• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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 <stdio.h>
13 #include <limits.h>
14 
15 #include "vpx/vpx_encoder.h"
16 #include "vpx_dsp/bitwriter_buffer.h"
17 #include "vpx_dsp/vpx_dsp_common.h"
18 #include "vpx_mem/vpx_mem.h"
19 #include "vpx_ports/mem_ops.h"
20 #include "vpx_ports/system_state.h"
21 #if CONFIG_BITSTREAM_DEBUG
22 #include "vpx_util/vpx_debug_util.h"
23 #endif  // CONFIG_BITSTREAM_DEBUG
24 
25 #include "vp9/common/vp9_entropy.h"
26 #include "vp9/common/vp9_entropymode.h"
27 #include "vp9/common/vp9_entropymv.h"
28 #include "vp9/common/vp9_mvref_common.h"
29 #include "vp9/common/vp9_pred_common.h"
30 #include "vp9/common/vp9_seg_common.h"
31 #include "vp9/common/vp9_tile_common.h"
32 
33 #include "vp9/encoder/vp9_cost.h"
34 #include "vp9/encoder/vp9_bitstream.h"
35 #include "vp9/encoder/vp9_encodemv.h"
36 #include "vp9/encoder/vp9_mcomp.h"
37 #include "vp9/encoder/vp9_segmentation.h"
38 #include "vp9/encoder/vp9_subexp.h"
39 #include "vp9/encoder/vp9_tokenize.h"
40 
41 static const struct vp9_token intra_mode_encodings[INTRA_MODES] = {
42   { 0, 1 },  { 6, 3 },   { 28, 5 },  { 30, 5 }, { 58, 6 },
43   { 59, 6 }, { 126, 7 }, { 127, 7 }, { 62, 6 }, { 2, 2 }
44 };
45 static const struct vp9_token
46     switchable_interp_encodings[SWITCHABLE_FILTERS] = { { 0, 1 },
47                                                         { 2, 2 },
48                                                         { 3, 2 } };
49 static const struct vp9_token partition_encodings[PARTITION_TYPES] = {
50   { 0, 1 }, { 2, 2 }, { 6, 3 }, { 7, 3 }
51 };
52 static const struct vp9_token inter_mode_encodings[INTER_MODES] = {
53   { 2, 2 }, { 6, 3 }, { 0, 1 }, { 7, 3 }
54 };
55 
write_intra_mode(vpx_writer * w,PREDICTION_MODE mode,const vpx_prob * probs)56 static void write_intra_mode(vpx_writer *w, PREDICTION_MODE mode,
57                              const vpx_prob *probs) {
58   vp9_write_token(w, vp9_intra_mode_tree, probs, &intra_mode_encodings[mode]);
59 }
60 
write_inter_mode(vpx_writer * w,PREDICTION_MODE mode,const vpx_prob * probs)61 static void write_inter_mode(vpx_writer *w, PREDICTION_MODE mode,
62                              const vpx_prob *probs) {
63   assert(is_inter_mode(mode));
64   vp9_write_token(w, vp9_inter_mode_tree, probs,
65                   &inter_mode_encodings[INTER_OFFSET(mode)]);
66 }
67 
encode_unsigned_max(struct vpx_write_bit_buffer * wb,int data,int max)68 static void encode_unsigned_max(struct vpx_write_bit_buffer *wb, int data,
69                                 int max) {
70   vpx_wb_write_literal(wb, data, get_unsigned_bits(max));
71 }
72 
prob_diff_update(const vpx_tree_index * tree,vpx_prob probs[],const unsigned int counts[],int n,vpx_writer * w)73 static void prob_diff_update(const vpx_tree_index *tree,
74                              vpx_prob probs[/*n - 1*/],
75                              const unsigned int counts[/*n - 1*/], int n,
76                              vpx_writer *w) {
77   int i;
78   unsigned int branch_ct[32][2];
79 
80   // Assuming max number of probabilities <= 32
81   assert(n <= 32);
82 
83   vp9_tree_probs_from_distribution(tree, branch_ct, counts);
84   for (i = 0; i < n - 1; ++i)
85     vp9_cond_prob_diff_update(w, &probs[i], branch_ct[i]);
86 }
87 
write_selected_tx_size(const VP9_COMMON * cm,const MACROBLOCKD * const xd,vpx_writer * w)88 static void write_selected_tx_size(const VP9_COMMON *cm,
89                                    const MACROBLOCKD *const xd, vpx_writer *w) {
90   TX_SIZE tx_size = xd->mi[0]->tx_size;
91   BLOCK_SIZE bsize = xd->mi[0]->sb_type;
92   const TX_SIZE max_tx_size = max_txsize_lookup[bsize];
93   const vpx_prob *const tx_probs =
94       get_tx_probs(max_tx_size, get_tx_size_context(xd), &cm->fc->tx_probs);
95   vpx_write(w, tx_size != TX_4X4, tx_probs[0]);
96   if (tx_size != TX_4X4 && max_tx_size >= TX_16X16) {
97     vpx_write(w, tx_size != TX_8X8, tx_probs[1]);
98     if (tx_size != TX_8X8 && max_tx_size >= TX_32X32)
99       vpx_write(w, tx_size != TX_16X16, tx_probs[2]);
100   }
101 }
102 
write_skip(const VP9_COMMON * cm,const MACROBLOCKD * const xd,int segment_id,const MODE_INFO * mi,vpx_writer * w)103 static int write_skip(const VP9_COMMON *cm, const MACROBLOCKD *const xd,
104                       int segment_id, const MODE_INFO *mi, vpx_writer *w) {
105   if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
106     return 1;
107   } else {
108     const int skip = mi->skip;
109     vpx_write(w, skip, vp9_get_skip_prob(cm, xd));
110     return skip;
111   }
112 }
113 
update_skip_probs(VP9_COMMON * cm,vpx_writer * w,FRAME_COUNTS * counts)114 static void update_skip_probs(VP9_COMMON *cm, vpx_writer *w,
115                               FRAME_COUNTS *counts) {
116   int k;
117 
118   for (k = 0; k < SKIP_CONTEXTS; ++k)
119     vp9_cond_prob_diff_update(w, &cm->fc->skip_probs[k], counts->skip[k]);
120 }
121 
update_switchable_interp_probs(VP9_COMMON * cm,vpx_writer * w,FRAME_COUNTS * counts)122 static void update_switchable_interp_probs(VP9_COMMON *cm, vpx_writer *w,
123                                            FRAME_COUNTS *counts) {
124   int j;
125   for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
126     prob_diff_update(vp9_switchable_interp_tree,
127                      cm->fc->switchable_interp_prob[j],
128                      counts->switchable_interp[j], SWITCHABLE_FILTERS, w);
129 }
130 
pack_mb_tokens(vpx_writer * w,TOKENEXTRA ** tp,const TOKENEXTRA * const stop,vpx_bit_depth_t bit_depth)131 static void pack_mb_tokens(vpx_writer *w, TOKENEXTRA **tp,
132                            const TOKENEXTRA *const stop,
133                            vpx_bit_depth_t bit_depth) {
134   const TOKENEXTRA *p;
135   const vp9_extra_bit *const extra_bits =
136 #if CONFIG_VP9_HIGHBITDEPTH
137       (bit_depth == VPX_BITS_12)
138           ? vp9_extra_bits_high12
139           : (bit_depth == VPX_BITS_10) ? vp9_extra_bits_high10 : vp9_extra_bits;
140 #else
141       vp9_extra_bits;
142   (void)bit_depth;
143 #endif  // CONFIG_VP9_HIGHBITDEPTH
144 
145   for (p = *tp; p < stop && p->token != EOSB_TOKEN; ++p) {
146     if (p->token == EOB_TOKEN) {
147       vpx_write(w, 0, p->context_tree[0]);
