• 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 static const int pos_shift_16x16[4][4] = {
45   { 9, 10, 13, 14 }, { 11, 12, 15, 16 }, { 17, 18, 21, 22 }, { 19, 20, 23, 24 }
46 };
47 
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)48 static int mv_refs_rt(VP9_COMP *cpi, const VP9_COMMON *cm, const MACROBLOCK *x,
49                       const MACROBLOCKD *xd, const TileInfo *const tile,
50                       MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
51                       int_mv *mv_ref_list, int_mv *base_mv, int mi_row,
52                       int mi_col, int use_base_mv) {
53   const int *ref_sign_bias = cm->ref_frame_sign_bias;
54   int i, refmv_count = 0;
55 
56   const POSITION *const mv_ref_search = mv_ref_blocks[mi->sb_type];
57 
58   int different_ref_found = 0;
59   int context_counter = 0;
60   int const_motion = 0;
61 
62   // Blank the reference vector list
63   memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
64 
65   // The nearest 2 blocks are treated differently
66   // if the size < 8x8 we get the mv from the bmi substructure,
67   // and we also need to keep a mode count.
68   for (i = 0; i < 2; ++i) {
69     const POSITION *const mv_ref = &mv_ref_search[i];
70     if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
71       const MODE_INFO *const candidate_mi =
72           xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
73       // Keep counts for entropy encoding.
74       context_counter += mode_2_counter[candidate_mi->mode];
75       different_ref_found = 1;
76 
77       if (candidate_mi->ref_frame[0] == ref_frame)
78         ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1),
79                         refmv_count, mv_ref_list, Done);
80     }
81   }
82 
83   const_motion = 1;
84 
85   // Check the rest of the neighbors in much the same way
86   // as before except we don't need to keep track of sub blocks or
87   // mode counts.
88   for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) {
89     const POSITION *const mv_ref = &mv_ref_search[i];
90     if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
91       const MODE_INFO *const candidate_mi =
92           xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
93       different_ref_found = 1;
94 
95       if (candidate_mi->ref_frame[0] == ref_frame)
96         ADD_MV_REF_LIST(candidate_mi->mv[0], refmv_count, mv_ref_list, Done);
97     }
98   }
99 
100   // Since we couldn't find 2 mvs from the same reference frame
101   // go back through the neighbors and find motion vectors from
102   // different reference frames.
103   if (different_ref_found && !refmv_count) {
104     for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
105       const POSITION *mv_ref = &mv_ref_search[i];
106       if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
107         const MODE_INFO *const candidate_mi =
108             xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
109 
110         // If the candidate is INTRA we don't want to consider its mv.
111         IF_DIFF_REF_FRAME_ADD_MV(candidate_mi, ref_frame, ref_sign_bias,
112                                  refmv_count, mv_ref_list, Done);
113       }
114     }
115   }
116   if (use_base_mv &&
117       !cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
118       ref_frame == LAST_FRAME) {
119     // Get base layer mv.
120     MV_REF *candidate =
121         &cm->prev_frame
122              ->mvs[(mi_col >> 1) + (mi_row >> 1) * (cm->mi_cols >> 1)];
123     if (candidate->mv[0].as_int != INVALID_MV) {
124       base_mv->as_mv.row = (candidate->mv[0].as_mv.row * 2);
125       base_mv->as_mv.col = (candidate->mv[0].as_mv.col * 2);
126       clamp_mv_ref(&base_mv->as_mv, xd);
127     } else {
128       base_mv->as_int = INVALID_MV;
129     }
130   }
131 
132 Done:
133 
134   x->mbmi_ext->mode_context[ref_frame] = counter_to_context[context_counter];
135 
136   // Clamp vectors
137   for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
138     clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
139 
140   return const_motion;
141 }
142 
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)143 static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
144                                   BLOCK_SIZE bsize, int mi_row, int mi_col,
145                                   int_mv *tmp_mv, int *rate_mv,
146                                   int64_t best_rd_sofar, int use_base_mv) {
147   MACROBLOCKD *xd = &x->e_mbd;
148   MODE_INFO *mi = xd->mi[0];
149   struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0 } };
150   const int step_param = cpi->sf.mv.fullpel_search_step_param;
151   const int sadpb = x->sadperbit16;
152   MV mvp_full;
153   const int ref = mi->ref_frame[0];
154   const MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
155   MV center_mv;
156   uint32_t dis;
157   int rate_mode;
158   const MvLimits tmp_mv_limits = x->mv_limits;
159   int rv = 0;
160   int cost_list[5];
161   const YV12_BUFFER_CONFIG *scaled_ref_frame =
162       vp9_get_scaled_ref_frame(cpi, ref);
163   if (scaled_ref_frame) {
164     int i;
165     // Swap out the reference frame for a version that's been scaled to
166     // match the resolution of the current frame, allowing the existing
167     // motion search code to be used without additional modifications.
168     for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
169     vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
170   }
171   vp9_set_mv_search_range(&x->mv_limits, &ref_mv);
172 
173   // Limit motion vector for large lightning change.
174   if (cpi->oxcf.speed > 5 && x->lowvar_highsumdiff) {
175     x->mv_limits.col_min = VPXMAX(x->mv_limits.col_min, -10);
176     x->mv_limits.row_min = VPXMAX(x->mv_limits.row_min, -10);
177     x->mv_limits.col_max = VPXMIN(x->mv_limits.col_max, 10);
178     x->mv_limits.row_max = VPXMIN(x->mv_limits.row_max, 10);
179   }
180 
181   assert(x->mv_best_ref_index[ref] <= 2);
182   if (x->mv_best_ref_index[ref] < 2)
183     mvp_full = x->mbmi_ext->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
184   else
185     mvp_full = x->pred_mv[ref];
186 
187   mvp_full.col >>= 3;
188   mvp_full.row >>= 3;
189 
190   if (!use_base_mv)
191     center_mv = ref_mv;
192   else
193     center_mv = tmp_mv->as_mv;
194 
195   vp9_full_pixel_search(
196       cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method, sadpb,
197       cond_cost_list(cpi, cost_list), &center_mv, &tmp_mv->as_mv, INT_MAX, 0);
198 
199   x->mv_limits = tmp_mv_limits;
200 
201   // calculate the bit cost on motion vector
202   mvp_full.row = tmp_mv->as_mv.row * 8;
203   mvp_full.col = tmp_mv->as_mv.col * 8;
204 
205   *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv, x->nmvjointcost, x->mvcost,
206                              MV_COST_WEIGHT);
207 
208   rate_mode =
209       cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref]][INTER_OFFSET(NEWMV)];
210   rv =
211       !(RDCOST(x->rdmult, x->rddiv, (*rate_mv + rate_mode), 0) > best_rd_sofar);
212 
213   if (rv) {
214     const int subpel_force_stop = cpi->sf.mv.subpel_force_stop;
215     cpi->find_fractional_mv_step(
216         x, &tmp_mv->as_mv, &ref_mv, cpi->common.allow_high_precision_mv,
217         x->errorperbit, &cpi->fn_ptr[bsize], subpel_force_stop,
218         cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
219         x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, 0, 0);
220     *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmvjointcost,
221                                x->mvcost, MV_COST_WEIGHT);
222   }
223 
224   if (scaled_ref_frame) {
225     int i;
226     for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
227   }
228   return rv;
229 }
230 
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)231 static void block_variance(const uint8_t *src, int src_stride,
232                            const uint8_t *ref, int ref_stride, int w, int h,
233                            unsigned int *sse, int *sum, int block_size,
234 #if CONFIG_VP9_HIGHBITDEPTH
235                            int use_highbitdepth, vpx_bit_depth_t bd,
236 #endif
237                            uint32_t *sse8x8, int *sum8x8, uint32_t *var8x8) {
238   int i, j, k = 0;
239 
240   *sse = 0;
241   *sum = 0;
242 
243   for (i = 0; i < h; i += block_size) {
244     for (j = 0; j < w; j += block_size) {
245 #if CONFIG_VP9_HIGHBITDEPTH
246       if (use_highbitdepth) {
247         switch (bd) {
248           case VPX_BITS_8:
249             vpx_highbd_8_get8x8var(src + src_stride * i + j, src_stride,
250                                    ref + ref_stride * i + j, ref_stride,
251                                    &sse8x8[k], &sum8x8[k]);
252             break;
253           case VPX_BITS_10:
254             vpx_highbd_10_get8x8var(src + src_stride * i + j, src_stride,
255                                     ref + ref_stride * i + j, ref_stride,
256                                     &sse8x8[k], &sum8x8[k]);
257             break;
258           case VPX_BITS_12:
259             vpx_highbd_12_get8x8var(src + src_stride * i + j, src_stride,
260                                     ref + ref_stride * i + j, ref_stride,
261                                     &sse8x8[k], &sum8x8[k]);
262             break;
263         }
264       } else {
265         vpx_get8x8var(src + src_stride * i + j, src_stride,
266                       ref + ref_stride * i + j, ref_stride, &sse8x8[k],
267                       &sum8x8[k]);
268       }
269 #else
270       vpx_get8x8var(src + src_stride * i + j, src_stride,
271                     ref + ref_stride * i + j, ref_stride, &sse8x8[k],
272                     &sum8x8[k]);
273 #endif
274       *sse += sse8x8[k];
275       *sum += sum8x8[k];
276       var8x8[k] = sse8x8[k] - (uint32_t)(((int64_t)sum8x8[k] * sum8x8[k]) >> 6);
277       k++;
278     }
279   }
280 }
281 
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)282 static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
283                                unsigned int *sse_i, int *sum_i,
284                                unsigned int *var_o, unsigned int *sse_o,
285                                int *sum_o) {
286   const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
287   const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
288   const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
289   int i, j, k = 0;
290 
291   for (i = 0; i < nh; i += 2) {
292     for (j = 0; j < nw; j += 2) {
293       sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] +
294                  sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1];
295       sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] +
296                  sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1];
297       var_o[k] = sse_o[k] - (uint32_t)(((int64_t)sum_o[k] * sum_o[k]) >>
298                                        (b_width_log2_lookup[unit_size] +
299                                         b_height_log2_lookup[unit_size] + 6));
300       k++;
301     }
302   }
303 }
304 
305 // 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)306 static int ac_thr_factor(const int speed, const int width, const int height,
307                          const int norm_sum) {
308   if (speed >= 8 && norm_sum < 5) {
309     if (width <= 640 && height <= 480)
310       return 4;
311     else
312       return 2;
313   }
314   return 1;
315 }
316 
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)317 static void model_rd_for_sb_y_large(VP9_COMP *cpi, BLOCK_SIZE bsize,
318                                     MACROBLOCK *x, MACROBLOCKD *xd,
319                                     int *out_rate_sum, int64_t *out_dist_sum,
320                                     unsigned int *var_y, unsigned int *sse_y,
321                                     int mi_row, int mi_col, int *early_term) {
322   // Note our transform coeffs are 8 times an orthogonal transform.
323   // Hence quantizer step is also 8 times. To get effective quantizer
324   // we need to divide by 8 before sending to modeling function.
325   unsigned int sse;
326   int rate;
327   int64_t dist;
328   struct macroblock_plane *const p = &x->plane[0];
329   struct macroblockd_plane *const pd = &xd->plane[0];
330   const uint32_t dc_quant = pd->dequant[0];
331   const uint32_t ac_quant = pd->dequant[1];
332   const int64_t dc_thr = dc_quant * dc_quant >> 6;
333   int64_t ac_thr = ac_quant * ac_quant >> 6;
334   unsigned int var;
335   int sum;
336   int skip_dc = 0;
337 
338   const int bw = b_width_log2_lookup[bsize];
339   const int bh = b_height_log2_lookup[bsize];
340   const int num8x8 = 1 << (bw + bh - 2);
341   unsigned int sse8x8[64] = { 0 };
342   int sum8x8[64] = { 0 };
343   unsigned int var8x8[64] = { 0 };
344   TX_SIZE tx_size;
345   int i, k;
346 #if CONFIG_VP9_HIGHBITDEPTH
347   const vpx_bit_depth_t bd = cpi->common.bit_depth;
348 #endif
349   // Calculate variance for whole partition, and also save 8x8 blocks' variance
350   // to be used in following transform skipping test.
351   block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
352                  4 << bw, 4 << bh, &sse, &sum, 8,
353 #if CONFIG_VP9_HIGHBITDEPTH
354                  cpi->common.use_highbitdepth, bd,
355 #endif
356                  sse8x8, sum8x8, var8x8);
357   var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
358 
359   *var_y = var;
360   *sse_y = sse;
361 
362 #if CONFIG_VP9_TEMPORAL_DENOISING
363   if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
364       cpi->oxcf.speed > 5)
365     ac_thr = vp9_scale_acskip_thresh(ac_thr, cpi->denoiser.denoising_level,
366                                      (abs(sum) >> (bw + bh)),
367                                      cpi->svc.temporal_layer_id);
368   else
369     ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
370                             cpi->common.height, abs(sum) >> (bw + bh));
371 #else
372   ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
373                           cpi->common.height, abs(sum) >> (bw + bh));
374 #endif
375 
376   if (cpi->common.tx_mode == TX_MODE_SELECT) {
377     if (sse > (var << 2))
378       tx_size = VPXMIN(max_txsize_lookup[bsize],
379                        tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
380     else
381       tx_size = TX_8X8;
382 
383     if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
384         cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
385       tx_size = TX_8X8;
386     else if (tx_size > TX_16X16)
387       tx_size = TX_16X16;
388   } else {
389     tx_size = VPXMIN(max_txsize_lookup[bsize],
390                      tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
391   }
392 
393   assert(tx_size >= TX_8X8);
394   xd->mi[0]->tx_size = tx_size;
395 
396   // Evaluate if the partition block is a skippable block in Y plane.
