1/*
2 *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
3 *
4 *  Use of this source code is governed by a BSD-style license
5 *  that can be found in the LICENSE file in the root of the source
6 *  tree. An additional intellectual property rights grant can be found
7 *  in the file PATENTS.  All contributing project authors may
8 *  be found in the AUTHORS file in the root of the source tree.
9 */
10
11#include <assert.h>
12#include <limits.h>
13#include <math.h>
14#include <stdio.h>
15
16#include "./vp9_rtcd.h"
17#include "./vpx_dsp_rtcd.h"
18
19#include "vpx_dsp/vpx_dsp_common.h"
20#include "vpx_mem/vpx_mem.h"
21#include "vpx_ports/mem.h"
22
23#include "vp9/common/vp9_blockd.h"
24#include "vp9/common/vp9_common.h"
25#include "vp9/common/vp9_mvref_common.h"
26#include "vp9/common/vp9_pred_common.h"
27#include "vp9/common/vp9_reconinter.h"
28#include "vp9/common/vp9_reconintra.h"
29#include "vp9/common/vp9_scan.h"
30
31#include "vp9/encoder/vp9_cost.h"
32#include "vp9/encoder/vp9_encoder.h"
33#include "vp9/encoder/vp9_pickmode.h"
34#include "vp9/encoder/vp9_ratectrl.h"
35#include "vp9/encoder/vp9_rd.h"
36
37typedef struct {
38  uint8_t *data;
39  int stride;
40  int in_use;
41} PRED_BUFFER;
42
43static int mv_refs_rt(const VP9_COMMON *cm, const MACROBLOCK *x,
44                      const MACROBLOCKD *xd,
45                      const TileInfo *const tile,
46                      MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
47                      int_mv *mv_ref_list,
48                      int mi_row, int mi_col) {
49  const int *ref_sign_bias = cm->ref_frame_sign_bias;
50  int i, refmv_count = 0;
51
52  const POSITION *const mv_ref_search = mv_ref_blocks[mi->mbmi.sb_type];
53
54  int different_ref_found = 0;
55  int context_counter = 0;
56  int const_motion = 0;
57
58  // Blank the reference vector list
59  memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
60
61  // The nearest 2 blocks are treated differently
62  // if the size < 8x8 we get the mv from the bmi substructure,
63  // and we also need to keep a mode count.
64  for (i = 0; i < 2; ++i) {
65    const POSITION *const mv_ref = &mv_ref_search[i];
66    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
67      const MODE_INFO *const candidate_mi = xd->mi[mv_ref->col + mv_ref->row *
68                                                   xd->mi_stride];
69      const MB_MODE_INFO *const candidate = &candidate_mi->mbmi;
70      // Keep counts for entropy encoding.
71      context_counter += mode_2_counter[candidate->mode];
72      different_ref_found = 1;
73
74      if (candidate->ref_frame[0] == ref_frame)
75        ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1),
76                        refmv_count, mv_ref_list, Done);
77    }
78  }
79
80  const_motion = 1;
81
82  // Check the rest of the neighbors in much the same way
83  // as before except we don't need to keep track of sub blocks or
84  // mode counts.
85  for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) {
86    const POSITION *const mv_ref = &mv_ref_search[i];
87    if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
88      const MB_MODE_INFO *const candidate = &xd->mi[mv_ref->col + mv_ref->row *
89                                                    xd->mi_stride]->mbmi;
90      different_ref_found = 1;
91
92      if (candidate->ref_frame[0] == ref_frame)
93        ADD_MV_REF_LIST(candidate->mv[0], refmv_count, mv_ref_list, Done);
94    }
95  }
96
97  // Since we couldn't find 2 mvs from the same reference frame
98  // go back through the neighbors and find motion vectors from
99  // different reference frames.
100  if (different_ref_found && !refmv_count) {
101    for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
102      const POSITION *mv_ref = &mv_ref_search[i];
103      if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
104        const MB_MODE_INFO *const candidate = &xd->mi[mv_ref->col + mv_ref->row
105                                              * xd->mi_stride]->mbmi;
106
107        // If the candidate is INTRA we don't want to consider its mv.
108        IF_DIFF_REF_FRAME_ADD_MV(candidate, ref_frame, ref_sign_bias,
109                                 refmv_count, mv_ref_list, Done);
110      }
111    }
112  }
113
114 Done:
115
116  x->mbmi_ext->mode_context[ref_frame] = counter_to_context[context_counter];
117
118  // Clamp vectors
119  for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
120    clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
121
122  return const_motion;
123}
124
125static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
126                                  BLOCK_SIZE bsize, int mi_row, int mi_col,
127                                  int_mv *tmp_mv, int *rate_mv,
128                                  int64_t best_rd_sofar) {
129  MACROBLOCKD *xd = &x->e_mbd;
130  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
131  struct buf_2d backup_yv12[MAX_MB_PLANE] = {{0, 0}};
132  const int step_param = cpi->sf.mv.fullpel_search_step_param;
133  const int sadpb = x->sadperbit16;
134  MV mvp_full;
135  const int ref = mbmi->ref_frame[0];
136  const MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
137  int dis;
138  int rate_mode;
139  const int tmp_col_min = x->mv_col_min;
140  const int tmp_col_max = x->mv_col_max;
141  const int tmp_row_min = x->mv_row_min;
142  const int tmp_row_max = x->mv_row_max;
143  int rv = 0;
144  int cost_list[5];
145  const YV12_BUFFER_CONFIG *scaled_ref_frame = vp9_get_scaled_ref_frame(cpi,
146                                                                        ref);
147  if (scaled_ref_frame) {
148    int i;
149    // Swap out the reference frame for a version that's been scaled to
150    // match the resolution of the current frame, allowing the existing
151    // motion search code to be used without additional modifications.
152    for (i = 0; i < MAX_MB_PLANE; i++)
153      backup_yv12[i] = xd->plane[i].pre[0];
154    vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
155  }
156  vp9_set_mv_search_range(x, &ref_mv);
157
158  assert(x->mv_best_ref_index[ref] <= 2);
159  if (x->mv_best_ref_index[ref] < 2)
160    mvp_full = x->mbmi_ext->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
161  else
162    mvp_full = x->pred_mv[ref];
163
164  mvp_full.col >>= 3;
165  mvp_full.row >>= 3;
166
167  vp9_full_pixel_search(cpi, x, bsize, &mvp_full, step_param, sadpb,
168                        cond_cost_list(cpi, cost_list),
169                        &ref_mv, &tmp_mv->as_mv, INT_MAX, 0);
170
171  x->mv_col_min = tmp_col_min;
172  x->mv_col_max = tmp_col_max;
173  x->mv_row_min = tmp_row_min;
174  x->mv_row_max = tmp_row_max;
175
176  // calculate the bit cost on motion vector
177  mvp_full.row = tmp_mv->as_mv.row * 8;
178  mvp_full.col = tmp_mv->as_mv.col * 8;
179
180  *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv,
181                             x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
182
183  rate_mode = cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref]]
184                                  [INTER_OFFSET(NEWMV)];
185  rv = !(RDCOST(x->rdmult, x->rddiv, (*rate_mv + rate_mode), 0) >
186         best_rd_sofar);
187
188  if (rv) {
189    cpi->find_fractional_mv_step(x, &tmp_mv->as_mv, &ref_mv,
190                                 cpi->common.allow_high_precision_mv,
191                                 x->errorperbit,
192                                 &cpi->fn_ptr[bsize],
193                                 cpi->sf.mv.subpel_force_stop,
194                                 cpi->sf.mv.subpel_iters_per_step,
195                                 cond_cost_list(cpi, cost_list),
196                                 x->nmvjointcost, x->mvcost,
197                                 &dis, &x->pred_sse[ref], NULL, 0, 0);
198    *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv,
199                               x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
200  }
201
202  if (scaled_ref_frame) {
203    int i;
204    for (i = 0; i < MAX_MB_PLANE; i++)
205      xd->plane[i].pre[0] = backup_yv12[i];
206  }
207  return rv;
208}
209
210static void block_variance(const uint8_t *src, int src_stride,
211                           const uint8_t *ref, int ref_stride,
212                           int w, int h, unsigned int *sse, int *sum,
213                           int block_size, unsigned int *sse8x8,
214                           int *sum8x8, unsigned int *var8x8) {
215  int i, j, k = 0;
216
217  *sse = 0;
218  *sum = 0;
219
220  for (i = 0; i < h; i += block_size) {
221    for (j = 0; j < w; j += block_size) {
222      vpx_get8x8var(src + src_stride * i + j, src_stride,
223                    ref + ref_stride * i + j, ref_stride,
224                    &sse8x8[k], &sum8x8[k]);
225      *sse += sse8x8[k];
226      *sum += sum8x8[k];
227      var8x8[k] = sse8x8[k] - (((unsigned int)sum8x8[k] * sum8x8[k]) >> 6);
228      k++;
229    }
230  }
231}
232
233static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
234                               unsigned int *sse_i, int *sum_i,
235                               unsigned int *var_o, unsigned int *sse_o,
236                               int *sum_o) {
237  const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
238  const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
239  const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
240  int i, j, k = 0;
241
242  for (i = 0; i < nh; i += 2) {
243    for (j = 0; j < nw; j += 2) {
244      sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] +
245          sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1];
246      sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] +
247          sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1];
248      var_o[k] = sse_o[k] - (((unsigned int)sum_o[k] * sum_o[k]) >>
249          (b_width_log2_lookup[unit_size] +
250              b_height_log2_lookup[unit_size] + 6));
251      k++;
252    }
253  }
254}
255
256static void model_rd_for_sb_y_large(VP9_COMP *cpi, BLOCK_SIZE bsize,
257                                    MACROBLOCK *x, MACROBLOCKD *xd,
258                                    int *out_rate_sum, int64_t *out_dist_sum,
259                                    unsigned int *var_y, unsigned int *sse_y,
260                                    int mi_row, int mi_col, int *early_term) {
261  // Note our transform coeffs are 8 times an orthogonal transform.