148       continue;
149     }
150     vpx_write(w, 1, p->context_tree[0]);
151     while (p->token == ZERO_TOKEN) {
152       vpx_write(w, 0, p->context_tree[1]);
153       ++p;
154       if (p == stop || p->token == EOSB_TOKEN) {
155         *tp = (TOKENEXTRA *)(uintptr_t)p + (p->token == EOSB_TOKEN);
156         return;
157       }
158     }
159 
160     {
161       const int t = p->token;
162       const vpx_prob *const context_tree = p->context_tree;
163       assert(t != ZERO_TOKEN);
164       assert(t != EOB_TOKEN);
165       assert(t != EOSB_TOKEN);
166       vpx_write(w, 1, context_tree[1]);
167       if (t == ONE_TOKEN) {
168         vpx_write(w, 0, context_tree[2]);
169         vpx_write_bit(w, p->extra & 1);
170       } else {  // t >= TWO_TOKEN && t < EOB_TOKEN
171         const struct vp9_token *const a = &vp9_coef_encodings[t];
172         const int v = a->value;
173         const int n = a->len;
174         const int e = p->extra;
175         vpx_write(w, 1, context_tree[2]);
176         vp9_write_tree(w, vp9_coef_con_tree,
177                        vp9_pareto8_full[context_tree[PIVOT_NODE] - 1], v,
178                        n - UNCONSTRAINED_NODES, 0);
179         if (t >= CATEGORY1_TOKEN) {
180           const vp9_extra_bit *const b = &extra_bits[t];
181           const unsigned char *pb = b->prob;
182           int v = e >> 1;
183           int n = b->len;  // number of bits in v, assumed nonzero
184           do {
185             const int bb = (v >> --n) & 1;
186             vpx_write(w, bb, *pb++);
187           } while (n);
188         }
189         vpx_write_bit(w, e & 1);
190       }
191     }
192   }
193   *tp = (TOKENEXTRA *)(uintptr_t)p + (p->token == EOSB_TOKEN);
194 }
195 
write_segment_id(vpx_writer * w,const struct segmentation * seg,int segment_id)196 static void write_segment_id(vpx_writer *w, const struct segmentation *seg,
197                              int segment_id) {
198   if (seg->enabled && seg->update_map)
199     vp9_write_tree(w, vp9_segment_tree, seg->tree_probs, segment_id, 3, 0);
200 }
201 
202 // This function encodes the reference frame
write_ref_frames(const VP9_COMMON * cm,const MACROBLOCKD * const xd,vpx_writer * w)203 static void write_ref_frames(const VP9_COMMON *cm, const MACROBLOCKD *const xd,
204                              vpx_writer *w) {
205   const MODE_INFO *const mi = xd->mi[0];
206   const int is_compound = has_second_ref(mi);
207   const int segment_id = mi->segment_id;
208 
209   // If segment level coding of this signal is disabled...
210   // or the segment allows multiple reference frame options
211   if (segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
212     assert(!is_compound);
213     assert(mi->ref_frame[0] ==
214            get_segdata(&cm->seg, segment_id, SEG_LVL_REF_FRAME));
215   } else {
216     // does the feature use compound prediction or not
217     // (if not specified at the frame/segment level)
218     if (cm->reference_mode == REFERENCE_MODE_SELECT) {
219       vpx_write(w, is_compound, vp9_get_reference_mode_prob(cm, xd));
220     } else {
221       assert((!is_compound) == (cm->reference_mode == SINGLE_REFERENCE));
222     }
223 
224     if (is_compound) {
225       const int idx = cm->ref_frame_sign_bias[cm->comp_fixed_ref];
226       vpx_write(w, mi->ref_frame[!idx] == cm->comp_var_ref[1],
227                 vp9_get_pred_prob_comp_ref_p(cm, xd));
228     } else {
229       const int bit0 = mi->ref_frame[0] != LAST_FRAME;
230       vpx_write(w, bit0, vp9_get_pred_prob_single_ref_p1(cm, xd));
231       if (bit0) {
232         const int bit1 = mi->ref_frame[0] != GOLDEN_FRAME;
233         vpx_write(w, bit1, vp9_get_pred_prob_single_ref_p2(cm, xd));
234       }
235     }
236   }
237 }
238 
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[MAX_REF_FRAMES][SWITCHABLE])239 static void pack_inter_mode_mvs(
240     VP9_COMP *cpi, const MACROBLOCKD *const xd,
241     const MB_MODE_INFO_EXT *const mbmi_ext, vpx_writer *w,
242     unsigned int *const max_mv_magnitude,
243     int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE]) {
244   VP9_COMMON *const cm = &cpi->common;
245   const nmv_context *nmvc = &cm->fc->nmvc;
246   const struct segmentation *const seg = &cm->seg;
247   const MODE_INFO *const mi = xd->mi[0];
248   const PREDICTION_MODE mode = mi->mode;
249   const int segment_id = mi->segment_id;
250   const BLOCK_SIZE bsize = mi->sb_type;
251   const int allow_hp = cm->allow_high_precision_mv;
252   const int is_inter = is_inter_block(mi);
253   const int is_compound = has_second_ref(mi);
254   int skip, ref;
255 
256   if (seg->update_map) {
257     if (seg->temporal_update) {
258       const int pred_flag = mi->seg_id_predicted;
259       vpx_prob pred_prob = vp9_get_pred_prob_seg_id(seg, xd);
260       vpx_write(w, pred_flag, pred_prob);
261       if (!pred_flag) write_segment_id(w, seg, segment_id);
262     } else {
263       write_segment_id(w, seg, segment_id);
264     }
265   }
266 
267   skip = write_skip(cm, xd, segment_id, mi, w);
268 
269   if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
270     vpx_write(w, is_inter, vp9_get_intra_inter_prob(cm, xd));
271 
272   if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT &&
273       !(is_inter && skip)) {
274     write_selected_tx_size(cm, xd, w);
275   }
276 
277   if (!is_inter) {
278     if (bsize >= BLOCK_8X8) {
279       write_intra_mode(w, mode, cm->fc->y_mode_prob[size_group_lookup[bsize]]);
280     } else {
281       int idx, idy;
282       const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
283       const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
284       for (idy = 0; idy < 2; idy += num_4x4_h) {
285         for (idx = 0; idx < 2; idx += num_4x4_w) {
286           const PREDICTION_MODE b_mode = mi->bmi[idy * 2 + idx].as_mode;
287           write_intra_mode(w, b_mode, cm->fc->y_mode_prob[0]);
288         }
289       }
290     }
291     write_intra_mode(w, mi->uv_mode, cm->fc->uv_mode_prob[mode]);
292   } else {
293     const int mode_ctx = mbmi_ext->mode_context[mi->ref_frame[0]];
294     const vpx_prob *const inter_probs = cm->fc->inter_mode_probs[mode_ctx];
295     write_ref_frames(cm, xd, w);
296 
297     // If segment skip is not enabled code the mode.
