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
22#include "vp9/common/vp9_entropy.h"
23#include "vp9/common/vp9_entropymode.h"
24#include "vp9/common/vp9_entropymv.h"
25#include "vp9/common/vp9_mvref_common.h"
26#include "vp9/common/vp9_pred_common.h"
27#include "vp9/common/vp9_seg_common.h"
28#include "vp9/common/vp9_tile_common.h"
29
30#include "vp9/encoder/vp9_cost.h"
31#include "vp9/encoder/vp9_bitstream.h"
32#include "vp9/encoder/vp9_encodemv.h"
33#include "vp9/encoder/vp9_mcomp.h"
34#include "vp9/encoder/vp9_segmentation.h"
35#include "vp9/encoder/vp9_subexp.h"
36#include "vp9/encoder/vp9_tokenize.h"
37
38static const struct vp9_token intra_mode_encodings[INTRA_MODES] = {
39  { 0, 1 },  { 6, 3 },   { 28, 5 },  { 30, 5 }, { 58, 6 },
40  { 59, 6 }, { 126, 7 }, { 127, 7 }, { 62, 6 }, { 2, 2 }
41};
42static const struct vp9_token switchable_interp_encodings[SWITCHABLE_FILTERS] =
43    { { 0, 1 }, { 2, 2 }, { 3, 2 } };
44static const struct vp9_token partition_encodings[PARTITION_TYPES] = {
45  { 0, 1 }, { 2, 2 }, { 6, 3 }, { 7, 3 }
46};
47static const struct vp9_token inter_mode_encodings[INTER_MODES] = {
48  { 2, 2 }, { 6, 3 }, { 0, 1 }, { 7, 3 }
49};
50
51static void write_intra_mode(vpx_writer *w, PREDICTION_MODE mode,
52                             const vpx_prob *probs) {
53  vp9_write_token(w, vp9_intra_mode_tree, probs, &intra_mode_encodings[mode]);
54}
55
56static void write_inter_mode(vpx_writer *w, PREDICTION_MODE mode,
57                             const vpx_prob *probs) {
58  assert(is_inter_mode(mode));
59  vp9_write_token(w, vp9_inter_mode_tree, probs,
60                  &inter_mode_encodings[INTER_OFFSET(mode)]);
61}
62
63static void encode_unsigned_max(struct vpx_write_bit_buffer *wb, int data,
64                                int max) {
65  vpx_wb_write_literal(wb, data, get_unsigned_bits(max));
66}
67
68static void prob_diff_update(const vpx_tree_index *tree,
69                             vpx_prob probs[/*n - 1*/],
70                             const unsigned int counts[/*n - 1*/], int n,
71                             vpx_writer *w) {
72  int i;
73  unsigned int branch_ct[32][2];
74
75  // Assuming max number of probabilities <= 32
76  assert(n <= 32);
77
78  vp9_tree_probs_from_distribution(tree, branch_ct, counts);
79  for (i = 0; i < n - 1; ++i)
80    vp9_cond_prob_diff_update(w, &probs[i], branch_ct[i]);
81}
82
83static void write_selected_tx_size(const VP9_COMMON *cm,
84                                   const MACROBLOCKD *const xd, vpx_writer *w) {
85  TX_SIZE tx_size = xd->mi[0]->tx_size;
86  BLOCK_SIZE bsize = xd->mi[0]->sb_type;
87  const TX_SIZE max_tx_size = max_txsize_lookup[bsize];
88  const vpx_prob *const tx_probs =
89      get_tx_probs2(max_tx_size, xd, &cm->fc->tx_probs);
90  vpx_write(w, tx_size != TX_4X4, tx_probs[0]);
91  if (tx_size != TX_4X4 && max_tx_size >= TX_16X16) {
92    vpx_write(w, tx_size != TX_8X8, tx_probs[1]);
93    if (tx_size != TX_8X8 && max_tx_size >= TX_32X32)
94      vpx_write(w, tx_size != TX_16X16, tx_probs[2]);
95  }
96}
97
98static int write_skip(const VP9_COMMON *cm, const MACROBLOCKD *const xd,
99                      int segment_id, const MODE_INFO *mi, vpx_writer *w) {
100  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
101    return 1;
102  } else {
103    const int skip = mi->skip;
104    vpx_write(w, skip, vp9_get_skip_prob(cm, xd));
105    return skip;
106  }
107}
108
109static void update_skip_probs(VP9_COMMON *cm, vpx_writer *w,
110                              FRAME_COUNTS *counts) {
111  int k;
112
113  for (k = 0; k < SKIP_CONTEXTS; ++k)
114    vp9_cond_prob_diff_update(w, &cm->fc->skip_probs[k], counts->skip[k]);
115}
116
117static void update_switchable_interp_probs(VP9_COMMON *cm, vpx_writer *w,
118                                           FRAME_COUNTS *counts) {
119  int j;
120  for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
121    prob_diff_update(vp9_switchable_interp_tree,
122                     cm->fc->switchable_interp_prob[j],
123                     counts->switchable_interp[j], SWITCHABLE_FILTERS, w);
124}
125
126static void pack_mb_tokens(vpx_writer *w, TOKENEXTRA **tp,
127                           const TOKENEXTRA *const stop,
128                           vpx_bit_depth_t bit_depth) {
129  const TOKENEXTRA *p;
130  const vp9_extra_bit *const extra_bits =
131#if CONFIG_VP9_HIGHBITDEPTH
132      (bit_depth == VPX_BITS_12)
133          ? vp9_extra_bits_high12
134          : (bit_depth == VPX_BITS_10) ? vp9_extra_bits_high10 : vp9_extra_bits;
135#else
136      vp9_extra_bits;
137  (void)bit_depth;
138#endif  // CONFIG_VP9_HIGHBITDEPTH
139
140  for (p = *tp; p < stop && p->token != EOSB_TOKEN; ++p) {
141    if (p->token == EOB_TOKEN) {
142      vpx_write(w, 0, p->context_tree[0]);
143      continue;
144    }
145    vpx_write(w, 1, p->context_tree[0]);
146    while (p->token == ZERO_TOKEN) {
147      vpx_write(w, 0, p->context_tree[1]);
148      ++p;
149      if (p == stop || p->token == EOSB_TOKEN) {
150        *tp = (TOKENEXTRA *)(uintptr_t)p + (p->token == EOSB_TOKEN);
151        return;
152      }
153    }
154
155    {
156      const int t = p->token;
157      const vpx_prob *const context_tree = p->context_tree;
158      assert(t != ZERO_TOKEN);
159      assert(t != EOB_TOKEN);
160      assert(t != EOSB_TOKEN);
161      vpx_write(w, 1, context_tree[1]);
162      if (t == ONE_TOKEN) {
163        vpx_write(w, 0, context_tree[2]);
164        vpx_write_bit(w, p->extra & 1);
165      } else {  // t >= TWO_TOKEN && t < EOB_TOKEN
166        const struct vp9_token *const a = &vp9_coef_encodings[t];
167        const int v = a->value;
168        const int n = a->len;
169        const int e = p->extra;
170        vpx_write(w, 1, context_tree[2]);
171        vp9_write_tree(w, vp9_coef_con_tree,
172                       vp9_pareto8_full[context_tree[PIVOT_NODE] - 1], v,
173                       n - UNCONSTRAINED_NODES, 0);
174        if (t >= CATEGORY1_TOKEN) {
175          const vp9_extra_bit *const b = &extra_bits[t];
176          const unsigned char *pb = b->prob;
177          int v = e >> 1;
178          int n = b->len;  // number of bits in v, assumed nonzero
179          do {
180            const int bb = (v >> --n) & 1;
181            vpx_write(w, bb, *pb++);
182          } while (n);
183        }
184        vpx_write_bit(w, e & 1);
185      }
186    }
187  }
188  *tp = (TOKENEXTRA *)(uintptr_t)p + (p->token == EOSB_TOKEN);
189}
190
191static void write_segment_id(vpx_writer *w, const struct segmentation *seg,
192                             int segment_id) {
193  if (seg->enabled && seg->update_map)
194    vp9_write_tree(w, vp9_segment_tree, seg->tree_probs, segment_id, 3, 0);
195}
196
197// This function encodes the reference frame
198static void write_ref_frames(const VP9_COMMON *cm, const MACROBLOCKD *const xd,
199                             vpx_writer *w) {
200  const MODE_INFO *const mi = xd->mi[0];
201  const int is_compound = has_second_ref(mi);
202  const int segment_id = mi->segment_id;
203
204  // If segment level coding of this signal is disabled...
