• 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/vpx_codec.h"
20 #include "vpx_dsp/vpx_dsp_common.h"
21 #include "vpx_mem/vpx_mem.h"
22 #include "vpx_ports/mem.h"
23 
24 #include "vp9/common/vp9_blockd.h"
25 #include "vp9/common/vp9_common.h"
26 #include "vp9/common/vp9_mvref_common.h"
27 #include "vp9/common/vp9_pred_common.h"
28 #include "vp9/common/vp9_reconinter.h"
29 #include "vp9/common/vp9_reconintra.h"
30 #include "vp9/common/vp9_scan.h"
31 
32 #include "vp9/encoder/vp9_cost.h"
33 #include "vp9/encoder/vp9_encoder.h"
34 #include "vp9/encoder/vp9_pickmode.h"
35 #include "vp9/encoder/vp9_ratectrl.h"
36 #include "vp9/encoder/vp9_rd.h"
37 
38 typedef struct {
39   uint8_t *data;
40   int stride;
41   int in_use;
42 } PRED_BUFFER;
43 
44 typedef struct {
45   PRED_BUFFER *best_pred;
46   PREDICTION_MODE best_mode;
47   TX_SIZE best_tx_size;
48   TX_SIZE best_intra_tx_size;
49   MV_REFERENCE_FRAME best_ref_frame;
50   MV_REFERENCE_FRAME best_second_ref_frame;
51   uint8_t best_mode_skip_txfm;
52   INTERP_FILTER best_pred_filter;
53 } BEST_PICKMODE;
54 
55 static const int pos_shift_16x16[4][4] = {
56   { 9, 10, 13, 14 }, { 11, 12, 15, 16 }, { 17, 18, 21, 22 }, { 19, 20, 23, 24 }
57 };
58 
mv_refs_rt(VP9_COMP * cpi,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_mv * base_mv,int mi_row,int mi_col,int use_base_mv)59 static int mv_refs_rt(VP9_COMP *cpi, const VP9_COMMON *cm, const MACROBLOCK *x,
60                       const MACROBLOCKD *xd, const TileInfo *const tile,
61                       MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
62                       int_mv *mv_ref_list, int_mv *base_mv, int mi_row,
63                       int mi_col, int use_base_mv) {
64   const int *ref_sign_bias = cm->ref_frame_sign_bias;
65   int i, refmv_count = 0;
66 
67   const POSITION *const mv_ref_search = mv_ref_blocks[mi->sb_type];
68 
69   int different_ref_found = 0;
70   int context_counter = 0;
71   int const_motion = 0;
72 
73   // Blank the reference vector list
74   memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
75 
76   // The nearest 2 blocks are treated differently
77   // if the size < 8x8 we get the mv from the bmi substructure,
78   // and we also need to keep a mode count.
79   for (i = 0; i < 2; ++i) {
80     const POSITION *const mv_ref = &mv_ref_search[i];
81     if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
82       const MODE_INFO *const candidate_mi =
83           xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
84       // Keep counts for entropy encoding.
85       context_counter += mode_2_counter[candidate_mi->mode];
86       different_ref_found = 1;
87 
88       if (candidate_mi->ref_frame[0] == ref_frame)
89         ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1),
90                         refmv_count, mv_ref_list, Done);
91     }
92   }
93 
94   const_motion = 1;
95 
96   // Check the rest of the neighbors in much the same way
97   // as before except we don't need to keep track of sub blocks or
98   // mode counts.
99   for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) {
100     const POSITION *const mv_ref = &mv_ref_search[i];
101     if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
102       const MODE_INFO *const candidate_mi =
103           xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
104       different_ref_found = 1;
105 
106       if (candidate_mi->ref_frame[0] == ref_frame)
107         ADD_MV_REF_LIST(candidate_mi->mv[0], refmv_count, mv_ref_list, Done);
108     }
109   }
110 
111   // Since we couldn't find 2 mvs from the same reference frame
112   // go back through the neighbors and find motion vectors from
113   // different reference frames.
114   if (different_ref_found && !refmv_count) {
115     for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
116       const POSITION *mv_ref = &mv_ref_search[i];
117       if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
118         const MODE_INFO *const candidate_mi =
119             xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
120 
121         // If the candidate is INTRA we don't want to consider its mv.
122         IF_DIFF_REF_FRAME_ADD_MV(candidate_mi, ref_frame, ref_sign_bias,
123                                  refmv_count, mv_ref_list, Done);
124       }
125     }
126   }
127   if (use_base_mv &&
128       !cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
129       ref_frame == LAST_FRAME) {
130     // Get base layer mv.
131     MV_REF *candidate =
132         &cm->prev_frame
133              ->mvs[(mi_col >> 1) + (mi_row >> 1) * (cm->mi_cols >> 1)];
134     if (candidate->mv[0].as_int != INVALID_MV) {
135       base_mv->as_mv.row = (candidate->mv[0].as_mv.row * 2);
136       base_mv->as_mv.col = (candidate->mv[0].as_mv.col * 2);
137       clamp_mv_ref(&base_mv->as_mv, xd);
138     } else {
139       base_mv->as_int = INVALID_MV;
140     }
141   }
142 
143 Done:
144 
145   x->mbmi_ext->mode_context[ref_frame] = counter_to_context[context_counter];
146 
147   // Clamp vectors
148   for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
149     clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
150 
151   return const_motion;
152 }
153 
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,int use_base_mv)154 static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
155                                   BLOCK_SIZE bsize, int mi_row, int mi_col,
156                                   int_mv *tmp_mv, int *rate_mv,
157                                   int64_t best_rd_sofar, int use_base_mv) {
158   MACROBLOCKD *xd = &x->e_mbd;
159   MODE_INFO *mi = xd->mi[0];
160   struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0 } };
161   const int step_param = cpi->sf.mv.fullpel_search_step_param;
162   const int sadpb = x->sadperbit16;
163   MV mvp_full;
164   const int ref = mi->ref_frame[0];
165   const MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
166   MV center_mv;
167   uint32_t dis;
168   int rate_mode;
169   const MvLimits tmp_mv_limits = x->mv_limits;
170   int rv = 0;
171   int cost_list[5];
172   int search_subpel = 1;
173   const YV12_BUFFER_CONFIG *scaled_ref_frame =
174       vp9_get_scaled_ref_frame(cpi, ref);
175   if (scaled_ref_frame) {
176     int i;
177     // Swap out the reference frame for a version that's been scaled to
178     // match the resolution of the current frame, allowing the existing
179     // motion search code to be used without additional modifications.
180     for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
181     vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
182   }
183   vp9_set_mv_search_range(&x->mv_limits, &ref_mv);
184 
185   // Limit motion vector for large lightning change.
186   if (cpi->oxcf.speed > 5 && x->lowvar_highsumdiff) {
187     x->mv_limits.col_min = VPXMAX(x->mv_limits.col_min, -10);
188     x->mv_limits.row_min = VPXMAX(x->mv_limits.row_min, -10);
189     x->mv_limits.col_max = VPXMIN(x->mv_limits.col_max, 10);
190     x->mv_limits.row_max = VPXMIN(x->mv_limits.row_max, 10);
191   }
192 
193   assert(x->mv_best_ref_index[ref] <= 2);
194   if (x->mv_best_ref_index[ref] < 2)
195     mvp_full = x->mbmi_ext->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
196   else
197     mvp_full = x->pred_mv[ref];
198 
199   mvp_full.col >>= 3;
200   mvp_full.row >>= 3;
201 
202   if (!use_base_mv)
203     center_mv = ref_mv;
204   else
205     center_mv = tmp_mv->as_mv;
206 
207   if (x->sb_use_mv_part) {
208     tmp_mv->as_mv.row = x->sb_mvrow_part >> 3;
209     tmp_mv->as_mv.col = x->sb_mvcol_part >> 3;
210   } else {
211     vp9_full_pixel_search(
212         cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method, sadpb,
213         cond_cost_list(cpi, cost_list), &center_mv, &tmp_mv->as_mv, INT_MAX, 0);
214   }
215 
216   x->mv_limits = tmp_mv_limits;
217 
218   // calculate the bit cost on motion vector
219   mvp_full.row = tmp_mv->as_mv.row * 8;
220   mvp_full.col = tmp_mv->as_mv.col * 8;
221 
222   *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv, x->nmvjointcost, x->mvcost,
223                              MV_COST_WEIGHT);
224 
225   rate_mode =
226       cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref]][INTER_OFFSET(NEWMV)];
227   rv =
228       !(RDCOST(x->rdmult, x->rddiv, (*rate_mv + rate_mode), 0) > best_rd_sofar);
229 
230   // For SVC on non-reference frame, avoid subpel for (0, 0) motion.
231   if (cpi->use_svc && cpi->svc.non_reference_frame) {
232     if (mvp_full.row == 0 && mvp_full.col == 0) search_subpel = 0;
233   }
234 
235   if (rv && search_subpel) {
236     SUBPEL_FORCE_STOP subpel_force_stop = cpi->sf.mv.subpel_force_stop;
237     if (use_base_mv && cpi->sf.base_mv_aggressive) subpel_force_stop = HALF_PEL;
238     if (cpi->sf.mv.enable_adaptive_subpel_force_stop) {
239       const int mv_thresh = cpi->sf.mv.adapt_subpel_force_stop.mv_thresh;
240       if (abs(tmp_mv->as_mv.row) >= mv_thresh ||
241           abs(tmp_mv->as_mv.col) >= mv_thresh)
242         subpel_force_stop = cpi->sf.mv.adapt_subpel_force_stop.force_stop_above;
243       else
244         subpel_force_stop = cpi->sf.mv.adapt_subpel_force_stop.force_stop_below;
245     }
246     cpi->find_fractional_mv_step(
247         x, &tmp_mv->as_mv, &ref_mv, cpi->common.allow_high_precision_mv,
248         x->errorperbit, &cpi->fn_ptr[bsize], subpel_force_stop,
249         cpi->sf.mv.subpel_search_level, cond_cost_list(cpi, cost_list),
250         x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, 0, 0,
251         cpi->sf.use_accurate_subpel_search);
252     *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmvjointcost,
253                                x->mvcost, MV_COST_WEIGHT);
254   }
255 
256   if (scaled_ref_frame) {
257     int i;
258     for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
259   }
260   return rv;
261 }
262 
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,int use_highbitdepth,vpx_bit_depth_t bd,uint32_t * sse8x8,int * sum8x8,uint32_t * var8x8)263 static void block_variance(const uint8_t *src, int src_stride,
264                            const uint8_t *ref, int ref_stride, int w, int h,
265                            unsigned int *sse, int *sum, int block_size,
266 #if CONFIG_VP9_HIGHBITDEPTH
267                            int use_highbitdepth, vpx_bit_depth_t bd,
268 #endif
269                            uint32_t *sse8x8, int *sum8x8, uint32_t *var8x8) {
270   int i, j, k = 0;
271 
272   *sse = 0;
273   *sum = 0;
274 
275   for (i = 0; i < h; i += block_size) {
276     for (j = 0; j < w; j += block_size) {
277 #if CONFIG_VP9_HIGHBITDEPTH
278       if (use_highbitdepth) {
279         switch (bd) {
280           case VPX_BITS_8:
281             vpx_highbd_8_get8x8var(src + src_stride * i + j, src_stride,
282                                    ref + ref_stride * i + j, ref_stride,
283                                    &sse8x8[k], &sum8x8[k]);
284             break;
285           case VPX_BITS_10:
286             vpx_highbd_10_get8x8var(src + src_stride * i + j, src_stride,
287                                     ref + ref_stride * i + j, ref_stride,
288                                     &sse8x8[k], &sum8x8[k]);
289             break;
290           case VPX_BITS_12:
291             vpx_highbd_12_get8x8var(src + src_stride * i + j, src_stride,
292                                     ref + ref_stride * i + j, ref_stride,
293                                     &sse8x8[k], &sum8x8[k]);
294             break;
295         }
296       } else {
297         vpx_get8x8var(src + src_stride * i + j, src_stride,
298                       ref + ref_stride * i + j, ref_stride, &sse8x8[k],
299                       &sum8x8[k]);
300       }
301 #else
302       vpx_get8x8var(src + src_stride * i + j, src_stride,
303                     ref + ref_stride * i + j, ref_stride, &sse8x8[k],
304                     &sum8x8[k]);
305 #endif
306       *sse += sse8x8[k];
307       *sum += sum8x8[k];
308       var8x8[k] = sse8x8[k] - (uint32_t)(((int64_t)sum8x8[k] * sum8x8[k]) >> 6);
309       k++;
310     }
311   }
312 }
313 
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)314 static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
315                                unsigned int *sse_i, int *sum_i,
316                                unsigned int *var_o, unsigned int *sse_o,
317                                int *sum_o) {
318   const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
319   const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
320   const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
321   int i, j, k = 0;
322 
323   for (i = 0; i < nh; i += 2) {
324     for (j = 0; j < nw; j += 2) {
325       sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] +
326                  sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1];
327       sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] +
328                  sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1];
329       var_o[k] = sse_o[k] - (uint32_t)(((int64_t)sum_o[k] * sum_o[k]) >>
330                                        (b_width_log2_lookup[unit_size] +
331                                         b_height_log2_lookup[unit_size] + 6));
332       k++;
333     }
334   }
335 }
336 
337 // Adjust the ac_thr according to speed, width, height and normalized sum
ac_thr_factor(const int speed,const int width,const int height,const int norm_sum)338 static int ac_thr_factor(const int speed, const int width, const int height,
339                          const int norm_sum) {
340   if (speed >= 8 && norm_sum < 5) {
341     if (width <= 640 && height <= 480)
342       return 4;
343     else
344       return 2;
345   }
346   return 1;
347 }
348 
calculate_tx_size(VP9_COMP * const cpi,BLOCK_SIZE bsize,MACROBLOCKD * const xd,unsigned int var,unsigned int sse,int64_t ac_thr)349 static TX_SIZE calculate_tx_size(VP9_COMP *const cpi, BLOCK_SIZE bsize,
350                                  MACROBLOCKD *const xd, unsigned int var,
351                                  unsigned int sse, int64_t ac_thr) {
352   TX_SIZE tx_size;
353   if (cpi->common.tx_mode == TX_MODE_SELECT) {
354     if (sse > (var << 2))
355       tx_size = VPXMIN(max_txsize_lookup[bsize],
356                        tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
357     else
358       tx_size = TX_8X8;
359 
360     if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
361         cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
362       tx_size = TX_8X8;
363     else if (tx_size > TX_16X16)
364       tx_size = TX_16X16;
365 
366     // For screen-content force 4X4 tx_size over 8X8, for large variance.
367     if (cpi->oxcf.content == VP9E_CONTENT_SCREEN && tx_size == TX_8X8 &&
368         bsize <= BLOCK_16X16 && var > (ac_thr << 6))
369       tx_size = TX_4X4;
370   } else {
371     tx_size = VPXMIN(max_txsize_lookup[bsize],
372                      tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
373   }
374 
375   return tx_size;
376 }
377 
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,int * flag_preduv_computed)378 static void model_rd_for_sb_y_large(VP9_COMP *cpi, BLOCK_SIZE bsize,
379                                     MACROBLOCK *x, MACROBLOCKD *xd,
380                                     int *out_rate_sum, int64_t *out_dist_sum,
381                                     unsigned int *var_y, unsigned int *sse_y,
382                                     int mi_row, int mi_col, int *early_term,
383                                     int *flag_preduv_computed) {
384   // Note our transform coeffs are 8 times an orthogonal transform.
385   // Hence quantizer step is also 8 times. To get effective quantizer
386   // we need to divide by 8 before sending to modeling function.
387   unsigned int sse;
388   int rate;
389   int64_t dist;
390   struct macroblock_plane *const p = &x->plane[0];
391   struct macroblockd_plane *const pd = &xd->plane[0];
392   const uint32_t dc_quant = pd->dequant[0];
393   const uint32_t ac_quant = pd->dequant[1];
394   int64_t dc_thr = dc_quant * dc_quant >> 6;
395   int64_t ac_thr = ac_quant * ac_quant >> 6;
396   unsigned int var;
397   int sum;
398   int skip_dc = 0;
399 
400   const int bw = b_width_log2_lookup[bsize];
401   const int bh = b_height_log2_lookup[bsize];
402   const int num8x8 = 1 << (bw + bh - 2);
403   unsigned int sse8x8[64] = { 0 };
404   int sum8x8[64] = { 0 };
405   unsigned int var8x8[64] = { 0 };
406   TX_SIZE tx_size;
407   int i, k;
408 #if CONFIG_VP9_HIGHBITDEPTH
409   const vpx_bit_depth_t bd = cpi->common.bit_depth;
410 #endif
411   // Calculate variance for whole partition, and also save 8x8 blocks' variance
412   // to be used in following transform skipping test.
