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/compiler_attributes.h"
23 
24 #include "vp9/common/vp9_blockd.h"
25 #include "vp9/common/vp9_common.h"
26 #include "vp9/common/vp9_mvref_common.h"
27 #include "vp9/common/vp9_pred_common.h"
28 #include "vp9/common/vp9_reconinter.h"
29 #include "vp9/common/vp9_reconintra.h"
30 #include "vp9/common/vp9_scan.h"
31 
32 #include "vp9/encoder/vp9_cost.h"
33 #include "vp9/encoder/vp9_encoder.h"
34 #include "vp9/encoder/vp9_pickmode.h"
35 #include "vp9/encoder/vp9_ratectrl.h"
36 #include "vp9/encoder/vp9_rd.h"
37 
38 typedef struct {
39   uint8_t *data;
40   int stride;
41   int in_use;
42 } PRED_BUFFER;
43 
44 typedef struct {
45   PRED_BUFFER *best_pred;
46   PREDICTION_MODE best_mode;
47   TX_SIZE best_tx_size;
48   TX_SIZE best_intra_tx_size;
49   MV_REFERENCE_FRAME best_ref_frame;
50   MV_REFERENCE_FRAME best_second_ref_frame;
51   uint8_t best_mode_skip_txfm;
52   INTERP_FILTER best_pred_filter;
53 } BEST_PICKMODE;
54 
55 static const int pos_shift_16x16[4][4] = {
56   { 9, 10, 13, 14 }, { 11, 12, 15, 16 }, { 17, 18, 21, 22 }, { 19, 20, 23, 24 }
57 };
58 
mv_refs_rt(VP9_COMP * cpi,const VP9_COMMON * cm,const MACROBLOCK * x,const MACROBLOCKD * xd,const TileInfo * const tile,MODE_INFO * mi,MV_REFERENCE_FRAME ref_frame,int_mv * mv_ref_list,int_mv * base_mv,int mi_row,int mi_col,int use_base_mv)59 static int mv_refs_rt(VP9_COMP *cpi, const VP9_COMMON *cm, const MACROBLOCK *x,
60                       const MACROBLOCKD *xd, const TileInfo *const tile,
61                       MODE_INFO *mi, MV_REFERENCE_FRAME ref_frame,
62                       int_mv *mv_ref_list, int_mv *base_mv, int mi_row,
63                       int mi_col, int use_base_mv) {
64   const int *ref_sign_bias = cm->ref_frame_sign_bias;
65   int i, refmv_count = 0;
66 
67   const POSITION *const mv_ref_search = mv_ref_blocks[mi->sb_type];
68 
69   int different_ref_found = 0;
70   int context_counter = 0;
71   int const_motion = 0;
72 
73   // Blank the reference vector list
74   memset(mv_ref_list, 0, sizeof(*mv_ref_list) * MAX_MV_REF_CANDIDATES);
75 
76   // The nearest 2 blocks are treated differently
77   // if the size < 8x8 we get the mv from the bmi substructure,
78   // and we also need to keep a mode count.
79   for (i = 0; i < 2; ++i) {
80     const POSITION *const mv_ref = &mv_ref_search[i];
81     if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
82       const MODE_INFO *const candidate_mi =
83           xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
84       // Keep counts for entropy encoding.
85       context_counter += mode_2_counter[candidate_mi->mode];
86       different_ref_found = 1;
87 
88       if (candidate_mi->ref_frame[0] == ref_frame)
89         ADD_MV_REF_LIST(get_sub_block_mv(candidate_mi, 0, mv_ref->col, -1),
90                         refmv_count, mv_ref_list, Done);
91     }
92   }
93 
94   const_motion = 1;
95 
96   // Check the rest of the neighbors in much the same way
97   // as before except we don't need to keep track of sub blocks or
98   // mode counts.
99   for (; i < MVREF_NEIGHBOURS && !refmv_count; ++i) {
100     const POSITION *const mv_ref = &mv_ref_search[i];
101     if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
102       const MODE_INFO *const candidate_mi =
103           xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
104       different_ref_found = 1;
105 
106       if (candidate_mi->ref_frame[0] == ref_frame)
107         ADD_MV_REF_LIST(candidate_mi->mv[0], refmv_count, mv_ref_list, Done);
108     }
109   }
110 
111   // Since we couldn't find 2 mvs from the same reference frame
112   // go back through the neighbors and find motion vectors from
113   // different reference frames.
114   if (different_ref_found && !refmv_count) {
115     for (i = 0; i < MVREF_NEIGHBOURS; ++i) {
116       const POSITION *mv_ref = &mv_ref_search[i];
117       if (is_inside(tile, mi_col, mi_row, cm->mi_rows, mv_ref)) {
118         const MODE_INFO *const candidate_mi =
119             xd->mi[mv_ref->col + mv_ref->row * xd->mi_stride];
120 
121         // If the candidate is INTRA we don't want to consider its mv.
122         IF_DIFF_REF_FRAME_ADD_MV(candidate_mi, ref_frame, ref_sign_bias,
123                                  refmv_count, mv_ref_list, Done);
124       }
125     }
126   }
127   if (use_base_mv &&
128       !cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
129       ref_frame == LAST_FRAME) {
130     // Get base layer mv.
131     MV_REF *candidate =
132         &cm->prev_frame
133              ->mvs[(mi_col >> 1) + (mi_row >> 1) * (cm->mi_cols >> 1)];
134     if (candidate->mv[0].as_int != INVALID_MV) {
135       base_mv->as_mv.row = (candidate->mv[0].as_mv.row * 2);
136       base_mv->as_mv.col = (candidate->mv[0].as_mv.col * 2);
137       clamp_mv_ref(&base_mv->as_mv, xd);
138     } else {
139       base_mv->as_int = INVALID_MV;
140     }
141   }
142 
143 Done:
144 
145   x->mbmi_ext->mode_context[ref_frame] = counter_to_context[context_counter];
146 
147   // Clamp vectors
148   for (i = 0; i < MAX_MV_REF_CANDIDATES; ++i)
149     clamp_mv_ref(&mv_ref_list[i].as_mv, xd);
150 
151   return const_motion;
152 }
153 
combined_motion_search(VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int mi_row,int mi_col,int_mv * tmp_mv,int * rate_mv,int64_t best_rd_sofar,int use_base_mv)154 static int combined_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
155                                   BLOCK_SIZE bsize, int mi_row, int mi_col,
156                                   int_mv *tmp_mv, int *rate_mv,
157                                   int64_t best_rd_sofar, int use_base_mv) {
158   MACROBLOCKD *xd = &x->e_mbd;
159   MODE_INFO *mi = xd->mi[0];
160   struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0 } };
161   const int step_param = cpi->sf.mv.fullpel_search_step_param;
162   const int sadpb = x->sadperbit16;
163   MV mvp_full;
164   const int ref = mi->ref_frame[0];
165   const MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
166   MV center_mv;
167   uint32_t dis;
168   int rate_mode;
169   const MvLimits tmp_mv_limits = x->mv_limits;
170   int rv = 0;
171   int cost_list[5];
172   int search_subpel = 1;
173   const YV12_BUFFER_CONFIG *scaled_ref_frame =
174       vp9_get_scaled_ref_frame(cpi, ref);
175   if (scaled_ref_frame) {
176     int i;
177     // Swap out the reference frame for a version that's been scaled to
178     // match the resolution of the current frame, allowing the existing
179     // motion search code to be used without additional modifications.
180     for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
181     vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
182   }
183   vp9_set_mv_search_range(&x->mv_limits, &ref_mv);
184 
185   // Limit motion vector for large lightning change.
186   if (cpi->oxcf.speed > 5 && x->lowvar_highsumdiff) {
187     x->mv_limits.col_min = VPXMAX(x->mv_limits.col_min, -10);
188     x->mv_limits.row_min = VPXMAX(x->mv_limits.row_min, -10);
189     x->mv_limits.col_max = VPXMIN(x->mv_limits.col_max, 10);
190     x->mv_limits.row_max = VPXMIN(x->mv_limits.row_max, 10);
191   }
192 
193   assert(x->mv_best_ref_index[ref] <= 2);
194   if (x->mv_best_ref_index[ref] < 2)
195     mvp_full = x->mbmi_ext->ref_mvs[ref][x->mv_best_ref_index[ref]].as_mv;
196   else
197     mvp_full = x->pred_mv[ref];
198 
199   mvp_full.col >>= 3;
200   mvp_full.row >>= 3;
201 
202   if (!use_base_mv)
203     center_mv = ref_mv;
204   else
205     center_mv = tmp_mv->as_mv;
206 
207   if (x->sb_use_mv_part) {
208     tmp_mv->as_mv.row = x->sb_mvrow_part >> 3;
209     tmp_mv->as_mv.col = x->sb_mvcol_part >> 3;
210   } else {
211     vp9_full_pixel_search(
212         cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method, sadpb,
213         cond_cost_list(cpi, cost_list), &center_mv, &tmp_mv->as_mv, INT_MAX, 0);
214   }
215 
216   x->mv_limits = tmp_mv_limits;
217 
218   // calculate the bit cost on motion vector
219   mvp_full.row = tmp_mv->as_mv.row * 8;
220   mvp_full.col = tmp_mv->as_mv.col * 8;
221 
222   *rate_mv = vp9_mv_bit_cost(&mvp_full, &ref_mv, x->nmvjointcost, x->mvcost,
223                              MV_COST_WEIGHT);
224 
225   rate_mode =
226       cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref]][INTER_OFFSET(NEWMV)];
227   rv =
228       !(RDCOST(x->rdmult, x->rddiv, (*rate_mv + rate_mode), 0) > best_rd_sofar);
229 
230   // For SVC on non-reference frame, avoid subpel for (0, 0) motion.
231   if (cpi->use_svc && cpi->svc.non_reference_frame) {
232     if (mvp_full.row == 0 && mvp_full.col == 0) search_subpel = 0;
233   }
234 
235   if (rv && search_subpel) {
236     SUBPEL_FORCE_STOP subpel_force_stop = cpi->sf.mv.subpel_force_stop;
237     if (use_base_mv && cpi->sf.base_mv_aggressive) subpel_force_stop = HALF_PEL;
238     if (cpi->sf.mv.enable_adaptive_subpel_force_stop) {
239       const int mv_thresh = cpi->sf.mv.adapt_subpel_force_stop.mv_thresh;
240       if (abs(tmp_mv->as_mv.row) >= mv_thresh ||
241           abs(tmp_mv->as_mv.col) >= mv_thresh)
242         subpel_force_stop = cpi->sf.mv.adapt_subpel_force_stop.force_stop_above;
243       else
244         subpel_force_stop = cpi->sf.mv.adapt_subpel_force_stop.force_stop_below;
245     }
246     cpi->find_fractional_mv_step(
247         x, &tmp_mv->as_mv, &ref_mv, cpi->common.allow_high_precision_mv,
248         x->errorperbit, &cpi->fn_ptr[bsize], subpel_force_stop,
249         cpi->sf.mv.subpel_search_level, cond_cost_list(cpi, cost_list),
250         x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, 0, 0,
251         cpi->sf.use_accurate_subpel_search);
252     *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmvjointcost,
253                                x->mvcost, MV_COST_WEIGHT);
254   }
255 
256   if (scaled_ref_frame) {
257     int i;
258     for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
259   }
260   return rv;
261 }
262 
block_variance(const uint8_t * src,int src_stride,const uint8_t * ref,int ref_stride,int w,int h,unsigned int * sse,int * sum,int block_size,int use_highbitdepth,vpx_bit_depth_t bd,uint32_t * sse8x8,int * sum8x8,uint32_t * var8x8)263 static void block_variance(const uint8_t *src, int src_stride,
264                            const uint8_t *ref, int ref_stride, int w, int h,
265                            unsigned int *sse, int *sum, int block_size,
266 #if CONFIG_VP9_HIGHBITDEPTH
267                            int use_highbitdepth, vpx_bit_depth_t bd,
268 #endif
269                            uint32_t *sse8x8, int *sum8x8, uint32_t *var8x8) {
270   int i, j, k = 0;
271 
272   *sse = 0;
273   *sum = 0;
274 
275   for (i = 0; i < h; i += block_size) {
276     for (j = 0; j < w; j += block_size) {
277 #if CONFIG_VP9_HIGHBITDEPTH
278       if (use_highbitdepth) {
279         switch (bd) {
280           case VPX_BITS_8:
281             vpx_highbd_8_get8x8var(src + src_stride * i + j, src_stride,
282                                    ref + ref_stride * i + j, ref_stride,
283                                    &sse8x8[k], &sum8x8[k]);
284             break;
285           case VPX_BITS_10:
286             vpx_highbd_10_get8x8var(src + src_stride * i + j, src_stride,
287                                     ref + ref_stride * i + j, ref_stride,
288                                     &sse8x8[k], &sum8x8[k]);
289             break;
290           case VPX_BITS_12:
291             vpx_highbd_12_get8x8var(src + src_stride * i + j, src_stride,
292                                     ref + ref_stride * i + j, ref_stride,
293                                     &sse8x8[k], &sum8x8[k]);
294             break;
295         }
296       } else {
297         vpx_get8x8var(src + src_stride * i + j, src_stride,
298                       ref + ref_stride * i + j, ref_stride, &sse8x8[k],
299                       &sum8x8[k]);
300       }
301 #else
302       vpx_get8x8var(src + src_stride * i + j, src_stride,
303                     ref + ref_stride * i + j, ref_stride, &sse8x8[k],
304                     &sum8x8[k]);
305 #endif
306       *sse += sse8x8[k];
307       *sum += sum8x8[k];
308       var8x8[k] = sse8x8[k] - (uint32_t)(((int64_t)sum8x8[k] * sum8x8[k]) >> 6);
309       k++;
310     }
311   }
312 }
313 
calculate_variance(int bw,int bh,TX_SIZE tx_size,unsigned int * sse_i,int * sum_i,unsigned int * var_o,unsigned int * sse_o,int * sum_o)314 static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
315                                unsigned int *sse_i, int *sum_i,
316                                unsigned int *var_o, unsigned int *sse_o,
317                                int *sum_o) {
318   const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
319   const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
320   const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
321   int i, j, k = 0;
322 
323   for (i = 0; i < nh; i += 2) {
324     for (j = 0; j < nw; j += 2) {
325       sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] +
326                  sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1];
327       sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] +
328                  sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1];
329       var_o[k] = sse_o[k] - (uint32_t)(((int64_t)sum_o[k] * sum_o[k]) >>
330                                        (b_width_log2_lookup[unit_size] +
331                                         b_height_log2_lookup[unit_size] + 6));
332       k++;
333     }
334   }
335 }
336 
337 // Adjust the ac_thr according to speed, width, height and normalized sum
ac_thr_factor(const int speed,const int width,const int height,const int norm_sum)338 static int ac_thr_factor(const int speed, const int width, const int height,
339                          const int norm_sum) {
340   if (speed >= 8 && norm_sum < 5) {
341     if (width <= 640 && height <= 480)
342       return 4;
343     else
344       return 2;
345   }
346   return 1;
347 }
348 
calculate_tx_size(VP9_COMP * const cpi,BLOCK_SIZE bsize,MACROBLOCKD * const xd,unsigned int var,unsigned int sse,int64_t ac_thr,unsigned int source_variance,int is_intra)349 static TX_SIZE calculate_tx_size(VP9_COMP *const cpi, BLOCK_SIZE bsize,
350                                  MACROBLOCKD *const xd, unsigned int var,
351                                  unsigned int sse, int64_t ac_thr,
352                                  unsigned int source_variance, int is_intra) {
353   // TODO(marpan): Tune selection for intra-modes, screen content, etc.
354   TX_SIZE tx_size;
355   unsigned int var_thresh = is_intra ? (unsigned int)ac_thr : 1;
356   int limit_tx = 1;
357   if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
358       (source_variance == 0 || var < var_thresh))
359     limit_tx = 0;
360   if (cpi->common.tx_mode == TX_MODE_SELECT) {
361     if (sse > (var << 2))
362       tx_size = VPXMIN(max_txsize_lookup[bsize],
363                        tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
364     else
365       tx_size = TX_8X8;
366     if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && limit_tx &&
367         cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
368       tx_size = TX_8X8;
369     else if (tx_size > TX_16X16 && limit_tx)
370       tx_size = TX_16X16;
371     // For screen-content force 4X4 tx_size over 8X8, for large variance.
372     if (cpi->oxcf.content == VP9E_CONTENT_SCREEN && tx_size == TX_8X8 &&
373         bsize <= BLOCK_16X16 && ((var >> 5) > (unsigned int)ac_thr))
374       tx_size = TX_4X4;
375   } else {
376     tx_size = VPXMIN(max_txsize_lookup[bsize],
377                      tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
378   }
379   return tx_size;
380 }
381 
compute_intra_yprediction(PREDICTION_MODE mode,BLOCK_SIZE bsize,MACROBLOCK * x,MACROBLOCKD * xd)382 static void compute_intra_yprediction(PREDICTION_MODE mode, BLOCK_SIZE bsize,
383                                       MACROBLOCK *x, MACROBLOCKD *xd) {
384   struct macroblockd_plane *const pd = &xd->plane[0];
385   struct macroblock_plane *const p = &x->plane[0];
386   uint8_t *const src_buf_base = p->src.buf;
387   uint8_t *const dst_buf_base = pd->dst.buf;
388   const int src_stride = p->src.stride;
389   const int dst_stride = pd->dst.stride;
390   // block and transform sizes, in number of 4x4 blocks log 2 ("*_b")
391   // 4x4=0, 8x8=2, 16x16=4, 32x32=6, 64x64=8
392   const TX_SIZE tx_size = max_txsize_lookup[bsize];
393   const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
394   const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
395   int row, col;
396   // If mb_to_right_edge is < 0 we are in a situation in which
397   // the current block size extends into the UMV and we won't
398   // visit the sub blocks that are wholly within the UMV.
