xref: /aosp_15_r20/external/libvpx/vp9/encoder/vp9_bitstream.c (revision fb1b10ab9aebc7c7068eedab379b749d7e3900be)
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 #include <stdint.h>
13 #include <stdio.h>
14 #include <limits.h>
15 
16 #include "vpx/vpx_encoder.h"
17 #include "vpx_dsp/bitwriter_buffer.h"
18 #include "vpx_dsp/vpx_dsp_common.h"
19 #include "vpx_mem/vpx_mem.h"
20 #include "vpx_ports/mem_ops.h"
21 #include "vpx_ports/system_state.h"
22 #if CONFIG_BITSTREAM_DEBUG
23 #include "vpx_util/vpx_debug_util.h"
24 #endif  // CONFIG_BITSTREAM_DEBUG
25 
26 #include "vp9/common/vp9_entropy.h"
27 #include "vp9/common/vp9_entropymode.h"
28 #include "vp9/common/vp9_entropymv.h"
29 #include "vp9/common/vp9_mvref_common.h"
30 #include "vp9/common/vp9_pred_common.h"
31 #include "vp9/common/vp9_seg_common.h"
32 #include "vp9/common/vp9_tile_common.h"
33 
34 #include "vp9/encoder/vp9_cost.h"
35 #include "vp9/encoder/vp9_bitstream.h"
36 #include "vp9/encoder/vp9_encodemv.h"
37 #include "vp9/encoder/vp9_mcomp.h"
38 #include "vp9/encoder/vp9_segmentation.h"
39 #include "vp9/encoder/vp9_subexp.h"
40 #include "vp9/encoder/vp9_tokenize.h"
41 
42 static const struct vp9_token intra_mode_encodings[INTRA_MODES] = {
43   { 0, 1 },  { 6, 3 },   { 28, 5 },  { 30, 5 }, { 58, 6 },
44   { 59, 6 }, { 126, 7 }, { 127, 7 }, { 62, 6 }, { 2, 2 }
45 };
46 static const struct vp9_token
47     switchable_interp_encodings[SWITCHABLE_FILTERS] = { { 0, 1 },
48                                                         { 2, 2 },
49                                                         { 3, 2 } };
50 static const struct vp9_token partition_encodings[PARTITION_TYPES] = {
51   { 0, 1 }, { 2, 2 }, { 6, 3 }, { 7, 3 }
52 };
53 static const struct vp9_token inter_mode_encodings[INTER_MODES] = {
54   { 2, 2 }, { 6, 3 }, { 0, 1 }, { 7, 3 }
55 };
56 
write_intra_mode(vpx_writer * w,PREDICTION_MODE mode,const vpx_prob * probs)57 static void write_intra_mode(vpx_writer *w, PREDICTION_MODE mode,
58                              const vpx_prob *probs) {
59   vp9_write_token(w, vp9_intra_mode_tree, probs, &intra_mode_encodings[mode]);
60 }
61 
write_inter_mode(vpx_writer * w,PREDICTION_MODE mode,const vpx_prob * probs)62 static void write_inter_mode(vpx_writer *w, PREDICTION_MODE mode,
63                              const vpx_prob *probs) {
64   assert(is_inter_mode(mode));
65   vp9_write_token(w, vp9_inter_mode_tree, probs,
66                   &inter_mode_encodings[INTER_OFFSET(mode)]);
67 }
68 
encode_unsigned_max(struct vpx_write_bit_buffer * wb,int data,int max)69 static void encode_unsigned_max(struct vpx_write_bit_buffer *wb, int data,
70                                 int max) {
71   vpx_wb_write_literal(wb, data, get_unsigned_bits(max));
72 }
73 
prob_diff_update(const vpx_tree_index * tree,vpx_prob probs[],const unsigned int counts[],int n,vpx_writer * w)74 static void prob_diff_update(const vpx_tree_index *tree,
75                              vpx_prob probs[/*n - 1*/],
76                              const unsigned int counts[/*n - 1*/], int n,
77                              vpx_writer *w) {
78   int i;
79   unsigned int branch_ct[32][2];
80 
81   // Assuming max number of probabilities <= 32
82   assert(n <= 32);
83 
84   vp9_tree_probs_from_distribution(tree, branch_ct, counts);
85   for (i = 0; i < n - 1; ++i)
86     vp9_cond_prob_diff_update(w, &probs[i], branch_ct[i]);
87 }
88 
write_selected_tx_size(const VP9_COMMON * cm,const MACROBLOCKD * const xd,vpx_writer * w)89 static void write_selected_tx_size(const VP9_COMMON *cm,
90                                    const MACROBLOCKD *const xd, vpx_writer *w) {
91   TX_SIZE tx_size = xd->mi[0]->tx_size;
92   BLOCK_SIZE bsize = xd->mi[0]->sb_type;
93   const TX_SIZE max_tx_size = max_txsize_lookup[bsize];
94   const vpx_prob *const tx_probs =
95       get_tx_probs(max_tx_size, get_tx_size_context(xd), &cm->fc->tx_probs);
96   vpx_write(w, tx_size != TX_4X4, tx_probs[0]);
97   if (tx_size != TX_4X4 && max_tx_size >= TX_16X16) {
98     vpx_write(w, tx_size != TX_8X8, tx_probs[1]);
99     if (tx_size != TX_8X8 && max_tx_size >= TX_32X32)
100       vpx_write(w, tx_size != TX_16X16, tx_probs[2]);
101   }
102 }
103 
write_skip(const VP9_COMMON * cm,const MACROBLOCKD * const xd,int segment_id,const MODE_INFO * mi,vpx_writer * w)104 static int write_skip(const VP9_COMMON *cm, const MACROBLOCKD *const xd,
105                       int segment_id, const MODE_INFO *mi, vpx_writer *w) {
106   if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
107     return 1;
108   } else {
109     const int skip = mi->skip;
110     vpx_write(w, skip, vp9_get_skip_prob(cm, xd));
111     return skip;
112   }
113 }
114 
update_skip_probs(VP9_COMMON * cm,vpx_writer * w,FRAME_COUNTS * counts)115 static void update_skip_probs(VP9_COMMON *cm, vpx_writer *w,
116                               FRAME_COUNTS *counts) {
117   int k;
118 
119   for (k = 0; k < SKIP_CONTEXTS; ++k)
120     vp9_cond_prob_diff_update(w, &cm->fc->skip_probs[k], counts->skip[k]);
121 }
122 
update_switchable_interp_probs(VP9_COMMON * cm,vpx_writer * w,FRAME_COUNTS * counts)123 static void update_switchable_interp_probs(VP9_COMMON *cm, vpx_writer *w,
124                                            FRAME_COUNTS *counts) {
125   int j;
126   for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
127     prob_diff_update(vp9_switchable_interp_tree,
128                      cm->fc->switchable_interp_prob[j],
129                      counts->switchable_interp[j], SWITCHABLE_FILTERS, w);
130 }
131 
pack_mb_tokens(vpx_writer * w,TOKENEXTRA ** tp,const TOKENEXTRA * const stop,vpx_bit_depth_t bit_depth)132 static void pack_mb_tokens(vpx_writer *w, TOKENEXTRA **tp,
133                            const TOKENEXTRA *const stop,
134                            vpx_bit_depth_t bit_depth) {
135   const TOKENEXTRA *p;
136   const vp9_extra_bit *const extra_bits =
137 #if CONFIG_VP9_HIGHBITDEPTH
138       (bit_depth == VPX_BITS_12)   ? vp9_extra_bits_high12
139       : (bit_depth == VPX_BITS_10) ? vp9_extra_bits_high10
140                                    : vp9_extra_bits;
141 #else
142       vp9_extra_bits;
143   (void)bit_depth;
144 #endif  // CONFIG_VP9_HIGHBITDEPTH
145 
146   for (p = *tp; p < stop && p->token != EOSB_TOKEN; ++p) {
147     if (p->token == EOB_TOKEN) {
148       vpx_write(w, 0, p->context_tree[0]);
149       continue;
150     }
151     vpx_write(w, 1, p->context_tree[0]);
152     while (p->token == ZERO_TOKEN) {
153       vpx_write(w, 0, p->context_tree[1]);
154       ++p;
155       if (p == stop || p->token == EOSB_TOKEN) {
156         *tp = (TOKENEXTRA *)(uintptr_t)p + (p->token == EOSB_TOKEN);
157         return;
158       }
159     }
160 
161     {
162       const int t = p->token;
163       const vpx_prob *const context_tree = p->context_tree;
164       assert(t != ZERO_TOKEN);
165       assert(t != EOB_TOKEN);
166       assert(t != EOSB_TOKEN);
167       vpx_write(w, 1, context_tree[1]);
168       if (t == ONE_TOKEN) {
169         vpx_write(w, 0, context_tree[2]);
170         vpx_write_bit(w, p->extra & 1);
171       } else {  // t >= TWO_TOKEN && t < EOB_TOKEN
172         const struct vp9_token *const a = &vp9_coef_encodings[t];
173         int v = a->value;
174         int n = a->len;
175         const int e = p->extra;
176         vpx_write(w, 1, context_tree[2]);
177         vp9_write_tree(w, vp9_coef_con_tree,
178                        vp9_pareto8_full[context_tree[PIVOT_NODE] - 1], v,
179                        n - UNCONSTRAINED_NODES, 0);
180         if (t >= CATEGORY1_TOKEN) {
181           const vp9_extra_bit *const b = &extra_bits[t];
182           const unsigned char *pb = b->prob;
183           v = e >> 1;
184           n = b->len;  // number of bits in v, assumed nonzero
185           do {
186             const int bb = (v >> --n) & 1;
187             vpx_write(w, bb, *pb++);
188           } while (n);
189         }
190         vpx_write_bit(w, e & 1);
191       }
192     }
193   }
194   *tp = (TOKENEXTRA *)(uintptr_t)p + (p->token == EOSB_TOKEN);
195 }
196 
write_segment_id(vpx_writer * w,const struct segmentation * seg,int segment_id)197 static void write_segment_id(vpx_writer *w, const struct segmentation *seg,
198                              int segment_id) {
199   if (seg->enabled && seg->update_map)
200     vp9_write_tree(w, vp9_segment_tree, seg->tree_probs, segment_id, 3, 0);
201 }
202 
203 // This function encodes the reference frame
write_ref_frames(const VP9_COMMON * cm,const MACROBLOCKD * const xd,vpx_writer * w)204 static void write_ref_frames(const VP9_COMMON *cm, const MACROBLOCKD *const xd,
205                              vpx_writer *w) {
206   const MODE_INFO *const mi = xd->mi[0];
207   const int is_compound = has_second_ref(mi);
208   const int segment_id = mi->segment_id;
209 
210   // If segment level coding of this signal is disabled...
