vp9_pickmode.c 29.7 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
68
69
70
71
72
73
74
75
76
77
78
79
80
      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)
        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1));
    }
  }

  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
81
                                                    xd->mi_stride].src_mi->mbmi;
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
      different_ref_found = 1;

      if (candidate->ref_frame[0] == ref_frame)
        ADD_MV_REF_LIST(candidate->mv[0]);
    }
  }

  // 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
97
                                              * xd->mi_stride].src_mi->mbmi;
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115

        // If the candidate is INTRA we don't want to consider its mv.
        IF_DIFF_REF_FRAME_ADD_MV(candidate);
      }
    }
  }

 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;
}

116
117
118
119
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) {
120
  MACROBLOCKD *xd = &x->e_mbd;
hkuang's avatar
hkuang committed
121
  MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
122
  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
123
124
  const int step_param = cpi->sf.mv.fullpel_search_step_param;
  const int sadpb = x->sadperbit16;
125
  MV mvp_full;
126
  const int ref = mbmi->ref_frame[0];
127
  const MV ref_mv = mbmi->ref_mvs[ref][0].as_mv;
128
129
130
131
132
133
134
  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;
135
  int sad_list[5];
136
137
  const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi,
                                                                        ref);
Yaowu Xu's avatar
Yaowu Xu committed
138
139
140
141
  if (cpi->common.show_frame &&
      (x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[LAST_FRAME])
    return rv;

142
143
144
145
146
147
148
  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];
149
    vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
150
  }
151
  vp9_set_mv_search_range(x, &ref_mv);
152

Yaowu Xu's avatar
Yaowu Xu committed
153
154
155
156
  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
157
    mvp_full = x->pred_mv[ref];
158
159
160
161

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

162
163
164
  vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
                        cond_sad_list(cpi, sad_list),
                        &ref_mv, &tmp_mv->as_mv, INT_MAX, 0);
165

166
167
168
169
170
  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;

171
172
173
  // 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;
174

175
176
  *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv,
                             x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
177

178
179
180
181
182
183
184
185
186
187
188
189
  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,
190
                                 cond_sad_list(cpi, sad_list),
191
                                 x->nmvjointcost, x->mvcost,
192
                                 &dis, &x->pred_sse[ref], NULL, 0, 0);
193
    x->pred_mv[ref] = tmp_mv->as_mv;
194
195
196
197
198
199
200
  }

  if (scaled_ref_frame) {
    int i;
    for (i = 0; i < MAX_MB_PLANE; i++)
      xd->plane[i].pre[0] = backup_yv12[i];
  }
201
  return rv;
202
203
}

204

205
206
static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize,
                              MACROBLOCK *x, MACROBLOCKD *xd,
207
208
                              int *out_rate_sum, int64_t *out_dist_sum,
                              unsigned int *var_y, unsigned int *sse_y) {
209
210
211
212
  // 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;
213
214
  int rate;
  int64_t dist;
215
216
  struct macroblock_plane *const p = &x->plane[0];
  struct macroblockd_plane *const pd = &xd->plane[0];
217
218
  const uint32_t dc_quant = pd->dequant[0];
  const uint32_t ac_quant = pd->dequant[1];
219
220
  unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
                                           pd->dst.buf, pd->dst.stride, &sse);
221
222
223
  *var_y = var;
  *sse_y = sse;

224
  if (sse < dc_quant * dc_quant >> 6)
225
    x->skip_txfm[0] = 1;
226
  else if (var < ac_quant * ac_quant >> 6)
227
    x->skip_txfm[0] = 2;
228
  else
229
    x->skip_txfm[0] = 0;
230

231
232
  if (cpi->common.tx_mode == TX_MODE_SELECT) {
    if (sse > (var << 2))
hkuang's avatar
hkuang committed
233
      xd->mi[0].src_mi->mbmi.tx_size =
234
235
          MIN(max_txsize_lookup[bsize],
              tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
236
    else
hkuang's avatar
hkuang committed
237
      xd->mi[0].src_mi->mbmi.tx_size = TX_8X8;
238
239
240
241

    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;
242
  } else {
hkuang's avatar
hkuang committed
243
    xd->mi[0].src_mi->mbmi.tx_size =
244
245
        MIN(max_txsize_lookup[bsize],
            tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
246
247
  }

