vp9_pickmode.c 44 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
/*
 *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include <assert.h>
12
13
14
#include <limits.h>
#include <math.h>
#include <stdio.h>
15

16
17
18
19
#include "./vp9_rtcd.h"

#include "vpx_mem/vpx_mem.h"

20
#include "vp9/common/vp9_blockd.h"
21
22
#include "vp9/common/vp9_common.h"
#include "vp9/common/vp9_mvref_common.h"
23
24
#include "vp9/common/vp9_reconinter.h"
#include "vp9/common/vp9_reconintra.h"
25

Dmitry Kovalev's avatar
Dmitry Kovalev committed
26
#include "vp9/encoder/vp9_encoder.h"
27
#include "vp9/encoder/vp9_pickmode.h"
28
#include "vp9/encoder/vp9_ratectrl.h"
29
#include "vp9/encoder/vp9_rd.h"
30

31
32
33
34
35
36
typedef struct {
  uint8_t *data;
  int stride;
  int in_use;
} PRED_BUFFER;

37
static int mv_refs_rt(const VP9_COMMON *cm, const MACROBLOCKD *xd,
38
39
40
41
                      const TileInfo *const tile,
                      MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
                      int_mv *mv_ref_list,
                      int mi_row, int mi_col) {
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
  const int *ref_sign_bias = cm->ref_frame_sign_bias;
  int i, refmv_count = 0;

  const POSITION *const mv_ref_search = mv_ref_blocks[mi->mbmi.sb_type];

  int different_ref_found = 0;
  int context_counter = 0;
  int const_motion = 0;

  // Blank the reference vector list
  vpx_memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);

  // The nearest 2 blocks are treated differently
  // if the size < 8x8 we get the mv from the bmi substructure,
  // and we also need to keep a mode count.
  for (i = 0; i < 2; ++i) {
    const POSITION *const mv_ref = &mv_ref_search[i];
    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
      const MODE_INFO *const candidate_mi = xd->mi[mv_ref->col + mv_ref->row *
hkuang's avatar
hkuang committed
61
                                                   xd->mi_stride].src_mi;
62
63
64
65
66
67
      const MB_MODE_INFO *const candidate = &candidate_mi->mbmi;
      // Keep counts for entropy encoding.
      context_counter += mode_2_counter[candidate->mode];
      different_ref_found = 1;

      if (candidate->ref_frame[0] == ref_frame)
68
69
        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1),
                        refmv_count, mv_ref_list, Done);
70
71
72
73
74
75
76
77
78
79
80
81
    }
  }

  const_motion = 1;

  // Check the rest of the neighbors in much the same way
  // as before except we don't need to keep track of sub blocks or
  // mode counts.
  for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) {
    const POSITION *const mv_ref = &mv_ref_search[i];
    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
      const MB_MODE_INFO *const candidate = &xd->mi[mv_ref->col + mv_ref->row *
hkuang's avatar
hkuang committed
82
                                                    xd->mi_stride].src_mi->mbmi;
83
84
85
      different_ref_found = 1;

      if (candidate->ref_frame[0] == ref_frame)
86
        ADD_MV_REF_LIST(candidate->mv[0], refmv_count, mv_ref_list, Done);
87
88
89
90
91
92
93
94
95
96
97
    }
  }

  // Since we couldn't find 2 mvs from the same reference frame
  // go back through the neighbors and find motion vectors from
  // different reference frames.
  if (different_ref_found && !refmv_count) {
    for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
      const POSITION *mv_ref = &mv_ref_search[i];
      if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
        const MB_MODE_INFO *const candidate = &xd->mi[mv_ref->col + mv_ref->row
hkuang's avatar
hkuang committed
98
                                              * xd->mi_stride].src_mi->mbmi;
99
100

        // If the candidate is INTRA we don't want to consider its mv.
101
102
        IF_DIFF_REF_FRAME_ADD_MV(candidate, ref_frame, ref_sign_bias,
                                 refmv_count, mv_ref_list, Done);
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
      }
    }
  }

 Done:

  mi->mbmi.mode_context[ref_frame] = counter_to_context[context_counter];

  // Clamp vectors
  for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
    clamp_mv_ref(&mv_ref_list[i].as_mv, xd);

  return const_motion;
}

118
119
120
121
static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
                                  BLOCK_SIZE bsize, int mi_row, int mi_col,
                                  int_mv *tmp_mv, int *rate_mv,
                                  int64_t best_rd_sofar) {
122
  MACROBLOCKD *xd = &x->e_mbd;
hkuang's avatar
hkuang committed
123
  MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
124
  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
125
126
  const int step_param = cpi->sf.mv.fullpel_search_step_param;
  const int sadpb = x->sadperbit16;
127
  MV mvp_full;
128
  const int ref = mbmi->ref_frame[0];
129
  const MV ref_mv = mbmi->ref_mvs[ref][0].as_mv;
130
131
132
133
134
135
136
  int dis;
  int rate_mode;
  const int tmp_col_min = x->mv_col_min;
  const int tmp_col_max = x->mv_col_max;
  const int tmp_row_min = x->mv_row_min;
  const int tmp_row_max = x->mv_row_max;
  int rv = 0;
137
  int cost_list[5];
138
139
  const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi,
                                                                        ref);
Yaowu Xu's avatar
Yaowu Xu committed
140
141
142
143
  if (cpi->common.show_frame &&
      (x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[LAST_FRAME])
    return rv;