399   const int max_blocks_wide =
400       num_4x4_w + (xd->mb_to_right_edge >= 0
401                        ? 0
402                        : xd->mb_to_right_edge >> (5 + pd->subsampling_x));
403   const int max_blocks_high =
404       num_4x4_h + (xd->mb_to_bottom_edge >= 0
405                        ? 0
406                        : xd->mb_to_bottom_edge >> (5 + pd->subsampling_y));
407 
408   // Keep track of the row and column of the blocks we use so that we know
409   // if we are in the unrestricted motion border.
410   for (row = 0; row < max_blocks_high; row += (1 << tx_size)) {
411     // Skip visiting the sub blocks that are wholly within the UMV.
412     for (col = 0; col < max_blocks_wide; col += (1 << tx_size)) {
413       p->src.buf = &src_buf_base[4 * (row * (int64_t)src_stride + col)];
414       pd->dst.buf = &dst_buf_base[4 * (row * (int64_t)dst_stride + col)];
415       vp9_predict_intra_block(xd, b_width_log2_lookup[bsize], tx_size, mode,
416                               x->skip_encode ? p->src.buf : pd->dst.buf,
417                               x->skip_encode ? src_stride : dst_stride,
418                               pd->dst.buf, dst_stride, col, row, 0);
419     }
420   }
421   p->src.buf = src_buf_base;
422   pd->dst.buf = dst_buf_base;
423 }
424 
model_rd_for_sb_y_large(VP9_COMP * cpi,BLOCK_SIZE bsize,MACROBLOCK * x,MACROBLOCKD * xd,int * out_rate_sum,int64_t * out_dist_sum,unsigned int * var_y,unsigned int * sse_y,int mi_row,int mi_col,int * early_term,int * flag_preduv_computed)425 static void model_rd_for_sb_y_large(VP9_COMP *cpi, BLOCK_SIZE bsize,
426                                     MACROBLOCK *x, MACROBLOCKD *xd,
427                                     int *out_rate_sum, int64_t *out_dist_sum,
428                                     unsigned int *var_y, unsigned int *sse_y,
429                                     int mi_row, int mi_col, int *early_term,
430                                     int *flag_preduv_computed) {
431   // Note our transform coeffs are 8 times an orthogonal transform.
432   // Hence quantizer step is also 8 times. To get effective quantizer
433   // we need to divide by 8 before sending to modeling function.
434   unsigned int sse;
435   int rate;
436   int64_t dist;
437   struct macroblock_plane *const p = &x->plane[0];
438   struct macroblockd_plane *const pd = &xd->plane[0];
439   const uint32_t dc_quant = pd->dequant[0];
440   const uint32_t ac_quant = pd->dequant[1];
441   int64_t dc_thr = dc_quant * dc_quant >> 6;
442   int64_t ac_thr = ac_quant * ac_quant >> 6;
443   unsigned int var;
444   int sum;
445   int skip_dc = 0;
446 
447   const int bw = b_width_log2_lookup[bsize];
448   const int bh = b_height_log2_lookup[bsize];
449   const int num8x8 = 1 << (bw + bh - 2);
450   unsigned int sse8x8[64] = { 0 };
451   int sum8x8[64] = { 0 };
452   unsigned int var8x8[64] = { 0 };
453   TX_SIZE tx_size;
454   int i, k;
455 #if CONFIG_VP9_HIGHBITDEPTH
456   const vpx_bit_depth_t bd = cpi->common.bit_depth;
457 #endif
458   // Calculate variance for whole partition, and also save 8x8 blocks' variance
459   // to be used in following transform skipping test.
460   block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
461                  4 << bw, 4 << bh, &sse, &sum, 8,
462 #if CONFIG_VP9_HIGHBITDEPTH
463                  cpi->common.use_highbitdepth, bd,
464 #endif
465                  sse8x8, sum8x8, var8x8);
466   var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
467 
468   *var_y = var;
469   *sse_y = sse;
470 
471 #if CONFIG_VP9_TEMPORAL_DENOISING
472   if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
473       cpi->oxcf.speed > 5)
474     ac_thr = vp9_scale_acskip_thresh(ac_thr, cpi->denoiser.denoising_level,
475                                      (abs(sum) >> (bw + bh)),
476                                      cpi->svc.temporal_layer_id);
477   else
478     ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
479                             cpi->common.height, abs(sum) >> (bw + bh));
480 #else
481   ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
482                           cpi->common.height, abs(sum) >> (bw + bh));
483 #endif
484 
485   tx_size = calculate_tx_size(cpi, bsize, xd, var, sse, ac_thr,
486                               x->source_variance, 0);
487   // The code below for setting skip flag assumes tranform size of at least 8x8,
488   // so force this lower limit on transform.
489   if (tx_size < TX_8X8) tx_size = TX_8X8;
490   xd->mi[0]->tx_size = tx_size;
491 
492   if (cpi->oxcf.content == VP9E_CONTENT_SCREEN && x->zero_temp_sad_source &&
493       x->source_variance == 0)
494     dc_thr = dc_thr << 1;
495 
496   // Evaluate if the partition block is a skippable block in Y plane.
497   {
498     unsigned int sse16x16[16] = { 0 };
499     int sum16x16[16] = { 0 };
500     unsigned int var16x16[16] = { 0 };
501     const int num16x16 = num8x8 >> 2;
502 
503     unsigned int sse32x32[4] = { 0 };
504     int sum32x32[4] = { 0 };
505     unsigned int var32x32[4] = { 0 };
506     const int num32x32 = num8x8 >> 4;
507 
508     int ac_test = 1;
509     int dc_test = 1;
510     const int num = (tx_size == TX_8X8)
511                         ? num8x8
512                         : ((tx_size == TX_16X16) ? num16x16 : num32x32);
513     const unsigned int *sse_tx =
514         (tx_size == TX_8X8) ? sse8x8
515                             : ((tx_size == TX_16X16) ? sse16x16 : sse32x32);
516     const unsigned int *var_tx =
517         (tx_size == TX_8X8) ? var8x8
518                             : ((tx_size == TX_16X16) ? var16x16 : var32x32);
519 
520     // Calculate variance if tx_size > TX_8X8
521     if (tx_size >= TX_16X16)
522       calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
523                          sum16x16);
524     if (tx_size == TX_32X32)
525       calculate_variance(bw, bh, TX_16X16, sse16x16, sum16x16, var32x32,
526                          sse32x32, sum32x32);
527 
528     // Skipping test
529     x->skip_txfm[0] = SKIP_TXFM_NONE;
530     for (k = 0; k < num; k++)
531       // Check if all ac coefficients can be quantized to zero.
532       if (!(var_tx[k] < ac_thr || var == 0)) {
533         ac_test = 0;
534         break;
535       }
536 
537     for (k = 0; k < num; k++)
538       // Check if dc coefficient can be quantized to zero.
539       if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
540         dc_test = 0;
541         break;
542       }
543 
544     if (ac_test) {
545       x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
546 
547       if (dc_test) x->skip_txfm[0] = SKIP_TXFM_AC_DC;
548     } else if (dc_test) {
549       skip_dc = 1;
550     }
551   }
552 
553   if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
554     int skip_uv[2] = { 0 };
555     unsigned int var_uv[2];
556     unsigned int sse_uv[2];
557 
558     *out_rate_sum = 0;
559     *out_dist_sum = sse << 4;
560 
561     // Transform skipping test in UV planes.
562     for (i = 1; i <= 2; i++) {
563       struct macroblock_plane *const p = &x->plane[i];
564       struct macroblockd_plane *const pd = &xd->plane[i];
565       const TX_SIZE uv_tx_size = get_uv_tx_size(xd->mi[0], pd);
566       const BLOCK_SIZE unit_size = txsize_to_bsize[uv_tx_size];
567       const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, pd);
568       const int uv_bw = b_width_log2_lookup[uv_bsize];
569       const int uv_bh = b_height_log2_lookup[uv_bsize];
570       const int sf = (uv_bw - b_width_log2_lookup[unit_size]) +
571                      (uv_bh - b_height_log2_lookup[unit_size]);
572       const uint32_t uv_dc_thr = pd->dequant[0] * pd->dequant[0] >> (6 - sf);
573       const uint32_t uv_ac_thr = pd->dequant[1] * pd->dequant[1] >> (6 - sf);
574       int j = i - 1;
575 
576       vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, i);
577       flag_preduv_computed[i - 1] = 1;
578       var_uv[j] = cpi->fn_ptr[uv_bsize].vf(
579           p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride, &sse_uv[j]);
580 
581       if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
582           (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
583         skip_uv[j] = 1;
584       else
585         break;
586     }
587 
588     // If the transform in YUV planes are skippable, the mode search checks
589     // fewer inter modes and doesn't check intra modes.
590     if (skip_uv[0] & skip_uv[1]) {
591       *early_term = 1;
592     }
593     return;
594   }
595 
596   if (!skip_dc) {
597 #if CONFIG_VP9_HIGHBITDEPTH
598     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
599                                  dc_quant >> (xd->bd - 5), &rate, &dist);
600 #else
601     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
602                                  dc_quant >> 3, &rate, &dist);
603 #endif  // CONFIG_VP9_HIGHBITDEPTH
604   }
605 
606   if (!skip_dc) {
607     *out_rate_sum = rate >> 1;
608     *out_dist_sum = dist << 3;
609   } else {
610     *out_rate_sum = 0;
611     *out_dist_sum = (sse - var) << 4;
612   }
613 
614 #if CONFIG_VP9_HIGHBITDEPTH
615   vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
616                                ac_quant >> (xd->bd - 5), &rate, &dist);
617 #else
618   vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
619                                &rate, &dist);
620 #endif  // CONFIG_VP9_HIGHBITDEPTH
621 
622   *out_rate_sum += rate;
623   *out_dist_sum += dist << 4;
624 }
625 
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,int is_intra)626 static void model_rd_for_sb_y(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
627                               MACROBLOCKD *xd, int *out_rate_sum,
628                               int64_t *out_dist_sum, unsigned int *var_y,
629                               unsigned int *sse_y, int is_intra) {
630   // Note our transform coeffs are 8 times an orthogonal transform.
631   // Hence quantizer step is also 8 times. To get effective quantizer
632   // we need to divide by 8 before sending to modeling function.
633   unsigned int sse;
634   int rate;
635   int64_t dist;
636   struct macroblock_plane *const p = &x->plane[0];
637   struct macroblockd_plane *const pd = &xd->plane[0];
638   const int64_t dc_thr = p->quant_thred[0] >> 6;
639   const int64_t ac_thr = p->quant_thred[1] >> 6;
640   const uint32_t dc_quant = pd->dequant[0];
641   const uint32_t ac_quant = pd->dequant[1];
642   unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
643                                            pd->dst.buf, pd->dst.stride, &sse);
644   int skip_dc = 0;
645 
646   *var_y = var;
647   *sse_y = sse;
648 
649   xd->mi[0]->tx_size = calculate_tx_size(cpi, bsize, xd, var, sse, ac_thr,
650                                          x->source_variance, is_intra);
651 
652   // Evaluate if the partition block is a skippable block in Y plane.
653   {
654     const BLOCK_SIZE unit_size = txsize_to_bsize[xd->mi[0]->tx_size];
655     const unsigned int num_blk_log2 =
656         (b_width_log2_lookup[bsize] - b_width_log2_lookup[unit_size]) +
657         (b_height_log2_lookup[bsize] - b_height_log2_lookup[unit_size]);
658     const unsigned int sse_tx = sse >> num_blk_log2;
659     const unsigned int var_tx = var >> num_blk_log2;
660 
661     x->skip_txfm[0] = SKIP_TXFM_NONE;
662     // Check if all ac coefficients can be quantized to zero.
663     if (var_tx < ac_thr || var == 0) {
664       x->skip_txfm[0] = SKIP_TXFM_AC_ONLY;
665       // Check if dc coefficient can be quantized to zero.
666       if (sse_tx - var_tx < dc_thr || sse == var)
667         x->skip_txfm[0] = SKIP_TXFM_AC_DC;
668     } else {
669       if (sse_tx - var_tx < dc_thr || sse == var) skip_dc = 1;
670     }
671   }
672 
673   if (x->skip_txfm[0] == SKIP_TXFM_AC_DC) {
674     *out_rate_sum = 0;
675     *out_dist_sum = sse << 4;
676     return;
677   }
678 
679   if (!skip_dc) {
680 #if CONFIG_VP9_HIGHBITDEPTH
681     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
682                                  dc_quant >> (xd->bd - 5), &rate, &dist);
683 #else
684     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bsize],
685                                  dc_quant >> 3, &rate, &dist);
686 #endif  // CONFIG_VP9_HIGHBITDEPTH
687   }
688 
689   if (!skip_dc) {
690     *out_rate_sum = rate >> 1;
691     *out_dist_sum = dist << 3;
692   } else {
693     *out_rate_sum = 0;
694     *out_dist_sum = (sse - var) << 4;
695   }
696 
697 #if CONFIG_VP9_HIGHBITDEPTH
698   vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize],
699                                ac_quant >> (xd->bd - 5), &rate, &dist);
700 #else
701   vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bsize], ac_quant >> 3,
702                                &rate, &dist);
703 #endif  // CONFIG_VP9_HIGHBITDEPTH
704 
705   *out_rate_sum += rate;
706   *out_dist_sum += dist << 4;
707 }
708 
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,int is_intra)709 static void block_yrd(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc,
710                       int *skippable, int64_t *sse, BLOCK_SIZE bsize,
711                       TX_SIZE tx_size, int rd_computed, int is_intra) {
712   MACROBLOCKD *xd = &x->e_mbd;
713   const struct macroblockd_plane *pd = &xd->plane[0];
714   struct macroblock_plane *const p = &x->plane[0];
715   const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
716   const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
717   const int step = 1 << (tx_size << 1);
718   const int block_step = (1 << tx_size);
719   int block = 0, r, c;
720   const int max_blocks_wide =
721       num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 : xd->mb_to_right_edge >> 5);
722   const int max_blocks_high =
723       num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 : xd->mb_to_bottom_edge >> 5);
724   int eob_cost = 0;
725   const int bw = 4 * num_4x4_w;
726   const int bh = 4 * num_4x4_h;
727 
728   if (cpi->sf.use_simple_block_yrd && cpi->common.frame_type != KEY_FRAME &&
729       (bsize < BLOCK_32X32 ||
730        (cpi->use_svc &&
731         (bsize < BLOCK_32X32 || cpi->svc.temporal_layer_id > 0)))) {
732     unsigned int var_y, sse_y;
733     (void)tx_size;
734     if (!rd_computed)
735       model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc->rate, &this_rdc->dist,
736                         &var_y, &sse_y, is_intra);
737     *sse = INT_MAX;
738     *skippable = 0;
739     return;
740   }
741 
742   (void)cpi;
743 
744   // The max tx_size passed in is TX_16X16.
745   assert(tx_size != TX_32X32);
746 #if CONFIG_VP9_HIGHBITDEPTH
747   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
748     vpx_highbd_subtract_block(bh, bw, p->src_diff, bw, p->src.buf,
749                               p->src.stride, pd->dst.buf, pd->dst.stride,
750                               x->e_mbd.bd);
751   } else {
752     vpx_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
753                        pd->dst.buf, pd->dst.stride);
754   }
755 #else
756   vpx_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
757                      pd->dst.buf, pd->dst.stride);
758 #endif
759   *skippable = 1;
760   // Keep track of the row and column of the blocks we use so that we know
761   // if we are in the unrestricted motion border.
