vp9_encodeframe.c 184 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"
Johann's avatar
Johann committed
16
#include "./vpx_dsp_rtcd.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
17
18
#include "./vpx_config.h"

19
#include "vpx_dsp/vpx_dsp_common.h"
20
#include "vpx_ports/mem.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
21
#include "vpx_ports/vpx_timer.h"
22
#include "vpx_ports/system_state.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
23

24
#include "vp9/common/vp9_common.h"
Yaowu Xu's avatar
Yaowu Xu committed
25
#include "vp9/common/vp9_entropy.h"
26
#include "vp9/common/vp9_entropymode.h"
27
#include "vp9/common/vp9_idct.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
28
29
30
#include "vp9/common/vp9_mvref_common.h"
#include "vp9/common/vp9_pred_common.h"
#include "vp9/common/vp9_quant_common.h"
31
#include "vp9/common/vp9_reconintra.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
32
#include "vp9/common/vp9_reconinter.h"
33
#include "vp9/common/vp9_seg_common.h"
34
#include "vp9/common/vp9_tile_common.h"
35

36
#include "vp9/encoder/vp9_aq_360.h"
37
#include "vp9/encoder/vp9_aq_complexity.h"
Marco Paniconi's avatar
Marco Paniconi committed
38
39
#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
#include "vp9/encoder/vp9_aq_variance.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
40
41
42
#include "vp9/encoder/vp9_encodeframe.h"
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_encodemv.h"
43
#include "vp9/encoder/vp9_ethread.h"
44
#include "vp9/encoder/vp9_extend.h"
45
#include "vp9/encoder/vp9_pickmode.h"
46
#include "vp9/encoder/vp9_rd.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
47
48
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_segmentation.h"
49
#include "vp9/encoder/vp9_tokenize.h"
50

clang-format's avatar
clang-format committed
51
52
53
static void encode_superblock(VP9_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
                              int output_enabled, int mi_row, int mi_col,
                              BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx);
54

55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
// Machine learning-based early termination parameters.
static const double train_mean[24] = {
  303501.697372, 3042630.372158, 24.694696, 1.392182,
  689.413511,    162.027012,     1.478213,  0.0,
  135382.260230, 912738.513263,  28.845217, 1.515230,
  544.158492,    131.807995,     1.436863,  0.0,
  43682.377587,  208131.711766,  28.084737, 1.356677,
  138.254122,    119.522553,     1.252322,  0.0
};

static const double train_stdm[24] = {
  673689.212982, 5996652.516628, 0.024449, 1.989792,
  985.880847,    0.014638,       2.001898, 0.0,
  208798.775332, 1812548.443284, 0.018693, 1.838009,
  396.986910,    0.015657,       1.332541, 0.0,
  55888.847031,  448587.962714,  0.017900, 1.904776,
  98.652832,     0.016598,       1.320992, 0.0
};

// Error tolerance: 0.01%-0.0.05%-0.1%
static const double classifiers[24] = {
  0.111736, 0.289977, 0.042219, 0.204765, 0.120410, -0.143863,
  0.282376, 0.847811, 0.637161, 0.131570, 0.018636, 0.202134,
  0.112797, 0.028162, 0.182450, 1.124367, 0.386133, 0.083700,
  0.050028, 0.150873, 0.061119, 0.109318, 0.127255, 0.625211
};

82
// This is used as a reference when computing the source variance for the
83
//  purpose of activity masking.
84
85
// Eventually this should be replaced by custom no-reference routines,
//  which will be faster.
86
static const uint8_t VP9_VAR_OFFS[64] = {
clang-format's avatar
clang-format committed
87
88
89
90
91
  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
92
93
};

94
95
#if CONFIG_VP9_HIGHBITDEPTH
static const uint16_t VP9_HIGH_VAR_OFFS_8[64] = {
clang-format's avatar
clang-format committed
96
97
98
99
100
  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
101
102
103
};

static const uint16_t VP9_HIGH_VAR_OFFS_10[64] = {
clang-format's avatar
clang-format committed
104
105
106
107
108
109
110
111
  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
112
113
114
};

static const uint16_t VP9_HIGH_VAR_OFFS_12[64] = {
clang-format's avatar
clang-format committed
115
116
117
118
119
120
121
122
123
124
  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
125
126
127
};
#endif  // CONFIG_VP9_HIGHBITDEPTH

128
129
130
unsigned int vp9_get_sby_perpixel_variance(VP9_COMP *cpi,
                                           const struct buf_2d *ref,
                                           BLOCK_SIZE bs) {
131
  unsigned int sse;
clang-format's avatar
clang-format committed
132
133
  const unsigned int var =
      cpi->fn_ptr[bs].vf(ref->buf, ref->stride, VP9_VAR_OFFS, 0, &sse);
134
  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
135
136
}

137
#if CONFIG_VP9_HIGHBITDEPTH
clang-format's avatar
clang-format committed
138
139
140
unsigned int vp9_high_get_sby_perpixel_variance(VP9_COMP *cpi,
                                                const struct buf_2d *ref,
                                                BLOCK_SIZE bs, int bd) {
141
142
143
  unsigned int var, sse;
  switch (bd) {
    case 10:
clang-format's avatar
clang-format committed
144
145
146
      var =
          cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
                             CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10), 0, &sse);
147
148
      break;
    case 12:
clang-format's avatar
clang-format committed
149
150
151
      var =
          cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
                             CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12), 0, &sse);
152
153
154
      break;
    case 8:
    default:
clang-format's avatar
clang-format committed
155
156
157
      var =
          cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
                             CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8), 0, &sse);
158
159
      break;
  }
160
161
  return (unsigned int)ROUND64_POWER_OF_TWO((int64_t)var,
                                            num_pels_log2_lookup[bs]);
162
163
164
}
#endif  // CONFIG_VP9_HIGHBITDEPTH