413   block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
414                  4 << bw, 4 << bh, &sse, &sum, 8,
415 #if CONFIG_VP9_HIGHBITDEPTH
416                  cpi->common.use_highbitdepth, bd,
417 #endif
418                  sse8x8, sum8x8, var8x8);
419   var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
420 
421   *var_y = var;
422   *sse_y = sse;
423 
424 #if CONFIG_VP9_TEMPORAL_DENOISING
425   if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
426       cpi->oxcf.speed > 5)
427     ac_thr = vp9_scale_acskip_thresh(ac_thr, cpi->denoiser.denoising_level,
428                                      (abs(sum) >> (bw + bh)),
429                                      cpi->svc.temporal_layer_id);
430   else
431     ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
432                             cpi->common.height, abs(sum) >> (bw + bh));
433 #else
434   ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
435                           cpi->common.height, abs(sum) >> (bw + bh));
436 #endif
437 
438   tx_size = calculate_tx_size(cpi, bsize, xd, var, sse, ac_thr);
439   // The code below for setting skip flag assumes tranform size of at least 8x8,
440   // so force this lower limit on transform.
441   if (tx_size < TX_8X8) tx_size = TX_8X8;
442   xd->mi[0]->tx_size = tx_size;
443 
444   if (cpi->oxcf.content == VP9E_CONTENT_SCREEN && x->zero_temp_sad_source &&
445       x->source_variance == 0)
446     dc_thr = dc_thr << 1;
447 
448   // Evaluate if the partition block is a skippable block in Y plane.
449   {
450     unsigned int sse16x16[16] = { 0 };
451     int sum16x16[16] = { 0 };
452     unsigned int var16x16[16] = { 0 };
453     const int num16x16 = num8x8 >> 2;
454 
455     unsigned int sse32x32[4] = { 0 };
456     int sum32x32[4] = { 0 };
457     unsigned int var32x32[4] = { 0 };
458     const int num32x32 = num8x8 >> 4;
459 
460     int ac_test = 1;
461     int dc_test = 1;
462     const int num = (tx_size == TX_8X8)
463                         ? num8x8
464                         : ((tx_size == TX_16X16) ? num16x16 : num32x32);
465     const unsigned int *sse_tx =
466         (tx_size == TX_8X8) ? sse8x8
467                             : ((tx_size == TX_16X16) ? sse16x16 : sse32x32);
468     const unsigned int *var_tx =
469         (tx_size == TX_8X8) ? var8x8
470                             : ((tx_size == TX_16X16) ? var16x16 : var32x32);
471 
472     // Calculate variance if tx_size > TX_8X8
473     if (tx_size >= TX_16X16)
474       calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
475                          sum16x16);
476     if (tx_size == TX_32X32)
477       calculate_variance(bw, bh, TX_16X16, sse16x16, sum16x16, var32x32,
478                          sse32x32, sum32x32);
479 
480     // Skipping test
481     x->skip_txfm[0] = SKIP_TXFM_NONE;
482     for (k = 0; k < num; k++)
483       // Check if all ac coefficients can be quantized to zero.
484       if (!(var_tx[k] < ac_thr || var == 0)) {
485         ac_test = 0;
486         break;
487       }
488 
489     for (k = 0; k < num; k++)
490       // Check if dc coefficient can be quantized to zero.
491       if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
492         dc_test = 0;
493         break;
494       }
495 
496     if (ac_test) {
497       x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
498 
499       if (dc_test) x->skip_txfm[0] = SKIP_TXFM_AC_DC;
500     } else if (dc_test) {
501       skip_dc = 1;
502     }
503   }
504 
505   if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
506     int skip_uv[2] = { 0 };
507     unsigned int var_uv[2];
508     unsigned int sse_uv[2];
509 
510     *out_rate_sum = 0;
511     *out_dist_sum = sse << 4;
512 
513     // Transform skipping test in UV planes.
514     for (i = 1; i <= 2; i++) {
515       if (cpi->oxcf.speed < 8 || x->color_sensitivity[i - 1]) {
516         struct macroblock_plane *const p = &x->plane[i];
517         struct macroblockd_plane *const pd = &xd->plane[i];
518         const TX_SIZE uv_tx_size = get_uv_tx_size(xd->mi[0], pd);
519         const BLOCK_SIZE unit_size = txsize_to_bsize[uv_tx_size];
520         const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, pd);
521         const int uv_bw = b_width_log2_lookup[uv_bsize];
522         const int uv_bh = b_height_log2_lookup[uv_bsize];
523         const int sf = (uv_bw - b_width_log2_lookup[unit_size]) +
524                        (uv_bh - b_height_log2_lookup[unit_size]);
525         const uint32_t uv_dc_thr = pd->dequant[0] * pd->dequant[0] >> (6 - sf);
526         const uint32_t uv_ac_thr = pd->dequant[1] * pd->dequant[1] >> (6 - sf);
527         int j = i - 1;
528 
529         vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, i);
530         flag_preduv_computed[i - 1] = 1;
531         var_uv[j] = cpi->fn_ptr[uv_bsize].vf(
532             p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, &sse_uv[j]);
533 
534         if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
535             (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
536           skip_uv[j] = 1;
537         else
538           break;
539       } else {
540         skip_uv[i - 1] = 1;
541       }
542     }
543 
544     // If the transform in YUV planes are skippable, the mode search checks
545     // fewer inter modes and doesn't check intra modes.
546     if (skip_uv[0] & skip_uv[1]) {
547       *early_term = 1;
548     }
549     return;
550   }
551 
552   if (!skip_dc) {
553 #if CONFIG_VP9_HIGHBITDEPTH
554     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
555                                  dc_quant >> (xd->bd - 5), &rate, &dist);
556 #else
557     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
558                                  dc_quant >> 3, &rate, &dist);
559 #endif  // CONFIG_VP9_HIGHBITDEPTH
560   }
561 
562   if (!skip_dc) {
563     *out_rate_sum = rate >> 1;
564     *out_dist_sum = dist << 3;
565   } else {
566     *out_rate_sum = 0;
567     *out_dist_sum = (sse - var) << 4;
568   }
569 
570 #if CONFIG_VP9_HIGHBITDEPTH
571   vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
572                                ac_quant >> (xd->bd - 5), &rate, &dist);
573 #else
574   vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
575                                &rate, &dist);
576 #endif  // CONFIG_VP9_HIGHBITDEPTH
577 
578   *out_rate_sum += rate;
579   *out_dist_sum += dist << 4;
580 }
581 
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)582 static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
583                               MACROBLOCKD *xd, int *out_rate_sum,
584                               int64_t *out_dist_sum, unsigned int *var_y,
585                               unsigned int *sse_y) {
586   // Note our transform coeffs are 8 times an orthogonal transform.
587   // Hence quantizer step is also 8 times. To get effective quantizer
588   // we need to divide by 8 before sending to modeling function.
589   unsigned int sse;
590   int rate;
591   int64_t dist;
592   struct macroblock_plane *const p = &x->plane[0];
593   struct macroblockd_plane *const pd = &xd->plane[0];
594   const int64_t dc_thr = p->quant_thred[0] >> 6;
595   const int64_t ac_thr = p->quant_thred[1] >> 6;
596   const uint32_t dc_quant = pd->dequant[0];
597   const uint32_t ac_quant = pd->dequant[1];
598   unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
599                                            pd->dst.buf, pd->dst.stride, &sse);
600   int skip_dc = 0;
601 
602   *var_y = var;
603   *sse_y = sse;
604 
605   xd->mi[0]->tx_size = calculate_tx_size(cpi, bsize, xd, var, sse, ac_thr);
606 
607   // Evaluate if the partition block is a skippable block in Y plane.
608   {
609     const BLOCK_SIZE unit_size = txsize_to_bsize[xd->mi[0]->tx_size];
610     const unsigned int num_blk_log2 =
611         (b_width_log2_lookup[bsize] - b_width_log2_lookup[unit_size]) +
612         (b_height_log2_lookup[bsize] - b_height_log2_lookup[unit_size]);
613     const unsigned int sse_tx = sse >> num_blk_log2;
614     const unsigned int var_tx = var >> num_blk_log2;
615 
616     x->skip_txfm[0] = SKIP_TXFM_NONE;
617     // Check if all ac coefficients can be quantized to zero.
618     if (var_tx < ac_thr || var == 0) {
619       x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
620       // Check if dc coefficient can be quantized to zero.
621       if (sse_tx - var_tx < dc_thr || sse == var)
622         x->skip_txfm[0] = SKIP_TXFM_AC_DC;
623     } else {
624       if (sse_tx - var_tx < dc_thr || sse == var) skip_dc = 1;
625     }
626   }
627 
628   if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
629     *out_rate_sum = 0;
630     *out_dist_sum = sse << 4;
631     return;
632   }
633 
634   if (!skip_dc) {
635 #if CONFIG_VP9_HIGHBITDEPTH
636     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
637                                  dc_quant >> (xd->bd - 5), &rate, &dist);
638 #else
639     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
640                                  dc_quant >> 3, &rate, &dist);
641 #endif  // CONFIG_VP9_HIGHBITDEPTH
642   }
643 
644   if (!skip_dc) {
645     *out_rate_sum = rate >> 1;
646     *out_dist_sum = dist << 3;
647   } else {
648     *out_rate_sum = 0;
649     *out_dist_sum = (sse - var) << 4;
650   }
651 
652 #if CONFIG_VP9_HIGHBITDEPTH
653   vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
654                                ac_quant >> (xd->bd - 5), &rate, &dist);
655 #else
656   vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
657                                &rate, &dist);
658 #endif  // CONFIG_VP9_HIGHBITDEPTH
659 
660   *out_rate_sum += rate;
661   *out_dist_sum += dist << 4;
662 }
663 
block_yrd(VP9_COMP * cpi,MACROBLOCK * x,RD_COST * this_rdc,int * skippable,int64_t * sse,BLOCK_SIZE bsize,TX_SIZE tx_size,int rd_computed)664 static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc,
665                       int *skippable, int64_t *sse, BLOCK_SIZE bsize,
666                       TX_SIZE tx_size, int rd_computed) {
667   MACROBLOCKD *xd = &x->e_mbd;
668   const struct macroblockd_plane *pd = &xd->plane[0];
669   struct macroblock_plane *const p = &x->plane[0];
670   const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
671   const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
672   const int step = 1 << (tx_size << 1);
673   const int block_step = (1 << tx_size);
674   int block = 0, r, c;
675   const int max_blocks_wide =
676       num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 : xd->mb_to_right_edge >> 5);
677   const int max_blocks_high =
678       num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 : xd->mb_to_bottom_edge >> 5);
679   int eob_cost = 0;
680   const int bw = 4 * num_4x4_w;
681   const int bh = 4 * num_4x4_h;
682 
683 #if CONFIG_VP9_HIGHBITDEPTH
684   // TODO(jingning): Implement the high bit-depth Hadamard transforms and
685   // remove this check condition.
686   // TODO(marpan): Use this path (model_rd) for 8bit under certain conditions
687   // for now, as the vp9_quantize_fp below for highbitdepth build is slow.
688   if (xd->bd != 8 ||
689       (cpi->oxcf.speed > 5 && cpi->common.frame_type != KEY_FRAME &&
690        bsize < BLOCK_32X32)) {
691     unsigned int var_y, sse_y;
692     (void)tx_size;
693     if (!rd_computed)
694       model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist,
695                         &var_y, &sse_y);
696     *sse = INT_MAX;
697     *skippable = 0;
698     return;
699   }
700 #endif
701 
702   if (cpi->sf.use_simple_block_yrd && cpi->common.frame_type != KEY_FRAME &&
703       (bsize < BLOCK_32X32 ||
704        (cpi->use_svc &&
705         (bsize < BLOCK_32X32 || cpi->svc.temporal_layer_id > 0)))) {
706     unsigned int var_y, sse_y;
707     (void)tx_size;
708     if (!rd_computed)
709       model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist,
710                         &var_y, &sse_y);
711     *sse = INT_MAX;
712     *skippable = 0;
713     return;
714   }
715 
716   (void)cpi;
717 
718   // The max tx_size passed in is TX_16X16.
719   assert(tx_size != TX_32X32);
720 
721   vpx_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
722                      pd->dst.buf, pd->dst.stride);
723   *skippable = 1;
724   // Keep track of the row and column of the blocks we use so that we know
725   // if we are in the unrestricted motion border.
726   for (r = 0; r < max_blocks_high; r += block_step) {
727     for (c = 0; c < num_4x4_w; c += block_step) {
728       if (c < max_blocks_wide) {
729         const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
730         tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
731         tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
732         tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
733         uint16_t *const eob = &p->eobs[block];
734         const int diff_stride = bw;
735         const int16_t *src_diff;
736         src_diff = &p->src_diff[(r * diff_stride + c) << 2];
737 
738         switch (tx_size) {
739           case TX_16X16:
740             vpx_hadamard_16x16(src_diff, diff_stride, coeff);
741             vp9_quantize_fp(coeff, 256, x->skip_block, p->round_fp, p->quant_fp,
742                             qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
743                             scan_order->iscan);
744             break;
745           case TX_8X8:
746             vpx_hadamard_8x8(src_diff, diff_stride, coeff);
747             vp9_quantize_fp(coeff, 64, x->skip_block, p->round_fp, p->quant_fp,
748                             qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
749                             scan_order->iscan);
750             break;
751           default:
752             assert(tx_size == TX_4X4);
753             x->fwd_txfm4x4(src_diff, coeff, diff_stride);
754             vp9_quantize_fp(coeff, 16, x->skip_block, p->round_fp, p->quant_fp,
755                             qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
756                             scan_order->iscan);
757             break;
758         }
759         *skippable &= (*eob == 0);
760         eob_cost += 1;
761       }
762       block += step;
763     }
764   }
765 
766   this_rdc->rate = 0;
767   if (*sse < INT64_MAX) {
768     *sse = (*sse << 6) >> 2;
769     if (*skippable) {
770       this_rdc->dist = *sse;
771       return;
772     }
773   }
774 
775   block = 0;
776   this_rdc->dist = 0;
777   for (r = 0; r < max_blocks_high; r += block_step) {
778     for (c = 0; c < num_4x4_w; c += block_step) {
779       if (c < max_blocks_wide) {
780         tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
781         tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
782         tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
783         uint16_t *const eob = &p->eobs[block];
784 
785         if (*eob == 1)
786           this_rdc->rate += (int)abs(qcoeff[0]);
787         else if (*eob > 1)
788           this_rdc->rate += vpx_satd(qcoeff, step << 4);
789 
790         this_rdc->dist += vp9_block_error_fp(coeff, dqcoeff, step << 4) >> 2;
791       }
792       block += step;
793     }
794   }
795 
796   // If skippable is set, rate gets clobbered later.
797   this_rdc->rate <<= (2 + VP9_PROB_COST_SHIFT);
798   this_rdc->rate += (eob_cost << VP9_PROB_COST_SHIFT);
799 }
800 
model_rd_for_sb_uv(VP9_COMP * cpi,BLOCK_SIZE plane_bsize,MACROBLOCK * x,MACROBLOCKD * xd,RD_COST * this_rdc,unsigned int * var_y,unsigned int * sse_y,int start_plane,int stop_plane)801 static void model_rd_for_sb_uv(VP9_COMP *cpi, BLOCK_SIZE plane_bsize,
802                                MACROBLOCK *x, MACROBLOCKD *xd,
803                                RD_COST *this_rdc, unsigned int *var_y,
804                                unsigned int *sse_y, int start_plane,
805                                int stop_plane) {
806   // Note our transform coeffs are 8 times an orthogonal transform.
807   // Hence quantizer step is also 8 times. To get effective quantizer
808   // we need to divide by 8 before sending to modeling function.
