vp9_encodeframe.c 140 KB
Newer Older
John Koleszar's avatar
John Koleszar committed
1
/*
2
 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
John Koleszar's avatar
John Koleszar committed
3
 *
4
 *  Use of this source code is governed by a BSD-style license
5
6
 *  that can be found in the LICENSE file in the root of the source
 *  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.
John Koleszar's avatar
John Koleszar committed
9
10
 */

Dmitry Kovalev's avatar
Dmitry Kovalev committed
11
12
13
14
#include <limits.h>
#include <math.h>
#include <stdio.h>

Jim Bankoski's avatar
Jim Bankoski committed
15
#include "./vp9_rtcd.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
16
17
18
19
#include "./vpx_config.h"

#include "vpx_ports/vpx_timer.h"

20
#include "vp9/common/vp9_common.h"
Yaowu Xu's avatar
Yaowu Xu committed
21
#include "vp9/common/vp9_entropy.h"
22
#include "vp9/common/vp9_entropymode.h"
23
#include "vp9/common/vp9_idct.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
24
25
26
#include "vp9/common/vp9_mvref_common.h"
#include "vp9/common/vp9_pred_common.h"
#include "vp9/common/vp9_quant_common.h"
27
#include "vp9/common/vp9_reconintra.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
28
#include "vp9/common/vp9_reconinter.h"
29
#include "vp9/common/vp9_seg_common.h"
30
#include "vp9/common/vp9_systemdependent.h"
31
#include "vp9/common/vp9_tile_common.h"
32

33
#include "vp9/encoder/vp9_aq_complexity.h"
Marco Paniconi's avatar
Marco Paniconi committed
34
35
#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
#include "vp9/encoder/vp9_aq_variance.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
36
37
38
#include "vp9/encoder/vp9_encodeframe.h"
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_encodemv.h"
39
#include "vp9/encoder/vp9_extend.h"
40
#include "vp9/encoder/vp9_pickmode.h"
41
#include "vp9/encoder/vp9_rd.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
42
43
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_segmentation.h"
44
#include "vp9/encoder/vp9_tokenize.h"
45

46
47
48
49
50
51
#define GF_ZEROMV_ZBIN_BOOST 0
#define LF_ZEROMV_ZBIN_BOOST 0
#define MV_ZBIN_BOOST        0
#define SPLIT_MV_ZBIN_BOOST  0
#define INTRA_ZBIN_BOOST     0

Jim Bankoski's avatar
Jim Bankoski committed
52
static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
53
54
                              int mi_row, int mi_col, BLOCK_SIZE bsize,
                              PICK_MODE_CONTEXT *ctx);
55

56
// Motion vector component magnitude threshold for defining fast motion.
57
#define FAST_MOTION_MV_THRESH 24
58

59
60
61
62
// This is used as a reference when computing the source variance for the
//  purposes of activity masking.
// Eventually this should be replaced by custom no-reference routines,
//  which will be faster.
63
static const uint8_t VP9_VAR_OFFS[64] = {
64
65
66
67
68
69
70
71
    128, 128, 128, 128, 128, 128, 128, 128,
    128, 128, 128, 128, 128, 128, 128, 128,
    128, 128, 128, 128, 128, 128, 128, 128,
    128, 128, 128, 128, 128, 128, 128, 128,
    128, 128, 128, 128, 128, 128, 128, 128,
    128, 128, 128, 128, 128, 128, 128, 128,
    128, 128, 128, 128, 128, 128, 128, 128,
    128, 128, 128, 128, 128, 128, 128, 128
72
73
};

74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
#if CONFIG_VP9_HIGHBITDEPTH
static const uint16_t VP9_HIGH_VAR_OFFS_8[64] = {
    128, 128, 128, 128, 128, 128, 128, 128,
    128, 128, 128, 128, 128, 128, 128, 128,
    128, 128, 128, 128, 128, 128, 128, 128,
    128, 128, 128, 128, 128, 128, 128, 128,
    128, 128, 128, 128, 128, 128, 128, 128,
    128, 128, 128, 128, 128, 128, 128, 128,
    128, 128, 128, 128, 128, 128, 128, 128,
    128, 128, 128, 128, 128, 128, 128, 128
};

static const uint16_t VP9_HIGH_VAR_OFFS_10[64] = {
    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
    128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4
};

static const uint16_t VP9_HIGH_VAR_OFFS_12[64] = {
    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
    128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16
};
#endif  // CONFIG_VP9_HIGHBITDEPTH

109
static unsigned int get_sby_perpixel_variance(VP9_COMP *cpi,
110
                                              const struct buf_2d *ref,
111
                                              BLOCK_SIZE bs) {
112
113
114
  unsigned int sse;
  const unsigned int var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
                                              VP9_VAR_OFFS, 0, &sse);
115
  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
116
117
}

118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
#if CONFIG_VP9_HIGHBITDEPTH
static unsigned int high_get_sby_perpixel_variance(
    VP9_COMP *cpi, const struct buf_2d *ref, BLOCK_SIZE bs, int bd) {
  unsigned int var, sse;
  switch (bd) {
    case 10:
      var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
                               CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10),
                               0, &sse);
      break;
    case 12:
      var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
                               CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12),
                               0, &sse);
      break;
    case 8:
    default:
      var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
                               CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8),
                               0, &sse);
      break;
  }
  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
}
#endif  // CONFIG_VP9_HIGHBITDEPTH