205  // or the segment allows multiple reference frame options
206  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
207    assert(!is_compound);
208    assert(mi->ref_frame[0] ==
209           get_segdata(&cm->seg, segment_id, SEG_LVL_REF_FRAME));
210  } else {
211    // does the feature use compound prediction or not
212    // (if not specified at the frame/segment level)
213    if (cm->reference_mode == REFERENCE_MODE_SELECT) {
214      vpx_write(w, is_compound, vp9_get_reference_mode_prob(cm, xd));
215    } else {
216      assert((!is_compound) == (cm->reference_mode == SINGLE_REFERENCE));
217    }
218
219    if (is_compound) {
220      vpx_write(w, mi->ref_frame[0] == GOLDEN_FRAME,
221                vp9_get_pred_prob_comp_ref_p(cm, xd));
222    } else {
223      const int bit0 = mi->ref_frame[0] != LAST_FRAME;
224      vpx_write(w, bit0, vp9_get_pred_prob_single_ref_p1(cm, xd));
225      if (bit0) {
226        const int bit1 = mi->ref_frame[0] != GOLDEN_FRAME;
227        vpx_write(w, bit1, vp9_get_pred_prob_single_ref_p2(cm, xd));
228      }
229    }
230  }
231}
232
233static void pack_inter_mode_mvs(
234    VP9_COMP *cpi, const MACROBLOCKD *const xd,
235    const MB_MODE_INFO_EXT *const mbmi_ext, vpx_writer *w,
236    unsigned int *const max_mv_magnitude,
237    int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE]) {
238  VP9_COMMON *const cm = &cpi->common;
239  const nmv_context *nmvc = &cm->fc->nmvc;
240  const struct segmentation *const seg = &cm->seg;
241  const MODE_INFO *const mi = xd->mi[0];
242  const PREDICTION_MODE mode = mi->mode;
243  const int segment_id = mi->segment_id;
244  const BLOCK_SIZE bsize = mi->sb_type;
245  const int allow_hp = cm->allow_high_precision_mv;
246  const int is_inter = is_inter_block(mi);
247  const int is_compound = has_second_ref(mi);
248  int skip, ref;
249
250  if (seg->update_map) {
251    if (seg->temporal_update) {
252      const int pred_flag = mi->seg_id_predicted;
253      vpx_prob pred_prob = vp9_get_pred_prob_seg_id(seg, xd);
254      vpx_write(w, pred_flag, pred_prob);
255      if (!pred_flag) write_segment_id(w, seg, segment_id);
256    } else {
257      write_segment_id(w, seg, segment_id);
258    }
259  }
260
261  skip = write_skip(cm, xd, segment_id, mi, w);
262
263  if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
264    vpx_write(w, is_inter, vp9_get_intra_inter_prob(cm, xd));
265
266  if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT &&
267      !(is_inter && skip)) {
268    write_selected_tx_size(cm, xd, w);
269  }
270
271  if (!is_inter) {
272    if (bsize >= BLOCK_8X8) {
273      write_intra_mode(w, mode, cm->fc->y_mode_prob[size_group_lookup[bsize]]);
274    } else {
275      int idx, idy;
276      const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
277      const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
278      for (idy = 0; idy < 2; idy += num_4x4_h) {
279        for (idx = 0; idx < 2; idx += num_4x4_w) {
280          const PREDICTION_MODE b_mode = mi->bmi[idy * 2 + idx].as_mode;
281          write_intra_mode(w, b_mode, cm->fc->y_mode_prob[0]);
282        }
283      }
284    }
285    write_intra_mode(w, mi->uv_mode, cm->fc->uv_mode_prob[mode]);
286  } else {
287    const int mode_ctx = mbmi_ext->mode_context[mi->ref_frame[0]];
288    const vpx_prob *const inter_probs = cm->fc->inter_mode_probs[mode_ctx];
289    write_ref_frames(cm, xd, w);
290
291    // If segment skip is not enabled code the mode.
292    if (!segfeature_active(seg, segment_id, SEG_LVL_SKIP)) {
293      if (bsize >= BLOCK_8X8) {
294        write_inter_mode(w, mode, inter_probs);
295      }
296    }
297
298    if (cm->interp_filter == SWITCHABLE) {
299      const int ctx = get_pred_context_switchable_interp(xd);
300      vp9_write_token(w, vp9_switchable_interp_tree,
301                      cm->fc->switchable_interp_prob[ctx],
302                      &switchable_interp_encodings[mi->interp_filter]);
303      ++interp_filter_selected[0][mi->interp_filter];
304    } else {
305      assert(mi->interp_filter == cm->interp_filter);
306    }
307
308    if (bsize < BLOCK_8X8) {
309      const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
310      const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
311      int idx, idy;
312      for (idy = 0; idy < 2; idy += num_4x4_h) {
313        for (idx = 0; idx < 2; idx += num_4x4_w) {
314          const int j = idy * 2 + idx;
315          const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
316          write_inter_mode(w, b_mode, inter_probs);
317          if (b_mode == NEWMV) {
318            for (ref = 0; ref < 1 + is_compound; ++ref)
319              vp9_encode_mv(cpi, w, &mi->bmi[j].as_mv[ref].as_mv,
320                            &mbmi_ext->ref_mvs[mi->ref_frame[ref]][0].as_mv,
321                            nmvc, allow_hp, max_mv_magnitude);
322          }
323        }
324      }
325    } else {
326      if (mode == NEWMV) {
327        for (ref = 0; ref < 1 + is_compound; ++ref)
328          vp9_encode_mv(cpi, w, &mi->mv[ref].as_mv,
329                        &mbmi_ext->ref_mvs[mi->ref_frame[ref]][0].as_mv, nmvc,
330                        allow_hp, max_mv_magnitude);
331      }
332    }
333  }
334}
335
336static void write_mb_modes_kf(const VP9_COMMON *cm, const MACROBLOCKD *xd,
337                              vpx_writer *w) {
338  const struct segmentation *const seg = &cm->seg;
339  const MODE_INFO *const mi = xd->mi[0];
340  const MODE_INFO *const above_mi = xd->above_mi;
341  const MODE_INFO *const left_mi = xd->left_mi;
342  const BLOCK_SIZE bsize = mi->sb_type;
343
344  if (seg->update_map) write_segment_id(w, seg, mi->segment_id);
345
346  write_skip(cm, xd, mi->segment_id, mi, w);
347
348  if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT)
349    write_selected_tx_size(cm, xd, w);
350
351  if (bsize >= BLOCK_8X8) {
352    write_intra_mode(w, mi->mode, get_y_mode_probs(mi, above_mi, left_mi, 0));
353  } else {
354    const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
355    const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
356    int idx, idy;
357
358    for (idy = 0; idy < 2; idy += num_4x4_h) {
359      for (idx = 0; idx < 2; idx += num_4x4_w) {
360        const int block = idy * 2 + idx;
361        write_intra_mode(w, mi->bmi[block].as_mode,
362                         get_y_mode_probs(mi, above_mi, left_mi, block));
363      }
364    }
365  }
366
367  write_intra_mode(w, mi->uv_mode, vp9_kf_uv_mode_prob[mi->mode]);
368}
369
370static void write_modes_b(
371    VP9_COMP *cpi, MACROBLOCKD *const xd, const TileInfo *const tile,
372    vpx_writer *w, TOKENEXTRA **tok, const TOKENEXTRA *const tok_end,
373    int mi_row, int mi_col, unsigned int *const max_mv_magnitude,
374    int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE]) {
375  const VP9_COMMON *const cm = &cpi->common;
376  const MB_MODE_INFO_EXT *const mbmi_ext =
