1 /*
2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
3 *
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.
9 */
10
11 #include <assert.h>
12
13 #include "bitwriter_buffer.h"
14 #include "vpx_dsp_common.h"
15
16 #include "vp9_treewriter.h"
17 #include "vp9_filter.h"
18 #include "vp9_enums.h"
19 #include "bitwriter.h"
20 #include "vp9_entropymode.h"
21 #include "vp9_tokenize.h"
22 #include "vpx_codec.h"
23 #include "vp9_common.h"
24 #include "vp9_subexp.h"
25 #include "vp9_onyxc_int.h"
26 #include "vp9_pred_common.h"
27 #include "vp9_block.h"
28 #include "vp9_encoder.h"
29 #include "vp9_encodemv.h"
30 #include "vp9_blockd.h"
31 #include "vp9_bitstream.h"
32 #include "vp9_seg_common.h"
33 #include "vp9_speed_features.h"
34 #include "vp9_cost.h"
35 #include "vp9_quant_common.h"
36
37 static const struct vp9_token intra_mode_encodings[INTRA_MODES] = {
38 { 0, 1 }, { 6, 3 }, { 28, 5 }, { 30, 5 }, { 58, 6 },
39 { 59, 6 }, { 126, 7 }, { 127, 7 }, { 62, 6 }, { 2, 2 }
40 };
41 static const struct vp9_token partition_encodings[PARTITION_TYPES] = {
42 { 0, 1 }, { 2, 2 }, { 6, 3 }, { 7, 3 }
43 };
44 static const struct vp9_token inter_mode_encodings[INTER_MODES] = {
45 { 2, 2 }, { 6, 3 }, { 0, 1 }, { 7, 3 }
46 };
47
write_intra_mode(VpxWriter * w,PREDICTION_MODE mode,const vpx_prob * probs)48 static void write_intra_mode(VpxWriter *w, PREDICTION_MODE mode,
49 const vpx_prob *probs) {
50 vp9_write_token(w, eb_vp9_intra_mode_tree, probs, &intra_mode_encodings[mode]);
51 }
52
write_inter_mode(VpxWriter * w,PREDICTION_MODE mode,const vpx_prob * probs)53 static void write_inter_mode(VpxWriter *w, PREDICTION_MODE mode,
54 const vpx_prob *probs) {
55 assert(is_inter_mode(mode));
56 vp9_write_token(w, eb_vp9_inter_mode_tree, probs,
57 &inter_mode_encodings[INTER_OFFSET(mode)]);
58 }
59 #if SEG_SUPPORT
encode_unsigned_max(struct vpx_write_bit_buffer * wb,int data,int max)60 static void encode_unsigned_max(struct vpx_write_bit_buffer *wb, int data,
61 int max) {
62 eb_vp9_wb_write_literal(wb, data, get_unsigned_bits(max));
63 }
64 #endif
65
prob_diff_update(const vpx_tree_index * tree,vpx_prob probs[],const unsigned int counts[],int n,VpxWriter * w)66 static void prob_diff_update(const vpx_tree_index *tree,
67 vpx_prob probs[/*n - 1*/],
68 const unsigned int counts[/*n - 1*/], int n,
69 VpxWriter *w) {
70 int i;
71 unsigned int branch_ct[32][2];
72
73 // Assuming max number of probabilities <= 32
74 assert(n <= 32);
75
76 eb_vp9_tree_probs_from_distribution(tree, branch_ct, counts);
77 for (i = 0; i < n - 1; ++i)
78 eb_vp9_cond_prob_diff_update(w, &probs[i], branch_ct[i]);
79 }
80
write_selected_tx_size(const VP9_COMMON * cm,const MACROBLOCKD * const xd,VpxWriter * w)81 static void write_selected_tx_size(const VP9_COMMON *cm,
82 const MACROBLOCKD *const xd, VpxWriter *w) {
83 TX_SIZE tx_size = xd->mi[0]->tx_size;
84 BLOCK_SIZE bsize = xd->mi[0]->sb_type;
85 const TX_SIZE max_tx_size = eb_vp9_max_txsize_lookup[bsize];
86 const vpx_prob *const tx_probs =
87 get_tx_probs(max_tx_size, get_tx_size_context(xd), &cm->fc->tx_probs);
88 vpx_write(w, tx_size != TX_4X4, tx_probs[0]);
89 if (tx_size != TX_4X4 && max_tx_size >= TX_16X16) {
90 vpx_write(w, tx_size != TX_8X8, tx_probs[1]);
91 if (tx_size != TX_8X8 && max_tx_size >= TX_32X32)
92 vpx_write(w, tx_size != TX_16X16, tx_probs[2]);
93 }
94 }
95
write_skip(const VP9_COMMON * cm,const MACROBLOCKD * const xd,int segment_id,const ModeInfo * mi,VpxWriter * w)96 static int write_skip(const VP9_COMMON *cm, const MACROBLOCKD *const xd,
97 int segment_id, const ModeInfo *mi, VpxWriter *w) {
98 if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
99 return 1;
100 } else {
101 const int skip = mi->skip;
102 vpx_write(w, skip, vp9_get_skip_prob(cm, xd));
103 return skip;
104 }
105 }
106
update_skip_probs(VP9_COMMON * cm,VpxWriter * w,FRAME_COUNTS * counts)107 static void update_skip_probs(VP9_COMMON *cm, VpxWriter *w,
108 FRAME_COUNTS *counts) {
109 int k;
110
111 for (k = 0; k < SKIP_CONTEXTS; ++k)
112 eb_vp9_cond_prob_diff_update(w, &cm->fc->skip_probs[k], counts->skip[k]);
113 }
114
pack_mb_tokens(VpxWriter * w,TOKENEXTRA ** tp,const TOKENEXTRA * const stop,vpx_bit_depth_t bit_depth)115 static void pack_mb_tokens(VpxWriter *w, TOKENEXTRA **tp,
116 const TOKENEXTRA *const stop,
117 vpx_bit_depth_t bit_depth) {
118 const TOKENEXTRA *p;
119 const vp9_extra_bit *const extra_bits =
120 #if CONFIG_VP9_HIGHBITDEPTH
121 (bit_depth == VPX_BITS_12)
122 ? eb_vp9_extra_bits_high12
123 : (bit_depth == VPX_BITS_10) ? eb_vp9_extra_bits_high10 : eb_vp9_extra_bits;
124 #else
125 eb_vp9_extra_bits;
126 (void)bit_depth;
127 #endif // CONFIG_VP9_HIGHBITDEPTH
128
129 for (p = *tp; p < stop && p->token != EOSB_TOKEN; ++p) {
130 if (p->token == EOB_TOKEN) {
131 vpx_write(w, 0, p->context_tree[0]);
132 continue;
133 }
134 vpx_write(w, 1, p->context_tree[0]);
135 while (p->token == ZERO_TOKEN) {
136 vpx_write(w, 0, p->context_tree[1]);
137 ++p;
138 if (p == stop || p->token == EOSB_TOKEN) {
139 *tp = (TOKENEXTRA *)(uintptr_t)p + (p->token == EOSB_TOKEN);
140 return;
141 }
142 }
143
144 {
145 const int t = p->token;
146 const vpx_prob *const context_tree = p->context_tree;
147 assert(t != ZERO_TOKEN);
148 assert(t != EOB_TOKEN);
149 assert(t != EOSB_TOKEN);
150 vpx_write(w, 1, context_tree[1]);
151 if (t == ONE_TOKEN) {
152 vpx_write(w, 0, context_tree[2]);
153 vpx_write_bit(w, p->extra & 1);
154 } else { // t >= TWO_TOKEN && t < EOB_TOKEN
155 const struct vp9_token *const a = &eb_vp9_coef_encodings[t];
156 const int v = a->value;
157 const int n = a->len;
158 const int e = p->extra;
159 vpx_write(w, 1, context_tree[2]);
160 vp9_write_tree(w, eb_vp9_coef_con_tree,
161 eb_vp9_pareto8_full[context_tree[PIVOT_NODE] - 1], v,
162 n - UNCONSTRAINED_NODES, 0);
163 if (t >= CATEGORY1_TOKEN) {
164 const vp9_extra_bit *const b = &extra_bits[t];
165 const unsigned char *pb = b->prob;
166 int v = e >> 1;
167 int n = b->len; // number of bits in v, assumed nonzero
168 do {
169 const int bb = (v >> --n) & 1;
170 vpx_write(w, bb, *pb++);
171 } while (n);
172 }
173 vpx_write_bit(w, e & 1);
174 }
175 }
176 }
177 *tp = (TOKENEXTRA *)(uintptr_t)p + (p->token == EOSB_TOKEN);
178 }
179 #if SEG_SUPPORT
write_segment_id(VpxWriter * w,const struct segmentation * seg,int segment_id)180 static void write_segment_id(VpxWriter *w, const struct segmentation *seg,
181 int segment_id) {
182 if (seg->enabled && seg->update_map)
183 vp9_write_tree(w, eb_vp9_segment_tree, seg->tree_probs, segment_id, 3, 0);
184 }
185 #endif
186 // This function encodes the reference frame
write_ref_frames(const VP9_COMMON * cm,const MACROBLOCKD * const xd,VpxWriter * w)187 static void write_ref_frames(const VP9_COMMON *cm, const MACROBLOCKD *const xd,
188 VpxWriter *w) {
189 const ModeInfo *const mi = xd->mi[0];
190 const int is_compound = has_second_ref(mi);
191 #if !SEG_SUPPORT // Hsan: segmentation not supported
192 const int segment_id = 0;
193 #else
194 const int segment_id = mi->segment_id;
195 #endif
196 // If segment level coding of this signal is disabled...