144
static unsigned int get_sby_perpixel_diff_variance(VP9_COMP *cpi,
145
146
                                                   const struct buf_2d *ref,
                                                   int mi_row, int mi_col,
147
                                                   BLOCK_SIZE bs) {
148
149
150
151
152
153
  const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);
  const uint8_t* last_y = &last->y_buffer[mi_row * MI_SIZE * last->y_stride +
                                              mi_col * MI_SIZE];
  unsigned int sse;
  const unsigned int var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
                                              last_y, last->y_stride, &sse);
154
155
156
157
158
159
  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
}

static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi,
                                                   int mi_row,
                                                   int mi_col) {
160
  unsigned int var = get_sby_perpixel_diff_variance(cpi, &cpi->mb.plane[0].src,
161
162
163
                                                    mi_row, mi_col,
                                                    BLOCK_64X64);
  if (var < 8)
164
    return BLOCK_64X64;
165
  else if (var < 128)
166
    return BLOCK_32X32;
167
168
169
170
  else if (var < 2048)
    return BLOCK_16X16;
  else
    return BLOCK_8X8;
171
172
}

173
174
175
static BLOCK_SIZE get_nonrd_var_based_fixed_partition(VP9_COMP *cpi,
                                                      int mi_row,
                                                      int mi_col) {
176
  unsigned int var = get_sby_perpixel_diff_variance(cpi, &cpi->mb.plane[0].src,
177
178
                                                    mi_row, mi_col,
                                                    BLOCK_64X64);
179
  if (var < 4)
180
    return BLOCK_64X64;
181
  else if (var < 10)
182
183
    return BLOCK_32X32;
  else
184
    return BLOCK_16X16;
185
186
}

187
188
// Lighter version of set_offsets that only sets the mode info
// pointers.
189
static INLINE void set_modeinfo_offsets(VP9_COMMON *const cm,
190
191
192
                                        MACROBLOCKD *const xd,
                                        int mi_row,
                                        int mi_col) {
193
  const int idx_str = xd->mi_stride * mi_row + mi_col;
hkuang's avatar
hkuang committed
194
195
  xd->mi = cm->mi + idx_str;
  xd->mi[0].src_mi = &xd->mi[0];
196
197
198
199
200
201
202
203
204
205
206
207
}

static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
                        int mi_row, int mi_col, BLOCK_SIZE bsize) {
  MACROBLOCK *const x = &cpi->mb;
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *mbmi;
  const int mi_width = num_8x8_blocks_wide_lookup[bsize];
  const int mi_height = num_8x8_blocks_high_lookup[bsize];
  const struct segmentation *const seg = &cm->seg;

208
  set_skip_context(xd, mi_row, mi_col);
209
210
211

  set_modeinfo_offsets(cm, xd, mi_row, mi_col);

hkuang's avatar
hkuang committed
212
  mbmi = &xd->mi[0].src_mi->mbmi;
213
214

  // Set up destination pointers.
215
  vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232

  // Set up limit values for MV components.
  // Mv beyond the range do not produce new/different prediction block.
  x->mv_row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
  x->mv_col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
  x->mv_row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
  x->mv_col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;

  // Set up distance of MB to edge of frame in 1/8th pel units.
  assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
  set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width,
                 cm->mi_rows, cm->mi_cols);

  // Set up source buffers.
  vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);

  // R/D setup.
Dmitry Kovalev's avatar
Dmitry Kovalev committed
233
234
  x->rddiv = cpi->rd.RDDIV;
  x->rdmult = cpi->rd.RDMULT;
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251

  // Setup segment ID.
  if (seg->enabled) {
    if (cpi->oxcf.aq_mode != VARIANCE_AQ) {
      const uint8_t *const map = seg->update_map ? cpi->segmentation_map
                                                 : cm->last_frame_seg_map;
      mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
    }
    vp9_init_plane_quantizers(cpi, x);

    x->encode_breakout = cpi->segment_encode_breakout[mbmi->segment_id];
  } else {
    mbmi->segment_id = 0;
    x->encode_breakout = cpi->encode_breakout;
  }
}

252
253
254
static void duplicate_mode_info_in_sb(VP9_COMMON *cm, MACROBLOCKD *xd,
                                      int mi_row, int mi_col,
                                      BLOCK_SIZE bsize) {
255
256
257
258
259
260
  const int block_width = num_8x8_blocks_wide_lookup[bsize];
  const int block_height = num_8x8_blocks_high_lookup[bsize];
  int i, j;
  for (j = 0; j < block_height; ++j)
    for (i = 0; i < block_width; ++i) {
      if (mi_row + j < cm->mi_rows && mi_col + i < cm->mi_cols)
hkuang's avatar
hkuang committed
261
        xd->mi[j * xd->mi_stride + i].src_mi = &xd->mi[0];
262
263
264
265
266
267
268
269
270
    }
}

static void set_block_size(VP9_COMP * const cpi,
                           int mi_row, int mi_col,
                           BLOCK_SIZE bsize) {
  if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
    MACROBLOCKD *const xd = &cpi->mb.e_mbd;
    set_modeinfo_offsets(&cpi->common, xd, mi_row, mi_col);
hkuang's avatar
hkuang committed
271
    xd->mi[0].src_mi->mbmi.sb_type = bsize;
272
    duplicate_mode_info_in_sb(&cpi->common, xd, mi_row, mi_col, bsize);
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
  }
}

typedef struct {
  int64_t sum_square_error;
  int64_t sum_error;
  int count;
  int variance;
} var;

typedef struct {
  var none;
  var horz[2];
  var vert[2];
} partition_variance;

typedef struct {
  partition_variance part_variances;
  var split[4];
} v8x8;

typedef struct {
  partition_variance part_variances;
  v8x8 split[4];
} v16x16;

typedef struct {
  partition_variance part_variances;
  v16x16 split[4];
} v32x32;

typedef struct {
  partition_variance part_variances;
  v32x32 split[4];
} v64x64;

typedef struct {
  partition_variance *part_variances;
  var *split[4];
} variance_node;

typedef enum {
  V16X16,
  V32X32,
  V64X64,
} TREE_LEVEL;