211   // or the segment allows multiple reference frame options
212   if (segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
213     assert(!is_compound);
214     assert(mi->ref_frame[0] ==
215            get_segdata(&cm->seg, segment_id, SEG_LVL_REF_FRAME));
216   } else {
217     // does the feature use compound prediction or not
218     // (if not specified at the frame/segment level)
219     if (cm->reference_mode == REFERENCE_MODE_SELECT) {
220       vpx_write(w, is_compound, vp9_get_reference_mode_prob(cm, xd));
221     } else {
222       assert((!is_compound) == (cm->reference_mode == SINGLE_REFERENCE));
223     }
224 
225     if (is_compound) {
226       const int idx = cm->ref_frame_sign_bias[cm->comp_fixed_ref];
227       vpx_write(w, mi->ref_frame[!idx] == cm->comp_var_ref[1],
228                 vp9_get_pred_prob_comp_ref_p(cm, xd));
229     } else {
230       const int bit0 = mi->ref_frame[0] != LAST_FRAME;
231       vpx_write(w, bit0, vp9_get_pred_prob_single_ref_p1(cm, xd));
232       if (bit0) {
233         const int bit1 = mi->ref_frame[0] != GOLDEN_FRAME;
234         vpx_write(w, bit1, vp9_get_pred_prob_single_ref_p2(cm, xd));
235       }
236     }
237   }
238 }
239 
pack_inter_mode_mvs(VP9_COMP * cpi,const MACROBLOCKD * const xd,const MB_MODE_INFO_EXT * const mbmi_ext,vpx_writer * w,unsigned int * const max_mv_magnitude,int interp_filter_selected[][SWITCHABLE])240 static void pack_inter_mode_mvs(VP9_COMP *cpi, const MACROBLOCKD *const xd,
241                                 const MB_MODE_INFO_EXT *const mbmi_ext,
242                                 vpx_writer *w,
243                                 unsigned int *const max_mv_magnitude,
244                                 int interp_filter_selected[][SWITCHABLE]) {
245   VP9_COMMON *const cm = &cpi->common;
246   const nmv_context *nmvc = &cm->fc->nmvc;
247   const struct segmentation *const seg = &cm->seg;
248   const MODE_INFO *const mi = xd->mi[0];
249   const PREDICTION_MODE mode = mi->mode;
250   const int segment_id = mi->segment_id;
251   const BLOCK_SIZE bsize = mi->sb_type;
252   const int allow_hp = cm->allow_high_precision_mv;
253   const int is_inter = is_inter_block(mi);
254   const int is_compound = has_second_ref(mi);
255   int skip, ref;
256 
257   if (seg->update_map) {
258     if (seg->temporal_update) {
259       const int pred_flag = mi->seg_id_predicted;
260       vpx_prob pred_prob = vp9_get_pred_prob_seg_id(seg, xd);
261       vpx_write(w, pred_flag, pred_prob);
262       if (!pred_flag) write_segment_id(w, seg, segment_id);
263     } else {
264       write_segment_id(w, seg, segment_id);
265     }
266   }
267 
268   skip = write_skip(cm, xd, segment_id, mi, w);
269 
270   if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
271     vpx_write(w, is_inter, vp9_get_intra_inter_prob(cm, xd));
272 
273   if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT &&
274       !(is_inter && skip)) {
275     write_selected_tx_size(cm, xd, w);
276   }
277 
278   if (!is_inter) {
279     if (bsize >= BLOCK_8X8) {
280       write_intra_mode(w, mode, cm->fc->y_mode_prob[size_group_lookup[bsize]]);
281     } else {
282       int idx, idy;
283       const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
284       const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
285       for (idy = 0; idy < 2; idy += num_4x4_h) {
286         for (idx = 0; idx < 2; idx += num_4x4_w) {
287           const PREDICTION_MODE b_mode = mi->bmi[idy * 2 + idx].as_mode;
288           write_intra_mode(w, b_mode, cm->fc->y_mode_prob[0]);
289         }
290       }
291     }
292     write_intra_mode(w, mi->uv_mode, cm->fc->uv_mode_prob[mode]);
293   } else {
294     const int mode_ctx = mbmi_ext->mode_context[mi->ref_frame[0]];
295     const vpx_prob *const inter_probs = cm->fc->inter_mode_probs[mode_ctx];
296     write_ref_frames(cm, xd, w);
297 
298     // If segment skip is not enabled code the mode.