144
145
146
147
148
149
150
  if (scaled_ref_frame) {
    int i;
    // Swap out the reference frame for a version that's been scaled to
    // match the resolution of the current frame, allowing the existing
    // motion search code to be used without additional modifications.
    for (i = 0; i < MAX_MB_PLANE; i++)
      backup_yv12[i] = xd->plane[i].pre[0];
151
    vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
152
  }
153
  vp9_set_mv_search_range(x, &ref_mv);
154

Yaowu Xu's avatar
Yaowu Xu committed
155
156
157
158
  assert(x->mv_best_ref_index[ref] <= 2);
  if (x->mv_best_ref_index[ref] < 2)
    mvp_full = mbmi->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
  else
Dmitry Kovalev's avatar
Dmitry Kovalev committed
159
    mvp_full = x->pred_mv[ref];
160
161
162
163

  mvp_full.col >>= 3;
  mvp_full.row >>= 3;

164
  vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
165
                        cond_cost_list(cpi, cost_list),
166
                        &ref_mv, &tmp_mv->as_mv, INT_MAX, 0);
167

168
169
170
171
172
  x->mv_col_min = tmp_col_min;
  x->mv_col_max = tmp_col_max;
  x->mv_row_min = tmp_row_min;
  x->mv_row_max = tmp_row_max;

173
174
175
  // calculate the bit cost on motion vector
  mvp_full.row = tmp_mv->as_mv.row * 8;
  mvp_full.col = tmp_mv->as_mv.col * 8;
176

177
178
  *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv,
                             x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
179

180
181
182
183
184
185
186
187
188
189
190
191
  rate_mode = cpi->inter_mode_cost[mbmi->mode_context[ref]]
                                  [INTER_OFFSET(NEWMV)];
  rv = !(RDCOST(x->rdmult, x->rddiv, (*rate_mv + rate_mode), 0) >
         best_rd_sofar);

  if (rv) {
    cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
                                 cpi->common.allow_high_precision_mv,
                                 x->errorperbit,
                                 &cpi->fn_ptr[bsize],
                                 cpi->sf.mv.subpel_force_stop,
                                 cpi->sf.mv.subpel_iters_per_step,
192
                                 cond_cost_list(cpi, cost_list),
193
                                 x->nmvjointcost, x->mvcost,
194
                                 &dis, &x->pred_sse[ref], NULL, 0, 0);
195
    x->pred_mv[ref] = tmp_mv->as_mv;
196
197
198
199
200
201
202
  }

  if (scaled_ref_frame) {
    int i;
    for (i = 0; i < MAX_MB_PLANE; i++)
      xd->plane[i].pre[0] = backup_yv12[i];
  }
203
  return rv;
204
205
}

206

207
208
static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize,
                              MACROBLOCK *x, MACROBLOCKD *xd,
209
210
                              int *out_rate_sum, int64_t *out_dist_sum,
                              unsigned int *var_y, unsigned int *sse_y) {
211
212
213
214
  // Note our transform coeffs are 8 times an orthogonal transform.
  // Hence quantizer step is also 8 times. To get effective quantizer
  // we need to divide by 8 before sending to modeling function.
  unsigned int sse;
215
216
  int rate;
  int64_t dist;
217
218
  struct macroblock_plane *const p = &x->plane[0];
  struct macroblockd_plane *const pd = &xd->plane[0];
219
220
  const uint32_t dc_quant = pd->dequant[0];
  const uint32_t ac_quant = pd->dequant[1];
221
222
  unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
                                           pd->dst.buf, pd->dst.stride, &sse);
223
224
225
  *var_y = var;
  *sse_y = sse;

226
  if (sse < dc_quant * dc_quant >> 6)
227
    x->skip_txfm[0] = 1;
228
  else if (var < ac_quant * ac_quant >> 6)
229
    x->skip_txfm[0] = 2;
230
  else
231
    x->skip_txfm[0] = 0;
232

233
234
  if (cpi->common.tx_mode == TX_MODE_SELECT) {
    if (sse > (var << 2))
hkuang's avatar
hkuang committed
235
      xd->mi[0].src_mi->mbmi.tx_size =
236
237
          MIN(max_txsize_lookup[bsize],
              tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
238
    else
hkuang's avatar
hkuang committed
239
      xd->mi[0].src_mi->mbmi.tx_size = TX_8X8;
240
241
242
243

    if (cpi->sf.partition_search_type == VAR_BASED_PARTITION &&
        xd->mi[0].src_mi->mbmi.tx_size > TX_16X16)
      xd->mi[0].src_mi->mbmi.tx_size = TX_16X16;
244
  } else {
hkuang's avatar
hkuang committed
245
    xd->mi[0].src_mi->mbmi.tx_size =
246
247
        MIN(max_txsize_lookup[bsize],
            tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
248
249
  }

250
251
#if CONFIG_VP9_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
252
    vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
253
254
                                 dc_quant >> (xd->bd - 5), &rate, &dist);
  } else {
255
    vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
256
257
258
                                 dc_quant >> 3, &rate, &dist);
  }
#else
259
  vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
260
                               dc_quant >> 3, &rate, &dist);
261
262
#endif  // CONFIG_VP9_HIGHBITDEPTH

263
  *out_rate_sum = rate >> 1;
264
  *out_dist_sum = dist << 3;
265

266
267
#if CONFIG_VP9_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
268
    vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
269
                                 ac_quant >> (xd->bd - 5), &rate, &dist);
270
  } else {
271
    vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
272
                                 ac_quant >> 3, &rate, &dist);
273
274
  }
#else
275
  vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
276
                               ac_quant >> 3, &rate, &dist);
277
278
#endif  // CONFIG_VP9_HIGHBITDEPTH