809   unsigned int sse;
810   int rate;
811   int64_t dist;
812   int i;
813 #if CONFIG_VP9_HIGHBITDEPTH
814   uint64_t tot_var = *var_y;
815   uint64_t tot_sse = *sse_y;
816 #else
817   uint32_t tot_var = *var_y;
818   uint32_t tot_sse = *sse_y;
819 #endif
820 
821   this_rdc->rate = 0;
822   this_rdc->dist = 0;
823 
824   for (i = start_plane; i <= stop_plane; ++i) {
825     struct macroblock_plane *const p = &x->plane[i];
826     struct macroblockd_plane *const pd = &xd->plane[i];
827     const uint32_t dc_quant = pd->dequant[0];
828     const uint32_t ac_quant = pd->dequant[1];
829     const BLOCK_SIZE bs = plane_bsize;
830     unsigned int var;
831     if (!x->color_sensitivity[i - 1]) continue;
832 
833     var = cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf,
834                              pd->dst.stride, &sse);
835     assert(sse >= var);
836     tot_var += var;
837     tot_sse += sse;
838 
839 #if CONFIG_VP9_HIGHBITDEPTH
840     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
841                                  dc_quant >> (xd->bd - 5), &rate, &dist);
842 #else
843     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
844                                  dc_quant >> 3, &rate, &dist);
845 #endif  // CONFIG_VP9_HIGHBITDEPTH
846 
847     this_rdc->rate += rate >> 1;
848     this_rdc->dist += dist << 3;
849 
850 #if CONFIG_VP9_HIGHBITDEPTH
851     vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
852                                  ac_quant >> (xd->bd - 5), &rate, &dist);
853 #else
854     vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs], ac_quant >> 3,
855                                  &rate, &dist);
856 #endif  // CONFIG_VP9_HIGHBITDEPTH
857 
858     this_rdc->rate += rate;
859     this_rdc->dist += dist << 4;
860   }
861 
862 #if CONFIG_VP9_HIGHBITDEPTH
863   *var_y = tot_var > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_var;
864   *sse_y = tot_sse > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_sse;
865 #else
866   *var_y = tot_var;
867   *sse_y = tot_sse;
868 #endif
869 }
870 
get_pred_buffer(PRED_BUFFER * p,int len)871 static int get_pred_buffer(PRED_BUFFER *p, int len) {
872   int i;
873 
874   for (i = 0; i < len; i++) {
875     if (!p[i].in_use) {
876       p[i].in_use = 1;
877       return i;
878     }
879   }
880   return -1;
881 }
882 
free_pred_buffer(PRED_BUFFER * p)883 static void free_pred_buffer(PRED_BUFFER *p) {
884   if (p != NULL) p->in_use = 0;
885 }
886 
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,int * flag_preduv_computed)887 static void encode_breakout_test(
888     VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, int mi_row, int mi_col,
889     MV_REFERENCE_FRAME ref_frame, PREDICTION_MODE this_mode, unsigned int var_y,
890     unsigned int sse_y, struct buf_2d yv12_mb[][MAX_MB_PLANE], int *rate,
891     int64_t *dist, int *flag_preduv_computed) {
892   MACROBLOCKD *xd = &x->e_mbd;
893   MODE_INFO *const mi = xd->mi[0];
894   const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
895   unsigned int var = var_y, sse = sse_y;
896   // Skipping threshold for ac.
897   unsigned int thresh_ac;
898   // Skipping threshold for dc.
899   unsigned int thresh_dc;
900   int motion_low = 1;
901 
902   if (cpi->use_svc && ref_frame == GOLDEN_FRAME) return;
903   if (mi->mv[0].as_mv.row > 64 || mi->mv[0].as_mv.row < -64 ||
904       mi->mv[0].as_mv.col > 64 || mi->mv[0].as_mv.col < -64)
905     motion_low = 0;
906   if (x->encode_breakout > 0 && motion_low == 1) {
907     // Set a maximum for threshold to avoid big PSNR loss in low bit rate
908     // case. Use extreme low threshold for static frames to limit
909     // skipping.
910     const unsigned int max_thresh = 36000;
911     // The encode_breakout input
912     const unsigned int min_thresh =
913         VPXMIN(((unsigned int)x->encode_breakout << 4), max_thresh);
914 #if CONFIG_VP9_HIGHBITDEPTH
915     const int shift = (xd->bd << 1) - 16;
916 #endif
917 
918     // Calculate threshold according to dequant value.
919     thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) >> 3;
920 #if CONFIG_VP9_HIGHBITDEPTH
921     if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
922       thresh_ac = ROUND_POWER_OF_TWO(thresh_ac, shift);
923     }
924 #endif  // CONFIG_VP9_HIGHBITDEPTH
925     thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
926 
927     // Adjust ac threshold according to partition size.
928     thresh_ac >>=
929         8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
930 
931     thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
932 #if CONFIG_VP9_HIGHBITDEPTH
933     if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
934       thresh_dc = ROUND_POWER_OF_TWO(thresh_dc, shift);
935     }
936 #endif  // CONFIG_VP9_HIGHBITDEPTH
937   } else {
938     thresh_ac = 0;
939     thresh_dc = 0;
940   }
941 
942   // Y skipping condition checking for ac and dc.
943   if (var <= thresh_ac && (sse - var) <= thresh_dc) {
944     unsigned int sse_u, sse_v;
945     unsigned int var_u, var_v;
946     unsigned int thresh_ac_uv = thresh_ac;
947     unsigned int thresh_dc_uv = thresh_dc;
948     if (x->sb_is_skin) {
949       thresh_ac_uv = 0;
950       thresh_dc_uv = 0;
951     }
952 
953     if (!flag_preduv_computed[0] || !flag_preduv_computed[1]) {
954       xd->plane[1].pre[0] = yv12_mb[ref_frame][1];
955       xd->plane[2].pre[0] = yv12_mb[ref_frame][2];
956       vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize);
957     }
958 
959     var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf, x->plane[1].src.stride,
960                                     xd->plane[1].dst.buf,
961                                     xd->plane[1].dst.stride, &sse_u);
962 
963     // U skipping condition checking
964     if (((var_u << 2) <= thresh_ac_uv) && (sse_u - var_u <= thresh_dc_uv)) {
965       var_v = cpi->fn_ptr[uv_size].vf(
966           x->plane[2].src.buf, x->plane[2].src.stride, xd->plane[2].dst.buf,
967           xd->plane[2].dst.stride, &sse_v);
968 
969       // V skipping condition checking
970       if (((var_v << 2) <= thresh_ac_uv) && (sse_v - var_v <= thresh_dc_uv)) {
971         x->skip = 1;
972 
973         // The cost of skip bit needs to be added.
974         *rate = cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
975                                     [INTER_OFFSET(this_mode)];
976 
977         // More on this part of rate
978         // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
979 
980         // Scaling factor for SSE from spatial domain to frequency
981         // domain is 16. Adjust distortion accordingly.
982         // TODO(yunqingwang): In this function, only y-plane dist is
983         // calculated.
984         *dist = (sse << 4);  // + ((sse_u + sse_v) << 4);
985 
986         // *disable_skip = 1;
987       }
988     }
989   }
990 }
991 
992 struct estimate_block_intra_args {
993   VP9_COMP *cpi;
994   MACROBLOCK *x;
995   PREDICTION_MODE mode;
996   int skippable;
997   RD_COST *rdc;
998 };
999 
estimate_block_intra(int plane,int block,int row,int col,BLOCK_SIZE plane_bsize,TX_SIZE tx_size,void * arg)1000 static void estimate_block_intra(int plane, int block, int row, int col,
1001                                  BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
1002                                  void *arg) {
1003   struct estimate_block_intra_args *const args = arg;
1004   VP9_COMP *const cpi = args->cpi;
1005   MACROBLOCK *const x = args->x;
1006   MACROBLOCKD *const xd = &x->e_mbd;
1007   struct macroblock_plane *const p = &x->plane[0];
1008   struct macroblockd_plane *const pd = &xd->plane[0];
1009   const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
1010   uint8_t *const src_buf_base = p->src.buf;
1011   uint8_t *const dst_buf_base = pd->dst.buf;
1012   const int src_stride = p->src.stride;
1013   const int dst_stride = pd->dst.stride;
1014   RD_COST this_rdc;
1015 
1016   (void)block;
1017 
1018   p->src.buf = &src_buf_base[4 * (row * src_stride + col)];
1019   pd->dst.buf = &dst_buf_base[4 * (row * dst_stride + col)];
1020   // Use source buffer as an approximation for the fully reconstructed buffer.
1021   vp9_predict_intra_block(xd, b_width_log2_lookup[plane_bsize], tx_size,
1022                           args->mode, x->skip_encode ? p->src.buf : pd->dst.buf,
1023                           x->skip_encode ? src_stride : dst_stride, pd->dst.buf,
1024                           dst_stride, col, row, plane);
1025 
1026   if (plane == 0) {
1027     int64_t this_sse = INT64_MAX;
1028     // TODO(jingning): This needs further refactoring.
1029     block_yrd(cpi, x, &this_rdc, &args->skippable, &this_sse, bsize_tx,
1030               VPXMIN(tx_size, TX_16X16), 0);
1031   } else {
1032     unsigned int var = 0;
1033     unsigned int sse = 0;
1034     model_rd_for_sb_uv(cpi, plane_bsize, x, xd, &this_rdc, &var, &sse, plane,
1035                        plane);
1036   }
1037 
1038   p->src.buf = src_buf_base;
1039   pd->dst.buf = dst_buf_base;
1040   args->rdc->rate += this_rdc.rate;
1041   args->rdc->dist += this_rdc.dist;
1042 }
1043 
1044 static const THR_MODES mode_idx[MAX_REF_FRAMES][4] = {
1045   { THR_DC, THR_V_PRED, THR_H_PRED, THR_TM },
1046   { THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV },
1047   { THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG },
1048   { THR_NEARESTA, THR_NEARA, THR_ZEROA, THR_NEWA },
1049 };
1050 
1051 static const PREDICTION_MODE intra_mode_list[] = { DC_PRED, V_PRED, H_PRED,
1052                                                    TM_PRED };
1053 
mode_offset(const PREDICTION_MODE mode)1054 static int mode_offset(const PREDICTION_MODE mode) {
1055   if (mode >= NEARESTMV) {
1056     return INTER_OFFSET(mode);
1057   } else {
1058     switch (mode) {
1059       case DC_PRED: return 0;
1060       case V_PRED: return 1;
1061       case H_PRED: return 2;
1062       case TM_PRED: return 3;
1063       default: return -1;
1064     }
1065   }
1066 }
1067 
rd_less_than_thresh_row_mt(int64_t best_rd,int thresh,const int * const thresh_fact)1068 static INLINE int rd_less_than_thresh_row_mt(int64_t best_rd, int thresh,
1069                                              const int *const thresh_fact) {
1070   int is_rd_less_than_thresh;
1071   is_rd_less_than_thresh =
1072       best_rd < ((int64_t)thresh * (*thresh_fact) >> 5) || thresh == INT_MAX;
1073   return is_rd_less_than_thresh;
1074 }
1075 
update_thresh_freq_fact_row_mt(VP9_COMP * cpi,TileDataEnc * tile_data,int source_variance,int thresh_freq_fact_idx,MV_REFERENCE_FRAME ref_frame,THR_MODES best_mode_idx,PREDICTION_MODE mode)1076 static INLINE void update_thresh_freq_fact_row_mt(
1077     VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance,
1078     int thresh_freq_fact_idx, MV_REFERENCE_FRAME ref_frame,
1079     THR_MODES best_mode_idx, PREDICTION_MODE mode) {
1080   THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1081   int freq_fact_idx = thresh_freq_fact_idx + thr_mode_idx;
1082   int *freq_fact = &tile_data->row_base_thresh_freq_fact[freq_fact_idx];
1083   if (thr_mode_idx == best_mode_idx)
1084     *freq_fact -= (*freq_fact >> 4);
1085   else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV &&
1086            ref_frame == LAST_FRAME && source_variance < 5) {
1087     *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32);
1088   } else {
1089     *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
1090                         cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1091   }
1092 }
1093 
update_thresh_freq_fact(VP9_COMP * cpi,TileDataEnc * tile_data,int source_variance,BLOCK_SIZE bsize,MV_REFERENCE_FRAME ref_frame,THR_MODES best_mode_idx,PREDICTION_MODE mode)1094 static INLINE void update_thresh_freq_fact(
1095     VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance,
1096     BLOCK_SIZE bsize, MV_REFERENCE_FRAME ref_frame, THR_MODES best_mode_idx,
1097     PREDICTION_MODE mode) {
1098   THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1099   int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx];
1100   if (thr_mode_idx == best_mode_idx)
1101     *freq_fact -= (*freq_fact >> 4);
1102   else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV &&
1103            ref_frame == LAST_FRAME && source_variance < 5) {
1104     *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32);
1105   } else {
1106     *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
1107                         cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1108   }
1109 }
1110 
vp9_pick_intra_mode(VP9_COMP * cpi,MACROBLOCK * x,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)1111 void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
1112                          BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1113   MACROBLOCKD *const xd = &x->e_mbd;
1114   MODE_INFO *const mi = xd->mi[0];
1115   RD_COST this_rdc, best_rdc;
1116   PREDICTION_MODE this_mode;
1117   struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
1118   const TX_SIZE intra_tx_size =
1119       VPXMIN(max_txsize_lookup[bsize],
1120              tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
1121   MODE_INFO *const mic = xd->mi[0];
1122   int *bmode_costs;
1123   const MODE_INFO *above_mi = xd->above_mi;
1124   const MODE_INFO *left_mi = xd->left_mi;
1125   const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1126   const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1127   bmode_costs = cpi->y_mode_costs[A][L];
1128 
1129   (void)ctx;
1130   vp9_rd_cost_reset(&best_rdc);
1131   vp9_rd_cost_reset(&this_rdc);
1132 
1133   mi->ref_frame[0] = INTRA_FRAME;
1134   // Initialize interp_filter here so we do not have to check for inter block
1135   // modes in get_pred_context_switchable_interp()
1136   mi->interp_filter = SWITCHABLE_FILTERS;
1137 
1138   mi->mv[0].as_int = INVALID_MV;
1139   mi->uv_mode = DC_PRED;
1140   memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
1141 
1142   // Change the limit of this loop to add other intra prediction
1143   // mode tests.
1144   for (this_mode = DC_PRED; this_mode <= H_PRED; ++this_mode) {
1145     this_rdc.dist = this_rdc.rate = 0;
1146     args.mode = this_mode;
1147     args.skippable = 1;
1148     args.rdc = &this_rdc;
1149     mi->tx_size = intra_tx_size;
1150     vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
1151                                            &args);
1152     if (args.skippable) {
1153       x->skip_txfm[0] = SKIP_TXFM_AC_DC;
1154       this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
1155     } else {
1156       x->skip_txfm[0] = SKIP_TXFM_NONE;
1157       this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
1158     }
1159     this_rdc.rate += bmode_costs[this_mode];
1160     this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1161 
1162     if (this_rdc.rdcost < best_rdc.rdcost) {
1163       best_rdc = this_rdc;
1164       mi->mode = this_mode;
1165     }
1166   }
1167 
1168   *rd_cost = best_rdc;
1169 }
1170 
init_ref_frame_cost(VP9_COMMON * const cm,MACROBLOCKD * const xd,int ref_frame_cost[MAX_REF_FRAMES])1171 static void init_ref_frame_cost(VP9_COMMON *const cm, MACROBLOCKD *const xd,
1172                                 int ref_frame_cost[MAX_REF_FRAMES]) {
1173   vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
1174   vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
1175   vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
1176 
1177   ref_frame_cost[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
1178   ref_frame_cost[LAST_FRAME] = ref_frame_cost[GOLDEN_FRAME] =
1179       ref_frame_cost[ALTREF_FRAME] = vp9_cost_bit(intra_inter_p, 1);
1180 
1181   ref_frame_cost[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
1182   ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1183   ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1184   ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
1185   ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
1186 }
1187 
1188 typedef struct {
1189   MV_REFERENCE_FRAME ref_frame;
1190   PREDICTION_MODE pred_mode;
1191 } REF_MODE;
1192 
1193 #define RT_INTER_MODES 12
1194 static const REF_MODE ref_mode_set[RT_INTER_MODES] = {
1195   { LAST_FRAME, ZEROMV },   { LAST_FRAME, NEARESTMV },
1196   { GOLDEN_FRAME, ZEROMV }, { LAST_FRAME, NEARMV },
1197   { LAST_FRAME, NEWMV },    { GOLDEN_FRAME, NEARESTMV },
1198   { GOLDEN_FRAME, NEARMV }, { GOLDEN_FRAME, NEWMV },
1199   { ALTREF_FRAME, ZEROMV }, { ALTREF_FRAME, NEARESTMV },
1200   { ALTREF_FRAME, NEARMV }, { ALTREF_FRAME, NEWMV }
1201 };
1202 
1203 #define RT_INTER_MODES_SVC 8
1204 static const REF_MODE ref_mode_set_svc[RT_INTER_MODES_SVC] = {
1205   { LAST_FRAME, ZEROMV },      { LAST_FRAME, NEARESTMV },
1206   { LAST_FRAME, NEARMV },      { GOLDEN_FRAME, ZEROMV },
1207   { GOLDEN_FRAME, NEARESTMV }, { GOLDEN_FRAME, NEARMV },
1208   { LAST_FRAME, NEWMV },       { GOLDEN_FRAME, NEWMV }
1209 };
1210 
find_predictors(VP9_COMP * cpi,MACROBLOCK * x,MV_REFERENCE_FRAME ref_frame,int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],int const_motion[MAX_REF_FRAMES],int * ref_frame_skip_mask,const int flag_list[4],TileDataEnc * tile_data,int mi_row,int mi_col,struct buf_2d yv12_mb[4][MAX_MB_PLANE],BLOCK_SIZE bsize,int force_skip_low_temp_var,int comp_pred_allowed)1211 static INLINE void find_predictors(
1212     VP9_COMP *cpi, MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame,
1213     int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1214     int const_motion[MAX_REF_FRAMES], int *ref_frame_skip_mask,
1215     const int flag_list[4], TileDataEnc *tile_data, int mi_row, int mi_col,
1216     struct buf_2d yv12_mb[4][MAX_MB_PLANE], BLOCK_SIZE bsize,
1217     int force_skip_low_temp_var, int comp_pred_allowed) {
1218   VP9_COMMON *const cm = &cpi->common;
1219   MACROBLOCKD *const xd = &x->e_mbd;
1220   const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
1221   TileInfo *const tile_info = &tile_data->tile_info;
1222   // TODO(jingning) placeholder for inter-frame non-RD mode decision.