377      cpi->td.mb.mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
378  MODE_INFO *m;
379
380  xd->mi = cm->mi_grid_visible + (mi_row * cm->mi_stride + mi_col);
381  m = xd->mi[0];
382
383  set_mi_row_col(xd, tile, mi_row, num_8x8_blocks_high_lookup[m->sb_type],
384                 mi_col, num_8x8_blocks_wide_lookup[m->sb_type], cm->mi_rows,
385                 cm->mi_cols);
386  if (frame_is_intra_only(cm)) {
387    write_mb_modes_kf(cm, xd, w);
388  } else {
389    pack_inter_mode_mvs(cpi, xd, mbmi_ext, w, max_mv_magnitude,
390                        interp_filter_selected);
391  }
392
393  assert(*tok < tok_end);
394  pack_mb_tokens(w, tok, tok_end, cm->bit_depth);
395}
396
397static void write_partition(const VP9_COMMON *const cm,
398                            const MACROBLOCKD *const xd, int hbs, int mi_row,
399                            int mi_col, PARTITION_TYPE p, BLOCK_SIZE bsize,
400                            vpx_writer *w) {
401  const int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
402  const vpx_prob *const probs = xd->partition_probs[ctx];
403  const int has_rows = (mi_row + hbs) < cm->mi_rows;
404  const int has_cols = (mi_col + hbs) < cm->mi_cols;
405
406  if (has_rows && has_cols) {
407    vp9_write_token(w, vp9_partition_tree, probs, &partition_encodings[p]);
408  } else if (!has_rows && has_cols) {
409    assert(p == PARTITION_SPLIT || p == PARTITION_HORZ);
410    vpx_write(w, p == PARTITION_SPLIT, probs[1]);
411  } else if (has_rows && !has_cols) {
412    assert(p == PARTITION_SPLIT || p == PARTITION_VERT);
413    vpx_write(w, p == PARTITION_SPLIT, probs[2]);
414  } else {
415    assert(p == PARTITION_SPLIT);
416  }
417}
418
419static void write_modes_sb(
420    VP9_COMP *cpi, MACROBLOCKD *const xd, const TileInfo *const tile,
421    vpx_writer *w, TOKENEXTRA **tok, const TOKENEXTRA *const tok_end,
422    int mi_row, int mi_col, BLOCK_SIZE bsize,
423    unsigned int *const max_mv_magnitude,
424    int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE]) {
425  const VP9_COMMON *const cm = &cpi->common;
426  const int bsl = b_width_log2_lookup[bsize];
427  const int bs = (1 << bsl) / 4;
428  PARTITION_TYPE partition;
429  BLOCK_SIZE subsize;
430  const MODE_INFO *m = NULL;
431
432  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
433
434  m = cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col];
435
436  partition = partition_lookup[bsl][m->sb_type];
437  write_partition(cm, xd, bs, mi_row, mi_col, partition, bsize, w);
438  subsize = get_subsize(bsize, partition);
439  if (subsize < BLOCK_8X8) {
440    write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
441                  max_mv_magnitude, interp_filter_selected);
442  } else {
443    switch (partition) {
444      case PARTITION_NONE:
445        write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
446                      max_mv_magnitude, interp_filter_selected);
447        break;
448      case PARTITION_HORZ:
449        write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
450                      max_mv_magnitude, interp_filter_selected);
451        if (mi_row + bs < cm->mi_rows)
452          write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row + bs, mi_col,
453                        max_mv_magnitude, interp_filter_selected);
454        break;
455      case PARTITION_VERT:
456        write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col,
457                      max_mv_magnitude, interp_filter_selected);
458        if (mi_col + bs < cm->mi_cols)
459          write_modes_b(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col + bs,
460                        max_mv_magnitude, interp_filter_selected);
461        break;
462      case PARTITION_SPLIT:
463        write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col, subsize,
464                       max_mv_magnitude, interp_filter_selected);
465        write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row, mi_col + bs,
466                       subsize, max_mv_magnitude, interp_filter_selected);
467        write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row + bs, mi_col,
468                       subsize, max_mv_magnitude, interp_filter_selected);
469        write_modes_sb(cpi, xd, tile, w, tok, tok_end, mi_row + bs, mi_col + bs,
470                       subsize, max_mv_magnitude, interp_filter_selected);
471        break;
472      default: assert(0);
473    }
474  }
475
476  // update partition context
477  if (bsize >= BLOCK_8X8 &&
478      (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
479    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
480}
481
482static void write_modes(
483    VP9_COMP *cpi, MACROBLOCKD *const xd, const TileInfo *const tile,
484    vpx_writer *w, int tile_row, int tile_col,
485    unsigned int *const max_mv_magnitude,
486    int interp_filter_selected[MAX_REF_FRAMES][SWITCHABLE]) {
487  const VP9_COMMON *const cm = &cpi->common;
488  int mi_row, mi_col, tile_sb_row;
489  TOKENEXTRA *tok = NULL;
490  TOKENEXTRA *tok_end = NULL;
491
492  set_partition_probs(cm, xd);
493
494  for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
495       mi_row += MI_BLOCK_SIZE) {
496    tile_sb_row = mi_cols_aligned_to_sb(mi_row - tile->mi_row_start) >>
497                  MI_BLOCK_SIZE_LOG2;
498    tok = cpi->tplist[tile_row][tile_col][tile_sb_row].start;
499    tok_end = tok + cpi->tplist[tile_row][tile_col][tile_sb_row].count;
500
501    vp9_zero(xd->left_seg_context);
502    for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
503         mi_col += MI_BLOCK_SIZE)
504      write_modes_sb(cpi, xd, tile, w, &tok, tok_end, mi_row, mi_col,
505                     BLOCK_64X64, max_mv_magnitude, interp_filter_selected);
506
507    assert(tok == cpi->tplist[tile_row][tile_col][tile_sb_row].stop);
508  }
509}
510
511static void build_tree_distribution(VP9_COMP *cpi, TX_SIZE tx_size,
512                                    vp9_coeff_stats *coef_branch_ct,
513                                    vp9_coeff_probs_model *coef_probs) {
514  vp9_coeff_count *coef_counts = cpi->td.rd_counts.coef_counts[tx_size];
515  unsigned int(*eob_branch_ct)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS] =
516      cpi->common.counts.eob_branch[tx_size];
517  int i, j, k, l, m;
518
519  for (i = 0; i < PLANE_TYPES; ++i) {
520    for (j = 0; j < REF_TYPES; ++j) {
521      for (k = 0; k < COEF_BANDS; ++k) {
522        for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
523          vp9_tree_probs_from_distribution(vp9_coef_tree,
524                                           coef_branch_ct[i][j][k][l],
525                                           coef_counts[i][j][k][l]);
526          coef_branch_ct[i][j][k][l][0][1] =
527              eob_branch_ct[i][j][k][l] - coef_branch_ct[i][j][k][l][0][0];
528          for (m = 0; m < UNCONSTRAINED_NODES; ++m)
529            coef_probs[i][j][k][l][m] =
530                get_binary_prob(coef_branch_ct[i][j][k][l][m][0],