279
280
  *out_rate_sum += rate;
  *out_dist_sum += dist << 4;
281
282
283

  if (*out_rate_sum == 0)
    x->skip_txfm[0] = 1;
284
285
}

286
287
288
289
290
291
292
293
294
295
296
297
298
static int get_pred_buffer(PRED_BUFFER *p, int len) {
  int i;

  for (i = 0; i < len; i++) {
    if (!p[i].in_use) {
      p[i].in_use = 1;
      return i;
    }
  }
  return -1;
}

static void free_pred_buffer(PRED_BUFFER *p) {
299
300
  if (p != NULL)
    p->in_use = 0;
301
302
}

303
304
305
306
307
308
309
310
static void encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x,
                                 BLOCK_SIZE bsize, int mi_row, int mi_col,
                                 MV_REFERENCE_FRAME ref_frame,
                                 PREDICTION_MODE this_mode,
                                 unsigned int var_y, unsigned int sse_y,
                                 struct buf_2d yv12_mb[][MAX_MB_PLANE],
                                 int *rate, int64_t *dist) {
  MACROBLOCKD *xd = &x->e_mbd;
hkuang's avatar
hkuang committed
311
  MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326

  const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
  unsigned int var = var_y, sse = sse_y;
  // Skipping threshold for ac.
  unsigned int thresh_ac;
  // Skipping threshold for dc.
  unsigned int thresh_dc;
  if (x->encode_breakout > 0) {
    // Set a maximum for threshold to avoid big PSNR loss in low bit rate
    // case. Use extreme low threshold for static frames to limit
    // skipping.
    const unsigned int max_thresh = 36000;
    // The encode_breakout input
    const unsigned int min_thresh =
        MIN(((unsigned int)x->encode_breakout << 4), max_thresh);
327
328
329
#if CONFIG_VP9_HIGHBITDEPTH
    const int shift = 2 * xd->bd - 16;
#endif
330
331
332

    // Calculate threshold according to dequant value.
    thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) / 9;
333
334
335
336
337
#if CONFIG_VP9_HIGHBITDEPTH
    if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
      thresh_ac = ROUND_POWER_OF_TWO(thresh_ac, shift);
    }
#endif  // CONFIG_VP9_HIGHBITDEPTH
338
339
340
341
    thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);

    // Adjust ac threshold according to partition size.
    thresh_ac >>=
342
        8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
343
344

    thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
345
346
347
348
349
#if CONFIG_VP9_HIGHBITDEPTH
    if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
      thresh_dc = ROUND_POWER_OF_TWO(thresh_dc, shift);
    }
#endif  // CONFIG_VP9_HIGHBITDEPTH
350
351
352
353
354
355
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
  } else {
    thresh_ac = 0;
    thresh_dc = 0;
  }

  // Y skipping condition checking for ac and dc.
  if (var <= thresh_ac && (sse - var) <= thresh_dc) {
    unsigned int sse_u, sse_v;
    unsigned int var_u, var_v;

    // Skip UV prediction unless breakout is zero (lossless) to save
    // computation with low impact on the result
    if (x->encode_breakout == 0) {
      xd->plane[1].pre[0] = yv12_mb[ref_frame][1];
      xd->plane[2].pre[0] = yv12_mb[ref_frame][2];
      vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize);
    }

    var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
                                    x->plane[1].src.stride,
                                    xd->plane[1].dst.buf,
                                    xd->plane[1].dst.stride, &sse_u);

    // U skipping condition checking
    if ((var_u * 4 <= thresh_ac) && (sse_u - var_u <= thresh_dc)) {
      var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
                                      x->plane[2].src.stride,
                                      xd->plane[2].dst.buf,
                                      xd->plane[2].dst.stride, &sse_v);

      // V skipping condition checking
      if ((var_v * 4 <= thresh_ac) && (sse_v - var_v <= thresh_dc)) {
        x->skip = 1;

        // The cost of skip bit needs to be added.
        *rate = cpi->inter_mode_cost[mbmi->mode_context[ref_frame]]
386
                                    [INTER_OFFSET(this_mode)];
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402

        // More on this part of rate
        // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);

        // Scaling factor for SSE from spatial domain to frequency
        // domain is 16. Adjust distortion accordingly.
        // TODO(yunqingwang): In this function, only y-plane dist is
        // calculated.
        *dist = (sse << 4);  // + ((sse_u + sse_v) << 4);

        // *disable_skip = 1;
      }
    }
  }
}

403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
struct estimate_block_intra_args {
  VP9_COMP *cpi;
  MACROBLOCK *x;
  PREDICTION_MODE mode;
  int rate;
  int64_t dist;
};

static void estimate_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
                                 TX_SIZE tx_size, void *arg) {
  struct estimate_block_intra_args* const args = arg;
  VP9_COMP *const cpi = args->cpi;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
  struct macroblock_plane *const p = &x->plane[0];
  struct macroblockd_plane *const pd = &xd->plane[0];
  const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
  uint8_t *const src_buf_base = p->src.buf;
  uint8_t *const dst_buf_base = pd->dst.buf;
  const int src_stride = p->src.stride;
  const int dst_stride = pd->dst.stride;
  int i, j;
  int rate;
  int64_t dist;
  unsigned int var_y, sse_y;
  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
  assert(plane == 0);
  (void) plane;