static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
  int i;
322
323
  node->part_variances = NULL;
  vpx_memset(node->split, 0, sizeof(node->split));
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
  switch (bsize) {
    case BLOCK_64X64: {
      v64x64 *vt = (v64x64 *) data;
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++)
        node->split[i] = &vt->split[i].part_variances.none;
      break;
    }
    case BLOCK_32X32: {
      v32x32 *vt = (v32x32 *) data;
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++)
        node->split[i] = &vt->split[i].part_variances.none;
      break;
    }
    case BLOCK_16X16: {
      v16x16 *vt = (v16x16 *) data;
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++)
        node->split[i] = &vt->split[i].part_variances.none;
      break;
    }
    case BLOCK_8X8: {
      v8x8 *vt = (v8x8 *) data;
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++)
        node->split[i] = &vt->split[i];
      break;
    }
    default: {
      assert(0);
355
      break;
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
    }
  }
}

// Set variance values given sum square error, sum error, count.
static void fill_variance(int64_t s2, int64_t s, int c, var *v) {
  v->sum_square_error = s2;
  v->sum_error = s;
  v->count = c;
  if (c > 0)
    v->variance = (int)(256 *
                        (v->sum_square_error - v->sum_error * v->sum_error /
                         v->count) / v->count);
  else
    v->variance = 0;
}

void sum_2_variances(const var *a, const var *b, var *r) {
  fill_variance(a->sum_square_error + b->sum_square_error,
                a->sum_error + b->sum_error, a->count + b->count, r);
}

static void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
  variance_node node;
  tree_to_node(data, bsize, &node);
  sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
  sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
  sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
  sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
  sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
                  &node.part_variances->none);
}

static int set_vt_partitioning(VP9_COMP *cpi,
                               void *data,
                               BLOCK_SIZE bsize,
                               int mi_row,
Yaowu Xu's avatar
Yaowu Xu committed
393
                               int mi_col) {
394
395
396
397
398
  VP9_COMMON * const cm = &cpi->common;
  variance_node vt;
  const int block_width = num_8x8_blocks_wide_lookup[bsize];
  const int block_height = num_8x8_blocks_high_lookup[bsize];
  // TODO(debargha): Choose this more intelligently.
399
400
401
402
  const int threshold_multiplier = cm->frame_type == KEY_FRAME ? 64 : 4;
  int64_t threshold =
      (int64_t)(threshold_multiplier *
                vp9_convert_qindex_to_q(cm->base_qindex, cm->bit_depth));
403
404
405
406
407
408
409
410
  assert(block_height == block_width);
  tree_to_node(data, bsize, &vt);

  // Split none is available only if we have more than half a block size
  // in width and height inside the visible image.
  if (mi_col + block_width / 2 < cm->mi_cols &&
      mi_row + block_height / 2 < cm->mi_rows &&
      vt.part_variances->none.variance < threshold) {
Yaowu Xu's avatar
Yaowu Xu committed
411
    set_block_size(cpi, mi_row, mi_col, bsize);
412
413
414
    return 1;
  }

415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
  // Only allow split for blocks above 16x16.
  if (bsize > BLOCK_16X16) {
    // Vertical split is available on all but the bottom border.
    if (mi_row + block_height / 2 < cm->mi_rows &&
        vt.part_variances->vert[0].variance < threshold &&
        vt.part_variances->vert[1].variance < threshold) {
      BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
      set_block_size(cpi, mi_row, mi_col, subsize);
      set_block_size(cpi, mi_row, mi_col + block_width / 2, subsize);
      return 1;
    }

    // Horizontal split is available on all but the right border.
    if (mi_col + block_width / 2 < cm->mi_cols &&
        vt.part_variances->horz[0].variance < threshold &&
        vt.part_variances->horz[1].variance < threshold) {
      BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
      set_block_size(cpi, mi_row, mi_col, subsize);
      set_block_size(cpi, mi_row + block_height / 2, mi_col, subsize);
      return 1;
    }
436
437
  }

438
439
440
441
442
443
444
445
  // This will only allow 8x8 if the 16x16 variance is very large.
  if (bsize == BLOCK_16X16) {
    if (mi_col + block_width / 2 < cm->mi_cols &&
        mi_row + block_height / 2 < cm->mi_rows &&
        vt.part_variances->none.variance < (threshold << 6)) {
      set_block_size(cpi, mi_row, mi_col, bsize);
      return 1;
    }
446
447
448
449
  }
  return 0;
}

450
451
452
453
454
455
// This function chooses partitioning based on the variance
// between source and reconstructed last, where variance is
// computed for 8x8 downsampled inputs. Some things to check:
// using the last source rather than reconstructed last, and
// allowing for small downsampling (4x4 or 2x2) for selection
// of smaller block sizes (i.e., < 16x16).
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
static void choose_partitioning(VP9_COMP *cpi,
                                const TileInfo *const tile,
                                int mi_row, int mi_col) {
  VP9_COMMON * const cm = &cpi->common;
  MACROBLOCK *x = &cpi->mb;
  MACROBLOCKD *xd = &cpi->mb.e_mbd;

  int i, j, k;
  v64x64 vt;
  uint8_t *s;
  const uint8_t *d;
  int sp;
  int dp;
  int pixels_wide = 64, pixels_high = 64;
  int_mv nearest_mv, near_mv;
  const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
  const struct scale_factors *const sf = &cm->frame_refs[LAST_FRAME - 1].sf;