299     if (!segfeature_active(seg, segment_id, SEG_LVL_SKIP)) {
300       if (bsize >= BLOCK_8X8) {
301         write_inter_mode(w, mode, inter_probs);
302       }
303     }
304 
305     if (cm->interp_filter == SWITCHABLE) {
306       const int ctx = get_pred_context_switchable_interp(xd);
307       vp9_write_token(w, vp9_switchable_interp_tree,
308                       cm->fc->switchable_interp_prob[ctx],
309                       &switchable_interp_encodings[mi->interp_filter]);
310       ++interp_filter_selected[0][mi->interp_filter];
311     } else {
312       assert(mi->interp_filter == cm->interp_filter);
313     }
314 
315     if (bsize < BLOCK_8X8) {
316       const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
317       const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
318       int idx, idy;
319       for (idy = 0; idy < 2; idy += num_4x4_h) {
320         for (idx = 0; idx < 2; idx += num_4x4_w) {
321           const int j = idy * 2 + idx;
322           const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
323           write_inter_mode(w, b_mode, inter_probs);
324           if (b_mode == NEWMV) {
325             for (ref = 0; ref < 1 + is_compound; ++ref)
326               vp9_encode_mv(cpi, w, &mi->bmi[j].as_mv[ref].as_mv,
327                             &mbmi_ext->ref_mvs[mi->ref_frame[ref]][0].as_mv,
328                             nmvc, allow_hp, max_mv_magnitude);
329           }
330         }
331       }
332     } else {
333       if (mode == NEWMV) {
334         for (ref = 0; ref < 1 + is_compound; ++ref)
335           vp9_encode_mv(cpi, w, &mi->mv[ref].as_mv,
336                         &mbmi_ext->ref_mvs[mi->ref_frame[ref]][0].as_mv, nmvc,
337                         allow_hp, max_mv_magnitude);
338       }
339     }
340   }
341 }
342 
write_mb_modes_kf(const VP9_COMMON * cm,const MACROBLOCKD * xd,vpx_writer * w)343 static void write_mb_modes_kf(const VP9_COMMON *cm, const MACROBLOCKD *xd,
344                               vpx_writer *w) {
345   const struct segmentation *const seg = &cm->seg;
346   const MODE_INFO *const mi = xd->mi[0];
347   const MODE_INFO *const above_mi = xd->above_mi;
348   const MODE_INFO *const left_mi = xd->left_mi;
349   const BLOCK_SIZE bsize = mi->sb_type;
350 
351   if (seg->update_map) write_segment_id(w, seg, mi->segment_id);
352 
353   write_skip(cm, xd, mi->segment_id, mi, w);
354 
355   if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT)
356     write_selected_tx_size(cm, xd, w);
357 
358   if (bsize >= BLOCK_8X8) {
359     write_intra_mode(w, mi->mode, get_y_mode_probs(mi, above_mi, left_mi, 0));
360   } else {
361     const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
362     const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
363     int idx, idy;
364 
365     for (idy = 0; idy < 2; idy += num_4x4_h) {
366       for (idx = 0; idx < 2; idx += num_4x4_w) {
367         const int block = idy * 2 + idx;
368         write_intra_mode(w, mi->bmi[block].as_mode,
369                          get_y_mode_probs(mi, above_mi, left_mi, block));
370       }
371     }
372   }
373 
374   write_intra_mode(w, mi->uv_mode, vp9_kf_uv_mode_prob[mi->mode]);
375 }
376 
write_modes_b(VP9_COMP * cpi,MACROBLOCKD * const xd,const TileInfo * const tile,vpx_writer * w,TOKENEXTRA ** tok,const TOKENEXTRA * const tok_end,int mi_row,int mi_col,unsigned int * const max_mv_magnitude,int interp_filter_selected[][SWITCHABLE])377 static void write_modes_b(VP9_COMP *cpi, MACROBLOCKD *const xd,
378                           const TileInfo *const tile, vpx_writer *w,
379                           TOKENEXTRA **tok, const TOKENEXTRA *const tok_end,
380                           int mi_row, int mi_col,
381                           unsigned int *const max_mv_magnitude,
382                           int interp_filter_selected[][SWITCHABLE]) {
383   const VP9_COMMON *const cm = &cpi->common;
384   const MB_MODE_INFO_EXT *const mbmi_ext =
385       cpi->td.mb.mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
386   MODE_INFO *m;
387 
388   xd->mi = cm->mi_grid_visible + (mi_row * cm->mi_stride + mi_col);
389   m = xd->mi[0];
390 
391   set_mi_row_col(xd, tile, mi_row, num_8x8_blocks_high_lookup[m->sb_type],
392                  mi_col, num_8x8_blocks_wide_lookup[m->sb_type], cm->mi_rows,
393                  cm->mi_cols);
394   if (frame_is_intra_only(cm)) {
395     write_mb_modes_kf(cm, xd, w);
396   } else {
397     pack_inter_mode_mvs(cpi, xd, mbmi_ext, w, max_mv_magnitude,
398                         interp_filter_selected);
399   }
400 
401   assert(*tok < tok_end);
402   pack_mb_tokens(w, tok, tok_end, cm->bit_depth);
403 }
404 
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,vpx_writer * w)405 static void write_partition(const VP9_COMMON *const cm,
406                             const MACROBLOCKD *const xd, int hbs, int mi_row,
407                             int mi_col, PARTITION_TYPE p, BLOCK_SIZE bsize,
408                             vpx_writer *w) {
409   const int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
410   const vpx_prob *const probs = xd->partition_probs[ctx];
411   const int has_rows = (mi_row + hbs) < cm->mi_rows;
412   const int has_cols = (mi_col + hbs) < cm->mi_cols;
413 
414   if (has_rows && has_cols) {
415     vp9_write_token(w, vp9_partition_tree, probs, &partition_encodings[p]);
416   } else if (!has_rows && has_cols) {
417     assert(p == PARTITION_SPLIT || p == PARTITION_HORZ);
418     vpx_write(w, p == PARTITION_SPLIT, probs[1]);
419   } else if (has_rows && !has_cols) {
420     assert(p == PARTITION_SPLIT || p == PARTITION_VERT);
421     vpx_write(w, p == PARTITION_SPLIT, probs[2]);
422   } else {
423     assert(p == PARTITION_SPLIT);
424   }
425 }
426 
write_modes_sb(VP9_COMP * cpi,MACROBLOCKD * const xd,const TileInfo * const tile,vpx_writer * w,TOKENEXTRA ** tok,const TOKENEXTRA * const tok_end,int mi_row,int mi_col,BLOCK_SIZE bsize,unsigned int * const max_mv_magnitude,int interp_filter_selected[][SWITCHABLE])427 static void write_modes_sb(VP9_COMP *cpi, MACROBLOCKD *const xd,
428                            const TileInfo *const tile, vpx_writer *w,
429                            TOKENEXTRA **tok, const TOKENEXTRA *const tok_end,
430                            int mi_row, int mi_col, BLOCK_SIZE bsize,
431                            unsigned int *const max_mv_magnitude,
432                            int interp_filter_selected[][SWITCHABLE]) {
433   const VP9_COMMON *const cm = &cpi->common;
434   const int bsl = b_width_log2_lookup[bsize];
435   const int bs = (1 << bsl) / 4;
436   PARTITION_TYPE partition;
437   BLOCK_SIZE subsize;
438   const MODE_INFO *m = NULL;
439 
440   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
441 
442   m = cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col];
443 
444   partition = partition_lookup[bsl][m->sb_type];
445   write_partition(cm, xd, bs, mi_row, mi_col, partition, bsize, w);
446   subsize = get_subsize(bsize, partition);
447   if (subsize < BLOCK_8X8) {
448     write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
449                   max_mv_magnitude, interp_filter_selected);
450   } else {
451     switch (partition) {
452       case PARTITION_NONE:
453         write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
454                       max_mv_magnitude, interp_filter_selected);
455         break;
456       case PARTITION_HORZ:
457         write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
458                       max_mv_magnitude, interp_filter_selected);
459         if (mi_row + bs < cm->mi_rows)
460           write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row + bs, mi_col,
461                         max_mv_magnitude, interp_filter_selected);
462         break;
463       case PARTITION_VERT:
464         write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
465                       max_mv_magnitude, interp_filter_selected);
466         if (mi_col + bs < cm->mi_cols)
467           write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col + bs,
468                         max_mv_magnitude, interp_filter_selected);
469         break;
470       default:
471         assert(partition == PARTITION_SPLIT);
472         write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col, subsize,
473                        max_mv_magnitude, interp_filter_selected);
474         write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col + bs,
475                        subsize, max_mv_magnitude, interp_filter_selected);
476         write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row + bs, mi_col,
477                        subsize, max_mv_magnitude, interp_filter_selected);
478         write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row + bs, mi_col + bs,
479                        subsize, max_mv_magnitude, interp_filter_selected);
480         break;
481     }
482   }
483 
484   // update partition context
485   if (bsize >= BLOCK_8X8 &&
486       (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
487     update_partition_context(xd, mi_row, mi_col, subsize, bsize);
488 }
489 
write_modes(VP9_COMP * cpi,MACROBLOCKD * const xd,const TileInfo * const tile,vpx_writer * w,int tile_row,int tile_col,unsigned int * const max_mv_magnitude,int interp_filter_selected[][SWITCHABLE])490 static void write_modes(VP9_COMP *cpi, MACROBLOCKD *const xd,
491                         const TileInfo *const tile, vpx_writer *w, int tile_row,
492                         int tile_col, unsigned int *const max_mv_magnitude,
493                         int interp_filter_selected[][SWITCHABLE]) {