  p->src.buf = &src_buf_base[4 * (j * src_stride + i)];
  pd->dst.buf = &dst_buf_base[4 * (j * dst_stride + i)];
  // Use source buffer as an approximation for the fully reconstructed buffer.
  vp9_predict_intra_block(xd, block >> (2 * tx_size),
436
                          b_width_log2_lookup[plane_bsize],
437
                          tx_size, args->mode,
438
439
                          x->skip_encode ? p->src.buf : pd->dst.buf,
                          x->skip_encode ? src_stride : dst_stride,
440
441
442
443
444
445
446
447
448
449
                          pd->dst.buf, dst_stride,
                          i, j, 0);
  // This procedure assumes zero offset from p->src.buf and pd->dst.buf.
  model_rd_for_sb_y(cpi, bsize_tx, x, xd, &rate, &dist, &var_y, &sse_y);
  p->src.buf = src_buf_base;
  pd->dst.buf = dst_buf_base;
  args->rate += rate;
  args->dist += dist;
}

450
static const THR_MODES mode_idx[MAX_REF_FRAMES - 1][4] = {
451
  {THR_DC, THR_H_PRED, THR_V_PRED, THR_TM},
452
453
454
455
  {THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV},
  {THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG},
};

456
457
458
459
static const PREDICTION_MODE intra_mode_list[] = {
  DC_PRED, V_PRED, H_PRED, TM_PRED
};

460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
                         BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
  RD_COST this_rdc, best_rdc;
  PREDICTION_MODE this_mode;
  struct estimate_block_intra_args args = { cpi, x, DC_PRED, 0, 0 };
  const TX_SIZE intra_tx_size =
      MIN(max_txsize_lookup[bsize],
          tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
  MODE_INFO *const mic = xd->mi[0].src_mi;
  int *bmode_costs;
  const MODE_INFO *above_mi = xd->mi[-xd->mi_stride].src_mi;
  const MODE_INFO *left_mi = xd->left_available ? xd->mi[-1].src_mi : NULL;
  const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
  const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
  bmode_costs = cpi->y_mode_costs[A][L];

  (void) ctx;
  vp9_rd_cost_reset(&best_rdc);
  vp9_rd_cost_reset(&this_rdc);

  mbmi->ref_frame[0] = INTRA_FRAME;
  mbmi->mv[0].as_int = INVALID_MV;
  mbmi->uv_mode = DC_PRED;
  vpx_memset(x->skip_txfm, 0, sizeof(x->skip_txfm));

  // Change the limit of this loop to add other intra prediction
  // mode tests.
  for (this_mode = DC_PRED; this_mode <= H_PRED; ++this_mode) {
    args.mode = this_mode;
    args.rate = 0;
    args.dist = 0;
    mbmi->tx_size = intra_tx_size;
    vp9_foreach_transformed_block_in_plane(xd, bsize, 0,
                                           estimate_block_intra, &args);
    this_rdc.rate = args.rate;
    this_rdc.dist = args.dist;
    this_rdc.rate += bmode_costs[this_mode];
    this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
                             this_rdc.rate, this_rdc.dist);

    if (this_rdc.rdcost < best_rdc.rdcost) {
      best_rdc = this_rdc;
      mbmi->mode = this_mode;
    }
  }

  *rd_cost = best_rdc;
}

511
512
// TODO(jingning) placeholder for inter-frame non-RD mode decision.
// this needs various further optimizations. to be continued..
513
void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
514
                         TileDataEnc *tile_data,
515
516
                         int mi_row, int mi_col, RD_COST *rd_cost,
                         BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
517
  VP9_COMMON *const cm = &cpi->common;
518
  TileInfo *const tile_info = &tile_data->tile_info;
519
  MACROBLOCKD *const xd = &x->e_mbd;
hkuang's avatar
hkuang committed
520
  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
521
  struct macroblockd_plane *const pd = &xd->plane[0];
522
  PREDICTION_MODE best_mode = ZEROMV;
523
  MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME;
524
  TX_SIZE best_tx_size = TX_SIZES;
525
  INTERP_FILTER best_pred_filter = EIGHTTAP;
526
527
528
529
  int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
  struct buf_2d yv12_mb[4][MAX_MB_PLANE];
  static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
                                    VP9_ALT_FLAG };
530
  RD_COST this_rdc, best_rdc;
531
  uint8_t skip_txfm = 0, best_mode_skip_txfm = 0;
532
533
534
  // var_y and sse_y are saved to be used in skipping checking
  unsigned int var_y = UINT_MAX;
  unsigned int sse_y = UINT_MAX;
535
536
537
  // Reduce the intra cost penalty for small blocks (<=16x16).
  const int reduction_fac =
      (cpi->sf.partition_search_type == VAR_BASED_PARTITION &&
538
       bsize <= BLOCK_16X16) ? 2 : 0;
539
  const int intra_cost_penalty = vp9_get_intra_cost_penalty(
540
      cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth) >> reduction_fac;
541
542
  const int64_t inter_mode_thresh = RDCOST(x->rdmult, x->rddiv,
                                           intra_cost_penalty, 0);
543
  const int *const rd_threshes = cpi->rd.threshes[mbmi->segment_id][bsize];
544
  const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
Yaowu Xu's avatar
Yaowu Xu committed
545
  INTERP_FILTER filter_ref;
546
  const int bsl = mi_width_log2_lookup[bsize];
547
  const int pred_filter_search = cm->interp_filter == SWITCHABLE ?
548
549
      (((mi_row + mi_col) >> bsl) +
       get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
550
  int const_motion[MAX_REF_FRAMES] = { 0 };
551
552
  const int bh = num_4x4_blocks_high_lookup[bsize] << 2;
  const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
553
554
  // For speed 6, the result of interp filter is reused later in actual encoding
  // process.
555
556
557
  // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.
  PRED_BUFFER tmp[4];
  DECLARE_ALIGNED_ARRAY(16, uint8_t, pred_buf, 3 * 64 * 64);
558
559
560
#if CONFIG_VP9_HIGHBITDEPTH
  DECLARE_ALIGNED_ARRAY(16, uint16_t, pred_buf_16, 3 * 64 * 64);
#endif
561
562
563
  struct buf_2d orig_dst = pd->dst;
  PRED_BUFFER *best_pred = NULL;
  PRED_BUFFER *this_mode_pred = NULL;
564
  const int pixels_in_block = bh * bw;
565
  int reuse_inter_pred = cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready;
566
  int ref_frame_skip_mask = 0;
567