397   {
398     unsigned int sse16x16[16] = { 0 };
399     int sum16x16[16] = { 0 };
400     unsigned int var16x16[16] = { 0 };
401     const int num16x16 = num8x8 >> 2;
402 
403     unsigned int sse32x32[4] = { 0 };
404     int sum32x32[4] = { 0 };
405     unsigned int var32x32[4] = { 0 };
406     const int num32x32 = num8x8 >> 4;
407 
408     int ac_test = 1;
409     int dc_test = 1;
410     const int num = (tx_size == TX_8X8)
411                         ? num8x8
412                         : ((tx_size == TX_16X16) ? num16x16 : num32x32);
413     const unsigned int *sse_tx =
414         (tx_size == TX_8X8) ? sse8x8
415                             : ((tx_size == TX_16X16) ? sse16x16 : sse32x32);
416     const unsigned int *var_tx =
417         (tx_size == TX_8X8) ? var8x8
418                             : ((tx_size == TX_16X16) ? var16x16 : var32x32);
419 
420     // Calculate variance if tx_size > TX_8X8
421     if (tx_size >= TX_16X16)
422       calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
423                          sum16x16);
424     if (tx_size == TX_32X32)
425       calculate_variance(bw, bh, TX_16X16, sse16x16, sum16x16, var32x32,
426                          sse32x32, sum32x32);
427 
428     // Skipping test
429     x->skip_txfm[0] = SKIP_TXFM_NONE;
430     for (k = 0; k < num; k++)
431       // Check if all ac coefficients can be quantized to zero.
432       if (!(var_tx[k] < ac_thr || var == 0)) {
433         ac_test = 0;
434         break;
435       }
436 
437     for (k = 0; k < num; k++)
438       // Check if dc coefficient can be quantized to zero.
439       if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
440         dc_test = 0;
441         break;
442       }
443 
444     if (ac_test) {
445       x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
446 
447       if (dc_test) x->skip_txfm[0] = SKIP_TXFM_AC_DC;
448     } else if (dc_test) {
449       skip_dc = 1;
450     }
451   }
452 
453   if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
454     int skip_uv[2] = { 0 };
455     unsigned int var_uv[2];
456     unsigned int sse_uv[2];
457 
458     *out_rate_sum = 0;
459     *out_dist_sum = sse << 4;
460 
461     // Transform skipping test in UV planes.
462     for (i = 1; i <= 2; i++) {
463       if (cpi->oxcf.speed < 8 || x->color_sensitivity[i - 1]) {
464         struct macroblock_plane *const p = &x->plane[i];
465         struct macroblockd_plane *const pd = &xd->plane[i];
466         const TX_SIZE uv_tx_size = get_uv_tx_size(xd->mi[0], pd);
467         const BLOCK_SIZE unit_size = txsize_to_bsize[uv_tx_size];
468         const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, pd);
469         const int uv_bw = b_width_log2_lookup[uv_bsize];
470         const int uv_bh = b_height_log2_lookup[uv_bsize];
471         const int sf = (uv_bw - b_width_log2_lookup[unit_size]) +
472                        (uv_bh - b_height_log2_lookup[unit_size]);
473         const uint32_t uv_dc_thr = pd->dequant[0] * pd->dequant[0] >> (6 - sf);
474         const uint32_t uv_ac_thr = pd->dequant[1] * pd->dequant[1] >> (6 - sf);
475         int j = i - 1;
476 
477         vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, i);
478         var_uv[j] = cpi->fn_ptr[uv_bsize].vf(
479             p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, &sse_uv[j]);
480 
481         if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
482             (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
483           skip_uv[j] = 1;
484         else
485           break;
486       } else {
487         skip_uv[i - 1] = 1;
488       }
489     }
490 
491     // If the transform in YUV planes are skippable, the mode search checks
492     // fewer inter modes and doesn't check intra modes.
493     if (skip_uv[0] & skip_uv[1]) {
494       *early_term = 1;
495     }
496     return;
497   }
498 
499   if (!skip_dc) {
500 #if CONFIG_VP9_HIGHBITDEPTH
501     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
502                                  dc_quant >> (xd->bd - 5), &rate, &dist);
503 #else
504     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
505                                  dc_quant >> 3, &rate, &dist);
506 #endif  // CONFIG_VP9_HIGHBITDEPTH
507   }
508 
509   if (!skip_dc) {
510     *out_rate_sum = rate >> 1;
511     *out_dist_sum = dist << 3;
512   } else {
513     *out_rate_sum = 0;
514     *out_dist_sum = (sse - var) << 4;
515   }
516 
517 #if CONFIG_VP9_HIGHBITDEPTH
518   vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
519                                ac_quant >> (xd->bd - 5), &rate, &dist);
520 #else
521   vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
522                                &rate, &dist);
523 #endif  // CONFIG_VP9_HIGHBITDEPTH
524 
525   *out_rate_sum += rate;
526   *out_dist_sum += dist << 4;
527 }
528 
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)529 static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
530                               MACROBLOCKD *xd, int *out_rate_sum,
531                               int64_t *out_dist_sum, unsigned int *var_y,
532                               unsigned int *sse_y) {
533   // Note our transform coeffs are 8 times an orthogonal transform.
534   // Hence quantizer step is also 8 times. To get effective quantizer
535   // we need to divide by 8 before sending to modeling function.
536   unsigned int sse;
537   int rate;
538   int64_t dist;
539   struct macroblock_plane *const p = &x->plane[0];
540   struct macroblockd_plane *const pd = &xd->plane[0];
541   const int64_t dc_thr = p->quant_thred[0] >> 6;
542   const int64_t ac_thr = p->quant_thred[1] >> 6;
543   const uint32_t dc_quant = pd->dequant[0];
544   const uint32_t ac_quant = pd->dequant[1];
545   unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
546                                            pd->dst.buf, pd->dst.stride, &sse);
547   int skip_dc = 0;
548 
549   *var_y = var;
550   *sse_y = sse;
551 
552   if (cpi->common.tx_mode == TX_MODE_SELECT) {
553     if (sse > (var << 2))
554       xd->mi[0]->tx_size =
555           VPXMIN(max_txsize_lookup[bsize],
556                  tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
557     else
558       xd->mi[0]->tx_size = TX_8X8;
559 
560     if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
561         cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
562       xd->mi[0]->tx_size = TX_8X8;
563     else if (xd->mi[0]->tx_size > TX_16X16)
564       xd->mi[0]->tx_size = TX_16X16;
565   } else {
566     xd->mi[0]->tx_size =
567         VPXMIN(max_txsize_lookup[bsize],
568                tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
569   }
570 
571   // Evaluate if the partition block is a skippable block in Y plane.
572   {
573     const BLOCK_SIZE unit_size = txsize_to_bsize[xd->mi[0]->tx_size];
574     const unsigned int num_blk_log2 =
575         (b_width_log2_lookup[bsize] - b_width_log2_lookup[unit_size]) +
576         (b_height_log2_lookup[bsize] - b_height_log2_lookup[unit_size]);
577     const unsigned int sse_tx = sse >> num_blk_log2;
578     const unsigned int var_tx = var >> num_blk_log2;
579 
580     x->skip_txfm[0] = SKIP_TXFM_NONE;
581     // Check if all ac coefficients can be quantized to zero.
582     if (var_tx < ac_thr || var == 0) {
583       x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
584       // Check if dc coefficient can be quantized to zero.
585       if (sse_tx - var_tx < dc_thr || sse == var)
586         x->skip_txfm[0] = SKIP_TXFM_AC_DC;
587     } else {
588       if (sse_tx - var_tx < dc_thr || sse == var) skip_dc = 1;
589     }
590   }
591 
592   if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
593     *out_rate_sum = 0;
594     *out_dist_sum = sse << 4;
595     return;
596   }
597 
598   if (!skip_dc) {
599 #if CONFIG_VP9_HIGHBITDEPTH
600     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
601                                  dc_quant >> (xd->bd - 5), &rate, &dist);
602 #else
603     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
604                                  dc_quant >> 3, &rate, &dist);
605 #endif  // CONFIG_VP9_HIGHBITDEPTH
606   }
607 
608   if (!skip_dc) {
609     *out_rate_sum = rate >> 1;
610     *out_dist_sum = dist << 3;
611   } else {
612     *out_rate_sum = 0;
613     *out_dist_sum = (sse - var) << 4;
614   }
615 
616 #if CONFIG_VP9_HIGHBITDEPTH
617   vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
618                                ac_quant >> (xd->bd - 5), &rate, &dist);
619 #else
620   vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
621                                &rate, &dist);
622 #endif  // CONFIG_VP9_HIGHBITDEPTH
623 
624   *out_rate_sum += rate;
625   *out_dist_sum += dist << 4;
626 }
627 
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)628 static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc,
629                       int *skippable, int64_t *sse, BLOCK_SIZE bsize,
630                       TX_SIZE tx_size, int rd_computed) {
631   MACROBLOCKD *xd = &x->e_mbd;
632   const struct macroblockd_plane *pd = &xd->plane[0];
633   struct macroblock_plane *const p = &x->plane[0];
634   const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
635   const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
636   const int step = 1 << (tx_size << 1);
637   const int block_step = (1 << tx_size);
638   int block = 0, r, c;
639   const int max_blocks_wide =
640       num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 : xd->mb_to_right_edge >> 5);
641   const int max_blocks_high =
642       num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 : xd->mb_to_bottom_edge >> 5);
643   int eob_cost = 0;
644   const int bw = 4 * num_4x4_w;
645   const int bh = 4 * num_4x4_h;
646 
647 #if CONFIG_VP9_HIGHBITDEPTH
648   // TODO(jingning): Implement the high bit-depth Hadamard transforms and
649   // remove this check condition.
650   // TODO(marpan): Use this path (model_rd) for 8bit under certain conditions
651   // for now, as the vp9_quantize_fp below for highbitdepth build is slow.
652   if (xd->bd != 8 ||
653       (cpi->oxcf.speed > 5 && cpi->common.frame_type != KEY_FRAME &&
654        bsize < BLOCK_32X32)) {
655     unsigned int var_y, sse_y;
656     (void)tx_size;
657     if (!rd_computed)
658       model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist,
659                         &var_y, &sse_y);
660     *sse = INT_MAX;
661     *skippable = 0;
662     return;
663   }
664 #endif
665 
666   if (cpi->sf.use_simple_block_yrd && cpi->common.frame_type != KEY_FRAME &&
667       bsize < BLOCK_32X32) {
668     unsigned int var_y, sse_y;
669     (void)tx_size;
670     if (!rd_computed)
671       model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist,
672                         &var_y, &sse_y);
673     *sse = INT_MAX;
674     *skippable = 0;
675     return;
676   }
677 
678   (void)cpi;
679 
680   // The max tx_size passed in is TX_16X16.
681   assert(tx_size != TX_32X32);
682 
683   vpx_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
684                      pd->dst.buf, pd->dst.stride);
685   *skippable = 1;
686   // Keep track of the row and column of the blocks we use so that we know
687   // if we are in the unrestricted motion border.
688   for (r = 0; r < max_blocks_high; r += block_step) {
689     for (c = 0; c < num_4x4_w; c += block_step) {
690       if (c < max_blocks_wide) {
691         const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
692         tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
693         tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
694         tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
695         uint16_t *const eob = &p->eobs[block];
696         const int diff_stride = bw;
697         const int16_t *src_diff;
698         src_diff = &p->src_diff[(r * diff_stride + c) << 2];
699 
700         switch (tx_size) {
701           case TX_16X16:
702             vpx_hadamard_16x16(src_diff, diff_stride, coeff);
703             vp9_quantize_fp(coeff, 256, x->skip_block, p->round_fp, p->quant_fp,
704                             qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
705                             scan_order->iscan);
706             break;
707           case TX_8X8:
708             vpx_hadamard_8x8(src_diff, diff_stride, coeff);
709             vp9_quantize_fp(coeff, 64, x->skip_block, p->round_fp, p->quant_fp,
710                             qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
711                             scan_order->iscan);
712             break;
713           case TX_4X4:
714             x->fwd_txm4x4(src_diff, coeff, diff_stride);
715             vp9_quantize_fp(coeff, 16, x->skip_block, p->round_fp, p->quant_fp,
716                             qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
717                             scan_order->iscan);
718             break;
719           default: assert(0); break;
720         }
721         *skippable &= (*eob == 0);
722         eob_cost += 1;
723       }
724       block += step;
725     }
726   }
727 
728   this_rdc->rate = 0;
729   if (*sse < INT64_MAX) {
730     *sse = (*sse << 6) >> 2;
731     if (*skippable) {
732       this_rdc->dist = *sse;
733       return;
734     }
735   }
736 
737   block = 0;
738   this_rdc->dist = 0;
739   for (r = 0; r < max_blocks_high; r += block_step) {
740     for (c = 0; c < num_4x4_w; c += block_step) {
741       if (c < max_blocks_wide) {
742         tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
743         tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
744         tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
745         uint16_t *const eob = &p->eobs[block];
746 
747         if (*eob == 1)
748           this_rdc->rate += (int)abs(qcoeff[0]);
749         else if (*eob > 1)
750           this_rdc->rate += vpx_satd(qcoeff, step << 4);
751 
752         this_rdc->dist += vp9_block_error_fp(coeff, dqcoeff, step << 4) >> 2;
753       }
754       block += step;
755     }
756   }
757 
758   // If skippable is set, rate gets clobbered later.