165
static unsigned int get_sby_perpixel_diff_variance(VP9_COMP *cpi,
166
167
                                                   const struct buf_2d *ref,
                                                   int mi_row, int mi_col,
168
                                                   BLOCK_SIZE bs) {
169
170
  unsigned int sse, var;
  uint8_t *last_y;
171
  const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);
172
173
174
175
176

  assert(last != NULL);
  last_y =
      &last->y_buffer[mi_row * MI_SIZE * last->y_stride + mi_col * MI_SIZE];
  var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride, last_y, last->y_stride, &sse);
177
178
179
  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
}

180
static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi, MACROBLOCK *x,
clang-format's avatar
clang-format committed
181
182
183
                                                   int mi_row, int mi_col) {
  unsigned int var = get_sby_perpixel_diff_variance(
      cpi, &x->plane[0].src, mi_row, mi_col, BLOCK_64X64);
184
  if (var < 8)
185
    return BLOCK_64X64;
186
  else if (var < 128)
187
    return BLOCK_32X32;
188
189
190
191
  else if (var < 2048)
    return BLOCK_16X16;
  else
    return BLOCK_8X8;
192
193
}

194
195
// Lighter version of set_offsets that only sets the mode info
// pointers.
Jingning Han's avatar
Jingning Han committed
196
static INLINE void set_mode_info_offsets(VP9_COMMON *const cm,
197
                                         MACROBLOCK *const x,
clang-format's avatar
clang-format committed
198
                                         MACROBLOCKD *const xd, int mi_row,
Jingning Han's avatar
Jingning Han committed
199
                                         int mi_col) {
200
  const int idx_str = xd->mi_stride * mi_row + mi_col;
201
202
  xd->mi = cm->mi_grid_visible + idx_str;
  xd->mi[0] = cm->mi + idx_str;
203
  x->mbmi_ext = x->mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
204
205
206
}

static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
207
208
                        MACROBLOCK *const x, int mi_row, int mi_col,
                        BLOCK_SIZE bsize) {
209
210
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;
Scott LaVarnway's avatar
Scott LaVarnway committed
211
  MODE_INFO *mi;
212
213
214
  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;
Alex Converse's avatar
Alex Converse committed
215
  MvLimits *const mv_limits = &x->mv_limits;
216

217
  set_skip_context(xd, mi_row, mi_col);
218

219
220
  set_mode_info_offsets(cm, x, xd, mi_row, mi_col);

Scott LaVarnway's avatar
Scott LaVarnway committed
221
  mi = xd->mi[0];
222
223

  // Set up destination pointers.
224
  vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
225
226
227

  // Set up limit values for MV components.
  // Mv beyond the range do not produce new/different prediction block.
Alex Converse's avatar
Alex Converse committed
228
229
230
231
  mv_limits->row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
  mv_limits->col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
  mv_limits->row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
  mv_limits->col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;
232
233
234

  // 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)));
clang-format's avatar
clang-format committed
235
236
  set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width, cm->mi_rows,
                 cm->mi_cols);
237
238
239
240
241

  // 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
242
243
  x->rddiv = cpi->rd.RDDIV;
  x->rdmult = cpi->rd.RDMULT;
244
245
246

  // Setup segment ID.
  if (seg->enabled) {
247
    if (cpi->oxcf.aq_mode != VARIANCE_AQ && cpi->oxcf.aq_mode != LOOKAHEAD_AQ &&
248
        cpi->oxcf.aq_mode != EQUATOR360_AQ) {
clang-format's avatar
clang-format committed
249
250
      const uint8_t *const map =
          seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
Scott LaVarnway's avatar
Scott LaVarnway committed
251
      mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
252
253
254
    }
    vp9_init_plane_quantizers(cpi, x);

Scott LaVarnway's avatar
Scott LaVarnway committed
255
    x->encode_breakout = cpi->segment_encode_breakout[mi->segment_id];
256
  } else {
Scott LaVarnway's avatar
Scott LaVarnway committed
257
    mi->segment_id = 0;
258
259
    x->encode_breakout = cpi->encode_breakout;
  }
Scott LaVarnway's avatar
Scott LaVarnway committed
260
261
262

  // required by vp9_append_sub8x8_mvs_for_idx() and vp9_find_best_ref_mvs()
  xd->tile = *tile;
263
264
}

265
266
267
static void duplicate_mode_info_in_sb(VP9_COMMON *cm, MACROBLOCKD *xd,
                                      int mi_row, int mi_col,
                                      BLOCK_SIZE bsize) {
clang-format's avatar
clang-format committed
268
269
270
271
  const int block_width =
      VPXMIN(num_8x8_blocks_wide_lookup[bsize], cm->mi_cols - mi_col);
  const int block_height =
      VPXMIN(num_8x8_blocks_high_lookup[bsize], cm->mi_rows - mi_row);
272
273
  const int mi_stride = xd->mi_stride;
  MODE_INFO *const src_mi = xd->mi[0];
274
  int i, j;
275