568
  if (reuse_inter_pred) {
569
    int i;
570
    for (i = 0; i < 3; i++) {
571
572
573
574
575
576
577
578
#if CONFIG_VP9_HIGHBITDEPTH
      if (cm->use_highbitdepth)
        tmp[i].data = CONVERT_TO_BYTEPTR(&pred_buf_16[pixels_in_block * i]);
      else
        tmp[i].data = &pred_buf[pixels_in_block * i];
#else
      tmp[i].data = &pred_buf[pixels_in_block * i];
#endif  // CONFIG_VP9_HIGHBITDEPTH
579
580
581
582
583
584
585
586
      tmp[i].stride = bw;
      tmp[i].in_use = 0;
    }
    tmp[3].data = pd->dst.buf;
    tmp[3].stride = pd->dst.stride;
    tmp[3].in_use = 0;
  }

587
588
  x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
  x->skip = 0;
589

Yaowu Xu's avatar
Yaowu Xu committed
590
591
592
593
594
595
596
  if (xd->up_available)
    filter_ref = xd->mi[-xd->mi_stride].src_mi->mbmi.interp_filter;
  else if (xd->left_available)
    filter_ref = xd->mi[-1].src_mi->mbmi.interp_filter;
  else
    filter_ref = cm->interp_filter;

597
  // initialize mode decisions
598
599
  vp9_rd_cost_reset(&best_rdc);
  vp9_rd_cost_reset(rd_cost);
600
601
602
603
  mbmi->sb_type = bsize;
  mbmi->ref_frame[0] = NONE;
  mbmi->ref_frame[1] = NONE;
  mbmi->tx_size = MIN(max_txsize_lookup[bsize],
604
605
606
                      tx_mode_to_biggest_tx_size[cm->tx_mode]);
  mbmi->interp_filter = cm->interp_filter == SWITCHABLE ?
                        EIGHTTAP : cm->interp_filter;
607

608
  for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
609
    x->pred_mv_sad[ref_frame] = INT_MAX;
610
611
612
    frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
    frame_mv[ZEROMV][ref_frame].as_int = 0;

613
    if (cpi->ref_frame_flags & flag_list[ref_frame]) {
614
615
616
      const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
      int_mv *const candidates = mbmi->ref_mvs[ref_frame];
      const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
617

618
619
620
      vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col,
                           sf, sf);

621
      if (cm->use_prev_frame_mvs)
622
        vp9_find_mv_refs(cm, xd, tile_info, xd->mi[0].src_mi, ref_frame,
623
624
                         candidates, mi_row, mi_col);
      else
625
626
        const_motion[ref_frame] = mv_refs_rt(cm, xd, tile_info,
                                             xd->mi[0].src_mi,
627
628
629
630
631
632
633
634
635
636
                                             ref_frame, candidates,
                                             mi_row, mi_col);

      vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
                            &frame_mv[NEARESTMV][ref_frame],
                            &frame_mv[NEARMV][ref_frame]);

      if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8)
        vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
                    ref_frame, bsize);
637
    } else {
638
      ref_frame_skip_mask |= (1 << ref_frame);
639
    }
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
  }

  for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
    PREDICTION_MODE this_mode;
    int i = (ref_frame == LAST_FRAME) ? GOLDEN_FRAME : LAST_FRAME;

    if (!(cpi->ref_frame_flags & flag_list[ref_frame]))
      continue;

    if (cpi->ref_frame_flags & flag_list[i])
      if (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[i] << 1))
        ref_frame_skip_mask |= (1 << ref_frame);

    if (ref_frame_skip_mask & (1 << ref_frame))
      continue;
655
656
657
658
659
660
661

    // Select prediction reference frames.
    xd->plane[0].pre[0] = yv12_mb[ref_frame][0];

    clamp_mv2(&frame_mv[NEARESTMV][ref_frame].as_mv, xd);
    clamp_mv2(&frame_mv[NEARMV][ref_frame].as_mv, xd);

662
663
    mbmi->ref_frame[0] = ref_frame;

664
    for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
665
      int rate_mv = 0;
666
      int mode_rd_thresh;
667
      int mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
668

669
      if (const_motion[ref_frame] && this_mode == NEARMV)
670
671
        continue;

672
      if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode)))
673
        continue;
Jim Bankoski's avatar
Jim Bankoski committed
674

675
676
      mode_rd_thresh = best_mode_skip_txfm ? rd_threshes[mode_index] << 1 :
                                             rd_threshes[mode_index];
677
      if (rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
678
                              rd_thresh_freq_fact[mode_index]))
679
680
        continue;