531                                coef_branch_ct[i][j][k][l][m][1]);
532        }
533      }
534    }
535  }
536}
537
538static void update_coef_probs_common(vpx_writer *const bc, VP9_COMP *cpi,
539                                     TX_SIZE tx_size,
540                                     vp9_coeff_stats *frame_branch_ct,
541                                     vp9_coeff_probs_model *new_coef_probs) {
542  vp9_coeff_probs_model *old_coef_probs = cpi->common.fc->coef_probs[tx_size];
543  const vpx_prob upd = DIFF_UPDATE_PROB;
544  const int entropy_nodes_update = UNCONSTRAINED_NODES;
545  int i, j, k, l, t;
546  int stepsize = cpi->sf.coeff_prob_appx_step;
547
548  switch (cpi->sf.use_fast_coef_updates) {
549    case TWO_LOOP: {
550      /* dry run to see if there is any update at all needed */
551      int savings = 0;
552      int update[2] = { 0, 0 };
553      for (i = 0; i < PLANE_TYPES; ++i) {
554        for (j = 0; j < REF_TYPES; ++j) {
555          for (k = 0; k < COEF_BANDS; ++k) {
556            for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
557              for (t = 0; t < entropy_nodes_update; ++t) {
558                vpx_prob newp = new_coef_probs[i][j][k][l][t];
559                const vpx_prob oldp = old_coef_probs[i][j][k][l][t];
560                int s;
561                int u = 0;
562                if (t == PIVOT_NODE)
563                  s = vp9_prob_diff_update_savings_search_model(
564                      frame_branch_ct[i][j][k][l][0], oldp, &newp, upd,
565                      stepsize);
566                else
567                  s = vp9_prob_diff_update_savings_search(
568                      frame_branch_ct[i][j][k][l][t], oldp, &newp, upd);
569                if (s > 0 && newp != oldp) u = 1;
570                if (u)
571                  savings += s - (int)(vp9_cost_zero(upd));
572                else
573                  savings -= (int)(vp9_cost_zero(upd));
574                update[u]++;
575              }
576            }
577          }
578        }
579      }
580
581      // printf("Update %d %d, savings %d\n", update[0], update[1], savings);
582      /* Is coef updated at all */
583      if (update[1] == 0 || savings < 0) {
584        vpx_write_bit(bc, 0);
585        return;
586      }
587      vpx_write_bit(bc, 1);
588      for (i = 0; i < PLANE_TYPES; ++i) {
589        for (j = 0; j < REF_TYPES; ++j) {
590          for (k = 0; k < COEF_BANDS; ++k) {
591            for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
592              // calc probs and branch cts for this frame only
593              for (t = 0; t < entropy_nodes_update; ++t) {
594                vpx_prob newp = new_coef_probs[i][j][k][l][t];
595                vpx_prob *oldp = old_coef_probs[i][j][k][l] + t;
596                const vpx_prob upd = DIFF_UPDATE_PROB;
597                int s;
598                int u = 0;
599                if (t == PIVOT_NODE)
600                  s = vp9_prob_diff_update_savings_search_model(
601                      frame_branch_ct[i][j][k][l][0], *oldp, &newp, upd,
602                      stepsize);
603                else
604                  s = vp9_prob_diff_update_savings_search(
605                      frame_branch_ct[i][j][k][l][t], *oldp, &newp, upd);
606                if (s > 0 && newp != *oldp) u = 1;
607                vpx_write(bc, u, upd);
608                if (u) {
609                  /* send/use new probability */
610                  vp9_write_prob_diff_update(bc, newp, *oldp);
611                  *oldp = newp;
612                }
613              }
614            }
615          }
616        }
617      }
618      return;
619    }
620
621    case ONE_LOOP_REDUCED: {
622      int updates = 0;
623      int noupdates_before_first = 0;
624      for (i = 0; i < PLANE_TYPES; ++i) {
625        for (j = 0; j < REF_TYPES; ++j) {
626          for (k = 0; k < COEF_BANDS; ++k) {
627            for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
628              // calc probs and branch cts for this frame only
629              for (t = 0; t < entropy_nodes_update; ++t) {
630                vpx_prob newp = new_coef_probs[i][j][k][l][t];
631                vpx_prob *oldp = old_coef_probs[i][j][k][l] + t;
632                int s;
633                int u = 0;
634
635                if (t == PIVOT_NODE) {
636                  s = vp9_prob_diff_update_savings_search_model(
637                      frame_branch_ct[i][j][k][l][0], *oldp, &newp, upd,
638                      stepsize);
639                } else {
640                  s = vp9_prob_diff_update_savings_search(
641                      frame_branch_ct[i][j][k][l][t], *oldp, &newp, upd);
642                }
643
644                if (s > 0 && newp != *oldp) u = 1;
645                updates += u;
646                if (u == 0 && updates == 0) {
647                  noupdates_before_first++;
648                  continue;
649                }
650                if (u == 1 && updates == 1) {
651                  int v;
652                  // first update
653                  vpx_write_bit(bc, 1);
654                  for (v = 0; v < noupdates_before_first; ++v)
655                    vpx_write(bc, 0, upd);
656                }
657                vpx_write(bc, u, upd);
658                if (u) {
659                  /* send/use new probability */
660                  vp9_write_prob_diff_update(bc, newp, *oldp);
661                  *oldp = newp;
662                }
663              }
664            }
665          }
666        }
667      }
668      if (updates == 0) {
669        vpx_write_bit(bc, 0);  // no updates
670      }
671      return;
672    }
673    default: assert(0);
674  }
675}
676
677static void update_coef_probs(VP9_COMP *cpi, vpx_writer *w) {
678  const TX_MODE tx_mode = cpi->common.tx_mode;
679  const TX_SIZE max_tx_size = tx_mode_to_biggest_tx_size[tx_mode];
680  TX_SIZE tx_size;
681  for (tx_size = TX_4X4; tx_size <= max_tx_size; ++tx_size) {
682    vp9_coeff_stats frame_branch_ct[PLANE_TYPES];
683    vp9_coeff_probs_model frame_coef_probs[PLANE_TYPES];
684    if (cpi->td.counts->tx.tx_totals[tx_size] <= 20 ||
685        (tx_size >= TX_16X16 && cpi->sf.tx_size_search_method == USE_TX_8X8)) {
686      vpx_write_bit(w, 0);
687    } else {
688      build_tree_distribution(cpi, tx_size, frame_branch_ct, frame_coef_probs);
689      update_coef_probs_common(w, cpi, tx_size, frame_branch_ct,
690                               frame_coef_probs);
691    }
692  }
693}
694
695static void encode_loopfilter(struct loopfilter *lf,
696                              struct vpx_write_bit_buffer *wb) {
697  int i;
698
699  // Encode the loop filter level and type
700  vpx_wb_write_literal(wb, lf->filter_level, 6);
701  vpx_wb_write_literal(wb, lf->sharpness_level, 3);
702
703  // Write out loop filter deltas applied at the MB level based on mode or
704  // ref frame (if they are enabled).