494   const VP9_COMMON *const cm = &cpi->common;
495   int mi_row, mi_col, tile_sb_row;
496   TOKENEXTRA *tok = NULL;
497   TOKENEXTRA *tok_end = NULL;
498 
499   set_partition_probs(cm, xd);
500 
501   for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
502        mi_row += MI_BLOCK_SIZE) {
503     tile_sb_row = mi_cols_aligned_to_sb(mi_row - tile->mi_row_start) >>
504                   MI_BLOCK_SIZE_LOG2;
505     tok = cpi->tplist[tile_row][tile_col][tile_sb_row].start;
506     tok_end = tok + cpi->tplist[tile_row][tile_col][tile_sb_row].count;
507 
508     vp9_zero(xd->left_seg_context);
509     for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
510          mi_col += MI_BLOCK_SIZE)
511       write_modes_sb(cpi, xd, tile, w, &tok, tok_end, mi_row, mi_col,
512                      BLOCK_64X64, max_mv_magnitude, interp_filter_selected);
513 
514     assert(tok == cpi->tplist[tile_row][tile_col][tile_sb_row].stop);
515   }
516 }
517 
build_tree_distribution(VP9_COMP * cpi,TX_SIZE tx_size,vp9_coeff_stats * coef_branch_ct,vp9_coeff_probs_model * coef_probs)518 static void build_tree_distribution(VP9_COMP *cpi, TX_SIZE tx_size,
519                                     vp9_coeff_stats *coef_branch_ct,
520                                     vp9_coeff_probs_model *coef_probs) {
521   vp9_coeff_count *coef_counts = cpi->td.rd_counts.coef_counts[tx_size];
522   unsigned int(*eob_branch_ct)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS] =
523       cpi->common.counts.eob_branch[tx_size];
524   int i, j, k, l, m;
525 
526   for (i = 0; i < PLANE_TYPES; ++i) {
527     for (j = 0; j < REF_TYPES; ++j) {
528       for (k = 0; k < COEF_BANDS; ++k) {
529         for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
530           vp9_tree_probs_from_distribution(vp9_coef_tree,
531                                            coef_branch_ct[i][j][k][l],
532                                            coef_counts[i][j][k][l]);
533           coef_branch_ct[i][j][k][l][0][1] =
534               eob_branch_ct[i][j][k][l] - coef_branch_ct[i][j][k][l][0][0];
535           for (m = 0; m < UNCONSTRAINED_NODES; ++m)
536             coef_probs[i][j][k][l][m] =
537                 get_binary_prob(coef_branch_ct[i][j][k][l][m][0],
538                                 coef_branch_ct[i][j][k][l][m][1]);
539         }
540       }
541     }
542   }
543 }
544 
update_coef_probs_common(vpx_writer * const bc,VP9_COMP * cpi,TX_SIZE tx_size,vp9_coeff_stats * frame_branch_ct,vp9_coeff_probs_model * new_coef_probs)545 static void update_coef_probs_common(vpx_writer *const bc, VP9_COMP *cpi,
546                                      TX_SIZE tx_size,
547                                      vp9_coeff_stats *frame_branch_ct,
548                                      vp9_coeff_probs_model *new_coef_probs) {
549   vp9_coeff_probs_model *old_coef_probs = cpi->common.fc->coef_probs[tx_size];
550   const vpx_prob upd = DIFF_UPDATE_PROB;
551   const int entropy_nodes_update = UNCONSTRAINED_NODES;
552   int i, j, k, l, t;
553   int stepsize = cpi->sf.coeff_prob_appx_step;
554 
555   switch (cpi->sf.use_fast_coef_updates) {
556     case TWO_LOOP: {
557       /* dry run to see if there is any update at all needed */
558       int64_t savings = 0;
559       int update[2] = { 0, 0 };
560       for (i = 0; i < PLANE_TYPES; ++i) {
561         for (j = 0; j < REF_TYPES; ++j) {
562           for (k = 0; k < COEF_BANDS; ++k) {
563             for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
564               for (t = 0; t < entropy_nodes_update; ++t) {
565                 vpx_prob newp = new_coef_probs[i][j][k][l][t];
566                 const vpx_prob oldp = old_coef_probs[i][j][k][l][t];
567                 int64_t s;
568                 int u = 0;
569                 if (t == PIVOT_NODE)
570                   s = vp9_prob_diff_update_savings_search_model(
571                       frame_branch_ct[i][j][k][l][0], oldp, &newp, upd,
572                       stepsize);
573                 else
574                   s = vp9_prob_diff_update_savings_search(
575                       frame_branch_ct[i][j][k][l][t], oldp, &newp, upd);
576                 if (s > 0 && newp != oldp) u = 1;
577                 if (u)
578                   savings += s - (int)(vp9_cost_zero(upd));
579                 else
580                   savings -= (int)(vp9_cost_zero(upd));
581                 update[u]++;
582               }
583             }
584           }
585         }
586       }
587 
588       // printf("Update %d %d, savings %d\n", update[0], update[1], savings);
589       /* Is coef updated at all */
590       if (update[1] == 0 || savings < 0) {
591         vpx_write_bit(bc, 0);
592         return;
593       }
594       vpx_write_bit(bc, 1);
595       for (i = 0; i < PLANE_TYPES; ++i) {
596         for (j = 0; j < REF_TYPES; ++j) {
597           for (k = 0; k < COEF_BANDS; ++k) {
598             for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
599               // calc probs and branch cts for this frame only
600               for (t = 0; t < entropy_nodes_update; ++t) {
601                 vpx_prob newp = new_coef_probs[i][j][k][l][t];
602                 vpx_prob *oldp = old_coef_probs[i][j][k][l] + t;
603                 int64_t s;
604                 int u = 0;
605                 if (t == PIVOT_NODE)
606                   s = vp9_prob_diff_update_savings_search_model(
607                       frame_branch_ct[i][j][k][l][0], *oldp, &newp, upd,
608                       stepsize);
609                 else
610                   s = vp9_prob_diff_update_savings_search(
611                       frame_branch_ct[i][j][k][l][t], *oldp, &newp, upd);
612                 if (s > 0 && newp != *oldp) u = 1;
613                 vpx_write(bc, u, upd);
614                 if (u) {
615                   /* send/use new probability */
616                   vp9_write_prob_diff_update(bc, newp, *oldp);
617                   *oldp = newp;
618                 }
619               }
620             }
621           }
622         }
623       }
624       return;
625     }
626 
627     default: {
628       int updates = 0;
629       int noupdates_before_first = 0;
630       assert(cpi->sf.use_fast_coef_updates == ONE_LOOP_REDUCED);
631       for (i = 0; i < PLANE_TYPES; ++i) {
632         for (j = 0; j < REF_TYPES; ++j) {
633           for (k = 0; k < COEF_BANDS; ++k) {
634             for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
635               // calc probs and branch cts for this frame only
636               for (t = 0; t < entropy_nodes_update; ++t) {
637                 vpx_prob newp = new_coef_probs[i][j][k][l][t];
638                 vpx_prob *oldp = old_coef_probs[i][j][k][l] + t;
639                 int64_t s;
640                 int u = 0;
641 
642                 if (t == PIVOT_NODE) {
643                   s = vp9_prob_diff_update_savings_search_model(
644                       frame_branch_ct[i][j][k][l][0], *oldp, &newp, upd,
645                       stepsize);
646                 } else {
647                   s = vp9_prob_diff_update_savings_search(
648                       frame_branch_ct[i][j][k][l][t], *oldp, &newp, upd);
649                 }
650 
651                 if (s > 0 && newp != *oldp) u = 1;
652                 updates += u;
653                 if (u == 0 && updates == 0) {
654                   noupdates_before_first++;
655                   continue;
656                 }
657                 if (u == 1 && updates == 1) {
658                   int v;
659                   // first update
660                   vpx_write_bit(bc, 1);
661                   for (v = 0; v < noupdates_before_first; ++v)
662                     vpx_write(bc, 0, upd);
663                 }
664                 vpx_write(bc, u, upd);
665                 if (u) {
666                   /* send/use new probability */
667                   vp9_write_prob_diff_update(bc, newp, *oldp);
668                   *oldp = newp;
669                 }
670               }
671             }
672           }
673         }
674       }
675       if (updates == 0) {
676         vpx_write_bit(bc, 0);  // no updates
677       }
678       return;
679     }
680   }
681 }
682 
update_coef_probs(VP9_COMP * cpi,vpx_writer * w)683 static void update_coef_probs(VP9_COMP *cpi, vpx_writer *w) {
684   const TX_MODE tx_mode = cpi->common.tx_mode;
685   const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
686   TX_SIZE tx_size;
687   for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size) {
688     vp9_coeff_stats frame_branch_ct[PLANE_TYPES];
689     vp9_coeff_probs_model frame_coef_probs[PLANE_TYPES];
690     if (cpi->td.counts->tx.tx_totals[tx_size] <= 20 ||
691         (tx_size >= TX_16X16 && cpi->sf.tx_size_search_method == USE_TX_8X8)) {
692       vpx_write_bit(w, 0);
693     } else {
694       build_tree_distribution(cpi, tx_size, frame_branch_ct, frame_coef_probs);
695       update_coef_probs_common(w, cpi, tx_size, frame_branch_ct,
696                                frame_coef_probs);
697     }
698   }
699 }
700 
encode_loopfilter(struct loopfilter * lf,struct vpx_write_bit_buffer * wb)701 static void encode_loopfilter(struct loopfilter *lf,
702                               struct vpx_write_bit_buffer *wb) {
703   int i;
704 
705   // Encode the loop filter level and type
706   vpx_wb_write_literal(wb, lf->filter_level, 6);
707   vpx_wb_write_literal(wb, lf->sharpness_level, 3);
708 
709   // Write out loop filter deltas applied at the MB level based on mode or
710   // ref frame (if they are enabled).