Jim Bankoski's avatar
Jim Bankoski committed
681
      if (this_mode == NEWMV) {
682
683
        if (ref_frame > LAST_FRAME)
          continue;
684
        if (cpi->sf.partition_search_type != VAR_BASED_PARTITION &&
685
            best_rdc.rdcost < (int64_t)(1 << num_pels_log2_lookup[bsize]))
686
          continue;
687
688
        if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
                                    &frame_mv[NEWMV][ref_frame],
689
                                    &rate_mv, best_rdc.rdcost))
Jim Bankoski's avatar
Jim Bankoski committed
690
691
692
          continue;
      }

693
694
695
      if (this_mode != NEARESTMV &&
          frame_mv[this_mode][ref_frame].as_int ==
              frame_mv[NEARESTMV][ref_frame].as_int)
696
        continue;
697

698
699
      mbmi->mode = this_mode;
      mbmi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
700

701
702
703
      // Search for the best prediction filter type, when the resulting
      // motion vector is at sub-pixel accuracy level for luma component, i.e.,
      // the last three bits are all zeros.
704
      if (reuse_inter_pred) {
705
        if (!this_mode_pred) {
706
707
708
709
710
711
712
713
          this_mode_pred = &tmp[3];
        } else {
          this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
          pd->dst.buf = this_mode_pred->data;
          pd->dst.stride = bw;
        }
      }

714
      if ((this_mode == NEWMV || filter_ref == SWITCHABLE) &&
715
          pred_filter_search &&
716
717
718
719
          ((mbmi->mv[0].as_mv.row & 0x07) != 0 ||
           (mbmi->mv[0].as_mv.col & 0x07) != 0)) {
        int pf_rate[3];
        int64_t pf_dist[3];
720
721
        unsigned int pf_var[3];
        unsigned int pf_sse[3];
722
        TX_SIZE pf_tx_size[3];
723
724
        int64_t best_cost = INT64_MAX;
        INTERP_FILTER best_filter = SWITCHABLE, filter;
725
        PRED_BUFFER *current_pred = this_mode_pred;
726
727
728
729
730
731

        for (filter = EIGHTTAP; filter <= EIGHTTAP_SHARP; ++filter) {
          int64_t cost;
          mbmi->interp_filter = filter;
          vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
          model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter],
732
                            &pf_dist[filter], &pf_var[filter], &pf_sse[filter]);
733
          cost = RDCOST(x->rdmult, x->rddiv,
734
                        vp9_get_switchable_rate(cpi, xd) + pf_rate[filter],
735
                        pf_dist[filter]);
736
          pf_tx_size[filter] = mbmi->tx_size;
737
          if (cost < best_cost) {
738
739
            best_filter = filter;
            best_cost = cost;
740
            skip_txfm = x->skip_txfm[0];
741

742
            if (reuse_inter_pred) {
743
744
745
746
747
748
749
750
751
752
753
              if (this_mode_pred != current_pred) {
                free_pred_buffer(this_mode_pred);
                this_mode_pred = current_pred;
              }

              if (filter < EIGHTTAP_SHARP) {
                current_pred = &tmp[get_pred_buffer(tmp, 3)];
                pd->dst.buf = current_pred->data;
                pd->dst.stride = bw;
              }
            }
754
          }
755
756
        }

757
        if (reuse_inter_pred && this_mode_pred != current_pred)
758
759
          free_pred_buffer(current_pred);

760
        mbmi->interp_filter = best_filter;
761
        mbmi->tx_size = pf_tx_size[mbmi->interp_filter];
762
763
        this_rdc.rate = pf_rate[mbmi->interp_filter];
        this_rdc.dist = pf_dist[mbmi->interp_filter];
764
765
        var_y = pf_var[mbmi->interp_filter];
        sse_y = pf_sse[mbmi->interp_filter];
766
        x->skip_txfm[0] = skip_txfm;
767
768
769
      } else {
        mbmi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP: filter_ref;
        vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
770
771
        model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
                          &var_y, &sse_y);
772
773
      }

774
775
      this_rdc.rate += rate_mv;
      this_rdc.rate += cpi->inter_mode_cost[mbmi->mode_context[ref_frame]]
776
                                  [INTER_OFFSET(this_mode)];
777
778
      this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
                               this_rdc.rate, this_rdc.dist);
779

780
781
      // Skipping checking: test to see if this block can be reconstructed by
      // prediction only.
782
      if (cpi->allow_encode_breakout) {
783
        encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame,
784
785
                             this_mode, var_y, sse_y, yv12_mb,
                             &this_rdc.rate, &this_rdc.dist);
786
        if (x->skip) {
787
788
789
          this_rdc.rate += rate_mv;
          this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
                                   this_rdc.rate, this_rdc.dist);
790
791
792
        }
      }

793
#if CONFIG_VP9_TEMPORAL_DENOISING
794
      if (cpi->oxcf.noise_sensitivity > 0)
795
        vp9_denoiser_update_frame_stats(mbmi, sse_y, this_mode, ctx);
796
797
#else
      (void)ctx;
Tim Kopp's avatar
Tim Kopp committed
798
799
#endif

800
801
      if (this_rdc.rdcost < best_rdc.rdcost || x->skip) {
        best_rdc = this_rdc;
802
        best_mode = this_mode;
803
        best_pred_filter = mbmi->interp_filter;
804
        best_tx_size = mbmi->tx_size;
805
        best_ref_frame = ref_frame;
806
        best_mode_skip_txfm = x->skip_txfm[0];
807