759   this_rdc->rate <<= (2 + VP9_PROB_COST_SHIFT);
760   this_rdc->rate += (eob_cost << VP9_PROB_COST_SHIFT);
761 }
762 
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)763 static void model_rd_for_sb_uv(VP9_COMP *cpi, BLOCK_SIZE plane_bsize,
764                                MACROBLOCK *x, MACROBLOCKD *xd,
765                                RD_COST *this_rdc, unsigned int *var_y,
766                                unsigned int *sse_y, int start_plane,
767                                int stop_plane) {
768   // Note our transform coeffs are 8 times an orthogonal transform.
769   // Hence quantizer step is also 8 times. To get effective quantizer
770   // we need to divide by 8 before sending to modeling function.
771   unsigned int sse;
772   int rate;
773   int64_t dist;
774   int i;
775 #if CONFIG_VP9_HIGHBITDEPTH
776   uint64_t tot_var = *var_y;
777   uint64_t tot_sse = *sse_y;
778 #else
779   uint32_t tot_var = *var_y;
780   uint32_t tot_sse = *sse_y;
781 #endif
782 
783   this_rdc->rate = 0;
784   this_rdc->dist = 0;
785 
786   for (i = start_plane; i <= stop_plane; ++i) {
787     struct macroblock_plane *const p = &x->plane[i];
788     struct macroblockd_plane *const pd = &xd->plane[i];
789     const uint32_t dc_quant = pd->dequant[0];
790     const uint32_t ac_quant = pd->dequant[1];
791     const BLOCK_SIZE bs = plane_bsize;
792     unsigned int var;
793     if (!x->color_sensitivity[i - 1]) continue;
794 
795     var = cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf,
796                              pd->dst.stride, &sse);
797     assert(sse >= var);
798     tot_var += var;
799     tot_sse += sse;
800 
801 #if CONFIG_VP9_HIGHBITDEPTH
802     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
803                                  dc_quant >> (xd->bd - 5), &rate, &dist);
804 #else
805     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
806                                  dc_quant >> 3, &rate, &dist);
807 #endif  // CONFIG_VP9_HIGHBITDEPTH
808 
809     this_rdc->rate += rate >> 1;
810     this_rdc->dist += dist << 3;
811 
812 #if CONFIG_VP9_HIGHBITDEPTH
813     vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
814                                  ac_quant >> (xd->bd - 5), &rate, &dist);
815 #else
816     vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs], ac_quant >> 3,
817                                  &rate, &dist);
818 #endif  // CONFIG_VP9_HIGHBITDEPTH
819 
820     this_rdc->rate += rate;
821     this_rdc->dist += dist << 4;
822   }
823 
824 #if CONFIG_VP9_HIGHBITDEPTH
825   *var_y = tot_var > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_var;
826   *sse_y = tot_sse > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_sse;
827 #else
828   *var_y = tot_var;
829   *sse_y = tot_sse;
830 #endif
831 }
832 
get_pred_buffer(PRED_BUFFER * p,int len)833 static int get_pred_buffer(PRED_BUFFER *p, int len) {
834   int i;
835 
836   for (i = 0; i < len; i++) {
837     if (!p[i].in_use) {
838       p[i].in_use = 1;
839       return i;
840     }
841   }
842   return -1;
843 }
844 
free_pred_buffer(PRED_BUFFER * p)845 static void free_pred_buffer(PRED_BUFFER *p) {
846   if (p != NULL) p->in_use = 0;
847 }
848 
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)849 static void encode_breakout_test(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
850                                  int mi_row, int mi_col,
851                                  MV_REFERENCE_FRAME ref_frame,
852                                  PREDICTION_MODE this_mode, unsigned int var_y,
853                                  unsigned int sse_y,
854                                  struct buf_2d yv12_mb[][MAX_MB_PLANE],
855                                  int *rate, int64_t *dist) {
856   MACROBLOCKD *xd = &x->e_mbd;
857   MODE_INFO *const mi = xd->mi[0];
858   const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
859   unsigned int var = var_y, sse = sse_y;
860   // Skipping threshold for ac.
861   unsigned int thresh_ac;
862   // Skipping threshold for dc.
863   unsigned int thresh_dc;
864   int motion_low = 1;
865   if (mi->mv[0].as_mv.row > 64 || mi->mv[0].as_mv.row < -64 ||
866       mi->mv[0].as_mv.col > 64 || mi->mv[0].as_mv.col < -64)
867     motion_low = 0;
868   if (x->encode_breakout > 0 && motion_low == 1) {
869     // Set a maximum for threshold to avoid big PSNR loss in low bit rate
870     // case. Use extreme low threshold for static frames to limit
871     // skipping.
872     const unsigned int max_thresh = 36000;
873     // The encode_breakout input
874     const unsigned int min_thresh =
875         VPXMIN(((unsigned int)x->encode_breakout << 4), max_thresh);
876 #if CONFIG_VP9_HIGHBITDEPTH
877     const int shift = (xd->bd << 1) - 16;
878 #endif
879 
880     // Calculate threshold according to dequant value.
881     thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) >> 3;
882 #if CONFIG_VP9_HIGHBITDEPTH
883     if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
884       thresh_ac = ROUND_POWER_OF_TWO(thresh_ac, shift);
885     }
886 #endif  // CONFIG_VP9_HIGHBITDEPTH
887     thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
888 
889     // Adjust ac threshold according to partition size.
890     thresh_ac >>=
891         8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
892 
893     thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
894 #if CONFIG_VP9_HIGHBITDEPTH
895     if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
896       thresh_dc = ROUND_POWER_OF_TWO(thresh_dc, shift);
897     }
898 #endif  // CONFIG_VP9_HIGHBITDEPTH
899   } else {
900     thresh_ac = 0;
901     thresh_dc = 0;
902   }
903 
904   // Y skipping condition checking for ac and dc.
905   if (var <= thresh_ac && (sse - var) <= thresh_dc) {
906     unsigned int sse_u, sse_v;
907     unsigned int var_u, var_v;
908     unsigned int thresh_ac_uv = thresh_ac;
909     unsigned int thresh_dc_uv = thresh_dc;
910     if (x->sb_is_skin) {
911       thresh_ac_uv = 0;
912       thresh_dc_uv = 0;
913     }
914 
915     // Skip UV prediction unless breakout is zero (lossless) to save
916     // computation with low impact on the result
917     if (x->encode_breakout == 0) {
918       xd->plane[1].pre[0] = yv12_mb[ref_frame][1];
919       xd->plane[2].pre[0] = yv12_mb[ref_frame][2];
920       vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize);
921     }
922 
923     var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf, x->plane[1].src.stride,
924                                     xd->plane[1].dst.buf,
925                                     xd->plane[1].dst.stride, &sse_u);
926 
927     // U skipping condition checking
928     if (((var_u << 2) <= thresh_ac_uv) && (sse_u - var_u <= thresh_dc_uv)) {
929       var_v = cpi->fn_ptr[uv_size].vf(
930           x->plane[2].src.buf, x->plane[2].src.stride, xd->plane[2].dst.buf,
931           xd->plane[2].dst.stride, &sse_v);
932 
933       // V skipping condition checking
934       if (((var_v << 2) <= thresh_ac_uv) && (sse_v - var_v <= thresh_dc_uv)) {
935         x->skip = 1;
936 
937         // The cost of skip bit needs to be added.
938         *rate = cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
939                                     [INTER_OFFSET(this_mode)];
940 
941         // More on this part of rate
942         // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
943 
944         // Scaling factor for SSE from spatial domain to frequency
945         // domain is 16. Adjust distortion accordingly.
946         // TODO(yunqingwang): In this function, only y-plane dist is
947         // calculated.
948         *dist = (sse << 4);  // + ((sse_u + sse_v) << 4);
949 
950         // *disable_skip = 1;
951       }
952     }
953   }
954 }
955 
956 struct estimate_block_intra_args {
957   VP9_COMP *cpi;
958   MACROBLOCK *x;
959   PREDICTION_MODE mode;
960   int skippable;
961   RD_COST *rdc;
962 };
963 
estimate_block_intra(int plane,int block,int row,int col,BLOCK_SIZE plane_bsize,TX_SIZE tx_size,void * arg)964 static void estimate_block_intra(int plane, int block, int row, int col,
965                                  BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
966                                  void *arg) {
967   struct estimate_block_intra_args *const args = arg;
968   VP9_COMP *const cpi = args->cpi;
969   MACROBLOCK *const x = args->x;
970   MACROBLOCKD *const xd = &x->e_mbd;
971   struct macroblock_plane *const p = &x->plane[0];
972   struct macroblockd_plane *const pd = &xd->plane[0];
973   const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
974   uint8_t *const src_buf_base = p->src.buf;
975   uint8_t *const dst_buf_base = pd->dst.buf;
976   const int src_stride = p->src.stride;
977   const int dst_stride = pd->dst.stride;
978   RD_COST this_rdc;
979 
980   (void)block;
981 
982   p->src.buf = &src_buf_base[4 * (row * src_stride + col)];
983   pd->dst.buf = &dst_buf_base[4 * (row * dst_stride + col)];
984   // Use source buffer as an approximation for the fully reconstructed buffer.
985   vp9_predict_intra_block(xd, b_width_log2_lookup[plane_bsize], tx_size,
986                           args->mode, x->skip_encode ? p->src.buf : pd->dst.buf,
987                           x->skip_encode ? src_stride : dst_stride, pd->dst.buf,
988                           dst_stride, col, row, plane);
989 
990   if (plane == 0) {
991     int64_t this_sse = INT64_MAX;
992     // TODO(jingning): This needs further refactoring.
993     block_yrd(cpi, x, &this_rdc, &args->skippable, &this_sse, bsize_tx,
994               VPXMIN(tx_size, TX_16X16), 0);
995   } else {
996     unsigned int var = 0;
997     unsigned int sse = 0;
998     model_rd_for_sb_uv(cpi, plane_bsize, x, xd, &this_rdc, &var, &sse, plane,
999                        plane);
1000   }
1001 
1002   p->src.buf = src_buf_base;
1003   pd->dst.buf = dst_buf_base;
1004   args->rdc->rate += this_rdc.rate;
1005   args->rdc->dist += this_rdc.dist;
1006 }
1007 
1008 static const THR_MODES mode_idx[MAX_REF_FRAMES][4] = {
1009   { THR_DC, THR_V_PRED, THR_H_PRED, THR_TM },
1010   { THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV },
1011   { THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG },
1012   { THR_NEARESTA, THR_NEARA, THR_ZEROA, THR_NEWA },
1013 };
1014 
1015 static const PREDICTION_MODE intra_mode_list[] = { DC_PRED, V_PRED, H_PRED,
1016                                                    TM_PRED };
1017 
mode_offset(const PREDICTION_MODE mode)1018 static int mode_offset(const PREDICTION_MODE mode) {
1019   if (mode >= NEARESTMV) {
1020     return INTER_OFFSET(mode);
1021   } else {
1022     switch (mode) {
1023       case DC_PRED: return 0;
1024       case V_PRED: return 1;
1025       case H_PRED: return 2;
1026       case TM_PRED: return 3;
1027       default: return -1;
1028     }
1029   }
1030 }
1031 
rd_less_than_thresh_row_mt(int64_t best_rd,int thresh,const int * const thresh_fact)1032 static INLINE int rd_less_than_thresh_row_mt(int64_t best_rd, int thresh,
1033                                              const int *const thresh_fact) {
1034   int is_rd_less_than_thresh;
1035   is_rd_less_than_thresh =
1036       best_rd < ((int64_t)thresh * (*thresh_fact) >> 5) || thresh == INT_MAX;
1037   return is_rd_less_than_thresh;
1038 }
1039 
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)1040 static INLINE void update_thresh_freq_fact_row_mt(
1041     VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance,
1042     int thresh_freq_fact_idx, MV_REFERENCE_FRAME ref_frame,
1043     THR_MODES best_mode_idx, PREDICTION_MODE mode) {
1044   THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1045   int freq_fact_idx = thresh_freq_fact_idx + thr_mode_idx;
1046   int *freq_fact = &tile_data->row_base_thresh_freq_fact[freq_fact_idx];
1047   if (thr_mode_idx == best_mode_idx)
1048     *freq_fact -= (*freq_fact >> 4);
1049   else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV &&
1050            ref_frame == LAST_FRAME && source_variance < 5) {
1051     *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32);
1052   } else {
1053     *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
1054                         cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1055   }
1056 }
1057 
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)1058 static INLINE void update_thresh_freq_fact(
1059     VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance,
1060     BLOCK_SIZE bsize, MV_REFERENCE_FRAME ref_frame, THR_MODES best_mode_idx,
1061     PREDICTION_MODE mode) {
1062   THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1063   int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx];
1064   if (thr_mode_idx == best_mode_idx)
1065     *freq_fact -= (*freq_fact >> 4);
1066   else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV &&
1067            ref_frame == LAST_FRAME && source_variance < 5) {
1068     *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32);
1069   } else {
1070     *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
1071                         cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1072   }
1073 }
1074 
vp9_pick_intra_mode(VP9_COMP * cpi,MACROBLOCK * x,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)1075 void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
1076                          BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1077   MACROBLOCKD *const xd = &x->e_mbd;
1078   MODE_INFO *const mi = xd->mi[0];
1079   RD_COST this_rdc, best_rdc;
1080   PREDICTION_MODE this_mode;
1081   struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
1082   const TX_SIZE intra_tx_size =
1083       VPXMIN(max_txsize_lookup[bsize],
1084              tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
1085   MODE_INFO *const mic = xd->mi[0];
1086   int *bmode_costs;
1087   const MODE_INFO *above_mi = xd->above_mi;
1088   const MODE_INFO *left_mi = xd->left_mi;
1089   const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1090   const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1091   bmode_costs = cpi->y_mode_costs[A][L];
1092 
1093   (void)ctx;
1094   vp9_rd_cost_reset(&best_rdc);
1095   vp9_rd_cost_reset(&this_rdc);
1096 
1097   mi->ref_frame[0] = INTRA_FRAME;
1098   // Initialize interp_filter here so we do not have to check for inter block
1099   // modes in get_pred_context_switchable_interp()
1100   mi->interp_filter = SWITCHABLE_FILTERS;
1101 
1102   mi->mv[0].as_int = INVALID_MV;
1103   mi->uv_mode = DC_PRED;
1104   memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
1105 
1106   // Change the limit of this loop to add other intra prediction
1107   // mode tests.