705  vpx_wb_write_bit(wb, lf->mode_ref_delta_enabled);
706
707  if (lf->mode_ref_delta_enabled) {
708    vpx_wb_write_bit(wb, lf->mode_ref_delta_update);
709    if (lf->mode_ref_delta_update) {
710      for (i = 0; i < MAX_REF_LF_DELTAS; i++) {
711        const int delta = lf->ref_deltas[i];
712        const int changed = delta != lf->last_ref_deltas[i];
713        vpx_wb_write_bit(wb, changed);
714        if (changed) {
715          lf->last_ref_deltas[i] = delta;
716          vpx_wb_write_literal(wb, abs(delta) & 0x3F, 6);
717          vpx_wb_write_bit(wb, delta < 0);
718        }
719      }
720
721      for (i = 0; i < MAX_MODE_LF_DELTAS; i++) {
722        const int delta = lf->mode_deltas[i];
723        const int changed = delta != lf->last_mode_deltas[i];
724        vpx_wb_write_bit(wb, changed);
725        if (changed) {
726          lf->last_mode_deltas[i] = delta;
727          vpx_wb_write_literal(wb, abs(delta) & 0x3F, 6);
728          vpx_wb_write_bit(wb, delta < 0);
729        }
730      }
731    }
732  }
733}
734
735static void write_delta_q(struct vpx_write_bit_buffer *wb, int delta_q) {
736  if (delta_q != 0) {
737    vpx_wb_write_bit(wb, 1);
738    vpx_wb_write_literal(wb, abs(delta_q), 4);
739    vpx_wb_write_bit(wb, delta_q < 0);
740  } else {
741    vpx_wb_write_bit(wb, 0);
742  }
743}
744
745static void encode_quantization(const VP9_COMMON *const cm,
746                                struct vpx_write_bit_buffer *wb) {
747  vpx_wb_write_literal(wb, cm->base_qindex, QINDEX_BITS);
748  write_delta_q(wb, cm->y_dc_delta_q);
749  write_delta_q(wb, cm->uv_dc_delta_q);
750  write_delta_q(wb, cm->uv_ac_delta_q);
751}
752
753static void encode_segmentation(VP9_COMMON *cm, MACROBLOCKD *xd,
754                                struct vpx_write_bit_buffer *wb) {
755  int i, j;
756
757  const struct segmentation *seg = &cm->seg;
758
759  vpx_wb_write_bit(wb, seg->enabled);
760  if (!seg->enabled) return;
761
762  // Segmentation map
763  vpx_wb_write_bit(wb, seg->update_map);
764  if (seg->update_map) {
765    // Select the coding strategy (temporal or spatial)
766    vp9_choose_segmap_coding_method(cm, xd);
767    // Write out probabilities used to decode unpredicted  macro-block segments
768    for (i = 0; i < SEG_TREE_PROBS; i++) {
769      const int prob = seg->tree_probs[i];
770      const int update = prob != MAX_PROB;
771      vpx_wb_write_bit(wb, update);
772      if (update) vpx_wb_write_literal(wb, prob, 8);
773    }
774
775    // Write out the chosen coding method.
776    vpx_wb_write_bit(wb, seg->temporal_update);
777    if (seg->temporal_update) {
778      for (i = 0; i < PREDICTION_PROBS; i++) {
779        const int prob = seg->pred_probs[i];
780        const int update = prob != MAX_PROB;
781        vpx_wb_write_bit(wb, update);
782        if (update) vpx_wb_write_literal(wb, prob, 8);
783      }
784    }
785  }
786
787  // Segmentation data
788  vpx_wb_write_bit(wb, seg->update_data);
789  if (seg->update_data) {
790    vpx_wb_write_bit(wb, seg->abs_delta);
791
792    for (i = 0; i < MAX_SEGMENTS; i++) {
793      for (j = 0; j < SEG_LVL_MAX; j++) {
794        const int active = segfeature_active(seg, i, j);
795        vpx_wb_write_bit(wb, active);
796        if (active) {
797          const int data = get_segdata(seg, i, j);
798          const int data_max = vp9_seg_feature_data_max(j);
799
800          if (vp9_is_segfeature_signed(j)) {
801            encode_unsigned_max(wb, abs(data), data_max);
802            vpx_wb_write_bit(wb, data < 0);
803          } else {
804            encode_unsigned_max(wb, data, data_max);
805          }
806        }
807      }
808    }
809  }
810}
811
812static void encode_txfm_probs(VP9_COMMON *cm, vpx_writer *w,
813                              FRAME_COUNTS *counts) {
814  // Mode
815  vpx_write_literal(w, VPXMIN(cm->tx_mode, ALLOW_32X32), 2);
816  if (cm->tx_mode >= ALLOW_32X32)
817    vpx_write_bit(w, cm->tx_mode == TX_MODE_SELECT);
818
819  // Probabilities
820  if (cm->tx_mode == TX_MODE_SELECT) {
821    int i, j;
822    unsigned int ct_8x8p[TX_SIZES - 3][2];
823    unsigned int ct_16x16p[TX_SIZES - 2][2];
824    unsigned int ct_32x32p[TX_SIZES - 1][2];
825
826    for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
827      tx_counts_to_branch_counts_8x8(counts->tx.p8x8[i], ct_8x8p);
828      for (j = 0; j < TX_SIZES - 3; j++)
829        vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p8x8[i][j], ct_8x8p[j]);
830    }
831
832    for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
833      tx_counts_to_branch_counts_16x16(counts->tx.p16x16[i], ct_16x16p);
834      for (j = 0; j < TX_SIZES - 2; j++)
835        vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p16x16[i][j],
836                                  ct_16x16p[j]);
837    }
838
839    for (i = 0; i < TX_SIZE_CONTEXTS; i++) {
840      tx_counts_to_branch_counts_32x32(counts->tx.p32x32[i], ct_32x32p);
841      for (j = 0; j < TX_SIZES - 1; j++)
842        vp9_cond_prob_diff_update(w, &cm->fc->tx_probs.p32x32[i][j],
843                                  ct_32x32p[j]);
844    }
845  }
846}
847
848static void write_interp_filter(INTERP_FILTER filter,
849                                struct vpx_write_bit_buffer *wb) {
850  const int filter_to_literal[] = { 1, 0, 2, 3 };
851
852  vpx_wb_write_bit(wb, filter == SWITCHABLE);
853  if (filter != SWITCHABLE)
854    vpx_wb_write_literal(wb, filter_to_literal[filter], 2);
855}
856
857static void fix_interp_filter(VP9_COMMON *cm, FRAME_COUNTS *counts) {
858  if (cm->interp_filter == SWITCHABLE) {
859    // Check to see if only one of the filters is actually used
860    int count[SWITCHABLE_FILTERS];
861    int i, j, c = 0;
862    for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
863      count[i] = 0;
864      for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
865        count[i] += counts->switchable_interp[j][i];
866      c += (count[i] > 0);
867    }
868    if (c == 1) {
869      // Only one filter is used. So set the filter at frame level
870      for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
871        if (count[i]) {
872          cm->interp_filter = i;
873          break;
874        }
875      }
876    }
877  }
878}
879
880static void write_tile_info(const VP9_COMMON *const cm,
881                            struct vpx_write_bit_buffer *wb) {
882  int min_log2_tile_cols, max_log2_tile_cols, ones;
883  vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
884
885  // columns
886  ones = cm->log2_tile_cols - min_log2_tile_cols;
887  while (ones--) vpx_wb_write_bit(wb, 1);
888
889  if (cm->log2_tile_cols < max_log2_tile_cols) vpx_wb_write_bit(wb, 0);
890
891  // rows
892  vpx_wb_write_bit(wb, cm->log2_tile_rows != 0);
893  if (cm->log2_tile_rows != 0) vpx_wb_write_bit(wb, cm->log2_tile_rows != 1);
894}
895
896int vp9_get_refresh_mask(VP9_COMP *cpi) {
897  if (vp9_preserve_existing_gf(cpi)) {
898    // We have decided to preserve the previously existing golden frame as our
899    // new ARF frame. However, in the short term we leave it in the GF slot and,
900    // if we're updating the GF with the current decoded frame, we save it
901    // instead to the ARF slot.