762   for (r = 0; r < max_blocks_high; r += block_step) {
763     for (c = 0; c < num_4x4_w; c += block_step) {
764       if (c < max_blocks_wide) {
765         const scan_order *const scan_order = &vp9_default_scan_orders[tx_size];
766         tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
767         tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
768         tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
769         uint16_t *const eob = &p->eobs[block];
770         const int diff_stride = bw;
771         const int16_t *src_diff;
772         src_diff = &p->src_diff[(r * diff_stride + c) << 2];
773 
774         switch (tx_size) {
775           case TX_16X16:
776             vpx_hadamard_16x16(src_diff, diff_stride, coeff);
777             vp9_quantize_fp(coeff, 256, x->skip_block, p->round_fp, p->quant_fp,
778                             qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
779                             scan_order->iscan);
780             break;
781           case TX_8X8:
782             vpx_hadamard_8x8(src_diff, diff_stride, coeff);
783             vp9_quantize_fp(coeff, 64, x->skip_block, p->round_fp, p->quant_fp,
784                             qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
785                             scan_order->iscan);
786             break;
787           default:
788             assert(tx_size == TX_4X4);
789             x->fwd_txfm4x4(src_diff, coeff, diff_stride);
790             vp9_quantize_fp(coeff, 16, x->skip_block, p->round_fp, p->quant_fp,
791                             qcoeff, dqcoeff, pd->dequant, eob, scan_order->scan,
792                             scan_order->iscan);
793             break;
794         }
795         *skippable &= (*eob == 0);
796         eob_cost += 1;
797       }
798       block += step;
799     }
800   }
801 
802   this_rdc->rate = 0;
803   if (*sse < INT64_MAX) {
804     *sse = (*sse << 6) >> 2;
805     if (*skippable) {
806       this_rdc->dist = *sse;
807       return;
808     }
809   }
810 
811   block = 0;
812   this_rdc->dist = 0;
813   for (r = 0; r < max_blocks_high; r += block_step) {
814     for (c = 0; c < num_4x4_w; c += block_step) {
815       if (c < max_blocks_wide) {
816         tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
817         tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
818         tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
819         uint16_t *const eob = &p->eobs[block];
820 
821         if (*eob == 1)
822           this_rdc->rate += (int)abs(qcoeff[0]);
823         else if (*eob > 1)
824           this_rdc->rate += vpx_satd(qcoeff, step << 4);
825 
826         this_rdc->dist += vp9_block_error_fp(coeff, dqcoeff, step << 4) >> 2;
827       }
828       block += step;
829     }
830   }
831 
832   // If skippable is set, rate gets clobbered later.
833   this_rdc->rate <<= (2 + VP9_PROB_COST_SHIFT);
834   this_rdc->rate += (eob_cost << VP9_PROB_COST_SHIFT);
835 }
836 
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)837 static void model_rd_for_sb_uv(VP9_COMP *cpi, BLOCK_SIZE plane_bsize,
838                                MACROBLOCK *x, MACROBLOCKD *xd,
839                                RD_COST *this_rdc, unsigned int *var_y,
840                                unsigned int *sse_y, int start_plane,
841                                int stop_plane) {
842   // Note our transform coeffs are 8 times an orthogonal transform.
843   // Hence quantizer step is also 8 times. To get effective quantizer
844   // we need to divide by 8 before sending to modeling function.
845   unsigned int sse;
846   int rate;
847   int64_t dist;
848   int i;
849 #if CONFIG_VP9_HIGHBITDEPTH
850   uint64_t tot_var = *var_y;
851   uint64_t tot_sse = *sse_y;
852 #else
853   uint32_t tot_var = *var_y;
854   uint32_t tot_sse = *sse_y;
855 #endif
856 
857   this_rdc->rate = 0;
858   this_rdc->dist = 0;
859 
860   for (i = start_plane; i <= stop_plane; ++i) {
861     struct macroblock_plane *const p = &x->plane[i];
862     struct macroblockd_plane *const pd = &xd->plane[i];
863     const uint32_t dc_quant = pd->dequant[0];
864     const uint32_t ac_quant = pd->dequant[1];
865     const BLOCK_SIZE bs = plane_bsize;
866     unsigned int var;
867     if (!x->color_sensitivity[i - 1]) continue;
868 
869     var = cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf,
870                              pd->dst.stride, &sse);
871     assert(sse >= var);
872     tot_var += var;
873     tot_sse += sse;
874 
875 #if CONFIG_VP9_HIGHBITDEPTH
876     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
877                                  dc_quant >> (xd->bd - 5), &rate, &dist);
878 #else
879     vp9_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
880                                  dc_quant >> 3, &rate, &dist);
881 #endif  // CONFIG_VP9_HIGHBITDEPTH
882 
883     this_rdc->rate += rate >> 1;
884     this_rdc->dist += dist << 3;
885 
886 #if CONFIG_VP9_HIGHBITDEPTH
887     vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs],
888                                  ac_quant >> (xd->bd - 5), &rate, &dist);
889 #else
890     vp9_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs], ac_quant >> 3,
891                                  &rate, &dist);
892 #endif  // CONFIG_VP9_HIGHBITDEPTH
893 
894     this_rdc->rate += rate;
895     this_rdc->dist += dist << 4;
896   }
897 
898 #if CONFIG_VP9_HIGHBITDEPTH
899   *var_y = tot_var > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_var;
900   *sse_y = tot_sse > UINT32_MAX ? UINT32_MAX : (uint32_t)tot_sse;
901 #else
902   *var_y = tot_var;
903   *sse_y = tot_sse;
904 #endif
905 }
906 
get_pred_buffer(PRED_BUFFER * p,int len)907 static int get_pred_buffer(PRED_BUFFER *p, int len) {
908   int i;
909 
910   for (i = 0; i < len; i++) {
911     if (!p[i].in_use) {
912       p[i].in_use = 1;
913       return i;
914     }
915   }
916   return -1;
917 }
918 
free_pred_buffer(PRED_BUFFER * p)919 static void free_pred_buffer(PRED_BUFFER *p) {
920   if (p != NULL) p->in_use = 0;
921 }
922 
encode_breakout_test(VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int mi_row,int mi_col,MV_REFERENCE_FRAME ref_frame,PREDICTION_MODE this_mode,unsigned int var_y,unsigned int sse_y,struct buf_2d yv12_mb[][MAX_MB_PLANE],int * rate,int64_t * dist,int * flag_preduv_computed)923 static void encode_breakout_test(
924     VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, int mi_row, int mi_col,
925     MV_REFERENCE_FRAME ref_frame, PREDICTION_MODE this_mode, unsigned int var_y,
926     unsigned int sse_y, struct buf_2d yv12_mb[][MAX_MB_PLANE], int *rate,
927     int64_t *dist, int *flag_preduv_computed) {
928   MACROBLOCKD *xd = &x->e_mbd;
929   MODE_INFO *const mi = xd->mi[0];
930   const BLOCK_SIZE uv_size = get_plane_block_size(bsize, &xd->plane[1]);
931   unsigned int var = var_y, sse = sse_y;
932   // Skipping threshold for ac.
933   unsigned int thresh_ac;
934   // Skipping threshold for dc.
935   unsigned int thresh_dc;
936   int motion_low = 1;
937 
938   if (cpi->use_svc && ref_frame == GOLDEN_FRAME) return;
939   if (mi->mv[0].as_mv.row > 64 || mi->mv[0].as_mv.row < -64 ||
940       mi->mv[0].as_mv.col > 64 || mi->mv[0].as_mv.col < -64)
941     motion_low = 0;
942   if (x->encode_breakout > 0 && motion_low == 1) {
943     // Set a maximum for threshold to avoid big PSNR loss in low bit rate
944     // case. Use extreme low threshold for static frames to limit
945     // skipping.
946     const unsigned int max_thresh = 36000;
947     // The encode_breakout input
948     const unsigned int min_thresh =
949         VPXMIN(((unsigned int)x->encode_breakout << 4), max_thresh);
950 #if CONFIG_VP9_HIGHBITDEPTH
951     const int shift = (xd->bd << 1) - 16;
952 #endif
953 
954     // Calculate threshold according to dequant value.
955     thresh_ac = (xd->plane[0].dequant[1] * xd->plane[0].dequant[1]) >> 3;
956 #if CONFIG_VP9_HIGHBITDEPTH
957     if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
958       thresh_ac = ROUND_POWER_OF_TWO(thresh_ac, shift);
959     }
960 #endif  // CONFIG_VP9_HIGHBITDEPTH
961     thresh_ac = clamp(thresh_ac, min_thresh, max_thresh);
962 
963     // Adjust ac threshold according to partition size.
964     thresh_ac >>=
965         8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
966 
967     thresh_dc = (xd->plane[0].dequant[0] * xd->plane[0].dequant[0] >> 6);
968 #if CONFIG_VP9_HIGHBITDEPTH
969     if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && shift > 0) {
970       thresh_dc = ROUND_POWER_OF_TWO(thresh_dc, shift);
971     }
972 #endif  // CONFIG_VP9_HIGHBITDEPTH
973   } else {
974     thresh_ac = 0;
975     thresh_dc = 0;
976   }
977 
978   // Y skipping condition checking for ac and dc.
979   if (var <= thresh_ac && (sse - var) <= thresh_dc) {
980     unsigned int sse_u, sse_v;
981     unsigned int var_u, var_v;
982     unsigned int thresh_ac_uv = thresh_ac;
983     unsigned int thresh_dc_uv = thresh_dc;
984     if (x->sb_is_skin) {
985       thresh_ac_uv = 0;
986       thresh_dc_uv = 0;
987     }
988 
989     if (!flag_preduv_computed[0] || !flag_preduv_computed[1]) {
990       xd->plane[1].pre[0] = yv12_mb[ref_frame][1];
991       xd->plane[2].pre[0] = yv12_mb[ref_frame][2];
992       vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, bsize);
993     }
994 
995     var_u = cpi->fn_ptr[uv_size].vf(x->plane[1].src.buf, x->plane[1].src.stride,
996                                     xd->plane[1].dst.buf,
997                                     xd->plane[1].dst.stride, &sse_u);
998 
999     // U skipping condition checking
1000     if (((var_u << 2) <= thresh_ac_uv) && (sse_u - var_u <= thresh_dc_uv)) {
1001       var_v = cpi->fn_ptr[uv_size].vf(
1002           x->plane[2].src.buf, x->plane[2].src.stride, xd->plane[2].dst.buf,
1003           xd->plane[2].dst.stride, &sse_v);
1004 
1005       // V skipping condition checking
1006       if (((var_v << 2) <= thresh_ac_uv) && (sse_v - var_v <= thresh_dc_uv)) {
1007         x->skip = 1;
1008 
1009         // The cost of skip bit needs to be added.
1010         *rate = cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
1011                                     [INTER_OFFSET(this_mode)];
1012 
1013         // More on this part of rate
1014         // rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
1015 
1016         // Scaling factor for SSE from spatial domain to frequency
1017         // domain is 16. Adjust distortion accordingly.
1018         // TODO(yunqingwang): In this function, only y-plane dist is
1019         // calculated.
1020         *dist = (sse << 4);  // + ((sse_u + sse_v) << 4);
1021 
1022         // *disable_skip = 1;
1023       }
1024     }
1025   }
1026 }
1027 
1028 struct estimate_block_intra_args {
1029   VP9_COMP *cpi;
1030   MACROBLOCK *x;
1031   PREDICTION_MODE mode;
1032   int skippable;
1033   RD_COST *rdc;
1034 };
1035 
estimate_block_intra(int plane,int block,int row,int col,BLOCK_SIZE plane_bsize,TX_SIZE tx_size,void * arg)1036 static void estimate_block_intra(int plane, int block, int row, int col,
1037                                  BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
1038                                  void *arg) {
1039   struct estimate_block_intra_args *const args = arg;
1040   VP9_COMP *const cpi = args->cpi;
1041   MACROBLOCK *const x = args->x;
1042   MACROBLOCKD *const xd = &x->e_mbd;
1043   struct macroblock_plane *const p = &x->plane[plane];
1044   struct macroblockd_plane *const pd = &xd->plane[plane];
1045   const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
1046   uint8_t *const src_buf_base = p->src.buf;
1047   uint8_t *const dst_buf_base = pd->dst.buf;
1048   const int src_stride = p->src.stride;
1049   const int dst_stride = pd->dst.stride;
1050   RD_COST this_rdc;
1051 
1052   (void)block;
1053 
1054   p->src.buf = &src_buf_base[4 * (row * (int64_t)src_stride + col)];
1055   pd->dst.buf = &dst_buf_base[4 * (row * (int64_t)dst_stride + col)];
1056   // Use source buffer as an approximation for the fully reconstructed buffer.
1057   vp9_predict_intra_block(xd, b_width_log2_lookup[plane_bsize], tx_size,
1058                           args->mode, x->skip_encode ? p->src.buf : pd->dst.buf,
1059                           x->skip_encode ? src_stride : dst_stride, pd->dst.buf,
1060                           dst_stride, col, row, plane);
1061 
1062   if (plane == 0) {
1063     int64_t this_sse = INT64_MAX;
1064     block_yrd(cpi, x, &this_rdc, &args->skippable, &this_sse, bsize_tx,
1065               VPXMIN(tx_size, TX_16X16), 0, 1);
1066   } else {
1067     unsigned int var = 0;
1068     unsigned int sse = 0;
1069     model_rd_for_sb_uv(cpi, bsize_tx, x, xd, &this_rdc, &var, &sse, plane,
1070                        plane);
1071   }
1072 
1073   p->src.buf = src_buf_base;
1074   pd->dst.buf = dst_buf_base;
1075   args->rdc->rate += this_rdc.rate;
1076   args->rdc->dist += this_rdc.dist;
1077 }
1078 
1079 static const THR_MODES mode_idx[MAX_REF_FRAMES][4] = {
1080   { THR_DC, THR_V_PRED, THR_H_PRED, THR_TM },
1081   { THR_NEARESTMV, THR_NEARMV, THR_ZEROMV, THR_NEWMV },
1082   { THR_NEARESTG, THR_NEARG, THR_ZEROG, THR_NEWG },
1083   { THR_NEARESTA, THR_NEARA, THR_ZEROA, THR_NEWA },
1084 };
1085 
1086 static const PREDICTION_MODE intra_mode_list[] = { DC_PRED, V_PRED, H_PRED,
1087                                                    TM_PRED };
1088 
mode_offset(const PREDICTION_MODE mode)1089 static int mode_offset(const PREDICTION_MODE mode) {
1090   if (mode >= NEARESTMV) {
1091     return INTER_OFFSET(mode);
1092   } else {
1093     switch (mode) {
1094       case DC_PRED: return 0;
1095       case V_PRED: return 1;
1096       case H_PRED: return 2;
1097       case TM_PRED: return 3;
1098       default: return -1;
1099     }
1100   }
1101 }
1102 
rd_less_than_thresh_row_mt(int64_t best_rd,int thresh,const int * const thresh_fact)1103 static INLINE int rd_less_than_thresh_row_mt(int64_t best_rd, int thresh,
1104                                              const int *const thresh_fact) {
1105   int is_rd_less_than_thresh;
1106   is_rd_less_than_thresh =
1107       best_rd < ((int64_t)thresh * (*thresh_fact) >> 5) || thresh == INT_MAX;
1108   return is_rd_less_than_thresh;
1109 }
1110 
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)1111 static INLINE void update_thresh_freq_fact_row_mt(
1112     VP9_COMP *cpi, TileDataEnc *tile_data, int source_variance,
1113     int thresh_freq_fact_idx, MV_REFERENCE_FRAME ref_frame,
1114     THR_MODES best_mode_idx, PREDICTION_MODE mode) {
1115   THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1116   int freq_fact_idx = thresh_freq_fact_idx + thr_mode_idx;
1117   int *freq_fact = &tile_data->row_base_thresh_freq_fact[freq_fact_idx];
1118   if (thr_mode_idx == best_mode_idx)
1119     *freq_fact -= (*freq_fact >> 4);
1120   else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV &&
1121            ref_frame == LAST_FRAME && source_variance < 5) {
1122     *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32);
1123   } else {
1124     *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
1125                         cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1126   }
1127 }
1128 
update_thresh_freq_fact(VP9_COMP * cpi,TileDataEnc * tile_data,unsigned int source_variance,BLOCK_SIZE bsize,MV_REFERENCE_FRAME ref_frame,THR_MODES best_mode_idx,PREDICTION_MODE mode)1129 static INLINE void update_thresh_freq_fact(
1130     VP9_COMP *cpi, TileDataEnc *tile_data, unsigned int source_variance,
1131     BLOCK_SIZE bsize, MV_REFERENCE_FRAME ref_frame, THR_MODES best_mode_idx,
1132     PREDICTION_MODE mode) {
1133   THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1134   int *freq_fact = &tile_data->thresh_freq_fact[bsize][thr_mode_idx];
1135   if (thr_mode_idx == best_mode_idx)
1136     *freq_fact -= (*freq_fact >> 4);
1137   else if (cpi->sf.limit_newmv_early_exit && mode == NEWMV &&
1138            ref_frame == LAST_FRAME && source_variance < 5) {
1139     *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC, 32);
1140   } else {
1141     *freq_fact = VPXMIN(*freq_fact + RD_THRESH_INC,
1142                         cpi->sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1143   }
1144 }
1145 
vp9_pick_intra_mode(VP9_COMP * cpi,MACROBLOCK * x,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)1146 void vp9_pick_intra_mode(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
1147                          BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1148   MACROBLOCKD *const xd = &x->e_mbd;
1149   MODE_INFO *const mi = xd->mi[0];
1150   RD_COST this_rdc, best_rdc;
1151   PREDICTION_MODE this_mode;
1152   struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
1153   const TX_SIZE intra_tx_size =
1154       VPXMIN(max_txsize_lookup[bsize],
1155              tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
1156   MODE_INFO *const mic = xd->mi[0];
1157   int *bmode_costs;
1158   const MODE_INFO *above_mi = xd->above_mi;
1159   const MODE_INFO *left_mi = xd->left_mi;
1160   const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1161   const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1162   bmode_costs = cpi->y_mode_costs[A][L];
1163 
1164   (void)ctx;
1165   vp9_rd_cost_reset(&best_rdc);
1166   vp9_rd_cost_reset(&this_rdc);
1167 
1168   mi->ref_frame[0] = INTRA_FRAME;
1169   // Initialize interp_filter here so we do not have to check for inter block
1170   // modes in get_pred_context_switchable_interp()
1171   mi->interp_filter = SWITCHABLE_FILTERS;
1172 
1173   mi->mv[0].as_int = INVALID_MV;
1174   mi->uv_mode = DC_PRED;
1175   memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
1176 
1177   // Change the limit of this loop to add other intra prediction
1178   // mode tests.