808
        if (reuse_inter_pred) {
809
          free_pred_buffer(best_pred);
810
811
812
          best_pred = this_mode_pred;
        }
      } else {
813
        if (reuse_inter_pred)
814
          free_pred_buffer(this_mode_pred);
815
      }
816
817
818

      if (x->skip)
        break;
819
    }
820
821
822

    // Check that a prediction mode has been selected.
    assert(best_rdc.rdcost < INT64_MAX);
823
824
825

    if (x->skip)
      break;
826
827
  }

828
  mbmi->mode          = best_mode;
829
  mbmi->interp_filter = best_pred_filter;
830
831
832
  mbmi->tx_size       = best_tx_size;
  mbmi->ref_frame[0]  = best_ref_frame;
  mbmi->mv[0].as_int  = frame_mv[best_mode][best_ref_frame].as_int;
hkuang's avatar
hkuang committed
833
  xd->mi[0].src_mi->bmi[0].as_mv[0].as_int = mbmi->mv[0].as_int;
834
  x->skip_txfm[0] = best_mode_skip_txfm;
835

836
837
  // Perform intra prediction search, if the best SAD is above a certain
  // threshold.
838
839
840
  if (best_rdc.rdcost == INT64_MAX ||
      (!x->skip && best_rdc.rdcost > inter_mode_thresh &&
       bsize <= cpi->sf.max_intra_bsize)) {
841
842
843
844
    struct estimate_block_intra_args args = { cpi, x, DC_PRED, 0, 0 };
    const TX_SIZE intra_tx_size =
        MIN(max_txsize_lookup[bsize],
            tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
845
    int i;
846
    TX_SIZE best_intra_tx_size = TX_SIZES;
847

848
849
850
    if (reuse_inter_pred && best_pred != NULL) {
      if (best_pred->data == orig_dst.buf) {
        this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
851
852
853
854
855
856
857
858
859
860
#if CONFIG_VP9_HIGHBITDEPTH
        if (cm->use_highbitdepth)
          vp9_highbd_convolve_copy(best_pred->data, best_pred->stride,
                                   this_mode_pred->data, this_mode_pred->stride,
                                   NULL, 0, NULL, 0, bw, bh, xd->bd);
        else
          vp9_convolve_copy(best_pred->data, best_pred->stride,
                          this_mode_pred->data, this_mode_pred->stride,
                          NULL, 0, NULL, 0, bw, bh);
#else
861
862
863
        vp9_convolve_copy(best_pred->data, best_pred->stride,
                          this_mode_pred->data, this_mode_pred->stride,
                          NULL, 0, NULL, 0, bw, bh);
864
#endif  // CONFIG_VP9_HIGHBITDEPTH
865
866
        best_pred = this_mode_pred;
      }
867
    }
868
    pd->dst = orig_dst;
869

870
871
872
873
    for (i = 0; i < 4; ++i) {
      const PREDICTION_MODE this_mode = intra_mode_list[i];
      if (!((1 << this_mode) & cpi->sf.intra_y_mode_mask[intra_tx_size]))
        continue;
874
875
876
877
878
879
      args.mode = this_mode;
      args.rate = 0;
      args.dist = 0;
      mbmi->tx_size = intra_tx_size;
      vp9_foreach_transformed_block_in_plane(xd, bsize, 0,
                                             estimate_block_intra, &args);
880
881
882
883
884
885
886
      this_rdc.rate = args.rate;
      this_rdc.dist = args.dist;
      this_rdc.rate += cpi->mbmode_cost[this_mode];
      this_rdc.rate += intra_cost_penalty;
      this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
                               this_rdc.rate, this_rdc.dist);

887
      if (this_rdc.rdcost < best_rdc.rdcost) {
888
        best_rdc = this_rdc;
889
        mbmi->mode = this_mode;
890
        best_intra_tx_size = mbmi->tx_size;
891
892
        mbmi->ref_frame[0] = INTRA_FRAME;
        mbmi->uv_mode = this_mode;
893
        mbmi->mv[0].as_int = INVALID_MV;
894
895
      }
    }
896
897
898
899
900

    // Reset mb_mode_info to the best inter mode.
    if (mbmi->ref_frame[0] != INTRA_FRAME) {
      x->skip_txfm[0] = best_mode_skip_txfm;
      mbmi->tx_size = best_tx_size;
901
902
    } else {
      mbmi->tx_size = best_intra_tx_size;
903
    }
904
905
906
  }

  pd->dst = orig_dst;
907
908
909

  if (reuse_inter_pred && best_pred != NULL) {
    if (best_pred->data != orig_dst.buf && is_inter_mode(mbmi->mode)) {
910
#if CONFIG_VP9_HIGHBITDEPTH
911
912
913
914
915
916
917
918
919
      if (cm->use_highbitdepth)
        vp9_highbd_convolve_copy(best_pred->data, best_pred->stride,
                                 pd->dst.buf, pd->dst.stride, NULL, 0,
                                 NULL, 0, bw, bh, xd->bd);
      else
        vp9_convolve_copy(best_pred->data, best_pred->stride,
                          pd->dst.buf, pd->dst.stride, NULL, 0,
                          NULL, 0, bw, bh);
#else
920
921
922
      vp9_convolve_copy(best_pred->data, best_pred->stride,
                        pd->dst.buf, pd->dst.stride, NULL, 0,
                        NULL, 0, bw, bh);
923
#endif  // CONFIG_VP9_HIGHBITDEPTH
924
    }
925
  }
926