197 // or the segment allows multiple reference frame options
198 if (segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
199 assert(!is_compound);
200 assert(mi->ref_frame[0] ==
201 get_segdata(&cm->seg, segment_id, SEG_LVL_REF_FRAME));
202 } else {
203 // does the feature use compound prediction or not
204 // (if not specified at the frame/segment level)
205 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
206 vpx_write(w, is_compound, vp9_get_reference_mode_prob(cm, xd));
207 } else {
208 assert((!is_compound) == (cm->reference_mode == SINGLE_REFERENCE));
209 }
210
211 if (is_compound) {
212 const int idx = cm->ref_frame_sign_bias[cm->comp_fixed_ref];
213 vpx_write(w, mi->ref_frame[!idx] == cm->comp_var_ref[1],
214 vp9_get_pred_prob_comp_ref_p(cm, xd));
215 } else {
216 const int bit0 = mi->ref_frame[0] != LAST_FRAME;
217 vpx_write(w, bit0, vp9_get_pred_prob_single_ref_p1(cm, xd));
218 if (bit0) {
219 const int bit1 = mi->ref_frame[0] != GOLDEN_FRAME;
220 vpx_write(w, bit1, vp9_get_pred_prob_single_ref_p2(cm, xd));
221 }
222 }
223 }
224 }
225
pack_inter_mode_mvs(VP9_COMP * cpi,const MACROBLOCKD * const xd,const MbModeInfoExt * const mbmi_ext,VpxWriter * w,unsigned int * const max_mv_magnitude)226 static void pack_inter_mode_mvs(
227 VP9_COMP *cpi, const MACROBLOCKD *const xd,
228 const MbModeInfoExt *const mbmi_ext, VpxWriter *w,
229 unsigned int *const max_mv_magnitude
230 #if 0
231 ,
232 int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE]
233 #endif
234 ) {
235 VP9_COMMON *const cm = &cpi->common;
236 const nmv_context *nmvc = &cm->fc->nmvc;
237 const struct segmentation *const seg = &cm->seg;
238 const ModeInfo *const mi = xd->mi[0];
239 const PREDICTION_MODE mode = mi->mode;
240 #if !SEG_SUPPORT // Hsan: segmentation not supported
241 const int segment_id = 0;
242 #else
243 const int segment_id = mi->segment_id;
244 #endif
245 const BLOCK_SIZE bsize = mi->sb_type;
246 const int allow_hp = cm->allow_high_precision_mv;
247 const int is_inter = is_inter_block(mi);
248 const int is_compound = has_second_ref(mi);
249 int skip, ref;
250 #if SEG_SUPPORT // Hsan: temporal_update not supported
251 if (seg->update_map) {
252 #if 0// Hsan: temporal_update not supported
253 if (seg->temporal_update) {
254 const int pred_flag = mi->seg_id_predicted;
255 vpx_prob pred_prob = vp9_get_pred_prob_seg_id(seg, xd);
256 vpx_write(w, pred_flag, pred_prob);
257 if (!pred_flag) write_segment_id(w, seg, segment_id);
258 } else
259 #endif
260 {
261 write_segment_id(w, seg, segment_id);
262 }
263 }
264 #endif
265 skip = write_skip(cm, xd, segment_id, mi, w);
266
267 if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
268 vpx_write(w, is_inter, vp9_get_intra_inter_prob(cm, xd));
269
270 if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT &&
271 !(is_inter && skip)) {
272 write_selected_tx_size(cm, xd, w);
273 }
274
275 if (!is_inter) {
276 if (bsize >= BLOCK_8X8) {
277 write_intra_mode(w, mode, cm->fc->y_mode_prob[eb_vp9_size_group_lookup[bsize]]);
278 } else {
279 int idx, idy;
280 const int num_4x4_w = eb_vp9_num_4x4_blocks_wide_lookup[bsize];
281 const int num_4x4_h = eb_vp9_num_4x4_blocks_high_lookup[bsize];
282 for (idy = 0; idy < 2; idy += num_4x4_h) {
283 for (idx = 0; idx < 2; idx += num_4x4_w) {
284 const PREDICTION_MODE b_mode = mi->bmi[idy * 2 + idx].as_mode;
285 write_intra_mode(w, b_mode, cm->fc->y_mode_prob[0]);
286 }
287 }
288 }
289 write_intra_mode(w, mi->uv_mode, cm->fc->uv_mode_prob[mode]);
290 } else {
291 const int mode_ctx = mbmi_ext->mode_context[mi->ref_frame[0]];
292 const vpx_prob *const inter_probs = cm->fc->inter_mode_probs[mode_ctx];
293 write_ref_frames(cm, xd, w);
294
295 // If segment skip is not enabled code the mode.
296 if (!segfeature_active(seg, segment_id, SEG_LVL_SKIP)) {
297 if (bsize >= BLOCK_8X8) {
298 write_inter_mode(w, mode, inter_probs);
299 }
300 }
301 #if 0 // Hsan: switchable interp_filter not supported
302 if (cm->interp_filter == SWITCHABLE) {
303 const int ctx = get_pred_context_switchable_interp(xd);
304 vp9_write_token(w, eb_vp9_switchable_interp_tree,
305 cm->fc->switchable_interp_prob[ctx],
306 &switchable_interp_encodings[mi->interp_filter]);
307 ++interp_filter_selected[0][mi->interp_filter];
308 } else {
309 assert(mi->interp_filter == cm->interp_filter);
310 }
311 #endif
312 if (bsize < BLOCK_8X8) {
313 const int num_4x4_w = eb_vp9_num_4x4_blocks_wide_lookup[bsize];
314 const int num_4x4_h = eb_vp9_num_4x4_blocks_high_lookup[bsize];
315 int idx, idy;
316 for (idy = 0; idy < 2; idy += num_4x4_h) {
317 for (idx = 0; idx < 2; idx += num_4x4_w) {
318 const int j = idy * 2 + idx;
319 const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
320 write_inter_mode(w, b_mode, inter_probs);
321 if (b_mode == NEWMV) {
322 for (ref = 0; ref < 1 + is_compound; ++ref)
323 eb_vp9_encode_mv(cpi, w, &mi->bmi[j].as_mv[ref].as_mv,
324 &mbmi_ext->ref_mvs[mi->ref_frame[ref]][0].as_mv,
325 nmvc, allow_hp, max_mv_magnitude);
326 }
327 }
328 }
329 } else {
330 if (mode == NEWMV) {
331 for (ref = 0; ref < 1 + is_compound; ++ref)
332 eb_vp9_encode_mv(cpi, w, &mi->mv[ref].as_mv,
333 &mbmi_ext->ref_mvs[mi->ref_frame[ref]][0].as_mv, nmvc,
334 allow_hp, max_mv_magnitude);
335 }
336 }
337 }
338 }
339
write_mb_modes_kf(const VP9_COMMON * cm,const MACROBLOCKD * xd,VpxWriter * w)340 static void write_mb_modes_kf(const VP9_COMMON *cm, const MACROBLOCKD *xd,
341 VpxWriter *w) {
342 #if SEG_SUPPORT
343 const struct segmentation *const seg = &cm->seg;
344 #endif
345 const ModeInfo *const mi = xd->mi[0];
346 const ModeInfo *const above_mi = xd->above_mi;
347 const ModeInfo *const left_mi = xd->left_mi;
348 const BLOCK_SIZE bsize = mi->sb_type;
349 #if !SEG_SUPPORT // Hsan: segmentation not supported
350 write_skip(cm, xd,0, mi, w);
351 #else
352 if (seg->update_map) write_segment_id(w, seg, mi->segment_id);
353
354 write_skip(cm, xd, mi->segment_id, mi, w);
355
356 #endif
357 if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT)
358 write_selected_tx_size(cm, xd, w);
359
360 if (bsize >= BLOCK_8X8) {
361 write_intra_mode(w, mi->mode, get_y_mode_probs(mi, above_mi, left_mi, 0));
362 } else {
363 const int num_4x4_w = eb_vp9_num_4x4_blocks_wide_lookup[bsize];
364 const int num_4x4_h = eb_vp9_num_4x4_blocks_high_lookup[bsize];
365 int idx, idy;
366
367 for (idy = 0; idy < 2; idy += num_4x4_h) {
368 for (idx = 0; idx < 2; idx += num_4x4_w) {
369 const int block = idy * 2 + idx;
370 write_intra_mode(w, mi->bmi[block].as_mode,
371 get_y_mode_probs(mi, above_mi, left_mi, block));
372 }
373 }
374 }