1223   x->pred_mv_sad[ref_frame] = INT_MAX;
1224   frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
1225   frame_mv[ZEROMV][ref_frame].as_int = 0;
1226   // this needs various further optimizations. to be continued..
1227   if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
1228     int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
1229     const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
1230     vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
1231     if (cm->use_prev_frame_mvs || comp_pred_allowed) {
1232       vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
1233                        x->mbmi_ext->mode_context);
1234     } else {
1235       const_motion[ref_frame] =
1236           mv_refs_rt(cpi, cm, x, xd, tile_info, xd->mi[0], ref_frame,
1237                      candidates, &frame_mv[NEWMV][ref_frame], mi_row, mi_col,
1238                      (int)(cpi->svc.use_base_mv && cpi->svc.spatial_layer_id));
1239     }
1240     vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
1241                           &frame_mv[NEARESTMV][ref_frame],
1242                           &frame_mv[NEARMV][ref_frame]);
1243     // Early exit for golden frame if force_skip_low_temp_var is set.
1244     if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8 &&
1245         !(force_skip_low_temp_var && ref_frame == GOLDEN_FRAME)) {
1246       vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
1247                   bsize);
1248     }
1249   } else {
1250     *ref_frame_skip_mask |= (1 << ref_frame);
1251   }
1252 }
1253 
vp9_NEWMV_diff_bias(const NOISE_ESTIMATE * ne,MACROBLOCKD * xd,PREDICTION_MODE this_mode,RD_COST * this_rdc,BLOCK_SIZE bsize,int mv_row,int mv_col,int is_last_frame,int lowvar_highsumdiff,int is_skin)1254 static void vp9_NEWMV_diff_bias(const NOISE_ESTIMATE *ne, MACROBLOCKD *xd,
1255                                 PREDICTION_MODE this_mode, RD_COST *this_rdc,
1256                                 BLOCK_SIZE bsize, int mv_row, int mv_col,
1257                                 int is_last_frame, int lowvar_highsumdiff,
1258                                 int is_skin) {
1259   // Bias against MVs associated with NEWMV mode that are very different from
1260   // top/left neighbors.
1261   if (this_mode == NEWMV) {
1262     int al_mv_average_row;
1263     int al_mv_average_col;
1264     int left_row, left_col;
1265     int row_diff, col_diff;
1266     int above_mv_valid = 0;
1267     int left_mv_valid = 0;
1268     int above_row = 0;
1269     int above_col = 0;
1270 
1271     if (xd->above_mi) {
1272       above_mv_valid = xd->above_mi->mv[0].as_int != INVALID_MV;
1273       above_row = xd->above_mi->mv[0].as_mv.row;
1274       above_col = xd->above_mi->mv[0].as_mv.col;
1275     }
1276     if (xd->left_mi) {
1277       left_mv_valid = xd->left_mi->mv[0].as_int != INVALID_MV;
1278       left_row = xd->left_mi->mv[0].as_mv.row;
1279       left_col = xd->left_mi->mv[0].as_mv.col;
1280     }
1281     if (above_mv_valid && left_mv_valid) {
1282       al_mv_average_row = (above_row + left_row + 1) >> 1;
1283       al_mv_average_col = (above_col + left_col + 1) >> 1;
1284     } else if (above_mv_valid) {
1285       al_mv_average_row = above_row;
1286       al_mv_average_col = above_col;
1287     } else if (left_mv_valid) {
1288       al_mv_average_row = left_row;
1289       al_mv_average_col = left_col;
1290     } else {
1291       al_mv_average_row = al_mv_average_col = 0;
1292     }
1293     row_diff = (al_mv_average_row - mv_row);
1294     col_diff = (al_mv_average_col - mv_col);
1295     if (row_diff > 48 || row_diff < -48 || col_diff > 48 || col_diff < -48) {
1296       if (bsize > BLOCK_32X32)
1297         this_rdc->rdcost = this_rdc->rdcost << 1;
1298       else
1299         this_rdc->rdcost = 3 * this_rdc->rdcost >> 1;
1300     }
1301   }
1302   // If noise estimation is enabled, and estimated level is above threshold,
1303   // add a bias to LAST reference with small motion, for large blocks.
1304   if (ne->enabled && ne->level >= kMedium && bsize >= BLOCK_32X32 &&
1305       is_last_frame && mv_row < 8 && mv_row > -8 && mv_col < 8 && mv_col > -8)
1306     this_rdc->rdcost = 7 * (this_rdc->rdcost >> 3);
1307   else if (lowvar_highsumdiff && !is_skin && bsize >= BLOCK_16X16 &&
1308            is_last_frame && mv_row < 16 && mv_row > -16 && mv_col < 16 &&
1309            mv_col > -16)
1310     this_rdc->rdcost = 7 * (this_rdc->rdcost >> 3);
1311 }
1312 
1313 #if CONFIG_VP9_TEMPORAL_DENOISING
vp9_pickmode_ctx_den_update(VP9_PICKMODE_CTX_DEN * ctx_den,int64_t zero_last_cost_orig,int ref_frame_cost[MAX_REF_FRAMES],int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],int reuse_inter_pred,BEST_PICKMODE * bp)1314 static void vp9_pickmode_ctx_den_update(
1315     VP9_PICKMODE_CTX_DEN *ctx_den, int64_t zero_last_cost_orig,
1316     int ref_frame_cost[MAX_REF_FRAMES],
1317     int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int reuse_inter_pred,
1318     BEST_PICKMODE *bp) {
1319   ctx_den->zero_last_cost_orig = zero_last_cost_orig;
1320   ctx_den->ref_frame_cost = ref_frame_cost;
1321   ctx_den->frame_mv = frame_mv;
1322   ctx_den->reuse_inter_pred = reuse_inter_pred;
1323   ctx_den->best_tx_size = bp->best_tx_size;
1324   ctx_den->best_mode = bp->best_mode;
1325   ctx_den->best_ref_frame = bp->best_ref_frame;
1326   ctx_den->best_pred_filter = bp->best_pred_filter;
1327   ctx_den->best_mode_skip_txfm = bp->best_mode_skip_txfm;
1328 }
1329 
recheck_zeromv_after_denoising(VP9_COMP * cpi,MODE_INFO * const mi,MACROBLOCK * x,MACROBLOCKD * const xd,VP9_DENOISER_DECISION decision,VP9_PICKMODE_CTX_DEN * ctx_den,struct buf_2d yv12_mb[4][MAX_MB_PLANE],RD_COST * best_rdc,BLOCK_SIZE bsize,int mi_row,int mi_col)1330 static void recheck_zeromv_after_denoising(
1331     VP9_COMP *cpi, MODE_INFO *const mi, MACROBLOCK *x, MACROBLOCKD *const xd,
1332     VP9_DENOISER_DECISION decision, VP9_PICKMODE_CTX_DEN *ctx_den,
1333     struct buf_2d yv12_mb[4][MAX_MB_PLANE], RD_COST *best_rdc, BLOCK_SIZE bsize,
1334     int mi_row, int mi_col) {
1335   // If INTRA or GOLDEN reference was selected, re-evaluate ZEROMV on
1336   // denoised result. Only do this under noise conditions, and if rdcost of
1337   // ZEROMV onoriginal source is not significantly higher than rdcost of best
1338   // mode.
1339   if (cpi->noise_estimate.enabled && cpi->noise_estimate.level > kLow &&
1340       ctx_den->zero_last_cost_orig < (best_rdc->rdcost << 3) &&
1341       ((ctx_den->best_ref_frame == INTRA_FRAME && decision >= FILTER_BLOCK) ||
1342        (ctx_den->best_ref_frame == GOLDEN_FRAME &&
1343         cpi->svc.number_spatial_layers == 1 &&
1344         decision == FILTER_ZEROMV_BLOCK))) {
1345     // Check if we should pick ZEROMV on denoised signal.
1346     int rate = 0;
1347     int64_t dist = 0;
1348     uint32_t var_y = UINT_MAX;
1349     uint32_t sse_y = UINT_MAX;
1350     RD_COST this_rdc;
1351     mi->mode = ZEROMV;
1352     mi->ref_frame[0] = LAST_FRAME;
1353     mi->ref_frame[1] = NONE;
1354     mi->mv[0].as_int = 0;
1355     mi->interp_filter = EIGHTTAP;
1356     if (cpi->sf.default_interp_filter == BILINEAR) mi->interp_filter = BILINEAR;
1357     xd->plane[0].pre[0] = yv12_mb[LAST_FRAME][0];
1358     vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1359     model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);
1360     this_rdc.rate = rate + ctx_den->ref_frame_cost[LAST_FRAME] +
1361                     cpi->inter_mode_cost[x->mbmi_ext->mode_context[LAST_FRAME]]
1362                                         [INTER_OFFSET(ZEROMV)];
1363     this_rdc.dist = dist;
1364     this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, rate, dist);
1365     // Don't switch to ZEROMV if the rdcost for ZEROMV on denoised source
1366     // is higher than best_ref mode (on original source).
1367     if (this_rdc.rdcost > best_rdc->rdcost) {
1368       this_rdc = *best_rdc;
1369       mi->mode = ctx_den->best_mode;
1370       mi->ref_frame[0] = ctx_den->best_ref_frame;
1371       mi->interp_filter = ctx_den->best_pred_filter;
1372       if (ctx_den->best_ref_frame == INTRA_FRAME) {
1373         mi->mv[0].as_int = INVALID_MV;
1374         mi->interp_filter = SWITCHABLE_FILTERS;
1375       } else if (ctx_den->best_ref_frame == GOLDEN_FRAME) {
1376         mi->mv[0].as_int =
1377             ctx_den->frame_mv[ctx_den->best_mode][ctx_den->best_ref_frame]
1378                 .as_int;
1379         if (ctx_den->reuse_inter_pred) {
1380           xd->plane[0].pre[0] = yv12_mb[GOLDEN_FRAME][0];
1381           vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1382         }
1383       }
1384       mi->tx_size = ctx_den->best_tx_size;
1385       x->skip_txfm[0] = ctx_den->best_mode_skip_txfm;
1386     } else {
1387       ctx_den->best_ref_frame = LAST_FRAME;
1388       *best_rdc = this_rdc;
1389     }
1390   }
1391 }
1392 #endif  // CONFIG_VP9_TEMPORAL_DENOISING
1393 
get_force_skip_low_temp_var(uint8_t * variance_low,int mi_row,int mi_col,BLOCK_SIZE bsize)1394 static INLINE int get_force_skip_low_temp_var(uint8_t *variance_low, int mi_row,
1395                                               int mi_col, BLOCK_SIZE bsize) {
1396   const int i = (mi_row & 0x7) >> 1;
1397   const int j = (mi_col & 0x7) >> 1;
1398   int force_skip_low_temp_var = 0;
1399   // Set force_skip_low_temp_var based on the block size and block offset.
1400   if (bsize == BLOCK_64X64) {
1401     force_skip_low_temp_var = variance_low[0];
1402   } else if (bsize == BLOCK_64X32) {
1403     if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1404       force_skip_low_temp_var = variance_low[1];
1405     } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1406       force_skip_low_temp_var = variance_low[2];
1407     }
1408   } else if (bsize == BLOCK_32X64) {
1409     if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1410       force_skip_low_temp_var = variance_low[3];
1411     } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1412       force_skip_low_temp_var = variance_low[4];
1413     }
1414   } else if (bsize == BLOCK_32X32) {
1415     if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1416       force_skip_low_temp_var = variance_low[5];
1417     } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1418       force_skip_low_temp_var = variance_low[6];
1419     } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1420       force_skip_low_temp_var = variance_low[7];
1421     } else if ((mi_col & 0x7) && (mi_row & 0x7)) {
1422       force_skip_low_temp_var = variance_low[8];
1423     }
1424   } else if (bsize == BLOCK_16X16) {
1425     force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]];
1426   } else if (bsize == BLOCK_32X16) {
1427     // The col shift index for the second 16x16 block.
1428     const int j2 = ((mi_col + 2) & 0x7) >> 1;
1429     // Only if each 16x16 block inside has low temporal variance.
1430     force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1431                               variance_low[pos_shift_16x16[i][j2]];
1432   } else if (bsize == BLOCK_16X32) {
1433     // The row shift index for the second 16x16 block.
