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