375 write_intra_mode(w, mi->uv_mode, eb_vp9_kf_uv_mode_prob[mi->mode]);
376 }
377
eb_vp9_write_modes_b(EntropyCodingContext * context_ptr,VP9_COMP * cpi,MACROBLOCKD * const xd,const TileInfo * const tile,VpxWriter * w,TOKENEXTRA ** tok,const TOKENEXTRA * const tok_end,int mi_row,int mi_col,unsigned int * const max_mv_magnitude,int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE])378 void eb_vp9_write_modes_b(
379 EntropyCodingContext *context_ptr,
380 VP9_COMP *cpi, MACROBLOCKD *const xd, const TileInfo *const tile,
381 VpxWriter *w, TOKENEXTRA **tok, const TOKENEXTRA *const tok_end,
382 int mi_row, int mi_col, unsigned int *const max_mv_magnitude,
383 int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE]) {
384
385 (void)mi_col;
386 (void)mi_row;
387 (void)tile;
388 (void)interp_filter_selected;
389 const VP9_COMMON *const cm = &cpi->common;
390 #if 1
391 const MbModeInfoExt *const mbmi_ext = context_ptr->block_ptr->mbmi_ext;
392 #else
393 const MbModeInfoExt *const mbmi_ext =
394 cpi->td.mb.mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
395 #endif
396 #if 0 // now done prior to eb_vp9_write_modes_b() call
397 ModeInfo *m;
398 xd->mi = cm->mi_grid_visible + (mi_row * cm->mi_stride + mi_col);
399
400 m = xd->mi[0];
401
402 set_mi_row_col(xd, tile, mi_row, eb_vp9_num_8x8_blocks_high_lookup[m->sb_type],
403 mi_col, eb_vp9_num_8x8_blocks_wide_lookup[m->sb_type], cm->mi_rows,
404 cm->mi_cols);
405 #endif
406 if (frame_is_intra_only(cm)) {
407 write_mb_modes_kf(cm, xd, w);
408 } else {
409 #if 0
410 pack_inter_mode_mvs(cpi, xd, mbmi_ext, w, max_mv_magnitude,
411 interp_filter_selected);
412 #else
413 pack_inter_mode_mvs(cpi, xd, mbmi_ext, w, max_mv_magnitude);
414 #endif
415 }
416
417 assert(*tok < tok_end);
418 pack_mb_tokens(w, tok, tok_end, cm->bit_depth);
419 }
420
write_partition(const VP9_COMMON * const cm,const MACROBLOCKD * const xd,int hbs,int mi_row,int mi_col,PARTITION_TYPE p,BLOCK_SIZE bsize,VpxWriter * w)421 void write_partition(const VP9_COMMON *const cm,
422 const MACROBLOCKD *const xd, int hbs, int mi_row,
423 int mi_col, PARTITION_TYPE p, BLOCK_SIZE bsize,
424 VpxWriter *w) {
425 const int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
426 const vpx_prob *const probs = xd->partition_probs[ctx];
427 const int has_rows = (mi_row + hbs) < cm->mi_rows;
428 const int has_cols = (mi_col + hbs) < cm->mi_cols;
429
430 if (has_rows && has_cols) {
431 vp9_write_token(w, eb_vp9_partition_tree, probs, &partition_encodings[p]);
432 } else if (!has_rows && has_cols) {
433 assert(p == PARTITION_SPLIT || p == PARTITION_HORZ);
434 vpx_write(w, p == PARTITION_SPLIT, probs[1]);
435 } else if (has_rows && !has_cols) {
436 assert(p == PARTITION_SPLIT || p == PARTITION_VERT);
437 vpx_write(w, p == PARTITION_SPLIT, probs[2]);
438 } else {
439 assert(p == PARTITION_SPLIT);
440 }
441 }
442 #if 0
443 static void write_modes_sb(
444 VP9_COMP *cpi, MACROBLOCKD *const xd, const TileInfo *const tile,
445 VpxWriter *w, TOKENEXTRA **tok, const TOKENEXTRA *const tok_end,
446 int mi_row, int mi_col, BLOCK_SIZE bsize,
447 unsigned int *const max_mv_magnitude,
448 int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE]) {
449 const VP9_COMMON *const cm = &cpi->common;
450 const int bsl = eb_vp9_b_width_log2_lookup[bsize];
451 const int bs = (1 << bsl) / 4;
452 PARTITION_TYPE partition;
453 BLOCK_SIZE subsize;
454 const ModeInfo *m = NULL;
455
456 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
457
458 m = cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col];
459
460 partition = eb_vp9_partition_lookup[bsl][m->sb_type];
461 write_partition(cm, xd, bs, mi_row, mi_col, partition, bsize, w);
462 subsize = get_subsize(bsize, partition);
463 if (subsize < BLOCK_8X8) {
464 eb_vp9_write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
465 max_mv_magnitude, interp_filter_selected);
466 } else {
467 switch (partition) {
468 case PARTITION_NONE:
469 eb_vp9_write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
470 max_mv_magnitude, interp_filter_selected);
471 break;
472 case PARTITION_HORZ:
473 eb_vp9_write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
474 max_mv_magnitude, interp_filter_selected);
475 if (mi_row + bs < cm->mi_rows)
476 eb_vp9_write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row + bs, mi_col,
477 max_mv_magnitude, interp_filter_selected);
478 break;
479 case PARTITION_VERT:
480 eb_vp9_write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
481 max_mv_magnitude, interp_filter_selected);
482 if (mi_col + bs < cm->mi_cols)
483 eb_vp9_write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col + bs,
484 max_mv_magnitude, interp_filter_selected);
485 break;
486 default:
487 assert(partition == PARTITION_SPLIT);
488 write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col, subsize,
489 max_mv_magnitude, interp_filter_selected);
490 write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col + bs,
491 subsize, max_mv_magnitude, interp_filter_selected);
492 write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row + bs, mi_col,
493 subsize, max_mv_magnitude, interp_filter_selected);
494 write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row + bs, mi_col + bs,
495 subsize, max_mv_magnitude, interp_filter_selected);
496 break;
497 }
498 }
499
500 // update partition context
501 if (bsize >= BLOCK_8X8 &&
502 (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
503 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
504 }
505
506 static void write_modes(
507 VP9_COMP *cpi, MACROBLOCKD *const xd, const TileInfo *const tile,
508 VpxWriter *w, int tile_row, int tile_col,
509 unsigned int *const max_mv_magnitude,
510 int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE]) {
511 const VP9_COMMON *const cm = &cpi->common;
512 int mi_row, mi_col, tile_sb_row;
513 TOKENEXTRA *tok = NULL;
514 TOKENEXTRA *tok_end = NULL;
515
516 set_partition_probs(cm, xd);
517
518 for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
519 mi_row += MI_BLOCK_SIZE) {
520 tile_sb_row = mi_cols_aligned_to_sb(mi_row - tile->mi_row_start) >>
521 MI_BLOCK_SIZE_LOG2;
522 tok = cpi->tplist[tile_row][tile_col][tile_sb_row].start;
523 tok_end = tok + cpi->tplist[tile_row][tile_col][tile_sb_row].count;
524
525 vp9_zero(xd->left_seg_context);
526 for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
527 mi_col += MI_BLOCK_SIZE)
528 write_modes_sb(cpi, xd, tile, w, &tok, tok_end, mi_row, mi_col,
529 BLOCK_64X64, max_mv_magnitude, interp_filter_selected);
530
531 assert(tok == cpi->tplist[tile_row][tile_col][tile_sb_row].stop);
532 }
533 }
534 #endif