262  // Hence quantizer step is also 8 times. To get effective quantizer
263  // we need to divide by 8 before sending to modeling function.
264  unsigned int sse;
265  int rate;
266  int64_t dist;
267  struct macroblock_plane *const p = &x->plane[0];
268  struct macroblockd_plane *const pd = &xd->plane[0];
269  const uint32_t dc_quant = pd->dequant[0];
270  const uint32_t ac_quant = pd->dequant[1];
271  const int64_t dc_thr = dc_quant * dc_quant >> 6;
272  const int64_t ac_thr = ac_quant * ac_quant >> 6;
273  unsigned int var;
274  int sum;
275  int skip_dc = 0;
276
277  const int bw = b_width_log2_lookup[bsize];
278  const int bh = b_height_log2_lookup[bsize];
279  const int num8x8 = 1 << (bw + bh - 2);
280  unsigned int sse8x8[64] = {0};
281  int sum8x8[64] = {0};
282  unsigned int var8x8[64] = {0};
283  TX_SIZE tx_size;
284  int i, k;
285
286  // Calculate variance for whole partition, and also save 8x8 blocks' variance
287  // to be used in following transform skipping test.
288  block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
289                 4 << bw, 4 << bh, &sse, &sum, 8, sse8x8, sum8x8, var8x8);
290  var = sse - (((int64_t)sum * sum) >> (bw + bh + 4));
291
292  *var_y = var;
293  *sse_y = sse;
294
295  if (cpi->common.tx_mode == TX_MODE_SELECT) {
296    if (sse > (var << 2))
297      tx_size = VPXMIN(max_txsize_lookup[bsize],
298                       tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
299    else
300      tx_size = TX_8X8;
301
302    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
303        cyclic_refresh_segment_id_boosted(xd->mi[0]->mbmi.segment_id))
304      tx_size = TX_8X8;
305    else if (tx_size > TX_16X16)
306      tx_size = TX_16X16;
307  } else {
308    tx_size = VPXMIN(max_txsize_lookup[bsize],
309                     tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
310  }
311
312  assert(tx_size >= TX_8X8);
313  xd->mi[0]->mbmi.tx_size = tx_size;
314
315  // Evaluate if the partition block is a skippable block in Y plane.
316  {
317    unsigned int sse16x16[16] = {0};
318    int sum16x16[16] = {0};
319    unsigned int var16x16[16] = {0};
320    const int num16x16 = num8x8 >> 2;
321
322    unsigned int sse32x32[4] = {0};
323    int sum32x32[4] = {0};
324    unsigned int var32x32[4] = {0};
325    const int num32x32 = num8x8 >> 4;
326
327    int ac_test = 1;
328    int dc_test = 1;
329    const int num = (tx_size == TX_8X8) ? num8x8 :
330        ((tx_size == TX_16X16) ? num16x16 : num32x32);
331    const unsigned int *sse_tx = (tx_size == TX_8X8) ? sse8x8 :
332        ((tx_size == TX_16X16) ? sse16x16 : sse32x32);
333    const unsigned int *var_tx = (tx_size == TX_8X8) ? var8x8 :
334        ((tx_size == TX_16X16) ? var16x16 : var32x32);
335
336    // Calculate variance if tx_size > TX_8X8
337    if (tx_size >= TX_16X16)
338      calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
339                         sum16x16);
340    if (tx_size == TX_32X32)
341      calculate_variance(bw, bh, TX_16X16, sse16x16, sum16x16, var32x32,
342                         sse32x32, sum32x32);
343
344    // Skipping test
345    x->skip_txfm[0] = SKIP_TXFM_NONE;
346    for (k = 0; k < num; k++)
347      // Check if all ac coefficients can be quantized to zero.
348      if (!(var_tx[k] < ac_thr || var == 0)) {
349        ac_test = 0;
350        break;
351      }
352
353    for (k = 0; k < num; k++)
354      // Check if dc coefficient can be quantized to zero.
355      if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
356        dc_test = 0;
357        break;
358      }
359
360    if (ac_test) {
361      x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
362
363      if (dc_test)
364        x->skip_txfm[0] = SKIP_TXFM_AC_DC;
365    } else if (dc_test) {
366      skip_dc = 1;
367    }
368  }
369
370  if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
371    int skip_uv[2] = {0};
372    unsigned int var_uv[2];
373    unsigned int sse_uv[2];
374
375    *out_rate_sum = 0;
376    *out_dist_sum = sse << 4;
377
378    // Transform skipping test in UV planes.
379    for (i = 1; i <= 2; i++) {
380      struct macroblock_plane *const p = &x->plane[i];
381      struct macroblockd_plane *const pd = &xd->plane[i];
382      const TX_SIZE uv_tx_size = get_uv_tx_size(&xd->mi[0]->mbmi, pd);
383      const BLOCK_SIZE unit_size = txsize_to_bsize[uv_tx_size];
384      const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, pd);
385      const int uv_bw = b_width_log2_lookup[uv_bsize];
386      const int uv_bh = b_height_log2_lookup[uv_bsize];
387      const int sf = (uv_bw - b_width_log2_lookup[unit_size]) +
388          (uv_bh - b_height_log2_lookup[unit_size]);
389      const uint32_t uv_dc_thr = pd->dequant[0] * pd->dequant[0] >> (6 - sf);
390      const uint32_t uv_ac_thr = pd->dequant[1] * pd->dequant[1] >> (6 - sf);
391      int j = i - 1;
392
393      vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, i);
394      var_uv[j] = cpi->fn_ptr[uv_bsize].vf(p->src.buf, p->src.stride,
395          pd->dst.buf, pd->dst.stride, &sse_uv[j]);
396
397      if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
398          (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
399        skip_uv[j] = 1;
400      else
401        break;
402    }
403
404    // If the transform in YUV planes are skippable, the mode search checks
405    // fewer inter modes and doesn't check intra modes.
406    if (skip_uv[0] & skip_uv[1]) {
407      *early_term = 1;
408    }
409
410    return;
411  }
412
413  if (!skip_dc) {
414#if CONFIG_VP9_HIGHBITDEPTH
415    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
416      vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
417                                   dc_quant >> (xd->bd - 5), &rate, &dist);
418    } else {
419      vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
420                                   dc_quant >> 3, &rate, &dist);
421    }
422#else
423    vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
424                                 dc_quant >> 3, &rate, &dist);
425#endif  // CONFIG_VP9_HIGHBITDEPTH
426  }
427
428  if (!skip_dc) {
429    *out_rate_sum = rate >> 1;
430    *out_dist_sum = dist << 3;
431  } else {
432    *out_rate_sum = 0;
433    *out_dist_sum = (sse - var) << 4;
434  }
435
436#if CONFIG_VP9_HIGHBITDEPTH
437  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
438    vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
439                                 ac_quant >> (xd->bd - 5), &rate, &dist);
440  } else {
441    vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
442                                 ac_quant >> 3, &rate, &dist);
443  }
444#else
445  vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
446                               ac_quant >> 3, &rate, &dist);
447#endif  // CONFIG_VP9_HIGHBITDEPTH
448
449  *out_rate_sum += rate;
450  *out_dist_sum += dist << 4;
451}
452
453static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize,
454                              MACROBLOCK *x, MACROBLOCKD *xd,
455                              int *out_rate_sum, int64_t *out_dist_sum,
456                              unsigned int *var_y, unsigned int *sse_y) {
457  // Note our transform coeffs are 8 times an orthogonal transform.
458  // Hence quantizer step is also 8 times. To get effective quantizer
459  // we need to divide by 8 before sending to modeling function.
460  unsigned int sse;
461  int rate;
462  int64_t dist;
463  struct macroblock_plane *const p = &x->plane[0];
464  struct macroblockd_plane *const pd = &xd->plane[0];
465  const int64_t dc_thr = p->quant_thred[0] >> 6;
466  const int64_t ac_thr = p->quant_thred[1] >> 6;
467  const uint32_t dc_quant = pd->dequant[0];
468  const uint32_t ac_quant = pd->dequant[1];
469  unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
470                                           pd->dst.buf, pd->dst.stride, &sse);
471  int skip_dc = 0;
472
473  *var_y = var;
474  *sse_y = sse;
475
476  if (cpi->common.tx_mode == TX_MODE_SELECT) {
477    if (sse > (var << 2))
478      xd->mi[0]->mbmi.tx_size =
479          VPXMIN(max_txsize_lookup[bsize],
480                 tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
481    else
482      xd->mi[0]->mbmi.tx_size = TX_8X8;
483
484    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
485        cyclic_refresh_segment_id_boosted(xd->mi[0]->mbmi.segment_id))
486      xd->mi[0]->mbmi.tx_size = TX_8X8;
487    else if (xd->mi[0]->mbmi.tx_size > TX_16X16)
488      xd->mi[0]->mbmi.tx_size = TX_16X16;
489  } else {
490    xd->mi[0]->mbmi.tx_size =
491        VPXMIN(max_txsize_lookup[bsize],
492               tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
493  }
494
495  // Evaluate if the partition block is a skippable block in Y plane.
496  {
497    const BLOCK_SIZE unit_size =
498        txsize_to_bsize[xd->mi[0]->mbmi.tx_size];
499    const unsigned int num_blk_log2 =
500        (b_width_log2_lookup[bsize] - b_width_log2_lookup[unit_size]) +
501        (b_height_log2_lookup[bsize] - b_height_log2_lookup[unit_size]);
502    const unsigned int sse_tx = sse >> num_blk_log2;
503    const unsigned int var_tx = var >> num_blk_log2;
504
505    x->skip_txfm[0] = SKIP_TXFM_NONE;
506    // Check if all ac coefficients can be quantized to zero.