711   vpx_wb_write_bit(wb, lf->mode_ref_delta_enabled);
712 
713   if (lf->mode_ref_delta_enabled) {
714     vpx_wb_write_bit(wb, lf->mode_ref_delta_update);
715     if (lf->mode_ref_delta_update) {
716       for (i = 0; i < MAX_REF_LF_DELTAS; i++) {
717         const int delta = lf->ref_deltas[i];
718         const int changed = delta != lf->last_ref_deltas[i];
719         vpx_wb_write_bit(wb, changed);
720         if (changed) {
721           lf->last_ref_deltas[i] = delta;
722           vpx_wb_write_literal(wb, abs(delta) & 0x3F, 6);
723           vpx_wb_write_bit(wb, delta < 0);
724         }
725       }
726 
727       for (i = 0; i < MAX_MODE_LF_DELTAS; i++) {
728         const int delta = lf->mode_deltas[i];
729         const int changed = delta != lf->last_mode_deltas[i];
730         vpx_wb_write_bit(wb, changed);
731         if (changed) {
732           lf->last_mode_deltas[i] = delta;
733           vpx_wb_write_literal(wb, abs(delta) & 0x3F, 6);
734           vpx_wb_write_bit(wb, delta < 0);
735         }
736       }
737     }
738   }
739 }
740 
write_delta_q(struct vpx_write_bit_buffer * wb,int delta_q)741 static void write_delta_q(struct vpx_write_bit_buffer *wb, int delta_q) {
742   if (delta_q != 0) {
743     vpx_wb_write_bit(wb, 1);
744     vpx_wb_write_literal(wb, abs(delta_q), 4);
745     vpx_wb_write_bit(wb, delta_q < 0);
746   } else {
747     vpx_wb_write_bit(wb, 0);
748   }
749 }
750 
encode_quantization(const VP9_COMMON * const cm,struct vpx_write_bit_buffer * wb)751 static void encode_quantization(const VP9_COMMON *const cm,
752                                 struct vpx_write_bit_buffer *wb) {
753   vpx_wb_write_literal(wb, cm->base_qindex, QINDEX_BITS);
754   write_delta_q(wb, cm->y_dc_delta_q);
755   write_delta_q(wb, cm->uv_dc_delta_q);
756   write_delta_q(wb, cm->uv_ac_delta_q);
757 }
758 
encode_segmentation(VP9_COMMON * cm,MACROBLOCKD * xd,struct vpx_write_bit_buffer * wb)759 static void encode_segmentation(VP9_COMMON *cm, MACROBLOCKD *xd,
760                                 struct vpx_write_bit_buffer *wb) {
761   int i, j;
762 
763   const struct segmentation *seg = &cm->seg;
764 
765   vpx_wb_write_bit(wb, seg->enabled);
766   if (!seg->enabled) return;
767 
768   // Segmentation map
769   vpx_wb_write_bit(wb, seg->update_map);
770   if (seg->update_map) {
771     // Select the coding strategy (temporal or spatial)
772     vp9_choose_segmap_coding_method(cm, xd);
773     // Write out probabilities used to decode unpredicted  macro-block segments
774     for (i = 0; i < SEG_TREE_PROBS; i++) {
775       const int prob = seg->tree_probs[i];
776       const int update = prob != MAX_PROB;
777       vpx_wb_write_bit(wb, update);
778       if (update) vpx_wb_write_literal(wb, prob, 8);
779     }
780 
781     // Write out the chosen coding method.
782     vpx_wb_write_bit(wb, seg->temporal_update);
783     if (seg->temporal_update) {
784       for (i = 0; i < PREDICTION_PROBS; i++) {
785         const int prob = seg->pred_probs[i];
786         const int update = prob != MAX_PROB;
787         vpx_wb_write_bit(wb, update);
788         if (update) vpx_wb_write_literal(wb, prob, 8);
789       }
790     }
791   }
792 
793   // Segmentation data
794   vpx_wb_write_bit(wb, seg->update_data);
795   if (seg->update_data) {
796     vpx_wb_write_bit(wb, seg->abs_delta);
797 
798     for (i = 0; i < MAX_SEGMENTS; i++) {
799       for (j = 0; j < SEG_LVL_MAX; j++) {
800         const int active = segfeature_active(seg, i, j);
801         vpx_wb_write_bit(wb, active);
802         if (active) {
803           const int data = get_segdata(seg, i, j);
804           const int data_max = vp9_seg_feature_data_max(j);
805 
806           if (vp9_is_segfeature_signed(j)) {
807             encode_unsigned_max(wb, abs(data), data_max);
808             vpx_wb_write_bit(wb, data < 0);
809           } else {
810             encode_unsigned_max(wb, data, data_max);
811           }
812         }
813       }
814     }
815   }
816 }
817 
encode_txfm_probs(VP9_COMMON * cm,vpx_writer * w,FRAME_COUNTS * counts)818 static void encode_txfm_probs(VP9_COMMON *cm, vpx_writer *w,
819                               FRAME_COUNTS *counts) {
820   // Mode
821   vpx_write_literal(w, VPXMIN(cm->tx_mode, ALLOW_32X32), 2);
822   if (cm->tx_mode >= ALLOW_32X32)
823     vpx_write_bit(w, cm->tx_mode == TX_MODE_SELECT);
824 
825   // Probabilities
826   if (cm->tx_mode == TX_MODE_SELECT) {
827     int i, j;
828     unsigned int ct_8x8p[TX_SIZES - 3][2];
829     unsigned int ct_16x16p[TX_SIZES - 2][2];
830     unsigned int ct_32x32p[TX_SIZES - 1][2];
831 
832     for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
833       tx_counts_to_branch_counts_8x8(counts->tx.p8x8[i], ct_8x8p);
834       for (j = 0; j < TX_SIZES - 3; j++)
835         vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p8x8[i][j], ct_8x8p[j]);
836     }
837 
838     for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
839       tx_counts_to_branch_counts_16x16(counts->tx.p16x16[i], ct_16x16p);
840       for (j = 0; j < TX_SIZES - 2; j++)
841         vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p16x16[i][j],
842                                   ct_16x16p[j]);
843     }
844 
845     for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
846       tx_counts_to_branch_counts_32x32(counts->tx.p32x32[i], ct_32x32p);
847       for (j = 0; j < TX_SIZES - 1; j++)
848         vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p32x32[i][j],
849                                   ct_32x32p[j]);
850     }
851   }
852 }
853 
write_interp_filter(INTERP_FILTER filter,struct vpx_write_bit_buffer * wb)854 static void write_interp_filter(INTERP_FILTER filter,
855                                 struct vpx_write_bit_buffer *wb) {
856   const int filter_to_literal[] = { 1, 0, 2, 3 };
857 
858   vpx_wb_write_bit(wb, filter == SWITCHABLE);
859   if (filter != SWITCHABLE)
860     vpx_wb_write_literal(wb, filter_to_literal[filter], 2);
861 }
862 
fix_interp_filter(VP9_COMMON * cm,FRAME_COUNTS * counts)863 static void fix_interp_filter(VP9_COMMON *cm, FRAME_COUNTS *counts) {
864   if (cm->interp_filter == SWITCHABLE) {
865     // Check to see if only one of the filters is actually used
866     int count[SWITCHABLE_FILTERS];
867     int i, j, c = 0;
868     for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
869       count[i] = 0;
870       for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
871         count[i] += counts->switchable_interp[j][i];
872       c += (count[i] > 0);
873     }
874     if (c == 1) {
875       // Only one filter is used. So set the filter at frame level
876       for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
877         if (count[i]) {
878           cm->interp_filter = i;
879           break;
880         }
881       }
882     }
883   }
884 }
885 
write_tile_info(const VP9_COMMON * const cm,struct vpx_write_bit_buffer * wb)886 static void write_tile_info(const VP9_COMMON *const cm,
887                             struct vpx_write_bit_buffer *wb) {
888   int min_log2_tile_cols, max_log2_tile_cols, ones;
889   vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
890 
891   // columns
892   ones = cm->log2_tile_cols - min_log2_tile_cols;
893   while (ones--) vpx_wb_write_bit(wb, 1);
894 
895   if (cm->log2_tile_cols < max_log2_tile_cols) vpx_wb_write_bit(wb, 0);
896 
897   // rows
898   vpx_wb_write_bit(wb, cm->log2_tile_rows != 0);
899   if (cm->log2_tile_rows != 0) vpx_wb_write_bit(wb, cm->log2_tile_rows != 1);
900 }
901 
vp9_get_refresh_mask(VP9_COMP * cpi)902 int vp9_get_refresh_mask(VP9_COMP *cpi) {
903   if (cpi->ext_ratectrl.ready &&
904       (cpi->ext_ratectrl.funcs.rc_type & VPX_RC_GOP) != 0 &&
905       cpi->ext_ratectrl.funcs.get_gop_decision != NULL) {
906     GF_GROUP *const gf_group = &cpi->twopass.gf_group;
907     const int this_gf_index = gf_group->index;
908     const int update_ref_idx = gf_group->update_ref_idx[this_gf_index];
909 
910     if (update_ref_idx != INVALID_IDX) {
911       return (1 << update_ref_idx);
912     } else {
913       return 0;
914     }
915   }
916   if (vp9_preserve_existing_gf(cpi)) {
917     // We have decided to preserve the previously existing golden frame as our
918     // new ARF frame. However, in the short term we leave it in the GF slot and,
919     // if we're updating the GF with the current decoded frame, we save it
920     // instead to the ARF slot.