248
249
250
251
252
253
254
255
256
#if CONFIG_VP9_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    vp9_model_rd_from_var_lapndz(sse - var, 1 << num_pels_log2_lookup[bsize],
                                 dc_quant >> (xd->bd - 5), &rate, &dist);
  } else {
    vp9_model_rd_from_var_lapndz(sse - var, 1 << num_pels_log2_lookup[bsize],
                                 dc_quant >> 3, &rate, &dist);
  }
#else
257
258
  vp9_model_rd_from_var_lapndz(sse - var, 1 << num_pels_log2_lookup[bsize],
                               dc_quant >> 3, &rate, &dist);
259
260
#endif  // CONFIG_VP9_HIGHBITDEPTH

261
  *out_rate_sum = rate >> 1;
262
  *out_dist_sum = dist << 3;
263

264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
#if CONFIG_VP9_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    vp9_model_rd_from_var_lapndz(var,
                                 1 << num_pels_log2_lookup[bsize],
                                 ac_quant >> (xd->bd - 5),
                                 &rate,
                                 &dist);
  } else {
    vp9_model_rd_from_var_lapndz(var,
                                 1 << num_pels_log2_lookup[bsize],
                                 ac_quant >> 3,
                                 &rate,
                                 &dist);
  }
#else
  vp9_model_rd_from_var_lapndz(var,
                               1 << num_pels_log2_lookup[bsize],
                               ac_quant >> 3,
                               &rate,
                               &dist);
#endif  // CONFIG_VP9_HIGHBITDEPTH

286
287
  *out_rate_sum += rate;
  *out_dist_sum += dist << 4;
288
289
}

290
291
292
293
294
295
296
297
298
299
300
301
302
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) {
303
304
  if (p != NULL)
    p->in_use = 0;
305
306
}

307
308
309
310
311
312
313
314
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
315
  MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330

  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);
331
332
333
#if CONFIG_VP9_HIGHBITDEPTH
    const int shift = 2 * xd->bd - 16;
#endif
334
335
336

    // Calculate threshold according to dequant value.
    thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) / 9;
337
338
339
340
341
#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
342
343
344
345
    thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);

    // Adjust ac threshold according to partition size.
    thresh_ac >>=
346
        8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
347
348

    thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
349
350
351
352
353
#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
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
386
387
388
389
  } 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]]
390
                                    [INTER_OFFSET(this_mode)];
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406

        // 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;
      }
    }
  }
}

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
436
437
438
439
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),
440
                          b_width_log2_lookup[plane_bsize],
441
442
443
444
445
446
447
448
449
450
451
452
                          tx_size, args->mode,
                          p->src.buf, src_stride,
                          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;
}

453
static const THR_MODES mode_idx[MAX_REF_FRAMES - 1][INTER_MODES] = {
454
455
456
457
458
  {THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV},
  {THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG},
  {THR_NEARESTA, THR_NEARA, THR_ZEROA, THR_NEWA},
};

459
460
// TODO(jingning) placeholder for inter-frame non-RD mode decision.
// this needs various further optimizations. to be continued..
461
462
463
464
465
466
467
void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
                         const TileInfo *const tile,
                         int mi_row, int mi_col,
                         int *returnrate,
                         int64_t *returndistortion,
                         BLOCK_SIZE bsize,
                         PICK_MODE_CONTEXT *ctx) {
468
469
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;
hkuang's avatar
hkuang committed
470
  MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
471
  struct macroblockd_plane *const pd = &xd->plane[0];
472
  PREDICTION_MODE best_mode = ZEROMV;
473
  MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME;
474
  TX_SIZE best_tx_size = MIN(max_txsize_lookup[bsize],
475
                             tx_mode_to_biggest_tx_size[cm->tx_mode]);
476
  INTERP_FILTER best_pred_filter = EIGHTTAP;
477
478
479
480
481
  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 };
  int64_t best_rd = INT64_MAX;
482
  int64_t this_rd = INT64_MAX;
483
  uint8_t skip_txfm = 0;
484
485
  int rate = INT_MAX;
  int64_t dist = INT64_MAX;
486
487
488
  // 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;
489
490
  const int intra_cost_penalty = vp9_get_intra_cost_penalty(
      cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth);
491
492
  const int64_t inter_mode_thresh = RDCOST(x->rdmult, x->rddiv,
                                           intra_cost_penalty, 0);
493
  const int intra_mode_cost = 50;
494