1108   for (this_mode = DC_PRED; this_mode <= H_PRED; ++this_mode) {
1109     this_rdc.dist = this_rdc.rate = 0;
1110     args.mode = this_mode;
1111     args.skippable = 1;
1112     args.rdc = &this_rdc;
1113     mi->tx_size = intra_tx_size;
1114     vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
1115                                            &args);
1116     if (args.skippable) {
1117       x->skip_txfm[0] = SKIP_TXFM_AC_DC;
1118       this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
1119     } else {
1120       x->skip_txfm[0] = SKIP_TXFM_NONE;
1121       this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
1122     }
1123     this_rdc.rate += bmode_costs[this_mode];
1124     this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1125 
1126     if (this_rdc.rdcost < best_rdc.rdcost) {
1127       best_rdc = this_rdc;
1128       mi->mode = this_mode;
1129     }
1130   }
1131 
1132   *rd_cost = best_rdc;
1133 }
1134 
init_ref_frame_cost(VP9_COMMON * const cm,MACROBLOCKD * const xd,int ref_frame_cost[MAX_REF_FRAMES])1135 static void init_ref_frame_cost(VP9_COMMON *const cm, MACROBLOCKD *const xd,
1136                                 int ref_frame_cost[MAX_REF_FRAMES]) {
1137   vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
1138   vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
1139   vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
1140 
1141   ref_frame_cost[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
1142   ref_frame_cost[LAST_FRAME] = ref_frame_cost[GOLDEN_FRAME] =
1143       ref_frame_cost[ALTREF_FRAME] = vp9_cost_bit(intra_inter_p, 1);
1144 
1145   ref_frame_cost[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
1146   ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1147   ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1148   ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
1149   ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
1150 }
1151 
1152 typedef struct {
1153   MV_REFERENCE_FRAME ref_frame;
1154   PREDICTION_MODE pred_mode;
1155 } REF_MODE;
1156 
1157 #define RT_INTER_MODES 12
1158 static const REF_MODE ref_mode_set[RT_INTER_MODES] = {
1159   { LAST_FRAME, ZEROMV },   { LAST_FRAME, NEARESTMV },
1160   { GOLDEN_FRAME, ZEROMV }, { LAST_FRAME, NEARMV },
1161   { LAST_FRAME, NEWMV },    { GOLDEN_FRAME, NEARESTMV },
1162   { GOLDEN_FRAME, NEARMV }, { GOLDEN_FRAME, NEWMV },
1163   { ALTREF_FRAME, ZEROMV }, { ALTREF_FRAME, NEARESTMV },
1164   { ALTREF_FRAME, NEARMV }, { ALTREF_FRAME, NEWMV }
1165 };
1166 static const REF_MODE ref_mode_set_svc[RT_INTER_MODES] = {
1167   { LAST_FRAME, ZEROMV },      { LAST_FRAME, NEARESTMV },
1168   { LAST_FRAME, NEARMV },      { GOLDEN_FRAME, ZEROMV },
1169   { GOLDEN_FRAME, NEARESTMV }, { GOLDEN_FRAME, NEARMV },
1170   { LAST_FRAME, NEWMV },       { GOLDEN_FRAME, NEWMV }
1171 };
1172 
set_intra_cost_penalty(const VP9_COMP * const cpi,BLOCK_SIZE bsize)1173 static int set_intra_cost_penalty(const VP9_COMP *const cpi, BLOCK_SIZE bsize) {
1174   const VP9_COMMON *const cm = &cpi->common;
1175   // Reduce the intra cost penalty for small blocks (<=16x16).
1176   int reduction_fac =
1177       (bsize <= BLOCK_16X16) ? ((bsize <= BLOCK_8X8) ? 4 : 2) : 0;
1178   if (cpi->noise_estimate.enabled && cpi->noise_estimate.level == kHigh)
1179     // Don't reduce intra cost penalty if estimated noise level is high.
1180     reduction_fac = 0;
1181   return vp9_get_intra_cost_penalty(cm->base_qindex, cm->y_dc_delta_q,
1182                                     cm->bit_depth) >>
1183          reduction_fac;
1184 }
1185 
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)1186 static INLINE void find_predictors(
1187     VP9_COMP *cpi, MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame,
1188     int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1189     int const_motion[MAX_REF_FRAMES], int *ref_frame_skip_mask,
1190     const int flag_list[4], TileDataEnc *tile_data, int mi_row, int mi_col,
1191     struct buf_2d yv12_mb[4][MAX_MB_PLANE], BLOCK_SIZE bsize,
1192     int force_skip_low_temp_var) {
1193   VP9_COMMON *const cm = &cpi->common;
1194   MACROBLOCKD *const xd = &x->e_mbd;
1195   const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
1196   TileInfo *const tile_info = &tile_data->tile_info;
1197   // TODO(jingning) placeholder for inter-frame non-RD mode decision.
1198   x->pred_mv_sad[ref_frame] = INT_MAX;
1199   frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
1200   frame_mv[ZEROMV][ref_frame].as_int = 0;
1201   // this needs various further optimizations. to be continued..
1202   if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
1203     int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
1204     const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
1205     vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
1206     if (cm->use_prev_frame_mvs) {
1207       vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
1208                        x->mbmi_ext->mode_context);
1209     } else {
1210       const_motion[ref_frame] =
1211           mv_refs_rt(cpi, cm, x, xd, tile_info, xd->mi[0], ref_frame,
1212                      candidates, &frame_mv[NEWMV][ref_frame], mi_row, mi_col,
1213                      (int)(cpi->svc.use_base_mv && cpi->svc.spatial_layer_id));
1214     }
1215     vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
1216                           &frame_mv[NEARESTMV][ref_frame],
1217                           &frame_mv[NEARMV][ref_frame]);
1218     // Early exit for golden frame if force_skip_low_temp_var is set.
1219     if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8 &&
1220         !(force_skip_low_temp_var && ref_frame == GOLDEN_FRAME)) {
1221       vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
1222                   bsize);
1223     }
1224   } else {
1225     *ref_frame_skip_mask |= (1 << ref_frame);
1226   }
1227 }
1228 
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)1229 static void vp9_NEWMV_diff_bias(const NOISE_ESTIMATE *ne, MACROBLOCKD *xd,
1230                                 PREDICTION_MODE this_mode, RD_COST *this_rdc,
1231                                 BLOCK_SIZE bsize, int mv_row, int mv_col,
1232                                 int is_last_frame, int lowvar_highsumdiff,
1233                                 int is_skin) {
1234   // Bias against MVs associated with NEWMV mode that are very different from
1235   // top/left neighbors.
1236   if (this_mode == NEWMV) {
1237     int al_mv_average_row;
1238     int al_mv_average_col;
1239     int left_row, left_col;
1240     int row_diff, col_diff;
1241     int above_mv_valid = 0;
1242     int left_mv_valid = 0;
1243     int above_row = 0;
1244     int above_col = 0;
1245 
1246     if (xd->above_mi) {
1247       above_mv_valid = xd->above_mi->mv[0].as_int != INVALID_MV;
1248       above_row = xd->above_mi->mv[0].as_mv.row;
1249       above_col = xd->above_mi->mv[0].as_mv.col;
1250     }
1251     if (xd->left_mi) {
1252       left_mv_valid = xd->left_mi->mv[0].as_int != INVALID_MV;
1253       left_row = xd->left_mi->mv[0].as_mv.row;
1254       left_col = xd->left_mi->mv[0].as_mv.col;
1255     }
1256     if (above_mv_valid && left_mv_valid) {
1257       al_mv_average_row = (above_row + left_row + 1) >> 1;
1258       al_mv_average_col = (above_col + left_col + 1) >> 1;
1259     } else if (above_mv_valid) {
1260       al_mv_average_row = above_row;
1261       al_mv_average_col = above_col;
1262     } else if (left_mv_valid) {
1263       al_mv_average_row = left_row;
1264       al_mv_average_col = left_col;
1265     } else {
1266       al_mv_average_row = al_mv_average_col = 0;
1267     }
1268     row_diff = (al_mv_average_row - mv_row);
1269     col_diff = (al_mv_average_col - mv_col);
1270     if (row_diff > 48 || row_diff < -48 || col_diff > 48 || col_diff < -48) {
1271       if (bsize > BLOCK_32X32)
1272         this_rdc->rdcost = this_rdc->rdcost << 1;
1273       else
1274         this_rdc->rdcost = 3 * this_rdc->rdcost >> 1;
1275     }
1276   }
1277   // If noise estimation is enabled, and estimated level is above threshold,
1278   // add a bias to LAST reference with small motion, for large blocks.
1279   if (ne->enabled && ne->level >= kMedium && bsize >= BLOCK_32X32 &&
1280       is_last_frame && mv_row < 8 && mv_row > -8 && mv_col < 8 && mv_col > -8)
1281     this_rdc->rdcost = 7 * (this_rdc->rdcost >> 3);
1282   else if (lowvar_highsumdiff && !is_skin && bsize >= BLOCK_16X16 &&
1283            is_last_frame && mv_row < 16 && mv_row > -16 && mv_col < 16 &&
1284            mv_col > -16)
1285     this_rdc->rdcost = 7 * (this_rdc->rdcost >> 3);
1286 }
1287 
1288 #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,TX_SIZE best_tx_size,PREDICTION_MODE best_mode,MV_REFERENCE_FRAME best_ref_frame,INTERP_FILTER best_pred_filter,uint8_t best_mode_skip_txfm)1289 static void vp9_pickmode_ctx_den_update(
1290     VP9_PICKMODE_CTX_DEN *ctx_den, int64_t zero_last_cost_orig,
1291     int ref_frame_cost[MAX_REF_FRAMES],
1292     int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int reuse_inter_pred,
1293     TX_SIZE best_tx_size, PREDICTION_MODE best_mode,
1294     MV_REFERENCE_FRAME best_ref_frame, INTERP_FILTER best_pred_filter,
1295     uint8_t best_mode_skip_txfm) {
1296   ctx_den->zero_last_cost_orig = zero_last_cost_orig;
1297   ctx_den->ref_frame_cost = ref_frame_cost;
1298   ctx_den->frame_mv = frame_mv;
1299   ctx_den->reuse_inter_pred = reuse_inter_pred;
1300   ctx_den->best_tx_size = best_tx_size;
1301   ctx_den->best_mode = best_mode;
1302   ctx_den->best_ref_frame = best_ref_frame;
1303   ctx_den->best_pred_filter = best_pred_filter;
1304   ctx_den->best_mode_skip_txfm = best_mode_skip_txfm;
1305 }
1306 
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)1307 static void recheck_zeromv_after_denoising(
1308     VP9_COMP *cpi, MODE_INFO *const mi, MACROBLOCK *x, MACROBLOCKD *const xd,
1309     VP9_DENOISER_DECISION decision, VP9_PICKMODE_CTX_DEN *ctx_den,
1310     struct buf_2d yv12_mb[4][MAX_MB_PLANE], RD_COST *best_rdc, BLOCK_SIZE bsize,
1311     int mi_row, int mi_col) {
1312   // If INTRA or GOLDEN reference was selected, re-evaluate ZEROMV on
1313   // denoised result. Only do this under noise conditions, and if rdcost of
1314   // ZEROMV onoriginal source is not significantly higher than rdcost of best
1315   // mode.
1316   if (cpi->noise_estimate.enabled && cpi->noise_estimate.level > kLow &&
1317       ctx_den->zero_last_cost_orig < (best_rdc->rdcost << 3) &&
1318       ((ctx_den->best_ref_frame == INTRA_FRAME && decision >= FILTER_BLOCK) ||
1319        (ctx_den->best_ref_frame == GOLDEN_FRAME &&
1320         cpi->svc.number_spatial_layers == 1 &&
1321         decision == FILTER_ZEROMV_BLOCK))) {
1322     // Check if we should pick ZEROMV on denoised signal.
1323     int rate = 0;
1324     int64_t dist = 0;
1325     uint32_t var_y = UINT_MAX;
1326     uint32_t sse_y = UINT_MAX;
1327     RD_COST this_rdc;
1328     mi->mode = ZEROMV;
1329     mi->ref_frame[0] = LAST_FRAME;
1330     mi->ref_frame[1] = NONE;
1331     mi->mv[0].as_int = 0;
1332     mi->interp_filter = EIGHTTAP;
1333     xd->plane[0].pre[0] = yv12_mb[LAST_FRAME][0];
1334     vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1335     model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y);
1336     this_rdc.rate = rate + ctx_den->ref_frame_cost[LAST_FRAME] +
1337                     cpi->inter_mode_cost[x->mbmi_ext->mode_context[LAST_FRAME]]
1338                                         [INTER_OFFSET(ZEROMV)];
1339     this_rdc.dist = dist;
1340     this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, rate, dist);
1341     // Don't switch to ZEROMV if the rdcost for ZEROMV on denoised source
1342     // is higher than best_ref mode (on original source).