1179   for (this_mode = DC_PRED; this_mode <= H_PRED; ++this_mode) {
1180     this_rdc.dist = this_rdc.rate = 0;
1181     args.mode = this_mode;
1182     args.skippable = 1;
1183     args.rdc = &this_rdc;
1184     mi->tx_size = intra_tx_size;
1185     vp9_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
1186                                            &args);
1187     if (args.skippable) {
1188       x->skip_txfm[0] = SKIP_TXFM_AC_DC;
1189       this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
1190     } else {
1191       x->skip_txfm[0] = SKIP_TXFM_NONE;
1192       this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
1193     }
1194     this_rdc.rate += bmode_costs[this_mode];
1195     this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
1196 
1197     if (this_rdc.rdcost < best_rdc.rdcost) {
1198       best_rdc = this_rdc;
1199       mi->mode = this_mode;
1200     }
1201   }
1202 
1203   *rd_cost = best_rdc;
1204 }
1205 
init_ref_frame_cost(VP9_COMMON * const cm,MACROBLOCKD * const xd,int ref_frame_cost[MAX_REF_FRAMES])1206 static void init_ref_frame_cost(VP9_COMMON *const cm, MACROBLOCKD *const xd,
1207                                 int ref_frame_cost[MAX_REF_FRAMES]) {
1208   vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
1209   vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
1210   vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
1211 
1212   ref_frame_cost[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
1213   ref_frame_cost[LAST_FRAME] = ref_frame_cost[GOLDEN_FRAME] =
1214       ref_frame_cost[ALTREF_FRAME] = vp9_cost_bit(intra_inter_p, 1);
1215 
1216   ref_frame_cost[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
1217   ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1218   ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
1219   ref_frame_cost[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
1220   ref_frame_cost[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
1221 }
1222 
1223 typedef struct {
1224   MV_REFERENCE_FRAME ref_frame;
1225   PREDICTION_MODE pred_mode;
1226 } REF_MODE;
1227 
1228 #define RT_INTER_MODES 12
1229 static const REF_MODE ref_mode_set[RT_INTER_MODES] = {
1230   { LAST_FRAME, ZEROMV },   { LAST_FRAME, NEARESTMV },
1231   { GOLDEN_FRAME, ZEROMV }, { LAST_FRAME, NEARMV },
1232   { LAST_FRAME, NEWMV },    { GOLDEN_FRAME, NEARESTMV },
1233   { GOLDEN_FRAME, NEARMV }, { GOLDEN_FRAME, NEWMV },
1234   { ALTREF_FRAME, ZEROMV }, { ALTREF_FRAME, NEARESTMV },
1235   { ALTREF_FRAME, NEARMV }, { ALTREF_FRAME, NEWMV }
1236 };
1237 
1238 #define RT_INTER_MODES_SVC 8
1239 static const REF_MODE ref_mode_set_svc[RT_INTER_MODES_SVC] = {
1240   { LAST_FRAME, ZEROMV },      { LAST_FRAME, NEARESTMV },
1241   { LAST_FRAME, NEARMV },      { GOLDEN_FRAME, ZEROMV },
1242   { GOLDEN_FRAME, NEARESTMV }, { GOLDEN_FRAME, NEARMV },
1243   { LAST_FRAME, NEWMV },       { GOLDEN_FRAME, NEWMV }
1244 };
1245 
find_predictors(VP9_COMP * cpi,MACROBLOCK * x,MV_REFERENCE_FRAME ref_frame,int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],int const_motion[MAX_REF_FRAMES],int * ref_frame_skip_mask,const int flag_list[4],TileDataEnc * tile_data,int mi_row,int mi_col,struct buf_2d yv12_mb[4][MAX_MB_PLANE],BLOCK_SIZE bsize,int force_skip_low_temp_var,int comp_pred_allowed)1246 static INLINE void find_predictors(
1247     VP9_COMP *cpi, MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame,
1248     int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1249     int const_motion[MAX_REF_FRAMES], int *ref_frame_skip_mask,
1250     const int flag_list[4], TileDataEnc *tile_data, int mi_row, int mi_col,
1251     struct buf_2d yv12_mb[4][MAX_MB_PLANE], BLOCK_SIZE bsize,
1252     int force_skip_low_temp_var, int comp_pred_allowed) {
1253   VP9_COMMON *const cm = &cpi->common;
1254   MACROBLOCKD *const xd = &x->e_mbd;
1255   const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
1256   TileInfo *const tile_info = &tile_data->tile_info;
1257   // TODO(jingning) placeholder for inter-frame non-RD mode decision.
1258   x->pred_mv_sad[ref_frame] = INT_MAX;
1259   frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
1260   frame_mv[ZEROMV][ref_frame].as_int = 0;
1261   // this needs various further optimizations. to be continued..
1262   if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
1263     int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
1264     const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
1265     vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
1266     if (cm->use_prev_frame_mvs || comp_pred_allowed) {
1267       vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
1268                        x->mbmi_ext->mode_context);
1269     } else {
1270       const_motion[ref_frame] =
1271           mv_refs_rt(cpi, cm, x, xd, tile_info, xd->mi[0], ref_frame,
1272                      candidates, &frame_mv[NEWMV][ref_frame], mi_row, mi_col,
1273                      (int)(cpi->svc.use_base_mv && cpi->svc.spatial_layer_id));
1274     }
1275     vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
1276                           &frame_mv[NEARESTMV][ref_frame],
1277                           &frame_mv[NEARMV][ref_frame]);
1278     // Early exit for golden frame if force_skip_low_temp_var is set.
1279     if (!vp9_is_scaled(sf) && bsize >= BLOCK_8X8 &&
1280         !(force_skip_low_temp_var && ref_frame == GOLDEN_FRAME)) {
1281       vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
1282                   bsize);
1283     }
1284   } else {
1285     *ref_frame_skip_mask |= (1 << ref_frame);
1286   }
1287 }
1288 
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)1289 static void vp9_NEWMV_diff_bias(const NOISE_ESTIMATE *ne, MACROBLOCKD *xd,
1290                                 PREDICTION_MODE this_mode, RD_COST *this_rdc,
1291                                 BLOCK_SIZE bsize, int mv_row, int mv_col,
1292                                 int is_last_frame, int lowvar_highsumdiff,
1293                                 int is_skin) {
1294   // Bias against MVs associated with NEWMV mode that are very different from
1295   // top/left neighbors.
1296   if (this_mode == NEWMV) {
1297     int al_mv_average_row;
1298     int al_mv_average_col;
1299     int left_row, left_col;
1300     int row_diff, col_diff;
1301     int above_mv_valid = 0;
1302     int left_mv_valid = 0;
1303     int above_row = 0;
1304     int above_col = 0;
1305 
1306     if (xd->above_mi) {
1307       above_mv_valid = xd->above_mi->mv[0].as_int != INVALID_MV;
1308       above_row = xd->above_mi->mv[0].as_mv.row;
1309       above_col = xd->above_mi->mv[0].as_mv.col;
1310     }
1311     if (xd->left_mi) {
1312       left_mv_valid = xd->left_mi->mv[0].as_int != INVALID_MV;
1313       left_row = xd->left_mi->mv[0].as_mv.row;
1314       left_col = xd->left_mi->mv[0].as_mv.col;
1315     }
1316     if (above_mv_valid && left_mv_valid) {
1317       al_mv_average_row = (above_row + left_row + 1) >> 1;
1318       al_mv_average_col = (above_col + left_col + 1) >> 1;
1319     } else if (above_mv_valid) {
1320       al_mv_average_row = above_row;
1321       al_mv_average_col = above_col;
1322     } else if (left_mv_valid) {
1323       al_mv_average_row = left_row;
1324       al_mv_average_col = left_col;
1325     } else {
1326       al_mv_average_row = al_mv_average_col = 0;
1327     }
1328     row_diff = (al_mv_average_row - mv_row);
1329     col_diff = (al_mv_average_col - mv_col);
1330     if (row_diff > 48 || row_diff < -48 || col_diff > 48 || col_diff < -48) {
1331       if (bsize > BLOCK_32X32)
1332         this_rdc->rdcost = this_rdc->rdcost << 1;
1333       else
1334         this_rdc->rdcost = 3 * this_rdc->rdcost >> 1;
1335     }
1336   }
1337   // If noise estimation is enabled, and estimated level is above threshold,
1338   // add a bias to LAST reference with small motion, for large blocks.
1339   if (ne->enabled && ne->level >= kMedium && bsize >= BLOCK_32X32 &&
1340       is_last_frame && mv_row < 8 && mv_row > -8 && mv_col < 8 && mv_col > -8)
1341     this_rdc->rdcost = 7 * (this_rdc->rdcost >> 3);
1342   else if (lowvar_highsumdiff && !is_skin && bsize >= BLOCK_16X16 &&
1343            is_last_frame && mv_row < 16 && mv_row > -16 && mv_col < 16 &&
1344            mv_col > -16)
1345     this_rdc->rdcost = 7 * (this_rdc->rdcost >> 3);
1346 }
1347 
1348 #if CONFIG_VP9_TEMPORAL_DENOISING
vp9_pickmode_ctx_den_update(VP9_PICKMODE_CTX_DEN * ctx_den,int64_t zero_last_cost_orig,int ref_frame_cost[MAX_REF_FRAMES],int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],int reuse_inter_pred,BEST_PICKMODE * bp)1349 static void vp9_pickmode_ctx_den_update(
1350     VP9_PICKMODE_CTX_DEN *ctx_den, int64_t zero_last_cost_orig,
1351     int ref_frame_cost[MAX_REF_FRAMES],
1352     int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES], int reuse_inter_pred,
1353     BEST_PICKMODE *bp) {
1354   ctx_den->zero_last_cost_orig = zero_last_cost_orig;
1355   ctx_den->ref_frame_cost = ref_frame_cost;
1356   ctx_den->frame_mv = frame_mv;
1357   ctx_den->reuse_inter_pred = reuse_inter_pred;
1358   ctx_den->best_tx_size = bp->best_tx_size;
1359   ctx_den->best_mode = bp->best_mode;
1360   ctx_den->best_ref_frame = bp->best_ref_frame;
1361   ctx_den->best_pred_filter = bp->best_pred_filter;
1362   ctx_den->best_mode_skip_txfm = bp->best_mode_skip_txfm;
1363 }
1364 
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)1365 static void recheck_zeromv_after_denoising(
1366     VP9_COMP *cpi, MODE_INFO *const mi, MACROBLOCK *x, MACROBLOCKD *const xd,
1367     VP9_DENOISER_DECISION decision, VP9_PICKMODE_CTX_DEN *ctx_den,
1368     struct buf_2d yv12_mb[4][MAX_MB_PLANE], RD_COST *best_rdc, BLOCK_SIZE bsize,
1369     int mi_row, int mi_col) {
1370   // If INTRA or GOLDEN reference was selected, re-evaluate ZEROMV on
1371   // denoised result. Only do this under noise conditions, and if rdcost of
1372   // ZEROMV onoriginal source is not significantly higher than rdcost of best
1373   // mode.
1374   if (cpi->noise_estimate.enabled && cpi->noise_estimate.level > kLow &&
1375       ctx_den->zero_last_cost_orig < (best_rdc->rdcost << 3) &&
1376       ((ctx_den->best_ref_frame == INTRA_FRAME && decision >= FILTER_BLOCK) ||
1377        (ctx_den->best_ref_frame == GOLDEN_FRAME &&
1378         cpi->svc.number_spatial_layers == 1 &&
1379         decision == FILTER_ZEROMV_BLOCK))) {
1380     // Check if we should pick ZEROMV on denoised signal.
1381     VP9_COMMON *const cm = &cpi->common;
1382     int rate = 0;
1383     int64_t dist = 0;
1384     uint32_t var_y = UINT_MAX;
1385     uint32_t sse_y = UINT_MAX;
1386     RD_COST this_rdc;
1387     mi->mode = ZEROMV;
1388     mi->ref_frame[0] = LAST_FRAME;
1389     mi->ref_frame[1] = NONE;
1390     set_ref_ptrs(cm, xd, mi->ref_frame[0], NONE);
1391     mi->mv[0].as_int = 0;
1392     mi->interp_filter = EIGHTTAP;
1393     if (cpi->sf.default_interp_filter == BILINEAR) mi->interp_filter = BILINEAR;
1394     xd->plane[0].pre[0] = yv12_mb[LAST_FRAME][0];
1395     vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1396     model_rd_for_sb_y(cpi, bsize, x, xd, &rate, &dist, &var_y, &sse_y, 0);
1397     this_rdc.rate = rate + ctx_den->ref_frame_cost[LAST_FRAME] +
1398                     cpi->inter_mode_cost[x->mbmi_ext->mode_context[LAST_FRAME]]
1399                                         [INTER_OFFSET(ZEROMV)];
1400     this_rdc.dist = dist;
1401     this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, rate, dist);
1402     // Don't switch to ZEROMV if the rdcost for ZEROMV on denoised source
1403     // is higher than best_ref mode (on original source).
1404     if (this_rdc.rdcost > best_rdc->rdcost) {
1405       this_rdc = *best_rdc;
1406       mi->mode = ctx_den->best_mode;
1407       mi->ref_frame[0] = ctx_den->best_ref_frame;
1408       set_ref_ptrs(cm, xd, mi->ref_frame[0], NONE);
1409       mi->interp_filter = ctx_den->best_pred_filter;
1410       if (ctx_den->best_ref_frame == INTRA_FRAME) {
1411         mi->mv[0].as_int = INVALID_MV;
1412         mi->interp_filter = SWITCHABLE_FILTERS;
1413       } else if (ctx_den->best_ref_frame == GOLDEN_FRAME) {
1414         mi->mv[0].as_int =
1415             ctx_den->frame_mv[ctx_den->best_mode][ctx_den->best_ref_frame]
1416                 .as_int;
1417         if (ctx_den->reuse_inter_pred) {
1418           xd->plane[0].pre[0] = yv12_mb[GOLDEN_FRAME][0];
1419           vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1420         }
1421       }
1422       mi->tx_size = ctx_den->best_tx_size;
1423       x->skip_txfm[0] = ctx_den->best_mode_skip_txfm;
1424     } else {
1425       ctx_den->best_ref_frame = LAST_FRAME;
1426       *best_rdc = this_rdc;
1427     }
1428   }
1429 }
1430 #endif  // CONFIG_VP9_TEMPORAL_DENOISING
1431 
get_force_skip_low_temp_var(uint8_t * variance_low,int mi_row,int mi_col,BLOCK_SIZE bsize)1432 static INLINE int get_force_skip_low_temp_var(uint8_t *variance_low, int mi_row,
1433                                               int mi_col, BLOCK_SIZE bsize) {
1434   const int i = (mi_row & 0x7) >> 1;
1435   const int j = (mi_col & 0x7) >> 1;
1436   int force_skip_low_temp_var = 0;
1437   // Set force_skip_low_temp_var based on the block size and block offset.
1438   if (bsize == BLOCK_64X64) {
1439     force_skip_low_temp_var = variance_low[0];
1440   } else if (bsize == BLOCK_64X32) {
1441     if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1442       force_skip_low_temp_var = variance_low[1];
1443     } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1444       force_skip_low_temp_var = variance_low[2];
1445     }
1446   } else if (bsize == BLOCK_32X64) {
1447     if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1448       force_skip_low_temp_var = variance_low[3];
1449     } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1450       force_skip_low_temp_var = variance_low[4];
1451     }
1452   } else if (bsize == BLOCK_32X32) {
1453     if (!(mi_col & 0x7) && !(mi_row & 0x7)) {
1454       force_skip_low_temp_var = variance_low[5];
1455     } else if ((mi_col & 0x7) && !(mi_row & 0x7)) {
1456       force_skip_low_temp_var = variance_low[6];
1457     } else if (!(mi_col & 0x7) && (mi_row & 0x7)) {
1458       force_skip_low_temp_var = variance_low[7];
1459     } else if ((mi_col & 0x7) && (mi_row & 0x7)) {
1460       force_skip_low_temp_var = variance_low[8];
1461     }
1462   } else if (bsize == BLOCK_16X16) {
1463     force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]];
1464   } else if (bsize == BLOCK_32X16) {
1465     // The col shift index for the second 16x16 block.
1466     const int j2 = ((mi_col + 2) & 0x7) >> 1;
1467     // Only if each 16x16 block inside has low temporal variance.
1468     force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1469                               variance_low[pos_shift_16x16[i][j2]];
1470   } else if (bsize == BLOCK_16X32) {
1471     // The row shift index for the second 16x16 block.