474
  vp9_clear_system_state();
475
476
477
478
479
480
481
482
483
484
485
486
487
488
  vp9_zero(vt);
  set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);

  if (xd->mb_to_right_edge < 0)
    pixels_wide += (xd->mb_to_right_edge >> 3);
  if (xd->mb_to_bottom_edge < 0)
    pixels_high += (xd->mb_to_bottom_edge >> 3);

  s = x->plane[0].src.buf;
  sp = x->plane[0].src.stride;

  if (cm->frame_type != KEY_FRAME) {
    vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col, sf);

hkuang's avatar
hkuang committed
489
490
    xd->mi[0].src_mi->mbmi.ref_frame[0] = LAST_FRAME;
    xd->mi[0].src_mi->mbmi.sb_type = BLOCK_64X64;
491
    vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv,
hkuang's avatar
hkuang committed
492
                          xd->mi[0].src_mi->mbmi.ref_mvs[LAST_FRAME],
493
494
                          &nearest_mv, &near_mv);

hkuang's avatar
hkuang committed
495
    xd->mi[0].src_mi->mbmi.mv[0] = nearest_mv;
496
497
498
499
500
501
502
    vp9_build_inter_predictors_sby(xd, mi_row, mi_col, BLOCK_64X64);

    d = xd->plane[0].dst.buf;
    dp = xd->plane[0].dst.stride;
  } else {
    d = VP9_VAR_OFFS;
    dp = 0;
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
#if CONFIG_VP9_HIGHBITDEPTH
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
      switch (xd->bd) {
        case 10:
          d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10);
          break;
        case 12:
          d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12);
          break;
        case 8:
        default:
          d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8);
          break;
      }
    }
#endif  // CONFIG_VP9_HIGHBITDEPTH
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
  }

  // Fill in the entire tree of 8x8 variances for splits.
  for (i = 0; i < 4; i++) {
    const int x32_idx = ((i & 1) << 5);
    const int y32_idx = ((i >> 1) << 5);
    for (j = 0; j < 4; j++) {
      const int x16_idx = x32_idx + ((j & 1) << 4);
      const int y16_idx = y32_idx + ((j >> 1) << 4);
      v16x16 *vst = &vt.split[i].split[j];
      for (k = 0; k < 4; k++) {
        int x_idx = x16_idx + ((k & 1) << 3);
        int y_idx = y16_idx + ((k >> 1) << 3);
        unsigned int sse = 0;
        int sum = 0;
534
535

        if (x_idx < pixels_wide && y_idx < pixels_high) {
536
537
538
539
540
541
542
543
544
545
546
547
548
          int s_avg, d_avg;
#if CONFIG_VP9_HIGHBITDEPTH
          if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
            s_avg = vp9_highbd_avg_8x8(s + y_idx * sp + x_idx, sp);
            d_avg = vp9_highbd_avg_8x8(d + y_idx * dp + x_idx, dp);
          } else {
            s_avg = vp9_avg_8x8(s + y_idx * sp + x_idx, sp);
            d_avg = vp9_avg_8x8(d + y_idx * dp + x_idx, dp);
          }
#else
          s_avg = vp9_avg_8x8(s + y_idx * sp + x_idx, sp);
          d_avg = vp9_avg_8x8(d + y_idx * dp + x_idx, dp);
#endif
549
550
551
552
553
554
555
          sum = s_avg - d_avg;
          sse = sum * sum;
        }
        // For an 8x8 block we have just one value the average of all 64
        // pixels,  so use 1.   This means of course that there is no variance
        // in an 8x8 block.
        fill_variance(sse, sum, 1, &vst->split[k].part_variances.none);
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
      }
    }
  }
  // Fill the rest of the variance tree by summing split partition values.
  for (i = 0; i < 4; i++) {
    for (j = 0; j < 4; j++) {
      fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
    }
    fill_variance_tree(&vt.split[i], BLOCK_32X32);
  }
  fill_variance_tree(&vt, BLOCK_64X64);

  // Now go through the entire structure,  splitting every block size until
  // we get to one that's got a variance lower than our threshold,  or we
  // hit 8x8.
571
572
  if ( mi_col + 8 > cm->mi_cols || mi_row + 8 > cm->mi_rows ||
      !set_vt_partitioning(cpi, &vt, BLOCK_64X64, mi_row, mi_col)) {
573
574
575
    for (i = 0; i < 4; ++i) {
      const int x32_idx = ((i & 1) << 2);
      const int y32_idx = ((i >> 1) << 2);
Yaowu Xu's avatar
Yaowu Xu committed
576
577
      if (!set_vt_partitioning(cpi, &vt.split[i], BLOCK_32X32,
                               (mi_row + y32_idx), (mi_col + x32_idx))) {
578
579
580
        for (j = 0; j < 4; ++j) {
          const int x16_idx = ((j & 1) << 1);
          const int y16_idx = ((j >> 1) << 1);
581
582
583
584
585
          // NOTE: Since this uses 8x8 downsampling for variance calculation
          // we cannot really select block size 8x8 (or even 8x16/16x8),
          // since we do not sufficient samples for variance.
          // For now, 8x8 partition is only set if the variance of the 16x16
          // block is very high. This is controlled in set_vt_partitioning.
586
          if (!set_vt_partitioning(cpi, &vt.split[i].split[j],
587
                                   BLOCK_16X16,
588
589
                                   mi_row + y32_idx + y16_idx,
                                   mi_col + x32_idx + x16_idx)) {
590
591
592
            for (k = 0; k < 4; ++k) {
              const int x8_idx = (k & 1);
              const int y8_idx = (k >> 1);
Yaowu Xu's avatar
Yaowu Xu committed
593
              set_block_size(cpi,
594
595
596
597
598
599
600
601
602
603
604
                             (mi_row + y32_idx + y16_idx + y8_idx),
                             (mi_col + x32_idx + x16_idx + x8_idx),
                             BLOCK_8X8);
            }
          }
        }
      }
    }
  }
}