507    if (var_tx < ac_thr || var == 0) {
508      x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
509      // Check if dc coefficient can be quantized to zero.
510      if (sse_tx - var_tx < dc_thr || sse == var)
511        x->skip_txfm[0] = SKIP_TXFM_AC_DC;
512    } else {
513      if (sse_tx - var_tx < dc_thr || sse == var)
514        skip_dc = 1;
515    }
516  }
517
518  if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
519    *out_rate_sum = 0;
520    *out_dist_sum = sse << 4;
521    return;
522  }
523
524  if (!skip_dc) {
525#if CONFIG_VP9_HIGHBITDEPTH
526    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
527      vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
528                                   dc_quant >> (xd->bd - 5), &rate, &dist);
529    } else {
530      vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
531                                   dc_quant >> 3, &rate, &dist);
532    }
533#else
534    vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
535                                 dc_quant >> 3, &rate, &dist);
536#endif  // CONFIG_VP9_HIGHBITDEPTH
537  }
538
539  if (!skip_dc) {
540    *out_rate_sum = rate >> 1;
541    *out_dist_sum = dist << 3;
542  } else {
543    *out_rate_sum = 0;
544    *out_dist_sum = (sse - var) << 4;
545  }
546
547#if CONFIG_VP9_HIGHBITDEPTH
548  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
549    vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
550                                 ac_quant >> (xd->bd - 5), &rate, &dist);
551  } else {
552    vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
553                                 ac_quant >> 3, &rate, &dist);
554  }
555#else
556  vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
557                               ac_quant >> 3, &rate, &dist);
558#endif  // CONFIG_VP9_HIGHBITDEPTH
559
560  *out_rate_sum += rate;
561  *out_dist_sum += dist << 4;
562}
563
564#if CONFIG_VP9_HIGHBITDEPTH
565static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate, int64_t *dist,
566                      int *skippable, int64_t *sse, int plane,
567                      BLOCK_SIZE bsize, TX_SIZE tx_size) {
568  MACROBLOCKD *xd = &x->e_mbd;
569  unsigned int var_y, sse_y;
570  (void)plane;
571  (void)tx_size;
572  model_rd_for_sb_y(cpi, bsize, x, xd, rate, dist, &var_y, &sse_y);
573  *sse = INT_MAX;
574  *skippable = 0;
575  return;
576}
577#else
578static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate, int64_t *dist,
579                      int *skippable, int64_t *sse, int plane,
580                      BLOCK_SIZE bsize, TX_SIZE tx_size) {
581  MACROBLOCKD *xd = &x->e_mbd;
582  const struct macroblockd_plane *pd = &xd->plane[plane];
583  const struct macroblock_plane *const p = &x->plane[plane];
584  const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
585  const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
586  const int step = 1 << (tx_size << 1);
587  const int block_step = (1 << tx_size);
588  int block = 0, r, c;
589  int shift = tx_size == TX_32X32 ? 0 : 2;
590  const int max_blocks_wide = num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 :
591      xd->mb_to_right_edge >> (5 + pd->subsampling_x));
592  const int max_blocks_high = num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 :
593      xd->mb_to_bottom_edge >> (5 + pd->subsampling_y));
594  int eob_cost = 0;
595
596  (void)cpi;
597  vp9_subtract_plane(x, bsize, plane);
598  *skippable = 1;
599  // Keep track of the row and column of the blocks we use so that we know
600  // if we are in the unrestricted motion border.
601  for (r = 0; r < max_blocks_high; r += block_step) {
602    for (c = 0; c < num_4x4_w; c += block_step) {
603      if (c < max_blocks_wide) {
604        const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
605        tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
606        tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
607        tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
608        uint16_t *const eob = &p->eobs[block];
609        const int diff_stride = 4 * num_4x4_blocks_wide_lookup[bsize];
610        const int16_t *src_diff;
611        src_diff = &p->src_diff[(r * diff_stride + c) << 2];
612
613        switch (tx_size) {
614          case TX_32X32:
615            vpx_fdct32x32_rd(src_diff, coeff, diff_stride);
616            vp9_quantize_fp_32x32(coeff, 1024, x->skip_block, p->zbin,
617                                  p->round_fp, p->quant_fp, p->quant_shift,
618                                  qcoeff, dqcoeff, pd->dequant, eob,
619                                  scan_order->scan, scan_order->iscan);
620            break;
621          case TX_16X16:
622            vp9_hadamard_16x16(src_diff, diff_stride, (int16_t *)coeff);
623            vp9_quantize_fp(coeff, 256, x->skip_block, p->zbin, p->round_fp,
624                            p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
625                            pd->dequant, eob,
626                            scan_order->scan, scan_order->iscan);
627            break;
628          case TX_8X8:
629            vp9_hadamard_8x8(src_diff, diff_stride, (int16_t *)coeff);
630            vp9_quantize_fp(coeff, 64, x->skip_block, p->zbin, p->round_fp,
631                            p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
632                            pd->dequant, eob,
633                            scan_order->scan, scan_order->iscan);
634            break;
635          case TX_4X4:
636            x->fwd_txm4x4(src_diff, coeff, diff_stride);
637            vp9_quantize_fp(coeff, 16, x->skip_block, p->zbin, p->round_fp,
638                            p->quant_fp, p->quant_shift, qcoeff, dqcoeff,
639                            pd->dequant, eob,
640                            scan_order->scan, scan_order->iscan);
641            break;
642          default:
643            assert(0);
644            break;
645        }
646        *skippable &= (*eob == 0);
647        eob_cost += 1;
648      }
649      block += step;
650    }
651  }
652
653  if (*skippable && *sse < INT64_MAX) {
654    *rate = 0;
655    *dist = (*sse << 6) >> shift;
656    *sse = *dist;
657    return;
658  }
659
660  block = 0;
661  *rate = 0;
662  *dist = 0;
663  if (*sse < INT64_MAX)
664    *sse = (*sse << 6) >> shift;
665  for (r = 0; r < max_blocks_high; r += block_step) {
666    for (c = 0; c < num_4x4_w; c += block_step) {
667      if (c < max_blocks_wide) {
668        tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
669        tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
670        tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
671        uint16_t *const eob = &p->eobs[block];
672
673        if (*eob == 1)
674          *rate += (int)abs(qcoeff[0]);
675        else if (*eob > 1)
676          *rate += (int)vp9_satd((const int16_t *)qcoeff, step << 4);
677
678        *dist += vp9_block_error_fp(coeff, dqcoeff, step << 4) >> shift;
679      }
680      block += step;
681    }
682  }
683
684  if (*skippable == 0) {
685    *rate <<= 10;
686    *rate += (eob_cost << 8);
687  }
688}
689#endif
690
691static void model_rd_for_sb_uv(VP9_COMP *cpi, BLOCK_SIZE plane_bsize,
692                               MACROBLOCK *x, MACROBLOCKD *xd,
693                               int *out_rate_sum, int64_t *out_dist_sum,
694                               unsigned int *var_y, unsigned int *sse_y,
695                               int start_plane, int stop_plane) {
696  // Note our transform coeffs are 8 times an orthogonal transform.
697  // Hence quantizer step is also 8 times. To get effective quantizer
698  // we need to divide by 8 before sending to modeling function.
699  unsigned int sse;
700  int rate;
701  int64_t dist;
702  int i;
703
704  *out_rate_sum = 0;
705  *out_dist_sum = 0;
706
707  for (i = start_plane; i <= stop_plane; ++i) {
708    struct macroblock_plane *const p = &x->plane[i];
709    struct macroblockd_plane *const pd = &xd->plane[i];
710    const uint32_t dc_quant = pd->dequant[0];
711    const uint32_t ac_quant = pd->dequant[1];
712    const BLOCK_SIZE bs = plane_bsize;
713    unsigned int var;
714
715    if (!x->color_sensitivity[i - 1])
716      continue;
717
718    var = cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride,
719                             pd->dst.buf, pd->dst.stride, &sse);
720    *var_y += var;
721    *sse_y += sse;
722
723  #if CONFIG_VP9_HIGHBITDEPTH
724    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
725      vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
726                                   dc_quant >> (xd->bd - 5), &rate, &dist);
727    } else {
728      vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
729                                   dc_quant >> 3, &rate, &dist);
730    }
731  #else
732    vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
733                                 dc_quant >> 3, &rate, &dist);
734  #endif  // CONFIG_VP9_HIGHBITDEPTH
735
736    *out_rate_sum += rate >> 1;
737    *out_dist_sum += dist << 3;
738
739  #if CONFIG_VP9_HIGHBITDEPTH
740    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
741      vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
742                                   ac_quant >> (xd->bd - 5), &rate, &dist);
743    } else {
744      vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
745                                   ac_quant >> 3, &rate, &dist);
746    }
747  #else
748    vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
749                                 ac_quant >> 3, &rate, &dist);
750  #endif  // CONFIG_VP9_HIGHBITDEPTH
751
752    *out_rate_sum += rate;
753    *out_dist_sum += dist << 4;
754  }
755}
756
757static int get_pred_buffer(PRED_BUFFER *p, int len) {
758  int i;
759
760  for (i = 0; i < len; i++) {
761    if (!p[i].in_use) {
762      p[i].in_use = 1;
763      return i;
764    }
765  }
766  return -1;
767}
768
769static void free_pred_buffer(PRED_BUFFER *p) {
770  if (p != NULL)
771    p->in_use = 0;
772}
773
774static void encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x,
775                                 BLOCK_SIZE bsize, int mi_row, int mi_col,
776                                 MV_REFERENCE_FRAME ref_frame,
777                                 PREDICTION_MODE this_mode,
778                                 unsigned int var_y, unsigned int sse_y,
779                                 struct buf_2d yv12_mb[][MAX_MB_PLANE],
780                                 int *rate, int64_t *dist) {
781  MACROBLOCKD *xd = &x->e_mbd;
782
783  const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
784  unsigned int var = var_y, sse = sse_y;
785  // Skipping threshold for ac.