495
  const int8_t segment_id = mbmi->segment_id;
Dmitry Kovalev's avatar
Dmitry Kovalev committed
496
497
  const int *const rd_threshes = cpi->rd.threshes[segment_id][bsize];
  const int *const rd_thresh_freq_fact = cpi->rd.thresh_freq_fact[bsize];
498
  INTERP_FILTER filter_ref = cm->interp_filter;
499
  const int bsl = mi_width_log2_lookup[bsize];
500
  const int pred_filter_search = cm->interp_filter == SWITCHABLE ?
501
502
      (((mi_row + mi_col) >> bsl) +
       get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
503
  int const_motion[MAX_REF_FRAMES] = { 0 };
504
505
  const int bh = num_4x4_blocks_high_lookup[bsize] << 2;
  const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
506
507
  // For speed 6, the result of interp filter is reused later in actual encoding
  // process.
508
509
510
  // 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);
511
512
513
#if CONFIG_VP9_HIGHBITDEPTH
  DECLARE_ALIGNED_ARRAY(16, uint16_t, pred_buf_16, 3 * 64 * 64);
#endif
514
515
516
  struct buf_2d orig_dst = pd->dst;
  PRED_BUFFER *best_pred = NULL;
  PRED_BUFFER *this_mode_pred = NULL;
517
  const int pixels_in_block = bh * bw;
518
519

  if (cpi->sf.reuse_inter_pred_sby) {
520
    int i;
521
    for (i = 0; i < 3; i++) {
522
523
524
525
526
527
528
529
#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
530
531
532
533
534
535
536
537
      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;
  }

538
539
  x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
  x->skip = 0;
540

541
542
  // initialize mode decisions
  *returnrate = INT_MAX;
543
  *returndistortion = INT64_MAX;
544
545
546
547
548
  vpx_memset(mbmi, 0, sizeof(MB_MODE_INFO));
  mbmi->sb_type = bsize;
  mbmi->ref_frame[0] = NONE;
  mbmi->ref_frame[1] = NONE;
  mbmi->tx_size = MIN(max_txsize_lookup[bsize],
549
550
551
                      tx_mode_to_biggest_tx_size[cm->tx_mode]);
  mbmi->interp_filter = cm->interp_filter == SWITCHABLE ?
                        EIGHTTAP : cm->interp_filter;
552
  mbmi->segment_id = segment_id;
553

554
  for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
555
    PREDICTION_MODE this_mode;
556
    x->pred_mv_sad[ref_frame] = INT_MAX;
557
558
559
560
    frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
    frame_mv[ZEROMV][ref_frame].as_int = 0;

    if (xd->up_available)
hkuang's avatar
hkuang committed
561
      filter_ref = xd->mi[-xd->mi_stride].src_mi->mbmi.interp_filter;
562
    else if (xd->left_available)
hkuang's avatar
hkuang committed
563
      filter_ref = xd->mi[-1].src_mi->mbmi.interp_filter;
564

565
    if (cpi->ref_frame_flags & flag_list[ref_frame]) {
566
567
568
569
570
571
      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;
      vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col,
                           sf, sf);

572
      if (!cm->error_resilient_mode)
hkuang's avatar
hkuang committed
573
        vp9_find_mv_refs(cm, xd, tile, xd->mi[0].src_mi, ref_frame,
574
575
                         candidates, mi_row, mi_col);
      else
hkuang's avatar
hkuang committed
576
        const_motion[ref_frame] = mv_refs_rt(cm, xd, tile, xd->mi[0].src_mi,
577
578
579
580
581
582
583
584
585
586
                                             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);
587
    } else {
588
      continue;
589
    }
590
591
592
593
594
595
596

    // 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);

597
598
    mbmi->ref_frame[0] = ref_frame;

599
    for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
600
      int rate_mv = 0;
601
      int mode_rd_thresh;
602

603
604
605
606
      if (const_motion[ref_frame] &&
          (this_mode == NEARMV || this_mode == ZEROMV))
        continue;