298     if (!segfeature_active(seg, segment_id, SEG_LVL_SKIP)) {
299       if (bsize >= BLOCK_8X8) {
300         write_inter_mode(w, mode, inter_probs);
301       }
302     }
303 
304     if (cm->interp_filter == SWITCHABLE) {
305       const int ctx = get_pred_context_switchable_interp(xd);
306       vp9_write_token(w, vp9_switchable_interp_tree,
307                       cm->fc->switchable_interp_prob[ctx],
308                       &switchable_interp_encodings[mi->interp_filter]);
309       ++interp_filter_selected[0][mi->interp_filter];
310     } else {
311       assert(mi->interp_filter == cm->interp_filter);
312     }
313 
314     if (bsize < BLOCK_8X8) {
315       const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
316       const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
317       int idx, idy;
318       for (idy = 0; idy < 2; idy += num_4x4_h) {
319         for (idx = 0; idx < 2; idx += num_4x4_w) {
320           const int j = idy * 2 + idx;
321           const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
322           write_inter_mode(w, b_mode, inter_probs);
323           if (b_mode == NEWMV) {
324             for (ref = 0; ref < 1 + is_compound; ++ref)
325               vp9_encode_mv(cpi, w, &mi->bmi[j].as_mv[ref].as_mv,
326                             &mbmi_ext->ref_mvs[mi->ref_frame[ref]][0].as_mv,
327                             nmvc, allow_hp, max_mv_magnitude);
328           }
329         }
330       }
331     } else {
332       if (mode == NEWMV) {
333         for (ref = 0; ref < 1 + is_compound; ++ref)
334           vp9_encode_mv(cpi, w, &mi->mv[ref].as_mv,
335                         &mbmi_ext->ref_mvs[mi->ref_frame[ref]][0].as_mv, nmvc,
336                         allow_hp, max_mv_magnitude);
337       }
338     }
339   }
340 }
341 
write_mb_modes_kf(const VP9_COMMON * cm,const MACROBLOCKD * xd,vpx_writer * w)342 static void write_mb_modes_kf(const VP9_COMMON *cm, const MACROBLOCKD *xd,
343                               vpx_writer *w) {
344   const struct segmentation *const seg = &cm->seg;
345   const MODE_INFO *const mi = xd->mi[0];
346   const MODE_INFO *const above_mi = xd->above_mi;
347   const MODE_INFO *const left_mi = xd->left_mi;
348   const BLOCK_SIZE bsize = mi->sb_type;
349 
350   if (seg->update_map) write_segment_id(w, seg, mi->segment_id);
351 
352   write_skip(cm, xd, mi->segment_id, mi, w);
353 
354   if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT)
355     write_selected_tx_size(cm, xd, w);
356 
357   if (bsize >= BLOCK_8X8) {
358     write_intra_mode(w, mi->mode, get_y_mode_probs(mi, above_mi, left_mi, 0));
359   } else {
360     const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
361     const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
362     int idx, idy;
363 
364     for (idy = 0; idy < 2; idy += num_4x4_h) {
365       for (idx = 0; idx < 2; idx += num_4x4_w) {
366         const int block = idy * 2 + idx;
367         write_intra_mode(w, mi->bmi[block].as_mode,
368                          get_y_mode_probs(mi, above_mi, left_mi, block));
369       }
370     }
371   }
372 
373   write_intra_mode(w, mi->uv_mode, vp9_kf_uv_mode_prob[mi->mode]);
374 }
375 
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[MAX_REF_FRAMES][SWITCHABLE])376 static void write_modes_b(
377     VP9_COMP *cpi, MACROBLOCKD *const xd, const TileInfo *const tile,
378     vpx_writer *w, TOKENEXTRA **tok, const TOKENEXTRA *const tok_end,
379     int mi_row, int mi_col, unsigned int *const max_mv_magnitude,
380     int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE]) {
381   const VP9_COMMON *const cm = &cpi->common;
382   const MB_MODE_INFO_EXT *const mbmi_ext =
383       cpi->td.mb.mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
384   MODE_INFO *m;
385 
386   xd->mi = cm->mi_grid_visible + (mi_row * cm->mi_stride + mi_col);
387   m = xd->mi[0];
388 
389   set_mi_row_col(xd, tile, mi_row, num_8x8_blocks_high_lookup[m->sb_type],
390                  mi_col, num_8x8_blocks_wide_lookup[m->sb_type], cm->mi_rows,
391                  cm->mi_cols);
392   if (frame_is_intra_only(cm)) {
393     write_mb_modes_kf(cm, xd, w);
394   } else {
395     pack_inter_mode_mvs(cpi, xd, mbmi_ext, w, max_mv_magnitude,
396                         interp_filter_selected);
397   }
398 
399   assert(*tok < tok_end);
400   pack_mb_tokens(w, tok, tok_end, cm->bit_depth);
401 }
402 
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)403 static void write_partition(const VP9_COMMON *const cm,
404                             const MACROBLOCKD *const xd, int hbs, int mi_row,
405                             int mi_col, PARTITION_TYPE p, BLOCK_SIZE bsize,
406                             vpx_writer *w) {
407   const int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
408   const vpx_prob *const probs = xd->partition_probs[ctx];
409   const int has_rows = (mi_row + hbs) < cm->mi_rows;
410   const int has_cols = (mi_col + hbs) < cm->mi_cols;
411 
412   if (has_rows && has_cols) {
413     vp9_write_token(w, vp9_partition_tree, probs, &partition_encodings[p]);
414   } else if (!has_rows && has_cols) {
415     assert(p == PARTITION_SPLIT || p == PARTITION_HORZ);
416     vpx_write(w, p == PARTITION_SPLIT, probs[1]);
417   } else if (has_rows && !has_cols) {
418     assert(p == PARTITION_SPLIT || p == PARTITION_VERT);
419     vpx_write(w, p == PARTITION_SPLIT, probs[2]);
420   } else {
421     assert(p == PARTITION_SPLIT);
422   }
423 }
424 
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[MAX_REF_FRAMES][SWITCHABLE])425 static void write_modes_sb(
426     VP9_COMP *cpi, MACROBLOCKD *const xd, const TileInfo *const tile,
427     vpx_writer *w, TOKENEXTRA **tok, const TOKENEXTRA *const tok_end,
428     int mi_row, int mi_col, BLOCK_SIZE bsize,
429     unsigned int *const max_mv_magnitude,
430     int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE]) {
431   const VP9_COMMON *const cm = &cpi->common;
432   const int bsl = b_width_log2_lookup[bsize];
433   const int bs = (1 << bsl) / 4;
434   PARTITION_TYPE partition;
435   BLOCK_SIZE subsize;
436   const MODE_INFO *m = NULL;
437 
438   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
439 
440   m = cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col];
441 
442   partition = partition_lookup[bsl][m->sb_type];