1472     const int i2 = ((mi_row + 2) & 0x7) >> 1;
1473     force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]] &&
1474                               variance_low[pos_shift_16x16[i2][j]];
1475   }
1476   return force_skip_low_temp_var;
1477 }
1478 
search_filter_ref(VP9_COMP * cpi,MACROBLOCK * x,RD_COST * this_rdc,int mi_row,int mi_col,PRED_BUFFER * tmp,BLOCK_SIZE bsize,int reuse_inter_pred,PRED_BUFFER ** this_mode_pred,unsigned int * var_y,unsigned int * sse_y,int force_smooth_filter,int * this_early_term,int * flag_preduv_computed,int use_model_yrd_large)1479 static void search_filter_ref(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *this_rdc,
1480                               int mi_row, int mi_col, PRED_BUFFER *tmp,
1481                               BLOCK_SIZE bsize, int reuse_inter_pred,
1482                               PRED_BUFFER **this_mode_pred, unsigned int *var_y,
1483                               unsigned int *sse_y, int force_smooth_filter,
1484                               int *this_early_term, int *flag_preduv_computed,
1485                               int use_model_yrd_large) {
1486   MACROBLOCKD *const xd = &x->e_mbd;
1487   MODE_INFO *const mi = xd->mi[0];
1488   struct macroblockd_plane *const pd = &xd->plane[0];
1489   const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
1490 
1491   int pf_rate[3] = { 0 };
1492   int64_t pf_dist[3] = { 0 };
1493   int curr_rate[3] = { 0 };
1494   unsigned int pf_var[3] = { 0 };
1495   unsigned int pf_sse[3] = { 0 };
1496   TX_SIZE pf_tx_size[3] = { 0 };
1497   int64_t best_cost = INT64_MAX;
1498   INTERP_FILTER best_filter = SWITCHABLE, filter;
1499   PRED_BUFFER *current_pred = *this_mode_pred;
1500   uint8_t skip_txfm = SKIP_TXFM_NONE;
1501   int best_early_term = 0;
1502   int best_flag_preduv_computed[2] = { 0 };
1503   INTERP_FILTER filter_start = force_smooth_filter ? EIGHTTAP_SMOOTH : EIGHTTAP;
1504   INTERP_FILTER filter_end = EIGHTTAP_SMOOTH;
1505   for (filter = filter_start; filter <= filter_end; ++filter) {
1506     int64_t cost;
1507     mi->interp_filter = filter;
1508     vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1509     // For large partition blocks, extra testing is done.
1510     if (use_model_yrd_large)
1511       model_rd_for_sb_y_large(cpi, bsize, x, xd, &pf_rate[filter],
1512                               &pf_dist[filter], &pf_var[filter],
1513                               &pf_sse[filter], mi_row, mi_col, this_early_term,
1514                               flag_preduv_computed);
1515     else
1516       model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[filter], &pf_dist[filter],
1517                         &pf_var[filter], &pf_sse[filter], 0);
1518     curr_rate[filter] = pf_rate[filter];
1519     pf_rate[filter] += vp9_get_switchable_rate(cpi, xd);
1520     cost = RDCOST(x->rdmult, x->rddiv, pf_rate[filter], pf_dist[filter]);
1521     pf_tx_size[filter] = mi->tx_size;
1522     if (cost < best_cost) {
1523       best_filter = filter;
1524       best_cost = cost;
1525       skip_txfm = x->skip_txfm[0];
1526       best_early_term = *this_early_term;
1527       best_flag_preduv_computed[0] = flag_preduv_computed[0];
1528       best_flag_preduv_computed[1] = flag_preduv_computed[1];
1529 
1530       if (reuse_inter_pred) {
1531         if (*this_mode_pred != current_pred) {
1532           free_pred_buffer(*this_mode_pred);
1533           *this_mode_pred = current_pred;
1534         }
1535         if (filter != filter_end) {
1536           current_pred = &tmp[get_pred_buffer(tmp, 3)];
1537           pd->dst.buf = current_pred->data;
1538           pd->dst.stride = bw;
1539         }
1540       }
1541     }
1542   }
1543 
1544   if (reuse_inter_pred && *this_mode_pred != current_pred)
1545     free_pred_buffer(current_pred);
1546 
1547   mi->interp_filter = best_filter;
1548   mi->tx_size = pf_tx_size[best_filter];
1549   this_rdc->rate = curr_rate[best_filter];
1550   this_rdc->dist = pf_dist[best_filter];
1551   *var_y = pf_var[best_filter];
1552   *sse_y = pf_sse[best_filter];
1553   x->skip_txfm[0] = skip_txfm;
1554   *this_early_term = best_early_term;
1555   flag_preduv_computed[0] = best_flag_preduv_computed[0];
1556   flag_preduv_computed[1] = best_flag_preduv_computed[1];
1557   if (reuse_inter_pred) {
1558     pd->dst.buf = (*this_mode_pred)->data;
1559     pd->dst.stride = (*this_mode_pred)->stride;
1560   } else if (best_filter < filter_end) {
1561     mi->interp_filter = best_filter;
1562     vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
1563   }
1564 }
1565 
search_new_mv(VP9_COMP * cpi,MACROBLOCK * x,int_mv frame_mv[][MAX_REF_FRAMES],MV_REFERENCE_FRAME ref_frame,int gf_temporal_ref,BLOCK_SIZE bsize,int mi_row,int mi_col,int best_pred_sad,int * rate_mv,unsigned int best_sse_sofar,RD_COST * best_rdc)1566 static int search_new_mv(VP9_COMP *cpi, MACROBLOCK *x,
1567                          int_mv frame_mv[][MAX_REF_FRAMES],
1568                          MV_REFERENCE_FRAME ref_frame, int gf_temporal_ref,
1569                          BLOCK_SIZE bsize, int mi_row, int mi_col,
1570                          int best_pred_sad, int *rate_mv,
1571                          unsigned int best_sse_sofar, RD_COST *best_rdc) {
1572   SVC *const svc = &cpi->svc;
1573   MACROBLOCKD *const xd = &x->e_mbd;
1574   MODE_INFO *const mi = xd->mi[0];
1575   SPEED_FEATURES *const sf = &cpi->sf;
1576 
1577   if (ref_frame > LAST_FRAME && gf_temporal_ref &&
1578       cpi->oxcf.rc_mode == VPX_CBR) {
1579     int tmp_sad;
1580     uint32_t dis;
1581     int cost_list[5] = { INT_MAX, INT_MAX, INT_MAX, INT_MAX, INT_MAX };
1582 
1583     if (bsize < BLOCK_16X16) return -1;
1584 
1585     tmp_sad = vp9_int_pro_motion_estimation(
1586         cpi, x, bsize, mi_row, mi_col,
1587         &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv);
1588 
1589     if (tmp_sad > x->pred_mv_sad[LAST_FRAME]) return -1;
1590     if (tmp_sad + (num_pels_log2_lookup[bsize] << 4) > best_pred_sad) return -1;
1591 
1592     frame_mv[NEWMV][ref_frame].as_int = mi->mv[0].as_int;
1593     *rate_mv = vp9_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv,
1594                                &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1595                                x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
1596     frame_mv[NEWMV][ref_frame].as_mv.row >>= 3;
1597     frame_mv[NEWMV][ref_frame].as_mv.col >>= 3;
1598 
1599     cpi->find_fractional_mv_step(
1600         x, &frame_mv[NEWMV][ref_frame].as_mv,
1601         &x->mbmi_ext->ref_mvs[ref_frame][0].as_mv,
1602         cpi->common.allow_high_precision_mv, x->errorperbit,
1603         &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
1604         cpi->sf.mv.subpel_search_level, cond_cost_list(cpi, cost_list),
1605         x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref_frame], NULL, 0, 0,
1606         cpi->sf.use_accurate_subpel_search);
1607   } else if (svc->use_base_mv && svc->spatial_layer_id) {
1608     if (frame_mv[NEWMV][ref_frame].as_int != INVALID_MV) {
1609       const int pre_stride = xd->plane[0].pre[0].stride;
1610       unsigned int base_mv_sse = UINT_MAX;
1611       int scale = (cpi->rc.avg_frame_low_motion > 60) ? 2 : 4;
1612       const uint8_t *const pre_buf =
1613           xd->plane[0].pre[0].buf +
1614           (frame_mv[NEWMV][ref_frame].as_mv.row >> 3) * pre_stride +
1615           (frame_mv[NEWMV][ref_frame].as_mv.col >> 3);
1616       cpi->fn_ptr[bsize].vf(x->plane[0].src.buf, x->plane[0].src.stride,
1617                             pre_buf, pre_stride, &base_mv_sse);
1618 
1619       // Exit NEWMV search if base_mv is (0,0) && bsize < BLOCK_16x16,
1620       // for SVC encoding.
1621       if (cpi->use_svc && svc->use_base_mv && bsize < BLOCK_16X16 &&
1622           frame_mv[NEWMV][ref_frame].as_mv.row == 0 &&
1623           frame_mv[NEWMV][ref_frame].as_mv.col == 0)
1624         return -1;
1625 
1626       // Exit NEWMV search if base_mv_sse is large.
1627       if (sf->base_mv_aggressive && base_mv_sse > (best_sse_sofar << scale))
1628         return -1;
1629       if (base_mv_sse < (best_sse_sofar << 1)) {
1630         // Base layer mv is good.
1631         // Exit NEWMV search if the base_mv is (0, 0) and sse is low, since
1632         // (0, 0) mode is already tested.
1633         unsigned int base_mv_sse_normalized =
1634             base_mv_sse >>
1635             (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
1636         if (sf->base_mv_aggressive && base_mv_sse <= best_sse_sofar &&
1637             base_mv_sse_normalized < 400 &&
1638             frame_mv[NEWMV][ref_frame].as_mv.row == 0 &&
1639             frame_mv[NEWMV][ref_frame].as_mv.col == 0)
1640           return -1;
1641         if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1642                                     &frame_mv[NEWMV][ref_frame], rate_mv,
1643                                     best_rdc->rdcost, 1)) {
1644           return -1;
1645         }
1646       } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1647                                          &frame_mv[NEWMV][ref_frame], rate_mv,
1648                                          best_rdc->rdcost, 0)) {
1649         return -1;
1650       }
1651     } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1652                                        &frame_mv[NEWMV][ref_frame], rate_mv,
1653                                        best_rdc->rdcost, 0)) {
1654       return -1;
1655     }
1656   } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
1657                                      &frame_mv[NEWMV][ref_frame], rate_mv,
1658                                      best_rdc->rdcost, 0)) {
1659     return -1;
1660   }
1661 
1662   return 0;
1663 }
1664 
init_best_pickmode(BEST_PICKMODE * bp)1665 static INLINE void init_best_pickmode(BEST_PICKMODE *bp) {
1666   bp->best_mode = ZEROMV;
1667   bp->best_ref_frame = LAST_FRAME;
1668   bp->best_tx_size = TX_SIZES;
1669   bp->best_intra_tx_size = TX_SIZES;
1670   bp->best_pred_filter = EIGHTTAP;
1671   bp->best_mode_skip_txfm = SKIP_TXFM_NONE;
1672   bp->best_second_ref_frame = NONE;
1673   bp->best_pred = NULL;
1674 }
1675 
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)1676 void vp9_pick_inter_mode(VP9_COMP *cpi, MACROBLOCK *x, TileDataEnc *tile_data,
1677                          int mi_row, int mi_col, RD_COST *rd_cost,
1678                          BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1679   VP9_COMMON *const cm = &cpi->common;
1680   SPEED_FEATURES *const sf = &cpi->sf;
1681   SVC *const svc = &cpi->svc;
1682   MACROBLOCKD *const xd = &x->e_mbd;
1683   MODE_INFO *const mi = xd->mi[0];
1684   struct macroblockd_plane *const pd = &xd->plane[0];
1685 
1686   BEST_PICKMODE best_pickmode;
1687 
1688   MV_REFERENCE_FRAME ref_frame;
1689   MV_REFERENCE_FRAME usable_ref_frame, second_ref_frame;
1690   int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
1691   uint8_t mode_checked[MB_MODE_COUNT][MAX_REF_FRAMES];
1692   struct buf_2d yv12_mb[4][MAX_MB_PLANE];
1693   static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
1694                                     VP9_ALT_FLAG };
1695   RD_COST this_rdc, best_rdc;
1696   // var_y and sse_y are saved to be used in skipping checking
1697   unsigned int var_y = UINT_MAX;
1698   unsigned int sse_y = UINT_MAX;
1699   const int intra_cost_penalty =
1700       vp9_get_intra_cost_penalty(cpi, bsize, cm->base_qindex, cm->y_dc_delta_q);
1701   int64_t inter_mode_thresh =
1702       RDCOST(x->rdmult, x->rddiv, intra_cost_penalty, 0);
1703   const int *const rd_threshes = cpi->rd.threshes[mi->segment_id][bsize];
1704   const int sb_row = mi_row >> MI_BLOCK_SIZE_LOG2;
1705   int thresh_freq_fact_idx = (sb_row * BLOCK_SIZES + bsize) * MAX_MODES;
1706   const int *const rd_thresh_freq_fact =
1707       (cpi->sf.adaptive_rd_thresh_row_mt)
1708           ? &(tile_data->row_base_thresh_freq_fact[thresh_freq_fact_idx])
1709           : tile_data->thresh_freq_fact[bsize];
1710 #if CONFIG_VP9_TEMPORAL_DENOISING
1711   const int denoise_recheck_zeromv = 1;
1712 #endif
1713   INTERP_FILTER filter_ref;
1714   int pred_filter_search = cm->interp_filter == SWITCHABLE;
1715   int const_motion[MAX_REF_FRAMES] = { 0 };
1716   const int bh = num_4x4_blocks_high_lookup[bsize] << 2;
1717   const int bw = num_4x4_blocks_wide_lookup[bsize] << 2;
1718   // For speed 6, the result of interp filter is reused later in actual encoding
1719   // process.
1720   // tmp[3] points to dst buffer, and the other 3 point to allocated buffers.
1721   PRED_BUFFER tmp[4];
1722   DECLARE_ALIGNED(16, uint8_t, pred_buf[3 * 64 * 64] VPX_UNINITIALIZED);
1723 #if CONFIG_VP9_HIGHBITDEPTH
1724   DECLARE_ALIGNED(16, uint16_t, pred_buf_16[3 * 64 * 64] VPX_UNINITIALIZED);
1725 #endif
1726   struct buf_2d orig_dst = pd->dst;
1727   PRED_BUFFER *this_mode_pred = NULL;
1728   const int pixels_in_block = bh * bw;
1729   int reuse_inter_pred = cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready;
1730   int ref_frame_skip_mask = 0;
1731   int idx;
1732   int best_pred_sad = INT_MAX;
1733   int best_early_term = 0;
1734   int ref_frame_cost[MAX_REF_FRAMES];
1735   int svc_force_zero_mode[3] = { 0 };
1736   int perform_intra_pred = 1;
1737   int use_golden_nonzeromv = 1;
1738   int force_skip_low_temp_var = 0;
1739   int skip_ref_find_pred[4] = { 0 };
1740   unsigned int sse_zeromv_normalized = UINT_MAX;
1741   unsigned int best_sse_sofar = UINT_MAX;
1742   int gf_temporal_ref = 0;
1743   int force_test_gf_zeromv = 0;
1744 #if CONFIG_VP9_TEMPORAL_DENOISING
1745   VP9_PICKMODE_CTX_DEN ctx_den;
1746   int64_t zero_last_cost_orig = INT64_MAX;
1747   int denoise_svc_pickmode = 1;
1748 #endif
1749   INTERP_FILTER filter_gf_svc = EIGHTTAP;
1750   MV_REFERENCE_FRAME inter_layer_ref = GOLDEN_FRAME;
1751   const struct segmentation *const seg = &cm->seg;
1752   int comp_modes = 0;
1753   int num_inter_modes = (cpi->use_svc) ? RT_INTER_MODES_SVC : RT_INTER_MODES;
1754   int flag_svc_subpel = 0;
1755   int svc_mv_col = 0;
1756   int svc_mv_row = 0;
1757   int no_scaling = 0;
1758   int large_block = 0;
1759   int use_model_yrd_large = 0;
1760   unsigned int thresh_svc_skip_golden = 500;
1761   unsigned int thresh_skip_golden = 500;
1762   int force_smooth_filter = cpi->sf.force_smooth_interpol;
1763   int scene_change_detected =
1764       cpi->rc.high_source_sad ||
1765       (cpi->use_svc && cpi->svc.high_source_sad_superframe);
1766 
1767   init_best_pickmode(&best_pickmode);
1768 
1769   x->encode_breakout = seg->enabled
1770                            ? cpi->segment_encode_breakout[mi->segment_id]
1771                            : cpi->encode_breakout;
1772 
1773   x->source_variance = UINT_MAX;
1774   if (cpi->sf.default_interp_filter == BILINEAR) {
1775     best_pickmode.best_pred_filter = BILINEAR;
1776     filter_gf_svc = BILINEAR;
1777   }
1778   if (cpi->use_svc && svc->spatial_layer_id > 0) {
1779     int layer =
1780         LAYER_IDS_TO_IDX(svc->spatial_layer_id - 1, svc->temporal_layer_id,
1781                          svc->number_temporal_layers);
1782     LAYER_CONTEXT *const lc = &svc->layer_context[layer];
1783     if (lc->scaling_factor_num == lc->scaling_factor_den) no_scaling = 1;
1784   }
1785   if (svc->spatial_layer_id > 0 &&
1786       (svc->high_source_sad_superframe || no_scaling))
1787     thresh_svc_skip_golden = 0;
1788   // Lower the skip threshold if lower spatial layer is better quality relative
1789   // to current layer.