786  unsigned int thresh_ac;
787  // Skipping threshold for dc.
788  unsigned int thresh_dc;
789  if (x->encode_breakout > 0) {
790    // Set a maximum for threshold to avoid big PSNR loss in low bit rate
791    // case. Use extreme low threshold for static frames to limit
792    // skipping.
793    const unsigned int max_thresh = 36000;
794    // The encode_breakout input
795    const unsigned int min_thresh =
796        VPXMIN(((unsigned int)x->encode_breakout << 4), max_thresh);
797#if CONFIG_VP9_HIGHBITDEPTH
798    const int shift = (xd->bd << 1) - 16;
799#endif
800
801    // Calculate threshold according to dequant value.
802    thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) >> 3;
803#if CONFIG_VP9_HIGHBITDEPTH
804    if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
805      thresh_ac = ROUND_POWER_OF_TWO(thresh_ac, shift);
806    }
807#endif  // CONFIG_VP9_HIGHBITDEPTH
808    thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
809
810    // Adjust ac threshold according to partition size.
811    thresh_ac >>=
812        8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
813
814    thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
815#if CONFIG_VP9_HIGHBITDEPTH
816    if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
817      thresh_dc = ROUND_POWER_OF_TWO(thresh_dc, shift);
818    }
819#endif  // CONFIG_VP9_HIGHBITDEPTH
820  } else {
821    thresh_ac = 0;
822    thresh_dc = 0;
823  }
824
825  // Y skipping condition checking for ac and dc.
826  if (var <= thresh_ac && (sse - var) <= thresh_dc) {
827    unsigned int sse_u, sse_v;
828    unsigned int var_u, var_v;
829
830    // Skip UV prediction unless breakout is zero (lossless) to save
831    // computation with low impact on the result
832    if (x->encode_breakout == 0) {
833      xd->plane[1].pre[0] = yv12_mb[ref_frame][1];
834      xd->plane[2].pre[0] = yv12_mb[ref_frame][2];
835      vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize);
836    }
837
838    var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf,
839                                    x->plane[1].src.stride,
840                                    xd->plane[1].dst.buf,
841                                    xd->plane[1].dst.stride, &sse_u);
842
843    // U skipping condition checking
844    if (((var_u << 2) <= thresh_ac) && (sse_u - var_u <= thresh_dc)) {
845      var_v = cpi->fn_ptr[uv_size].vf(x->plane[2].src.buf,
846                                      x->plane[2].src.stride,
847                                      xd->plane[2].dst.buf,
848                                      xd->plane[2].dst.stride, &sse_v);
849
850      // V skipping condition checking
851      if (((var_v << 2) <= thresh_ac) && (sse_v - var_v <= thresh_dc)) {
852        x->skip = 1;
853
854        // The cost of skip bit needs to be added.
855        *rate = cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
856                                    [INTER_OFFSET(this_mode)];
857
858        // More on this part of rate
859        // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
860
861        // Scaling factor for SSE from spatial domain to frequency
862        // domain is 16. Adjust distortion accordingly.
863        // TODO(yunqingwang): In this function, only y-plane dist is
864        // calculated.
865        *dist = (sse << 4);  // + ((sse_u + sse_v) << 4);
866
867        // *disable_skip = 1;
868      }
869    }
870  }
871}
872
873struct estimate_block_intra_args {
874  VP9_COMP *cpi;
875  MACROBLOCK *x;
876  PREDICTION_MODE mode;
877  int rate;
878  int64_t dist;
879};
880
881static void estimate_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
882                                 TX_SIZE tx_size, void *arg) {
883  struct estimate_block_intra_args* const args = arg;
884  VP9_COMP *const cpi = args->cpi;
885  MACROBLOCK *const x = args->x;
886  MACROBLOCKD *const xd = &x->e_mbd;
887  struct macroblock_plane *const p = &x->plane[0];
888  struct macroblockd_plane *const pd = &xd->plane[0];
889  const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
890  uint8_t *const src_buf_base = p->src.buf;
891  uint8_t *const dst_buf_base = pd->dst.buf;
892  const int src_stride = p->src.stride;
893  const int dst_stride = pd->dst.stride;
894  int i, j;
895  int rate;
896  int64_t dist;
897
898  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
899
900  p->src.buf = &src_buf_base[4 * (j * src_stride + i)];
901  pd->dst.buf = &dst_buf_base[4 * (j * dst_stride + i)];
902  // Use source buffer as an approximation for the fully reconstructed buffer.
903  vp9_predict_intra_block(xd, b_width_log2_lookup[plane_bsize],
904                          tx_size, args->mode,
905                          x->skip_encode ? p->src.buf : pd->dst.buf,
906                          x->skip_encode ? src_stride : dst_stride,
907                          pd->dst.buf, dst_stride,
908                          i, j, plane);
909
910  if (plane == 0) {
911    int64_t this_sse = INT64_MAX;
912    int is_skippable;
913    // TODO(jingning): This needs further refactoring.
914    block_yrd(cpi, x, &rate, &dist, &is_skippable, &this_sse, 0,
915              bsize_tx, VPXMIN(tx_size, TX_16X16));
916    x->skip_txfm[0] = is_skippable;
917    // TODO(jingning): Skip is signalled per prediciton block not per tx block.
918    rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), is_skippable);
919  } else {
920    unsigned int var, sse;
921    model_rd_for_sb_uv(cpi, plane_bsize, x, xd, &rate, &dist, &var, &sse,
922                       plane, plane);
923  }
924
925  p->src.buf = src_buf_base;
926  pd->dst.buf = dst_buf_base;
927  args->rate += rate;
928  args->dist += dist;
929}
930
931static const THR_MODES mode_idx[MAX_REF_FRAMES - 1][4] = {
932  {THR_DC, THR_V_PRED, THR_H_PRED, THR_TM},
933  {THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV},
934  {THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG},
935};
936
937static const PREDICTION_MODE intra_mode_list[] = {
938  DC_PRED, V_PRED, H_PRED, TM_PRED
939};
940
941static int mode_offset(const PREDICTION_MODE mode) {
942  if (mode >= NEARESTMV) {
943    return INTER_OFFSET(mode);
944  } else {
945    switch (mode) {
946      case DC_PRED:
947        return 0;
948      case V_PRED:
949        return 1;
950      case H_PRED:
951        return 2;
952      case TM_PRED:
953        return 3;
954      default:
955        return -1;
956    }
957  }
958}
959
960static INLINE void update_thresh_freq_fact(VP9_COMP *cpi,
961                                           TileDataEnc *tile_data,
962                                           BLOCK_SIZE bsize,
963                                           MV_REFERENCE_FRAME ref_frame,
964                                           THR_MODES best_mode_idx,
965                                           PREDICTION_MODE mode) {
966  THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
967  int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx];
968  if (thr_mode_idx == best_mode_idx)
969    *freq_fact -= (*freq_fact >> 4);
970  else
971    *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
972                        cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
973}
974
975void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
976                         BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
977  MACROBLOCKD *const xd = &x->e_mbd;
978  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
979  RD_COST this_rdc, best_rdc;
980  PREDICTION_MODE this_mode;
981  struct estimate_block_intra_args args = { cpi, x, DC_PRED, 0, 0 };
982  const TX_SIZE intra_tx_size =
983      VPXMIN(max_txsize_lookup[bsize],
984             tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
985  MODE_INFO *const mic = xd->mi[0];
986  int *bmode_costs;
987  const MODE_INFO *above_mi = xd->mi[-xd->mi_stride];
988  const MODE_INFO *left_mi = xd->left_available ? xd->mi[-1] : NULL;
989  const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
990  const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
991  bmode_costs = cpi->y_mode_costs[A][L];
992
993  (void) ctx;
994  vp9_rd_cost_reset(&best_rdc);
995  vp9_rd_cost_reset(&this_rdc);
996
997  mbmi->ref_frame[0] = INTRA_FRAME;
998  mbmi->mv[0].as_int = INVALID_MV;
999  mbmi->uv_mode = DC_PRED;
1000  memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
1001
1002  // Change the limit of this loop to add other intra prediction
1003  // mode tests.
1004  for (this_mode = DC_PRED; this_mode <= H_PRED; ++this_mode) {
1005    args.mode = this_mode;
1006    args.rate = 0;
1007    args.dist = 0;
1008    mbmi->tx_size = intra_tx_size;
1009    vp9_foreach_transformed_block_in_plane(xd, bsize, 0,
1010                                           estimate_block_intra, &args);
1011    this_rdc.rate = args.rate;
1012    this_rdc.dist = args.dist;
1013    this_rdc.rate += bmode_costs[this_mode];
1014    this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
1015                             this_rdc.rate, this_rdc.dist);
1016
1017    if (this_rdc.rdcost < best_rdc.rdcost) {
1018      best_rdc = this_rdc;
1019      mbmi->mode = this_mode;
1020    }
1021  }
1022
1023  *rd_cost = best_rdc;
1024}
1025
1026static void init_ref_frame_cost(VP9_COMMON *const cm,
1027                                MACROBLOCKD *const xd,
1028                                int ref_frame_cost[MAX_REF_FRAMES]) {
1029  vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
1030  vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
1031  vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
1032
1033  ref_frame_cost[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
1034  ref_frame_cost[LAST_FRAME] = ref_frame_cost[GOLDEN_FRAME] =
1035    ref_frame_cost[ALTREF_FRAME] = vp9_cost_bit(intra_inter_p, 1);
1036
1037  ref_frame_cost[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
1038  ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1039  ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1040  ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
1041  ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
1042}
1043
1044typedef struct {
1045  MV_REFERENCE_FRAME ref_frame;
1046  PREDICTION_MODE pred_mode;
1047} REF_MODE;
1048
1049#define RT_INTER_MODES 8
1050static const REF_MODE ref_mode_set[RT_INTER_MODES] = {
1051    {LAST_FRAME, ZEROMV},
1052    {LAST_FRAME, NEARESTMV},
1053    {GOLDEN_FRAME, ZEROMV},
1054    {LAST_FRAME, NEARMV},
1055    {LAST_FRAME, NEWMV},
1056    {GOLDEN_FRAME, NEARESTMV},
1057    {GOLDEN_FRAME, NEARMV},
1058    {GOLDEN_FRAME, NEWMV}
1059};
1060static const REF_MODE ref_mode_set_svc[RT_INTER_MODES] = {
1061    {LAST_FRAME, ZEROMV},
1062    {GOLDEN_FRAME, ZEROMV},
1063    {LAST_FRAME, NEARESTMV},
1064    {LAST_FRAME, NEARMV},
1065    {GOLDEN_FRAME, NEARESTMV},
1066    {GOLDEN_FRAME, NEARMV},
1067    {LAST_FRAME, NEWMV},
1068    {GOLDEN_FRAME, NEWMV}
1069};
1070
1071// TODO(jingning) placeholder for inter-frame non-RD mode decision.
