row < tile_rows - 1)
959 vp9_start_encode(&residual_bc, data_ptr + total_size + 4);
961 vp9_start_encode(&residual_bc, data_ptr + total_size);
963 write_modes(cpi, &tile, &residual_bc, &tok[tile_row][tile_col], tok_end);
964 assert(tok[tile_row][tile_col] == tok_end);
965 vp9_stop_encode(&residual_bc);
966 if (tile_col < tile_cols - 1 || tile_row < tile_rows - 1) {
968 mem_put_be32(data_ptr + total_size, residual_bc.pos);
972 total_size += residual_bc.pos;
979 static void write_display_size(const VP9_COMMON *cm,
980 struct vp9_write_bit_buffer *wb) {
981 const int scaling_active = cm->width != cm->display_width ||
982 cm->height != cm->display_height;
983 vp9_wb_write_bit(wb, scaling_active);
984 if (scaling_active) {
985 vp9_wb_write_literal(wb, cm->display_width - 1, 16);
986 vp9_wb_write_literal(wb, cm->display_height - 1, 16);
990 static void write_frame_size(const VP9_COMMON *cm,
991 struct vp9_write_bit_buffer *wb) {
992 vp9_wb_write_literal(wb, cm->width - 1, 16);
993 vp9_wb_write_literal(wb, cm->height - 1, 16);
995 write_display_size(cm, wb);
998 static void write_frame_size_with_refs(VP9_COMP *cpi,
999 struct vp9_write_bit_buffer *wb) {
1000 VP9_COMMON *const cm = &cpi->common;
1003 MV_REFERENCE_FRAME ref_frame;
1004 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
1005 YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi, ref_frame);
1006 found = cm->width == cfg->y_crop_width &&
1007 cm->height == cfg->y_crop_height;
1009 // Set "found" to 0 for temporal svc and for spatial svc key frame
1011 (cpi->svc.number_spatial_layers == 1 ||
1012 cpi->svc.layer_context[cpi->svc.spatial_layer_id].is_key_frame)) {
1015 vp9_wb_write_bit(wb, found);
1022 vp9_wb_write_literal(wb, cm->width - 1, 16);
1023 vp9_wb_write_literal(wb, cm->height - 1, 16);
1026 write_display_size(cm, wb);
1029 static void write_sync_code(struct vp9_write_bit_buffer *wb) {
1030 vp9_wb_write_literal(wb, VP9_SYNC_CODE_0, 8);
1031 vp9_wb_write_literal(wb, VP9_SYNC_CODE_1, 8);
1032 vp9_wb_write_literal(wb, VP9_SYNC_CODE_2, 8);
1035 static void write_profile(BITSTREAM_PROFILE profile,
1036 struct vp9_write_bit_buffer *wb) {
1037 assert(profile < MAX_PROFILES);
1038 vp9_wb_write_bit(wb, profile & 1);
1039 vp9_wb_write_bit(wb, profile >> 1);
1042 static void write_uncompressed_header(VP9_COMP *cpi,
1043 struct vp9_write_bit_buffer *wb) {
1044 VP9_COMMON *const cm = &cpi->common;
1046 vp9_wb_write_literal(wb, VP9_FRAME_MARKER, 2);
1048 write_profile(cm->profile, wb);
1050 vp9_wb_write_bit(wb, 0); // show_existing_frame
1051 vp9_wb_write_bit(wb, cm->frame_type);
1052 vp9_wb_write_bit(wb, cm->show_frame);
1053 vp9_wb_write_bit(wb, cm->error_resilient_mode);
1055 if (cm->frame_type == KEY_FRAME) {
1056 const COLOR_SPACE cs = UNKNOWN;
1057 write_sync_code(wb);
1058 if (cm->profile > PROFILE_1) {
1059 assert(cm->bit_depth > BITS_8);
1060 vp9_wb_write_bit(wb, cm->bit_depth - BITS_10);
1062 vp9_wb_write_literal(wb, cs, 3);
1064 vp9_wb_write_bit(wb, 0); // 0: [16, 235] (i.e. xvYCC), 1: [0, 255]
1065 if (cm->profile >= PROFILE_1) {
1066 vp9_wb_write_bit(wb, cm->subsampling_x);
1067 vp9_wb_write_bit(wb, cm->subsampling_y);
1068 vp9_wb_write_bit(wb, 0); // has extra plane
1071 assert(cm->profile == PROFILE_1);
1072 vp9_wb_write_bit(wb, 0); // has extra plane
1075 write_frame_size(cm, wb);
1077 if (!cm->show_frame)
1078 vp9_wb_write_bit(wb, cm->intra_only);
1080 if (!cm->error_resilient_mode)
1081 vp9_wb_write_literal(wb, cm->reset_frame_context, 2);
1083 if (cm->intra_only) {
1084 write_sync_code(wb);
1086 vp9_wb_write_literal(wb, get_refresh_mask(cpi), REF_FRAMES);
1087 write_frame_size(cm, wb);
1089 MV_REFERENCE_FRAME ref_frame;
1090 vp9_wb_write_literal(wb, get_refresh_mask(cpi), REF_FRAMES);
1091 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
1092 vp9_wb_write_literal(wb, get_ref_frame_idx(cpi, ref_frame),