607
      if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode)))
608
        continue;
Jim Bankoski's avatar
Jim Bankoski committed
609

610
611
612
      mode_rd_thresh =
          rd_threshes[mode_idx[ref_frame -
                               LAST_FRAME][INTER_OFFSET(this_mode)]];
613
      if (rd_less_than_thresh(best_rd, mode_rd_thresh,
614
                              rd_thresh_freq_fact[this_mode]))
615
616
        continue;

Jim Bankoski's avatar
Jim Bankoski committed
617
      if (this_mode == NEWMV) {
618
619
        if (cpi->sf.partition_search_type != VAR_BASED_PARTITION &&
            this_rd < (int64_t)(1 << num_pels_log2_lookup[bsize]))
620
          continue;
621
622
623
        if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
                                    &frame_mv[NEWMV][ref_frame],
                                    &rate_mv, best_rd))
Jim Bankoski's avatar
Jim Bankoski committed
624
625
626
          continue;
      }

627
628
629
      if (this_mode != NEARESTMV &&
          frame_mv[this_mode][ref_frame].as_int ==
              frame_mv[NEARESTMV][ref_frame].as_int)
630
        continue;
631

632
633
      mbmi->mode = this_mode;
      mbmi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
634

635
636
637
      // 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.
638
639
640
641
642
643
644
645
646
647
      if (cpi->sf.reuse_inter_pred_sby) {
        if (this_mode == NEARESTMV) {
          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;
        }
      }

648
      if ((this_mode == NEWMV || filter_ref == SWITCHABLE) &&
649
          pred_filter_search &&
650
651
652
653
          ((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];
654
655
        unsigned int pf_var[3];
        unsigned int pf_sse[3];
656
        TX_SIZE pf_tx_size[3];
657
658
        int64_t best_cost = INT64_MAX;
        INTERP_FILTER best_filter = SWITCHABLE, filter;
659
        PRED_BUFFER *current_pred = this_mode_pred;
660
661
662
663
664
665

        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],
666
                            &pf_dist[filter], &pf_var[filter], &pf_sse[filter]);
667
668
669
          cost = RDCOST(x->rdmult, x->rddiv,
                        vp9_get_switchable_rate(cpi) + pf_rate[filter],
                        pf_dist[filter]);
670
          pf_tx_size[filter] = mbmi->tx_size;
671
          if (cost < best_cost) {
672
673
            best_filter = filter;
            best_cost = cost;
674
            skip_txfm = x->skip_txfm[0];
675
676
677
678
679
680
681
682
683
684
685
686
687

            if (cpi->sf.reuse_inter_pred_sby) {
              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;
              }
            }
688
          }
689
690
        }

691
692
693
        if (cpi->sf.reuse_inter_pred_sby && this_mode_pred != current_pred)
          free_pred_buffer(current_pred);

694
        mbmi->interp_filter = best_filter;
695
        mbmi->tx_size = pf_tx_size[mbmi->interp_filter];
696
697
        rate = pf_rate[mbmi->interp_filter];
        dist = pf_dist[mbmi->interp_filter];
698
699
        var_y = pf_var[mbmi->interp_filter];
        sse_y = pf_sse[mbmi->interp_filter];
700
        x->skip_txfm[0] = skip_txfm;
701
702
703
      } else {
        mbmi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP: filter_ref;
        vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
704
        model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);
705
706
      }

707
      rate += rate_mv;
708
      rate += cpi->inter_mode_cost[mbmi->mode_context[ref_frame]]
709
                                  [INTER_OFFSET(this_mode)];
710
      this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
711

712
713
      // Skipping checking: test to see if this block can be reconstructed by
      // prediction only.
714
      if (cpi->allow_encode_breakout) {
715
716
717
718
719
        encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame,
                             this_mode, var_y, sse_y, yv12_mb, &rate, &dist);
        if (x->skip) {
          rate += rate_mv;
          this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
720
721
722
        }
      }

723
#if CONFIG_VP9_TEMPORAL_DENOISING
724
      if (cpi->oxcf.noise_sensitivity > 0) {
725
        vp9_denoiser_update_frame_stats(mbmi, sse_y, this_mode, ctx);
726
      }
727
728
#else
      (void)ctx;
Tim Kopp's avatar
Tim Kopp committed
729
730
#endif