276
  for (j = 0; j < block_height; ++j)
clang-format's avatar
clang-format committed
277
    for (i = 0; i < block_width; ++i) xd->mi[j * mi_stride + i] = src_mi;
278
279
}

clang-format's avatar
clang-format committed
280
281
static void set_block_size(VP9_COMP *const cpi, MACROBLOCK *const x,
                           MACROBLOCKD *const xd, int mi_row, int mi_col,
282
283
                           BLOCK_SIZE bsize) {
  if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
284
    set_mode_info_offsets(&cpi->common, x, xd, mi_row, mi_col);
Scott LaVarnway's avatar
Scott LaVarnway committed
285
    xd->mi[0]->sb_type = bsize;
286
287
288
289
290
291
  }
}

typedef struct {
  int64_t sum_square_error;
  int64_t sum_error;
292
  int log2_count;
293
294
295
296
297
298
299
300
301
302
303
304
  int variance;
} var;

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

typedef struct {
  partition_variance part_variances;
  var split[4];
305
306
307
308
309
} v4x4;

typedef struct {
  partition_variance part_variances;
  v4x4 split[4];
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
} 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;
340
  node->part_variances = NULL;
341
342
  switch (bsize) {
    case BLOCK_64X64: {
clang-format's avatar
clang-format committed
343
      v64x64 *vt = (v64x64 *)data;
344
345
346
347
348
349
      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: {
clang-format's avatar
clang-format committed
350
      v32x32 *vt = (v32x32 *)data;
351
352
353
354
355
356
      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: {
clang-format's avatar
clang-format committed
357
      v16x16 *vt = (v16x16 *)data;
358
359
360
361
362
363
      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: {
clang-format's avatar
clang-format committed
364
      v8x8 *vt = (v8x8 *)data;
365
      node->part_variances = &vt->part_variances;
366
367
368
369
370
      for (i = 0; i < 4; i++)
        node->split[i] = &vt->split[i].part_variances.none;
      break;
    }
    case BLOCK_4X4: {
clang-format's avatar
clang-format committed
371
      v4x4 *vt = (v4x4 *)data;
372
      node->part_variances = &vt->part_variances;
clang-format's avatar
clang-format committed
373
      for (i = 0; i < 4; i++) node->split[i] = &vt->split[i];
374
375
376
377
      break;
    }
    default: {
      assert(0);
378
      break;
379
380
381
382
383
384
385
386
    }
  }
}

// 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;
387
  v->log2_count = c;
388
389
390
}

static void get_variance(var *v) {
clang-format's avatar
clang-format committed
391
392
393
394
  v->variance =
      (int)(256 * (v->sum_square_error -
                   ((v->sum_error * v->sum_error) >> v->log2_count)) >>
            v->log2_count);
395
396
}

397
static void sum_2_variances(const var *a, const var *b, var *r) {
398
  assert(a->log2_count == b->log2_count);
399
  fill_variance(a->sum_square_error + b->sum_square_error,
400
                a->sum_error + b->sum_error, a->log2_count + 1, r);
401
402
403
404
}

static void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
  variance_node node;
405
  memset(&node, 0, sizeof(node));
406
407
408
409
410
411
412
413
414
  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);
}

clang-format's avatar
clang-format committed
415
416
417
418
static int set_vt_partitioning(VP9_COMP *cpi, MACROBLOCK *const x,
                               MACROBLOCKD *const xd, void *data,
                               BLOCK_SIZE bsize, int mi_row, int mi_col,
                               int64_t threshold, BLOCK_SIZE bsize_min,
Marco's avatar
Marco committed
419
                               int force_split) {
clang-format's avatar
clang-format committed
420
  VP9_COMMON *const cm = &cpi->common;
421
422
423
  variance_node vt;
  const int block_width = num_8x8_blocks_wide_lookup[bsize];
  const int block_height = num_8x8_blocks_high_lookup[bsize];
424

425
426
427
  assert(block_height == block_width);
  tree_to_node(data, bsize, &vt);

clang-format's avatar
clang-format committed
428
  if (force_split == 1) return 0;
429

430
  // For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
431
432
  // variance is below threshold, otherwise split will be selected.
  // No check for vert/horiz split as too few samples for variance.
433
  if (bsize == bsize_min) {
434
    // Variance already computed to set the force_split.
clang-format's avatar
clang-format committed
435
    if (cm->frame_type == KEY_FRAME) get_variance(&vt.part_variances->none);
436
437
    if (mi_col + block_width / 2 < cm->mi_cols &&
        mi_row + block_height / 2 < cm->mi_rows &&
438
        vt.part_variances->none.variance < threshold) {
439
      set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
440
441
442
      return 1;
    }
    return 0;
443
  } else if (bsize > bsize_min) {
444
    // Variance already computed to set the force_split.
clang-format's avatar
clang-format committed
445
    if (cm->frame_type == KEY_FRAME) get_variance(&vt.part_variances->none);
446
    // For key frame: take split for bsize above 32X32 or very high variance.
447
448
    if (cm->frame_type == KEY_FRAME &&
        (bsize > BLOCK_32X32 ||
clang-format's avatar
clang-format committed
449
         vt.part_variances->none.variance > (threshold << 4))) {
450
451
452
453
454
      return 0;
    }
    // If variance is low, take the bsize (no split).
    if (mi_col + block_width / 2 < cm->mi_cols &&
        mi_row + block_height / 2 < cm->mi_rows &&
455
        vt.part_variances->none.variance < threshold) {
456
      set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
457
458
      return 1;
    }
459