1072// this needs various further optimizations. to be continued..
1073void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x,
1074                         TileDataEnc *tile_data,
1075                         int mi_row, int mi_col, RD_COST *rd_cost,
1076                         BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1077  VP9_COMMON *const cm = &cpi->common;
1078  SPEED_FEATURES *const sf = &cpi->sf;
1079  TileInfo *const tile_info = &tile_data->tile_info;
1080  MACROBLOCKD *const xd = &x->e_mbd;
1081  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1082  struct macroblockd_plane *const pd = &xd->plane[0];
1083  PREDICTION_MODE best_mode = ZEROMV;
1084  MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME;
1085  MV_REFERENCE_FRAME usable_ref_frame;
1086  TX_SIZE best_tx_size = TX_SIZES;
1087  INTERP_FILTER best_pred_filter = EIGHTTAP;
1088  int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1089  struct buf_2d yv12_mb[4][MAX_MB_PLANE];
1090  static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
1091                                    VP9_ALT_FLAG };
1092  RD_COST this_rdc, best_rdc;
1093  uint8_t skip_txfm = SKIP_TXFM_NONE, best_mode_skip_txfm = SKIP_TXFM_NONE;
1094  // var_y and sse_y are saved to be used in skipping checking
1095  unsigned int var_y = UINT_MAX;
1096  unsigned int sse_y = UINT_MAX;
1097  // Reduce the intra cost penalty for small blocks (<=16x16).
1098  const int reduction_fac = (bsize <= BLOCK_16X16) ?
1099      ((bsize <= BLOCK_8X8) ? 4 : 2) : 0;
1100  const int intra_cost_penalty = vp9_get_intra_cost_penalty(
1101      cm->base_qindex, cm->y_dc_delta_q, cm->bit_depth) >> reduction_fac;
1102  const int64_t inter_mode_thresh = RDCOST(x->rdmult, x->rddiv,
1103                                           intra_cost_penalty, 0);
1104  const int *const rd_threshes = cpi->rd.threshes[mbmi->segment_id][bsize];
1105  const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
1106  INTERP_FILTER filter_ref;
1107  const int bsl = mi_width_log2_lookup[bsize];
1108  const int pred_filter_search = cm->interp_filter == SWITCHABLE ?
1109      (((mi_row + mi_col) >> bsl) +
1110       get_chessboard_index(cm->current_video_frame)) & 0x1 : 0;
1111  int const_motion[MAX_REF_FRAMES] = { 0 };
1112  const int bh = num_4x4_blocks_high_lookup[bsize] << 2;
1113  const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
1114  // For speed 6, the result of interp filter is reused later in actual encoding
1115  // process.
1116  // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.
1117  PRED_BUFFER tmp[4];
1118  DECLARE_ALIGNED(16, uint8_t, pred_buf[3 * 64 * 64]);
1119#if CONFIG_VP9_HIGHBITDEPTH
1120  DECLARE_ALIGNED(16, uint16_t, pred_buf_16[3 * 64 * 64]);
1121#endif
1122  struct buf_2d orig_dst = pd->dst;
1123  PRED_BUFFER *best_pred = NULL;
1124  PRED_BUFFER *this_mode_pred = NULL;
1125  const int pixels_in_block = bh * bw;
1126  int reuse_inter_pred = cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready;
1127  int ref_frame_skip_mask = 0;
1128  int idx;
1129  int best_pred_sad = INT_MAX;
1130  int best_early_term = 0;
1131  int ref_frame_cost[MAX_REF_FRAMES];
1132
1133  init_ref_frame_cost(cm, xd, ref_frame_cost);
1134
1135  if (reuse_inter_pred) {
1136    int i;
1137    for (i = 0; i < 3; i++) {
1138#if CONFIG_VP9_HIGHBITDEPTH
1139      if (cm->use_highbitdepth)
1140        tmp[i].data = CONVERT_TO_BYTEPTR(&pred_buf_16[pixels_in_block * i]);
1141      else
1142        tmp[i].data = &pred_buf[pixels_in_block * i];
1143#else
1144      tmp[i].data = &pred_buf[pixels_in_block * i];
1145#endif  // CONFIG_VP9_HIGHBITDEPTH
1146      tmp[i].stride = bw;
1147      tmp[i].in_use = 0;
1148    }
1149    tmp[3].data = pd->dst.buf;
1150    tmp[3].stride = pd->dst.stride;
1151    tmp[3].in_use = 0;
1152  }
1153
1154  x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
1155  x->skip = 0;
1156
1157  if (xd->up_available)
1158    filter_ref = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
1159  else if (xd->left_available)
1160    filter_ref = xd->mi[-1]->mbmi.interp_filter;
1161  else
1162    filter_ref = cm->interp_filter;
1163
1164  // initialize mode decisions
1165  vp9_rd_cost_reset(&best_rdc);
1166  vp9_rd_cost_reset(rd_cost);
1167  mbmi->sb_type = bsize;
1168  mbmi->ref_frame[0] = NONE;
1169  mbmi->ref_frame[1] = NONE;
1170  mbmi->tx_size = VPXMIN(max_txsize_lookup[bsize],
1171                         tx_mode_to_biggest_tx_size[cm->tx_mode]);
1172
1173#if CONFIG_VP9_TEMPORAL_DENOISING
1174  vp9_denoiser_reset_frame_stats(ctx);
1175#endif
1176
1177  if (cpi->rc.frames_since_golden == 0 && !cpi->use_svc) {
1178    usable_ref_frame = LAST_FRAME;
1179  } else {
1180    usable_ref_frame = GOLDEN_FRAME;
1181  }
1182  for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) {
1183    const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
1184
1185    x->pred_mv_sad[ref_frame] = INT_MAX;
1186    frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
1187    frame_mv[ZEROMV][ref_frame].as_int = 0;
1188
1189    if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
1190      int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
1191      const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
1192
1193      vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col,
1194                           sf, sf);
1195
1196      if (cm->use_prev_frame_mvs)
1197        vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame,
1198                         candidates, mi_row, mi_col, NULL, NULL,
1199                         x->mbmi_ext->mode_context);
1200      else
1201        const_motion[ref_frame] = mv_refs_rt(cm, x, xd, tile_info,
1202                                             xd->mi[0],
1203                                             ref_frame, candidates,
1204                                             mi_row, mi_col);
1205
1206      vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
1207                            &frame_mv[NEARESTMV][ref_frame],
1208                            &frame_mv[NEARMV][ref_frame]);
1209
1210      if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8)
1211        vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride,
1212                    ref_frame, bsize);
1213    } else {
1214      ref_frame_skip_mask |= (1 << ref_frame);
1215    }
1216  }
1217
1218  for (idx = 0; idx < RT_INTER_MODES; ++idx) {
1219    int rate_mv = 0;
1220    int mode_rd_thresh;
1221    int mode_index;
1222    int i;
1223    int64_t this_sse;
1224    int is_skippable;
1225    int this_early_term = 0;
1226    PREDICTION_MODE this_mode = ref_mode_set[idx].pred_mode;
1227    if (cpi->use_svc)
1228      this_mode = ref_mode_set_svc[idx].pred_mode;
1229
1230    if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode)))
1231      continue;
1232
1233    ref_frame = ref_mode_set[idx].ref_frame;
1234    if (cpi->use_svc)
1235      ref_frame = ref_mode_set_svc[idx].ref_frame;
1236    if (!(cpi->ref_frame_flags & flag_list[ref_frame]))
1237      continue;
1238    if (const_motion[ref_frame] && this_mode == NEARMV)
1239      continue;
1240
1241    if (!(this_mode == ZEROMV && ref_frame == LAST_FRAME)) {
1242      i = (ref_frame == LAST_FRAME) ? GOLDEN_FRAME : LAST_FRAME;
1243      if ((cpi->ref_frame_flags & flag_list[i]) && sf->reference_masking)
1244        if (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[i] << 1))
1245          ref_frame_skip_mask |= (1 << ref_frame);
1246    }
1247    if (ref_frame_skip_mask & (1 << ref_frame))
1248      continue;
1249
1250    // Select prediction reference frames.
1251    for (i = 0; i < MAX_MB_PLANE; i++)
1252      xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
1253
1254    mbmi->ref_frame[0] = ref_frame;
1255    set_ref_ptrs(cm, xd, ref_frame, NONE);
1256
1257    mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
1258    mode_rd_thresh = best_mode_skip_txfm ?