build_tree_distribution(VP9_COMP * cpi,TX_SIZE tx_size,vp9_coeff_stats * coef_branch_ct,vp9_coeff_probs_model * coef_probs)535 static void build_tree_distribution(VP9_COMP *cpi, TX_SIZE tx_size,
536 vp9_coeff_stats *coef_branch_ct,
537 vp9_coeff_probs_model *coef_probs) {
538 vp9_coeff_count *coef_counts = cpi->td.rd_counts.coef_counts[tx_size];
539 unsigned int(*eob_branch_ct)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS] =
540 cpi->common.counts.eob_branch[tx_size];
541 int i, j, k, l, m;
542
543 for (i = 0; i < PLANE_TYPES; ++i) {
544 for (j = 0; j < REF_TYPES; ++j) {
545 for (k = 0; k < COEF_BANDS; ++k) {
546 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
547 eb_vp9_tree_probs_from_distribution(eb_vp9_coef_tree,
548 coef_branch_ct[i][j][k][l],
549 coef_counts[i][j][k][l]);
550 coef_branch_ct[i][j][k][l][0][1] =
551 eob_branch_ct[i][j][k][l] - coef_branch_ct[i][j][k][l][0][0];
552 for (m = 0; m < UNCONSTRAINED_NODES; ++m)
553 coef_probs[i][j][k][l][m] =
554 get_binary_prob(coef_branch_ct[i][j][k][l][m][0],
555 coef_branch_ct[i][j][k][l][m][1]);
556 }
557 }
558 }
559 }
560 }
561
update_coef_probs_common(VpxWriter * const bc,VP9_COMP * cpi,TX_SIZE tx_size,vp9_coeff_stats * frame_branch_ct,vp9_coeff_probs_model * new_coef_probs)562 static void update_coef_probs_common(VpxWriter *const bc, VP9_COMP *cpi,
563 TX_SIZE tx_size,
564 vp9_coeff_stats *frame_branch_ct,
565 vp9_coeff_probs_model *new_coef_probs) {
566 vp9_coeff_probs_model *old_coef_probs = cpi->common.fc->coef_probs[tx_size];
567 const vpx_prob upd = DIFF_UPDATE_PROB;
568 const int entropy_nodes_update = UNCONSTRAINED_NODES;
569 int i, j, k, l, t;
570 int stepsize = cpi->sf.coeff_prob_appx_step;
571
572 switch (cpi->sf.use_fast_coef_updates) {
573 case TWO_LOOP: {
574 /* dry run to see if there is any update at all needed */
575 int savings = 0;
576 int update[2] = { 0, 0 };
577 for (i = 0; i < PLANE_TYPES; ++i) {
578 for (j = 0; j < REF_TYPES; ++j) {
579 for (k = 0; k < COEF_BANDS; ++k) {
580 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
581 for (t = 0; t < entropy_nodes_update; ++t) {
582 vpx_prob newp = new_coef_probs[i][j][k][l][t];
583 const vpx_prob oldp = old_coef_probs[i][j][k][l][t];
584 int s;
585 int u = 0;
586 if (t == PIVOT_NODE)
587 s = eb_vp9_prob_diff_update_savings_search_model(
588 frame_branch_ct[i][j][k][l][0], oldp, &newp, upd,
589 stepsize);
590 else
591 s = eb_vp9_prob_diff_update_savings_search(
592 frame_branch_ct[i][j][k][l][t], oldp, &newp, upd);
593 if (s > 0 && newp != oldp) u = 1;
594 if (u)
595 savings += s - (int)(vp9_cost_zero(upd));
596 else
597 savings -= (int)(vp9_cost_zero(upd));
598 update[u]++;
599 }
600 }
601 }
602 }
603 }
604
605 // SVT_LOG("Update %d %d, savings %d\n", update[0], update[1], savings);
606 /* Is coef updated at all */
607 if (update[1] == 0 || savings < 0) {
608 vpx_write_bit(bc, 0);
609 return;
610 }
611 vpx_write_bit(bc, 1);
612 for (i = 0; i < PLANE_TYPES; ++i) {
613 for (j = 0; j < REF_TYPES; ++j) {
614 for (k = 0; k < COEF_BANDS; ++k) {
615 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
616 // calc probs and branch cts for this frame only
617 for (t = 0; t < entropy_nodes_update; ++t) {
618 vpx_prob newp = new_coef_probs[i][j][k][l][t];
619 vpx_prob *oldp = old_coef_probs[i][j][k][l] + t;
620 const vpx_prob upd = DIFF_UPDATE_PROB;
621 int s;
622 int u = 0;
623 if (t == PIVOT_NODE)
624 s = eb_vp9_prob_diff_update_savings_search_model(
625 frame_branch_ct[i][j][k][l][0], *oldp, &newp, upd,
626 stepsize);
627 else
628 s = eb_vp9_prob_diff_update_savings_search(
629 frame_branch_ct[i][j][k][l][t], *oldp, &newp, upd);
630 if (s > 0 && newp != *oldp) u = 1;
631 vpx_write(bc, u, upd);
632 if (u) {
633 /* send/use new probability */
634 eb_vp9_write_prob_diff_update(bc, newp, *oldp);
635 *oldp = newp;
636 }
637 }
638 }
639 }
640 }
641 }
642 return;
643 }
644
645 default: {
646 int updates = 0;
647 int noupdates_before_first = 0;
648 assert(cpi->sf.use_fast_coef_updates == ONE_LOOP_REDUCED);
649 for (i = 0; i < PLANE_TYPES; ++i) {
650 for (j = 0; j < REF_TYPES; ++j) {
651 for (k = 0; k < COEF_BANDS; ++k) {
652 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
653 // calc probs and branch cts for this frame only
654 for (t = 0; t < entropy_nodes_update; ++t) {
655 vpx_prob newp = new_coef_probs[i][j][k][l][t];
656 vpx_prob *oldp = old_coef_probs[i][j][k][l] + t;
657 int s;
658 int u = 0;
659
660 if (t == PIVOT_NODE) {
661 s = eb_vp9_prob_diff_update_savings_search_model(
662 frame_branch_ct[i][j][k][l][0], *oldp, &newp, upd,
663 stepsize);
664 } else {
665 s = eb_vp9_prob_diff_update_savings_search(
666 frame_branch_ct[i][j][k][l][t], *oldp, &newp, upd);
667 }
668
669 if (s > 0 && newp != *oldp) u = 1;
670 updates += u;
671 if (u == 0 && updates == 0) {
672 noupdates_before_first++;
673 continue;
674 }
675 if (u == 1 && updates == 1) {
676 int v;
677 // first update
678 vpx_write_bit(bc, 1);
679 for (v = 0; v < noupdates_before_first; ++v)
680 vpx_write(bc, 0, upd);
681 }
682 vpx_write(bc, u, upd);
683 if (u) {
684 /* send/use new probability */
685 eb_vp9_write_prob_diff_update(bc, newp, *oldp);
686 *oldp = newp;
687 }
688 }
689 }
690 }
691 }
692 }
693 if (updates == 0) {
694 vpx_write_bit(bc, 0); // no updates
695 }
696 return;
697 }
698 }
699 }
700
update_coef_probs(VP9_COMP * cpi,VpxWriter * w)701 static void update_coef_probs(VP9_COMP *cpi, VpxWriter *w) {
702 const TX_MODE tx_mode = cpi->common.tx_mode;
703 if (tx_mode >= 5) {
704 return;
705 }
706 const TX_SIZE max_tx_size = eb_vp9_tx_mode_to_biggest_tx_size[tx_mode];
707 TX_SIZE tx_size;
708 for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size) {
709 vp9_coeff_stats frame_branch_ct[PLANE_TYPES];
710 vp9_coeff_probs_model frame_coef_probs[PLANE_TYPES];
711 if (cpi->td.counts->tx.tx_totals[tx_size] <= 20 ||
712 (tx_size >= TX_16X16 && cpi->sf.tx_size_search_method == USE_TX_8X8)) {
713 vpx_write_bit(w, 0);
714 } else {
715 build_tree_distribution(cpi, tx_size, frame_branch_ct, frame_coef_probs);
716 update_coef_probs_common(w, cpi, tx_size, frame_branch_ct,
717 frame_coef_probs);
718 }
719 }
720 }
721
encode_loopfilter(struct loop_filter * lf,struct vpx_write_bit_buffer * wb)722 static void encode_loopfilter(struct loop_filter *lf,
723 struct vpx_write_bit_buffer *wb) {
724 int i;
725
726 // Encode the loop filter level and type
727 eb_vp9_wb_write_literal(wb, lf->filter_level, 6);
728 eb_vp9_wb_write_literal(wb, lf->sharpness_level, 3);
729
730 // Write out loop filter deltas applied at the MB level based on mode or
731 // ref frame (if they are enabled).