1434     const int i2 = ((mi_row + 2) & 0x7) >> 1;
1435     force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1436                               variance_low[pos_shift_16x16[i2][j]];
1437   }
1438   return force_skip_low_temp_var;
1439 }
1440 
search_filter_ref(VP9_COMP * cpi,MACROBLOCK * x,RD_COST * this_rdc,int mi_row,int mi_col,PRED_BUFFER * tmp,BLOCK_SIZE bsize,int reuse_inter_pred,PRED_BUFFER ** this_mode_pred,unsigned int * var_y,unsigned int * sse_y)1441 static void search_filter_ref(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc,
1442                               int mi_row, int mi_col, PRED_BUFFER *tmp,
1443                               BLOCK_SIZE bsize, int reuse_inter_pred,
1444                               PRED_BUFFER **this_mode_pred, unsigned int *var_y,
1445                               unsigned int *sse_y) {
1446   MACROBLOCKD *const xd = &x->e_mbd;
1447   MODE_INFO *const mi = xd->mi[0];
1448   struct macroblockd_plane *const pd = &xd->plane[0];
1449   const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
1450 
1451   int pf_rate[3] = { 0 };
1452   int64_t pf_dist[3] = { 0 };
1453   int curr_rate[3] = { 0 };
1454   unsigned int pf_var[3] = { 0 };
1455   unsigned int pf_sse[3] = { 0 };
1456   TX_SIZE pf_tx_size[3] = { 0 };
1457   int64_t best_cost = INT64_MAX;
1458   INTERP_FILTER best_filter = SWITCHABLE, filter;
1459   PRED_BUFFER *current_pred = *this_mode_pred;
1460   uint8_t skip_txfm = SKIP_TXFM_NONE;
1461 
1462   for (filter = EIGHTTAP; filter <= EIGHTTAP_SMOOTH; ++filter) {
1463     int64_t cost;
1464     mi->interp_filter = filter;
1465     vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1466     model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter], &pf_dist[filter],
1467                       &pf_var[filter], &pf_sse[filter]);
1468     curr_rate[filter] = pf_rate[filter];
1469     pf_rate[filter] += vp9_get_switchable_rate(cpi, xd);
1470     cost = RDCOST(x->rdmult, x->rddiv, pf_rate[filter], pf_dist[filter]);
1471     pf_tx_size[filter] = mi->tx_size;
1472     if (cost < best_cost) {
1473       best_filter = filter;
1474       best_cost = cost;
1475       skip_txfm = x->skip_txfm[0];
1476 
1477       if (reuse_inter_pred) {
1478         if (*this_mode_pred != current_pred) {
1479           free_pred_buffer(*this_mode_pred);
1480           *this_mode_pred = current_pred;
1481         }
1482         current_pred = &tmp[get_pred_buffer(tmp, 3)];
1483         pd->dst.buf = current_pred->data;
1484         pd->dst.stride = bw;
1485       }
1486     }
1487   }
1488 
1489   if (reuse_inter_pred && *this_mode_pred != current_pred)
1490     free_pred_buffer(current_pred);
1491 
1492   mi->interp_filter = best_filter;
1493   mi->tx_size = pf_tx_size[best_filter];
1494   this_rdc->rate = curr_rate[best_filter];
1495   this_rdc->dist = pf_dist[best_filter];
1496   *var_y = pf_var[best_filter];
1497   *sse_y = pf_sse[best_filter];
1498   x->skip_txfm[0] = skip_txfm;
1499   if (reuse_inter_pred) {
1500     pd->dst.buf = (*this_mode_pred)->data;
1501     pd->dst.stride = (*this_mode_pred)->stride;
1502   }
1503 }
1504 
search_new_mv(VP9_COMP * cpi,MACROBLOCK * x,int_mv frame_mv[][MAX_REF_FRAMES],MV_REFERENCE_FRAME ref_frame,int gf_temporal_ref,BLOCK_SIZE bsize,int mi_row,int mi_col,int best_pred_sad,int * rate_mv,unsigned int best_sse_sofar,RD_COST * best_rdc)1505 static int search_new_mv(VP9_COMP *cpi, MACROBLOCK *x,
1506                          int_mv frame_mv[][MAX_REF_FRAMES],
1507                          MV_REFERENCE_FRAME ref_frame, int gf_temporal_ref,
1508                          BLOCK_SIZE bsize, int mi_row, int mi_col,
1509                          int best_pred_sad, int *rate_mv,
1510                          unsigned int best_sse_sofar, RD_COST *best_rdc) {
1511   SVC *const svc = &cpi->svc;
1512   MACROBLOCKD *const xd = &x->e_mbd;
1513   MODE_INFO *const mi = xd->mi[0];
1514   SPEED_FEATURES *const sf = &cpi->sf;
1515 
1516   if (ref_frame > LAST_FRAME && gf_temporal_ref &&
1517       cpi->oxcf.rc_mode == VPX_CBR) {
1518     int tmp_sad;
1519     uint32_t dis;
1520     int cost_list[5] = { INT_MAX, INT_MAX, INT_MAX, INT_MAX, INT_MAX };
1521 
1522     if (bsize < BLOCK_16X16) return -1;
1523 
1524     tmp_sad = vp9_int_pro_motion_estimation(
1525         cpi, x, bsize, mi_row, mi_col,
1526         &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv);
1527 
1528     if (tmp_sad > x->pred_mv_sad[LAST_FRAME]) return -1;
1529     if (tmp_sad + (num_pels_log2_lookup[bsize] << 4) > best_pred_sad) return -1;
1530 
1531     frame_mv[NEWMV][ref_frame].as_int = mi->mv[0].as_int;
1532     *rate_mv = vp9_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv,
1533                                &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1534                                x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1535     frame_mv[NEWMV][ref_frame].as_mv.row >>= 3;
1536     frame_mv[NEWMV][ref_frame].as_mv.col >>= 3;
1537 
1538     cpi->find_fractional_mv_step(
1539         x, &frame_mv[NEWMV][ref_frame].as_mv,
1540         &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1541         cpi->common.allow_high_precision_mv, x->errorperbit,
1542         &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
1543         cpi->sf.mv.subpel_search_level, cond_cost_list(cpi, cost_list),
1544         x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref_frame], NULL, 0, 0,
1545         cpi->sf.use_accurate_subpel_search);
1546   } else if (svc->use_base_mv && svc->spatial_layer_id) {
1547     if (frame_mv[NEWMV][ref_frame].as_int != INVALID_MV) {
1548       const int pre_stride = xd->plane[0].pre[0].stride;
1549       unsigned int base_mv_sse = UINT_MAX;
1550       int scale = (cpi->rc.avg_frame_low_motion > 60) ? 2 : 4;
1551       const uint8_t *const pre_buf =
1552           xd->plane[0].pre[0].buf +
1553           (frame_mv[NEWMV][ref_frame].as_mv.row >> 3) * pre_stride +
1554           (frame_mv[NEWMV][ref_frame].as_mv.col >> 3);
1555       cpi->fn_ptr[bsize].vf(x->plane[0].src.buf, x->plane[0].src.stride,
1556                             pre_buf, pre_stride, &base_mv_sse);
1557 
1558       // Exit NEWMV search if base_mv is (0,0) && bsize < BLOCK_16x16,
1559       // for SVC encoding.
1560       if (cpi->use_svc && svc->use_base_mv && bsize < BLOCK_16X16 &&
1561           frame_mv[NEWMV][ref_frame].as_mv.row == 0 &&
1562           frame_mv[NEWMV][ref_frame].as_mv.col == 0)
1563         return -1;
1564 
1565       // Exit NEWMV search if base_mv_sse is large.
1566       if (sf->base_mv_aggressive && base_mv_sse > (best_sse_sofar << scale))
1567         return -1;
1568       if (base_mv_sse < (best_sse_sofar << 1)) {
1569         // Base layer mv is good.
1570         // Exit NEWMV search if the base_mv is (0, 0) and sse is low, since
1571         // (0, 0) mode is already tested.
1572         unsigned int base_mv_sse_normalized =
1573             base_mv_sse >>
1574             (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
1575         if (sf->base_mv_aggressive && base_mv_sse <= best_sse_sofar &&
1576             base_mv_sse_normalized < 400 &&
1577             frame_mv[NEWMV][ref_frame].as_mv.row == 0 &&
1578             frame_mv[NEWMV][ref_frame].as_mv.col == 0)
1579           return -1;
1580         if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1581                                     &frame_mv[NEWMV][ref_frame], rate_mv,
1582                                     best_rdc->rdcost, 1)) {
1583           return -1;
1584         }
1585       } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1586                                          &frame_mv[NEWMV][ref_frame], rate_mv,
1587                                          best_rdc->rdcost, 0)) {
1588         return -1;
1589       }
1590     } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1591                                        &frame_mv[NEWMV][ref_frame], rate_mv,
1592                                        best_rdc->rdcost, 0)) {
1593       return -1;
1594     }
1595   } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1596                                      &frame_mv[NEWMV][ref_frame], rate_mv,
1597                                      best_rdc->rdcost, 0)) {
1598     return -1;
1599   }
1600 
1601   return 0;
1602 }
1603 
init_best_pickmode(BEST_PICKMODE * bp)1604 static INLINE void init_best_pickmode(BEST_PICKMODE *bp) {
1605   bp->best_mode = ZEROMV;
1606   bp->best_ref_frame = LAST_FRAME;
1607   bp->best_tx_size = TX_SIZES;
1608   bp->best_intra_tx_size = TX_SIZES;
1609   bp->best_pred_filter = EIGHTTAP;
1610   bp->best_mode_skip_txfm = SKIP_TXFM_NONE;
1611   bp->best_second_ref_frame = NONE;
1612   bp->best_pred = NULL;
1613 }
1614 
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)1615 void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data,
1616                          int mi_row, int mi_col, RD_COST *rd_cost,
1617                          BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1618   VP9_COMMON *const cm = &cpi->common;
1619   SPEED_FEATURES *const sf = &cpi->sf;
1620   SVC *const svc = &cpi->svc;
1621   MACROBLOCKD *const xd = &x->e_mbd;
1622   MODE_INFO *const mi = xd->mi[0];
1623   struct macroblockd_plane *const pd = &xd->plane[0];
1624 
1625   BEST_PICKMODE best_pickmode;
1626 
1627   MV_REFERENCE_FRAME ref_frame;
1628   MV_REFERENCE_FRAME usable_ref_frame, second_ref_frame;
1629   int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1630   uint8_t mode_checked[MB_MODE_COUNT][MAX_REF_FRAMES];
1631   struct buf_2d yv12_mb[4][MAX_MB_PLANE];
1632   static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
1633                                     VP9_ALT_FLAG };
1634   RD_COST this_rdc, best_rdc;
1635   // var_y and sse_y are saved to be used in skipping checking
1636   unsigned int var_y = UINT_MAX;
1637   unsigned int sse_y = UINT_MAX;
1638   const int intra_cost_penalty =
1639       vp9_get_intra_cost_penalty(cpi, bsize, cm->base_qindex, cm->y_dc_delta_q);
1640   int64_t inter_mode_thresh =
1641       RDCOST(x->rdmult, x->rddiv, intra_cost_penalty, 0);
1642   const int *const rd_threshes = cpi->rd.threshes[mi->segment_id][bsize];
1643   const int sb_row = mi_row >> MI_BLOCK_SIZE_LOG2;
1644   int thresh_freq_fact_idx = (sb_row * BLOCK_SIZES + bsize) * MAX_MODES;
1645   const int *const rd_thresh_freq_fact =
1646       (cpi->sf.adaptive_rd_thresh_row_mt)
1647           ? &(tile_data->row_base_thresh_freq_fact[thresh_freq_fact_idx])
1648           : tile_data->thresh_freq_fact[bsize];
1649 
1650   INTERP_FILTER filter_ref;
1651   const int bsl = mi_width_log2_lookup[bsize];
1652   const int pred_filter_search =
1653       cm->interp_filter == SWITCHABLE
1654           ? (((mi_row + mi_col) >> bsl) +
1655              get_chessboard_index(cm->current_video_frame)) &
1656                 0x1
1657           : 0;
1658   int const_motion[MAX_REF_FRAMES] = { 0 };
1659   const int bh = num_4x4_blocks_high_lookup[bsize] << 2;
1660   const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
1661   // For speed 6, the result of interp filter is reused later in actual encoding
1662   // process.
1663   // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.
1664   PRED_BUFFER tmp[4];
1665   DECLARE_ALIGNED(16, uint8_t, pred_buf[3 * 64 * 64]);
1666 #if CONFIG_VP9_HIGHBITDEPTH
1667   DECLARE_ALIGNED(16, uint16_t, pred_buf_16[3 * 64 * 64]);
1668 #endif
1669   struct buf_2d orig_dst = pd->dst;
1670   PRED_BUFFER *this_mode_pred = NULL;
1671   const int pixels_in_block = bh * bw;
1672   int reuse_inter_pred = cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready;
1673   int ref_frame_skip_mask = 0;
1674   int idx;
1675   int best_pred_sad = INT_MAX;
1676   int best_early_term = 0;
1677   int ref_frame_cost[MAX_REF_FRAMES];
1678   int svc_force_zero_mode[3] = { 0 };
1679   int perform_intra_pred = 1;
1680   int use_golden_nonzeromv = 1;
1681   int force_skip_low_temp_var = 0;
1682   int skip_ref_find_pred[4] = { 0 };
1683   unsigned int sse_zeromv_normalized = UINT_MAX;
1684   unsigned int best_sse_sofar = UINT_MAX;
1685   int gf_temporal_ref = 0;
1686 #if CONFIG_VP9_TEMPORAL_DENOISING
1687   VP9_PICKMODE_CTX_DEN ctx_den;
1688   int64_t zero_last_cost_orig = INT64_MAX;
1689   int denoise_svc_pickmode = 1;
1690 #endif
1691   INTERP_FILTER filter_gf_svc = EIGHTTAP;
1692   MV_REFERENCE_FRAME inter_layer_ref = GOLDEN_FRAME;
1693   const struct segmentation *const seg = &cm->seg;
1694   int comp_modes = 0;
1695   int num_inter_modes = (cpi->use_svc) ? RT_INTER_MODES_SVC : RT_INTER_MODES;
1696   int flag_svc_subpel = 0;
1697   int svc_mv_col = 0;
1698   int svc_mv_row = 0;
1699   int no_scaling = 0;
1700   unsigned int thresh_svc_skip_golden = 500;
1701   int scene_change_detected =
1702       cpi->rc.high_source_sad ||
1703       (cpi->use_svc && cpi->svc.high_source_sad_superframe);
1704 
1705   init_best_pickmode(&best_pickmode);
1706 
1707   x->encode_breakout = seg->enabled
1708                            ? cpi->segment_encode_breakout[mi->segment_id]
1709                            : cpi->encode_breakout;
1710 
1711   x->source_variance = UINT_MAX;
1712   if (cpi->sf.default_interp_filter == BILINEAR) {
1713     best_pickmode.best_pred_filter = BILINEAR;
1714     filter_gf_svc = BILINEAR;
1715   }
1716   if (cpi->use_svc && svc->spatial_layer_id > 0) {
1717     int layer =
1718         LAYER_IDS_TO_IDX(svc->spatial_layer_id - 1, svc->temporal_layer_id,
1719                          svc->number_temporal_layers);
1720     LAYER_CONTEXT *const lc = &svc->layer_context[layer];
1721     if (lc->scaling_factor_num == lc->scaling_factor_den) no_scaling = 1;
1722   }
1723   if (svc->spatial_layer_id > 0 &&
1724       (svc->high_source_sad_superframe || no_scaling))
1725     thresh_svc_skip_golden = 0;
1726   // Lower the skip threshold if lower spatial layer is better quality relative
1727   // to current layer.
1728   else if (svc->spatial_layer_id > 0 && cm->base_qindex > 150 &&
1729            cm->base_qindex > svc->lower_layer_qindex + 15)
1730     thresh_svc_skip_golden = 100;
1731   // Increase skip threshold if lower spatial layer is lower quality relative
1732   // to current layer.
1733   else if (svc->spatial_layer_id > 0 && cm->base_qindex < 140 &&
1734            cm->base_qindex < svc->lower_layer_qindex - 20)
1735     thresh_svc_skip_golden = 1000;
1736 
1737   if (!cpi->use_svc ||
1738       (svc->use_gf_temporal_ref_current_layer &&
1739        !svc->layer_context[svc->temporal_layer_id].is_key_frame)) {
1740     struct scale_factors *const sf_last = &cm->frame_refs[LAST_FRAME - 1].sf;
1741     struct scale_factors *const sf_golden =
1742         &cm->frame_refs[GOLDEN_FRAME - 1].sf;
1743     gf_temporal_ref = 1;
1744     // For temporal long term prediction, check that the golden reference
1745     // is same scale as last reference, otherwise disable.
1746     if ((sf_last->x_scale_fp != sf_golden->x_scale_fp) ||
1747         (sf_last->y_scale_fp != sf_golden->y_scale_fp)) {
1748       gf_temporal_ref = 0;
1749     } else {
1750       if (cpi->rc.avg_frame_low_motion > 70)
1751         thresh_svc_skip_golden = 500;
1752       else
1753         thresh_svc_skip_golden = 0;
1754     }
1755   }
1756 
1757   init_ref_frame_cost(cm, xd, ref_frame_cost);
1758   memset(&mode_checked[0][0], 0, MB_MODE_COUNT * MAX_REF_FRAMES);
1759 
1760   if (reuse_inter_pred) {
1761     int i;
1762     for (i = 0; i < 3; i++) {
1763 #if CONFIG_VP9_HIGHBITDEPTH
1764       if (cm->use_highbitdepth)
1765         tmp[i].data = CONVERT_TO_BYTEPTR(&pred_buf_16[pixels_in_block * i]);
1766       else
1767         tmp[i].data = &pred_buf[pixels_in_block * i];
1768 #else
1769       tmp[i].data = &pred_buf[pixels_in_block * i];
1770 #endif  // CONFIG_VP9_HIGHBITDEPTH
1771       tmp[i].stride = bw;
1772       tmp[i].in_use = 0;
1773     }
1774     tmp[3].data = pd->dst.buf;
1775     tmp[3].stride = pd->dst.stride;
1776     tmp[3].in_use = 0;
1777   }
1778 
1779   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
1780   x->skip = 0;
1781 
1782   // Instead of using vp9_get_pred_context_switchable_interp(xd) to assign
1783   // filter_ref, we use a less strict condition on assigning filter_ref.
1784   // This is to reduce the probabily of entering the flow of not assigning
1785   // filter_ref and then skip filter search.