1259            rd_threshes[mode_index] << 1 : rd_threshes[mode_index];
1260    if (rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
1261                            rd_thresh_freq_fact[mode_index]))
1262      continue;
1263
1264    if (this_mode == NEWMV) {
1265      if (ref_frame > LAST_FRAME && !cpi->use_svc) {
1266        int tmp_sad;
1267        int dis, cost_list[5];
1268
1269        if (bsize < BLOCK_16X16)
1270          continue;
1271
1272        tmp_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
1273
1274        if (tmp_sad > x->pred_mv_sad[LAST_FRAME])
1275          continue;
1276        if (tmp_sad + (num_pels_log2_lookup[bsize] << 4) > best_pred_sad)
1277          continue;
1278
1279        frame_mv[NEWMV][ref_frame].as_int = mbmi->mv[0].as_int;
1280        rate_mv = vp9_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv,
1281          &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1282          x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1283        frame_mv[NEWMV][ref_frame].as_mv.row >>= 3;
1284        frame_mv[NEWMV][ref_frame].as_mv.col >>= 3;
1285
1286        cpi->find_fractional_mv_step(x, &frame_mv[NEWMV][ref_frame].as_mv,
1287          &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1288          cpi->common.allow_high_precision_mv,
1289          x->errorperbit,
1290          &cpi->fn_ptr[bsize],
1291          cpi->sf.mv.subpel_force_stop,
1292          cpi->sf.mv.subpel_iters_per_step,
1293          cond_cost_list(cpi, cost_list),
1294          x->nmvjointcost, x->mvcost, &dis,
1295          &x->pred_sse[ref_frame], NULL, 0, 0);
1296      } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1297        &frame_mv[NEWMV][ref_frame], &rate_mv, best_rdc.rdcost)) {
1298        continue;
1299      }
1300    }
1301
1302    if (this_mode == NEWMV && ref_frame == LAST_FRAME &&
1303        frame_mv[NEWMV][LAST_FRAME].as_int != INVALID_MV) {
1304      const int pre_stride = xd->plane[0].pre[0].stride;
1305      const uint8_t * const pre_buf = xd->plane[0].pre[0].buf +
1306          (frame_mv[NEWMV][LAST_FRAME].as_mv.row >> 3) * pre_stride +
1307          (frame_mv[NEWMV][LAST_FRAME].as_mv.col >> 3);
1308      best_pred_sad = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf,
1309                                   x->plane[0].src.stride,
1310                                   pre_buf, pre_stride);
1311      x->pred_mv_sad[LAST_FRAME] = best_pred_sad;
1312    }
1313
1314    if (cpi->use_svc) {
1315      if (this_mode == NEWMV && ref_frame == GOLDEN_FRAME &&
1316          frame_mv[NEWMV][GOLDEN_FRAME].as_int != INVALID_MV) {
1317        const int pre_stride = xd->plane[0].pre[0].stride;
1318        const uint8_t * const pre_buf = xd->plane[0].pre[0].buf +
1319            (frame_mv[NEWMV][GOLDEN_FRAME].as_mv.row >> 3) * pre_stride +
1320            (frame_mv[NEWMV][GOLDEN_FRAME].as_mv.col >> 3);
1321        best_pred_sad = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf,
1322                                               x->plane[0].src.stride,
1323                                               pre_buf, pre_stride);
1324        x->pred_mv_sad[GOLDEN_FRAME] = best_pred_sad;
1325      }
1326    }
1327
1328
1329    if (this_mode != NEARESTMV &&
1330        frame_mv[this_mode][ref_frame].as_int ==
1331            frame_mv[NEARESTMV][ref_frame].as_int)
1332      continue;
1333
1334    mbmi->mode = this_mode;
1335    mbmi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
1336
1337    // Search for the best prediction filter type, when the resulting
1338    // motion vector is at sub-pixel accuracy level for luma component, i.e.,
1339    // the last three bits are all zeros.
1340    if (reuse_inter_pred) {
1341      if (!this_mode_pred) {
1342        this_mode_pred = &tmp[3];
1343      } else {
1344        this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
1345        pd->dst.buf = this_mode_pred->data;
1346        pd->dst.stride = bw;
1347      }
1348    }
1349
1350    if ((this_mode == NEWMV || filter_ref == SWITCHABLE) && pred_filter_search
1351        && (ref_frame == LAST_FRAME ||
1352            (ref_frame == GOLDEN_FRAME && cpi->use_svc))
1353        && (((mbmi->mv[0].as_mv.row | mbmi->mv[0].as_mv.col) & 0x07) != 0)) {
1354      int pf_rate[3];
1355      int64_t pf_dist[3];
1356      unsigned int pf_var[3];
1357      unsigned int pf_sse[3];
1358      TX_SIZE pf_tx_size[3];
1359      int64_t best_cost = INT64_MAX;
1360      INTERP_FILTER best_filter = SWITCHABLE, filter;
1361      PRED_BUFFER *current_pred = this_mode_pred;
1362
1363      for (filter = EIGHTTAP; filter <= EIGHTTAP_SMOOTH; ++filter) {
1364        int64_t cost;
1365        mbmi->interp_filter = filter;
1366        vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1367        model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter], &pf_dist[filter],
1368                          &pf_var[filter], &pf_sse[filter]);
1369        pf_rate[filter] += vp9_get_switchable_rate(cpi, xd);
1370        cost = RDCOST(x->rdmult, x->rddiv, pf_rate[filter], pf_dist[filter]);
1371        pf_tx_size[filter] = mbmi->tx_size;
1372        if (cost < best_cost) {
1373          best_filter = filter;
1374          best_cost = cost;
1375          skip_txfm = x->skip_txfm[0];
1376
1377          if (reuse_inter_pred) {
1378            if (this_mode_pred != current_pred) {
1379              free_pred_buffer(this_mode_pred);
1380              this_mode_pred = current_pred;
1381            }
1382
1383            if (filter < EIGHTTAP_SHARP) {
1384              current_pred = &tmp[get_pred_buffer(tmp, 3)];
1385              pd->dst.buf = current_pred->data;
1386              pd->dst.stride = bw;
1387            }
1388          }
1389        }
1390      }
1391
1392      if (reuse_inter_pred && this_mode_pred != current_pred)
1393        free_pred_buffer(current_pred);
1394
1395      mbmi->interp_filter = best_filter;
1396      mbmi->tx_size = pf_tx_size[best_filter];
1397      this_rdc.rate = pf_rate[best_filter];
1398      this_rdc.dist = pf_dist[best_filter];
1399      var_y = pf_var[best_filter];
1400      sse_y = pf_sse[best_filter];
1401      x->skip_txfm[0] = skip_txfm;
1402      if (reuse_inter_pred) {
1403        pd->dst.buf = this_mode_pred->data;
1404        pd->dst.stride = this_mode_pred->stride;
1405      }
1406    } else {
1407      mbmi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP : filter_ref;
1408      vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1409
1410      // For large partition blocks, extra testing is done.
1411      if (bsize > BLOCK_32X32 &&
1412        !cyclic_refresh_segment_id_boosted(xd->mi[0]->mbmi.segment_id) &&
1413        cm->base_qindex) {
1414        model_rd_for_sb_y_large(cpi, bsize, x, xd, &this_rdc.rate,
1415                                &this_rdc.dist, &var_y, &sse_y, mi_row, mi_col,
1416                                &this_early_term);
1417      } else {
1418        model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
1419                          &var_y, &sse_y);
1420      }
1421    }
1422
1423    if (!this_early_term) {
1424      this_sse = (int64_t)sse_y;
1425      block_yrd(cpi, x, &this_rdc.rate, &this_rdc.dist, &is_skippable,
1426                &this_sse, 0, bsize, VPXMIN(mbmi->tx_size, TX_16X16));
1427      x->skip_txfm[0] = is_skippable;
1428      if (is_skippable) {
1429        this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1430      } else {
1431        if (RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist) <
1432            RDCOST(x->rdmult, x->rddiv, 0, this_sse)) {
1433          this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
1434        } else {
1435          this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1436          this_rdc.dist = this_sse;
1437          x->skip_txfm[0] = SKIP_TXFM_AC_DC;
1438        }
1439      }
1440
1441      if (cm->interp_filter == SWITCHABLE) {
1442        if ((mbmi->mv[0].as_mv.row | mbmi->mv[0].as_mv.col) & 0x07)
1443          this_rdc.rate += vp9_get_switchable_rate(cpi, xd);
1444      }
1445    } else {
1446      this_rdc.rate += cm->interp_filter == SWITCHABLE ?
1447          vp9_get_switchable_rate(cpi, xd) : 0;
1448      this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1449    }
1450
1451    if (x->color_sensitivity[0] || x->color_sensitivity[1]) {
1452      int uv_rate = 0;
1453      int64_t uv_dist = 0;
1454      const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, &xd->plane[1]);
1455      if (x->color_sensitivity[0])
1456        vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 1);
1457      if (x->color_sensitivity[1])
1458        vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 2);
1459      model_rd_for_sb_uv(cpi, uv_bsize, x, xd, &uv_rate, &uv_dist,
1460                         &var_y, &sse_y, 1, 2);
1461      this_rdc.rate += uv_rate;
1462      this_rdc.dist += uv_dist;
1463    }
1464
1465    this_rdc.rate += rate_mv;
1466    this_rdc.rate +=
1467        cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]][INTER_OFFSET(
1468            this_mode)];
1469    this_rdc.rate += ref_frame_cost[ref_frame];
1470    this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1471
1472    // Skipping checking: test to see if this block can be reconstructed by
1473    // prediction only.