460
    // Check vertical split.
461
    if (mi_row + block_height / 2 < cm->mi_rows) {
Yaowu Xu's avatar
Yaowu Xu committed
462
      BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
463
464
      get_variance(&vt.part_variances->vert[0]);
      get_variance(&vt.part_variances->vert[1]);
465
      if (vt.part_variances->vert[0].variance < threshold &&
Yaowu Xu's avatar
Yaowu Xu committed
466
467
          vt.part_variances->vert[1].variance < threshold &&
          get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
468
469
        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
        set_block_size(cpi, x, xd, mi_row, mi_col + block_width / 2, subsize);
470
471
        return 1;
      }
472
    }
473
    // Check horizontal split.
474
    if (mi_col + block_width / 2 < cm->mi_cols) {
Yaowu Xu's avatar
Yaowu Xu committed
475
      BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
476
477
      get_variance(&vt.part_variances->horz[0]);
      get_variance(&vt.part_variances->horz[1]);
478
      if (vt.part_variances->horz[0].variance < threshold &&
Yaowu Xu's avatar
Yaowu Xu committed
479
480
          vt.part_variances->horz[1].variance < threshold &&
          get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
481
482
        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
        set_block_size(cpi, x, xd, mi_row + block_height / 2, mi_col, subsize);
483
484
        return 1;
      }
485
    }
486

487
    return 0;
488
489
490
491
  }
  return 0;
}

492
493
494
495
496
497
int64_t scale_part_thresh_sumdiff(int64_t threshold_base, int speed, int width,
                                  int height, int content_state) {
  if (speed >= 8) {
    if (width <= 640 && height <= 480)
      return (5 * threshold_base) >> 2;
    else if ((content_state == kLowSadLowSumdiff) ||
498
499
             (content_state == kHighSadLowSumdiff) ||
             (content_state == kLowVarHighSumdiff))
500
501
502
      return (5 * threshold_base) >> 2;
  } else if (speed == 7) {
    if ((content_state == kLowSadLowSumdiff) ||
503
504
        (content_state == kHighSadLowSumdiff) ||
        (content_state == kLowVarHighSumdiff)) {
505
506
507
508
509
510
      return (5 * threshold_base) >> 2;
    }
  }
  return threshold_base;
}

Marco's avatar
Marco committed
511
512
513
514
// Set the variance split thresholds for following the block sizes:
// 0 - threshold_64x64, 1 - threshold_32x32, 2 - threshold_16x16,
// 3 - vbp_threshold_8x8. vbp_threshold_8x8 (to split to 4x4 partition) is
// currently only used on key frame.
515
516
static void set_vbp_thresholds(VP9_COMP *cpi, int64_t thresholds[], int q,
                               int content_state) {
Marco's avatar
Marco committed
517
518
519
  VP9_COMMON *const cm = &cpi->common;
  const int is_key_frame = (cm->frame_type == KEY_FRAME);
  const int threshold_multiplier = is_key_frame ? 20 : 1;
clang-format's avatar
clang-format committed
520
521
  int64_t threshold_base =
      (int64_t)(threshold_multiplier * cpi->y_dequant[q][1]);
522

Marco's avatar
Marco committed
523
524
525
526
527
528
  if (is_key_frame) {
    thresholds[0] = threshold_base;
    thresholds[1] = threshold_base >> 2;
    thresholds[2] = threshold_base >> 2;
    thresholds[3] = threshold_base << 2;
  } else {
529
530
    // Increase base variance threshold based on estimated noise level.
    if (cpi->noise_estimate.enabled && cm->width >= 640 && cm->height >= 480) {
clang-format's avatar
clang-format committed
531
532
      NOISE_LEVEL noise_level =
          vp9_noise_estimate_extract_level(&cpi->noise_estimate);
533
      if (noise_level == kHigh)
534
        threshold_base = 3 * threshold_base;
535
536
      else if (noise_level == kMedium)
        threshold_base = threshold_base << 1;
537
      else if (noise_level < kLow)
538
        threshold_base = (7 * threshold_base) >> 3;
539
    }
540
#if CONFIG_VP9_TEMPORAL_DENOISING
541
542
    if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
        cpi->oxcf.speed > 5 && cpi->denoiser.denoising_level >= kDenLow)
543
544
545
      threshold_base =
          vp9_scale_part_thresh(threshold_base, cpi->denoiser.denoising_level,
                                content_state, cpi->svc.temporal_layer_id);
546
    else
547
      threshold_base =
548
549
          scale_part_thresh_sumdiff(threshold_base, cpi->oxcf.speed, cm->width,
                                    cm->height, content_state);
550
#else
551
552
553
    // Increase base variance threshold based on content_state/sum_diff level.
    threshold_base = scale_part_thresh_sumdiff(
        threshold_base, cpi->oxcf.speed, cm->width, cm->height, content_state);
554
#endif
555
556
    thresholds[0] = threshold_base;
    thresholds[2] = threshold_base << cpi->oxcf.speed;
Marco's avatar
Marco committed
557
    if (cm->width <= 352 && cm->height <= 288) {
558
559
      thresholds[0] = threshold_base >> 3;
      thresholds[1] = threshold_base >> 1;
560
      thresholds[2] = threshold_base << 3;
561
    } else if (cm->width < 1280 && cm->height < 720) {
Marco's avatar
Marco committed
562
      thresholds[1] = (5 * threshold_base) >> 2;
563
564
565
566
    } else if (cm->width < 1920 && cm->height < 1080) {
      thresholds[1] = threshold_base << 1;
    } else {
      thresholds[1] = (5 * threshold_base) >> 1;
Marco's avatar
Marco committed
567
568
569
570
    }
  }
}