1343     if (this_rdc.rdcost > best_rdc->rdcost) {
1344       this_rdc = *best_rdc;
1345       mi->mode = ctx_den->best_mode;
1346       mi->ref_frame[0] = ctx_den->best_ref_frame;
1347       mi->interp_filter = ctx_den->best_pred_filter;
1348       if (ctx_den->best_ref_frame == INTRA_FRAME) {
1349         mi->mv[0].as_int = INVALID_MV;
1350         mi->interp_filter = SWITCHABLE_FILTERS;
1351       } else if (ctx_den->best_ref_frame == GOLDEN_FRAME) {
1352         mi->mv[0].as_int =
1353             ctx_den->frame_mv[ctx_den->best_mode][ctx_den->best_ref_frame]
1354                 .as_int;
1355         if (ctx_den->reuse_inter_pred) {
1356           xd->plane[0].pre[0] = yv12_mb[GOLDEN_FRAME][0];
1357           vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1358         }
1359       }
1360       mi->tx_size = ctx_den->best_tx_size;
1361       x->skip_txfm[0] = ctx_den->best_mode_skip_txfm;
1362     } else {
1363       ctx_den->best_ref_frame = LAST_FRAME;
1364       *best_rdc = this_rdc;
1365     }
1366   }
1367 }
1368 #endif  // CONFIG_VP9_TEMPORAL_DENOISING
1369 
get_force_skip_low_temp_var(uint8_t * variance_low,int mi_row,int mi_col,BLOCK_SIZE bsize)1370 static INLINE int get_force_skip_low_temp_var(uint8_t *variance_low, int mi_row,
1371                                               int mi_col, BLOCK_SIZE bsize) {
1372   const int i = (mi_row & 0x7) >> 1;
1373   const int j = (mi_col & 0x7) >> 1;
1374   int force_skip_low_temp_var = 0;
1375   // Set force_skip_low_temp_var based on the block size and block offset.
1376   if (bsize == BLOCK_64X64) {
1377     force_skip_low_temp_var = variance_low[0];
1378   } else if (bsize == BLOCK_64X32) {
1379     if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1380       force_skip_low_temp_var = variance_low[1];
1381     } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1382       force_skip_low_temp_var = variance_low[2];
1383     }
1384   } else if (bsize == BLOCK_32X64) {
1385     if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1386       force_skip_low_temp_var = variance_low[3];
1387     } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1388       force_skip_low_temp_var = variance_low[4];
1389     }
1390   } else if (bsize == BLOCK_32X32) {
1391     if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1392       force_skip_low_temp_var = variance_low[5];
1393     } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1394       force_skip_low_temp_var = variance_low[6];
1395     } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1396       force_skip_low_temp_var = variance_low[7];
1397     } else if ((mi_col & 0x7) && (mi_row & 0x7)) {
1398       force_skip_low_temp_var = variance_low[8];
1399     }
1400   } else if (bsize == BLOCK_16X16) {
1401     force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]];
1402   } else if (bsize == BLOCK_32X16) {
1403     // The col shift index for the second 16x16 block.
1404     const int j2 = ((mi_col + 2) & 0x7) >> 1;
1405     // Only if each 16x16 block inside has low temporal variance.
1406     force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1407                               variance_low[pos_shift_16x16[i][j2]];
1408   } else if (bsize == BLOCK_16X32) {
1409     // The row shift index for the second 16x16 block.
1410     const int i2 = ((mi_row + 2) & 0x7) >> 1;
1411     force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1412                               variance_low[pos_shift_16x16[i2][j]];
1413   }
1414   return force_skip_low_temp_var;
1415 }
1416 
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)1417 void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data,
1418                          int mi_row, int mi_col, RD_COST *rd_cost,
1419                          BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1420   VP9_COMMON *const cm = &cpi->common;
1421   SPEED_FEATURES *const sf = &cpi->sf;
1422   const SVC *const svc = &cpi->svc;
1423   MACROBLOCKD *const xd = &x->e_mbd;
1424   MODE_INFO *const mi = xd->mi[0];
1425   struct macroblockd_plane *const pd = &xd->plane[0];
1426   PREDICTION_MODE best_mode = ZEROMV;
1427   MV_REFERENCE_FRAME ref_frame, best_ref_frame = LAST_FRAME;
1428   MV_REFERENCE_FRAME usable_ref_frame;
1429   TX_SIZE best_tx_size = TX_SIZES;
1430   INTERP_FILTER best_pred_filter = EIGHTTAP;
1431   int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1432   struct buf_2d yv12_mb[4][MAX_MB_PLANE];
1433   static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
1434                                     VP9_ALT_FLAG };
1435   RD_COST this_rdc, best_rdc;
1436   uint8_t skip_txfm = SKIP_TXFM_NONE, best_mode_skip_txfm = SKIP_TXFM_NONE;
1437   // var_y and sse_y are saved to be used in skipping checking
1438   unsigned int var_y = UINT_MAX;
1439   unsigned int sse_y = UINT_MAX;
1440   const int intra_cost_penalty = set_intra_cost_penalty(cpi, bsize);
1441   int64_t inter_mode_thresh =
1442       RDCOST(x->rdmult, x->rddiv, intra_cost_penalty, 0);
1443   const int *const rd_threshes = cpi->rd.threshes[mi->segment_id][bsize];
1444   const int sb_row = mi_row >> MI_BLOCK_SIZE_LOG2;
1445   int thresh_freq_fact_idx = (sb_row * BLOCK_SIZES + bsize) * MAX_MODES;
1446   const int *const rd_thresh_freq_fact =
1447       (cpi->sf.adaptive_rd_thresh_row_mt)
1448           ? &(tile_data->row_base_thresh_freq_fact[thresh_freq_fact_idx])
1449           : tile_data->thresh_freq_fact[bsize];
1450 
1451   INTERP_FILTER filter_ref;
1452   const int bsl = mi_width_log2_lookup[bsize];
1453   const int pred_filter_search =
1454       cm->interp_filter == SWITCHABLE
1455           ? (((mi_row + mi_col) >> bsl) +
1456              get_chessboard_index(cm->current_video_frame)) &
1457                 0x1
1458           : 0;
1459   int const_motion[MAX_REF_FRAMES] = { 0 };
1460   const int bh = num_4x4_blocks_high_lookup[bsize] << 2;
1461   const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
1462   // For speed 6, the result of interp filter is reused later in actual encoding
1463   // process.
1464   // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.
1465   PRED_BUFFER tmp[4];
1466   DECLARE_ALIGNED(16, uint8_t, pred_buf[3 * 64 * 64]);
1467 #if CONFIG_VP9_HIGHBITDEPTH
1468   DECLARE_ALIGNED(16, uint16_t, pred_buf_16[3 * 64 * 64]);
1469 #endif
1470   struct buf_2d orig_dst = pd->dst;
1471   PRED_BUFFER *best_pred = NULL;
1472   PRED_BUFFER *this_mode_pred = NULL;
1473   const int pixels_in_block = bh * bw;
1474   int reuse_inter_pred = cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready;
1475   int ref_frame_skip_mask = 0;
1476   int idx;
1477   int best_pred_sad = INT_MAX;
1478   int best_early_term = 0;
1479   int ref_frame_cost[MAX_REF_FRAMES];
1480   int svc_force_zero_mode[3] = { 0 };
1481   int perform_intra_pred = 1;
1482   int use_golden_nonzeromv = 1;
1483   int force_skip_low_temp_var = 0;
1484   int skip_ref_find_pred[4] = { 0 };
1485   unsigned int sse_zeromv_normalized = UINT_MAX;
1486   unsigned int thresh_svc_skip_golden = 500;
1487 #if CONFIG_VP9_TEMPORAL_DENOISING
1488   VP9_PICKMODE_CTX_DEN ctx_den;
1489   int64_t zero_last_cost_orig = INT64_MAX;
1490   int denoise_svc_pickmode = 1;
1491 #endif
1492   INTERP_FILTER filter_gf_svc = EIGHTTAP;
1493 
1494   init_ref_frame_cost(cm, xd, ref_frame_cost);
1495 
1496   if (reuse_inter_pred) {
1497     int i;
1498     for (i = 0; i < 3; i++) {
1499 #if CONFIG_VP9_HIGHBITDEPTH
1500       if (cm->use_highbitdepth)
1501         tmp[i].data = CONVERT_TO_BYTEPTR(&pred_buf_16[pixels_in_block * i]);
1502       else
1503         tmp[i].data = &pred_buf[pixels_in_block * i];
1504 #else
1505       tmp[i].data = &pred_buf[pixels_in_block * i];
1506 #endif  // CONFIG_VP9_HIGHBITDEPTH
1507       tmp[i].stride = bw;
1508       tmp[i].in_use = 0;
1509     }
1510     tmp[3].data = pd->dst.buf;
1511     tmp[3].stride = pd->dst.stride;
1512     tmp[3].in_use = 0;
1513   }
1514 
1515   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
1516   x->skip = 0;
1517 
1518   // Instead of using vp9_get_pred_context_switchable_interp(xd) to assign
1519   // filter_ref, we use a less strict condition on assigning filter_ref.
1520   // This is to reduce the probabily of entering the flow of not assigning
1521   // filter_ref and then skip filter search.
1522   if (xd->above_mi && is_inter_block(xd->above_mi))
1523     filter_ref = xd->above_mi->interp_filter;
1524   else if (xd->left_mi && is_inter_block(xd->left_mi))
1525     filter_ref = xd->left_mi->interp_filter;
1526   else
1527     filter_ref = cm->interp_filter;
1528 
1529   // initialize mode decisions
1530   vp9_rd_cost_reset(&best_rdc);
1531   vp9_rd_cost_reset(rd_cost);
1532   mi->sb_type = bsize;
1533   mi->ref_frame[0] = NONE;
1534   mi->ref_frame[1] = NONE;
1535 
1536   mi->tx_size =
1537       VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[cm->tx_mode]);
1538 
1539   if (sf->short_circuit_flat_blocks || sf->limit_newmv_early_exit) {
1540 #if CONFIG_VP9_HIGHBITDEPTH
1541     if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
1542       x->source_variance = vp9_high_get_sby_perpixel_variance(
1543           cpi, &x->plane[0].src, bsize, xd->bd);
1544     else
1545 #endif  // CONFIG_VP9_HIGHBITDEPTH
1546       x->source_variance =
1547           vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1548   }
1549 
1550 #if CONFIG_VP9_TEMPORAL_DENOISING
1551   if (cpi->oxcf.noise_sensitivity > 0) {
1552     if (cpi->use_svc) {
1553       int layer = LAYER_IDS_TO_IDX(cpi->svc.spatial_layer_id,
1554                                    cpi->svc.temporal_layer_id,
1555                                    cpi->svc.number_temporal_layers);
1556       LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer];
1557       denoise_svc_pickmode = denoise_svc(cpi) && !lc->is_key_frame;
1558     }
1559     if (cpi->denoiser.denoising_level > kDenLowLow && denoise_svc_pickmode)
1560       vp9_denoiser_reset_frame_stats(ctx);
1561   }
1562 #endif
1563 
1564   if (cpi->rc.frames_since_golden == 0 && !cpi->use_svc) {
1565     usable_ref_frame = LAST_FRAME;
1566   } else {
1567     usable_ref_frame = GOLDEN_FRAME;
1568   }
1569 
1570   if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
1571     if (cpi->rc.alt_ref_gf_group || cpi->rc.is_src_frame_alt_ref)
1572       usable_ref_frame = ALTREF_FRAME;
1573 
1574     if (cpi->rc.is_src_frame_alt_ref) {
1575       skip_ref_find_pred[LAST_FRAME] = 1;
1576       skip_ref_find_pred[GOLDEN_FRAME] = 1;
1577     }
1578   }
1579 
1580   // For svc mode, on spatial_layer_id > 0: if the reference has different scale
1581   // constrain the inter mode to only test zero motion.
1582   if (cpi->use_svc && svc->force_zero_mode_spatial_ref &&
1583       cpi->svc.spatial_layer_id > 0) {
1584     if (cpi->ref_frame_flags & flag_list[LAST_FRAME]) {
1585       struct scale_factors *const sf = &cm->frame_refs[LAST_FRAME - 1].sf;
1586       if (vp9_is_scaled(sf)) svc_force_zero_mode[LAST_FRAME - 1] = 1;
1587     }
1588     if (cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) {
1589       struct scale_factors *const sf = &cm->frame_refs[GOLDEN_FRAME - 1].sf;
1590       if (vp9_is_scaled(sf)) svc_force_zero_mode[GOLDEN_FRAME - 1] = 1;
1591     }
1592   }
1593 
1594   if (cpi->sf.short_circuit_low_temp_var) {
1595     force_skip_low_temp_var =
1596         get_force_skip_low_temp_var(&x->variance_low[0], mi_row, mi_col, bsize);
1597     // If force_skip_low_temp_var is set, and for short circuit mode = 1 and 3,
1598     // skip golden reference.