731
      if (this_rd < best_rd || x->skip) {
732
        best_rd = this_rd;
733
734
        *returnrate = rate;
        *returndistortion = dist;
735
        best_mode = this_mode;
736
        best_pred_filter = mbmi->interp_filter;
737
        best_tx_size = mbmi->tx_size;
738
        best_ref_frame = ref_frame;
739
        skip_txfm = x->skip_txfm[0];
740
741

        if (cpi->sf.reuse_inter_pred_sby) {
742
          free_pred_buffer(best_pred);
743
744
745
746
747
748

          best_pred = this_mode_pred;
        }
      } else {
        if (cpi->sf.reuse_inter_pred_sby)
          free_pred_buffer(this_mode_pred);
749
      }
750
751
752

      if (x->skip)
        break;
753
    }
754
755
756
757
    // If the current reference frame is valid and we found a usable mode,
    // we are done.
    if (best_rd < INT64_MAX)
      break;
758
759
  }

760
761
  // If best prediction is not in dst buf, then copy the prediction block from
  // temp buf to dst buf.
762
763
  if (best_pred != NULL && cpi->sf.reuse_inter_pred_sby &&
      best_pred->data != orig_dst.buf) {
764
    pd->dst = orig_dst;
765
766
#if CONFIG_VP9_HIGHBITDEPTH
    if (cm->use_highbitdepth) {
767
768
      vp9_highbd_convolve_copy(best_pred->data, bw, pd->dst.buf, pd->dst.stride,
                               NULL, 0, NULL, 0, bw, bh, xd->bd);
769
770
771
772
773
    } else {
      vp9_convolve_copy(best_pred->data, bw, pd->dst.buf, pd->dst.stride,
                        NULL, 0, NULL, 0, bw, bh);
    }
#else
774
775
    vp9_convolve_copy(best_pred->data, bw, pd->dst.buf, pd->dst.stride, NULL, 0,
                      NULL, 0, bw, bh);
776
#endif  // CONFIG_VP9_HIGHBITDEPTH
777
  }
Tim Kopp's avatar
Tim Kopp committed
778

779
  mbmi->mode          = best_mode;
780
  mbmi->interp_filter = best_pred_filter;
781
782
783
  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
784
  xd->mi[0].src_mi->bmi[0].as_mv[0].as_int = mbmi->mv[0].as_int;
785
  x->skip_txfm[0] = skip_txfm;
786

787
788
  // Perform intra prediction search, if the best SAD is above a certain
  // threshold.
789
  if (!x->skip && best_rd > inter_mode_thresh &&
790
      bsize <= cpi->sf.max_intra_bsize) {
791
    PREDICTION_MODE this_mode;
792
793
794
795
    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]);
796

797
798
799
800
    if (cpi->sf.reuse_inter_pred_sby) {
      pd->dst.buf = tmp[0].data;
      pd->dst.stride = bw;
    }
801

802
    for (this_mode = DC_PRED; this_mode <= DC_PRED; ++this_mode) {
803
804
805
806
807
808
809
810
811
812
      const TX_SIZE saved_tx_size = mbmi->tx_size;
      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);
      mbmi->tx_size = saved_tx_size;
      rate = args.rate;
      dist = args.dist;
813
      rate += cpi->mbmode_cost[this_mode];
814
      rate += intra_cost_penalty;
815
      this_rd = RDCOST(x->rdmult, x->rddiv, rate, dist);
816
817
818

      if (this_rd + intra_mode_cost < best_rd) {
        best_rd = this_rd;
819
820
        *returnrate = rate;
        *returndistortion = dist;
821
        mbmi->mode = this_mode;
822
        mbmi->tx_size = intra_tx_size;
823
824
        mbmi->ref_frame[0] = INTRA_FRAME;
        mbmi->uv_mode = this_mode;
825
        mbmi->mv[0].as_int = INVALID_MV;
826
      } else {
827
        x->skip_txfm[0] = skip_txfm;
828
829
      }
    }
830
831
    if (cpi->sf.reuse_inter_pred_sby)
      pd->dst = orig_dst;
832
  }
833
}