Jim Bankoski's avatar
Jim Bankoski committed
605
static void update_state(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
606
607
                         int mi_row, int mi_col, BLOCK_SIZE bsize,
                         int output_enabled) {
Ronald S. Bultje's avatar
Ronald S. Bultje committed
608
  int i, x_idx, y;
609
  VP9_COMMON *const cm = &cpi->common;
Dmitry Kovalev's avatar
Dmitry Kovalev committed
610
  RD_OPT *const rd_opt = &cpi->rd;
611
612
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
613
614
  struct macroblock_plane *const p = x->plane;
  struct macroblockd_plane *const pd = xd->plane;
John Koleszar's avatar
John Koleszar committed
615
  MODE_INFO *mi = &ctx->mic;
hkuang's avatar
hkuang committed
616
617
  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
  MODE_INFO *mi_addr = &xd->mi[0];
618
  const struct segmentation *const seg = &cm->seg;
619

620
  const int mis = cm->mi_stride;
Jim Bankoski's avatar
Jim Bankoski committed
621
  const int mi_width = num_8x8_blocks_wide_lookup[bsize];
Jim Bankoski's avatar
Jim Bankoski committed
622
  const int mi_height = num_8x8_blocks_high_lookup[bsize];
623
  int max_plane;
Adrian Grange's avatar
Adrian Grange committed
624

625
  assert(mi->mbmi.sb_type == bsize);
626

627
  *mi_addr = *mi;
hkuang's avatar
hkuang committed
628
  mi_addr->src_mi = mi_addr;
629

Paul Wilkins's avatar
Paul Wilkins committed
630
631
632
633
634
635
636
637
638
639
640
  // If segmentation in use
  if (seg->enabled && output_enabled) {
    // For in frame complexity AQ copy the segment id from the segment map.
    if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
      const uint8_t *const map = seg->update_map ? cpi->segmentation_map
                                                 : cm->last_frame_seg_map;
      mi_addr->mbmi.segment_id =
        vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
    }
    // Else for cyclic refresh mode update the segment map, set the segment id
    // and then update the quantizer.
641
    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
hkuang's avatar
hkuang committed
642
      vp9_cyclic_refresh_update_segment(cpi, &xd->mi[0].src_mi->mbmi,
Paul Wilkins's avatar
Paul Wilkins committed
643
644
                                        mi_row, mi_col, bsize, 1);
    }
645
  }
646

647
648
  max_plane = is_inter_block(mbmi) ? MAX_MB_PLANE : 1;
  for (i = 0; i < max_plane; ++i) {
649
    p[i].coeff = ctx->coeff_pbuf[i][1];
650
    p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
651
    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
652
    p[i].eobs = ctx->eobs_pbuf[i][1];
653
654
  }

655
656
  for (i = max_plane; i < MAX_MB_PLANE; ++i) {
    p[i].coeff = ctx->coeff_pbuf[i][2];
657
    p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
658
    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
659
    p[i].eobs = ctx->eobs_pbuf[i][2];
660
661
  }

John Koleszar's avatar
John Koleszar committed
662
663
  // Restore the coding context of the MB to that that was in place
  // when the mode was picked for it
664
665
  for (y = 0; y < mi_height; y++)
    for (x_idx = 0; x_idx < mi_width; x_idx++)
James Zern's avatar
James Zern committed
666
      if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx
667
        && (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
hkuang's avatar
hkuang committed
668
        xd->mi[x_idx + y * mis].src_mi = mi_addr;
669
      }
670

671
  if (cpi->oxcf.aq_mode)
Dmitry Kovalev's avatar
Dmitry Kovalev committed
672
    vp9_init_plane_quantizers(cpi, x);
673

674
675
  // FIXME(rbultje) I'm pretty sure this should go to the end of this block
  // (i.e. after the output_enabled)
676
677
  if (bsize < BLOCK_32X32) {
    if (bsize < BLOCK_16X16)
678
679
      ctx->tx_rd_diff[ALLOW_16X16] = ctx->tx_rd_diff[ALLOW_8X8];
    ctx->tx_rd_diff[ALLOW_32X32] = ctx->tx_rd_diff[ALLOW_16X16];
Ronald S. Bultje's avatar
Ronald S. Bultje committed
680
  }
Adrian Grange's avatar
Adrian Grange committed
681

682
  if (is_inter_block(mbmi) && mbmi->sb_type < BLOCK_8X8) {
683
684
    mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
    mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
John Koleszar's avatar
John Koleszar committed
685
686
  }

687
  x->skip = ctx->skip;
688
  vpx_memcpy(x->zcoeff_blk[mbmi->tx_size], ctx->zcoeff_blk,
689
             sizeof(uint8_t) * ctx->num_4x4_blk);
690

Ronald S. Bultje's avatar
Ronald S. Bultje committed
691
692
693
  if (!output_enabled)
    return;

694
  if (!vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
695
    for (i = 0; i < TX_MODES; i++)
Dmitry Kovalev's avatar
Dmitry Kovalev committed
696
      rd_opt->tx_select_diff[i] += ctx->tx_rd_diff[i];
697
698
  }