921     // Later, in the function vp9_encoder.c:vp9_update_reference_frames() we
922     // will swap gld_fb_idx and alt_fb_idx to achieve our objective. We do it
923     // there so that it can be done outside of the recode loop.
924     // Note: This is highly specific to the use of ARF as a forward reference,
925     // and this needs to be generalized as other uses are implemented
926     // (like RTC/temporal scalability).
927     return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
928            (cpi->refresh_golden_frame << cpi->alt_fb_idx);
929   } else {
930     int arf_idx = cpi->alt_fb_idx;
931     GF_GROUP *const gf_group = &cpi->twopass.gf_group;
932 
933     if (cpi->multi_layer_arf) {
934       for (arf_idx = 0; arf_idx < REF_FRAMES; ++arf_idx) {
935         if (arf_idx != cpi->alt_fb_idx && arf_idx != cpi->lst_fb_idx &&
936             arf_idx != cpi->gld_fb_idx) {
937           int idx;
938           for (idx = 0; idx < gf_group->stack_size; ++idx)
939             if (arf_idx == gf_group->arf_index_stack[idx]) break;
940           if (idx == gf_group->stack_size) break;
941         }
942       }
943     }
944     cpi->twopass.gf_group.top_arf_idx = arf_idx;
945 
946     if (cpi->use_svc && cpi->svc.use_set_ref_frame_config &&
947         cpi->svc.temporal_layering_mode == VP9E_TEMPORAL_LAYERING_MODE_BYPASS)
948       return cpi->svc.update_buffer_slot[cpi->svc.spatial_layer_id];
949     return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
950            (cpi->refresh_golden_frame << cpi->gld_fb_idx) |
951            (cpi->refresh_alt_ref_frame << arf_idx);
952   }
953 }
954 
encode_tile_worker(void * arg1,void * arg2)955 static int encode_tile_worker(void *arg1, void *arg2) {
956   VP9_COMP *cpi = (VP9_COMP *)arg1;
957   VP9BitstreamWorkerData *data = (VP9BitstreamWorkerData *)arg2;
958   MACROBLOCKD *const xd = &data->xd;
959   const int tile_row = 0;
960   vpx_start_encode(&data->bit_writer, data->dest, data->dest_size);
961   write_modes(cpi, xd, &cpi->tile_data[data->tile_idx].tile_info,
962               &data->bit_writer, tile_row, data->tile_idx,
963               &data->max_mv_magnitude, data->interp_filter_selected);
964   return vpx_stop_encode(&data->bit_writer) == 0;
965 }
966 
vp9_bitstream_encode_tiles_buffer_dealloc(VP9_COMP * const cpi)967 void vp9_bitstream_encode_tiles_buffer_dealloc(VP9_COMP *const cpi) {
968   if (cpi->vp9_bitstream_worker_data) {
969     int i;
970     for (i = 1; i < cpi->num_workers; ++i) {
971       vpx_free(cpi->vp9_bitstream_worker_data[i].dest);
972     }
973     vpx_free(cpi->vp9_bitstream_worker_data);
974     cpi->vp9_bitstream_worker_data = NULL;
975   }
976 }
977 
encode_tiles_buffer_alloc_size(const VP9_COMP * cpi)978 static size_t encode_tiles_buffer_alloc_size(const VP9_COMP *cpi) {
979   const VP9_COMMON *cm = &cpi->common;
980   const int image_bps =
981       (8 + 2 * (8 >> (cm->subsampling_x + cm->subsampling_y))) *
982       (1 + (cm->bit_depth > 8));
983   const int64_t size =
984       (int64_t)cpi->oxcf.width * cpi->oxcf.height * image_bps / 8;
985   return (size_t)size;
986 }
987 
encode_tiles_buffer_alloc(VP9_COMP * const cpi,size_t buffer_alloc_size)988 static void encode_tiles_buffer_alloc(VP9_COMP *const cpi,
989                                       size_t buffer_alloc_size) {
990   VP9_COMMON *const cm = &cpi->common;
991   int i;
992   const size_t worker_data_size =
993       cpi->num_workers * sizeof(*cpi->vp9_bitstream_worker_data);
994   CHECK_MEM_ERROR(&cm->error, cpi->vp9_bitstream_worker_data,
995                   vpx_memalign(16, worker_data_size));
996   memset(cpi->vp9_bitstream_worker_data, 0, worker_data_size);
997   for (i = 1; i < cpi->num_workers; ++i) {
998     CHECK_MEM_ERROR(&cm->error, cpi->vp9_bitstream_worker_data[i].dest,
999                     vpx_malloc(buffer_alloc_size));
1000     cpi->vp9_bitstream_worker_data[i].dest_size = buffer_alloc_size;
1001   }
1002 }
1003 
encode_tiles_mt(VP9_COMP * cpi,uint8_t * data_ptr,size_t data_size)1004 static size_t encode_tiles_mt(VP9_COMP *cpi, uint8_t *data_ptr,
1005                               size_t data_size) {
1006   const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
1007   VP9_COMMON *const cm = &cpi->common;
1008   const int tile_cols = 1 << cm->log2_tile_cols;
1009   const int num_workers = cpi->num_workers;
1010   size_t total_size = 0;
1011   int tile_col = 0;
1012   int error = 0;
1013 
1014   const size_t buffer_alloc_size = encode_tiles_buffer_alloc_size(cpi);
1015   if (!cpi->vp9_bitstream_worker_data ||
1016       cpi->vp9_bitstream_worker_data[1].dest_size != buffer_alloc_size) {
1017     vp9_bitstream_encode_tiles_buffer_dealloc(cpi);
1018     encode_tiles_buffer_alloc(cpi, buffer_alloc_size);
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         const size_t offset = total_size + (tile_col == tile_cols - 1 ? 0 : 4);
1039         if (data_size < offset) {
1040           vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
1041                              "encode_tiles_mt: output buffer full");
1042         }
1043         data->dest = data_ptr + offset;
1044         data->dest_size = data_size - offset;
1045       }
1046       worker->data1 = cpi;
1047       worker->data2 = data;
1048       worker->hook = encode_tile_worker;
1049       worker->had_error = 0;
1050 
1051       if (i < num_workers - 1) {
1052         winterface->launch(worker);
1053       } else {
1054         winterface->execute(worker);
1055       }
1056       ++tile_col;
1057     }
1058     for (j = 0; j < i; ++j) {
1059       VPxWorker *const worker = &cpi->workers[j];
1060       VP9BitstreamWorkerData *const data =
1061           (VP9BitstreamWorkerData *)worker->data2;
1062       uint32_t tile_size;
1063       int k;
1064 
1065       if (!winterface->sync(worker)) {
1066         error = 1;
1067         continue;
1068       }
1069 
1070       tile_size = data->bit_writer.pos;
1071 
1072       // Aggregate per-thread bitstream stats.
1073       cpi->max_mv_magnitude =
1074           VPXMAX(cpi->max_mv_magnitude, data->max_mv_magnitude);
1075       for (k = 0; k < SWITCHABLE; ++k) {
1076         cpi->interp_filter_selected[0][k] += data->interp_filter_selected[0][k];
1077       }
1078 
1079       // Prefix the size of the tile on all but the last.