1474    if (cpi->allow_encode_breakout) {
1475      encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame, this_mode,
1476                           var_y, sse_y, yv12_mb, &this_rdc.rate,
1477                           &this_rdc.dist);
1478      if (x->skip) {
1479        this_rdc.rate += rate_mv;
1480        this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate,
1481                                 this_rdc.dist);
1482      }
1483    }
1484
1485#if CONFIG_VP9_TEMPORAL_DENOISING
1486    if (cpi->oxcf.noise_sensitivity > 0)
1487      vp9_denoiser_update_frame_stats(mbmi, sse_y, this_mode, ctx);
1488#else
1489    (void)ctx;
1490#endif
1491
1492    if (this_rdc.rdcost < best_rdc.rdcost || x->skip) {
1493      best_rdc = this_rdc;
1494      best_mode = this_mode;
1495      best_pred_filter = mbmi->interp_filter;
1496      best_tx_size = mbmi->tx_size;
1497      best_ref_frame = ref_frame;
1498      best_mode_skip_txfm = x->skip_txfm[0];
1499      best_early_term = this_early_term;
1500
1501      if (reuse_inter_pred) {
1502        free_pred_buffer(best_pred);
1503        best_pred = this_mode_pred;
1504      }
1505    } else {
1506      if (reuse_inter_pred)
1507        free_pred_buffer(this_mode_pred);
1508    }
1509
1510    if (x->skip)
1511      break;
1512
1513    // If early termination flag is 1 and at least 2 modes are checked,
1514    // the mode search is terminated.
1515    if (best_early_term && idx > 0) {
1516      x->skip = 1;
1517      break;
1518    }
1519  }
1520
1521  mbmi->mode          = best_mode;
1522  mbmi->interp_filter = best_pred_filter;
1523  mbmi->tx_size       = best_tx_size;
1524  mbmi->ref_frame[0]  = best_ref_frame;
1525  mbmi->mv[0].as_int  = frame_mv[best_mode][best_ref_frame].as_int;
1526  xd->mi[0]->bmi[0].as_mv[0].as_int = mbmi->mv[0].as_int;
1527  x->skip_txfm[0] = best_mode_skip_txfm;
1528
1529  // Perform intra prediction search, if the best SAD is above a certain
1530  // threshold.
1531  if (best_rdc.rdcost == INT64_MAX ||
1532      (!x->skip && best_rdc.rdcost > inter_mode_thresh &&
1533       bsize <= cpi->sf.max_intra_bsize)) {
1534    struct estimate_block_intra_args args = { cpi, x, DC_PRED, 0, 0 };
1535    int i;
1536    TX_SIZE best_intra_tx_size = TX_SIZES;
1537    TX_SIZE intra_tx_size =
1538        VPXMIN(max_txsize_lookup[bsize],
1539               tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
1540    if (cpi->oxcf.content != VP9E_CONTENT_SCREEN && intra_tx_size > TX_16X16)
1541      intra_tx_size = TX_16X16;
1542
1543    if (reuse_inter_pred && best_pred != NULL) {
1544      if (best_pred->data == orig_dst.buf) {
1545        this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
1546#if CONFIG_VP9_HIGHBITDEPTH
1547        if (cm->use_highbitdepth)
1548          vpx_highbd_convolve_copy(best_pred->data, best_pred->stride,
1549                                   this_mode_pred->data, this_mode_pred->stride,
1550                                   NULL, 0, NULL, 0, bw, bh, xd->bd);
1551        else
1552          vpx_convolve_copy(best_pred->data, best_pred->stride,
1553                          this_mode_pred->data, this_mode_pred->stride,
1554                          NULL, 0, NULL, 0, bw, bh);
1555#else
1556        vpx_convolve_copy(best_pred->data, best_pred->stride,
1557                          this_mode_pred->data, this_mode_pred->stride,
1558                          NULL, 0, NULL, 0, bw, bh);
1559#endif  // CONFIG_VP9_HIGHBITDEPTH
1560        best_pred = this_mode_pred;
1561      }
1562    }
1563    pd->dst = orig_dst;
1564
1565    for (i = 0; i < 4; ++i) {
1566      const PREDICTION_MODE this_mode = intra_mode_list[i];
1567      THR_MODES mode_index = mode_idx[INTRA_FRAME][mode_offset(this_mode)];
1568      int mode_rd_thresh = rd_threshes[mode_index];
1569
1570      if (!((1 << this_mode) & cpi->sf.intra_y_mode_bsize_mask[bsize]))
1571        continue;
1572
1573      if (rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
1574                              rd_thresh_freq_fact[mode_index]))
1575        continue;
1576
1577      mbmi->mode = this_mode;
1578      mbmi->ref_frame[0] = INTRA_FRAME;
1579      args.mode = this_mode;
1580      args.rate = 0;
1581      args.dist = 0;
1582      mbmi->tx_size = intra_tx_size;
1583      vp9_foreach_transformed_block_in_plane(xd, bsize, 0,
1584                                             estimate_block_intra, &args);
1585      // Inter and intra RD will mismatch in scale for non-screen content.
1586      if (cpi->oxcf.content == VP9E_CONTENT_SCREEN) {
1587        if (x->color_sensitivity[0])
1588          vp9_foreach_transformed_block_in_plane(xd, bsize, 1,
1589                                                 estimate_block_intra, &args);
1590        if (x->color_sensitivity[1])
1591          vp9_foreach_transformed_block_in_plane(xd, bsize, 2,
1592                                                 estimate_block_intra, &args);
1593      }
1594      this_rdc.rate = args.rate;
1595      this_rdc.dist = args.dist;
1596      this_rdc.rate += cpi->mbmode_cost[this_mode];
1597      this_rdc.rate += ref_frame_cost[INTRA_FRAME];
1598      this_rdc.rate += intra_cost_penalty;
1599      this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
1600                               this_rdc.rate, this_rdc.dist);
1601
1602      if (this_rdc.rdcost < best_rdc.rdcost) {
1603        best_rdc = this_rdc;
1604        best_mode = this_mode;
1605        best_intra_tx_size = mbmi->tx_size;
1606        best_ref_frame = INTRA_FRAME;
1607        mbmi->uv_mode = this_mode;
1608        mbmi->mv[0].as_int = INVALID_MV;
1609        best_mode_skip_txfm = x->skip_txfm[0];
1610      }
1611    }
1612
1613    // Reset mb_mode_info to the best inter mode.
1614    if (best_ref_frame != INTRA_FRAME) {
1615      mbmi->tx_size = best_tx_size;
1616    } else {
1617      mbmi->tx_size = best_intra_tx_size;
1618    }
1619  }
1620
1621  pd->dst = orig_dst;
1622  mbmi->mode = best_mode;
1623  mbmi->ref_frame[0] = best_ref_frame;
1624  x->skip_txfm[0] = best_mode_skip_txfm;
1625
1626  if (reuse_inter_pred && best_pred != NULL) {
1627    if (best_pred->data != orig_dst.buf && is_inter_mode(mbmi->mode)) {
1628#if CONFIG_VP9_HIGHBITDEPTH
1629      if (cm->use_highbitdepth)
1630        vpx_highbd_convolve_copy(best_pred->data, best_pred->stride,
1631                                 pd->dst.buf, pd->dst.stride, NULL, 0,
1632                                 NULL, 0, bw, bh, xd->bd);
1633      else
1634        vpx_convolve_copy(best_pred->data, best_pred->stride,
1635                          pd->dst.buf, pd->dst.stride, NULL, 0,
1636                          NULL, 0, bw, bh);
1637#else
1638      vpx_convolve_copy(best_pred->data, best_pred->stride,
1639                        pd->dst.buf, pd->dst.stride, NULL, 0,
1640                        NULL, 0, bw, bh);
1641#endif  // CONFIG_VP9_HIGHBITDEPTH
1642    }
1643  }
1644
1645  if (cpi->sf.adaptive_rd_thresh) {
1646    THR_MODES best_mode_idx = mode_idx[best_ref_frame][mode_offset(mbmi->mode)];
1647
1648    if (best_ref_frame == INTRA_FRAME) {
1649      // Only consider the modes that are included in the intra_mode_list.
1650      int intra_modes = sizeof(intra_mode_list)/sizeof(PREDICTION_MODE);
1651      int i;
1652
1653      // TODO(yunqingwang): Check intra mode mask and only update freq_fact
1654      // for those valid modes.
1655      for (i = 0; i < intra_modes; i++) {
1656        update_thresh_freq_fact(cpi, tile_data, bsize, INTRA_FRAME,
1657                                best_mode_idx, intra_mode_list[i]);
1658      }
1659    } else {
1660      for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
1661        PREDICTION_MODE this_mode;
1662        if (best_ref_frame != ref_frame) continue;
1663        for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1664          update_thresh_freq_fact(cpi, tile_data, bsize, ref_frame,
1665                                  best_mode_idx, this_mode);
1666        }
1667      }
1668    }
1669  }
1670
1671  *rd_cost = best_rdc;
1672}
1673
1674void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x,
1675                                int mi_row, int mi_col, RD_COST *rd_cost,
1676                                BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1677  VP9_COMMON *const cm = &cpi->common;
1678  SPEED_FEATURES *const sf = &cpi->sf;
1679  MACROBLOCKD *const xd = &x->e_mbd;
1680  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1681  MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1682  const struct segmentation *const seg = &cm->seg;
1683  MV_REFERENCE_FRAME ref_frame, second_ref_frame = NONE;
1684  MV_REFERENCE_FRAME best_ref_frame = NONE;
1685  unsigned char segment_id = mbmi->segment_id;
1686  struct buf_2d yv12_mb[4][MAX_MB_PLANE];
1687  static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
1688                                    VP9_ALT_FLAG };
1689  int64_t best_rd = INT64_MAX;
1690  b_mode_info bsi[MAX_REF_FRAMES][4];
1691  int ref_frame_skip_mask = 0;
1692  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1693  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1694  int idx, idy;
1695
1696  x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
1697  ctx->pred_pixel_ready = 0;
1698
1699  for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
1700    const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
1701    int_mv dummy_mv[2];
1702    x->pred_mv_sad[ref_frame] = INT_MAX;
1703
1704    if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
1705      int_mv *const candidates = mbmi_ext->ref_mvs[ref_frame];
1706      const struct scale_factors *const sf =
1707                             &cm->frame_refs[ref_frame - 1].sf;
1708      vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col,
1709                           sf, sf);
1710      vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame,
1711                       candidates, mi_row, mi_col, NULL, NULL,
1712                       mbmi_ext->mode_context);
1713
1714      vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
1715                            &dummy_mv[0], &dummy_mv[1]);
1716    } else {
1717      ref_frame_skip_mask |= (1 << ref_frame);
1718    }
1719  }
1720
1721  mbmi->sb_type = bsize;
1722  mbmi->tx_size = TX_4X4;
1723  mbmi->uv_mode = DC_PRED;
1724  mbmi->ref_frame[0] = LAST_FRAME;
1725  mbmi->ref_frame[1] = NONE;
1726  mbmi->interp_filter = cm->interp_filter == SWITCHABLE ? EIGHTTAP
1727                                                        : cm->interp_filter;
1728
1729  for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
1730    int64_t this_rd = 0;
1731    int plane;
1732
1733    if (ref_frame_skip_mask & (1 << ref_frame))
1734      continue;
1735
1736    // TODO(jingning, agrange): Scaling reference frame not supported for
1737    // sub8x8 blocks. Is this supported now?