927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
  if (cpi->sf.adaptive_rd_thresh) {
    THR_MODES best_mode_idx = is_inter_block(mbmi) ?
        mode_idx[best_ref_frame][INTER_OFFSET(mbmi->mode)] :
        mode_idx[INTRA_FRAME][mbmi->mode];
    PREDICTION_MODE this_mode;
    for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
      for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
        THR_MODES thr_mode_idx = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
        int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx];
        if (thr_mode_idx == best_mode_idx)
          *freq_fact -= (*freq_fact >> 4);
        else
          *freq_fact = MIN(*freq_fact + RD_THRESH_INC,
                           cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
      }
    }
  }
944

945
  *rd_cost = best_rdc;
946
}
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974

void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
                                TileDataEnc *tile_data,
                                int mi_row, int mi_col, RD_COST *rd_cost,
                                BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
  VP9_COMMON *const cm = &cpi->common;
  TileInfo *const tile_info = &tile_data->tile_info;
  SPEED_FEATURES *const sf = &cpi->sf;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
  const struct segmentation *const seg = &cm->seg;
  MV_REFERENCE_FRAME ref_frame, second_ref_frame = NONE;
  MV_REFERENCE_FRAME best_ref_frame = NONE;
  unsigned char segment_id = mbmi->segment_id;
  struct buf_2d yv12_mb[4][MAX_MB_PLANE];
  static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
                                    VP9_ALT_FLAG };
  int64_t best_rd = INT64_MAX;
  b_mode_info bsi[MAX_REF_FRAMES][4];
  int ref_frame_skip_mask = 0;
  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
  int idx, idy;

  x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
  ctx->pred_pixel_ready = 0;

  for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ref_frame++) {
975
976
977
978
979
980
981
    int_mv dummy_mv[2];
    x->pred_mv_sad[ref_frame] = INT_MAX;

    if (cpi->ref_frame_flags & flag_list[ref_frame]) {
      const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
      int_mv *const candidates = mbmi->ref_mvs[ref_frame];
      const struct scale_factors *const sf =
982
                             &cm->frame_refs[ref_frame - 1].sf;
983
984
985
986
      vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col,
                           sf, sf);
      vp9_find_mv_refs(cm, xd, tile_info, xd->mi[0].src_mi, ref_frame,
                       candidates, mi_row, mi_col);
987

988
989
990
991
992
      vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
                            &dummy_mv[0], &dummy_mv[1]);
    } else {
      ref_frame_skip_mask |= (1 << ref_frame);
    }
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
  }

  mbmi->sb_type = bsize;
  mbmi->tx_size = TX_4X4;
  mbmi->uv_mode = DC_PRED;
  mbmi->ref_frame[0] = LAST_FRAME;
  mbmi->ref_frame[1] = NONE;
  mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
                                                        : cm->interp_filter;

  for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
    int64_t this_rd = 0;
    int plane;

    if (ref_frame_skip_mask & (1 << ref_frame))
      continue;

    // TODO(jingning, agrange): Scaling reference frame not supported for
    // sub8x8 blocks. Is this supported now?
    if (ref_frame > INTRA_FRAME &&
        vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
      continue;

    // If the segment reference frame feature is enabled....
    // then do nothing if the current ref frame is not allowed..
    if (vp9_segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
        vp9_get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
      continue;

    mbmi->ref_frame[0] = ref_frame;
    x->skip = 0;
    set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);

    // Select prediction reference frames.
    for (plane = 0; plane < MAX_MB_PLANE; plane++)
      xd->plane[plane].pre[0] = yv12_mb[ref_frame][plane];

    for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
      for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1032
        int_mv b_mv[MB_MODE_COUNT];
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
        int64_t b_best_rd = INT64_MAX;
        const int i = idy * 2 + idx;
        PREDICTION_MODE this_mode;
        int b_rate = 0;
        int64_t b_dist = 0;
        RD_COST this_rdc;
        unsigned int var_y, sse_y;

        struct macroblock_plane *p = &x->plane[0];
        struct macroblockd_plane *pd = &xd->plane[0];

        const struct buf_2d orig_src = p->src;
        const struct buf_2d orig_dst = pd->dst;
        struct buf_2d orig_pre[2];
        vpx_memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre));

        // set buffer pointers for sub8x8 motion search.
        p->src.buf =
            &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
        pd->dst.buf =
            &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
        pd->pre[0].buf =
            &pd->pre[0].buf[vp9_raster_block_offset(BLOCK_8X8,
                                                    i, pd->pre[0].stride)];

1058
1059
        b_mv[ZEROMV].as_int = 0;
        b_mv[NEWMV].as_int = INVALID_MV;
1060
        vp9_append_sub8x8_mvs_for_idx(cm, xd, tile_info, i, 0, mi_row, mi_col,
1061
1062
                                      &b_mv[NEARESTMV],
                                      &b_mv[NEARMV]);
1063
1064

        for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1065
          xd->mi[0].bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int;
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078

          if (this_mode == NEWMV) {
            const int step_param = cpi->sf.mv.fullpel_search_step_param;
            MV mvp_full;
            MV tmp_mv;
            int cost_list[5];
            const int tmp_col_min = x->mv_col_min;
            const int tmp_col_max = x->mv_col_max;
            const int tmp_row_min = x->mv_row_min;
            const int tmp_row_max = x->mv_row_max;
            int dummy_dist;

            if (i == 0) {
1079
1080
              mvp_full.row = b_mv[NEARESTMV].as_mv.row >> 3;
              mvp_full.col = b_mv[NEARESTMV].as_mv.col >> 3;
1081
1082
1083
1084
1085
1086
1087
1088