1790   else if (svc->spatial_layer_id > 0 && cm->base_qindex > 150 &&
1791            cm->base_qindex > svc->lower_layer_qindex + 15)
1792     thresh_svc_skip_golden = 100;
1793   // Increase skip threshold if lower spatial layer is lower quality relative
1794   // to current layer.
1795   else if (svc->spatial_layer_id > 0 && cm->base_qindex < 140 &&
1796            cm->base_qindex < svc->lower_layer_qindex - 20)
1797     thresh_svc_skip_golden = 1000;
1798 
1799   if (!cpi->use_svc ||
1800       (svc->use_gf_temporal_ref_current_layer &&
1801        !svc->layer_context[svc->temporal_layer_id].is_key_frame)) {
1802     struct scale_factors *const sf_last = &cm->frame_refs[LAST_FRAME - 1].sf;
1803     struct scale_factors *const sf_golden =
1804         &cm->frame_refs[GOLDEN_FRAME - 1].sf;
1805     gf_temporal_ref = 1;
1806     // For temporal long term prediction, check that the golden reference
1807     // is same scale as last reference, otherwise disable.
1808     if ((sf_last->x_scale_fp != sf_golden->x_scale_fp) ||
1809         (sf_last->y_scale_fp != sf_golden->y_scale_fp)) {
1810       gf_temporal_ref = 0;
1811     } else {
1812       if (cpi->rc.avg_frame_low_motion > 70)
1813         thresh_svc_skip_golden = 500;
1814       else
1815         thresh_svc_skip_golden = 0;
1816     }
1817   }
1818 
1819   init_ref_frame_cost(cm, xd, ref_frame_cost);
1820   memset(&mode_checked[0][0], 0, MB_MODE_COUNT * MAX_REF_FRAMES);
1821 
1822   if (reuse_inter_pred) {
1823     int i;
1824     for (i = 0; i < 3; i++) {
1825 #if CONFIG_VP9_HIGHBITDEPTH
1826       if (cm->use_highbitdepth)
1827         tmp[i].data = CONVERT_TO_BYTEPTR(&pred_buf_16[pixels_in_block * i]);
1828       else
1829         tmp[i].data = &pred_buf[pixels_in_block * i];
1830 #else
1831       tmp[i].data = &pred_buf[pixels_in_block * i];
1832 #endif  // CONFIG_VP9_HIGHBITDEPTH
1833       tmp[i].stride = bw;
1834       tmp[i].in_use = 0;
1835     }
1836     tmp[3].data = pd->dst.buf;
1837     tmp[3].stride = pd->dst.stride;
1838     tmp[3].in_use = 0;
1839   }
1840 
1841   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
1842   x->skip = 0;
1843 
1844   if (cpi->sf.cb_pred_filter_search) {
1845     const int bsl = mi_width_log2_lookup[bsize];
1846     pred_filter_search = cm->interp_filter == SWITCHABLE
1847                              ? (((mi_row + mi_col) >> bsl) +
1848                                 get_chessboard_index(cm->current_video_frame)) &
1849                                    0x1
1850                              : 0;
1851   }
1852   // Instead of using vp9_get_pred_context_switchable_interp(xd) to assign
1853   // filter_ref, we use a less strict condition on assigning filter_ref.
1854   // This is to reduce the probabily of entering the flow of not assigning
1855   // filter_ref and then skip filter search.
1856   filter_ref = cm->interp_filter;
1857   if (cpi->sf.default_interp_filter != BILINEAR) {
1858     if (xd->above_mi && is_inter_block(xd->above_mi))
1859       filter_ref = xd->above_mi->interp_filter;
1860     else if (xd->left_mi && is_inter_block(xd->left_mi))
1861       filter_ref = xd->left_mi->interp_filter;
1862   }
1863 
1864   // initialize mode decisions
1865   vp9_rd_cost_reset(&best_rdc);
1866   vp9_rd_cost_reset(rd_cost);
1867   mi->sb_type = bsize;
1868   mi->ref_frame[0] = NONE;
1869   mi->ref_frame[1] = NONE;
1870 
1871   mi->tx_size =
1872       VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[cm->tx_mode]);
1873 
1874   if (sf->short_circuit_flat_blocks || sf->limit_newmv_early_exit) {
1875 #if CONFIG_VP9_HIGHBITDEPTH
1876     if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
1877       x->source_variance = vp9_high_get_sby_perpixel_variance(
1878           cpi, &x->plane[0].src, bsize, xd->bd);
1879     else
1880 #endif  // CONFIG_VP9_HIGHBITDEPTH
1881       x->source_variance =
1882           vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1883 
1884     if (cpi->oxcf.content == VP9E_CONTENT_SCREEN &&
1885         cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && mi->segment_id > 0 &&
1886         x->zero_temp_sad_source && x->source_variance == 0) {
1887       mi->segment_id = 0;
1888       vp9_init_plane_quantizers(cpi, x);
1889     }
1890   }
1891 
1892 #if CONFIG_VP9_TEMPORAL_DENOISING
1893   if (cpi->oxcf.noise_sensitivity > 0) {
1894     if (cpi->use_svc) denoise_svc_pickmode = vp9_denoise_svc_non_key(cpi);
1895     if (cpi->denoiser.denoising_level > kDenLowLow && denoise_svc_pickmode)
1896       vp9_denoiser_reset_frame_stats(ctx);
1897   }
1898 #endif
1899 
1900   if (cpi->rc.frames_since_golden == 0 && gf_temporal_ref &&
1901       !cpi->rc.alt_ref_gf_group && !cpi->rc.last_frame_is_src_altref) {
1902     usable_ref_frame = LAST_FRAME;
1903   } else {
1904     usable_ref_frame = GOLDEN_FRAME;
1905   }
1906 
1907   if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
1908     if (cpi->rc.alt_ref_gf_group || cpi->rc.is_src_frame_alt_ref)
1909       usable_ref_frame = ALTREF_FRAME;
1910 
1911     if (cpi->rc.is_src_frame_alt_ref) {
1912       skip_ref_find_pred[LAST_FRAME] = 1;
1913       skip_ref_find_pred[GOLDEN_FRAME] = 1;
1914     }
1915     if (!cm->show_frame) {
1916       if (cpi->rc.frames_since_key == 1) {
1917         usable_ref_frame = LAST_FRAME;
1918         skip_ref_find_pred[GOLDEN_FRAME] = 1;
1919         skip_ref_find_pred[ALTREF_FRAME] = 1;
1920       }
1921     }
1922   }
1923 
1924   // For svc mode, on spatial_layer_id > 0: if the reference has different scale
1925   // constrain the inter mode to only test zero motion.
1926   if (cpi->use_svc && svc->force_zero_mode_spatial_ref &&
1927       svc->spatial_layer_id > 0 && !gf_temporal_ref) {
1928     if (cpi->ref_frame_flags & flag_list[LAST_FRAME]) {
1929       struct scale_factors *const sf = &cm->frame_refs[LAST_FRAME - 1].sf;
1930       if (vp9_is_scaled(sf)) {
1931         svc_force_zero_mode[LAST_FRAME - 1] = 1;
1932         inter_layer_ref = LAST_FRAME;
1933       }
1934     }
1935     if (cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) {
1936       struct scale_factors *const sf = &cm->frame_refs[GOLDEN_FRAME - 1].sf;
1937       if (vp9_is_scaled(sf)) {
1938         svc_force_zero_mode[GOLDEN_FRAME - 1] = 1;
1939         inter_layer_ref = GOLDEN_FRAME;
1940       }
1941     }
1942   }
1943 
1944   if (cpi->sf.short_circuit_low_temp_var) {
1945     force_skip_low_temp_var =
1946         get_force_skip_low_temp_var(&x->variance_low[0], mi_row, mi_col, bsize);
1947     // If force_skip_low_temp_var is set, and for short circuit mode = 1 and 3,
1948     // skip golden reference.
1949     if ((cpi->sf.short_circuit_low_temp_var == 1 ||
1950          cpi->sf.short_circuit_low_temp_var == 3) &&
1951         force_skip_low_temp_var) {
1952       usable_ref_frame = LAST_FRAME;
1953     }
1954   }
1955 
1956   if (sf->disable_golden_ref && (x->content_state_sb != kVeryHighSad ||
1957                                  cpi->rc.avg_frame_low_motion < 60))
1958     usable_ref_frame = LAST_FRAME;
1959 
1960   if (!((cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) &&
1961         !svc_force_zero_mode[GOLDEN_FRAME - 1] && !force_skip_low_temp_var))
1962     use_golden_nonzeromv = 0;
1963 
1964   if (cpi->oxcf.speed >= 8 && !cpi->use_svc &&
1965       ((cpi->rc.frames_since_golden + 1) < x->last_sb_high_content ||
1966        x->last_sb_high_content > 40 || cpi->rc.frames_since_golden > 120))
1967     usable_ref_frame = LAST_FRAME;
1968 
1969   // Compound prediction modes: (0,0) on LAST/GOLDEN and ARF.
1970   if (cm->reference_mode == REFERENCE_MODE_SELECT &&
1971       cpi->sf.use_compound_nonrd_pickmode && usable_ref_frame == ALTREF_FRAME)
1972     comp_modes = 2;
1973 
1974   // If the segment reference frame feature is enabled and it's set to GOLDEN
1975   // reference, then make sure we don't skip checking GOLDEN, this is to
1976   // prevent possibility of not picking any mode.
1977   if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME) &&
1978       get_segdata(seg, mi->segment_id, SEG_LVL_REF_FRAME) == GOLDEN_FRAME) {
1979     usable_ref_frame = GOLDEN_FRAME;
1980     skip_ref_find_pred[GOLDEN_FRAME] = 0;
1981     thresh_svc_skip_golden = 0;
1982   }
1983 
1984   for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) {
1985     // Skip find_predictor if the reference frame is not in the
1986     // ref_frame_flags (i.e., not used as a reference for this frame).
1987     skip_ref_find_pred[ref_frame] =
1988         !(cpi->ref_frame_flags & flag_list[ref_frame]);
1989     if (!skip_ref_find_pred[ref_frame]) {
1990       find_predictors(cpi, x, ref_frame, frame_mv, const_motion,
1991                       &ref_frame_skip_mask, flag_list, tile_data, mi_row,
1992                       mi_col, yv12_mb, bsize, force_skip_low_temp_var,
1993                       comp_modes > 0);
1994     }
1995   }
1996 
1997   if (cpi->use_svc || cpi->oxcf.speed <= 7 || bsize < BLOCK_32X32)
1998     x->sb_use_mv_part = 0;
1999 
2000   // Set the flag_svc_subpel to 1 for SVC if the lower spatial layer used
2001   // an averaging filter for downsampling (phase = 8). If so, we will test
2002   // a nonzero motion mode on the spatial reference.
2003   // The nonzero motion is half pixel shifted to left and top (-4, -4).
2004   if (cpi->use_svc && svc->spatial_layer_id > 0 &&
2005       svc_force_zero_mode[inter_layer_ref - 1] &&
2006       svc->downsample_filter_phase[svc->spatial_layer_id - 1] == 8 &&
2007       !gf_temporal_ref) {
2008     svc_mv_col = -4;
2009     svc_mv_row = -4;
2010     flag_svc_subpel = 1;
2011   }
2012 
2013   // For SVC with quality layers, when QP of lower layer is lower
2014   // than current layer: force check of GF-ZEROMV before early exit
2015   // due to skip flag.
2016   if (svc->spatial_layer_id > 0 && no_scaling &&
2017       (cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) &&
2018       cm->base_qindex > svc->lower_layer_qindex + 10)
2019     force_test_gf_zeromv = 1;
2020 
2021   // For low motion content use x->sb_is_skin in addition to VeryHighSad
2022   // for setting large_block.
2023   large_block = (x->content_state_sb == kVeryHighSad ||
2024                  (x->sb_is_skin && cpi->rc.avg_frame_low_motion > 70) ||
2025                  cpi->oxcf.speed < 7)
2026                     ? bsize > BLOCK_32X32
2027                     : bsize >= BLOCK_32X32;
2028   use_model_yrd_large =
2029       cpi->oxcf.rc_mode == VPX_CBR && large_block &&
2030       !cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id) &&
2031       cm->base_qindex;
2032 
2033   for (idx = 0; idx < num_inter_modes + comp_modes; ++idx) {
2034     int rate_mv = 0;
2035     int mode_rd_thresh;
2036     int mode_index;
2037     int i;
2038     int64_t this_sse;
2039     int is_skippable;
2040     int this_early_term = 0;
2041     int rd_computed = 0;
2042     int flag_preduv_computed[2] = { 0 };
2043     int inter_mv_mode = 0;
2044     int skip_this_mv = 0;
2045     int comp_pred = 0;
2046     int force_mv_inter_layer = 0;
2047     PREDICTION_MODE this_mode;
2048     second_ref_frame = NONE;
2049 
2050     if (idx < num_inter_modes) {
2051       this_mode = ref_mode_set[idx].pred_mode;
2052       ref_frame = ref_mode_set[idx].ref_frame;
2053 
2054       if (cpi->use_svc) {
2055         this_mode = ref_mode_set_svc[idx].pred_mode;
2056         ref_frame = ref_mode_set_svc[idx].ref_frame;
2057       }
2058     } else {
2059       // Add (0,0) compound modes.
2060       this_mode = ZEROMV;
2061       ref_frame = LAST_FRAME;
2062       if (idx == num_inter_modes + comp_modes - 1) ref_frame = GOLDEN_FRAME;
2063       second_ref_frame = ALTREF_FRAME;
2064       comp_pred = 1;
2065     }
2066 
2067     if (ref_frame > usable_ref_frame) continue;
2068     if (skip_ref_find_pred[ref_frame]) continue;
2069 
2070     if (svc->previous_frame_is_intra_only) {
2071       if (ref_frame != LAST_FRAME || frame_mv[this_mode][ref_frame].as_int != 0)
2072         continue;
2073     }
2074 
2075     // If the segment reference frame feature is enabled then do nothing if the
2076     // current ref frame is not allowed.
2077     if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME) &&
2078         get_segdata(seg, mi->segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
2079       continue;
2080 
2081     if (flag_svc_subpel && ref_frame == inter_layer_ref) {
2082       force_mv_inter_layer = 1;
2083       // Only test mode if NEARESTMV/NEARMV is (svc_mv_col, svc_mv_row),
2084       // otherwise set NEWMV to (svc_mv_col, svc_mv_row).
2085       if (this_mode == NEWMV) {
2086         frame_mv[this_mode][ref_frame].as_mv.col = svc_mv_col;
2087         frame_mv[this_mode][ref_frame].as_mv.row = svc_mv_row;
2088       } else if (frame_mv[this_mode][ref_frame].as_mv.col != svc_mv_col ||
2089                  frame_mv[this_mode][ref_frame].as_mv.row != svc_mv_row) {
2090         continue;
2091       }
2092     }
2093 
2094     if (comp_pred) {
2095       if (!cpi->allow_comp_inter_inter) continue;
2096       // Skip compound inter modes if ARF is not available.
2097       if (!(cpi->ref_frame_flags & flag_list[second_ref_frame])) continue;
2098       // Do not allow compound prediction if the segment level reference frame
2099       // feature is in use as in this case there can only be one reference.
2100       if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME)) continue;
2101     }
2102 
2103     // For CBR mode: skip the golden reference search if sse of zeromv_last is
2104     // below threshold.
2105     if (ref_frame == GOLDEN_FRAME && cpi->oxcf.rc_mode == VPX_CBR &&
2106         ((cpi->use_svc && sse_zeromv_normalized < thresh_svc_skip_golden) ||
2107          (!cpi->use_svc && sse_zeromv_normalized < thresh_skip_golden)))
2108       continue;
2109 
2110     if (!(cpi->ref_frame_flags & flag_list[ref_frame])) continue;
2111 
2112     // For screen content. If zero_temp_sad source is computed: skip
2113     // non-zero motion check for stationary blocks. If the superblock is
2114     // non-stationary then for flat blocks skip the zero last check (keep golden
2115     // as it may be inter-layer reference). Otherwise (if zero_temp_sad_source
2116     // is not computed) skip non-zero motion check for flat blocks.
2117     // TODO(marpan): Compute zero_temp_sad_source per coding block.