732 eb_vp9_wb_write_bit(wb, lf->mode_ref_delta_enabled);
733
734 if (lf->mode_ref_delta_enabled) {
735 eb_vp9_wb_write_bit(wb, lf->mode_ref_delta_update);
736 if (lf->mode_ref_delta_update) {
737 for (i = 0; i < MAX_REF_LF_DELTAS; i++) {
738 const int delta = lf->ref_deltas[i];
739 const int changed = delta != lf->last_ref_deltas[i];
740 eb_vp9_wb_write_bit(wb, changed);
741 if (changed) {
742 lf->last_ref_deltas[i] = (signed char)delta;
743 eb_vp9_wb_write_literal(wb, abs(delta) & 0x3F, 6);
744 eb_vp9_wb_write_bit(wb, delta < 0);
745 }
746 }
747
748 for (i = 0; i < MAX_MODE_LF_DELTAS; i++) {
749 const int delta = lf->mode_deltas[i];
750 const int changed = delta != lf->last_mode_deltas[i];
751 eb_vp9_wb_write_bit(wb, changed);
752 if (changed) {
753 lf->last_mode_deltas[i] = (signed char)delta;
754 eb_vp9_wb_write_literal(wb, abs(delta) & 0x3F, 6);
755 eb_vp9_wb_write_bit(wb, delta < 0);
756 }
757 }
758 }
759 }
760 }
761
write_delta_q(struct vpx_write_bit_buffer * wb,int delta_q)762 static void write_delta_q(struct vpx_write_bit_buffer *wb, int delta_q) {
763 if (delta_q != 0) {
764 eb_vp9_wb_write_bit(wb, 1);
765 eb_vp9_wb_write_literal(wb, abs(delta_q), 4);
766 eb_vp9_wb_write_bit(wb, delta_q < 0);
767 } else {
768 eb_vp9_wb_write_bit(wb, 0);
769 }
770 }
771
encode_quantization(const VP9_COMMON * const cm,struct vpx_write_bit_buffer * wb)772 static void encode_quantization(const VP9_COMMON *const cm,
773 struct vpx_write_bit_buffer *wb) {
774 eb_vp9_wb_write_literal(wb, cm->base_qindex, QINDEX_BITS);
775 write_delta_q(wb, cm->y_dc_delta_q);
776 write_delta_q(wb, cm->uv_dc_delta_q);
777 write_delta_q(wb, cm->uv_ac_delta_q);
778 }
779
encode_segmentation(VP9_COMMON * cm,MACROBLOCKD * xd,struct vpx_write_bit_buffer * wb)780 static void encode_segmentation(VP9_COMMON *cm, MACROBLOCKD *xd,
781 struct vpx_write_bit_buffer *wb) {
782 (void)xd;
783 #if SEG_SUPPORT
784 int i, j;
785 #endif
786 const struct segmentation *seg = &cm->seg;
787
788 eb_vp9_wb_write_bit(wb, seg->enabled);
789 if (!seg->enabled) return;
790 #if SEG_SUPPORT
791 // Segmentation map
792 eb_vp9_wb_write_bit(wb, seg->update_map);
793 if (seg->update_map) {
794
795 for (i = 0; i < SEG_TREE_PROBS; i++) {
796 const int prob = seg->tree_probs[i];
797 const int update = prob != MAX_PROB;
798 eb_vp9_wb_write_bit(wb, update);
799 if (update) eb_vp9_wb_write_literal(wb, prob, 8);
800 }
801
802 // Write out the chosen coding method.
803 eb_vp9_wb_write_bit(wb, seg->temporal_update);
804 if (seg->temporal_update) {
805 for (i = 0; i < PREDICTION_PROBS; i++) {
806 const int prob = seg->pred_probs[i];
807 const int update = prob != MAX_PROB;
808 eb_vp9_wb_write_bit(wb, update);
809 if (update) eb_vp9_wb_write_literal(wb, prob, 8);
810 }
811 }
812 }
813
814 // Segmentation data
815 eb_vp9_wb_write_bit(wb, seg->update_data);
816 if (seg->update_data) {
817 eb_vp9_wb_write_bit(wb, seg->abs_delta);
818
819 for (i = 0; i < MAX_SEGMENTS; i++) {
820 for (j = 0; j < SEG_LVL_MAX; j++) {
821 const int active = segfeature_active(seg, i, (SEG_LVL_FEATURES)j);
822 eb_vp9_wb_write_bit(wb, active);
823 if (active) {
824 const int data = get_segdata(seg, i, (SEG_LVL_FEATURES)j);
825 const int data_max = eb_vp9_seg_feature_data_max((SEG_LVL_FEATURES)j);
826
827 if (eb_vp9_is_segfeature_signed((SEG_LVL_FEATURES)j)) {
828 encode_unsigned_max(wb, abs(data), data_max);
829 eb_vp9_wb_write_bit(wb, data < 0);
830 } else {
831 encode_unsigned_max(wb, data, data_max);
832 }
833 }
834 }
835 }
836 }
837 #endif
838 }
839
encode_txfm_probs(VP9_COMMON * cm,VpxWriter * w,FRAME_COUNTS * counts)840 static void encode_txfm_probs(VP9_COMMON *cm, VpxWriter *w,
841 FRAME_COUNTS *counts) {
842 // Mode
843 vpx_write_literal(w, VPXMIN(cm->tx_mode, ALLOW_32X32), 2);
844 if (cm->tx_mode >= ALLOW_32X32)
845 vpx_write_bit(w, cm->tx_mode == TX_MODE_SELECT);
846
847 // Probabilities
848 if (cm->tx_mode == TX_MODE_SELECT) {
849 int i, j;
850 unsigned int ct_8x8p[TX_SIZES - 3][2];
851 unsigned int ct_16x16p[TX_SIZES - 2][2];
852 unsigned int ct_32x32p[TX_SIZES - 1][2];
853
854 for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
855 eb_vp9_tx_counts_to_branch_counts_8x8(counts->tx.p8x8[i], ct_8x8p);
856 for (j = 0; j < TX_SIZES - 3; j++)
857 eb_vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p8x8[i][j], ct_8x8p[j]);
858 }
859
860 for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
861 eb_vp9_tx_counts_to_branch_counts_16x16(counts->tx.p16x16[i], ct_16x16p);
862 for (j = 0; j < TX_SIZES - 2; j++)
863 eb_vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p16x16[i][j],
864 ct_16x16p[j]);
865 }
866
867 for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
868 eb_vp9_tx_counts_to_branch_counts_32x32(counts->tx.p32x32[i], ct_32x32p);
869 for (j = 0; j < TX_SIZES - 1; j++)
870 eb_vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p32x32[i][j],
871 ct_32x32p[j]);
872 }
873 }
874 }
875
write_interp_filter(INTERP_FILTER filter,struct vpx_write_bit_buffer * wb)876 static void write_interp_filter(INTERP_FILTER filter,
877 struct vpx_write_bit_buffer *wb) {
878 const int filter_to_literal[] = { 1, 0, 2, 3 };
879
880 eb_vp9_wb_write_bit(wb, filter == SWITCHABLE);
881 if (filter != SWITCHABLE)
882 eb_vp9_wb_write_literal(wb, filter_to_literal[filter], 2);
883 }
884 #if 0 // Hsan: switchable interp_filter not supported
885 static void fix_interp_filter(VP9_COMMON *cm, FRAME_COUNTS *counts) {
886 if (cm->interp_filter == SWITCHABLE) {
887 // Check to see if only one of the filters is actually used
888 int count[SWITCHABLE_FILTERS];
889 int i, j, c = 0;
890 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
891 count[i] = 0;
892 for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
893 count[i] += counts->switchable_interp[j][i];
894 c += (count[i] > 0);
895 }
896 if (c == 1) {
897 // Only one filter is used. So set the filter at frame level
898 for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
899 if (count[i]) {
900 cm->interp_filter = i;
901 break;
902 }
903 }
904 }
905 }
906 }
907 #endif
write_tile_info(const VP9_COMMON * const cm,struct vpx_write_bit_buffer * wb)908 static void write_tile_info(const VP9_COMMON *const cm,
909 struct vpx_write_bit_buffer *wb) {
910 int min_log2_tile_cols, max_log2_tile_cols, ones;
911 eb_vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
912
913 // columns
914 ones = cm->log2_tile_cols - min_log2_tile_cols;
915 while (ones--) eb_vp9_wb_write_bit(wb, 1);
916
917 if (cm->log2_tile_cols < max_log2_tile_cols) eb_vp9_wb_write_bit(wb, 0);
918
919 // rows
920 eb_vp9_wb_write_bit(wb, cm->log2_tile_rows != 0);
921 if (cm->log2_tile_rows != 0) eb_vp9_wb_write_bit(wb, cm->log2_tile_rows != 1);
922 }
923 #if 0
924 int vp9_get_refresh_mask(VP9_COMP *cpi) {
925 if (vp9_preserve_existing_gf(cpi)) {
926 // We have decided to preserve the previously existing golden frame as our
927 // new ARF frame. However, in the short term we leave it in the GF slot and,
928 // if we're updating the GF with the current decoded frame, we save it
929 // instead to the ARF slot.