1599     if ((cpi->sf.short_circuit_low_temp_var == 1 ||
1600          cpi->sf.short_circuit_low_temp_var == 3) &&
1601         force_skip_low_temp_var) {
1602       usable_ref_frame = LAST_FRAME;
1603     }
1604   }
1605 
1606   if (!((cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) &&
1607         !svc_force_zero_mode[GOLDEN_FRAME - 1] && !force_skip_low_temp_var))
1608     use_golden_nonzeromv = 0;
1609 
1610   if (cpi->oxcf.speed >= 8 && !cpi->use_svc &&
1611       ((cpi->rc.frames_since_golden + 1) < x->last_sb_high_content ||
1612        x->last_sb_high_content > 40))
1613     usable_ref_frame = LAST_FRAME;
1614 
1615   for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) {
1616     if (!skip_ref_find_pred[ref_frame]) {
1617       find_predictors(cpi, x, ref_frame, frame_mv, const_motion,
1618                       &ref_frame_skip_mask, flag_list, tile_data, mi_row,
1619                       mi_col, yv12_mb, bsize, force_skip_low_temp_var);
1620     }
1621   }
1622 
1623   for (idx = 0; idx < RT_INTER_MODES; ++idx) {
1624     int rate_mv = 0;
1625     int mode_rd_thresh;
1626     int mode_index;
1627     int i;
1628     int64_t this_sse;
1629     int is_skippable;
1630     int this_early_term = 0;
1631     int rd_computed = 0;
1632 
1633     PREDICTION_MODE this_mode = ref_mode_set[idx].pred_mode;
1634 
1635     ref_frame = ref_mode_set[idx].ref_frame;
1636 
1637     if (cpi->use_svc) {
1638       this_mode = ref_mode_set_svc[idx].pred_mode;
1639       ref_frame = ref_mode_set_svc[idx].ref_frame;
1640     }
1641     if (ref_frame > usable_ref_frame) continue;
1642     if (skip_ref_find_pred[ref_frame]) continue;
1643 
1644     // For SVC, skip the golden (spatial) reference search if sse of zeromv_last
1645     // is below threshold.
1646     if (cpi->use_svc && ref_frame == GOLDEN_FRAME &&
1647         sse_zeromv_normalized < thresh_svc_skip_golden)
1648       continue;
1649 
1650     if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
1651         this_mode != NEARESTMV) {
1652       continue;
1653     }
1654 
1655     if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode))) continue;
1656 
1657     if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
1658       if (cpi->rc.is_src_frame_alt_ref &&
1659           (ref_frame != ALTREF_FRAME ||
1660            frame_mv[this_mode][ref_frame].as_int != 0))
1661         continue;
1662 
1663       if (cpi->rc.alt_ref_gf_group &&
1664           cpi->rc.frames_since_golden > (cpi->rc.baseline_gf_interval >> 1) &&
1665           ref_frame == GOLDEN_FRAME &&
1666           frame_mv[this_mode][ref_frame].as_int != 0)
1667         continue;
1668 
1669       if (cpi->rc.alt_ref_gf_group &&
1670           cpi->rc.frames_since_golden < (cpi->rc.baseline_gf_interval >> 1) &&
1671           ref_frame == ALTREF_FRAME &&
1672           frame_mv[this_mode][ref_frame].as_int != 0)
1673         continue;
1674     }
1675 
1676     if (!(cpi->ref_frame_flags & flag_list[ref_frame])) continue;
1677 
1678     if (const_motion[ref_frame] && this_mode == NEARMV) continue;
1679 
1680     // Skip non-zeromv mode search for golden frame if force_skip_low_temp_var
1681     // is set. If nearestmv for golden frame is 0, zeromv mode will be skipped
1682     // later.
1683     if (force_skip_low_temp_var && ref_frame == GOLDEN_FRAME &&
1684         frame_mv[this_mode][ref_frame].as_int != 0) {
1685       continue;
1686     }
1687 
1688     if ((cpi->sf.short_circuit_low_temp_var >= 2 ||
1689          (cpi->sf.short_circuit_low_temp_var == 1 && bsize == BLOCK_64X64)) &&
1690         force_skip_low_temp_var && ref_frame == LAST_FRAME &&
1691         this_mode == NEWMV) {
1692       continue;
1693     }
1694 
1695     if (cpi->use_svc) {
1696       if (svc_force_zero_mode[ref_frame - 1] &&
1697           frame_mv[this_mode][ref_frame].as_int != 0)
1698         continue;
1699     }
1700 
1701     if (sf->reference_masking &&
1702         !(frame_mv[this_mode][ref_frame].as_int == 0 &&
1703           ref_frame == LAST_FRAME)) {
1704       if (usable_ref_frame < ALTREF_FRAME) {
1705         if (!force_skip_low_temp_var && usable_ref_frame > LAST_FRAME) {
1706           i = (ref_frame == LAST_FRAME) ? GOLDEN_FRAME : LAST_FRAME;
1707           if ((cpi->ref_frame_flags & flag_list[i]))
1708             if (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[i] << 1))
1709               ref_frame_skip_mask |= (1 << ref_frame);
1710         }
1711       } else if (!cpi->rc.is_src_frame_alt_ref &&
1712                  !(frame_mv[this_mode][ref_frame].as_int == 0 &&
1713                    ref_frame == ALTREF_FRAME)) {
1714         int ref1 = (ref_frame == GOLDEN_FRAME) ? LAST_FRAME : GOLDEN_FRAME;
1715         int ref2 = (ref_frame == ALTREF_FRAME) ? LAST_FRAME : ALTREF_FRAME;
1716         if (((cpi->ref_frame_flags & flag_list[ref1]) &&
1717              (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref1] << 1))) ||
1718             ((cpi->ref_frame_flags & flag_list[ref2]) &&
1719              (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref2] << 1))))
1720           ref_frame_skip_mask |= (1 << ref_frame);
1721       }
1722     }
1723     if (ref_frame_skip_mask & (1 << ref_frame)) continue;
1724 
1725     // Select prediction reference frames.
1726     for (i = 0; i < MAX_MB_PLANE; i++)
1727       xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
1728 
1729     mi->ref_frame[0] = ref_frame;
1730     set_ref_ptrs(cm, xd, ref_frame, NONE);
1731 
1732     mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
1733     mode_rd_thresh = best_mode_skip_txfm ? rd_threshes[mode_index] << 1
1734                                          : rd_threshes[mode_index];
1735 
1736     // Increase mode_rd_thresh value for GOLDEN_FRAME for improved encoding
1737     // speed with little/no subjective quality loss.
1738     if (cpi->sf.bias_golden && ref_frame == GOLDEN_FRAME &&
1739         cpi->rc.frames_since_golden > 4)
1740       mode_rd_thresh = mode_rd_thresh << 3;
1741 
1742     if ((cpi->sf.adaptive_rd_thresh_row_mt &&
1743          rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh,
1744                                     &rd_thresh_freq_fact[mode_index])) ||
1745         (!cpi->sf.adaptive_rd_thresh_row_mt &&
1746          rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
1747                              &rd_thresh_freq_fact[mode_index])))
1748       continue;
1749 
1750     if (this_mode == NEWMV) {
1751       if (ref_frame > LAST_FRAME && !cpi->use_svc &&
1752           cpi->oxcf.rc_mode == VPX_CBR) {
1753         int tmp_sad;
1754         uint32_t dis;
1755         int cost_list[5];
1756 
1757         if (bsize < BLOCK_16X16) continue;
1758 
1759         tmp_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
1760 
1761         if (tmp_sad > x->pred_mv_sad[LAST_FRAME]) continue;
1762         if (tmp_sad + (num_pels_log2_lookup[bsize] << 4) > best_pred_sad)
1763           continue;
1764 
1765         frame_mv[NEWMV][ref_frame].as_int = mi->mv[0].as_int;
1766         rate_mv = vp9_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv,
1767                                   &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1768                                   x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1769         frame_mv[NEWMV][ref_frame].as_mv.row >>= 3;
1770         frame_mv[NEWMV][ref_frame].as_mv.col >>= 3;
1771 
1772         cpi->find_fractional_mv_step(
1773             x, &frame_mv[NEWMV][ref_frame].as_mv,
1774             &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1775             cpi->common.allow_high_precision_mv, x->errorperbit,
1776             &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
1777             cpi->sf.mv.subpel_iters_per_step, cond_cost_list(cpi, cost_list),
1778             x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref_frame], NULL, 0,
1779             0);
1780       } else if (svc->use_base_mv && svc->spatial_layer_id) {
1781         if (frame_mv[NEWMV][ref_frame].as_int != INVALID_MV) {
1782           const int pre_stride = xd->plane[0].pre[0].stride;
1783           int base_mv_sad = INT_MAX;
1784           const float base_mv_bias = sf->base_mv_aggressive ? 1.5f : 1.0f;
1785           const uint8_t *const pre_buf =
1786               xd->plane[0].pre[0].buf +
1787               (frame_mv[NEWMV][ref_frame].as_mv.row >> 3) * pre_stride +
1788               (frame_mv[NEWMV][ref_frame].as_mv.col >> 3);
1789           base_mv_sad = cpi->fn_ptr[bsize].sdf(
1790               x->plane[0].src.buf, x->plane[0].src.stride, pre_buf, pre_stride);
1791 
1792           if (base_mv_sad < (int)(base_mv_bias * x->pred_mv_sad[ref_frame])) {
1793             // Base layer mv is good.
1794             if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1795                                         &frame_mv[NEWMV][ref_frame], &rate_mv,
1796                                         best_rdc.rdcost, 1)) {
1797               continue;
1798             }
1799           } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1800                                              &frame_mv[NEWMV][ref_frame],
1801                                              &rate_mv, best_rdc.rdcost, 0)) {
1802             continue;
1803           }
1804         } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1805                                            &frame_mv[NEWMV][ref_frame],
1806                                            &rate_mv, best_rdc.rdcost, 0)) {
1807           continue;
1808         }
1809       } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1810                                          &frame_mv[NEWMV][ref_frame], &rate_mv,
1811                                          best_rdc.rdcost, 0)) {
1812         continue;
1813       }
1814     }
1815 
1816     // If use_golden_nonzeromv is false, NEWMV mode is skipped for golden, no
1817     // need to compute best_pred_sad which is only used to skip golden NEWMV.
1818     if (use_golden_nonzeromv && this_mode == NEWMV && ref_frame == LAST_FRAME &&
1819         frame_mv[NEWMV][LAST_FRAME].as_int != INVALID_MV) {
1820       const int pre_stride = xd->plane[0].pre[0].stride;
1821       const uint8_t *const pre_buf =
1822           xd->plane[0].pre[0].buf +
1823           (frame_mv[NEWMV][LAST_FRAME].as_mv.row >> 3) * pre_stride +
1824           (frame_mv[NEWMV][LAST_FRAME].as_mv.col >> 3);
1825       best_pred_sad = cpi->fn_ptr[bsize].sdf(
1826           x->plane[0].src.buf, x->plane[0].src.stride, pre_buf, pre_stride);
1827       x->pred_mv_sad[LAST_FRAME] = best_pred_sad;
1828     }
1829 
1830     if (this_mode != NEARESTMV &&
1831         frame_mv[this_mode][ref_frame].as_int ==
1832             frame_mv[NEARESTMV][ref_frame].as_int)
1833       continue;
1834 
1835     mi->mode = this_mode;
1836     mi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
1837 
1838     // Search for the best prediction filter type, when the resulting
1839     // motion vector is at sub-pixel accuracy level for luma component, i.e.,
1840     // the last three bits are all zeros.
1841     if (reuse_inter_pred) {
1842       if (!this_mode_pred) {
1843         this_mode_pred = &tmp[3];
1844       } else {
1845         this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
1846         pd->dst.buf = this_mode_pred->data;
1847         pd->dst.stride = bw;
1848       }
1849     }
1850 
1851     if ((this_mode == NEWMV || filter_ref == SWITCHABLE) &&
1852         pred_filter_search &&
1853         (ref_frame == LAST_FRAME ||
1854          (ref_frame == GOLDEN_FRAME &&
1855           (cpi->use_svc || cpi->oxcf.rc_mode == VPX_VBR))) &&
1856         (((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07) != 0)) {
1857       int pf_rate[3];
1858       int64_t pf_dist[3];
1859       int curr_rate[3];
1860       unsigned int pf_var[3];
1861       unsigned int pf_sse[3];
1862       TX_SIZE pf_tx_size[3];
1863       int64_t best_cost = INT64_MAX;
1864       INTERP_FILTER best_filter = SWITCHABLE, filter;
1865       PRED_BUFFER *current_pred = this_mode_pred;
1866       rd_computed = 1;
1867 
1868       for (filter = EIGHTTAP; filter <= EIGHTTAP_SMOOTH; ++filter) {
1869         int64_t cost;
1870         mi->interp_filter = filter;
1871         vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1872         model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter], &pf_dist[filter],
1873                           &pf_var[filter], &pf_sse[filter]);
1874         curr_rate[filter] = pf_rate[filter];
1875         pf_rate[filter] += vp9_get_switchable_rate(cpi, xd);
1876         cost = RDCOST(x->rdmult, x->rddiv, pf_rate[filter], pf_dist[filter]);
1877         pf_tx_size[filter] = mi->tx_size;
1878         if (cost < best_cost) {
1879           best_filter = filter;
1880           best_cost = cost;
1881           skip_txfm = x->skip_txfm[0];
1882 
1883           if (reuse_inter_pred) {
1884             if (this_mode_pred != current_pred) {
1885               free_pred_buffer(this_mode_pred);
1886               this_mode_pred = current_pred;
1887             }
1888             current_pred = &tmp[get_pred_buffer(tmp, 3)];
1889             pd->dst.buf = current_pred->data;
1890             pd->dst.stride = bw;
1891           }
1892         }
1893       }
1894 
1895       if (reuse_inter_pred && this_mode_pred != current_pred)
1896         free_pred_buffer(current_pred);
1897 
1898       mi->interp_filter = best_filter;
1899       mi->tx_size = pf_tx_size[best_filter];
1900       this_rdc.rate = curr_rate[best_filter];
1901       this_rdc.dist = pf_dist[best_filter];
1902       var_y = pf_var[best_filter];
1903       sse_y = pf_sse[best_filter];
1904       x->skip_txfm[0] = skip_txfm;
1905       if (reuse_inter_pred) {
1906         pd->dst.buf = this_mode_pred->data;
1907         pd->dst.stride = this_mode_pred->stride;
1908       }
1909     } else {
1910       const int large_block = (x->sb_is_skin || cpi->oxcf.speed < 7)
1911                                   ? bsize > BLOCK_32X32
1912                                   : bsize >= BLOCK_32X32;
1913       mi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP : filter_ref;
1914 
1915       if (cpi->use_svc && ref_frame == GOLDEN_FRAME &&
1916           svc_force_zero_mode[ref_frame - 1])
1917         mi->interp_filter = filter_gf_svc;
1918 
1919       vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1920 
1921       // For large partition blocks, extra testing is done.