902    // Later, in the function vp9_encoder.c:vp9_update_reference_frames() we
903    // will swap gld_fb_idx and alt_fb_idx to achieve our objective. We do it
904    // there so that it can be done outside of the recode loop.
905    // Note: This is highly specific to the use of ARF as a forward reference,
906    // and this needs to be generalized as other uses are implemented
907    // (like RTC/temporal scalability).
908    return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
909           (cpi->refresh_golden_frame << cpi->alt_fb_idx);
910  } else {
911    int arf_idx = cpi->alt_fb_idx;
912    if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
913      const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
914      arf_idx = gf_group->arf_update_idx[gf_group->index];
915    }
916    return (cpi->refresh_last_frame << cpi->lst_fb_idx) |
917           (cpi->refresh_golden_frame << cpi->gld_fb_idx) |
918           (cpi->refresh_alt_ref_frame << arf_idx);
919  }
920}
921
922static int encode_tile_worker(void *arg1, void *arg2) {
923  VP9_COMP *cpi = (VP9_COMP *)arg1;
924  VP9BitstreamWorkerData *data = (VP9BitstreamWorkerData *)arg2;
925  MACROBLOCKD *const xd = &data->xd;
926  const int tile_row = 0;
927  vpx_start_encode(&data->bit_writer, data->dest);
928  write_modes(cpi, xd, &cpi->tile_data[data->tile_idx].tile_info,
929              &data->bit_writer, tile_row, data->tile_idx,
930              &data->max_mv_magnitude, data->interp_filter_selected);
931  vpx_stop_encode(&data->bit_writer);
932  return 1;
933}
934
935void vp9_bitstream_encode_tiles_buffer_dealloc(VP9_COMP *const cpi) {
936  if (cpi->vp9_bitstream_worker_data) {
937    int i;
938    for (i = 1; i < cpi->num_workers; ++i) {
939      vpx_free(cpi->vp9_bitstream_worker_data[i].dest);
940    }
941    vpx_free(cpi->vp9_bitstream_worker_data);
942    cpi->vp9_bitstream_worker_data = NULL;
943  }
944}
945
946static int encode_tiles_buffer_alloc(VP9_COMP *const cpi) {
947  int i;
948  const size_t worker_data_size =
949      cpi->num_workers * sizeof(*cpi->vp9_bitstream_worker_data);
950  cpi->vp9_bitstream_worker_data = vpx_memalign(16, worker_data_size);
951  memset(cpi->vp9_bitstream_worker_data, 0, worker_data_size);
952  if (!cpi->vp9_bitstream_worker_data) return 1;
953  for (i = 1; i < cpi->num_workers; ++i) {
954    cpi->vp9_bitstream_worker_data[i].dest_size =
955        cpi->oxcf.width * cpi->oxcf.height;
956    cpi->vp9_bitstream_worker_data[i].dest =
957        vpx_malloc(cpi->vp9_bitstream_worker_data[i].dest_size);
958    if (!cpi->vp9_bitstream_worker_data[i].dest) return 1;
959  }
960  return 0;
961}
962
963static size_t encode_tiles_mt(VP9_COMP *cpi, uint8_t *data_ptr) {
964  const VPxWorkerInterface *const winterface = vpx_get_worker_interface();
965  VP9_COMMON *const cm = &cpi->common;
966  const int tile_cols = 1 << cm->log2_tile_cols;
967  const int num_workers = cpi->num_workers;
968  size_t total_size = 0;
969  int tile_col = 0;
970
971  if (!cpi->vp9_bitstream_worker_data ||
972      cpi->vp9_bitstream_worker_data[1].dest_size >
973          (cpi->oxcf.width * cpi->oxcf.height)) {
974    vp9_bitstream_encode_tiles_buffer_dealloc(cpi);
975    if (encode_tiles_buffer_alloc(cpi)) return 0;
976  }
977
978  while (tile_col < tile_cols) {
979    int i, j;
980    for (i = 0; i < num_workers && tile_col < tile_cols; ++i) {
981      VPxWorker *const worker = &cpi->workers[i];
982      VP9BitstreamWorkerData *const data = &cpi->vp9_bitstream_worker_data[i];
983
984      // Populate the worker data.
985      data->xd = cpi->td.mb.e_mbd;
986      data->tile_idx = tile_col;
987      data->max_mv_magnitude = cpi->max_mv_magnitude;
988      memset(data->interp_filter_selected, 0,
989             sizeof(data->interp_filter_selected[0][0]) * SWITCHABLE);
990
991      // First thread can directly write into the output buffer.
992      if (i == 0) {
993        // If this worker happens to be for the last tile, then do not offset it
994        // by 4 for the tile size.
995        data->dest =
996            data_ptr + total_size + (tile_col == tile_cols - 1 ? 0 : 4);
997      }
998      worker->data1 = cpi;
999      worker->data2 = data;
1000      worker->hook = encode_tile_worker;
1001      worker->had_error = 0;
1002
1003      if (i < num_workers - 1) {
1004        winterface->launch(worker);
1005      } else {
1006        winterface->execute(worker);
1007      }
1008      ++tile_col;
1009    }
1010    for (j = 0; j < i; ++j) {
1011      VPxWorker *const worker = &cpi->workers[j];
1012      VP9BitstreamWorkerData *const data =
1013          (VP9BitstreamWorkerData *)worker->data2;
1014      uint32_t tile_size;
1015      int k;
1016
1017      if (!winterface->sync(worker)) return 0;
1018      tile_size = data->bit_writer.pos;
1019
1020      // Aggregate per-thread bitstream stats.
1021      cpi->max_mv_magnitude =
1022          VPXMAX(cpi->max_mv_magnitude, data->max_mv_magnitude);
1023      for (k = 0; k < SWITCHABLE; ++k) {
1024        cpi->interp_filter_selected[0][k] += data->interp_filter_selected[0][k];
1025      }
1026
1027      // Prefix the size of the tile on all but the last.