930 // Later, in the function vp9_encoder.c:vp9_update_reference_frames() we
931 // will swap gld_fb_idx and alt_fb_idx to achieve our objective. We do it
932 // there so that it can be done outside of the recode loop.
933 // Note: This is highly specific to the use of ARF as a forward reference,
934 // and this needs to be generalized as other uses are implemented
935 // (like RTC/temporal scalability).
936 return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
937 (cpi->refresh_golden_frame << cpi->alt_fb_idx);
938 } else {
939 int arf_idx = cpi->alt_fb_idx;
940 GF_GROUP *const gf_group = &cpi->twopass.gf_group;
941
942 if (cpi->multi_layer_arf) {
943 for (arf_idx = 0; arf_idx < REF_FRAMES; ++arf_idx) {
944 if (arf_idx != cpi->alt_fb_idx && arf_idx != cpi->lst_fb_idx &&
945 arf_idx != cpi->gld_fb_idx) {
946 int idx;
947 for (idx = 0; idx < gf_group->stack_size; ++idx)
948 if (arf_idx == gf_group->arf_index_stack[idx]) break;
949 if (idx == gf_group->stack_size) break;
950 }
951 }
952 }
953 cpi->twopass.gf_group.top_arf_idx = arf_idx;
954
955 if (cpi->use_svc && cpi->svc.use_set_ref_frame_config &&
956 cpi->svc.temporal_layering_mode == VP9E_TEMPORAL_LAYERING_MODE_BYPASS)
957 return cpi->svc.update_buffer_slot[cpi->svc.spatial_layer_id];
958 return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
959 (cpi->refresh_golden_frame << cpi->gld_fb_idx) |
960 (cpi->refresh_alt_ref_frame << arf_idx);
961 }
962 }
963 #endif
964 #if 0
965 static int encode_tile_worker(void *arg1, void *arg2) {
966 VP9_COMP *cpi = (VP9_COMP *)arg1;
967 VP9BitstreamWorkerData *data = (VP9BitstreamWorkerData *)arg2;
968 MACROBLOCKD *const xd = &data->xd;
969 const int tile_row = 0;
970 eb_vp9_start_encode(&data->bit_writer, data->dest);
971 write_modes(cpi, xd, &cpi->tile_data[data->tile_idx].tile_info,
972 &data->bit_writer, tile_row, data->tile_idx,
973 &data->max_mv_magnitude, data->interp_filter_selected);
974 eb_vp9_stop_encode(&data->bit_writer);
975 return 1;
976 }
977
978 void vp9_bitstream_encode_tiles_buffer_dealloc(VP9_COMP *const cpi) {
979 if (cpi->vp9_bitstream_worker_data) {
980 int i;
981 for (i = 1; i < cpi->num_workers; ++i) {
982 vpx_free(cpi->vp9_bitstream_worker_data[i].dest);
983 }
984 vpx_free(cpi->vp9_bitstream_worker_data);
985 cpi->vp9_bitstream_worker_data = NULL;
986 }
987 }
988
989 static int encode_tiles_buffer_alloc(VP9_COMP *const cpi) {
990 int i;
991 const size_t worker_data_size =
992 cpi->num_workers * sizeof(*cpi->vp9_bitstream_worker_data);
993 cpi->vp9_bitstream_worker_data = vpx_memalign(16, worker_data_size);
994 memset(cpi->vp9_bitstream_worker_data, 0, worker_data_size);
995 if (!cpi->vp9_bitstream_worker_data) return 1;
996 for (i = 1; i < cpi->num_workers; ++i) {
997 cpi->vp9_bitstream_worker_data[i].dest_size =
998 cpi->oxcf.width * cpi->oxcf.height;
999 cpi->vp9_bitstream_worker_data[i].dest =
1000 vpx_malloc(cpi->vp9_bitstream_worker_data[i].dest_size);
1001 if (!cpi->vp9_bitstream_worker_data[i].dest) return 1;
1002 }
1003 return 0;
1004 }
1005
1006 static size_t encode_tiles_mt(VP9_COMP *cpi, uint8_t *data_ptr) {
1007 const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
1008 VP9_COMMON *const cm = &cpi->common;
1009 const int tile_cols = 1 << cm->log2_tile_cols;
1010 const int num_workers = cpi->num_workers;
1011 size_t total_size = 0;
1012 int tile_col = 0;
1013
1014 if (!cpi->vp9_bitstream_worker_data ||
1015 cpi->vp9_bitstream_worker_data[1].dest_size >
1016 (cpi->oxcf.width * cpi->oxcf.height)) {
1017 vp9_bitstream_encode_tiles_buffer_dealloc(cpi);
1018 if (encode_tiles_buffer_alloc(cpi)) return 0;
1019 }
1020
1021 while (tile_col < tile_cols) {
1022 int i, j;
1023 for (i = 0; i < num_workers && tile_col < tile_cols; ++i) {
1024 VPxWorker *const worker = &cpi->workers[i];
1025 VP9BitstreamWorkerData *const data = &cpi->vp9_bitstream_worker_data[i];
1026
1027 // Populate the worker data.
1028 data->xd = cpi->td.mb.e_mbd;
1029 data->tile_idx = tile_col;
1030 data->max_mv_magnitude = cpi->max_mv_magnitude;
1031 memset(data->interp_filter_selected, 0,
1032 sizeof(data->interp_filter_selected[0][0]) * SWITCHABLE);
1033
1034 // First thread can directly write into the output buffer.
1035 if (i == 0) {
1036 // If this worker happens to be for the last tile, then do not offset it
1037 // by 4 for the tile size.
1038 data->dest =
1039 data_ptr + total_size + (tile_col == tile_cols - 1 ? 0 : 4);
1040 }
1041 worker->data1 = cpi;
1042 worker->data2 = data;
1043 worker->hook = encode_tile_worker;
1044 worker->had_error = 0;
1045
1046 if (i < num_workers - 1) {
1047 winterface->launch(worker);
1048 } else {
1049 winterface->execute(worker);
1050 }
1051 ++tile_col;
1052 }
1053 for (j = 0; j < i; ++j) {
1054 VPxWorker *const worker = &cpi->workers[j];
1055 VP9BitstreamWorkerData *const data =
1056 (VP9BitstreamWorkerData *)worker->data2;
1057 uint32_t tile_size;
1058 int k;
1059
1060 if (!winterface->sync(worker)) return 0;
1061 tile_size = data->bit_writer.pos;
1062
1063 // Aggregate per-thread bitstream stats.
1064 cpi->max_mv_magnitude =
1065 VPXMAX(cpi->max_mv_magnitude, data->max_mv_magnitude);
1066 for (k = 0; k < SWITCHABLE; ++k) {
1067 cpi->interp_filter_selected[0][k] += data->interp_filter_selected[0][k];
1068 }
1069
1070 // Prefix the size of the tile on all but the last.
1071 if (tile_col != tile_cols || j < i - 1) {
1072 mem_put_be32(data_ptr + total_size, tile_size);
1073 total_size += 4;
1074 }
1075 if (j > 0) {
1076 memcpy(data_ptr + total_size, data->dest, tile_size);
1077 }
1078 total_size += tile_size;
1079 }
1080 }
1081 return total_size;
1082 }
1083
1084 static size_t encode_tiles(VP9_COMP *cpi, uint8_t *data_ptr) {
1085 VP9_COMMON *const cm = &cpi->common;
1086 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
1087 VpxWriter residual_bc;
1088 int tile_row, tile_col;
1089 size_t total_size = 0;
1090 const int tile_cols = 1 << cm->log2_tile_cols;
1091 const int tile_rows = 1 << cm->log2_tile_rows;
1092
1093 memset(cm->above_seg_context, 0,
1094 sizeof(*cm->above_seg_context) * mi_cols_aligned_to_sb(cm->mi_cols));
1095
1096 // Encoding tiles in parallel is done only for realtime mode now. In other
1097 // modes the speed up is insignificant and requires further testing to ensure
1098 // that it does not make the overall process worse in any case.