1738    if (ref_frame > INTRA_FRAME &&
1739        vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
1740      continue;
1741
1742    // If the segment reference frame feature is enabled....
1743    // then do nothing if the current ref frame is not allowed..
1744    if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
1745        get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
1746      continue;
1747
1748    mbmi->ref_frame[0] = ref_frame;
1749    x->skip = 0;
1750    set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
1751
1752    // Select prediction reference frames.
1753    for (plane = 0; plane < MAX_MB_PLANE; plane++)
1754      xd->plane[plane].pre[0] = yv12_mb[ref_frame][plane];
1755
1756    for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1757      for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1758        int_mv b_mv[MB_MODE_COUNT];
1759        int64_t b_best_rd = INT64_MAX;
1760        const int i = idy * 2 + idx;
1761        PREDICTION_MODE this_mode;
1762        RD_COST this_rdc;
1763        unsigned int var_y, sse_y;
1764
1765        struct macroblock_plane *p = &x->plane[0];
1766        struct macroblockd_plane *pd = &xd->plane[0];
1767
1768        const struct buf_2d orig_src = p->src;
1769        const struct buf_2d orig_dst = pd->dst;
1770        struct buf_2d orig_pre[2];
1771        memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre));
1772
1773        // set buffer pointers for sub8x8 motion search.
1774        p->src.buf =
1775            &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1776        pd->dst.buf =
1777            &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
1778        pd->pre[0].buf =
1779            &pd->pre[0].buf[vp9_raster_block_offset(BLOCK_8X8,
1780                                                    i, pd->pre[0].stride)];
1781
1782        b_mv[ZEROMV].as_int = 0;
1783        b_mv[NEWMV].as_int = INVALID_MV;
1784        vp9_append_sub8x8_mvs_for_idx(cm, xd, i, 0, mi_row, mi_col,
1785                                      &b_mv[NEARESTMV],
1786                                      &b_mv[NEARMV],
1787                                      mbmi_ext->mode_context);
1788
1789        for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
1790          int b_rate = 0;
1791          xd->mi[0]->bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int;
1792
1793          if (this_mode == NEWMV) {
1794            const int step_param = cpi->sf.mv.fullpel_search_step_param;
1795            MV mvp_full;
1796            MV tmp_mv;
1797            int cost_list[5];
1798            const int tmp_col_min = x->mv_col_min;
1799            const int tmp_col_max = x->mv_col_max;
1800            const int tmp_row_min = x->mv_row_min;
1801            const int tmp_row_max = x->mv_row_max;
1802            int dummy_dist;
1803
1804            if (i == 0) {
1805              mvp_full.row = b_mv[NEARESTMV].as_mv.row >> 3;
1806              mvp_full.col = b_mv[NEARESTMV].as_mv.col >> 3;
1807            } else {
1808              mvp_full.row = xd->mi[0]->bmi[0].as_mv[0].as_mv.row >> 3;
1809              mvp_full.col = xd->mi[0]->bmi[0].as_mv[0].as_mv.col >> 3;
1810            }
1811
1812            vp9_set_mv_search_range(x, &mbmi_ext->ref_mvs[0]->as_mv);
1813
1814            vp9_full_pixel_search(
1815                cpi, x, bsize, &mvp_full, step_param, x->sadperbit4,
1816                cond_cost_list(cpi, cost_list),
1817                &mbmi_ext->ref_mvs[ref_frame][0].as_mv, &tmp_mv,
1818                INT_MAX, 0);
1819
1820            x->mv_col_min = tmp_col_min;
1821            x->mv_col_max = tmp_col_max;
1822            x->mv_row_min = tmp_row_min;
1823            x->mv_row_max = tmp_row_max;
1824
1825            // calculate the bit cost on motion vector
1826            mvp_full.row = tmp_mv.row * 8;
1827            mvp_full.col = tmp_mv.col * 8;
1828
1829            b_rate += vp9_mv_bit_cost(&mvp_full,
1830                                      &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1831                                      x->nmvjointcost, x->mvcost,
1832                                      MV_COST_WEIGHT);
1833
1834            b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
1835                                          [INTER_OFFSET(NEWMV)];
1836            if (RDCOST(x->rdmult, x->rddiv, b_rate, 0) > b_best_rd)
1837              continue;
1838
1839            cpi->find_fractional_mv_step(x, &tmp_mv,
1840                                         &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1841                                         cpi->common.allow_high_precision_mv,
1842                                         x->errorperbit,
1843                                         &cpi->fn_ptr[bsize],
1844                                         cpi->sf.mv.subpel_force_stop,
1845                                         cpi->sf.mv.subpel_iters_per_step,
1846                                         cond_cost_list(cpi, cost_list),
1847                                         x->nmvjointcost, x->mvcost,
1848                                         &dummy_dist,
1849                                         &x->pred_sse[ref_frame], NULL, 0, 0);
1850
1851            xd->mi[0]->bmi[i].as_mv[0].as_mv = tmp_mv;
1852          } else {
1853            b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
1854                                          [INTER_OFFSET(this_mode)];
1855          }
1856
1857#if CONFIG_VP9_HIGHBITDEPTH
1858          if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1859            vp9_highbd_build_inter_predictor(pd->pre[0].buf, pd->pre[0].stride,
1860                                    pd->dst.buf, pd->dst.stride,
1861                                    &xd->mi[0]->bmi[i].as_mv[0].as_mv,
1862                                    &xd->block_refs[0]->sf,
1863                                    4 * num_4x4_blocks_wide,
1864                                    4 * num_4x4_blocks_high, 0,
1865                                    vp9_filter_kernels[mbmi->interp_filter],
1866                                    MV_PRECISION_Q3,
1867                                    mi_col * MI_SIZE + 4 * (i & 0x01),
1868                                    mi_row * MI_SIZE + 4 * (i >> 1), xd->bd);
1869          } else {
1870#endif
1871            vp9_build_inter_predictor(pd->pre[0].buf, pd->pre[0].stride,
1872                                     pd->dst.buf, pd->dst.stride,
1873                                     &xd->mi[0]->bmi[i].as_mv[0].as_mv,
1874                                     &xd->block_refs[0]->sf,
1875                                     4 * num_4x4_blocks_wide,
1876                                     4 * num_4x4_blocks_high, 0,
1877                                     vp9_filter_kernels[mbmi->interp_filter],
1878                                     MV_PRECISION_Q3,
1879                                     mi_col * MI_SIZE + 4 * (i & 0x01),
1880                                     mi_row * MI_SIZE + 4 * (i >> 1));
1881
1882#if CONFIG_VP9_HIGHBITDEPTH
1883          }
1884#endif
1885
1886          model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
1887                            &var_y, &sse_y);
1888
1889          this_rdc.rate += b_rate;
1890          this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
1891                                   this_rdc.rate, this_rdc.dist);
1892          if (this_rdc.rdcost < b_best_rd) {
1893            b_best_rd = this_rdc.rdcost;
1894            bsi[ref_frame][i].as_mode = this_mode;
1895            bsi[ref_frame][i].as_mv[0].as_mv = xd->mi[0]->bmi[i].as_mv[0].as_mv;
1896          }
1897        }  // mode search
1898
1899        // restore source and prediction buffer pointers.
1900        p->src = orig_src;
1901        pd->pre[0] = orig_pre[0];
1902        pd->dst = orig_dst;
1903        this_rd += b_best_rd;
1904
1905        xd->mi[0]->bmi[i] = bsi[ref_frame][i];
1906        if (num_4x4_blocks_wide > 1)
1907          xd->mi[0]->bmi[i + 1] = xd->mi[0]->bmi[i];
1908        if (num_4x4_blocks_high > 1)
1909          xd->mi[0]->bmi[i + 2] = xd->mi[0]->bmi[i];
1910      }
1911    }  // loop through sub8x8 blocks
1912
1913    if (this_rd < best_rd) {
1914      best_rd = this_rd;
1915      best_ref_frame = ref_frame;
1916    }
1917  }  // reference frames
1918
1919  mbmi->tx_size = TX_4X4;
1920  mbmi->ref_frame[0] = best_ref_frame;
1921  for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1922    for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1923      const int block = idy * 2 + idx;
1924      xd->mi[0]->bmi[block] = bsi[best_ref_frame][block];
1925      if (num_4x4_blocks_wide > 1)
1926        xd->mi[0]->bmi[block + 1] = bsi[best_ref_frame][block];
1927      if (num_4x4_blocks_high > 1)
1928        xd->mi[0]->bmi[block + 2] = bsi[best_ref_frame][block];
1929    }
1930  }
1931  mbmi->mode = xd->mi[0]->bmi[3].as_mode;
1932  ctx->mic = *(xd->mi[0]);
1933  ctx->mbmi_ext = *x->mbmi_ext;
1934  ctx->skip_txfm[0] = SKIP_TXFM_NONE;
1935  ctx->skip = 0;
1936  // Dummy assignment for speed -5. No effect in speed -6.
1937  rd_cost->rdcost = best_rd;
1938}
1939