571
572
void vp9_set_variance_partition_thresholds(VP9_COMP *cpi, int q,
                                           int content_state) {
Marco's avatar
Marco committed
573
  VP9_COMMON *const cm = &cpi->common;
Yaowu Xu's avatar
Yaowu Xu committed
574
  SPEED_FEATURES *const sf = &cpi->sf;
Marco's avatar
Marco committed
575
  const int is_key_frame = (cm->frame_type == KEY_FRAME);
576
577
  if (sf->partition_search_type != VAR_BASED_PARTITION &&
      sf->partition_search_type != REFERENCE_PARTITION) {
Yaowu Xu's avatar
Yaowu Xu committed
578
579
    return;
  } else {
580
    set_vbp_thresholds(cpi, cpi->vbp_thresholds, q, content_state);
Marco's avatar
Marco committed
581
    // The thresholds below are not changed locally.
582
    if (is_key_frame) {
583
      cpi->vbp_threshold_sad = 0;
584
      cpi->vbp_threshold_copy = 0;
585
586
      cpi->vbp_bsize_min = BLOCK_8X8;
    } else {
Marco's avatar
Marco committed
587
      if (cm->width <= 352 && cm->height <= 288)
588
        cpi->vbp_threshold_sad = 10;
Marco's avatar
Marco committed
589
      else
clang-format's avatar
clang-format committed
590
591
592
        cpi->vbp_threshold_sad = (cpi->y_dequant[q][1] << 1) > 1000
                                     ? (cpi->y_dequant[q][1] << 1)
                                     : 1000;
593
      cpi->vbp_bsize_min = BLOCK_16X16;
594
      if (cm->width <= 352 && cm->height <= 288)
595
596
        cpi->vbp_threshold_copy = 4000;
      else if (cm->width <= 640 && cm->height <= 360)
597
        cpi->vbp_threshold_copy = 8000;
598
      else
599
600
601
        cpi->vbp_threshold_copy = (cpi->y_dequant[q][1] << 3) > 8000
                                      ? (cpi->y_dequant[q][1] << 3)
                                      : 8000;
602
    }
603
    cpi->vbp_threshold_minmax = 15 + (q >> 3);
Yaowu Xu's avatar
Yaowu Xu committed
604
605
606
  }
}

607
608
609
610
611
612
// Compute the minmax over the 8x8 subblocks.
static int compute_minmax_8x8(const uint8_t *s, int sp, const uint8_t *d,
                              int dp, int x16_idx, int y16_idx,
#if CONFIG_VP9_HIGHBITDEPTH
                              int highbd_flag,
#endif
clang-format's avatar
clang-format committed
613
                              int pixels_wide, int pixels_high) {
614
615
616
617
618
619
620
621
622
623
624
625
  int k;
  int minmax_max = 0;
  int minmax_min = 255;
  // Loop over the 4 8x8 subblocks.
  for (k = 0; k < 4; k++) {
    int x8_idx = x16_idx + ((k & 1) << 3);
    int y8_idx = y16_idx + ((k >> 1) << 3);
    int min = 0;
    int max = 0;
    if (x8_idx < pixels_wide && y8_idx < pixels_high) {
#if CONFIG_VP9_HIGHBITDEPTH
      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
James Zern's avatar
James Zern committed
626
        vpx_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
clang-format's avatar
clang-format committed
627
                              d + y8_idx * dp + x8_idx, dp, &min, &max);
628
      } else {
clang-format's avatar
clang-format committed
629
630
        vpx_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx,
                       dp, &min, &max);
631
632
      }
#else
clang-format's avatar
clang-format committed
633
      vpx_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx, dp,
634
635
                     &min, &max);
#endif
clang-format's avatar
clang-format committed
636
637
      if ((max - min) > minmax_max) minmax_max = (max - min);
      if ((max - min) < minmax_min) minmax_min = (max - min);
638
639
640
641
642
    }
  }
  return (minmax_max - minmax_min);
}

643
644
645
646
647
static void fill_variance_4x4avg(const uint8_t *s, int sp, const uint8_t *d,
                                 int dp, int x8_idx, int y8_idx, v8x8 *vst,
#if CONFIG_VP9_HIGHBITDEPTH
                                 int highbd_flag,
#endif
clang-format's avatar
clang-format committed
648
                                 int pixels_wide, int pixels_high,
649
650
651
652
653
654
655
656
657
658
659
660
                                 int is_key_frame) {
  int k;
  for (k = 0; k < 4; k++) {
    int x4_idx = x8_idx + ((k & 1) << 2);
    int y4_idx = y8_idx + ((k >> 1) << 2);
    unsigned int sse = 0;
    int sum = 0;
    if (x4_idx < pixels_wide && y4_idx < pixels_high) {
      int s_avg;
      int d_avg = 128;
#if CONFIG_VP9_HIGHBITDEPTH
      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
James Zern's avatar
James Zern committed
661
        s_avg = vpx_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
662
        if (!is_key_frame)
James Zern's avatar
James Zern committed
663
          d_avg = vpx_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
664
      } else {
James Zern's avatar
James Zern committed
665
        s_avg = vpx_avg_4x4(s + y4_idx * sp + x4_idx, sp);
clang-format's avatar
clang-format committed
666
        if (!is_key_frame) d_avg = vpx_avg_4x4(d + y4_idx * dp + x4_idx, dp);
667
668
      }
#else
James Zern's avatar
James Zern committed
669
      s_avg = vpx_avg_4x4(s + y4_idx * sp + x4_idx, sp);
clang-format's avatar
clang-format committed
670
      if (!is_key_frame) d_avg = vpx_avg_4x4(d + y4_idx * dp + x4_idx, dp);
671
672
673
674
675
676
677
678
679
680
681
682
683
#endif
      sum = s_avg - d_avg;
      sse = sum * sum;
    }
    fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
  }
}