1099 if (cpi->oxcf.mode == REALTIME && cpi->num_workers > 1 && tile_rows == 1 &&
1100 tile_cols > 1) {
1101 return encode_tiles_mt(cpi, data_ptr);
1102 }
1103
1104 for (tile_row = 0; tile_row < tile_rows; tile_row++) {
1105 for (tile_col = 0; tile_col < tile_cols; tile_col++) {
1106 int tile_idx = tile_row * tile_cols + tile_col;
1107
1108 if (tile_col < tile_cols - 1 || tile_row < tile_rows - 1)
1109 eb_vp9_start_encode(&residual_bc, data_ptr + total_size + 4);
1110 else
1111 eb_vp9_start_encode(&residual_bc, data_ptr + total_size);
1112
1113 write_modes(cpi, xd, &cpi->tile_data[tile_idx].tile_info, &residual_bc,
1114 tile_row, tile_col, &cpi->max_mv_magnitude,
1115 cpi->interp_filter_selected);
1116
1117 eb_vp9_stop_encode(&residual_bc);
1118 if (tile_col < tile_cols - 1 || tile_row < tile_rows - 1) {
1119 // size of this tile
1120 mem_put_be32(data_ptr + total_size, residual_bc.pos);
1121 total_size += 4;
1122 }
1123
1124 total_size += residual_bc.pos;
1125 }
1126 }
1127 return total_size;
1128 }
1129 #endif
1130
write_render_size(const VP9_COMMON * cm,struct vpx_write_bit_buffer * wb)1131 static void write_render_size(const VP9_COMMON *cm,
1132 struct vpx_write_bit_buffer *wb) {
1133 const int scaling_active =
1134 cm->width != cm->render_width || cm->height != cm->render_height;
1135 eb_vp9_wb_write_bit(wb, scaling_active);
1136 if (scaling_active) {
1137 eb_vp9_wb_write_literal(wb, cm->render_width - 1, 16);
1138 eb_vp9_wb_write_literal(wb, cm->render_height - 1, 16);
1139 }
1140 }
1141
write_frame_size(const VP9_COMMON * cm,struct vpx_write_bit_buffer * wb)1142 static void write_frame_size(const VP9_COMMON *cm,
1143 struct vpx_write_bit_buffer *wb) {
1144 eb_vp9_wb_write_literal(wb, cm->width - 1, 16);
1145 eb_vp9_wb_write_literal(wb, cm->height - 1, 16);
1146
1147 write_render_size(cm, wb);
1148 }
1149
write_frame_size_with_refs(VP9_COMP * cpi,struct vpx_write_bit_buffer * wb)1150 static void write_frame_size_with_refs(VP9_COMP *cpi,
1151 struct vpx_write_bit_buffer *wb) {
1152 VP9_COMMON *const cm = &cpi->common;
1153 int found = 0;
1154
1155 MV_REFERENCE_FRAME ref_frame;
1156 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
1157 #if 0
1158 YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi, ref_frame);
1159
1160 // Set "found" to 0 for temporal svc and for spatial svc key frame
1161 if (cpi->use_svc &&
1162 ((cpi->svc.number_temporal_layers > 1 &&
1163 cpi->oxcf.rc_mode == VPX_CBR) ||
1164 (cpi->svc.number_spatial_layers > 1 &&
1165 cpi->svc.layer_context[cpi->svc.spatial_layer_id].is_key_frame))) {
1166 found = 0;
1167 } else if (cfg != NULL) {
1168 found =
1169 cm->width == cfg->y_crop_width && cm->height == cfg->y_crop_height;
1170 }
1171 #endif
1172 eb_vp9_wb_write_bit(wb, found);
1173 if (found) {
1174 break;
1175 }
1176 }
1177
1178 if (!found) {
1179 eb_vp9_wb_write_literal(wb, cm->width - 1, 16);
1180 eb_vp9_wb_write_literal(wb, cm->height - 1, 16);
1181 }
1182
1183 write_render_size(cm, wb);
1184 }
1185
write_sync_code(struct vpx_write_bit_buffer * wb)1186 static void write_sync_code(struct vpx_write_bit_buffer *wb) {
1187 eb_vp9_wb_write_literal(wb, VP9_SYNC_CODE_0, 8);
1188 eb_vp9_wb_write_literal(wb, VP9_SYNC_CODE_1, 8);
1189 eb_vp9_wb_write_literal(wb, VP9_SYNC_CODE_2, 8);
1190 }
1191
write_profile(BITSTREAM_PROFILE profile,struct vpx_write_bit_buffer * wb)1192 static void write_profile(BITSTREAM_PROFILE profile,
1193 struct vpx_write_bit_buffer *wb) {
1194 switch (profile) {
1195 case PROFILE_0: eb_vp9_wb_write_literal(wb, 0, 2); break;
1196 case PROFILE_1: eb_vp9_wb_write_literal(wb, 2, 2); break;
1197 case PROFILE_2: eb_vp9_wb_write_literal(wb, 1, 2); break;
1198 default:
1199 assert(profile == PROFILE_3);
1200 eb_vp9_wb_write_literal(wb, 6, 3);
1201 break;
1202 }
1203 }
1204
write_bitdepth_colorspace_sampling(VP9_COMMON * const cm,struct vpx_write_bit_buffer * wb)1205 static void write_bitdepth_colorspace_sampling(
1206 VP9_COMMON *const cm, struct vpx_write_bit_buffer *wb) {
1207 if (cm->profile >= PROFILE_2) {
1208 assert(cm->bit_depth > VPX_BITS_8);
1209 eb_vp9_wb_write_bit(wb, cm->bit_depth == VPX_BITS_10 ? 0 : 1);
1210 }
1211 eb_vp9_wb_write_literal(wb, cm->color_space, 3);
1212 if (cm->color_space != VPX_CS_SRGB) {
1213 // 0: [16, 235] (i.e. xvYCC), 1: [0, 255]
1214 eb_vp9_wb_write_bit(wb, cm->color_range);
1215 if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
1216 assert(cm->subsampling_x != 1 || cm->subsampling_y != 1);
1217 eb_vp9_wb_write_bit(wb, cm->subsampling_x);
1218 eb_vp9_wb_write_bit(wb, cm->subsampling_y);
1219 eb_vp9_wb_write_bit(wb, 0); // unused
1220 } else {
1221 assert(cm->subsampling_x == 1 && cm->subsampling_y == 1);
1222 }
1223 } else {
1224 assert(cm->profile == PROFILE_1 || cm->profile == PROFILE_3);
1225 eb_vp9_wb_write_bit(wb, 0); // unused
1226 }
1227 }
1228
write_uncompressed_header(PictureControlSet * picture_control_set_ptr,VP9_COMP * cpi,struct vpx_write_bit_buffer * wb,int show_existing_frame,int show_existing_frame_index)1229 static void write_uncompressed_header(
1230 PictureControlSet *picture_control_set_ptr,
1231 VP9_COMP *cpi,
1232 struct vpx_write_bit_buffer *wb,
1233 int show_existing_frame,
1234 int show_existing_frame_index) {
1235
1236 VP9_COMMON *const cm = &cpi->common;
1237 #if 0
1238 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
1239 #endif
1240
1241 eb_vp9_wb_write_literal(wb, VP9_FRAME_MARKER, 2);
1242
1243 write_profile(cm->profile, wb);
1244
1245 // If to use show existing frame.
1246 eb_vp9_wb_write_bit(wb, show_existing_frame);
1247 if (show_existing_frame) {
1248 #if 1 // Hsan: from SVT
1249 eb_vp9_wb_write_literal(wb, picture_control_set_ptr->parent_pcs_ptr->show_existing_frame_index_array[show_existing_frame_index], 3);
1250 #else
1251 eb_vp9_wb_write_literal(wb, cpi->alt_fb_idx, 3);
1252 #endif
1253
1254 return;
1255 }
1256
1257 eb_vp9_wb_write_bit(wb, cm->frame_type);
1258 eb_vp9_wb_write_bit(wb, cm->show_frame);
1259 eb_vp9_wb_write_bit(wb, cm->error_resilient_mode);
1260
1261 if (cm->frame_type == KEY_FRAME) {
1262 write_sync_code(wb);
1263 write_bitdepth_colorspace_sampling(cm, wb);
1264 write_frame_size(cm, wb);
1265 } else {
1266 if (!cm->show_frame) eb_vp9_wb_write_bit(wb, cm->intra_only);
1267
1268 if (!cm->error_resilient_mode)
1269 eb_vp9_wb_write_literal(wb, cm->reset_frame_context, 2);
1270
1271 if (cm->intra_only) {
1272 write_sync_code(wb);
1273
1274 // Note for profile 0, 420 8bpp is assumed.