699
#if CONFIG_INTERNAL_STATS
700
  if (frame_is_intra_only(cm)) {
John Koleszar's avatar
John Koleszar committed
701
    static const int kf_mode_index[] = {
702
703
704
705
      THR_DC        /*DC_PRED*/,
      THR_V_PRED    /*V_PRED*/,
      THR_H_PRED    /*H_PRED*/,
      THR_D45_PRED  /*D45_PRED*/,
John Koleszar's avatar
John Koleszar committed
706
707
708
      THR_D135_PRED /*D135_PRED*/,
      THR_D117_PRED /*D117_PRED*/,
      THR_D153_PRED /*D153_PRED*/,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
709
      THR_D207_PRED /*D207_PRED*/,
710
711
      THR_D63_PRED  /*D63_PRED*/,
      THR_TM        /*TM_PRED*/,
John Koleszar's avatar
John Koleszar committed
712
    };
713
    ++cpi->mode_chosen_counts[kf_mode_index[mbmi->mode]];
John Koleszar's avatar
John Koleszar committed
714
715
  } else {
    // Note how often each mode chosen as best
716
717
718
719
    ++cpi->mode_chosen_counts[ctx->best_mode_index];
  }
#endif
  if (!frame_is_intra_only(cm)) {
720
    if (is_inter_block(mbmi)) {
721
      vp9_update_mv_count(cm, xd);
722
723
724
725
726

      if (cm->interp_filter == SWITCHABLE) {
        const int ctx = vp9_get_pred_context_switchable_interp(xd);
        ++cm->counts.switchable_interp[ctx][mbmi->interp_filter];
      }
727
    }
Adrian Grange's avatar
Adrian Grange committed
728

Dmitry Kovalev's avatar
Dmitry Kovalev committed
729
730
731
    rd_opt->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
    rd_opt->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
    rd_opt->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
732

733
    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
Dmitry Kovalev's avatar
Dmitry Kovalev committed
734
      rd_opt->filter_diff[i] += ctx->best_filter_diff[i];
John Koleszar's avatar
John Koleszar committed
735
  }
Adrian Grange's avatar
Adrian Grange committed
736
737
}

Jim Bankoski's avatar
Jim Bankoski committed
738
void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
739
                          int mi_row, int mi_col) {
740
741
  uint8_t *const buffers[3] = {src->y_buffer, src->u_buffer, src->v_buffer };
  const int strides[3] = {src->y_stride, src->uv_stride, src->uv_stride };
742
743
  int i;

744
745
746
  // Set current frame pointer.
  x->e_mbd.cur_buf = src;

747
748
  for (i = 0; i < MAX_MB_PLANE; i++)
    setup_pred_plane(&x->plane[i].src, buffers[i], strides[i], mi_row, mi_col,
Jim Bankoski's avatar
Jim Bankoski committed
749
                     NULL, x->e_mbd.plane[i].subsampling_x,
750
                     x->e_mbd.plane[i].subsampling_y);
John Koleszar's avatar
John Koleszar committed
751
752
}

753
754
static void set_mode_info_seg_skip(MACROBLOCK *x, TX_MODE tx_mode,
                                   RD_COST *rd_cost, BLOCK_SIZE bsize) {
755
  MACROBLOCKD *const xd = &x->e_mbd;
hkuang's avatar
hkuang committed
756
  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
757
758
759
  INTERP_FILTER filter_ref;

  if (xd->up_available)
hkuang's avatar
hkuang committed
760
    filter_ref = xd->mi[-xd->mi_stride].src_mi->mbmi.interp_filter;
761
  else if (xd->left_available)
hkuang's avatar
hkuang committed
762
    filter_ref = xd->mi[-1].src_mi->mbmi.interp_filter;
763
764
765
766
767
768
769
770
771
772
773
774
775
776
  else
    filter_ref = EIGHTTAP;

  mbmi->sb_type = bsize;
  mbmi->mode = ZEROMV;
  mbmi->tx_size = MIN(max_txsize_lookup[bsize],
                      tx_mode_to_biggest_tx_size[tx_mode]);
  mbmi->skip = 1;
  mbmi->uv_mode = DC_PRED;
  mbmi->ref_frame[0] = LAST_FRAME;
  mbmi->ref_frame[1] = NONE;
  mbmi->mv[0].as_int = 0;
  mbmi->interp_filter = filter_ref;

hkuang's avatar
hkuang committed
777
  xd->mi[0].src_mi->bmi[0].as_mv[0].as_int = 0;
778
779
  x->skip = 1;

780
  vp9_rd_cost_init(rd_cost);
781
782
}

783
static void rd_pick_sb_modes(VP9_COMP *cpi,
784
                             TileDataEnc *tile_data,
785
                             int mi_row, int mi_col, RD_COST *rd_cost,
786
                             BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
787
                             int64_t best_rd) {
788
  VP9_COMMON *const cm = &cpi->common;
789
  TileInfo *const tile_info = &tile_data->tile_info;
790
791
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
792
  MB_MODE_INFO *mbmi;
793
794
  struct macroblock_plane *const p = x->plane;
  struct macroblockd_plane *const pd = xd->plane;
795
  const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
796
  int i, orig_rdmult;
797
798
  double rdmult_ratio;

799
  vp9_clear_system_state();
800
  rdmult_ratio = 1.0;  // avoid uninitialized warnings
Ronald S. Bultje's avatar
Ronald S. Bultje committed
801

802
  // Use the lower precision, but faster, 32x32 fdct for mode selection.
803
  x->use_lp32x32fdct = 1;
804

805
  set_offsets(cpi, tile_info, mi_row, mi_col, bsize);
hkuang's avatar
hkuang committed
806
  mbmi = &xd->mi[0].src_mi->mbmi;
807
  mbmi->sb_type = bsize;
808

809
  for (i = 0; i < MAX_MB_PLANE; ++i) {
810
    p[i].coeff = ctx->coeff_pbuf[i][0];
811
    p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
812
    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
813
    p[i].eobs = ctx->eobs_pbuf[i][0];
814
  }
815
  ctx->is_coded = 0;
816
  ctx->skippable = 0;
817
  ctx->pred_pixel_ready = 0;
818
  x->skip_recode = 0;
819

820
  // Set to zero to make sure we do not use the previous encoded frame stats
821
  mbmi->skip = 0;
822

823
824
825
826
827
828
829
830
831
#if CONFIG_VP9_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    x->source_variance =
        high_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize, xd->bd);
  } else {
    x->source_variance =
        get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
  }
#else
832
  x->source_variance = get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
833
#endif  // CONFIG_VP9_HIGHBITDEPTH
834