static void fill_variance_8x8avg(const uint8_t *s, int sp, const uint8_t *d,
                                 int dp, int x16_idx, int y16_idx, v16x16 *vst,
#if CONFIG_VP9_HIGHBITDEPTH
                                 int highbd_flag,
#endif
clang-format's avatar
clang-format committed
684
                                 int pixels_wide, int pixels_high,
685
686
687
688
689
690
691
692
693
694
695
696
                                 int is_key_frame) {
  int k;
  for (k = 0; k < 4; k++) {
    int x8_idx = x16_idx + ((k & 1) << 3);
    int y8_idx = y16_idx + ((k >> 1) << 3);
    unsigned int sse = 0;
    int sum = 0;
    if (x8_idx < pixels_wide && y8_idx < pixels_high) {
      int s_avg;
      int d_avg = 128;
#if CONFIG_VP9_HIGHBITDEPTH
      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
James Zern's avatar
James Zern committed
697
        s_avg = vpx_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
698
        if (!is_key_frame)
James Zern's avatar
James Zern committed
699
          d_avg = vpx_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
700
      } else {
James Zern's avatar
James Zern committed
701
        s_avg = vpx_avg_8x8(s + y8_idx * sp + x8_idx, sp);
clang-format's avatar
clang-format committed
702
        if (!is_key_frame) d_avg = vpx_avg_8x8(d + y8_idx * dp + x8_idx, dp);
703
704
      }
#else
James Zern's avatar
James Zern committed
705
      s_avg = vpx_avg_8x8(s + y8_idx * sp + x8_idx, sp);
clang-format's avatar
clang-format committed
706
      if (!is_key_frame) d_avg = vpx_avg_8x8(d + y8_idx * dp + x8_idx, dp);
707
708
709
710
711
712
713
714
#endif
      sum = s_avg - d_avg;
      sse = sum * sum;
    }
    fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
  }
}

JackyChen's avatar
JackyChen committed
715
716
717
718
// Check if most of the superblock is skin content, and if so, force split to
// 32x32, and set x->sb_is_skin for use in mode selection.
static int skin_sb_split(VP9_COMP *cpi, MACROBLOCK *x, const int low_res,
                         int mi_row, int mi_col, int *force_split) {
clang-format's avatar
clang-format committed
719
  VP9_COMMON *const cm = &cpi->common;
720
721
722
#if CONFIG_VP9_HIGHBITDEPTH
  if (cm->use_highbitdepth) return 0;
#endif
JackyChen's avatar
JackyChen committed
723
724
725
726
  // Avoid checking superblocks on/near boundary and avoid low resolutions.
  // Note superblock may still pick 64X64 if y_sad is very small
  // (i.e., y_sad < cpi->vbp_threshold_sad) below. For now leave this as is.
  if (!low_res && (mi_col >= 8 && mi_col + 8 < cm->mi_cols && mi_row >= 8 &&
clang-format's avatar
clang-format committed
727
                   mi_row + 8 < cm->mi_rows)) {
JackyChen's avatar
JackyChen committed
728
729
730
731
732
733
734
735
736
737
738
739
740
741
    int num_16x16_skin = 0;
    int num_16x16_nonskin = 0;
    uint8_t *ysignal = x->plane[0].src.buf;
    uint8_t *usignal = x->plane[1].src.buf;
    uint8_t *vsignal = x->plane[2].src.buf;
    int sp = x->plane[0].src.stride;
    int spuv = x->plane[1].src.stride;
    const int block_index = mi_row * cm->mi_cols + mi_col;
    const int bw = num_8x8_blocks_wide_lookup[BLOCK_64X64];
    const int bh = num_8x8_blocks_high_lookup[BLOCK_64X64];
    const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
    const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
    // Loop through the 16x16 sub-blocks.
    int i, j;
clang-format's avatar
clang-format committed
742
743
    for (i = 0; i < ymis; i += 2) {
      for (j = 0; j < xmis; j += 2) {
JackyChen's avatar
JackyChen committed
744
745
746
747
        int bl_index = block_index + i * cm->mi_cols + j;
        int bl_index1 = bl_index + 1;
        int bl_index2 = bl_index + cm->mi_cols;
        int bl_index3 = bl_index2 + 1;
clang-format's avatar
clang-format committed
748
749
750
751
752
753
754
        int consec_zeromv =
            VPXMIN(cpi->consec_zero_mv[bl_index],
                   VPXMIN(cpi->consec_zero_mv[bl_index1],
                          VPXMIN(cpi->consec_zero_mv[bl_index2],
                                 cpi->consec_zero_mv[bl_index3])));
        int is_skin = vp9_compute_skin_block(
            ysignal, usignal, vsignal, sp, spuv, BLOCK_16X16, consec_zeromv, 0);
JackyChen's avatar
JackyChen committed
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
        num_16x16_skin += is_skin;
        num_16x16_nonskin += (1 - is_skin);
        if (num_16x16_nonskin > 3) {
          // Exit loop if at least 4 of the 16x16 blocks are not skin.
          i = ymis;
          break;
        }
        ysignal += 16;
        usignal += 8;
        vsignal += 8;
      }
      ysignal += (sp << 4) - 64;
      usignal += (spuv << 3) - 32;
      vsignal += (spuv << 3) - 32;
    }
    if (num_16x16_skin > 12) {
      *force_split = 1;
      return 1;
    }
  }
  return 0;
}