1275 if (cm->profile > PROFILE_0) {
1276 write_bitdepth_colorspace_sampling(cm, wb);
1277 }
1278 #if 1 // Hsan: from SVT
1279 eb_vp9_wb_write_literal(wb, picture_control_set_ptr->parent_pcs_ptr->ref_signal.refresh_frame_mask, REF_FRAMES);
1280 #else
1281 eb_vp9_wb_write_literal(wb, vp9_get_refresh_mask(cpi), REF_FRAMES);
1282 #endif
1283
1284 write_frame_size(cm, wb);
1285 } else {
1286 MV_REFERENCE_FRAME ref_frame;
1287 #if 1 // Hsan: from SVT
1288 eb_vp9_wb_write_literal(wb, picture_control_set_ptr->parent_pcs_ptr->ref_signal.refresh_frame_mask, REF_FRAMES);
1289 #else
1290 eb_vp9_wb_write_literal(wb, vp9_get_refresh_mask(cpi), REF_FRAMES);
1291 #endif
1292 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
1293 #if 1 // Hsan: from SVT
1294 eb_vp9_wb_write_literal(wb, picture_control_set_ptr->parent_pcs_ptr->ref_signal.ref_dpb_index[ref_frame - LAST_FRAME],
1295 REF_FRAMES_LOG2);
1296 #else
1297 assert(get_ref_frame_map_idx(cpi, ref_frame) != INVALID_IDX);
1298 eb_vp9_wb_write_literal(wb, get_ref_frame_map_idx(cpi, ref_frame),
1299 REF_FRAMES_LOG2);
1300 #endif
1301 eb_vp9_wb_write_bit(wb, cm->ref_frame_sign_bias[ref_frame]);
1302 }
1303
1304 write_frame_size_with_refs(cpi, wb);
1305
1306 eb_vp9_wb_write_bit(wb, cm->allow_high_precision_mv);
1307 #if 1 // Hsan: switchable interp_filter not supported
1308 write_interp_filter(0, wb);
1309 #else
1310 fix_interp_filter(cm, cpi->td.counts);
1311 write_interp_filter(cm->interp_filter, wb);
1312 #endif
1313 }
1314 }
1315
1316 if (!cm->error_resilient_mode) {
1317 eb_vp9_wb_write_bit(wb, cm->refresh_frame_context);
1318 eb_vp9_wb_write_bit(wb, cm->frame_parallel_decoding_mode);
1319 }
1320
1321 eb_vp9_wb_write_literal(wb, cm->frame_context_idx, FRAME_CONTEXTS_LOG2);
1322
1323 encode_loopfilter(&cm->lf, wb);
1324 encode_quantization(cm, wb);
1325 #if 1
1326 encode_segmentation(cm, (MACROBLOCKD *) EB_NULL, wb);
1327 #else
1328 encode_segmentation(cm, xd, wb);
1329 #endif
1330
1331 write_tile_info(cm, wb);
1332 }
1333
write_compressed_header(VP9_COMP * cpi,uint8_t * data)1334 size_t write_compressed_header(VP9_COMP *cpi, uint8_t *data) {
1335 VP9_COMMON *const cm = &cpi->common;
1336 #if 0 // xd
1337 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
1338 #endif
1339 FRAME_CONTEXT *const fc = cm->fc;
1340
1341 FRAME_COUNTS *counts = cpi->td.counts;
1342
1343 VpxWriter header_bc;
1344
1345 eb_vp9_start_encode(&header_bc, data);
1346 #if 1 // xd
1347 encode_txfm_probs(cm, &header_bc, counts);
1348 #else
1349 if (xd->lossless)
1350 cm->tx_mode = ONLY_4X4;
1351 else
1352 encode_txfm_probs(cm, &header_bc, counts);
1353 #endif
1354
1355 update_coef_probs(cpi, &header_bc);
1356 update_skip_probs(cm, &header_bc, counts);
1357
1358 if (!frame_is_intra_only(cm)) {
1359 int i;
1360
1361 for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
1362 prob_diff_update(eb_vp9_inter_mode_tree, cm->fc->inter_mode_probs[i],
1363 counts->inter_mode[i], INTER_MODES, &header_bc);
1364 #if 0 // Hsan: switchable interp_filter not supported
1365 if (cm->interp_filter == SWITCHABLE)
1366 update_switchable_interp_probs(cm, &header_bc, counts);
1367 #endif
1368 for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
1369 eb_vp9_cond_prob_diff_update(&header_bc, &fc->intra_inter_prob[i],
1370 counts->intra_inter[i]);
1371
1372 if (cpi->allow_comp_inter_inter) {
1373 const int use_compound_pred = cm->reference_mode != SINGLE_REFERENCE;
1374 const int use_hybrid_pred = cm->reference_mode == REFERENCE_MODE_SELECT;
1375
1376 vpx_write_bit(&header_bc, use_compound_pred);
1377 if (use_compound_pred) {
1378 vpx_write_bit(&header_bc, use_hybrid_pred);
1379 if (use_hybrid_pred)
1380 for (i = 0; i < COMP_INTER_CONTEXTS; i++)
1381 eb_vp9_cond_prob_diff_update(&header_bc, &fc->comp_inter_prob[i],
1382 counts->comp_inter[i]);
1383 }
1384 }
1385 if (cm->reference_mode != COMPOUND_REFERENCE) {
1386 for (i = 0; i < REF_CONTEXTS; i++) {
1387 eb_vp9_cond_prob_diff_update(&header_bc, &fc->single_ref_prob[i][0],
1388 counts->single_ref[i][0]);
1389 eb_vp9_cond_prob_diff_update(&header_bc, &fc->single_ref_prob[i][1],
1390 counts->single_ref[i][1]);
1391 }
1392 }
1393
1394 if (cm->reference_mode != SINGLE_REFERENCE)
1395 for (i = 0; i < REF_CONTEXTS; i++)
1396 eb_vp9_cond_prob_diff_update(&header_bc, &fc->comp_ref_prob[i],
1397 counts->comp_ref[i]);
1398
1399 for (i = 0; i < BLOCK_SIZE_GROUPS; ++i)
1400 prob_diff_update(eb_vp9_intra_mode_tree, cm->fc->y_mode_prob[i],
1401 counts->y_mode[i], INTRA_MODES, &header_bc);
1402
1403 for (i = 0; i < PARTITION_CONTEXTS; ++i)
1404 prob_diff_update(eb_vp9_partition_tree, fc->partition_prob[i],
1405 counts->partition[i], PARTITION_TYPES, &header_bc);
1406
1407 eb_vp9_write_nmv_probs(cm, cm->allow_high_precision_mv, &header_bc,
1408 &counts->mv);
1409 }
1410
1411 eb_vp9_stop_encode(&header_bc);
1412 assert(header_bc.pos <= 0xffff);
1413
1414 return header_bc.pos;
1415 }
1416
eb_vp9_pack_bitstream(PictureControlSet * picture_control_set_ptr,VP9_COMP * cpi,uint8_t * dest,size_t * size,int show_existing_frame,int show_existing_frame_index)1417 void eb_vp9_pack_bitstream(
1418 PictureControlSet *picture_control_set_ptr,
1419 VP9_COMP *cpi,
1420 uint8_t *dest,
1421 size_t *size,
1422 int show_existing_frame,
1423 int show_existing_frame_index) {
1424
1425 uint8_t *data = dest;
1426 size_t first_part_size, uncompressed_hdr_size;
1427 struct vpx_write_bit_buffer wb = { data, 0 };
1428 struct vpx_write_bit_buffer saved_wb;
1429
1430 write_uncompressed_header(
1431 picture_control_set_ptr,
1432 cpi,
1433 &wb,
1434 show_existing_frame,
1435 show_existing_frame_index);
1436
1437 // Skip the rest coding process if use show existing frame.
1438 if (show_existing_frame) {
1439 *size = eb_vp9_wb_bytes_written(&wb);
1440 return;
1441 }
1442
1443 saved_wb = wb;
1444 eb_vp9_wb_write_literal(&wb, 0, 16); // don't know in advance first part. size
1445
1446 uncompressed_hdr_size = eb_vp9_wb_bytes_written(&wb);
1447 data += uncompressed_hdr_size;
1448 #if 0 // Hsan
1449 vpx_clear_system_state();
1450 #endif
1451 first_part_size = write_compressed_header(cpi, data);
1452 data += first_part_size;
1453 // TODO(jbb): Figure out what to do if first_part_size > 16 bits.
1454 eb_vp9_wb_write_literal(&saved_wb, (int)first_part_size, 16);
1455
1456 #if 1 // Hsan ------------------------------------
1457 // Link data from EC stream to final stream.
1458 unsigned int ecOutputBitstreamSize = picture_control_set_ptr->entropy_coder_ptr->residual_bc.pos;
1459
1460 OutputBitstreamUnit *ec_output_bitstream_ptr = (OutputBitstreamUnit *)picture_control_set_ptr->entropy_coder_ptr->ec_output_bitstream_ptr;
1461
1462 // Copy from EC stream to frame stream
1463 EB_MEMCPY(data, ec_output_bitstream_ptr->buffer_begin, ecOutputBitstreamSize);
1464 data += ecOutputBitstreamSize;
1465 #else
1466 data += encode_tiles(cpi, data);
1467 #endif
1468
1469 *size = data - dest;
1470 }
1471