1922       if (cpi->oxcf.rc_mode == VPX_CBR && large_block &&
1923           !cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id) &&
1924           cm->base_qindex) {
1925         model_rd_for_sb_y_large(cpi, bsize, x, xd, &this_rdc.rate,
1926                                 &this_rdc.dist, &var_y, &sse_y, mi_row, mi_col,
1927                                 &this_early_term);
1928       } else {
1929         rd_computed = 1;
1930         model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
1931                           &var_y, &sse_y);
1932       }
1933       // Save normalized sse (between current and last frame) for (0, 0) motion.
1934       if (cpi->use_svc && ref_frame == LAST_FRAME &&
1935           frame_mv[this_mode][ref_frame].as_int == 0) {
1936         sse_zeromv_normalized =
1937             sse_y >> (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
1938       }
1939     }
1940 
1941     if (!this_early_term) {
1942       this_sse = (int64_t)sse_y;
1943       block_yrd(cpi, x, &this_rdc, &is_skippable, &this_sse, bsize,
1944                 VPXMIN(mi->tx_size, TX_16X16), rd_computed);
1945 
1946       x->skip_txfm[0] = is_skippable;
1947       if (is_skippable) {
1948         this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1949       } else {
1950         if (RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist) <
1951             RDCOST(x->rdmult, x->rddiv, 0, this_sse)) {
1952           this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
1953         } else {
1954           this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1955           this_rdc.dist = this_sse;
1956           x->skip_txfm[0] = SKIP_TXFM_AC_DC;
1957         }
1958       }
1959 
1960       if (cm->interp_filter == SWITCHABLE) {
1961         if ((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07)
1962           this_rdc.rate += vp9_get_switchable_rate(cpi, xd);
1963       }
1964     } else {
1965       this_rdc.rate += cm->interp_filter == SWITCHABLE
1966                            ? vp9_get_switchable_rate(cpi, xd)
1967                            : 0;
1968       this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1969     }
1970 
1971     if (x->color_sensitivity[0] || x->color_sensitivity[1]) {
1972       RD_COST rdc_uv;
1973       const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, &xd->plane[1]);
1974       if (x->color_sensitivity[0])
1975         vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 1);
1976       if (x->color_sensitivity[1])
1977         vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 2);
1978       model_rd_for_sb_uv(cpi, uv_bsize, x, xd, &rdc_uv, &var_y, &sse_y, 1, 2);
1979       this_rdc.rate += rdc_uv.rate;
1980       this_rdc.dist += rdc_uv.dist;
1981     }
1982 
1983     this_rdc.rate += rate_mv;
1984     this_rdc.rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
1985                                          [INTER_OFFSET(this_mode)];
1986     this_rdc.rate += ref_frame_cost[ref_frame];
1987     this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1988 
1989     // Bias against NEWMV that is very different from its neighbors, and bias
1990     // to small motion-lastref for noisy input.
1991     if (cpi->oxcf.rc_mode == VPX_CBR && cpi->oxcf.speed >= 5 &&
1992         cpi->oxcf.content != VP9E_CONTENT_SCREEN) {
1993       vp9_NEWMV_diff_bias(&cpi->noise_estimate, xd, this_mode, &this_rdc, bsize,
1994                           frame_mv[this_mode][ref_frame].as_mv.row,
1995                           frame_mv[this_mode][ref_frame].as_mv.col,
1996                           ref_frame == LAST_FRAME, x->lowvar_highsumdiff,
1997                           x->sb_is_skin);
1998     }
1999 
2000     // Skipping checking: test to see if this block can be reconstructed by
2001     // prediction only.
2002     if (cpi->allow_encode_breakout) {
2003       encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame, this_mode,
2004                            var_y, sse_y, yv12_mb, &this_rdc.rate,
2005                            &this_rdc.dist);
2006       if (x->skip) {
2007         this_rdc.rate += rate_mv;
2008         this_rdc.rdcost =
2009             RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2010       }
2011     }
2012 
2013 #if CONFIG_VP9_TEMPORAL_DENOISING
2014     if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc_pickmode &&
2015         cpi->denoiser.denoising_level > kDenLowLow) {
2016       vp9_denoiser_update_frame_stats(mi, sse_y, this_mode, ctx);
2017       // Keep track of zero_last cost.
2018       if (ref_frame == LAST_FRAME && frame_mv[this_mode][ref_frame].as_int == 0)
2019         zero_last_cost_orig = this_rdc.rdcost;
2020     }
2021 #else
2022     (void)ctx;
2023 #endif
2024 
2025     if (this_rdc.rdcost < best_rdc.rdcost || x->skip) {
2026       best_rdc = this_rdc;
2027       best_mode = this_mode;
2028       best_pred_filter = mi->interp_filter;
2029       best_tx_size = mi->tx_size;
2030       best_ref_frame = ref_frame;
2031       best_mode_skip_txfm = x->skip_txfm[0];
2032       best_early_term = this_early_term;
2033 
2034       if (reuse_inter_pred) {
2035         free_pred_buffer(best_pred);
2036         best_pred = this_mode_pred;
2037       }
2038     } else {
2039       if (reuse_inter_pred) free_pred_buffer(this_mode_pred);
2040     }
2041 
2042     if (x->skip) break;
2043 
2044     // If early termination flag is 1 and at least 2 modes are checked,
2045     // the mode search is terminated.
2046     if (best_early_term && idx > 0) {
2047       x->skip = 1;
2048       break;
2049     }
2050   }
2051 
2052   mi->mode = best_mode;
2053   mi->interp_filter = best_pred_filter;
2054   mi->tx_size = best_tx_size;
2055   mi->ref_frame[0] = best_ref_frame;
2056   mi->mv[0].as_int = frame_mv[best_mode][best_ref_frame].as_int;
2057   xd->mi[0]->bmi[0].as_mv[0].as_int = mi->mv[0].as_int;
2058   x->skip_txfm[0] = best_mode_skip_txfm;
2059 
2060   // For spatial enhancemanent layer: perform intra prediction only if base
2061   // layer is chosen as the reference. Always perform intra prediction if
2062   // LAST is the only reference or is_key_frame is set.
2063   if (cpi->svc.spatial_layer_id) {
2064     perform_intra_pred =
2065         cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame ||
2066         !(cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) ||
2067         (!cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
2068          svc_force_zero_mode[best_ref_frame - 1]);
2069     inter_mode_thresh = (inter_mode_thresh << 1) + inter_mode_thresh;
2070   }
2071   if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR &&
2072       cpi->rc.is_src_frame_alt_ref)
2073     perform_intra_pred = 0;
2074   // Perform intra prediction search, if the best SAD is above a certain
2075   // threshold.
2076   if (best_rdc.rdcost == INT64_MAX ||
2077       ((!force_skip_low_temp_var || bsize < BLOCK_32X32) &&
2078        perform_intra_pred && !x->skip && best_rdc.rdcost > inter_mode_thresh &&
2079        bsize <= cpi->sf.max_intra_bsize && !x->skip_low_source_sad &&
2080        !x->lowvar_highsumdiff)) {
2081     struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
2082     int i;
2083     TX_SIZE best_intra_tx_size = TX_SIZES;
2084     TX_SIZE intra_tx_size =
2085         VPXMIN(max_txsize_lookup[bsize],
2086                tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
2087     if (cpi->oxcf.content != VP9E_CONTENT_SCREEN && intra_tx_size > TX_16X16)
2088       intra_tx_size = TX_16X16;
2089 
2090     if (reuse_inter_pred && best_pred != NULL) {
2091       if (best_pred->data == orig_dst.buf) {
2092         this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
2093 #if CONFIG_VP9_HIGHBITDEPTH
2094         if (cm->use_highbitdepth)
2095           vpx_highbd_convolve_copy(
2096               CONVERT_TO_SHORTPTR(best_pred->data), best_pred->stride,
2097               CONVERT_TO_SHORTPTR(this_mode_pred->data), this_mode_pred->stride,
2098               NULL, 0, NULL, 0, bw, bh, xd->bd);
2099         else
2100           vpx_convolve_copy(best_pred->data, best_pred->stride,
2101                             this_mode_pred->data, this_mode_pred->stride, NULL,
2102                             0, NULL, 0, bw, bh);
2103 #else
2104         vpx_convolve_copy(best_pred->data, best_pred->stride,
2105                           this_mode_pred->data, this_mode_pred->stride, NULL, 0,
2106                           NULL, 0, bw, bh);
2107 #endif  // CONFIG_VP9_HIGHBITDEPTH
2108         best_pred = this_mode_pred;
2109       }
2110     }
2111     pd->dst = orig_dst;
2112 
2113     for (i = 0; i < 4; ++i) {
2114       const PREDICTION_MODE this_mode = intra_mode_list[i];
2115       THR_MODES mode_index = mode_idx[INTRA_FRAME][mode_offset(this_mode)];
2116       int mode_rd_thresh = rd_threshes[mode_index];
2117       if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
2118           this_mode != DC_PRED) {
2119         continue;
2120       }
2121 
2122       if (!((1 << this_mode) & cpi->sf.intra_y_mode_bsize_mask[bsize]))
2123         continue;
2124 
2125       if ((cpi->sf.adaptive_rd_thresh_row_mt &&
2126            rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh,
2127                                       &rd_thresh_freq_fact[mode_index])) ||
2128           (!cpi->sf.adaptive_rd_thresh_row_mt &&
2129            rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
2130                                &rd_thresh_freq_fact[mode_index])))
2131         continue;
2132 
2133       mi->mode = this_mode;
2134       mi->ref_frame[0] = INTRA_FRAME;
2135       this_rdc.dist = this_rdc.rate = 0;
2136       args.mode = this_mode;
2137       args.skippable = 1;
2138       args.rdc = &this_rdc;
2139       mi->tx_size = intra_tx_size;
2140       vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
2141                                              &args);
2142       // Check skip cost here since skippable is not set for for uv, this
2143       // mirrors the behavior used by inter
2144       if (args.skippable) {
2145         x->skip_txfm[0] = SKIP_TXFM_AC_DC;
2146         this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
2147       } else {
2148         x->skip_txfm[0] = SKIP_TXFM_NONE;
2149         this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
2150       }
2151       // Inter and intra RD will mismatch in scale for non-screen content.
2152       if (cpi->oxcf.content == VP9E_CONTENT_SCREEN) {
2153         if (x->color_sensitivity[0])
2154           vp9_foreach_transformed_block_in_plane(xd, bsize, 1,
2155                                                  estimate_block_intra, &args);
2156         if (x->color_sensitivity[1])
2157           vp9_foreach_transformed_block_in_plane(xd, bsize, 2,
2158                                                  estimate_block_intra, &args);
2159       }
2160       this_rdc.rate += cpi->mbmode_cost[this_mode];
2161       this_rdc.rate += ref_frame_cost[INTRA_FRAME];
2162       this_rdc.rate += intra_cost_penalty;
2163       this_rdc.rdcost =
2164           RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2165 
2166       if (this_rdc.rdcost < best_rdc.rdcost) {
2167         best_rdc = this_rdc;
2168         best_mode = this_mode;
2169         best_intra_tx_size = mi->tx_size;
2170         best_ref_frame = INTRA_FRAME;
2171         mi->uv_mode = this_mode;
2172         mi->mv[0].as_int = INVALID_MV;
2173         best_mode_skip_txfm = x->skip_txfm[0];
2174       }
2175     }
2176 
2177     // Reset mb_mode_info to the best inter mode.