1094 vp9_wb_write_bit(wb, cm->ref_frame_sign_bias[ref_frame]);
1097 write_frame_size_with_refs(cpi, wb);
1099 vp9_wb_write_bit(wb, cm->allow_high_precision_mv);
1101 fix_interp_filter(cm);
1102 write_interp_filter(cm->interp_filter, wb);
1106 if (!cm->error_resilient_mode) {
1107 vp9_wb_write_bit(wb, cm->refresh_frame_context);
1108 vp9_wb_write_bit(wb, cm->frame_parallel_decoding_mode);
1111 vp9_wb_write_literal(wb, cm->frame_context_idx, FRAME_CONTEXTS_LOG2);
1113 encode_loopfilter(&cm->lf, wb);
1114 encode_quantization(cm, wb);
1115 encode_segmentation(cpi, wb);
1117 write_tile_info(cm, wb);
1120 static size_t write_compressed_header(VP9_COMP *cpi, uint8_t *data) {
1121 VP9_COMMON *const cm = &cpi->common;
1122 MACROBLOCKD *const xd = &cpi->mb.e_mbd;
1123 FRAME_CONTEXT *const fc = &cm->fc;
1124 vp9_writer header_bc;
1126 vp9_start_encode(&header_bc, data);
1129 cm->tx_mode = ONLY_4X4;
1131 encode_txfm_probs(cm, &header_bc);
1133 update_coef_probs(cpi, &header_bc);
1134 update_skip_probs(cm, &header_bc);
1136 if (!frame_is_intra_only(cm)) {
1139 for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
1140 prob_diff_update(vp9_inter_mode_tree, cm->fc.inter_mode_probs[i],
1141 cm->counts.inter_mode[i], INTER_MODES, &header_bc);
1143 vp9_zero(cm->counts.inter_mode);
1145 if (cm->interp_filter == SWITCHABLE)
1146 update_switchable_interp_probs(cm, &header_bc);
1148 for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
1149 vp9_cond_prob_diff_update(&header_bc, &fc->intra_inter_prob[i],
1150 cm->counts.intra_inter[i]);
1152 if (cm->allow_comp_inter_inter) {
1153 const int use_compound_pred = cm->reference_mode != SINGLE_REFERENCE;
1154 const int use_hybrid_pred = cm->reference_mode == REFERENCE_MODE_SELECT;
1156 vp9_write_bit(&header_bc, use_compound_pred);
1157 if (use_compound_pred) {
1158 vp9_write_bit(&header_bc, use_hybrid_pred);
1159 if (use_hybrid_pred)
1160 for (i = 0; i < COMP_INTER_CONTEXTS; i++)
1161 vp9_cond_prob_diff_update(&header_bc, &fc->comp_inter_prob[i],
1162 cm->counts.comp_inter[i]);
1166 if (cm->reference_mode != COMPOUND_REFERENCE) {
1167 for (i = 0; i < REF_CONTEXTS; i++) {
1168 vp9_cond_prob_diff_update(&header_bc, &fc->single_ref_prob[i][0],
1169 cm->counts.single_ref[i][0]);
1170 vp9_cond_prob_diff_update(&header_bc, &fc->single_ref_prob[i][1],
1171 cm->counts.single_ref[i][1]);
1175 if (cm->reference_mode != SINGLE_REFERENCE)
1176 for (i = 0; i < REF_CONTEXTS; i++)
1177 vp9_cond_prob_diff_update(&header_bc, &fc->comp_ref_prob[i],
1178 cm->counts.comp_ref[i]);
1180 for (i = 0; i < BLOCK_SIZE_GROUPS; ++i)
1181 prob_diff_update(vp9_intra_mode_tree, cm->fc.y_mode_prob[i],
1182 cm->counts.y_mode[i], INTRA_MODES, &header_bc);
1184 for (i = 0; i < PARTITION_CONTEXTS; ++i)
1185 prob_diff_update(vp9_partition_tree, fc->partition_prob[i],
1186 cm->counts.partition[i], PARTITION_TYPES, &header_bc);
1188 vp9_write_nmv_probs(cm, cm->allow_high_precision_mv, &header_bc);
1191 vp9_stop_encode(&header_bc);
1192 assert(header_bc.pos <= 0xffff);
1194 return header_bc.pos;
1197 void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, size_t *size) {
1198 uint8_t *data = dest;
1199 size_t first_part_size, uncompressed_hdr_size;
1200 struct vp9_write_bit_buffer wb = {data, 0};
1201 struct vp9_write_bit_buffer saved_wb;
1203 write_uncompressed_header(cpi, &wb);
1205 vp9_wb_write_literal(&wb, 0, 16); // don't know in advance first part. size
1207 uncompressed_hdr_size = vp9_rb_bytes_written(&wb);
1208 data += uncompressed_hdr_size;
1210 vp9_compute_update_table();
1212 vp9_clear_system_state();
1214 first_part_size = write_compressed_header(cpi, data);
1215 data += first_part_size;
1216 // TODO(jbb): Figure out what to do if first_part_size > 16 bits.
1217 vp9_wb_write_literal(&saved_wb, (int)first_part_size, 16);
1219 data += encode_tiles(cpi, data);
1221 *size = data - dest;