1080       if (tile_col != tile_cols || j < i - 1) {
1081         if (data_size - total_size < 4) {
1082           error = 1;
1083           continue;
1084         }
1085         mem_put_be32(data_ptr + total_size, tile_size);
1086         total_size += 4;
1087       }
1088       if (j > 0) {
1089         if (data_size - total_size < tile_size) {
1090           error = 1;
1091           continue;
1092         }
1093         memcpy(data_ptr + total_size, data->dest, tile_size);
1094       }
1095       total_size += tile_size;
1096     }
1097     if (error) {
1098       vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
1099                          "encode_tiles_mt: output buffer full");
1100     }
1101   }
1102   return total_size;
1103 }
1104 
encode_tiles(VP9_COMP * cpi,uint8_t * data_ptr,size_t data_size)1105 static size_t encode_tiles(VP9_COMP *cpi, uint8_t *data_ptr, size_t data_size) {
1106   VP9_COMMON *const cm = &cpi->common;
1107   MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
1108   vpx_writer residual_bc;
1109   int tile_row, tile_col;
1110   size_t total_size = 0;
1111   const int tile_cols = 1 << cm->log2_tile_cols;
1112   const int tile_rows = 1 << cm->log2_tile_rows;
1113 
1114   memset(cm->above_seg_context, 0,
1115          sizeof(*cm->above_seg_context) * mi_cols_aligned_to_sb(cm->mi_cols));
1116 
1117   // Encoding tiles in parallel is done only for realtime mode now. In other
1118   // modes the speed up is insignificant and requires further testing to ensure
1119   // that it does not make the overall process worse in any case.
1120   if (cpi->oxcf.mode == REALTIME && cpi->num_workers > 1 && tile_rows == 1 &&
1121       tile_cols > 1) {
1122     return encode_tiles_mt(cpi, data_ptr, data_size);
1123   }
1124 
1125   for (tile_row = 0; tile_row < tile_rows; tile_row++) {
1126     for (tile_col = 0; tile_col < tile_cols; tile_col++) {
1127       int tile_idx = tile_row * tile_cols + tile_col;
1128 
1129       size_t offset;
1130       if (tile_col < tile_cols - 1 || tile_row < tile_rows - 1)
1131         offset = total_size + 4;
1132       else
1133         offset = total_size;
1134       if (data_size < offset) {
1135         vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
1136                            "encode_tiles: output buffer full");
1137       }
1138       vpx_start_encode(&residual_bc, data_ptr + offset, data_size - offset);
1139 
1140       write_modes(cpi, xd, &cpi->tile_data[tile_idx].tile_info, &residual_bc,
1141                   tile_row, tile_col, &cpi->max_mv_magnitude,
1142                   cpi->interp_filter_selected);
1143 
1144       if (vpx_stop_encode(&residual_bc)) {
1145         vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
1146                            "encode_tiles: output buffer full");
1147       }
1148       if (tile_col < tile_cols - 1 || tile_row < tile_rows - 1) {
1149         // size of this tile
1150         mem_put_be32(data_ptr + total_size, residual_bc.pos);
1151         total_size += 4;
1152       }
1153 
1154       total_size += residual_bc.pos;
1155     }
1156   }
1157   return total_size;
1158 }
1159 
write_render_size(const VP9_COMMON * cm,struct vpx_write_bit_buffer * wb)1160 static void write_render_size(const VP9_COMMON *cm,
1161                               struct vpx_write_bit_buffer *wb) {
1162   const int scaling_active =
1163       cm->width != cm->render_width || cm->height != cm->render_height;
1164   vpx_wb_write_bit(wb, scaling_active);
1165   if (scaling_active) {
1166     vpx_wb_write_literal(wb, cm->render_width - 1, 16);
1167     vpx_wb_write_literal(wb, cm->render_height - 1, 16);
1168   }
1169 }
1170 
write_frame_size(const VP9_COMMON * cm,struct vpx_write_bit_buffer * wb)1171 static void write_frame_size(const VP9_COMMON *cm,
1172                              struct vpx_write_bit_buffer *wb) {
1173   vpx_wb_write_literal(wb, cm->width - 1, 16);
1174   vpx_wb_write_literal(wb, cm->height - 1, 16);
1175 
1176   write_render_size(cm, wb);
1177 }
1178 
write_frame_size_with_refs(VP9_COMP * cpi,struct vpx_write_bit_buffer * wb)1179 static void write_frame_size_with_refs(VP9_COMP *cpi,
1180                                        struct vpx_write_bit_buffer *wb) {
1181   VP9_COMMON *const cm = &cpi->common;
1182   int found = 0;
1183 
1184   MV_REFERENCE_FRAME ref_frame;
1185   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
1186     YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi, ref_frame);
1187 
1188     // Set "found" to 0 for temporal svc and for spatial svc key frame
1189     if (cpi->use_svc &&
1190         ((cpi->svc.number_temporal_layers > 1 &&
1191           cpi->oxcf.rc_mode == VPX_CBR) ||
1192          (cpi->svc.number_spatial_layers > 1 &&
1193           cpi->svc.layer_context[cpi->svc.spatial_layer_id].is_key_frame))) {
1194       found = 0;
1195     } else if (cfg != NULL) {
1196       found =
1197           cm->width == cfg->y_crop_width && cm->height == cfg->y_crop_height;
1198     }
1199     vpx_wb_write_bit(wb, found);
1200     if (found) {
1201       break;
1202     }
1203   }
1204 
1205   if (!found) {
1206     vpx_wb_write_literal(wb, cm->width - 1, 16);
1207     vpx_wb_write_literal(wb, cm->height - 1, 16);
1208   }
1209 
1210   write_render_size(cm, wb);
1211 }
1212 
write_sync_code(struct vpx_write_bit_buffer * wb)1213 static void write_sync_code(struct vpx_write_bit_buffer *wb) {
1214   vpx_wb_write_literal(wb, VP9_SYNC_CODE_0, 8);
1215   vpx_wb_write_literal(wb, VP9_SYNC_CODE_1, 8);
1216   vpx_wb_write_literal(wb, VP9_SYNC_CODE_2, 8);
1217 }
1218 
write_profile(BITSTREAM_PROFILE profile,struct vpx_write_bit_buffer * wb)1219 static void write_profile(BITSTREAM_PROFILE profile,
1220                           struct vpx_write_bit_buffer *wb) {
1221   switch (profile) {
1222     case PROFILE_0: vpx_wb_write_literal(wb, 0, 2); break;
1223     case PROFILE_1: vpx_wb_write_literal(wb, 2, 2); break;
1224     case PROFILE_2: vpx_wb_write_literal(wb, 1, 2); break;
1225     default:
1226       assert(profile == PROFILE_3);
1227       vpx_wb_write_literal(wb, 6, 3);
1228       break;
1229   }
1230 }
1231 
write_bitdepth_colorspace_sampling(VP9_COMMON * const cm,struct vpx_write_bit_buffer * wb)1232 static void write_bitdepth_colorspace_sampling(
1233     VP9_COMMON *const cm, struct vpx_write_bit_buffer *wb) {
1234   if (cm->profile >= PROFILE_2) {
1235     assert(cm->bit_depth > VPX_BITS_8);
1236     vpx_wb_write_bit(wb, cm->bit_depth == VPX_BITS_10 ? 0 : 1);
1237   }
1238   vpx_wb_write_literal(wb, cm->color_space, 3);
1239   if (cm->color_space != VPX_CS_SRGB) {
1240     // 0: [16, 235] (i.e. xvYCC), 1: [0, 255]
1241     vpx_wb_write_bit(wb, cm->color_range);
1242     if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
1243       assert(cm->subsampling_x != 1 || cm->subsampling_y != 1);
1244       vpx_wb_write_bit(wb, cm->subsampling_x);
1245       vpx_wb_write_bit(wb, cm->subsampling_y);
1246       vpx_wb_write_bit(wb, 0);  // unused
1247     } else {
1248       assert(cm->subsampling_x == 1 && cm->subsampling_y == 1);
1249     }
1250   } else {
1251     assert(cm->profile == PROFILE_1 || cm->profile == PROFILE_3);
1252     vpx_wb_write_bit(wb, 0);  // unused
1253   }
1254 }
1255 
write_uncompressed_header(VP9_COMP * cpi,struct vpx_write_bit_buffer * wb)1256 static void write_uncompressed_header(VP9_COMP *cpi,
1257                                       struct vpx_write_bit_buffer *wb) {
1258   VP9_COMMON *const cm = &cpi->common;
1259   MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
1260 
1261   vpx_wb_write_literal(wb, VP9_FRAME_MARKER, 2);
1262 
1263   write_profile(cm->profile, wb);
1264 
1265   // If to use show existing frame.