2178     if (best_ref_frame != INTRA_FRAME) {
2179       mi->tx_size = best_tx_size;
2180     } else {
2181       mi->tx_size = best_intra_tx_size;
2182     }
2183   }
2184 
2185   pd->dst = orig_dst;
2186   mi->mode = best_mode;
2187   mi->ref_frame[0] = best_ref_frame;
2188   x->skip_txfm[0] = best_mode_skip_txfm;
2189 
2190   if (!is_inter_block(mi)) {
2191     mi->interp_filter = SWITCHABLE_FILTERS;
2192   }
2193 
2194   if (reuse_inter_pred && best_pred != NULL) {
2195     if (best_pred->data != orig_dst.buf && is_inter_mode(mi->mode)) {
2196 #if CONFIG_VP9_HIGHBITDEPTH
2197       if (cm->use_highbitdepth)
2198         vpx_highbd_convolve_copy(
2199             CONVERT_TO_SHORTPTR(best_pred->data), best_pred->stride,
2200             CONVERT_TO_SHORTPTR(pd->dst.buf), pd->dst.stride, NULL, 0, NULL, 0,
2201             bw, bh, xd->bd);
2202       else
2203         vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2204                           pd->dst.stride, NULL, 0, NULL, 0, bw, bh);
2205 #else
2206       vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2207                         pd->dst.stride, NULL, 0, NULL, 0, bw, bh);
2208 #endif  // CONFIG_VP9_HIGHBITDEPTH
2209     }
2210   }
2211 
2212 #if CONFIG_VP9_TEMPORAL_DENOISING
2213   if (cpi->oxcf.noise_sensitivity > 0 && cpi->resize_pending == 0 &&
2214       denoise_svc_pickmode && cpi->denoiser.denoising_level > kDenLowLow &&
2215       cpi->denoiser.reset == 0) {
2216     VP9_DENOISER_DECISION decision = COPY_BLOCK;
2217     vp9_pickmode_ctx_den_update(&ctx_den, zero_last_cost_orig, ref_frame_cost,
2218                                 frame_mv, reuse_inter_pred, best_tx_size,
2219                                 best_mode, best_ref_frame, best_pred_filter,
2220                                 best_mode_skip_txfm);
2221     vp9_denoiser_denoise(cpi, x, mi_row, mi_col, bsize, ctx, &decision);
2222     recheck_zeromv_after_denoising(cpi, mi, x, xd, decision, &ctx_den, yv12_mb,
2223                                    &best_rdc, bsize, mi_row, mi_col);
2224     best_ref_frame = ctx_den.best_ref_frame;
2225   }
2226 #endif
2227 
2228   if (cpi->sf.adaptive_rd_thresh) {
2229     THR_MODES best_mode_idx = mode_idx[best_ref_frame][mode_offset(mi->mode)];
2230 
2231     if (best_ref_frame == INTRA_FRAME) {
2232       // Only consider the modes that are included in the intra_mode_list.
2233       int intra_modes = sizeof(intra_mode_list) / sizeof(PREDICTION_MODE);
2234       int i;
2235 
2236       // TODO(yunqingwang): Check intra mode mask and only update freq_fact
2237       // for those valid modes.
2238       for (i = 0; i < intra_modes; i++) {
2239         if (cpi->sf.adaptive_rd_thresh_row_mt)
2240           update_thresh_freq_fact_row_mt(cpi, tile_data, x->source_variance,
2241                                          thresh_freq_fact_idx, INTRA_FRAME,
2242                                          best_mode_idx, intra_mode_list[i]);
2243         else
2244           update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2245                                   INTRA_FRAME, best_mode_idx,
2246                                   intra_mode_list[i]);
2247       }
2248     } else {
2249       for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2250         PREDICTION_MODE this_mode;
2251         if (best_ref_frame != ref_frame) continue;
2252         for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2253           if (cpi->sf.adaptive_rd_thresh_row_mt)
2254             update_thresh_freq_fact_row_mt(cpi, tile_data, x->source_variance,
2255                                            thresh_freq_fact_idx, ref_frame,
2256                                            best_mode_idx, this_mode);
2257           else
2258             update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2259                                     ref_frame, best_mode_idx, this_mode);
2260         }
2261       }
2262     }
2263   }
2264 
2265   *rd_cost = best_rdc;
2266 }
2267 
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)2268 void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x, int mi_row,
2269                                 int mi_col, RD_COST *rd_cost, BLOCK_SIZE bsize,
2270                                 PICK_MODE_CONTEXT *ctx) {
2271   VP9_COMMON *const cm = &cpi->common;
2272   SPEED_FEATURES *const sf = &cpi->sf;
2273   MACROBLOCKD *const xd = &x->e_mbd;
2274   MODE_INFO *const mi = xd->mi[0];
2275   MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2276   const struct segmentation *const seg = &cm->seg;
2277   MV_REFERENCE_FRAME ref_frame, second_ref_frame = NONE;
2278   MV_REFERENCE_FRAME best_ref_frame = NONE;
2279   unsigned char segment_id = mi->segment_id;
2280   struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2281   static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2282                                     VP9_ALT_FLAG };
2283   int64_t best_rd = INT64_MAX;
2284   b_mode_info bsi[MAX_REF_FRAMES][4];
2285   int ref_frame_skip_mask = 0;
2286   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
2287   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
2288   int idx, idy;
2289 
2290   x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
2291   ctx->pred_pixel_ready = 0;
2292 
2293   for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2294     const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2295     int_mv dummy_mv[2];
2296     x->pred_mv_sad[ref_frame] = INT_MAX;
2297 
2298     if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
2299       int_mv *const candidates = mbmi_ext->ref_mvs[ref_frame];
2300       const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2301       vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf,
2302                            sf);
2303       vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
2304                        mbmi_ext->mode_context);
2305 
2306       vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2307                             &dummy_mv[0], &dummy_mv[1]);
2308     } else {
2309       ref_frame_skip_mask |= (1 << ref_frame);
2310     }
2311   }
2312 
2313   mi->sb_type = bsize;
2314   mi->tx_size = TX_4X4;
2315   mi->uv_mode = DC_PRED;
2316   mi->ref_frame[0] = LAST_FRAME;
2317   mi->ref_frame[1] = NONE;
2318   mi->interp_filter =
2319       cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
2320 
2321   for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2322     int64_t this_rd = 0;
2323     int plane;
2324 
2325     if (ref_frame_skip_mask & (1 << ref_frame)) continue;
2326 
2327 #if CONFIG_BETTER_HW_COMPATIBILITY
2328     if ((bsize == BLOCK_8X4 || bsize == BLOCK_4X8) && ref_frame > INTRA_FRAME &&
2329         vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2330       continue;
2331 #endif
2332 
2333     // TODO(jingning, agrange): Scaling reference frame not supported for
2334     // sub8x8 blocks. Is this supported now?
2335     if (ref_frame > INTRA_FRAME &&
2336         vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2337       continue;
2338 
2339     // If the segment reference frame feature is enabled....
2340     // then do nothing if the current ref frame is not allowed..
2341     if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
2342         get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
2343       continue;
2344 
2345     mi->ref_frame[0] = ref_frame;
2346     x->skip = 0;
2347     set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
2348 
2349     // Select prediction reference frames.
2350     for (plane = 0; plane < MAX_MB_PLANE; plane++)
2351       xd->plane[plane].pre[0] = yv12_mb[ref_frame][plane];
2352 
2353     for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2354       for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2355         int_mv b_mv[MB_MODE_COUNT];
2356         int64_t b_best_rd = INT64_MAX;
2357         const int i = idy * 2 + idx;
2358         PREDICTION_MODE this_mode;
2359         RD_COST this_rdc;
2360         unsigned int var_y, sse_y;
2361 
2362         struct macroblock_plane *p = &x->plane[0];
2363         struct macroblockd_plane *pd = &xd->plane[0];
2364 
2365         const struct buf_2d orig_src = p->src;
2366         const struct buf_2d orig_dst = pd->dst;
2367         struct buf_2d orig_pre[2];
2368         memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre));
2369 
2370         // set buffer pointers for sub8x8 motion search.
2371         p->src.buf =
2372             &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
2373         pd->dst.buf =
2374             &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
2375         pd->pre[0].buf =
2376             &pd->pre[0]
2377                  .buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
2378 
2379         b_mv[ZEROMV].as_int = 0;
2380         b_mv[NEWMV].as_int = INVALID_MV;
2381         vp9_append_sub8x8_mvs_for_idx(cm, xd, i, 0, mi_row, mi_col,
2382                                       &b_mv[NEARESTMV], &b_mv[NEARMV],
2383                                       mbmi_ext->mode_context);
2384 
2385         for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2386           int b_rate = 0;
2387           xd->mi[0]->bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int;
2388 
2389           if (this_mode == NEWMV) {
2390             const int step_param = cpi->sf.mv.fullpel_search_step_param;
2391             MV mvp_full;
2392             MV tmp_mv;
2393             int cost_list[5];
2394             const MvLimits tmp_mv_limits = x->mv_limits;
2395             uint32_t dummy_dist;
2396 
2397             if (i == 0) {
2398               mvp_full.row = b_mv[NEARESTMV].as_mv.row >> 3;
2399               mvp_full.col = b_mv[NEARESTMV].as_mv.col >> 3;
2400             } else {
2401               mvp_full.row = xd->mi[0]->bmi[0].as_mv[0].as_mv.row >> 3;
2402               mvp_full.col = xd->mi[0]->bmi[0].as_mv[0].as_mv.col >> 3;
2403             }
2404 
2405             vp9_set_mv_search_range(&x->mv_limits,
2406                                     &mbmi_ext->ref_mvs[ref_frame][0].as_mv);
2407 
2408             vp9_full_pixel_search(
2409                 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method,
2410                 x->sadperbit4, cond_cost_list(cpi, cost_list),
2411                 &mbmi_ext->ref_mvs[ref_frame][0].as_mv, &tmp_mv, INT_MAX, 0);
2412 
2413             x->mv_limits = tmp_mv_limits;
2414 
2415             // calculate the bit cost on motion vector
2416             mvp_full.row = tmp_mv.row * 8;
2417             mvp_full.col = tmp_mv.col * 8;
2418 
2419             b_rate += vp9_mv_bit_cost(
2420                 &mvp_full, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2421                 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2422 
2423             b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2424                                           [INTER_OFFSET(NEWMV)];
2425             if (RDCOST(x->rdmult, x->rddiv, b_rate, 0) > b_best_rd) continue;
2426 
2427             cpi->find_fractional_mv_step(
2428                 x, &tmp_mv, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2429                 cpi->common.allow_high_precision_mv, x->errorperbit,
2430                 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
2431                 cpi->sf.mv.subpel_iters_per_step,
2432                 cond_cost_list(cpi, cost_list), x->nmvjointcost, x->mvcost,
2433                 &dummy_dist, &x->pred_sse[ref_frame], NULL, 0, 0);
2434 
2435             xd->mi[0]->bmi[i].as_mv[0].as_mv = tmp_mv;
2436           } else {
2437             b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2438                                           [INTER_OFFSET(this_mode)];
2439           }
2440 
2441 #if CONFIG_VP9_HIGHBITDEPTH
2442           if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2443             vp9_highbd_build_inter_predictor(
2444                 CONVERT_TO_SHORTPTR(pd->pre[0].buf), pd->pre[0].stride,
2445                 CONVERT_TO_SHORTPTR(pd->dst.buf), pd->dst.stride,
2446                 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2447                 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2448                 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2449                 mi_col * MI_SIZE + 4 * (i & 0x01),
2450                 mi_row * MI_SIZE + 4 * (i >> 1), xd->bd);
2451           } else {
2452 #endif
2453             vp9_build_inter_predictor(
2454                 pd->pre[0].buf, pd->pre[0].stride, pd->dst.buf, pd->dst.stride,
2455                 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2456                 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2457                 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2458                 mi_col * MI_SIZE + 4 * (i & 0x01),
2459                 mi_row * MI_SIZE + 4 * (i >> 1));
2460 
2461 #if CONFIG_VP9_HIGHBITDEPTH
2462           }
2463 #endif
2464 
2465           model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
2466                             &var_y, &sse_y);
2467 
2468           this_rdc.rate += b_rate;
2469           this_rdc.rdcost =
2470               RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2471           if (this_rdc.rdcost < b_best_rd) {
2472             b_best_rd = this_rdc.rdcost;
2473             bsi[ref_frame][i].as_mode = this_mode;
2474             bsi[ref_frame][i].as_mv[0].as_mv = xd->mi[0]->bmi[i].as_mv[0].as_mv;
2475           }
2476         }  // mode search
2477 
2478         // restore source and prediction buffer pointers.
2479         p->src = orig_src;
2480         pd->pre[0] = orig_pre[0];
2481         pd->dst = orig_dst;
2482         this_rd += b_best_rd;
2483 
2484         xd->mi[0]->bmi[i] = bsi[ref_frame][i];
2485         if (num_4x4_blocks_wide > 1) xd->mi[0]->bmi[i + 1] = xd->mi[0]->bmi[i];
2486         if (num_4x4_blocks_high > 1) xd->mi[0]->bmi[i + 2] = xd->mi[0]->bmi[i];
2487       }
2488     }  // loop through sub8x8 blocks
2489 
2490     if (this_rd < best_rd) {
2491       best_rd = this_rd;
2492       best_ref_frame = ref_frame;
2493     }
2494   }  // reference frames
2495 
2496   mi->tx_size = TX_4X4;
2497   mi->ref_frame[0] = best_ref_frame;
2498   for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2499     for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2500       const int block = idy * 2 + idx;
2501       xd->mi[0]->bmi[block] = bsi[best_ref_frame][block];
2502       if (num_4x4_blocks_wide > 1)
2503         xd->mi[0]->bmi[block + 1] = bsi[best_ref_frame][block];
2504       if (num_4x4_blocks_high > 1)
2505         xd->mi[0]->bmi[block + 2] = bsi[best_ref_frame][block];
2506     }
2507   }
2508   mi->mode = xd->mi[0]->bmi[3].as_mode;
2509   ctx->mic = *(xd->mi[0]);
2510   ctx->mbmi_ext = *x->mbmi_ext;
2511   ctx->skip_txfm[0] = SKIP_TXFM_NONE;
2512   ctx->skip = 0;
2513   // Dummy assignment for speed -5. No effect in speed -6.
2514   rd_cost->rdcost = best_rd;
2515 }
2516