1786   filter_ref = cm->interp_filter;
1787   if (cpi->sf.default_interp_filter != BILINEAR) {
1788     if (xd->above_mi && is_inter_block(xd->above_mi))
1789       filter_ref = xd->above_mi->interp_filter;
1790     else if (xd->left_mi && is_inter_block(xd->left_mi))
1791       filter_ref = xd->left_mi->interp_filter;
1792   }
1793 
1794   // initialize mode decisions
1795   vp9_rd_cost_reset(&best_rdc);
1796   vp9_rd_cost_reset(rd_cost);
1797   mi->sb_type = bsize;
1798   mi->ref_frame[0] = NONE;
1799   mi->ref_frame[1] = NONE;
1800 
1801   mi->tx_size =
1802       VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[cm->tx_mode]);
1803 
1804   if (sf->short_circuit_flat_blocks || sf->limit_newmv_early_exit) {
1805 #if CONFIG_VP9_HIGHBITDEPTH
1806     if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
1807       x->source_variance = vp9_high_get_sby_perpixel_variance(
1808           cpi, &x->plane[0].src, bsize, xd->bd);
1809     else
1810 #endif  // CONFIG_VP9_HIGHBITDEPTH
1811       x->source_variance =
1812           vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1813 
1814     if (cpi->oxcf.content == VP9E_CONTENT_SCREEN && mi->segment_id > 0 &&
1815         x->zero_temp_sad_source && x->source_variance == 0) {
1816       mi->segment_id = 0;
1817       vp9_init_plane_quantizers(cpi, x);
1818     }
1819   }
1820 
1821 #if CONFIG_VP9_TEMPORAL_DENOISING
1822   if (cpi->oxcf.noise_sensitivity > 0) {
1823     if (cpi->use_svc) denoise_svc_pickmode = vp9_denoise_svc_non_key(cpi);
1824     if (cpi->denoiser.denoising_level > kDenLowLow && denoise_svc_pickmode)
1825       vp9_denoiser_reset_frame_stats(ctx);
1826   }
1827 #endif
1828 
1829   if (cpi->rc.frames_since_golden == 0 && gf_temporal_ref &&
1830       !cpi->rc.alt_ref_gf_group && !cpi->rc.last_frame_is_src_altref) {
1831     usable_ref_frame = LAST_FRAME;
1832   } else {
1833     usable_ref_frame = GOLDEN_FRAME;
1834   }
1835 
1836   if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
1837     if (cpi->rc.alt_ref_gf_group || cpi->rc.is_src_frame_alt_ref)
1838       usable_ref_frame = ALTREF_FRAME;
1839 
1840     if (cpi->rc.is_src_frame_alt_ref) {
1841       skip_ref_find_pred[LAST_FRAME] = 1;
1842       skip_ref_find_pred[GOLDEN_FRAME] = 1;
1843     }
1844     if (!cm->show_frame) {
1845       if (cpi->rc.frames_since_key == 1) {
1846         usable_ref_frame = LAST_FRAME;
1847         skip_ref_find_pred[GOLDEN_FRAME] = 1;
1848         skip_ref_find_pred[ALTREF_FRAME] = 1;
1849       }
1850     }
1851   }
1852 
1853   // For svc mode, on spatial_layer_id > 0: if the reference has different scale
1854   // constrain the inter mode to only test zero motion.
1855   if (cpi->use_svc && svc->force_zero_mode_spatial_ref &&
1856       svc->spatial_layer_id > 0 && !gf_temporal_ref) {
1857     if (cpi->ref_frame_flags & flag_list[LAST_FRAME]) {
1858       struct scale_factors *const sf = &cm->frame_refs[LAST_FRAME - 1].sf;
1859       if (vp9_is_scaled(sf)) {
1860         svc_force_zero_mode[LAST_FRAME - 1] = 1;
1861         inter_layer_ref = LAST_FRAME;
1862       }
1863     }
1864     if (cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) {
1865       struct scale_factors *const sf = &cm->frame_refs[GOLDEN_FRAME - 1].sf;
1866       if (vp9_is_scaled(sf)) {
1867         svc_force_zero_mode[GOLDEN_FRAME - 1] = 1;
1868         inter_layer_ref = GOLDEN_FRAME;
1869       }
1870     }
1871   }
1872 
1873   if (cpi->sf.short_circuit_low_temp_var) {
1874     force_skip_low_temp_var =
1875         get_force_skip_low_temp_var(&x->variance_low[0], mi_row, mi_col, bsize);
1876     // If force_skip_low_temp_var is set, and for short circuit mode = 1 and 3,
1877     // skip golden reference.
1878     if ((cpi->sf.short_circuit_low_temp_var == 1 ||
1879          cpi->sf.short_circuit_low_temp_var == 3) &&
1880         force_skip_low_temp_var) {
1881       usable_ref_frame = LAST_FRAME;
1882     }
1883   }
1884 
1885   if (sf->disable_golden_ref && (x->content_state_sb != kVeryHighSad ||
1886                                  cpi->rc.avg_frame_low_motion < 60))
1887     usable_ref_frame = LAST_FRAME;
1888 
1889   if (!((cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) &&
1890         !svc_force_zero_mode[GOLDEN_FRAME - 1] && !force_skip_low_temp_var))
1891     use_golden_nonzeromv = 0;
1892 
1893   if (cpi->oxcf.speed >= 8 && !cpi->use_svc &&
1894       ((cpi->rc.frames_since_golden + 1) < x->last_sb_high_content ||
1895        x->last_sb_high_content > 40 || cpi->rc.frames_since_golden > 120))
1896     usable_ref_frame = LAST_FRAME;
1897 
1898   // Compound prediction modes: (0,0) on LAST/GOLDEN and ARF.
1899   if (cm->reference_mode == REFERENCE_MODE_SELECT &&
1900       cpi->sf.use_compound_nonrd_pickmode && usable_ref_frame == ALTREF_FRAME)
1901     comp_modes = 2;
1902 
1903   // If the segment reference frame feature is enabled and it's set to GOLDEN
1904   // reference, then make sure we don't skip checking GOLDEN, this is to
1905   // prevent possibility of not picking any mode.
1906   if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME) &&
1907       get_segdata(seg, mi->segment_id, SEG_LVL_REF_FRAME) == GOLDEN_FRAME) {
1908     usable_ref_frame = GOLDEN_FRAME;
1909     skip_ref_find_pred[GOLDEN_FRAME] = 0;
1910     thresh_svc_skip_golden = 0;
1911   }
1912 
1913   for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) {
1914     // Skip find_predictor if the reference frame is not in the
1915     // ref_frame_flags (i.e., not used as a reference for this frame).
1916     skip_ref_find_pred[ref_frame] =
1917         !(cpi->ref_frame_flags & flag_list[ref_frame]);
1918     if (!skip_ref_find_pred[ref_frame]) {
1919       find_predictors(cpi, x, ref_frame, frame_mv, const_motion,
1920                       &ref_frame_skip_mask, flag_list, tile_data, mi_row,
1921                       mi_col, yv12_mb, bsize, force_skip_low_temp_var,
1922                       comp_modes > 0);
1923     }
1924   }
1925 
1926   if (cpi->use_svc || cpi->oxcf.speed <= 7 || bsize < BLOCK_32X32)
1927     x->sb_use_mv_part = 0;
1928 
1929   // Set the flag_svc_subpel to 1 for SVC if the lower spatial layer used
1930   // an averaging filter for downsampling (phase = 8). If so, we will test
1931   // a nonzero motion mode on the spatial reference.
1932   // The nonzero motion is half pixel shifted to left and top (-4, -4).
1933   if (cpi->use_svc && svc->spatial_layer_id > 0 &&
1934       svc_force_zero_mode[inter_layer_ref - 1] &&
1935       svc->downsample_filter_phase[svc->spatial_layer_id - 1] == 8 &&
1936       !gf_temporal_ref) {
1937     svc_mv_col = -4;
1938     svc_mv_row = -4;
1939     flag_svc_subpel = 1;
1940   }
1941 
1942   for (idx = 0; idx < num_inter_modes + comp_modes; ++idx) {
1943     int rate_mv = 0;
1944     int mode_rd_thresh;
1945     int mode_index;
1946     int i;
1947     int64_t this_sse;
1948     int is_skippable;
1949     int this_early_term = 0;
1950     int rd_computed = 0;
1951     int flag_preduv_computed[2] = { 0 };
1952     int inter_mv_mode = 0;
1953     int skip_this_mv = 0;
1954     int comp_pred = 0;
1955     int force_mv_inter_layer = 0;
1956     PREDICTION_MODE this_mode;
1957     second_ref_frame = NONE;
1958 
1959     if (idx < num_inter_modes) {
1960       this_mode = ref_mode_set[idx].pred_mode;
1961       ref_frame = ref_mode_set[idx].ref_frame;
1962 
1963       if (cpi->use_svc) {
1964         this_mode = ref_mode_set_svc[idx].pred_mode;
1965         ref_frame = ref_mode_set_svc[idx].ref_frame;
1966       }
1967     } else {
1968       // Add (0,0) compound modes.
1969       this_mode = ZEROMV;
1970       ref_frame = LAST_FRAME;
1971       if (idx == num_inter_modes + comp_modes - 1) ref_frame = GOLDEN_FRAME;
1972       second_ref_frame = ALTREF_FRAME;
1973       comp_pred = 1;
1974     }
1975 
1976     if (ref_frame > usable_ref_frame) continue;
1977     if (skip_ref_find_pred[ref_frame]) continue;
1978 
1979     if (svc->previous_frame_is_intra_only) {
1980       if (ref_frame != LAST_FRAME || frame_mv[this_mode][ref_frame].as_int != 0)
1981         continue;
1982     }
1983 
1984     // If the segment reference frame feature is enabled then do nothing if the
1985     // current ref frame is not allowed.
1986     if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME) &&
1987         get_segdata(seg, mi->segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
1988       continue;
1989 
1990     if (flag_svc_subpel && ref_frame == inter_layer_ref) {
1991       force_mv_inter_layer = 1;
1992       // Only test mode if NEARESTMV/NEARMV is (svc_mv_col, svc_mv_row),
1993       // otherwise set NEWMV to (svc_mv_col, svc_mv_row).
1994       if (this_mode == NEWMV) {
1995         frame_mv[this_mode][ref_frame].as_mv.col = svc_mv_col;
1996         frame_mv[this_mode][ref_frame].as_mv.row = svc_mv_row;
1997       } else if (frame_mv[this_mode][ref_frame].as_mv.col != svc_mv_col ||
1998                  frame_mv[this_mode][ref_frame].as_mv.row != svc_mv_row) {
1999         continue;
2000       }
2001     }
2002 
2003     if (comp_pred) {
2004       if (!cpi->allow_comp_inter_inter) continue;
2005       // Skip compound inter modes if ARF is not available.
2006       if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) continue;
2007       // Do not allow compound prediction if the segment level reference frame
2008       // feature is in use as in this case there can only be one reference.
2009       if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME)) continue;
2010     }
2011 
2012     // For SVC, skip the golden (spatial) reference search if sse of zeromv_last
2013     // is below threshold.
2014     if (cpi->use_svc && ref_frame == GOLDEN_FRAME &&
2015         sse_zeromv_normalized < thresh_svc_skip_golden)
2016       continue;
2017 
2018     if (!(cpi->ref_frame_flags & flag_list[ref_frame])) continue;
2019 
2020     if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
2021         (frame_mv[this_mode][ref_frame].as_int != 0 ||
2022          (cpi->oxcf.content == VP9E_CONTENT_SCREEN && !svc->spatial_layer_id &&
2023           !x->zero_temp_sad_source))) {
2024       continue;
2025     }
2026 
2027     if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode))) continue;
2028 
2029     if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
2030       if (cpi->rc.is_src_frame_alt_ref &&
2031           (ref_frame != ALTREF_FRAME ||
2032            frame_mv[this_mode][ref_frame].as_int != 0))
2033         continue;
2034 
2035       if (!cm->show_frame && ref_frame == ALTREF_FRAME &&
2036           frame_mv[this_mode][ref_frame].as_int != 0)
2037         continue;
2038 
2039       if (cpi->rc.alt_ref_gf_group && cm->show_frame &&
2040           cpi->rc.frames_since_golden > (cpi->rc.baseline_gf_interval >> 1) &&
2041           ref_frame == GOLDEN_FRAME &&
2042           frame_mv[this_mode][ref_frame].as_int != 0)
2043         continue;
2044 
2045       if (cpi->rc.alt_ref_gf_group && cm->show_frame &&
2046           cpi->rc.frames_since_golden > 0 &&
2047           cpi->rc.frames_since_golden < (cpi->rc.baseline_gf_interval >> 1) &&
2048           ref_frame == ALTREF_FRAME &&
2049           frame_mv[this_mode][ref_frame].as_int != 0)
2050         continue;
2051     }
2052 
2053     if (const_motion[ref_frame] && this_mode == NEARMV) continue;
2054 
2055     // Skip non-zeromv mode search for golden frame if force_skip_low_temp_var
2056     // is set. If nearestmv for golden frame is 0, zeromv mode will be skipped
2057     // later.
2058     if (!force_mv_inter_layer && force_skip_low_temp_var &&
2059         ref_frame == GOLDEN_FRAME &&
2060         frame_mv[this_mode][ref_frame].as_int != 0) {
2061       continue;
2062     }
2063 
2064     if (x->content_state_sb != kVeryHighSad &&
2065         (cpi->sf.short_circuit_low_temp_var >= 2 ||
2066          (cpi->sf.short_circuit_low_temp_var == 1 && bsize == BLOCK_64X64)) &&
2067         force_skip_low_temp_var && ref_frame == LAST_FRAME &&
2068         this_mode == NEWMV) {
2069       continue;
2070     }
2071 
2072     if (cpi->use_svc) {
2073       if (!force_mv_inter_layer && svc_force_zero_mode[ref_frame - 1] &&
2074           frame_mv[this_mode][ref_frame].as_int != 0)
2075         continue;
2076     }
2077 
2078     // Disable this drop out case if the ref frame segment level feature is
2079     // enabled for this segment. This is to prevent the possibility that we end
2080     // up unable to pick any mode.
2081     if (!segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME)) {
2082       if (sf->reference_masking &&
2083           !(frame_mv[this_mode][ref_frame].as_int == 0 &&
2084             ref_frame == LAST_FRAME)) {
2085         if (usable_ref_frame < ALTREF_FRAME) {
2086           if (!force_skip_low_temp_var && usable_ref_frame > LAST_FRAME) {
2087             i = (ref_frame == LAST_FRAME) ? GOLDEN_FRAME : LAST_FRAME;
2088             if ((cpi->ref_frame_flags & flag_list[i]))
2089               if (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[i] << 1))
2090                 ref_frame_skip_mask |= (1 << ref_frame);
2091           }
2092         } else if (!cpi->rc.is_src_frame_alt_ref &&
2093                    !(frame_mv[this_mode][ref_frame].as_int == 0 &&
2094                      ref_frame == ALTREF_FRAME)) {
2095           int ref1 = (ref_frame == GOLDEN_FRAME) ? LAST_FRAME : GOLDEN_FRAME;
2096           int ref2 = (ref_frame == ALTREF_FRAME) ? LAST_FRAME : ALTREF_FRAME;
2097           if (((cpi->ref_frame_flags & flag_list[ref1]) &&
2098                (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref1] << 1))) ||
2099               ((cpi->ref_frame_flags & flag_list[ref2]) &&
2100                (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref2] << 1))))
2101             ref_frame_skip_mask |= (1 << ref_frame);
2102         }
2103       }
2104       if (ref_frame_skip_mask & (1 << ref_frame)) continue;
2105     }
2106 
2107     // Select prediction reference frames.
2108     for (i = 0; i < MAX_MB_PLANE; i++) {
2109       xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
2110       if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
2111     }
2112 
2113     mi->ref_frame[0] = ref_frame;
2114     mi->ref_frame[1] = second_ref_frame;
2115     set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
2116 
2117     mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
2118     mode_rd_thresh = best_pickmode.best_mode_skip_txfm
2119                          ? rd_threshes[mode_index] << 1
2120                          : rd_threshes[mode_index];
2121 
2122     // Increase mode_rd_thresh value for GOLDEN_FRAME for improved encoding
2123     // speed with little/no subjective quality loss.
2124     if (cpi->sf.bias_golden && ref_frame == GOLDEN_FRAME &&
2125         cpi->rc.frames_since_golden > 4)
2126       mode_rd_thresh = mode_rd_thresh << 3;
2127 
2128     if ((cpi->sf.adaptive_rd_thresh_row_mt &&
2129          rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh,
2130                                     &rd_thresh_freq_fact[mode_index])) ||
2131         (!cpi->sf.adaptive_rd_thresh_row_mt &&
2132          rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
2133                              &rd_thresh_freq_fact[mode_index])))
2134       if (frame_mv[this_mode][ref_frame].as_int != 0) continue;
2135 
2136     if (this_mode == NEWMV && !force_mv_inter_layer) {
2137       if (search_new_mv(cpi, x, frame_mv, ref_frame, gf_temporal_ref, bsize,
2138                         mi_row, mi_col, best_pred_sad, &rate_mv, best_sse_sofar,
2139                         &best_rdc))
2140         continue;
2141     }
2142 
2143     // TODO(jianj): Skipping the testing of (duplicate) non-zero motion vector
2144     // causes some regression, leave it for duplicate zero-mv for now, until
2145     // regression issue is resolved.