443   write_partition(cm, xd, bs, mi_row, mi_col, partition, bsize, w);
444   subsize = get_subsize(bsize, partition);
445   if (subsize < BLOCK_8X8) {
446     write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
447                   max_mv_magnitude, interp_filter_selected);
448   } else {
449     switch (partition) {
450       case PARTITION_NONE:
451         write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
452                       max_mv_magnitude, interp_filter_selected);
453         break;
454       case PARTITION_HORZ:
455         write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
456                       max_mv_magnitude, interp_filter_selected);
457         if (mi_row + bs < cm->mi_rows)
458           write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row + bs, mi_col,
459                         max_mv_magnitude, interp_filter_selected);
460         break;
461       case PARTITION_VERT:
462         write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
463                       max_mv_magnitude, interp_filter_selected);
464         if (mi_col + bs < cm->mi_cols)
465           write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col + bs,
466                         max_mv_magnitude, interp_filter_selected);
467         break;
468       default:
469         assert(partition == PARTITION_SPLIT);
470         write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col, subsize,
471                        max_mv_magnitude, interp_filter_selected);
472         write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col + bs,
473                        subsize, max_mv_magnitude, interp_filter_selected);
474         write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row + bs, mi_col,
475                        subsize, max_mv_magnitude, interp_filter_selected);
476         write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row + bs, mi_col + bs,
477                        subsize, max_mv_magnitude, interp_filter_selected);
478         break;
479     }
480   }
481 
482   // update partition context
483   if (bsize >= BLOCK_8X8 &&
484       (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
485     update_partition_context(xd, mi_row, mi_col, subsize, bsize);
486 }
487 
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[MAX_REF_FRAMES][SWITCHABLE])488 static void write_modes(
489     VP9_COMP *cpi, MACROBLOCKD *const xd, const TileInfo *const tile,
490     vpx_writer *w, int tile_row, int tile_col,
491     unsigned int *const max_mv_magnitude,
492     int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE]) {
493   const VP9_COMMON *const cm = &cpi->common;
494   int mi_row, mi_col, tile_sb_row;
495   TOKENEXTRA *tok = NULL;
496   TOKENEXTRA *tok_end = NULL;
497 
498   set_partition_probs(cm, xd);
499 
500   for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
501        mi_row += MI_BLOCK_SIZE) {
502     tile_sb_row = mi_cols_aligned_to_sb(mi_row - tile->mi_row_start) >>
503                   MI_BLOCK_SIZE_LOG2;
504     tok = cpi->tplist[tile_row][tile_col][tile_sb_row].start;
505     tok_end = tok + cpi->tplist[tile_row][tile_col][tile_sb_row].count;
506 
507     vp9_zero(xd->left_seg_context);
508     for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
509          mi_col += MI_BLOCK_SIZE)
510       write_modes_sb(cpi, xd, tile, w, &tok, tok_end, mi_row, mi_col,
511                      BLOCK_64X64, max_mv_magnitude, interp_filter_selected);
512 
513     assert(tok == cpi->tplist[tile_row][tile_col][tile_sb_row].stop);
514   }
515 }
516 
build_tree_distribution(VP9_COMP * cpi,TX_SIZE tx_size,vp9_coeff_stats * coef_branch_ct,vp9_coeff_probs_model * coef_probs)517 static void build_tree_distribution(VP9_COMP *cpi, TX_SIZE tx_size,
518                                     vp9_coeff_stats *coef_branch_ct,
519                                     vp9_coeff_probs_model *coef_probs) {
520   vp9_coeff_count *coef_counts = cpi->td.rd_counts.coef_counts[tx_size];
521   unsigned int(*eob_branch_ct)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS] =
522       cpi->common.counts.eob_branch[tx_size];
523   int i, j, k, l, m;
524 
525   for (i = 0; i < PLANE_TYPES; ++i) {
526     for (j = 0; j < REF_TYPES; ++j) {
527       for (k = 0; k < COEF_BANDS; ++k) {
528         for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
529           vp9_tree_probs_from_distribution(vp9_coef_tree,
530                                            coef_branch_ct[i][j][k][l],
531                                            coef_counts[i][j][k][l]);
532           coef_branch_ct[i][j][k][l][0][1] =
533               eob_branch_ct[i][j][k][l] - coef_branch_ct[i][j][k][l][0][0];
534           for (m = 0; m < UNCONSTRAINED_NODES; ++m)
535             coef_probs[i][j][k][l][m] =
536                 get_binary_prob(coef_branch_ct[i][j][k][l][m][0],
537                                 coef_branch_ct[i][j][k][l][m][1]);
538         }
539       }
540     }
541   }
542 }
543 
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)544 static void update_coef_probs_common(vpx_writer *const bc, VP9_COMP *cpi,
545                                      TX_SIZE tx_size,
546                                      vp9_coeff_stats *frame_branch_ct,
547                                      vp9_coeff_probs_model *new_coef_probs) {
548   vp9_coeff_probs_model *old_coef_probs = cpi->common.fc->coef_probs[tx_size];
549   const vpx_prob upd = DIFF_UPDATE_PROB;
550   const int entropy_nodes_update = UNCONSTRAINED_NODES;
551   int i, j, k, l, t;
552   int stepsize = cpi->sf.coeff_prob_appx_step;
553 
554   switch (cpi->sf.use_fast_coef_updates) {
555     case TWO_LOOP: {
556       /* dry run to see if there is any update at all needed */
557       int savings = 0;
558       int update[2] = { 0, 0 };
559       for (i = 0; i < PLANE_TYPES; ++i) {
560         for (j = 0; j < REF_TYPES; ++j) {
561           for (k = 0; k < COEF_BANDS; ++k) {
562             for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
563               for (t = 0; t < entropy_nodes_update; ++t) {
564                 vpx_prob newp = new_coef_probs[i][j][k][l][t];
565                 const vpx_prob oldp = old_coef_probs[i][j][k][l][t];
566                 int s;
567                 int u = 0;
568                 if (t == PIVOT_NODE)
569                   s = vp9_prob_diff_update_savings_search_model(