1266   vpx_wb_write_bit(wb, cm->show_existing_frame);
1267   if (cm->show_existing_frame) {
1268     vpx_wb_write_literal(wb, cpi->alt_fb_idx, 3);
1269     return;
1270   }
1271 
1272   vpx_wb_write_bit(wb, cm->frame_type);
1273   vpx_wb_write_bit(wb, cm->show_frame);
1274   vpx_wb_write_bit(wb, cm->error_resilient_mode);
1275 
1276   if (cm->frame_type == KEY_FRAME) {
1277     write_sync_code(wb);
1278     write_bitdepth_colorspace_sampling(cm, wb);
1279     write_frame_size(cm, wb);
1280   } else {
1281     if (!cm->show_frame) vpx_wb_write_bit(wb, cm->intra_only);
1282 
1283     if (!cm->error_resilient_mode)
1284       vpx_wb_write_literal(wb, cm->reset_frame_context, 2);
1285 
1286     if (cm->intra_only) {
1287       write_sync_code(wb);
1288 
1289       // Note for profile 0, 420 8bpp is assumed.
1290       if (cm->profile > PROFILE_0) {
1291         write_bitdepth_colorspace_sampling(cm, wb);
1292       }
1293 
1294       vpx_wb_write_literal(wb, vp9_get_refresh_mask(cpi), REF_FRAMES);
1295       write_frame_size(cm, wb);
1296     } else {
1297       MV_REFERENCE_FRAME ref_frame;
1298       vpx_wb_write_literal(wb, vp9_get_refresh_mask(cpi), REF_FRAMES);
1299       for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
1300         assert(get_ref_frame_map_idx(cpi, ref_frame) != INVALID_IDX);
1301         vpx_wb_write_literal(wb, get_ref_frame_map_idx(cpi, ref_frame),
1302                              REF_FRAMES_LOG2);
1303         vpx_wb_write_bit(wb, cm->ref_frame_sign_bias[ref_frame]);
1304       }
1305 
1306       write_frame_size_with_refs(cpi, wb);
1307 
1308       vpx_wb_write_bit(wb, cm->allow_high_precision_mv);
1309 
1310       fix_interp_filter(cm, cpi->td.counts);
1311       write_interp_filter(cm->interp_filter, wb);
1312     }
1313   }
1314 
1315   if (!cm->error_resilient_mode) {
1316     vpx_wb_write_bit(wb, cm->refresh_frame_context);
1317     vpx_wb_write_bit(wb, cm->frame_parallel_decoding_mode);
1318   }
1319 
1320   vpx_wb_write_literal(wb, cm->frame_context_idx, FRAME_CONTEXTS_LOG2);
1321 
1322   encode_loopfilter(&cm->lf, wb);
1323   encode_quantization(cm, wb);
1324   encode_segmentation(cm, xd, wb);
1325 
1326   write_tile_info(cm, wb);
1327 }
1328 
write_compressed_header(VP9_COMP * cpi,uint8_t * data,size_t data_size)1329 static size_t write_compressed_header(VP9_COMP *cpi, uint8_t *data,
1330                                       size_t data_size) {
1331   VP9_COMMON *const cm = &cpi->common;
1332   MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
1333   FRAME_CONTEXT *const fc = cm->fc;
1334   FRAME_COUNTS *counts = cpi->td.counts;
1335   vpx_writer header_bc;
1336 
1337   vpx_start_encode(&header_bc, data, data_size);
1338 
1339   if (xd->lossless)
1340     cm->tx_mode = ONLY_4X4;
1341   else
1342     encode_txfm_probs(cm, &header_bc, counts);
1343 
1344   update_coef_probs(cpi, &header_bc);
1345   update_skip_probs(cm, &header_bc, counts);
1346 
1347   if (!frame_is_intra_only(cm)) {
1348     int i;
1349 
1350     for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
1351       prob_diff_update(vp9_inter_mode_tree, cm->fc->inter_mode_probs[i],
1352                        counts->inter_mode[i], INTER_MODES, &header_bc);
1353 
1354     if (cm->interp_filter == SWITCHABLE)
1355       update_switchable_interp_probs(cm, &header_bc, counts);
1356 
1357     for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
1358       vp9_cond_prob_diff_update(&header_bc, &fc->intra_inter_prob[i],
1359                                 counts->intra_inter[i]);
1360 
1361     if (cpi->allow_comp_inter_inter) {
1362       const int use_compound_pred = cm->reference_mode != SINGLE_REFERENCE;
1363       const int use_hybrid_pred = cm->reference_mode == REFERENCE_MODE_SELECT;
1364 
1365       vpx_write_bit(&header_bc, use_compound_pred);
1366       if (use_compound_pred) {
1367         vpx_write_bit(&header_bc, use_hybrid_pred);
1368         if (use_hybrid_pred)
1369           for (i = 0; i < COMP_INTER_CONTEXTS; i++)
1370             vp9_cond_prob_diff_update(&header_bc, &fc->comp_inter_prob[i],
1371                                       counts->comp_inter[i]);
1372       }
1373     }
1374 
1375     if (cm->reference_mode != COMPOUND_REFERENCE) {
1376       for (i = 0; i < REF_CONTEXTS; i++) {
1377         vp9_cond_prob_diff_update(&header_bc, &fc->single_ref_prob[i][0],
1378                                   counts->single_ref[i][0]);
1379         vp9_cond_prob_diff_update(&header_bc, &fc->single_ref_prob[i][1],
1380                                   counts->single_ref[i][1]);
1381       }
1382     }
1383 
1384     if (cm->reference_mode != SINGLE_REFERENCE)
1385       for (i = 0; i < REF_CONTEXTS; i++)
1386         vp9_cond_prob_diff_update(&header_bc, &fc->comp_ref_prob[i],
1387                                   counts->comp_ref[i]);
1388 
1389     for (i = 0; i < BLOCK_SIZE_GROUPS; ++i)
1390       prob_diff_update(vp9_intra_mode_tree, cm->fc->y_mode_prob[i],
1391                        counts->y_mode[i], INTRA_MODES, &header_bc);
1392 
1393     for (i = 0; i < PARTITION_CONTEXTS; ++i)
1394       prob_diff_update(vp9_partition_tree, fc->partition_prob[i],
1395                        counts->partition[i], PARTITION_TYPES, &header_bc);
1396 
1397     vp9_write_nmv_probs(cm, cm->allow_high_precision_mv, &header_bc,
1398                         &counts->mv);
1399   }
1400 
1401   if (vpx_stop_encode(&header_bc)) {
1402     vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
1403                        "write_compressed_header: output buffer full");
1404   }
1405 
1406   return header_bc.pos;
1407 }
1408 
vp9_pack_bitstream(VP9_COMP * cpi,uint8_t * dest,size_t dest_size,size_t * size)1409 void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, size_t dest_size,
1410                         size_t *size) {
1411   VP9_COMMON *const cm = &cpi->common;
1412   uint8_t *data = dest;
1413   size_t data_size = dest_size;
1414   size_t uncompressed_hdr_size, compressed_hdr_size;
1415   struct vpx_write_bit_buffer wb;
1416   struct vpx_write_bit_buffer saved_wb;
1417 
1418 #if CONFIG_BITSTREAM_DEBUG
1419   bitstream_queue_reset_write();
1420 #endif
1421 
1422   vpx_wb_init(&wb, data, data_size);
1423   write_uncompressed_header(cpi, &wb);
1424   if (vpx_wb_has_error(&wb)) {
1425     vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
1426                        "vp9_pack_bitstream: output buffer full");
1427   }
1428 
1429   // Skip the rest coding process if use show existing frame.
1430   if (cm->show_existing_frame) {
1431     uncompressed_hdr_size = vpx_wb_bytes_written(&wb);
1432     data += uncompressed_hdr_size;
1433     *size = data - dest;
1434     return;
1435   }
1436 
1437   saved_wb = wb;
1438   // don't know in advance compressed header size
1439   vpx_wb_write_literal(&wb, 0, 16);
1440   if (vpx_wb_has_error(&wb)) {
1441     vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
1442                        "vp9_pack_bitstream: output buffer full");
1443   }
1444 
1445   uncompressed_hdr_size = vpx_wb_bytes_written(&wb);
1446   data += uncompressed_hdr_size;
1447   data_size -= uncompressed_hdr_size;
1448 
1449   vpx_clear_system_state();
1450 
1451   compressed_hdr_size = write_compressed_header(cpi, data, data_size);
1452   data += compressed_hdr_size;
1453   data_size -= compressed_hdr_size;
1454   if (compressed_hdr_size > UINT16_MAX) {
1455     vpx_internal_error(&cm->error, VPX_CODEC_ERROR,
1456                        "compressed_hdr_size > 16 bits");
1457   }
1458   vpx_wb_write_literal(&saved_wb, (int)compressed_hdr_size, 16);
1459   assert(!vpx_wb_has_error(&saved_wb));
1460 
1461   data += encode_tiles(cpi, data, data_size);
1462 
1463   *size = data - dest;
1464 }
1465