2118     if (cpi->oxcf.content == VP9E_CONTENT_SCREEN) {
2119       if (cpi->compute_source_sad_onepass && cpi->sf.use_source_sad) {
2120         if ((frame_mv[this_mode][ref_frame].as_int != 0 &&
2121              x->zero_temp_sad_source) ||
2122             (frame_mv[this_mode][ref_frame].as_int == 0 &&
2123              x->source_variance == 0 && ref_frame == LAST_FRAME &&
2124              !x->zero_temp_sad_source))
2125           continue;
2126       } else if (frame_mv[this_mode][ref_frame].as_int != 0 &&
2127                  x->source_variance == 0) {
2128         continue;
2129       }
2130     }
2131 
2132     if (!(cpi->sf.inter_mode_mask[bsize] & (1 << this_mode))) continue;
2133 
2134     if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) {
2135       if (cpi->rc.is_src_frame_alt_ref &&
2136           (ref_frame != ALTREF_FRAME ||
2137            frame_mv[this_mode][ref_frame].as_int != 0))
2138         continue;
2139 
2140       if (!cm->show_frame && ref_frame == ALTREF_FRAME &&
2141           frame_mv[this_mode][ref_frame].as_int != 0)
2142         continue;
2143 
2144       if (cpi->rc.alt_ref_gf_group && cm->show_frame &&
2145           cpi->rc.frames_since_golden > (cpi->rc.baseline_gf_interval >> 1) &&
2146           ref_frame == GOLDEN_FRAME &&
2147           frame_mv[this_mode][ref_frame].as_int != 0)
2148         continue;
2149 
2150       if (cpi->rc.alt_ref_gf_group && cm->show_frame &&
2151           cpi->rc.frames_since_golden > 0 &&
2152           cpi->rc.frames_since_golden < (cpi->rc.baseline_gf_interval >> 1) &&
2153           ref_frame == ALTREF_FRAME &&
2154           frame_mv[this_mode][ref_frame].as_int != 0)
2155         continue;
2156     }
2157 
2158     if (const_motion[ref_frame] && this_mode == NEARMV) continue;
2159 
2160     // Skip non-zeromv mode search for golden frame if force_skip_low_temp_var
2161     // is set. If nearestmv for golden frame is 0, zeromv mode will be skipped
2162     // later.
2163     if (!force_mv_inter_layer && force_skip_low_temp_var &&
2164         ref_frame == GOLDEN_FRAME &&
2165         frame_mv[this_mode][ref_frame].as_int != 0) {
2166       continue;
2167     }
2168 
2169     if (x->content_state_sb != kVeryHighSad &&
2170         (cpi->sf.short_circuit_low_temp_var >= 2 ||
2171          (cpi->sf.short_circuit_low_temp_var == 1 && bsize == BLOCK_64X64)) &&
2172         force_skip_low_temp_var && ref_frame == LAST_FRAME &&
2173         this_mode == NEWMV) {
2174       continue;
2175     }
2176 
2177     if (cpi->use_svc) {
2178       if (!force_mv_inter_layer && svc_force_zero_mode[ref_frame - 1] &&
2179           frame_mv[this_mode][ref_frame].as_int != 0)
2180         continue;
2181     }
2182 
2183     // Disable this drop out case if the ref frame segment level feature is
2184     // enabled for this segment. This is to prevent the possibility that we end
2185     // up unable to pick any mode.
2186     if (!segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME)) {
2187       if (sf->reference_masking &&
2188           !(frame_mv[this_mode][ref_frame].as_int == 0 &&
2189             ref_frame == LAST_FRAME)) {
2190         if (usable_ref_frame < ALTREF_FRAME) {
2191           if (!force_skip_low_temp_var && usable_ref_frame > LAST_FRAME) {
2192             i = (ref_frame == LAST_FRAME) ? GOLDEN_FRAME : LAST_FRAME;
2193             if ((cpi->ref_frame_flags & flag_list[i]))
2194               if (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[i] << 1))
2195                 ref_frame_skip_mask |= (1 << ref_frame);
2196           }
2197         } else if (!cpi->rc.is_src_frame_alt_ref &&
2198                    !(frame_mv[this_mode][ref_frame].as_int == 0 &&
2199                      ref_frame == ALTREF_FRAME)) {
2200           int ref1 = (ref_frame == GOLDEN_FRAME) ? LAST_FRAME : GOLDEN_FRAME;
2201           int ref2 = (ref_frame == ALTREF_FRAME) ? LAST_FRAME : ALTREF_FRAME;
2202           if (((cpi->ref_frame_flags & flag_list[ref1]) &&
2203                (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref1] << 1))) ||
2204               ((cpi->ref_frame_flags & flag_list[ref2]) &&
2205                (x->pred_mv_sad[ref_frame] > (x->pred_mv_sad[ref2] << 1))))
2206             ref_frame_skip_mask |= (1 << ref_frame);
2207         }
2208       }
2209       if (ref_frame_skip_mask & (1 << ref_frame)) continue;
2210     }
2211 
2212     // Select prediction reference frames.
2213     for (i = 0; i < MAX_MB_PLANE; i++) {
2214       xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
2215       if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
2216     }
2217 
2218     mi->ref_frame[0] = ref_frame;
2219     mi->ref_frame[1] = second_ref_frame;
2220     set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
2221 
2222     mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
2223     mode_rd_thresh = best_pickmode.best_mode_skip_txfm
2224                          ? rd_threshes[mode_index] << 1
2225                          : rd_threshes[mode_index];
2226 
2227     // Increase mode_rd_thresh value for GOLDEN_FRAME for improved encoding
2228     // speed with little/no subjective quality loss.
2229     if (cpi->sf.bias_golden && ref_frame == GOLDEN_FRAME &&
2230         cpi->rc.frames_since_golden > 4)
2231       mode_rd_thresh = mode_rd_thresh << 3;
2232 
2233     if ((cpi->sf.adaptive_rd_thresh_row_mt &&
2234          rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh,
2235                                     &rd_thresh_freq_fact[mode_index])) ||
2236         (!cpi->sf.adaptive_rd_thresh_row_mt &&
2237          rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
2238                              &rd_thresh_freq_fact[mode_index])))
2239       if (frame_mv[this_mode][ref_frame].as_int != 0) continue;
2240 
2241     if (this_mode == NEWMV && !force_mv_inter_layer) {
2242       if (search_new_mv(cpi, x, frame_mv, ref_frame, gf_temporal_ref, bsize,
2243                         mi_row, mi_col, best_pred_sad, &rate_mv, best_sse_sofar,
2244                         &best_rdc))
2245         continue;
2246     }
2247 
2248     // TODO(jianj): Skipping the testing of (duplicate) non-zero motion vector
2249     // causes some regression, leave it for duplicate zero-mv for now, until
2250     // regression issue is resolved.
2251     for (inter_mv_mode = NEARESTMV; inter_mv_mode <= NEWMV; inter_mv_mode++) {
2252       if (inter_mv_mode == this_mode || comp_pred) continue;
2253       if (mode_checked[inter_mv_mode][ref_frame] &&
2254           frame_mv[this_mode][ref_frame].as_int ==
2255               frame_mv[inter_mv_mode][ref_frame].as_int &&
2256           frame_mv[inter_mv_mode][ref_frame].as_int == 0) {
2257         skip_this_mv = 1;
2258         break;
2259       }
2260     }
2261 
2262     if (skip_this_mv) continue;
2263 
2264     // If use_golden_nonzeromv is false, NEWMV mode is skipped for golden, no
2265     // need to compute best_pred_sad which is only used to skip golden NEWMV.
2266     if (use_golden_nonzeromv && this_mode == NEWMV && ref_frame == LAST_FRAME &&
2267         frame_mv[NEWMV][LAST_FRAME].as_int != INVALID_MV) {
2268       const int pre_stride = xd->plane[0].pre[0].stride;
2269       const uint8_t *const pre_buf =
2270           xd->plane[0].pre[0].buf +
2271           (frame_mv[NEWMV][LAST_FRAME].as_mv.row >> 3) * pre_stride +
2272           (frame_mv[NEWMV][LAST_FRAME].as_mv.col >> 3);
2273       best_pred_sad = cpi->fn_ptr[bsize].sdf(
2274           x->plane[0].src.buf, x->plane[0].src.stride, pre_buf, pre_stride);
2275       x->pred_mv_sad[LAST_FRAME] = best_pred_sad;
2276     }
2277 
2278     if (this_mode != NEARESTMV && !comp_pred &&
2279         frame_mv[this_mode][ref_frame].as_int ==
2280             frame_mv[NEARESTMV][ref_frame].as_int)
2281       continue;
2282 
2283     mi->mode = this_mode;
2284     mi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
2285     mi->mv[1].as_int = 0;
2286 
2287     // Search for the best prediction filter type, when the resulting
2288     // motion vector is at sub-pixel accuracy level for luma component, i.e.,
2289     // the last three bits are all zeros.
2290     if (reuse_inter_pred) {
2291       if (!this_mode_pred) {
2292         this_mode_pred = &tmp[3];
2293       } else {
2294         this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
2295         pd->dst.buf = this_mode_pred->data;
2296         pd->dst.stride = bw;
2297       }
2298     }
2299 
2300     if ((this_mode == NEWMV || filter_ref == SWITCHABLE) &&
2301         pred_filter_search &&
2302         (ref_frame == LAST_FRAME ||
2303          (ref_frame == GOLDEN_FRAME && !force_mv_inter_layer &&
2304           (cpi->use_svc || cpi->oxcf.rc_mode == VPX_VBR))) &&
2305         (((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07) != 0)) {
2306       rd_computed = 1;
2307       search_filter_ref(cpi, x, &this_rdc, mi_row, mi_col, tmp, bsize,
2308                         reuse_inter_pred, &this_mode_pred, &var_y, &sse_y,
2309                         force_smooth_filter, &this_early_term,
2310                         flag_preduv_computed, use_model_yrd_large);
2311     } else {
2312       mi->interp_filter = (filter_ref == SWITCHABLE) ? EIGHTTAP : filter_ref;
2313 
2314       if (cpi->use_svc && ref_frame == GOLDEN_FRAME &&
2315           svc_force_zero_mode[ref_frame - 1])
2316         mi->interp_filter = filter_gf_svc;
2317 
2318       vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
2319 
2320       // For large partition blocks, extra testing is done.
2321       if (use_model_yrd_large) {
2322         rd_computed = 1;
2323         model_rd_for_sb_y_large(cpi, bsize, x, xd, &this_rdc.rate,
2324                                 &this_rdc.dist, &var_y, &sse_y, mi_row, mi_col,
2325                                 &this_early_term, flag_preduv_computed);
2326       } else {
2327         rd_computed = 1;
2328         model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
2329                           &var_y, &sse_y, 0);
2330       }
2331       // Save normalized sse (between current and last frame) for (0, 0) motion.
2332       if (ref_frame == LAST_FRAME &&
2333           frame_mv[this_mode][ref_frame].as_int == 0) {
2334         sse_zeromv_normalized =
2335             sse_y >> (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
2336       }
2337       if (sse_y < best_sse_sofar) best_sse_sofar = sse_y;
2338     }
2339 
2340     if (!this_early_term) {
2341       this_sse = (int64_t)sse_y;
2342       block_yrd(cpi, x, &this_rdc, &is_skippable, &this_sse, bsize,
2343                 VPXMIN(mi->tx_size, TX_16X16), rd_computed, 0);
2344       x->skip_txfm[0] = is_skippable;
2345       if (is_skippable) {
2346         this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2347       } else {
2348         if (RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist) <
2349             RDCOST(x->rdmult, x->rddiv, 0, this_sse)) {
2350           this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
2351         } else {
2352           this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2353           this_rdc.dist = this_sse;
2354           x->skip_txfm[0] = SKIP_TXFM_AC_DC;
2355         }
2356       }
2357 
2358       if (cm->interp_filter == SWITCHABLE) {
2359         if ((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07)
2360           this_rdc.rate += vp9_get_switchable_rate(cpi, xd);
2361       }
2362     } else {
2363       if (cm->interp_filter == SWITCHABLE) {
2364         if ((mi->mv[0].as_mv.row | mi->mv[0].as_mv.col) & 0x07)
2365           this_rdc.rate += vp9_get_switchable_rate(cpi, xd);
2366       }
2367       this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
2368     }
2369 
2370     if (!this_early_term &&
2371         (x->color_sensitivity[0] || x->color_sensitivity[1])) {
2372       RD_COST rdc_uv;
2373       const BLOCK_SIZE uv_bsize = get_plane_block_size(bsize, &xd->plane[1]);
2374       if (x->color_sensitivity[0] && !flag_preduv_computed[0]) {
2375         vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 1);
2376         flag_preduv_computed[0] = 1;
2377       }
2378       if (x->color_sensitivity[1] && !flag_preduv_computed[1]) {
2379         vp9_build_inter_predictors_sbp(xd, mi_row, mi_col, bsize, 2);
2380         flag_preduv_computed[1] = 1;
2381       }
2382       model_rd_for_sb_uv(cpi, uv_bsize, x, xd, &rdc_uv, &var_y, &sse_y, 1, 2);
2383       this_rdc.rate += rdc_uv.rate;
2384       this_rdc.dist += rdc_uv.dist;
2385     }
2386 
2387     this_rdc.rate += rate_mv;
2388     this_rdc.rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2389                                          [INTER_OFFSET(this_mode)];
2390     // TODO(marpan): Add costing for compound mode.
2391     this_rdc.rate += ref_frame_cost[ref_frame];
2392     this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2393 
2394     // Bias against NEWMV that is very different from its neighbors, and bias
2395     // to small motion-lastref for noisy input.
2396     if (cpi->oxcf.rc_mode == VPX_CBR && cpi->oxcf.speed >= 5 &&
2397         cpi->oxcf.content != VP9E_CONTENT_SCREEN) {
2398       vp9_NEWMV_diff_bias(&cpi->noise_estimate, xd, this_mode, &this_rdc, bsize,
2399                           frame_mv[this_mode][ref_frame].as_mv.row,
2400                           frame_mv[this_mode][ref_frame].as_mv.col,
2401                           ref_frame == LAST_FRAME, x->lowvar_highsumdiff,
2402                           x->sb_is_skin);
2403     }
2404 
2405     // Skipping checking: test to see if this block can be reconstructed by
2406     // prediction only.
2407     if (cpi->allow_encode_breakout && !xd->lossless && !scene_change_detected &&
2408         !svc->high_num_blocks_with_motion) {
2409       encode_breakout_test(cpi, x, bsize, mi_row, mi_col, ref_frame, this_mode,
2410                            var_y, sse_y, yv12_mb, &this_rdc.rate,
2411                            &this_rdc.dist, flag_preduv_computed);
2412       if (x->skip) {
2413         this_rdc.rate += rate_mv;
2414         this_rdc.rdcost =
2415             RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2416       }
2417     }
2418 
2419     // On spatially flat blocks for screne content: bias against zero-last
2420     // if the sse_y is non-zero. Only on scene change or high motion frames.
2421     if (cpi->oxcf.content == VP9E_CONTENT_SCREEN &&
2422         (scene_change_detected || svc->high_num_blocks_with_motion) &&
2423         ref_frame == LAST_FRAME && frame_mv[this_mode][ref_frame].as_int == 0 &&
2424         svc->spatial_layer_id == 0 && x->source_variance == 0 && sse_y > 0) {
2425       this_rdc.rdcost = this_rdc.rdcost << 2;
2426     }
2427 
2428 #if CONFIG_VP9_TEMPORAL_DENOISING
2429     if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc_pickmode &&
2430         cpi->denoiser.denoising_level > kDenLowLow) {
2431       vp9_denoiser_update_frame_stats(mi, sse_y, this_mode, ctx);
2432       // Keep track of zero_last cost.
2433       if (ref_frame == LAST_FRAME && frame_mv[this_mode][ref_frame].as_int == 0)
2434         zero_last_cost_orig = this_rdc.rdcost;
2435     }
2436 #else
2437     (void)ctx;
2438 #endif
2439 
2440     mode_checked[this_mode][ref_frame] = 1;
2441 
2442     if (this_rdc.rdcost < best_rdc.rdcost || x->skip) {
2443       best_rdc = this_rdc;
2444       best_early_term = this_early_term;
2445       best_pickmode.best_mode = this_mode;
2446       best_pickmode.best_pred_filter = mi->interp_filter;
2447       best_pickmode.best_tx_size = mi->tx_size;
2448       best_pickmode.best_ref_frame = ref_frame;
2449       best_pickmode.best_mode_skip_txfm = x->skip_txfm[0];
2450       best_pickmode.best_second_ref_frame = second_ref_frame;
2451 
2452       if (reuse_inter_pred) {
2453         free_pred_buffer(best_pickmode.best_pred);
2454         best_pickmode.best_pred = this_mode_pred;
2455       }
2456     } else {
2457       if (reuse_inter_pred) free_pred_buffer(this_mode_pred);
2458     }
2459 
2460     if (x->skip &&
2461         (!force_test_gf_zeromv || mode_checked[ZEROMV][GOLDEN_FRAME]))
2462       break;
2463 
2464     // If early termination flag is 1 and at least 2 modes are checked,
2465     // the mode search is terminated.