570                       frame_branch_ct[i][j][k][l][0], oldp, &newp, upd,
571                       stepsize);
572                 else
573                   s = vp9_prob_diff_update_savings_search(
574                       frame_branch_ct[i][j][k][l][t], oldp, &newp, upd);
575                 if (s > 0 && newp != oldp) u = 1;
576                 if (u)
577                   savings += s - (int)(vp9_cost_zero(upd));
578                 else
579                   savings -= (int)(vp9_cost_zero(upd));
580                 update[u]++;
581               }
582             }
583           }
584         }
585       }
586 
587       // printf("Update %d %d, savings %d\n", update[0], update[1], savings);
588       /* Is coef updated at all */
589       if (update[1] == 0 || savings < 0) {
590         vpx_write_bit(bc, 0);
591         return;
592       }
593       vpx_write_bit(bc, 1);
594       for (i = 0; i < PLANE_TYPES; ++i) {
595         for (j = 0; j < REF_TYPES; ++j) {
596           for (k = 0; k < COEF_BANDS; ++k) {
597             for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
598               // calc probs and branch cts for this frame only
599               for (t = 0; t < entropy_nodes_update; ++t) {
600                 vpx_prob newp = new_coef_probs[i][j][k][l][t];
601                 vpx_prob *oldp = old_coef_probs[i][j][k][l] + t;
602                 const vpx_prob upd = DIFF_UPDATE_PROB;
603                 int 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                 int 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 (vp9_preserve_existing_gf(cpi)) {
904     // We have decided to preserve the previously existing golden frame as our
905     // new ARF frame. However, in the short term we leave it in the GF slot and,
906     // if we're updating the GF with the current decoded frame, we save it
907     // instead to the ARF slot.
908     // Later, in the function vp9_encoder.c:vp9_update_reference_frames() we
909     // will swap gld_fb_idx and alt_fb_idx to achieve our objective. We do it
910     // there so that it can be done outside of the recode loop.
911     // Note: This is highly specific to the use of ARF as a forward reference,
912     // and this needs to be generalized as other uses are implemented
913     // (like RTC/temporal scalability).
914     return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
915            (cpi->refresh_golden_frame << cpi->alt_fb_idx);
916   } else {
917     int arf_idx = cpi->alt_fb_idx;
918     GF_GROUP *const gf_group = &cpi->twopass.gf_group;
919 
920     if (cpi->multi_layer_arf) {
921       for (arf_idx = 0; arf_idx < REF_FRAMES; ++arf_idx) {
922         if (arf_idx != cpi->alt_fb_idx && arf_idx != cpi->lst_fb_idx &&
923             arf_idx != cpi->gld_fb_idx) {
924           int idx;
925           for (idx = 0; idx < gf_group->stack_size; ++idx)
926             if (arf_idx == gf_group->arf_index_stack[idx]) break;
927           if (idx == gf_group->stack_size) break;
928         }
929       }
930     }
931     cpi->twopass.gf_group.top_arf_idx = arf_idx;
932 
933     if (cpi->use_svc && cpi->svc.use_set_ref_frame_config &&
934         cpi->svc.temporal_layering_mode == VP9E_TEMPORAL_LAYERING_MODE_BYPASS)
935       return cpi->svc.update_buffer_slot[cpi->svc.spatial_layer_id];
936     return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
937            (cpi->refresh_golden_frame << cpi->gld_fb_idx) |
938            (cpi->refresh_alt_ref_frame << arf_idx);
939   }
940 }
941 
encode_tile_worker(void * arg1,void * arg2)942 static int encode_tile_worker(void *arg1, void *arg2) {
943   VP9_COMP *cpi = (VP9_COMP *)arg1;
944   VP9BitstreamWorkerData *data = (VP9BitstreamWorkerData *)arg2;
945   MACROBLOCKD *const xd = &data->xd;
946   const int tile_row = 0;
947   vpx_start_encode(&data->bit_writer, data->dest);
948   write_modes(cpi, xd, &cpi->tile_data[data->tile_idx].tile_info,
949               &data->bit_writer, tile_row, data->tile_idx,
950               &data->max_mv_magnitude, data->interp_filter_selected);
951   vpx_stop_encode(&data->bit_writer);
952   return 1;
953 }
954 
vp9_bitstream_encode_tiles_buffer_dealloc(VP9_COMP * const cpi)955 void vp9_bitstream_encode_tiles_buffer_dealloc(VP9_COMP *const cpi) {
956   if (cpi->vp9_bitstream_worker_data) {
957     int i;
958     for (i = 1; i < cpi->num_workers; ++i) {
959       vpx_free(cpi->vp9_bitstream_worker_data[i].dest);
960     }
961     vpx_free(cpi->vp9_bitstream_worker_data);
962     cpi->vp9_bitstream_worker_data = NULL;
963   }
964 }
965 
encode_tiles_buffer_alloc(VP9_COMP * const cpi)966 static int encode_tiles_buffer_alloc(VP9_COMP *const cpi) {
967   int i;
968   const size_t worker_data_size =
969       cpi->num_workers * sizeof(*cpi->vp9_bitstream_worker_data);
970   cpi->vp9_bitstream_worker_data = vpx_memalign(16, worker_data_size);
971   memset(cpi->vp9_bitstream_worker_data, 0, worker_data_size);
972   if (!cpi->vp9_bitstream_worker_data) return 1;
973   for (i = 1; i < cpi->num_workers; ++i) {
974     cpi->vp9_bitstream_worker_data[i].dest_size =
975         cpi->oxcf.width * cpi->oxcf.height;
976     cpi->vp9_bitstream_worker_data[i].dest =
977         vpx_malloc(cpi->vp9_bitstream_worker_data[i].dest_size);
978     if (!cpi->vp9_bitstream_worker_data[i].dest) return 1;
979   }
980   return 0;
981 }
982 
encode_tiles_mt(VP9_COMP * cpi,uint8_t * data_ptr)983 static size_t encode_tiles_mt(VP9_COMP *cpi, uint8_t *data_ptr) {
984   const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
985   VP9_COMMON *const cm = &cpi->common;
986   const int tile_cols = 1 << cm->log2_tile_cols;
987   const int num_workers = cpi->num_workers;
988   size_t total_size = 0;
989   int tile_col = 0;
990 
991   if (!cpi->vp9_bitstream_worker_data ||
992       cpi->vp9_bitstream_worker_data[1].dest_size >
993           (cpi->oxcf.width * cpi->oxcf.height)) {
994     vp9_bitstream_encode_tiles_buffer_dealloc(cpi);
995     if (encode_tiles_buffer_alloc(cpi)) return 0;
996   }
997 
998   while (tile_col < tile_cols) {
999     int i, j;
1000     for (i = 0; i < num_workers && tile_col < tile_cols; ++i) {
1001       VPxWorker *const worker = &cpi->workers[i];
1002       VP9BitstreamWorkerData *const data = &cpi->vp9_bitstream_worker_data[i];
1003 
1004       // Populate the worker data.