1028      if (tile_col != tile_cols || j < i - 1) {
1029        mem_put_be32(data_ptr + total_size, tile_size);
1030        total_size += 4;
1031      }
1032      if (j > 0) {
1033        memcpy(data_ptr + total_size, data->dest, tile_size);
1034      }
1035      total_size += tile_size;
1036    }
1037  }
1038  return total_size;
1039}
1040
1041static size_t encode_tiles(VP9_COMP *cpi, uint8_t *data_ptr) {
1042  VP9_COMMON *const cm = &cpi->common;
1043  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
1044  vpx_writer residual_bc;
1045  int tile_row, tile_col;
1046  size_t total_size = 0;
1047  const int tile_cols = 1 << cm->log2_tile_cols;
1048  const int tile_rows = 1 << cm->log2_tile_rows;
1049
1050  memset(cm->above_seg_context, 0,
1051         sizeof(*cm->above_seg_context) * mi_cols_aligned_to_sb(cm->mi_cols));
1052
1053  // Encoding tiles in parallel is done only for realtime mode now. In other
1054  // modes the speed up is insignificant and requires further testing to ensure
1055  // that it does not make the overall process worse in any case.
1056  if (cpi->oxcf.mode == REALTIME && cpi->num_workers > 1 && tile_rows == 1 &&
1057      tile_cols > 1) {
1058    return encode_tiles_mt(cpi, data_ptr);
1059  }
1060
1061  for (tile_row = 0; tile_row < tile_rows; tile_row++) {
1062    for (tile_col = 0; tile_col < tile_cols; tile_col++) {
1063      int tile_idx = tile_row * tile_cols + tile_col;
1064
1065      if (tile_col < tile_cols - 1 || tile_row < tile_rows - 1)
1066        vpx_start_encode(&residual_bc, data_ptr + total_size + 4);
1067      else
1068        vpx_start_encode(&residual_bc, data_ptr + total_size);
1069
1070      write_modes(cpi, xd, &cpi->tile_data[tile_idx].tile_info, &residual_bc,
1071                  tile_row, tile_col, &cpi->max_mv_magnitude,
1072                  cpi->interp_filter_selected);
1073
1074      vpx_stop_encode(&residual_bc);
1075      if (tile_col < tile_cols - 1 || tile_row < tile_rows - 1) {
1076        // size of this tile
1077        mem_put_be32(data_ptr + total_size, residual_bc.pos);
1078        total_size += 4;
1079      }
1080
1081      total_size += residual_bc.pos;
1082    }
1083  }
1084  return total_size;
1085}
1086
1087static void write_render_size(const VP9_COMMON *cm,
1088                              struct vpx_write_bit_buffer *wb) {
1089  const int scaling_active =
1090      cm->width != cm->render_width || cm->height != cm->render_height;
1091  vpx_wb_write_bit(wb, scaling_active);
1092  if (scaling_active) {
1093    vpx_wb_write_literal(wb, cm->render_width - 1, 16);
1094    vpx_wb_write_literal(wb, cm->render_height - 1, 16);
1095  }
1096}
1097
1098static void write_frame_size(const VP9_COMMON *cm,
1099                             struct vpx_write_bit_buffer *wb) {
1100  vpx_wb_write_literal(wb, cm->width - 1, 16);
1101  vpx_wb_write_literal(wb, cm->height - 1, 16);
1102
1103  write_render_size(cm, wb);
1104}
1105
1106static void write_frame_size_with_refs(VP9_COMP *cpi,
1107                                       struct vpx_write_bit_buffer *wb) {
1108  VP9_COMMON *const cm = &cpi->common;
1109  int found = 0;
1110
1111  MV_REFERENCE_FRAME ref_frame;
1112  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
1113    YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi, ref_frame);
1114
1115    // Set "found" to 0 for temporal svc and for spatial svc key frame
1116    if (cpi->use_svc &&
1117        ((cpi->svc.number_temporal_layers > 1 &&
1118          cpi->oxcf.rc_mode == VPX_CBR) ||
1119         (cpi->svc.number_spatial_layers > 1 &&
1120          cpi->svc.layer_context[cpi->svc.spatial_layer_id].is_key_frame) ||
1121         (is_two_pass_svc(cpi) &&
1122          cpi->svc.encode_empty_frame_state == ENCODING &&
1123          cpi->svc.layer_context[0].frames_from_key_frame <
1124              cpi->svc.number_temporal_layers + 1))) {
1125      found = 0;
1126    } else if (cfg != NULL) {
1127      found =
1128          cm->width == cfg->y_crop_width && cm->height == cfg->y_crop_height;
1129    }
1130    vpx_wb_write_bit(wb, found);
1131    if (found) {
1132      break;
1133    }
1134  }
1135
1136  if (!found) {
1137    vpx_wb_write_literal(wb, cm->width - 1, 16);
1138    vpx_wb_write_literal(wb, cm->height - 1, 16);
1139  }
1140
1141  write_render_size(cm, wb);
1142}
1143
1144static void write_sync_code(struct vpx_write_bit_buffer *wb) {
1145  vpx_wb_write_literal(wb, VP9_SYNC_CODE_0, 8);
1146  vpx_wb_write_literal(wb, VP9_SYNC_CODE_1, 8);
1147  vpx_wb_write_literal(wb, VP9_SYNC_CODE_2, 8);
1148}
1149
1150static void write_profile(BITSTREAM_PROFILE profile,
1151                          struct vpx_write_bit_buffer *wb) {
1152  switch (profile) {
1153    case PROFILE_0: vpx_wb_write_literal(wb, 0, 2); break;
1154    case PROFILE_1: vpx_wb_write_literal(wb, 2, 2); break;
1155    case PROFILE_2: vpx_wb_write_literal(wb, 1, 2); break;
1156    case PROFILE_3: vpx_wb_write_literal(wb, 6, 3); break;
1157    default: assert(0);
1158  }
1159}
1160
1161static void write_bitdepth_colorspace_sampling(
1162    VP9_COMMON *const cm, struct vpx_write_bit_buffer *wb) {
1163  if (cm->profile >= PROFILE_2) {
1164    assert(cm->bit_depth > VPX_BITS_8);
1165    vpx_wb_write_bit(wb, cm->bit_depth == VPX_BITS_10 ? 0 : 1);
1166  }
1167  vpx_wb_write_literal(wb, cm->color_space, 3);
1168  if (cm->color_space != VPX_CS_SRGB) {
1169    // 0: [16, 235] (i.e. xvYCC), 1: [0, 255]
1170    vpx_wb_write_bit(wb, cm->color_range);
1171    if (cm->profile == PROFILE_1 || cm->profile == PROFILE_3) {
1172      assert(cm->subsampling_x != 1 || cm->subsampling_y != 1);
1173      vpx_wb_write_bit(wb, cm->subsampling_x);
1174      vpx_wb_write_bit(wb, cm->subsampling_y);
1175      vpx_wb_write_bit(wb, 0);  // unused
1176    } else {
1177      assert(cm->subsampling_x == 1 && cm->subsampling_y == 1);
1178    }
1179  } else {
1180    assert(cm->profile == PROFILE_1 || cm->profile == PROFILE_3);
1181    vpx_wb_write_bit(wb, 0);  // unused
1182  }
1183}
1184
1185static void write_uncompressed_header(VP9_COMP *cpi,
1186                                      struct vpx_write_bit_buffer *wb) {
1187  VP9_COMMON *const cm = &cpi->common;
1188  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
1189
1190  vpx_wb_write_literal(wb, VP9_FRAME_MARKER, 2);
1191
1192  write_profile(cm->profile, wb);
1193
1194  vpx_wb_write_bit(wb, 0);  // show_existing_frame
1195  vpx_wb_write_bit(wb, cm->frame_type);
1196  vpx_wb_write_bit(wb, cm->show_frame);
1197  vpx_wb_write_bit(wb, cm->error_resilient_mode);
1198
1199  if (cm->frame_type == KEY_FRAME) {
1200    write_sync_code(wb);
1201    write_bitdepth_colorspace_sampling(cm, wb);
1202    write_frame_size(cm, wb);
1203  } else {
1204    // In spatial svc if it's not error_resilient_mode then we need to code all
1205    // visible frames as invisible. But we need to keep the show_frame flag so
1206    // that the publisher could know whether it is supposed to be visible.