2146     for (inter_mv_mode = NEARESTMV; inter_mv_mode <= NEWMV; inter_mv_mode++) {
2147       if (inter_mv_mode == this_mode || comp_pred) continue;
2148       if (mode_checked[inter_mv_mode][ref_frame] &&
2149           frame_mv[this_mode][ref_frame].as_int ==
2150               frame_mv[inter_mv_mode][ref_frame].as_int &&
2151           frame_mv[inter_mv_mode][ref_frame].as_int == 0) {
2152         skip_this_mv = 1;
2153         break;
2154       }
2155     }
2156 
2157     if (skip_this_mv) continue;
2158 
2159     // If use_golden_nonzeromv is false, NEWMV mode is skipped for golden, no
2160     // need to compute best_pred_sad which is only used to skip golden NEWMV.
2161     if (use_golden_nonzeromv && this_mode == NEWMV && ref_frame == LAST_FRAME &&
2162         frame_mv[NEWMV][LAST_FRAME].as_int != INVALID_MV) {
2163       const int pre_stride = xd->plane[0].pre[0].stride;
2164       const uint8_t *const pre_buf =
2165           xd->plane[0].pre[0].buf +
2166           (frame_mv[NEWMV][LAST_FRAME].as_mv.row >> 3) * pre_stride +
2167           (frame_mv[NEWMV][LAST_FRAME].as_mv.col >> 3);
2168       best_pred_sad = cpi->fn_ptr[bsize].sdf(
2169           x->plane[0].src.buf, x->plane[0].src.stride, pre_buf, pre_stride);
2170       x->pred_mv_sad[LAST_FRAME] = best_pred_sad;
2171     }
2172 
2173     if (this_mode != NEARESTMV && !comp_pred &&
2174         frame_mv[this_mode][ref_frame].as_int ==
2175             frame_mv[NEARESTMV][ref_frame].as_int)
2176       continue;
2177 
2178     mi->mode = this_mode;
2179     mi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
2180     mi->mv[1].as_int = 0;
2181 
2182     // Search for the best prediction filter type, when the resulting
2183     // motion vector is at sub-pixel accuracy level for luma component, i.e.,
2184     // the last three bits are all zeros.
2185     if (reuse_inter_pred) {
2186       if (!this_mode_pred) {
2187         this_mode_pred = &tmp[3];
2188       } else {
2189         this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
2190         pd->dst.buf = this_mode_pred->data;
2191         pd->dst.stride = bw;
2192       }
2193     }
2194 
2195     if ((this_mode == NEWMV || filter_ref == SWITCHABLE) &&
2196         pred_filter_search &&
2197         (ref_frame == LAST_FRAME ||
2198          (ref_frame == GOLDEN_FRAME && !force_mv_inter_layer &&
2199           (cpi->use_svc || cpi->oxcf.rc_mode == VPX_VBR))) &&
2200         (((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07) != 0)) {
2201       rd_computed = 1;
2202       search_filter_ref(cpi, x, &this_rdc, mi_row, mi_col, tmp, bsize,
2203                         reuse_inter_pred, &this_mode_pred, &var_y, &sse_y);
2204     } else {
2205       // For low motion content use x->sb_is_skin in addition to VeryHighSad
2206       // for setting large_block.
2207       const int large_block =
2208           (x->content_state_sb == kVeryHighSad ||
2209            (x->sb_is_skin && cpi->rc.avg_frame_low_motion > 70) ||
2210            cpi->oxcf.speed < 7)
2211               ? bsize > BLOCK_32X32
2212               : bsize >= BLOCK_32X32;
2213       mi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP : filter_ref;
2214 
2215       if (cpi->use_svc && ref_frame == GOLDEN_FRAME &&
2216           svc_force_zero_mode[ref_frame - 1])
2217         mi->interp_filter = filter_gf_svc;
2218 
2219       vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
2220 
2221       // For large partition blocks, extra testing is done.
2222       if (cpi->oxcf.rc_mode == VPX_CBR && large_block &&
2223           !cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id) &&
2224           cm->base_qindex) {
2225         model_rd_for_sb_y_large(cpi, bsize, x, xd, &this_rdc.rate,
2226                                 &this_rdc.dist, &var_y, &sse_y, mi_row, mi_col,
2227                                 &this_early_term, flag_preduv_computed);
2228       } else {
2229         rd_computed = 1;
2230         model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
2231                           &var_y, &sse_y);
2232       }
2233       // Save normalized sse (between current and last frame) for (0, 0) motion.
2234       if (cpi->use_svc && ref_frame == LAST_FRAME &&
2235           frame_mv[this_mode][ref_frame].as_int == 0) {
2236         sse_zeromv_normalized =
2237             sse_y >> (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
2238       }
2239       if (sse_y < best_sse_sofar) best_sse_sofar = sse_y;
2240     }
2241 
2242     if (!this_early_term) {
2243       this_sse = (int64_t)sse_y;
2244       block_yrd(cpi, x, &this_rdc, &is_skippable, &this_sse, bsize,
2245                 VPXMIN(mi->tx_size, TX_16X16), rd_computed);
2246 
2247       x->skip_txfm[0] = is_skippable;
2248       if (is_skippable) {
2249         this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2250       } else {
2251         if (RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist) <
2252             RDCOST(x->rdmult, x->rddiv, 0, this_sse)) {
2253           this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
2254         } else {
2255           this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2256           this_rdc.dist = this_sse;
2257           x->skip_txfm[0] = SKIP_TXFM_AC_DC;
2258         }
2259       }
2260 
2261       if (cm->interp_filter == SWITCHABLE) {
2262         if ((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07)
2263           this_rdc.rate += vp9_get_switchable_rate(cpi, xd);
2264       }
2265     } else {
2266       this_rdc.rate += cm->interp_filter == SWITCHABLE
2267                            ? vp9_get_switchable_rate(cpi, xd)
2268                            : 0;
2269       this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2270     }
2271 
2272     if (!this_early_term &&
2273         (x->color_sensitivity[0] || x->color_sensitivity[1])) {
2274       RD_COST rdc_uv;
2275       const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, &xd->plane[1]);
2276       if (x->color_sensitivity[0] && !flag_preduv_computed[0]) {
2277         vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 1);
2278         flag_preduv_computed[0] = 1;
2279       }
2280       if (x->color_sensitivity[1] && !flag_preduv_computed[1]) {
2281         vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 2);
2282         flag_preduv_computed[1] = 1;
2283       }
2284       model_rd_for_sb_uv(cpi, uv_bsize, x, xd, &rdc_uv, &var_y, &sse_y, 1, 2);
2285       this_rdc.rate += rdc_uv.rate;
2286       this_rdc.dist += rdc_uv.dist;
2287     }
2288 
2289     this_rdc.rate += rate_mv;
2290     this_rdc.rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2291                                          [INTER_OFFSET(this_mode)];
2292     // TODO(marpan): Add costing for compound mode.
2293     this_rdc.rate += ref_frame_cost[ref_frame];
2294     this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2295 
2296     // Bias against NEWMV that is very different from its neighbors, and bias
2297     // to small motion-lastref for noisy input.
2298     if (cpi->oxcf.rc_mode == VPX_CBR && cpi->oxcf.speed >= 5 &&
2299         cpi->oxcf.content != VP9E_CONTENT_SCREEN) {
2300       vp9_NEWMV_diff_bias(&cpi->noise_estimate, xd, this_mode, &this_rdc, bsize,
2301                           frame_mv[this_mode][ref_frame].as_mv.row,
2302                           frame_mv[this_mode][ref_frame].as_mv.col,
2303                           ref_frame == LAST_FRAME, x->lowvar_highsumdiff,
2304                           x->sb_is_skin);
2305     }
2306 
2307     // Skipping checking: test to see if this block can be reconstructed by
2308     // prediction only.
2309     if (cpi->allow_encode_breakout && !xd->lossless && !scene_change_detected) {
2310       encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame, this_mode,
2311                            var_y, sse_y, yv12_mb, &this_rdc.rate,
2312                            &this_rdc.dist, flag_preduv_computed);
2313       if (x->skip) {
2314         this_rdc.rate += rate_mv;
2315         this_rdc.rdcost =
2316             RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2317       }
2318     }
2319 
2320 #if CONFIG_VP9_TEMPORAL_DENOISING
2321     if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc_pickmode &&
2322         cpi->denoiser.denoising_level > kDenLowLow) {
2323       vp9_denoiser_update_frame_stats(mi, sse_y, this_mode, ctx);
2324       // Keep track of zero_last cost.
2325       if (ref_frame == LAST_FRAME && frame_mv[this_mode][ref_frame].as_int == 0)
2326         zero_last_cost_orig = this_rdc.rdcost;
2327     }
2328 #else
2329     (void)ctx;
2330 #endif
2331 
2332     mode_checked[this_mode][ref_frame] = 1;
2333 
2334     if (this_rdc.rdcost < best_rdc.rdcost || x->skip) {
2335       best_rdc = this_rdc;
2336       best_early_term = this_early_term;
2337       best_pickmode.best_mode = this_mode;
2338       best_pickmode.best_pred_filter = mi->interp_filter;
2339       best_pickmode.best_tx_size = mi->tx_size;
2340       best_pickmode.best_ref_frame = ref_frame;
2341       best_pickmode.best_mode_skip_txfm = x->skip_txfm[0];
2342       best_pickmode.best_second_ref_frame = second_ref_frame;
2343 
2344       if (reuse_inter_pred) {
2345         free_pred_buffer(best_pickmode.best_pred);
2346         best_pickmode.best_pred = this_mode_pred;
2347       }
2348     } else {
2349       if (reuse_inter_pred) free_pred_buffer(this_mode_pred);
2350     }
2351 
2352     if (x->skip) break;
2353 
2354     // If early termination flag is 1 and at least 2 modes are checked,
2355     // the mode search is terminated.
2356     if (best_early_term && idx > 0 && !scene_change_detected) {
2357       x->skip = 1;
2358       break;
2359     }
2360   }
2361 
2362   mi->mode = best_pickmode.best_mode;
2363   mi->interp_filter = best_pickmode.best_pred_filter;
2364   mi->tx_size = best_pickmode.best_tx_size;
2365   mi->ref_frame[0] = best_pickmode.best_ref_frame;
2366   mi->mv[0].as_int =
2367       frame_mv[best_pickmode.best_mode][best_pickmode.best_ref_frame].as_int;
2368   xd->mi[0]->bmi[0].as_mv[0].as_int = mi->mv[0].as_int;
2369   x->skip_txfm[0] = best_pickmode.best_mode_skip_txfm;
2370   mi->ref_frame[1] = best_pickmode.best_second_ref_frame;
2371 
2372   // For spatial enhancemanent layer: perform intra prediction only if base
2373   // layer is chosen as the reference. Always perform intra prediction if
2374   // LAST is the only reference, or is_key_frame is set, or on base
2375   // temporal layer.
2376   if (svc->spatial_layer_id && !gf_temporal_ref) {
2377     perform_intra_pred =
2378         svc->temporal_layer_id == 0 ||
2379         svc->layer_context[svc->temporal_layer_id].is_key_frame ||
2380         !(cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) ||
2381         (!svc->layer_context[svc->temporal_layer_id].is_key_frame &&
2382          svc_force_zero_mode[best_pickmode.best_ref_frame - 1]);
2383     inter_mode_thresh = (inter_mode_thresh << 1) + inter_mode_thresh;
2384   }
2385   if ((cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR &&
2386        cpi->rc.is_src_frame_alt_ref) ||
2387       svc->previous_frame_is_intra_only)
2388     perform_intra_pred = 0;
2389 
2390   // If the segment reference frame feature is enabled and set then
2391   // skip the intra prediction.
2392   if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME) &&
2393       get_segdata(seg, mi->segment_id, SEG_LVL_REF_FRAME) > 0)
2394     perform_intra_pred = 0;
2395 
2396   // Perform intra prediction search, if the best SAD is above a certain
2397   // threshold.
2398   if (best_rdc.rdcost == INT64_MAX ||
2399       (scene_change_detected && perform_intra_pred) ||
2400       ((!force_skip_low_temp_var || bsize < BLOCK_32X32 ||
2401         x->content_state_sb == kVeryHighSad) &&
2402        perform_intra_pred && !x->skip && best_rdc.rdcost > inter_mode_thresh &&
2403        bsize <= cpi->sf.max_intra_bsize && !x->skip_low_source_sad &&
2404        !x->lowvar_highsumdiff)) {
2405     struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
2406     int i;
2407     PRED_BUFFER *const best_pred = best_pickmode.best_pred;
2408     TX_SIZE intra_tx_size =
2409         VPXMIN(max_txsize_lookup[bsize],
2410                tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
2411     if (cpi->oxcf.content != VP9E_CONTENT_SCREEN && intra_tx_size > TX_16X16)
2412       intra_tx_size = TX_16X16;
2413 
2414     if (reuse_inter_pred && best_pred != NULL) {
2415       if (best_pred->data == orig_dst.buf) {
2416         this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
2417 #if CONFIG_VP9_HIGHBITDEPTH
2418         if (cm->use_highbitdepth)
2419           vpx_highbd_convolve_copy(
2420               CONVERT_TO_SHORTPTR(best_pred->data), best_pred->stride,
2421               CONVERT_TO_SHORTPTR(this_mode_pred->data), this_mode_pred->stride,
2422               NULL, 0, 0, 0, 0, bw, bh, xd->bd);
2423         else
2424           vpx_convolve_copy(best_pred->data, best_pred->stride,
2425                             this_mode_pred->data, this_mode_pred->stride, NULL,
2426                             0, 0, 0, 0, bw, bh);
2427 #else
2428         vpx_convolve_copy(best_pred->data, best_pred->stride,
2429                           this_mode_pred->data, this_mode_pred->stride, NULL, 0,
2430                           0, 0, 0, bw, bh);
2431 #endif  // CONFIG_VP9_HIGHBITDEPTH
2432         best_pickmode.best_pred = this_mode_pred;
2433       }
2434     }
2435     pd->dst = orig_dst;
2436 
2437     for (i = 0; i < 4; ++i) {
2438       const PREDICTION_MODE this_mode = intra_mode_list[i];
2439       THR_MODES mode_index = mode_idx[INTRA_FRAME][mode_offset(this_mode)];
2440       int mode_rd_thresh = rd_threshes[mode_index];
2441       if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
2442           this_mode != DC_PRED) {
2443         continue;
2444       }
2445 
2446       if (!((1 << this_mode) & cpi->sf.intra_y_mode_bsize_mask[bsize]))
2447         continue;
2448 
2449       if ((cpi->sf.adaptive_rd_thresh_row_mt &&
2450            rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh,
2451                                       &rd_thresh_freq_fact[mode_index])) ||
2452           (!cpi->sf.adaptive_rd_thresh_row_mt &&
2453            rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
2454                                &rd_thresh_freq_fact[mode_index])))
2455         continue;
2456 
2457       mi->mode = this_mode;
2458       mi->ref_frame[0] = INTRA_FRAME;
2459       this_rdc.dist = this_rdc.rate = 0;
2460       args.mode = this_mode;
2461       args.skippable = 1;
2462       args.rdc = &this_rdc;
2463       mi->tx_size = intra_tx_size;
2464       vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
2465                                              &args);
2466       // Check skip cost here since skippable is not set for for uv, this
2467       // mirrors the behavior used by inter
2468       if (args.skippable) {
2469         x->skip_txfm[0] = SKIP_TXFM_AC_DC;
2470         this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
2471       } else {
2472         x->skip_txfm[0] = SKIP_TXFM_NONE;
2473         this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
2474       }
2475       // Inter and intra RD will mismatch in scale for non-screen content.