1005       data->xd = cpi->td.mb.e_mbd;
1006       data->tile_idx = tile_col;
1007       data->max_mv_magnitude = cpi->max_mv_magnitude;
1008       memset(data->interp_filter_selected, 0,
1009              sizeof(data->interp_filter_selected[0][0]) * SWITCHABLE);
1010 
1011       // First thread can directly write into the output buffer.
1012       if (i == 0) {
1013         // If this worker happens to be for the last tile, then do not offset it
1014         // by 4 for the tile size.
1015         data->dest =
1016             data_ptr + total_size + (tile_col == tile_cols - 1 ? 0 : 4);
1017       }
1018       worker->data1 = cpi;
1019       worker->data2 = data;
1020       worker->hook = encode_tile_worker;
1021       worker->had_error = 0;
1022 
1023       if (i < num_workers - 1) {
1024         winterface->launch(worker);
1025       } else {
1026         winterface->execute(worker);
1027       }
1028       ++tile_col;
1029     }
1030     for (j = 0; j < i; ++j) {
1031       VPxWorker *const worker = &cpi->workers[j];
1032       VP9BitstreamWorkerData *const data =
1033           (VP9BitstreamWorkerData *)worker->data2;
1034       uint32_t tile_size;
1035       int k;
1036 
1037       if (!winterface->sync(worker)) return 0;
1038       tile_size = data->bit_writer.pos;
1039 
1040       // Aggregate per-thread bitstream stats.
1041       cpi->max_mv_magnitude =
1042           VPXMAX(cpi->max_mv_magnitude, data->max_mv_magnitude);
1043       for (k = 0; k < SWITCHABLE; ++k) {
1044         cpi->interp_filter_selected[0][k] += data->interp_filter_selected[0][k];
1045       }
1046 
1047       // Prefix the size of the tile on all but the last.
1048       if (tile_col != tile_cols || j < i - 1) {
1049         mem_put_be32(data_ptr + total_size, tile_size);
1050         total_size += 4;
1051       }
1052       if (j > 0) {
1053         memcpy(data_ptr + total_size, data->dest, tile_size);
1054       }
1055       total_size += tile_size;
1056     }
1057   }
1058   return total_size;
1059 }
1060 
encode_tiles(VP9_COMP * cpi,uint8_t * data_ptr)1061 static size_t encode_tiles(VP9_COMP *cpi, uint8_t *data_ptr) {
1062   VP9_COMMON *const cm = &cpi->common;
1063   MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
1064   vpx_writer residual_bc;
1065   int tile_row, tile_col;
1066   size_t total_size = 0;
1067   const int tile_cols = 1 << cm->log2_tile_cols;
1068   const int tile_rows = 1 << cm->log2_tile_rows;
1069 
1070   memset(cm->above_seg_context, 0,
1071          sizeof(*cm->above_seg_context) * mi_cols_aligned_to_sb(cm->mi_cols));
1072 
1073   // Encoding tiles in parallel is done only for realtime mode now. In other
1074   // modes the speed up is insignificant and requires further testing to ensure
1075   // that it does not make the overall process worse in any case.