1207    // So we will code the show_frame flag as it is. Then code the intra_only
1208    // bit here. This will make the bitstream incompatible. In the player we
1209    // will change to show_frame flag to 0, then add an one byte frame with
1210    // show_existing_frame flag which tells the decoder which frame we want to
1211    // show.
1212    if (!cm->show_frame) vpx_wb_write_bit(wb, cm->intra_only);
1213
1214    if (!cm->error_resilient_mode)
1215      vpx_wb_write_literal(wb, cm->reset_frame_context, 2);
1216
1217    if (cm->intra_only) {
1218      write_sync_code(wb);
1219
1220      // Note for profile 0, 420 8bpp is assumed.
1221      if (cm->profile > PROFILE_0) {
1222        write_bitdepth_colorspace_sampling(cm, wb);
1223      }
1224
1225      vpx_wb_write_literal(wb, vp9_get_refresh_mask(cpi), REF_FRAMES);
1226      write_frame_size(cm, wb);
1227    } else {
1228      MV_REFERENCE_FRAME ref_frame;
1229      vpx_wb_write_literal(wb, vp9_get_refresh_mask(cpi), REF_FRAMES);
1230      for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
1231        assert(get_ref_frame_map_idx(cpi, ref_frame) != INVALID_IDX);
1232        vpx_wb_write_literal(wb, get_ref_frame_map_idx(cpi, ref_frame),
1233                             REF_FRAMES_LOG2);
1234        vpx_wb_write_bit(wb, cm->ref_frame_sign_bias[ref_frame]);
1235      }
1236
1237      write_frame_size_with_refs(cpi, wb);
1238
1239      vpx_wb_write_bit(wb, cm->allow_high_precision_mv);
1240
1241      fix_interp_filter(cm, cpi->td.counts);
1242      write_interp_filter(cm->interp_filter, wb);
1243    }
1244  }
1245
1246  if (!cm->error_resilient_mode) {
1247    vpx_wb_write_bit(wb, cm->refresh_frame_context);
1248    vpx_wb_write_bit(wb, cm->frame_parallel_decoding_mode);
1249  }
1250
1251  vpx_wb_write_literal(wb, cm->frame_context_idx, FRAME_CONTEXTS_LOG2);
1252
1253  encode_loopfilter(&cm->lf, wb);
1254  encode_quantization(cm, wb);
1255  encode_segmentation(cm, xd, wb);
1256
1257  write_tile_info(cm, wb);
1258}
1259
1260static size_t write_compressed_header(VP9_COMP *cpi, uint8_t *data) {
1261  VP9_COMMON *const cm = &cpi->common;
1262  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
1263  FRAME_CONTEXT *const fc = cm->fc;
1264  FRAME_COUNTS *counts = cpi->td.counts;
1265  vpx_writer header_bc;
1266
1267  vpx_start_encode(&header_bc, data);
1268
1269  if (xd->lossless)
1270    cm->tx_mode = ONLY_4X4;
1271  else
1272    encode_txfm_probs(cm, &header_bc, counts);
1273
1274  update_coef_probs(cpi, &header_bc);
1275  update_skip_probs(cm, &header_bc, counts);
1276
1277  if (!frame_is_intra_only(cm)) {
1278    int i;
1279
1280    for (i = 0; i < INTER_MODE_CONTEXTS; ++i)
1281      prob_diff_update(vp9_inter_mode_tree, cm->fc->inter_mode_probs[i],
1282                       counts->inter_mode[i], INTER_MODES, &header_bc);
1283
1284    if (cm->interp_filter == SWITCHABLE)
1285      update_switchable_interp_probs(cm, &header_bc, counts);
1286
1287    for (i = 0; i < INTRA_INTER_CONTEXTS; i++)
1288      vp9_cond_prob_diff_update(&header_bc, &fc->intra_inter_prob[i],
1289                                counts->intra_inter[i]);
1290
1291    if (cpi->allow_comp_inter_inter) {
1292      const int use_compound_pred = cm->reference_mode != SINGLE_REFERENCE;
1293      const int use_hybrid_pred = cm->reference_mode == REFERENCE_MODE_SELECT;
1294
1295      vpx_write_bit(&header_bc, use_compound_pred);
1296      if (use_compound_pred) {
1297        vpx_write_bit(&header_bc, use_hybrid_pred);
1298        if (use_hybrid_pred)
1299          for (i = 0; i < COMP_INTER_CONTEXTS; i++)
1300            vp9_cond_prob_diff_update(&header_bc, &fc->comp_inter_prob[i],
1301                                      counts->comp_inter[i]);
1302      }
1303    }
1304
1305    if (cm->reference_mode != COMPOUND_REFERENCE) {
1306      for (i = 0; i < REF_CONTEXTS; i++) {
1307        vp9_cond_prob_diff_update(&header_bc, &fc->single_ref_prob[i][0],
1308                                  counts->single_ref[i][0]);
1309        vp9_cond_prob_diff_update(&header_bc, &fc->single_ref_prob[i][1],
1310                                  counts->single_ref[i][1]);
1311      }
1312    }
1313
1314    if (cm->reference_mode != SINGLE_REFERENCE)
1315      for (i = 0; i < REF_CONTEXTS; i++)
1316        vp9_cond_prob_diff_update(&header_bc, &fc->comp_ref_prob[i],
1317                                  counts->comp_ref[i]);
1318
1319    for (i = 0; i < BLOCK_SIZE_GROUPS; ++i)
1320      prob_diff_update(vp9_intra_mode_tree, cm->fc->y_mode_prob[i],
1321                       counts->y_mode[i], INTRA_MODES, &header_bc);
1322
1323    for (i = 0; i < PARTITION_CONTEXTS; ++i)
1324      prob_diff_update(vp9_partition_tree, fc->partition_prob[i],
1325                       counts->partition[i], PARTITION_TYPES, &header_bc);
1326
1327    vp9_write_nmv_probs(cm, cm->allow_high_precision_mv, &header_bc,
1328                        &counts->mv);
1329  }
1330
1331  vpx_stop_encode(&header_bc);
1332  assert(header_bc.pos <= 0xffff);
1333
1334  return header_bc.pos;
1335}
1336
1337void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, size_t *size) {
1338  uint8_t *data = dest;
1339  size_t first_part_size, uncompressed_hdr_size;
1340  struct vpx_write_bit_buffer wb = { data, 0 };
1341  struct vpx_write_bit_buffer saved_wb;
1342
1343  write_uncompressed_header(cpi, &wb);
1344  saved_wb = wb;
1345  vpx_wb_write_literal(&wb, 0, 16);  // don't know in advance first part. size
1346
1347  uncompressed_hdr_size = vpx_wb_bytes_written(&wb);
1348  data += uncompressed_hdr_size;
1349
1350  vpx_clear_system_state();
1351
1352  first_part_size = write_compressed_header(cpi, data);
1353  data += first_part_size;
1354  // TODO(jbb): Figure out what to do if first_part_size > 16 bits.
1355  vpx_wb_write_literal(&saved_wb, (int)first_part_size, 16);
1356
1357  data += encode_tiles(cpi, data);
1358
1359  *size = data - dest;
1360}
1361