835
836
837
  // Save rdmult before it might be changed, so it can be restored later.
  orig_rdmult = x->rdmult;

838
  if (aq_mode == VARIANCE_AQ) {
839
840
    const int energy = bsize <= BLOCK_16X16 ? x->mb_energy
                                            : vp9_block_energy(cpi, x, bsize);
Paul Wilkins's avatar
Paul Wilkins committed
841
842
843
    if (cm->frame_type == KEY_FRAME ||
        cpi->refresh_alt_ref_frame ||
        (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
844
      mbmi->segment_id = vp9_vaq_segment_id(energy);
Paul Wilkins's avatar
Paul Wilkins committed
845
846
847
    } else {
      const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
                                                    : cm->last_frame_seg_map;
848
      mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
Paul Wilkins's avatar
Paul Wilkins committed
849
850
    }

851
    rdmult_ratio = vp9_vaq_rdmult_ratio(energy);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
852
    vp9_init_plane_quantizers(cpi, x);
853
    vp9_clear_system_state();
854
    x->rdmult = (int)round(x->rdmult * rdmult_ratio);
855
  } else if (aq_mode == COMPLEXITY_AQ) {
856
857
    const int mi_offset = mi_row * cm->mi_cols + mi_col;
    unsigned char complexity = cpi->complexity_map[mi_offset];
Paul Wilkins's avatar
Paul Wilkins committed
858
859
    const int is_edge = (mi_row <= 1) || (mi_row >= (cm->mi_rows - 2)) ||
                        (mi_col <= 1) || (mi_col >= (cm->mi_cols - 2));
860
861
    if (!is_edge && (complexity > 128))
      x->rdmult += ((x->rdmult * (complexity - 128)) / 256);
862
  } else if (aq_mode == CYCLIC_REFRESH_AQ) {
863
    const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
864
                                                  : cm->last_frame_seg_map;
865
866
    // If segment 1, use rdmult for that segment.
    if (vp9_get_segment_id(cm, map, bsize, mi_row, mi_col))
867
      x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
868
  }
869

870
871
  // Find best coding mode & reconstruct the MB so it is available
  // as a predictor for MBs that follow in the SB
872
  if (frame_is_intra_only(cm)) {
873
    vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd);
874
  } else {
875
876
    if (bsize >= BLOCK_8X8) {
      if (vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
877
        vp9_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, rd_cost, bsize,
878
                                           ctx, best_rd);
879
      else
880
        vp9_rd_pick_inter_mode_sb(cpi, tile_data, x, mi_row, mi_col,
881
                                  rd_cost, bsize, ctx, best_rd);
882
    } else {
883
      vp9_rd_pick_inter_mode_sub8x8(cpi, tile_data, x, mi_row, mi_col,
884
                                    rd_cost, bsize, ctx, best_rd);
885
    }
886
  }
887

888
  if (aq_mode == VARIANCE_AQ && rd_cost->rate != INT_MAX) {
889
    vp9_clear_system_state();
890
891
    rd_cost->rate = (int)round(rd_cost->rate * rdmult_ratio);
    rd_cost->rdcost = RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
Paul Wilkins's avatar
Paul Wilkins committed
892
  }
893
894
895
896
897
898
899

  x->rdmult = orig_rdmult;

  // TODO(jingning) The rate-distortion optimization flow needs to be
  // refactored to provide proper exit/return handle.
  if (rd_cost->rate == INT_MAX)
    rd_cost->rdcost = INT64_MAX;
Ronald S. Bultje's avatar
Ronald S. Bultje committed
900
}
Adrian Grange's avatar
Adrian Grange committed
901

902
static void update_stats(VP9_COMMON *cm, const MACROBLOCK *x) {
903
  const MACROBLOCKD *const xd = &x->e_mbd;
hkuang's avatar
hkuang committed
904
  const MODE_INFO *const mi = xd->mi[0].src_mi;
905
  const MB_MODE_INFO *const mbmi = &mi->mbmi;
Adrian Grange's avatar
Adrian Grange committed
906

907
  if (!frame_is_intra_only(cm)) {
908
    const int seg_ref_active = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
909
                                                     SEG_LVL_REF_FRAME);
910
911
912
913
914
915
916
917
918
919
920
    if (!seg_ref_active) {
      FRAME_COUNTS *const counts = &cm->counts;
      const int inter_block = is_inter_block(mbmi);

      counts->intra_inter[vp9_get_intra_inter_context(xd)][inter_block]++;

      // If the segment reference feature is enabled we have only a single
      // reference frame allowed for the segment so exclude it from
      // the reference frame counts used to work out probabilities.
      if (inter_block) {
        const MV_REFERENCE_FRAME ref0 = mbmi->ref_frame[0];
921

922
923
924
925
926
927
928
929
930
931
932
933
934
935
        if (cm->reference_mode == REFERENCE_MODE_SELECT)
          counts->comp_inter[vp9_get_reference_mode_context(cm, xd)]
                            [has_second_ref(mbmi)]++;

        if (has_second_ref(mbmi)) {
          counts->comp_ref[vp9_get_pred_context_comp_ref_p(cm, xd)]
                          [ref0 == GOLDEN_FRAME]++;
        } else {
          counts->single_ref[vp9_get_pred_context_single_ref_p1(xd)][0]
                            [ref0 != LAST_FRAME]++;
          if (ref0 != LAST_FRAME)
            counts->single_ref[vp9_get_pred_context_single_ref_p2(xd)][1]
                              [ref0 != GOLDEN_FRAME]++;
        }
Ronald S. Bultje's avatar
Ronald S. Bultje committed
936
      }
Ronald S. Bultje's avatar
Ronald S. Bultje committed
937
    }
Ronald S. Bultje's avatar
Ronald S. Bultje committed
938
939
  }
}
John Koleszar's avatar
John Koleszar committed
940