1076   if (cpi->oxcf.mode == REALTIME && cpi->num_workers > 1 && tile_rows == 1 &&
1077       tile_cols > 1) {
1078     return encode_tiles_mt(cpi, data_ptr);
1079   }
1080 
1081   for (tile_row = 0; tile_row < tile_rows; tile_row++) {
1082     for (tile_col = 0; tile_col < tile_cols; tile_col++) {
1083       int tile_idx = tile_row * tile_cols + tile_col;
1084 
1085       if (tile_col < tile_cols - 1 || tile_row < tile_rows - 1)
1086         vpx_start_encode(&residual_bc, data_ptr + total_size + 4);
1087       else
1088         vpx_start_encode(&residual_bc, data_ptr + total_size);
1089 
1090       write_modes(cpi, xd, &cpi->tile_data[tile_idx].tile_info, &residual_bc,
1091                   tile_row, tile_col, &cpi->max_mv_magnitude,
1092                   cpi->interp_filter_selected);
1093 
1094       vpx_stop_encode(&residual_bc);
1095       if (tile_col < tile_cols - 1 || tile_row < tile_rows - 1) {
1096         // size of this tile
1097         mem_put_be32(data_ptr + total_size, residual_bc.pos);
1098         total_size += 4;
1099       }
1100 
1101       total_size += residual_bc.pos;
1102     }
1103   }
1104   return total_size;
1105 }
1106 
write_render_size(const VP9_COMMON * cm,struct vpx_write_bit_buffer * wb)1107 static void write_render_size(const VP9_COMMON *cm,
1108                               struct vpx_write_bit_buffer *wb) {
1109   const int scaling_active =
1110       cm->width != cm->render_width || cm->height != cm->render_height;
1111   vpx_wb_write_bit(wb, scaling_active);
1112   if (scaling_active) {
1113     vpx_wb_write_literal(wb, cm->render_width - 1, 16);
1114     vpx_wb_write_literal(wb, cm->render_height - 1, 16);
1115   }
1116 }
1117 
write_frame_size(const VP9_COMMON * cm,struct vpx_write_bit_buffer * wb)1118 static void write_frame_size(const VP9_COMMON *cm,
1119                              struct vpx_write_bit_buffer *wb) {
1120   vpx_wb_write_literal(wb, cm->width - 1, 16);
1121   vpx_wb_write_literal(wb, cm->height - 1, 16);
1122 
1123   write_render_size(cm, wb);
1124 }
1125 
write_frame_size_with_refs(VP9_COMP * cpi,struct vpx_write_bit_buffer * wb)1126 static void write_frame_size_with_refs(VP9_COMP *cpi,
1127                                        struct vpx_write_bit_buffer *wb) {
1128   VP9_COMMON *const cm = &cpi->common;
1129   int found = 0;
1130 
1131   MV_REFERENCE_FRAME ref_frame;
1132   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
1133     YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi, ref_frame);
1134 
1135     // Set "found" to 0 for temporal svc and for spatial svc key frame
1136     if (cpi->use_svc &&
1137         ((cpi->svc.number_temporal_layers > 1 &&
1138           cpi->oxcf.rc_mode == VPX_CBR) ||
1139          (cpi->svc.number_spatial_layers > 1 &&
1140           cpi->svc.layer_context[cpi->svc.spatial_layer_id].is_key_frame))) {
1141       found = 0;
1142     } else if (cfg != NULL) {
1143       found =
1144           cm->width == cfg->y_crop_width && cm->height == cfg->y_crop_height;
1145     }
1146     vpx_wb_write_bit(wb, found);
1147     if (found) {
1148       break;
1149     }
1150   }
1151 
1152   if (!found) {
1153     vpx_wb_write_literal(wb, cm->width - 1, 16);
1154     vpx_wb_write_literal(wb, cm->height - 1, 16);
1155   }
1156 
1157   write_render_size(cm, wb);
1158 }
1159 
write_sync_code(struct vpx_write_bit_buffer * wb)1160 static void write_sync_code(struct vpx_write_bit_buffer *wb) {
1161   vpx_wb_write_literal(wb, VP9_SYNC_CODE_0, 8);
1162   vpx_wb_write_literal(wb, VP9_SYNC_CODE_1, 8);
1163   vpx_wb_write_literal(wb, VP9_SYNC_CODE_2, 8);
1164 }
1165 
write_profile(BITSTREAM_PROFILE profile,struct vpx_write_bit_buffer * wb)1166 static void write_profile(BITSTREAM_PROFILE profile,
1167                           struct vpx_write_bit_buffer *wb) {
1168   switch (profile) {
1169     case PROFILE_0: vpx_wb_write_literal(wb, 0, 2); break;
1170     case PROFILE_1: vpx_wb_write_literal(wb, 2, 2); break;
1171     case PROFILE_2: vpx_wb_write_literal(wb, 1, 2); break;
1172     default:
1173       assert(profile == PROFILE_3);
1174       vpx_wb_write_literal(wb, 6, 3);
1175       break;
1176   }
1177 }
1178 
write_bitdepth_colorspace_sampling(VP9_COMMON * const cm,struct vpx_write_bit_buffer * wb)1179 static void write_bitdepth_colorspace_sampling(
1180     VP9_COMMON *const cm, struct vpx_write_bit_buffer *wb) {
1181   if (cm->profile >= PROFILE_2) {
1182     assert(cm->bit_depth > VPX_BITS_8);
1183     vpx_wb_write_bit(wb, cm->bit_depth == VPX_BITS_10 ? 0 : 1);
1184   }
1185   vpx_wb_write_literal(wb, cm->color_space, 3);
1186   if (cm->color_space != VPX_CS_SRGB) {
1187     // 0: [16, 235] (i.e. xvYCC), 1: [0, 255]
1188     vpx_wb_write_bit(wb, cm->color_range);
1189     if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
1190       assert(cm->subsampling_x != 1 || cm->subsampling_y != 1);
1191       vpx_wb_write_bit(wb, cm->subsampling_x);
1192       vpx_wb_write_bit(wb, cm->subsampling_y);
1193       vpx_wb_write_bit(wb, 0);  // unused
1194     } else {
1195       assert(cm->subsampling_x == 1 && cm->subsampling_y == 1);
1196     }
1197   } else {
1198     assert(cm->profile == PROFILE_1 || cm->profile == PROFILE_3);
1199     vpx_wb_write_bit(wb, 0);  // unused
1200   }
1201 }
1202 
write_uncompressed_header(VP9_COMP * cpi,struct vpx_write_bit_buffer * wb)1203 static void write_uncompressed_header(VP9_COMP *cpi,
1204                                       struct vpx_write_bit_buffer *wb) {
1205   VP9_COMMON *const cm = &cpi->common;
1206   MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
1207 
1208   vpx_wb_write_literal(wb, VP9_FRAME_MARKER, 2);
1209 
1210   write_profile(cm->profile, wb);
1211 
1212   // If to use show existing frame.
1213   vpx_wb_write_bit(wb, cm->show_existing_frame);
1214   if (cm->show_existing_frame) {
1215     vpx_wb_write_literal(wb, cpi->alt_fb_idx, 3);
1216     return;
1217   }
1218 
1219   vpx_wb_write_bit(wb, cm->frame_type);
1220   vpx_wb_write_bit(wb, cm->show_frame);
1221   vpx_wb_write_bit(wb, cm->error_resilient_mode);
1222 
1223   if (cm->frame_type == KEY_FRAME) {
1224     write_sync_code(wb);
1225     write_bitdepth_colorspace_sampling(cm, wb);
1226     write_frame_size(cm, wb);
1227   } else {
1228     if (!cm->show_frame) vpx_wb_write_bit(wb, cm->intra_only);
1229 
1230     if (!cm->error_resilient_mode)
1231       vpx_wb_write_literal(wb, cm->reset_frame_context, 2);
1232 
1233     if (cm->intra_only) {
1234       write_sync_code(wb);
1235 
1236       // Note for profile 0, 420 8bpp is assumed.