2476       if (cpi->oxcf.content == VP9E_CONTENT_SCREEN) {
2477         if (x->color_sensitivity[0])
2478           vp9_foreach_transformed_block_in_plane(xd, bsize, 1,
2479                                                  estimate_block_intra, &args);
2480         if (x->color_sensitivity[1])
2481           vp9_foreach_transformed_block_in_plane(xd, bsize, 2,
2482                                                  estimate_block_intra, &args);
2483       }
2484       this_rdc.rate += cpi->mbmode_cost[this_mode];
2485       this_rdc.rate += ref_frame_cost[INTRA_FRAME];
2486       this_rdc.rate += intra_cost_penalty;
2487       this_rdc.rdcost =
2488           RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2489 
2490       if (this_rdc.rdcost < best_rdc.rdcost) {
2491         best_rdc = this_rdc;
2492         best_pickmode.best_mode = this_mode;
2493         best_pickmode.best_intra_tx_size = mi->tx_size;
2494         best_pickmode.best_ref_frame = INTRA_FRAME;
2495         best_pickmode.best_second_ref_frame = NONE;
2496         mi->uv_mode = this_mode;
2497         mi->mv[0].as_int = INVALID_MV;
2498         mi->mv[1].as_int = INVALID_MV;
2499         best_pickmode.best_mode_skip_txfm = x->skip_txfm[0];
2500       }
2501     }
2502 
2503     // Reset mb_mode_info to the best inter mode.
2504     if (best_pickmode.best_ref_frame != INTRA_FRAME) {
2505       mi->tx_size = best_pickmode.best_tx_size;
2506     } else {
2507       mi->tx_size = best_pickmode.best_intra_tx_size;
2508     }
2509   }
2510 
2511   pd->dst = orig_dst;
2512   mi->mode = best_pickmode.best_mode;
2513   mi->ref_frame[0] = best_pickmode.best_ref_frame;
2514   mi->ref_frame[1] = best_pickmode.best_second_ref_frame;
2515   x->skip_txfm[0] = best_pickmode.best_mode_skip_txfm;
2516 
2517   if (!is_inter_block(mi)) {
2518     mi->interp_filter = SWITCHABLE_FILTERS;
2519   }
2520 
2521   if (reuse_inter_pred && best_pickmode.best_pred != NULL) {
2522     PRED_BUFFER *const best_pred = best_pickmode.best_pred;
2523     if (best_pred->data != orig_dst.buf && is_inter_mode(mi->mode)) {
2524 #if CONFIG_VP9_HIGHBITDEPTH
2525       if (cm->use_highbitdepth)
2526         vpx_highbd_convolve_copy(
2527             CONVERT_TO_SHORTPTR(best_pred->data), best_pred->stride,
2528             CONVERT_TO_SHORTPTR(pd->dst.buf), pd->dst.stride, NULL, 0, 0, 0, 0,
2529             bw, bh, xd->bd);
2530       else
2531         vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2532                           pd->dst.stride, NULL, 0, 0, 0, 0, bw, bh);
2533 #else
2534       vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2535                         pd->dst.stride, NULL, 0, 0, 0, 0, bw, bh);
2536 #endif  // CONFIG_VP9_HIGHBITDEPTH
2537     }
2538   }
2539 
2540 #if CONFIG_VP9_TEMPORAL_DENOISING
2541   if (cpi->oxcf.noise_sensitivity > 0 && cpi->resize_pending == 0 &&
2542       denoise_svc_pickmode && cpi->denoiser.denoising_level > kDenLowLow &&
2543       cpi->denoiser.reset == 0) {
2544     VP9_DENOISER_DECISION decision = COPY_BLOCK;
2545     ctx->sb_skip_denoising = 0;
2546     // TODO(marpan): There is an issue with denoising when the
2547     // superblock partitioning scheme is based on the pickmode.
2548     // Remove this condition when the issue is resolved.
2549     if (x->sb_pickmode_part) ctx->sb_skip_denoising = 1;
2550     vp9_pickmode_ctx_den_update(&ctx_den, zero_last_cost_orig, ref_frame_cost,
2551                                 frame_mv, reuse_inter_pred, &best_pickmode);
2552     vp9_denoiser_denoise(cpi, x, mi_row, mi_col, bsize, ctx, &decision,
2553                          gf_temporal_ref);
2554     recheck_zeromv_after_denoising(cpi, mi, x, xd, decision, &ctx_den, yv12_mb,
2555                                    &best_rdc, bsize, mi_row, mi_col);
2556     best_pickmode.best_ref_frame = ctx_den.best_ref_frame;
2557   }
2558 #endif
2559 
2560   if (best_pickmode.best_ref_frame == ALTREF_FRAME ||
2561       best_pickmode.best_second_ref_frame == ALTREF_FRAME)
2562     x->arf_frame_usage++;
2563   else if (best_pickmode.best_ref_frame != INTRA_FRAME)
2564     x->lastgolden_frame_usage++;
2565 
2566   if (cpi->sf.adaptive_rd_thresh) {
2567     THR_MODES best_mode_idx =
2568         mode_idx[best_pickmode.best_ref_frame][mode_offset(mi->mode)];
2569 
2570     if (best_pickmode.best_ref_frame == INTRA_FRAME) {
2571       // Only consider the modes that are included in the intra_mode_list.
2572       int intra_modes = sizeof(intra_mode_list) / sizeof(PREDICTION_MODE);
2573       int i;
2574 
2575       // TODO(yunqingwang): Check intra mode mask and only update freq_fact
2576       // for those valid modes.
2577       for (i = 0; i < intra_modes; i++) {
2578         if (cpi->sf.adaptive_rd_thresh_row_mt)
2579           update_thresh_freq_fact_row_mt(cpi, tile_data, x->source_variance,
2580                                          thresh_freq_fact_idx, INTRA_FRAME,
2581                                          best_mode_idx, intra_mode_list[i]);
2582         else
2583           update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2584                                   INTRA_FRAME, best_mode_idx,
2585                                   intra_mode_list[i]);
2586       }
2587     } else {
2588       for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2589         PREDICTION_MODE this_mode;
2590         if (best_pickmode.best_ref_frame != ref_frame) continue;
2591         for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2592           if (cpi->sf.adaptive_rd_thresh_row_mt)
2593             update_thresh_freq_fact_row_mt(cpi, tile_data, x->source_variance,
2594                                            thresh_freq_fact_idx, ref_frame,
2595                                            best_mode_idx, this_mode);
2596           else
2597             update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2598                                     ref_frame, best_mode_idx, this_mode);
2599         }
2600       }
2601     }
2602   }
2603 
2604   *rd_cost = best_rdc;
2605 }
2606 
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)2607 void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x, int mi_row,
2608                                 int mi_col, RD_COST *rd_cost, BLOCK_SIZE bsize,
2609                                 PICK_MODE_CONTEXT *ctx) {
2610   VP9_COMMON *const cm = &cpi->common;
2611   SPEED_FEATURES *const sf = &cpi->sf;
2612   MACROBLOCKD *const xd = &x->e_mbd;
2613   MODE_INFO *const mi = xd->mi[0];
2614   MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2615   const struct segmentation *const seg = &cm->seg;
2616   MV_REFERENCE_FRAME ref_frame, second_ref_frame = NONE;
2617   MV_REFERENCE_FRAME best_ref_frame = NONE;
2618   unsigned char segment_id = mi->segment_id;
2619   struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2620   static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2621                                     VP9_ALT_FLAG };
2622   int64_t best_rd = INT64_MAX;
2623   b_mode_info bsi[MAX_REF_FRAMES][4];
2624   int ref_frame_skip_mask = 0;
2625   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
2626   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
2627   int idx, idy;
2628 
2629   x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
2630   ctx->pred_pixel_ready = 0;
2631 
2632   for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2633     const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2634     int_mv dummy_mv[2];
2635     x->pred_mv_sad[ref_frame] = INT_MAX;
2636 
2637     if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
2638       int_mv *const candidates = mbmi_ext->ref_mvs[ref_frame];
2639       const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2640       vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf,
2641                            sf);
2642       vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
2643                        mbmi_ext->mode_context);
2644 
2645       vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2646                             &dummy_mv[0], &dummy_mv[1]);
2647     } else {
2648       ref_frame_skip_mask |= (1 << ref_frame);
2649     }
2650   }
2651 
2652   mi->sb_type = bsize;
2653   mi->tx_size = TX_4X4;
2654   mi->uv_mode = DC_PRED;
2655   mi->ref_frame[0] = LAST_FRAME;
2656   mi->ref_frame[1] = NONE;
2657   mi->interp_filter =
2658       cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
2659 
2660   for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2661     int64_t this_rd = 0;
2662     int plane;
2663 
2664     if (ref_frame_skip_mask & (1 << ref_frame)) continue;
2665 
2666 #if CONFIG_BETTER_HW_COMPATIBILITY
2667     if ((bsize == BLOCK_8X4 || bsize == BLOCK_4X8) && ref_frame > INTRA_FRAME &&
2668         vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2669       continue;
2670 #endif
2671 
2672     // TODO(jingning, agrange): Scaling reference frame not supported for
2673     // sub8x8 blocks. Is this supported now?
2674     if (ref_frame > INTRA_FRAME &&
2675         vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2676       continue;
2677 
2678     // If the segment reference frame feature is enabled....
2679     // then do nothing if the current ref frame is not allowed..
2680     if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
2681         get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
2682       continue;
2683 
2684     mi->ref_frame[0] = ref_frame;
2685     x->skip = 0;
2686     set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
2687 
2688     // Select prediction reference frames.
2689     for (plane = 0; plane < MAX_MB_PLANE; plane++)
2690       xd->plane[plane].pre[0] = yv12_mb[ref_frame][plane];
2691 
2692     for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2693       for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2694         int_mv b_mv[MB_MODE_COUNT];
2695         int64_t b_best_rd = INT64_MAX;
2696         const int i = idy * 2 + idx;
2697         PREDICTION_MODE this_mode;
2698         RD_COST this_rdc;
2699         unsigned int var_y, sse_y;
2700 
2701         struct macroblock_plane *p = &x->plane[0];
2702         struct macroblockd_plane *pd = &xd->plane[0];
2703 
2704         const struct buf_2d orig_src = p->src;
2705         const struct buf_2d orig_dst = pd->dst;
2706         struct buf_2d orig_pre[2];
2707         memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre));
2708 
2709         // set buffer pointers for sub8x8 motion search.
2710         p->src.buf =
2711             &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
2712         pd->dst.buf =
2713             &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
2714         pd->pre[0].buf =
2715             &pd->pre[0]
2716                  .buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
2717 
2718         b_mv[ZEROMV].as_int = 0;
2719         b_mv[NEWMV].as_int = INVALID_MV;
2720         vp9_append_sub8x8_mvs_for_idx(cm, xd, i, 0, mi_row, mi_col,
2721                                       &b_mv[NEARESTMV], &b_mv[NEARMV],
2722                                       mbmi_ext->mode_context);
2723 
2724         for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2725           int b_rate = 0;
2726           xd->mi[0]->bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int;
2727 
2728           if (this_mode == NEWMV) {
2729             const int step_param = cpi->sf.mv.fullpel_search_step_param;
2730             MV mvp_full;
2731             MV tmp_mv;
2732             int cost_list[5];
2733             const MvLimits tmp_mv_limits = x->mv_limits;
2734             uint32_t dummy_dist;
2735 
2736             if (i == 0) {
2737               mvp_full.row = b_mv[NEARESTMV].as_mv.row >> 3;
2738               mvp_full.col = b_mv[NEARESTMV].as_mv.col >> 3;
2739             } else {
2740               mvp_full.row = xd->mi[0]->bmi[0].as_mv[0].as_mv.row >> 3;
2741               mvp_full.col = xd->mi[0]->bmi[0].as_mv[0].as_mv.col >> 3;
2742             }
2743 
2744             vp9_set_mv_search_range(&x->mv_limits,
2745                                     &mbmi_ext->ref_mvs[ref_frame][0].as_mv);
2746 
2747             vp9_full_pixel_search(
2748                 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method,
2749                 x->sadperbit4, cond_cost_list(cpi, cost_list),
2750                 &mbmi_ext->ref_mvs[ref_frame][0].as_mv, &tmp_mv, INT_MAX, 0);
2751 
2752             x->mv_limits = tmp_mv_limits;
2753 
2754             // calculate the bit cost on motion vector
2755             mvp_full.row = tmp_mv.row * 8;
2756             mvp_full.col = tmp_mv.col * 8;
2757 
2758             b_rate += vp9_mv_bit_cost(
2759                 &mvp_full, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2760                 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2761 
2762             b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2763                                           [INTER_OFFSET(NEWMV)];
2764             if (RDCOST(x->rdmult, x->rddiv, b_rate, 0) > b_best_rd) continue;
2765 
2766             cpi->find_fractional_mv_step(
2767                 x, &tmp_mv, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2768                 cpi->common.allow_high_precision_mv, x->errorperbit,
2769                 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
2770                 cpi->sf.mv.subpel_search_level, cond_cost_list(cpi, cost_list),
2771                 x->nmvjointcost, x->mvcost, &dummy_dist,
2772                 &x->pred_sse[ref_frame], NULL, 0, 0,
2773                 cpi->sf.use_accurate_subpel_search);
2774 
2775             xd->mi[0]->bmi[i].as_mv[0].as_mv = tmp_mv;
2776           } else {
2777             b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2778                                           [INTER_OFFSET(this_mode)];
2779           }
2780 
2781 #if CONFIG_VP9_HIGHBITDEPTH
2782           if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2783             vp9_highbd_build_inter_predictor(
2784                 CONVERT_TO_SHORTPTR(pd->pre[0].buf), pd->pre[0].stride,
2785                 CONVERT_TO_SHORTPTR(pd->dst.buf), pd->dst.stride,
2786                 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2787                 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2788                 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2789                 mi_col * MI_SIZE + 4 * (i & 0x01),
2790                 mi_row * MI_SIZE + 4 * (i >> 1), xd->bd);
2791           } else {
2792 #endif
2793             vp9_build_inter_predictor(
2794                 pd->pre[0].buf, pd->pre[0].stride, pd->dst.buf, pd->dst.stride,
2795                 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2796                 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2797                 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2798                 mi_col * MI_SIZE + 4 * (i & 0x01),
2799                 mi_row * MI_SIZE + 4 * (i >> 1));
2800 
2801 #if CONFIG_VP9_HIGHBITDEPTH
2802           }
2803 #endif
2804 
2805           model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
2806                             &var_y, &sse_y);
2807 
2808           this_rdc.rate += b_rate;
2809           this_rdc.rdcost =
2810               RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2811           if (this_rdc.rdcost < b_best_rd) {
2812             b_best_rd = this_rdc.rdcost;
2813             bsi[ref_frame][i].as_mode = this_mode;
2814             bsi[ref_frame][i].as_mv[0].as_mv = xd->mi[0]->bmi[i].as_mv[0].as_mv;
2815           }
2816         }  // mode search
2817 
2818         // restore source and prediction buffer pointers.
2819         p->src = orig_src;
2820         pd->pre[0] = orig_pre[0];
2821         pd->dst = orig_dst;
2822         this_rd += b_best_rd;
2823 
2824         xd->mi[0]->bmi[i] = bsi[ref_frame][i];
2825         if (num_4x4_blocks_wide > 1) xd->mi[0]->bmi[i + 1] = xd->mi[0]->bmi[i];
2826         if (num_4x4_blocks_high > 1) xd->mi[0]->bmi[i + 2] = xd->mi[0]->bmi[i];
2827       }
2828     }  // loop through sub8x8 blocks
2829 
2830     if (this_rd < best_rd) {
2831       best_rd = this_rd;
2832       best_ref_frame = ref_frame;
2833     }
2834   }  // reference frames
2835 
2836   mi->tx_size = TX_4X4;
2837   mi->ref_frame[0] = best_ref_frame;
2838   for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2839     for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2840       const int block = idy * 2 + idx;
2841       xd->mi[0]->bmi[block] = bsi[best_ref_frame][block];
2842       if (num_4x4_blocks_wide > 1)
2843         xd->mi[0]->bmi[block + 1] = bsi[best_ref_frame][block];
2844       if (num_4x4_blocks_high > 1)
2845         xd->mi[0]->bmi[block + 2] = bsi[best_ref_frame][block];
2846     }
2847   }
2848   mi->mode = xd->mi[0]->bmi[3].as_mode;
2849   ctx->mic = *(xd->mi[0]);
2850   ctx->mbmi_ext = *x->mbmi_ext;
2851   ctx->skip_txfm[0] = SKIP_TXFM_NONE;
2852   ctx->skip = 0;
2853   // Dummy assignment for speed -5. No effect in speed -6.
2854   rd_cost->rdcost = best_rd;
2855 }
2856