941
942
943
static void restore_context(VP9_COMP *cpi, int mi_row, int mi_col,
                            ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
                            ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
Jim Bankoski's avatar
Jim Bankoski committed
944
                            PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
945
                            BLOCK_SIZE bsize) {
Jim Bankoski's avatar
Jim Bankoski committed
946
947
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
948
  int p;
Jim Bankoski's avatar
Jim Bankoski committed
949
950
  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
Jim Bankoski's avatar
Jim Bankoski committed
951
952
  int mi_width = num_8x8_blocks_wide_lookup[bsize];
  int mi_height = num_8x8_blocks_high_lookup[bsize];
953
  for (p = 0; p < MAX_MB_PLANE; p++) {
Jim Bankoski's avatar
Jim Bankoski committed
954
    vpx_memcpy(
955
        xd->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
Jim Bankoski's avatar
Jim Bankoski committed
956
957
958
        a + num_4x4_blocks_wide * p,
        (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
        xd->plane[p].subsampling_x);
Jim Bankoski's avatar
Jim Bankoski committed
959
    vpx_memcpy(
960
        xd->left_context[p]
Jim Bankoski's avatar
Jim Bankoski committed
961
962
963
964
965
            + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
        l + num_4x4_blocks_high * p,
        (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
        xd->plane[p].subsampling_y);
  }
966
967
968
969
  vpx_memcpy(xd->above_seg_context + mi_col, sa,
             sizeof(*xd->above_seg_context) * mi_width);
  vpx_memcpy(xd->left_seg_context + (mi_row & MI_MASK), sl,
             sizeof(xd->left_seg_context[0]) * mi_height);
970
}
Jingning Han's avatar
Jingning Han committed
971

Jim Bankoski's avatar
Jim Bankoski committed
972
static void save_context(VP9_COMP *cpi, int mi_row, int mi_col,
Jim Bankoski's avatar
Jim Bankoski committed
973
974
975
                         ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
                         ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
                         PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
976
                         BLOCK_SIZE bsize) {
Jim Bankoski's avatar
Jim Bankoski committed
977
978
  const MACROBLOCK *const x = &cpi->mb;
  const MACROBLOCKD *const xd = &x->e_mbd;
Jim Bankoski's avatar
Jim Bankoski committed
979
  int p;
Jim Bankoski's avatar
Jim Bankoski committed
980
981
  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
Jim Bankoski's avatar
Jim Bankoski committed
982
983
  int mi_width = num_8x8_blocks_wide_lookup[bsize];
  int mi_height = num_8x8_blocks_high_lookup[bsize];
Jim Bankoski's avatar
Jim Bankoski committed
984
985
986

  // buffer the above/left context information of the block in search.
  for (p = 0; p < MAX_MB_PLANE; ++p) {
Jim Bankoski's avatar
Jim Bankoski committed
987
    vpx_memcpy(
Jim Bankoski's avatar
Jim Bankoski committed
988
        a + num_4x4_blocks_wide * p,
989
        xd->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
Jim Bankoski's avatar
Jim Bankoski committed
990
991
        (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
        xd->plane[p].subsampling_x);
Jim Bankoski's avatar
Jim Bankoski committed
992
    vpx_memcpy(
Jim Bankoski's avatar
Jim Bankoski committed
993
        l + num_4x4_blocks_high * p,
994
        xd->left_context[p]
Jim Bankoski's avatar
Jim Bankoski committed
995
996
997
998
            + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
        (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
        xd->plane[p].subsampling_y);
  }
999
1000
1001
1002
  vpx_memcpy(sa, xd->above_seg_context + mi_col,
             sizeof(*xd->above_seg_context) * mi_width);
  vpx_memcpy(sl, xd->left_seg_context + (mi_row & MI_MASK),
             sizeof(xd->left_seg_context[0]) * mi_height);
Jim Bankoski's avatar
Jim Bankoski committed
1003
}
1004

James Zern's avatar
James Zern committed
1005
1006
static void encode_b(VP9_COMP *cpi, const TileInfo *const tile,
                     TOKENEXTRA **tp, int mi_row, int mi_col,
1007
1008
                     int output_enabled, BLOCK_SIZE bsize,
                     PICK_MODE_CONTEXT *ctx) {
James Zern's avatar
James Zern committed
1009
  set_offsets(cpi, tile, mi_row, mi_col, bsize);
1010
1011
  update_state(cpi, ctx, mi_row, mi_col, bsize, output_enabled);
  encode_superblock(cpi, tp, output_enabled, mi_row, mi_col, bsize, ctx);
John Koleszar's avatar
John Koleszar committed
1012

1013
  if (output_enabled) {
1014
    update_stats(&cpi->common, &cpi->mb);
1015

1016
1017
    (*tp)->token = EOSB_TOKEN;
    (*tp)++;
John Koleszar's avatar
John Koleszar committed
1018
  }
Ronald S. Bultje's avatar
Ronald S. Bultje committed
1019
1020
}

James Zern's avatar
James Zern committed
1021
1022
static void encode_sb(VP9_COMP *cpi, const TileInfo *const tile,
                      TOKENEXTRA **tp, int mi_row, int mi_col,
1023
1024
                      int output_enabled, BLOCK_SIZE bsize,
                      PC_TREE *pc_tree) {
1025
1026
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
1027
1028
  MACROBLOCKD *const xd = &x->e_mbd;

1029
  const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
1030
  int ctx;
Jim Bankoski's avatar
Jim Bankoski committed
1031
  PARTITION_TYPE partition;
1032
  BLOCK_SIZE subsize = bsize;