1237       if (cm->profile > PROFILE_0) {
1238         write_bitdepth_colorspace_sampling(cm, wb);
1239       }
1240 
1241       vpx_wb_write_literal(wb, vp9_get_refresh_mask(cpi), REF_FRAMES);
1242       write_frame_size(cm, wb);
1243     } else {
1244       MV_REFERENCE_FRAME ref_frame;
1245       vpx_wb_write_literal(wb, vp9_get_refresh_mask(cpi), REF_FRAMES);
1246       for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
1247         assert(get_ref_frame_map_idx(cpi, ref_frame) != INVALID_IDX);
1248         vpx_wb_write_literal(wb, get_ref_frame_map_idx(cpi, ref_frame),
1249                              REF_FRAMES_LOG2);
1250         vpx_wb_write_bit(wb, cm->ref_frame_sign_bias[ref_frame]);
1251       }
1252 
1253       write_frame_size_with_refs(cpi, wb);
1254 
1255       vpx_wb_write_bit(wb, cm->allow_high_precision_mv);
1256 
1257       fix_interp_filter(cm, cpi->td.counts);
1258       write_interp_filter(cm->interp_filter, wb);
1259     }
1260   }
1261 
1262   if (!cm->error_resilient_mode) {
1263     vpx_wb_write_bit(wb, cm->refresh_frame_context);
1264     vpx_wb_write_bit(wb, cm->frame_parallel_decoding_mode);
1265   }
1266 
1267   vpx_wb_write_literal(wb, cm->frame_context_idx, FRAME_CONTEXTS_LOG2);
1268 
1269   encode_loopfilter(&cm->lf, wb);
1270   encode_quantization(cm, wb);
1271   encode_segmentation(cm, xd, wb);
1272 
1273   write_tile_info(cm, wb);
1274 }
1275 
write_compressed_header(VP9_COMP * cpi,uint8_t * data)1276 static size_t write_compressed_header(VP9_COMP *cpi, uint8_t *data) {
1277   VP9_COMMON *const cm = &cpi->common;
1278   MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
1279   FRAME_CONTEXT *const fc = cm->fc;
1280   FRAME_COUNTS *counts = cpi->td.counts;
1281   vpx_writer header_bc;
1282 
1283   vpx_start_encode(&header_bc, data);
1284 
1285   if (xd->lossless)
1286     cm->tx_mode = ONLY_4X4;
1287   else
1288     encode_txfm_probs(cm, &header_bc, counts);
1289 
1290   update_coef_probs(cpi, &header_bc);
1291   update_skip_probs(cm, &header_bc, counts);
1292 
1293   if (!frame_is_intra_only(cm)) {
1294     int i;
1295 
1296     for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
1297       prob_diff_update(vp9_inter_mode_tree, cm->fc->inter_mode_probs[i],
1298                        counts->inter_mode[i], INTER_MODES, &header_bc);
1299 
1300     if (cm->interp_filter == SWITCHABLE)
1301       update_switchable_interp_probs(cm, &header_bc, counts);
1302 
1303     for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
1304       vp9_cond_prob_diff_update(&header_bc, &fc->intra_inter_prob[i],
1305                                 counts->intra_inter[i]);
1306 
1307     if (cpi->allow_comp_inter_inter) {
1308       const int use_compound_pred = cm->reference_mode != SINGLE_REFERENCE;
1309       const int use_hybrid_pred = cm->reference_mode == REFERENCE_MODE_SELECT;
1310 
1311       vpx_write_bit(&header_bc, use_compound_pred);
1312       if (use_compound_pred) {
1313         vpx_write_bit(&header_bc, use_hybrid_pred);
1314         if (use_hybrid_pred)
1315           for (i = 0; i < COMP_INTER_CONTEXTS; i++)
1316             vp9_cond_prob_diff_update(&header_bc, &fc->comp_inter_prob[i],
1317                                       counts->comp_inter[i]);
1318       }
1319     }
1320 
1321     if (cm->reference_mode != COMPOUND_REFERENCE) {
1322       for (i = 0; i < REF_CONTEXTS; i++) {
1323         vp9_cond_prob_diff_update(&header_bc, &fc->single_ref_prob[i][0],
1324                                   counts->single_ref[i][0]);
1325         vp9_cond_prob_diff_update(&header_bc, &fc->single_ref_prob[i][1],
1326                                   counts->single_ref[i][1]);
1327       }
1328     }
1329 
1330     if (cm->reference_mode != SINGLE_REFERENCE)
1331       for (i = 0; i < REF_CONTEXTS; i++)
1332         vp9_cond_prob_diff_update(&header_bc, &fc->comp_ref_prob[i],
1333                                   counts->comp_ref[i]);
1334 
1335     for (i = 0; i < BLOCK_SIZE_GROUPS; ++i)
1336       prob_diff_update(vp9_intra_mode_tree, cm->fc->y_mode_prob[i],
1337                        counts->y_mode[i], INTRA_MODES, &header_bc);
1338 
1339     for (i = 0; i < PARTITION_CONTEXTS; ++i)
1340       prob_diff_update(vp9_partition_tree, fc->partition_prob[i],
1341                        counts->partition[i], PARTITION_TYPES, &header_bc);
1342 
1343     vp9_write_nmv_probs(cm, cm->allow_high_precision_mv, &header_bc,
1344                         &counts->mv);
1345   }
1346 
1347   vpx_stop_encode(&header_bc);
1348   assert(header_bc.pos <= 0xffff);
1349 
1350   return header_bc.pos;
1351 }
1352 
vp9_pack_bitstream(VP9_COMP * cpi,uint8_t * dest,size_t * size)1353 void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, size_t *size) {
1354   uint8_t *data = dest;
1355   size_t first_part_size, uncompressed_hdr_size;
1356   struct vpx_write_bit_buffer wb = { data, 0 };
1357   struct vpx_write_bit_buffer saved_wb;
1358 
1359 #if CONFIG_BITSTREAM_DEBUG
1360   bitstream_queue_reset_write();
1361 #endif
1362 
1363   write_uncompressed_header(cpi, &wb);
1364 
1365   // Skip the rest coding process if use show existing frame.
1366   if (cpi->common.show_existing_frame) {
1367     uncompressed_hdr_size = vpx_wb_bytes_written(&wb);
1368     data += uncompressed_hdr_size;
1369     *size = data - dest;
1370     return;
1371   }
1372 
1373   saved_wb = wb;
1374   vpx_wb_write_literal(&wb, 0, 16);  // don't know in advance first part. size
1375 
1376   uncompressed_hdr_size = vpx_wb_bytes_written(&wb);
1377   data += uncompressed_hdr_size;
1378 
1379   vpx_clear_system_state();
1380 
1381   first_part_size = write_compressed_header(cpi, data);
1382   data += first_part_size;
1383   // TODO(jbb): Figure out what to do if first_part_size > 16 bits.
1384   vpx_wb_write_literal(&saved_wb, (int)first_part_size, 16);
1385 
1386   data += encode_tiles(cpi, data);
1387 
1388   *size = data - dest;
1389 }
1390