clang-format's avatar
clang-format committed
778
static void set_low_temp_var_flag(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
779
                                  v64x64 *vt, int64_t thresholds[],
780
781
782
                                  MV_REFERENCE_FRAME ref_frame_partition,
                                  int mi_col, int mi_row) {
  int i, j;
clang-format's avatar
clang-format committed
783
  VP9_COMMON *const cm = &cpi->common;
784
785
786
787
788
789
790
791
792
793
794
  const int mv_thr = cm->width > 640 ? 8 : 4;
  // Check temporal variance for bsize >= 16x16, if LAST_FRAME was selected and
  // int_pro mv is small. If the temporal variance is small set the flag
  // variance_low for the block. The variance threshold can be adjusted, the
  // higher the more aggressive.
  if (ref_frame_partition == LAST_FRAME &&
      (cpi->sf.short_circuit_low_temp_var == 1 ||
       (xd->mi[0]->mv[0].as_mv.col < mv_thr &&
        xd->mi[0]->mv[0].as_mv.col > -mv_thr &&
        xd->mi[0]->mv[0].as_mv.row < mv_thr &&
        xd->mi[0]->mv[0].as_mv.row > -mv_thr))) {
795
796
797
    if (xd->mi[0]->sb_type == BLOCK_64X64) {
      if ((vt->part_variances).none.variance < (thresholds[0] >> 1))
        x->variance_low[0] = 1;
798
799
800
801
802
803
804
805
806
807
808
809
    } else if (xd->mi[0]->sb_type == BLOCK_64X32) {
      for (i = 0; i < 2; i++) {
        if (vt->part_variances.horz[i].variance < (thresholds[0] >> 2))
          x->variance_low[i + 1] = 1;
      }
    } else if (xd->mi[0]->sb_type == BLOCK_32X64) {
      for (i = 0; i < 2; i++) {
        if (vt->part_variances.vert[i].variance < (thresholds[0] >> 2))
          x->variance_low[i + 3] = 1;
      }
    } else {
      for (i = 0; i < 4; i++) {
810
811
812
813
814
815
816
817
818
819
        const int idx[4][2] = { { 0, 0 }, { 0, 4 }, { 4, 0 }, { 4, 4 } };
        const int idx_str =
            cm->mi_stride * (mi_row + idx[i][0]) + mi_col + idx[i][1];
        MODE_INFO **this_mi = cm->mi_grid_visible + idx_str;

        if (cm->mi_cols <= mi_col + idx[i][1] ||
            cm->mi_rows <= mi_row + idx[i][0])
          continue;

        if ((*this_mi)->sb_type == BLOCK_32X32) {
820
821
          int64_t threshold_32x32 = (cpi->sf.short_circuit_low_temp_var == 1 ||
                                     cpi->sf.short_circuit_low_temp_var == 3)
822
                                        ? ((5 * thresholds[1]) >> 3)
823
824
                                        : (thresholds[1] >> 1);
          if (vt->split[i].part_variances.none.variance < threshold_32x32)
825
            x->variance_low[i + 5] = 1;
826
        } else if (cpi->sf.short_circuit_low_temp_var >= 2) {
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
          // For 32x16 and 16x32 blocks, the flag is set on each 16x16 block
          // inside.
          if ((*this_mi)->sb_type == BLOCK_16X16 ||
              (*this_mi)->sb_type == BLOCK_32X16 ||
              (*this_mi)->sb_type == BLOCK_16X32) {
            for (j = 0; j < 4; j++) {
              if (vt->split[i].split[j].part_variances.none.variance <
                  (thresholds[2] >> 8))
                x->variance_low[(i << 2) + j + 9] = 1;
            }
          }
        }
      }
    }
  }
}

844
845
static void copy_partitioning_helper(VP9_COMP *cpi, MACROBLOCK *x,
                                     MACROBLOCKD *xd, BLOCK_SIZE bsize,
846
                                     int mi_row, int mi_col) {
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
  VP9_COMMON *const cm = &cpi->common;
  BLOCK_SIZE *prev_part = cpi->prev_partition;
  int start_pos = mi_row * cm->mi_stride + mi_col;

  const int bsl = b_width_log2_lookup[bsize];
  const int bs = (1 << bsl) / 4;
  BLOCK_SIZE subsize;
  PARTITION_TYPE partition;

  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;

  partition = partition_lookup[bsl][prev_part[start_pos]];
  subsize = get_subsize(bsize, partition);

  if (subsize < BLOCK_8X8) {
862
    set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
863
864
  } else {
    switch (partition) {
865
866
867
      case PARTITION_NONE:
        set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
        break;
868
      case PARTITION_HORZ:
869
870
        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
        set_block_size(cpi, x, xd, mi_row + bs, mi_col, subsize);
871
872
        break;
      case PARTITION_VERT:
873
874
        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
        set_block_size(cpi, x, xd, mi_row, mi_col + bs, subsize);
875
876
        break;
      case PARTITION_SPLIT:
877
878
879
880
        copy_partitioning_helper(cpi, x, xd, subsize, mi_row, mi_col);
        copy_partitioning_helper(cpi, x, xd, subsize, mi_row + bs, mi_col);
        copy_partitioning_helper(cpi, x, xd, subsize, mi_row, mi_col + bs);
        copy_partitioning_helper(cpi, x, xd, subsize, mi_row + bs, mi_col + bs);
881
882
883
884
885
886
        break;
      default: assert(0);
    }
  }
}

887
888
889
static int copy_partitioning(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
                             int mi_row, int mi_col, int segment_id,
                             int sb_offset) {
890
  int svc_copy_allowed = 1;
891
  int frames_since_key_thresh = 1;
892
893
894
  if (cpi->use_svc) {
    // For SVC, don't allow copy if base spatial layer is key frame, or if
    // frame is not a temporal enhancement layer frame.
895
896
897
    int layer = LAYER_IDS_TO_IDX(0, cpi->svc.temporal_layer_id,
                                 cpi->svc.number_temporal_layers);
    const LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer];
898
899
900
901
    if (lc->is_key_frame ||
        (cpi->svc.temporal_layer_id != cpi->svc.number_temporal_layers - 1 &&
         cpi->svc.number_temporal_layers > 1))
      svc_copy_allowed = 0;
902
903
    frames_since_key_thresh = cpi->svc.number_spatial_layers << 1;
  }
904
905
  if (cpi->rc.frames_since_key > frames_since_key_thresh && svc_copy_allowed &&
      !cpi->resize_pending && segment_id == CR_SEGMENT_ID_BASE &&
906
907
908
      cpi->prev_segment_id[sb_offset] == CR_SEGMENT_ID_BASE &&
      cpi->copied_frame_cnt[sb_offset] < cpi->max_copied_frame) {
    if (cpi->prev_partition != NULL) {
909
      copy_partitioning_helper(cpi, x, xd, BLOCK_64X64, mi_row, mi_col);
910
911
912
913
914
915
916
917
918
919
      cpi->copied_frame_cnt[sb_offset] += 1;
      memcpy(x->variance_low, &(cpi->prev_variance_low[sb_offset * 25]),
             sizeof(x->variance_low));
      return 1;
    }
  }

  return 0;
}

920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
static void update_prev_partition(VP9_COMP *cpi, BLOCK_SIZE bsize, int mi_row,
                                  int mi_col) {
  VP9_COMMON *const cm = &cpi->common;
  BLOCK_SIZE *prev_part = cpi->prev_partition;
  int start_pos = mi_row * cm->mi_stride + mi_col;
  const int bsl = b_width_log2_lookup[bsize];
  const int bs = (1 << bsl) / 4;
  BLOCK_SIZE subsize;
  PARTITION_TYPE partition;
  const MODE_INFO *mi = NULL;

  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;

  mi = cm->mi_grid_visible[start_pos];
  partition = partition_lookup[bsl][mi->sb_type];
  subsize = get_subsize(bsize, partition);
  if (subsize < BLOCK_8X8) {
    prev_part[start_pos] = bsize;
  } else {
    switch (partition) {
      case PARTITION_NONE: prev_part[start_pos] = bsize; break;
      case PARTITION_HORZ:
        prev_part[start_pos] = subsize;
        if (mi_row + bs < cm->mi_rows)
          prev_part[start_pos + bs * cm->mi_stride] = subsize;
        break;
      case PARTITION_VERT:
        prev_part[start_pos] = subsize;
        if (mi_col + bs < cm->mi_cols) prev_part[start_pos + bs] = subsize;
        break;
      case PARTITION_SPLIT:
        update_prev_partition(cpi, subsize, mi_row, mi_col);
        update_prev_partition(cpi, subsize, mi_row + bs, mi_col);
        update_prev_partition(cpi, subsize, mi_row, mi_col + bs);
        update_prev_partition(cpi, subsize, mi_row + bs, mi_col + bs);
        break;
      default: assert(0);
    }
  }
}

961
962
963
964
static void chroma_check(VP9_COMP *cpi, MACROBLOCK *x, int bsize,
                         unsigned int y_sad, int is_key_frame) {
  int i;
  MACROBLOCKD *xd = &x->e_mbd;
965
966
967

  if (is_key_frame) return;

968
  // For speed >= 8, avoid the chroma check if y_sad is above threshold.
969
970
971
972
973
974
  if (cpi->oxcf.speed >= 8) {
    if (y_sad > cpi->vbp_thresholds[1] &&
        (!cpi->noise_estimate.enabled ||
         vp9_noise_estimate_extract_level(&cpi->noise_estimate) < kMedium))
      return;
  }
975
976
977

  for (i = 1; i <= 2; ++i) {
    unsigned int uv_sad = UINT_MAX;
clang-format's avatar
clang-format committed
978
    struct macroblock_plane *p = &x->plane[i];
979
980
981
982
    struct macroblockd_plane *pd = &xd->plane[i];
    const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);

    if (bs != BLOCK_INVALID)
clang-format's avatar
clang-format committed
983
984
      uv_sad = cpi->fn_ptr[bs].sdf(p->src.buf, p->src.stride, pd->dst.buf,
                                   pd->dst.stride);
985
986
987
988
989
990
991

    // TODO(marpan): Investigate if we should lower this threshold if