2466     if (best_early_term && idx > 0 && !scene_change_detected &&
2467         (!force_test_gf_zeromv || mode_checked[ZEROMV][GOLDEN_FRAME])) {
2468       x->skip = 1;
2469       break;
2470     }
2471   }
2472 
2473   mi->mode = best_pickmode.best_mode;
2474   mi->interp_filter = best_pickmode.best_pred_filter;
2475   mi->tx_size = best_pickmode.best_tx_size;
2476   mi->ref_frame[0] = best_pickmode.best_ref_frame;
2477   mi->mv[0].as_int =
2478       frame_mv[best_pickmode.best_mode][best_pickmode.best_ref_frame].as_int;
2479   xd->mi[0]->bmi[0].as_mv[0].as_int = mi->mv[0].as_int;
2480   x->skip_txfm[0] = best_pickmode.best_mode_skip_txfm;
2481   mi->ref_frame[1] = best_pickmode.best_second_ref_frame;
2482 
2483   // For spatial enhancemanent layer: perform intra prediction only if base
2484   // layer is chosen as the reference. Always perform intra prediction if
2485   // LAST is the only reference, or is_key_frame is set, or on base
2486   // temporal layer.
2487   if (svc->spatial_layer_id && !gf_temporal_ref) {
2488     perform_intra_pred =
2489         svc->temporal_layer_id == 0 ||
2490         svc->layer_context[svc->temporal_layer_id].is_key_frame ||
2491         !(cpi->ref_frame_flags & flag_list[GOLDEN_FRAME]) ||
2492         (!svc->layer_context[svc->temporal_layer_id].is_key_frame &&
2493          svc_force_zero_mode[best_pickmode.best_ref_frame - 1]);
2494     inter_mode_thresh = (inter_mode_thresh << 1) + inter_mode_thresh;
2495   }
2496   if ((cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR &&
2497        cpi->rc.is_src_frame_alt_ref) ||
2498       svc->previous_frame_is_intra_only)
2499     perform_intra_pred = 0;
2500 
2501   // If the segment reference frame feature is enabled and set then
2502   // skip the intra prediction.
2503   if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME) &&
2504       get_segdata(seg, mi->segment_id, SEG_LVL_REF_FRAME) > 0)
2505     perform_intra_pred = 0;
2506 
2507   // Perform intra prediction search, if the best SAD is above a certain
2508   // threshold.
2509   if (best_rdc.rdcost == INT64_MAX ||
2510       (cpi->oxcf.content == VP9E_CONTENT_SCREEN && x->source_variance == 0) ||
2511       (scene_change_detected && perform_intra_pred) ||
2512       ((!force_skip_low_temp_var || bsize < BLOCK_32X32 ||
2513         x->content_state_sb == kVeryHighSad) &&
2514        perform_intra_pred && !x->skip && best_rdc.rdcost > inter_mode_thresh &&
2515        bsize <= cpi->sf.max_intra_bsize && !x->skip_low_source_sad &&
2516        !x->lowvar_highsumdiff)) {
2517     struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
2518     int64_t this_sse = INT64_MAX;
2519     int i;
2520     PRED_BUFFER *const best_pred = best_pickmode.best_pred;
2521     TX_SIZE intra_tx_size =
2522         VPXMIN(max_txsize_lookup[bsize],
2523                tx_mode_to_biggest_tx_size[cpi->common.tx_mode]);
2524 
2525     if (reuse_inter_pred && best_pred != NULL) {
2526       if (best_pred->data == orig_dst.buf) {
2527         this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
2528 #if CONFIG_VP9_HIGHBITDEPTH
2529         if (cm->use_highbitdepth)
2530           vpx_highbd_convolve_copy(
2531               CONVERT_TO_SHORTPTR(best_pred->data), best_pred->stride,
2532               CONVERT_TO_SHORTPTR(this_mode_pred->data), this_mode_pred->stride,
2533               NULL, 0, 0, 0, 0, bw, bh, xd->bd);
2534         else
2535           vpx_convolve_copy(best_pred->data, best_pred->stride,
2536                             this_mode_pred->data, this_mode_pred->stride, NULL,
2537                             0, 0, 0, 0, bw, bh);
2538 #else
2539         vpx_convolve_copy(best_pred->data, best_pred->stride,
2540                           this_mode_pred->data, this_mode_pred->stride, NULL, 0,
2541                           0, 0, 0, bw, bh);
2542 #endif  // CONFIG_VP9_HIGHBITDEPTH
2543         best_pickmode.best_pred = this_mode_pred;
2544       }
2545     }
2546     pd->dst = orig_dst;
2547 
2548     for (i = 0; i < 4; ++i) {
2549       const PREDICTION_MODE this_mode = intra_mode_list[i];
2550       THR_MODES mode_index = mode_idx[INTRA_FRAME][mode_offset(this_mode)];
2551       int mode_rd_thresh = rd_threshes[mode_index];
2552       // For spatially flat blocks, under short_circuit_flat_blocks flag:
2553       // only check DC mode for stationary blocks, otherwise also check
2554       // H and V mode.
2555       if (sf->short_circuit_flat_blocks && x->source_variance == 0 &&
2556           ((x->zero_temp_sad_source && this_mode != DC_PRED) || i > 2)) {
2557         continue;
2558       }
2559 
2560       if (!((1 << this_mode) & cpi->sf.intra_y_mode_bsize_mask[bsize]))
2561         continue;
2562 
2563       if (cpi->sf.rt_intra_dc_only_low_content && this_mode != DC_PRED &&
2564           x->content_state_sb != kVeryHighSad)
2565         continue;
2566 
2567       if ((cpi->sf.adaptive_rd_thresh_row_mt &&
2568            rd_less_than_thresh_row_mt(best_rdc.rdcost, mode_rd_thresh,
2569                                       &rd_thresh_freq_fact[mode_index])) ||
2570           (!cpi->sf.adaptive_rd_thresh_row_mt &&
2571            rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
2572                                &rd_thresh_freq_fact[mode_index]))) {
2573         // Avoid this early exit for screen on base layer, for scene
2574         // changes or high motion frames.
2575         if (cpi->oxcf.content != VP9E_CONTENT_SCREEN ||
2576             svc->spatial_layer_id > 0 ||
2577             (!scene_change_detected && !svc->high_num_blocks_with_motion))
2578           continue;
2579       }
2580 
2581       mi->mode = this_mode;
2582       mi->ref_frame[0] = INTRA_FRAME;
2583       this_rdc.dist = this_rdc.rate = 0;
2584       args.mode = this_mode;
2585       args.skippable = 1;
2586       args.rdc = &this_rdc;
2587       mi->tx_size = intra_tx_size;
2588 
2589       compute_intra_yprediction(this_mode, bsize, x, xd);
2590       model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
2591                         &var_y, &sse_y, 1);
2592       block_yrd(cpi, x, &this_rdc, &args.skippable, &this_sse, bsize,
2593                 VPXMIN(mi->tx_size, TX_16X16), 1, 1);
2594 
2595       // Check skip cost here since skippable is not set for for uv, this
2596       // mirrors the behavior used by inter
2597       if (args.skippable) {
2598         x->skip_txfm[0] = SKIP_TXFM_AC_DC;
2599         this_rdc.rate = vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 1);
2600       } else {
2601         x->skip_txfm[0] = SKIP_TXFM_NONE;
2602         this_rdc.rate += vp9_cost_bit(vp9_get_skip_prob(&cpi->common, xd), 0);
2603       }
2604       // Inter and intra RD will mismatch in scale for non-screen content.
2605       if (cpi->oxcf.content == VP9E_CONTENT_SCREEN) {
2606         if (x->color_sensitivity[0])
2607           vp9_foreach_transformed_block_in_plane(xd, bsize, 1,
2608                                                  estimate_block_intra, &args);
2609         if (x->color_sensitivity[1])
2610           vp9_foreach_transformed_block_in_plane(xd, bsize, 2,
2611                                                  estimate_block_intra, &args);
2612       }
2613       this_rdc.rate += cpi->mbmode_cost[this_mode];
2614       this_rdc.rate += ref_frame_cost[INTRA_FRAME];
2615       this_rdc.rate += intra_cost_penalty;
2616       this_rdc.rdcost =
2617           RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2618 
2619       if (this_rdc.rdcost < best_rdc.rdcost) {
2620         best_rdc = this_rdc;
2621         best_pickmode.best_mode = this_mode;
2622         best_pickmode.best_intra_tx_size = mi->tx_size;
2623         best_pickmode.best_ref_frame = INTRA_FRAME;
2624         best_pickmode.best_second_ref_frame = NONE;
2625         mi->uv_mode = this_mode;
2626         mi->mv[0].as_int = INVALID_MV;
2627         mi->mv[1].as_int = INVALID_MV;
2628         best_pickmode.best_mode_skip_txfm = x->skip_txfm[0];
2629       }
2630     }
2631 
2632     // Reset mb_mode_info to the best inter mode.
2633     if (best_pickmode.best_ref_frame != INTRA_FRAME) {
2634       mi->tx_size = best_pickmode.best_tx_size;
2635     } else {
2636       mi->tx_size = best_pickmode.best_intra_tx_size;
2637     }
2638   }
2639 
2640   pd->dst = orig_dst;
2641   mi->mode = best_pickmode.best_mode;
2642   mi->ref_frame[0] = best_pickmode.best_ref_frame;
2643   mi->ref_frame[1] = best_pickmode.best_second_ref_frame;
2644   x->skip_txfm[0] = best_pickmode.best_mode_skip_txfm;
2645 
2646   if (!is_inter_block(mi)) {
2647     mi->interp_filter = SWITCHABLE_FILTERS;
2648   }
2649 
2650   if (reuse_inter_pred && best_pickmode.best_pred != NULL) {
2651     PRED_BUFFER *const best_pred = best_pickmode.best_pred;
2652     if (best_pred->data != orig_dst.buf && is_inter_mode(mi->mode)) {
2653 #if CONFIG_VP9_HIGHBITDEPTH
2654       if (cm->use_highbitdepth)
2655         vpx_highbd_convolve_copy(
2656             CONVERT_TO_SHORTPTR(best_pred->data), best_pred->stride,
2657             CONVERT_TO_SHORTPTR(pd->dst.buf), pd->dst.stride, NULL, 0, 0, 0, 0,
2658             bw, bh, xd->bd);
2659       else
2660         vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2661                           pd->dst.stride, NULL, 0, 0, 0, 0, bw, bh);
2662 #else
2663       vpx_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2664                         pd->dst.stride, NULL, 0, 0, 0, 0, bw, bh);
2665 #endif  // CONFIG_VP9_HIGHBITDEPTH
2666     }
2667   }
2668 
2669 #if CONFIG_VP9_TEMPORAL_DENOISING
2670   if (cpi->oxcf.noise_sensitivity > 0 && cpi->resize_pending == 0 &&
2671       denoise_svc_pickmode && cpi->denoiser.denoising_level > kDenLowLow &&
2672       cpi->denoiser.reset == 0) {
2673     VP9_DENOISER_DECISION decision = COPY_BLOCK;
2674     ctx->sb_skip_denoising = 0;
2675     // TODO(marpan): There is an issue with denoising when the
2676     // superblock partitioning scheme is based on the pickmode.
2677     // Remove this condition when the issue is resolved.
2678     if (x->sb_pickmode_part) ctx->sb_skip_denoising = 1;
2679     vp9_pickmode_ctx_den_update(&ctx_den, zero_last_cost_orig, ref_frame_cost,
2680                                 frame_mv, reuse_inter_pred, &best_pickmode);
2681     vp9_denoiser_denoise(cpi, x, mi_row, mi_col, bsize, ctx, &decision,
2682                          gf_temporal_ref);
2683     if (denoise_recheck_zeromv)
2684       recheck_zeromv_after_denoising(cpi, mi, x, xd, decision, &ctx_den,
2685                                      yv12_mb, &best_rdc, bsize, mi_row, mi_col);
2686     best_pickmode.best_ref_frame = ctx_den.best_ref_frame;
2687   }
2688 #endif
2689 
2690   if (best_pickmode.best_ref_frame == ALTREF_FRAME ||
2691       best_pickmode.best_second_ref_frame == ALTREF_FRAME)
2692     x->arf_frame_usage++;
2693   else if (best_pickmode.best_ref_frame != INTRA_FRAME)
2694     x->lastgolden_frame_usage++;
2695 
2696   if (cpi->sf.adaptive_rd_thresh) {
2697     THR_MODES best_mode_idx =
2698         mode_idx[best_pickmode.best_ref_frame][mode_offset(mi->mode)];
2699 
2700     if (best_pickmode.best_ref_frame == INTRA_FRAME) {
2701       // Only consider the modes that are included in the intra_mode_list.
2702       int intra_modes = sizeof(intra_mode_list) / sizeof(PREDICTION_MODE);
2703       int i;
2704 
2705       // TODO(yunqingwang): Check intra mode mask and only update freq_fact
2706       // for those valid modes.
2707       for (i = 0; i < intra_modes; i++) {
2708         if (cpi->sf.adaptive_rd_thresh_row_mt)
2709           update_thresh_freq_fact_row_mt(cpi, tile_data, x->source_variance,
2710                                          thresh_freq_fact_idx, INTRA_FRAME,
2711                                          best_mode_idx, intra_mode_list[i]);
2712         else
2713           update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2714                                   INTRA_FRAME, best_mode_idx,
2715                                   intra_mode_list[i]);
2716       }
2717     } else {
2718       for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2719         PREDICTION_MODE this_mode;
2720         if (best_pickmode.best_ref_frame != ref_frame) continue;
2721         for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2722           if (cpi->sf.adaptive_rd_thresh_row_mt)
2723             update_thresh_freq_fact_row_mt(cpi, tile_data, x->source_variance,
2724                                            thresh_freq_fact_idx, ref_frame,
2725                                            best_mode_idx, this_mode);
2726           else
2727             update_thresh_freq_fact(cpi, tile_data, x->source_variance, bsize,
2728                                     ref_frame, best_mode_idx, this_mode);
2729         }
2730       }
2731     }
2732   }
2733 
2734   *rd_cost = best_rdc;
2735 }
2736 
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)2737 void vp9_pick_inter_mode_sub8x8(VP9_COMP *cpi, MACROBLOCK *x, int mi_row,
2738                                 int mi_col, RD_COST *rd_cost, BLOCK_SIZE bsize,
2739                                 PICK_MODE_CONTEXT *ctx) {
2740   VP9_COMMON *const cm = &cpi->common;
2741   SPEED_FEATURES *const sf = &cpi->sf;
2742   MACROBLOCKD *const xd = &x->e_mbd;
2743   MODE_INFO *const mi = xd->mi[0];
2744   MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2745   const struct segmentation *const seg = &cm->seg;
2746   MV_REFERENCE_FRAME ref_frame, second_ref_frame = NONE;
2747   MV_REFERENCE_FRAME best_ref_frame = NONE;
2748   unsigned char segment_id = mi->segment_id;
2749   struct buf_2d yv12_mb[4][MAX_MB_PLANE];
2750   static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
2751                                     VP9_ALT_FLAG };
2752   int64_t best_rd = INT64_MAX;
2753   b_mode_info bsi[MAX_REF_FRAMES][4];
2754   int ref_frame_skip_mask = 0;
2755   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
2756   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
2757   int idx, idy;
2758 
2759   x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
2760   ctx->pred_pixel_ready = 0;
2761 
2762   for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2763     const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2764     int_mv dummy_mv[2];
2765     x->pred_mv_sad[ref_frame] = INT_MAX;
2766 
2767     if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
2768       int_mv *const candidates = mbmi_ext->ref_mvs[ref_frame];
2769       const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2770       vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf,
2771                            sf);
2772       vp9_find_mv_refs(cm, xd, xd->mi[0], ref_frame, candidates, mi_row, mi_col,
2773                        mbmi_ext->mode_context);
2774 
2775       vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2776                             &dummy_mv[0], &dummy_mv[1]);
2777     } else {
2778       ref_frame_skip_mask |= (1 << ref_frame);
2779     }
2780   }
2781 
2782   mi->sb_type = bsize;
2783   mi->tx_size = TX_4X4;
2784   mi->uv_mode = DC_PRED;
2785   mi->ref_frame[0] = LAST_FRAME;
2786   mi->ref_frame[1] = NONE;
2787   mi->interp_filter =
2788       cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
2789 
2790   for (ref_frame = LAST_FRAME; ref_frame <= GOLDEN_FRAME; ++ref_frame) {
2791     int64_t this_rd = 0;
2792     int plane;
2793 
2794     if (ref_frame_skip_mask & (1 << ref_frame)) continue;
2795 
2796 #if CONFIG_BETTER_HW_COMPATIBILITY
2797     if ((bsize == BLOCK_8X4 || bsize == BLOCK_4X8) && ref_frame > INTRA_FRAME &&
2798         vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2799       continue;
2800 #endif
2801 
2802     // TODO(jingning, agrange): Scaling reference frame not supported for
2803     // sub8x8 blocks. Is this supported now?
2804     if (ref_frame > INTRA_FRAME &&
2805         vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf))
2806       continue;
2807 
2808     // If the segment reference frame feature is enabled....
2809     // then do nothing if the current ref frame is not allowed..
2810     if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
2811         get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
2812       continue;
2813 
2814     mi->ref_frame[0] = ref_frame;
2815     x->skip = 0;
2816     set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
2817 
2818     // Select prediction reference frames.
2819     for (plane = 0; plane < MAX_MB_PLANE; plane++)
2820       xd->plane[plane].pre[0] = yv12_mb[ref_frame][plane];
2821 
2822     for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2823       for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2824         int_mv b_mv[MB_MODE_COUNT];
2825         int64_t b_best_rd = INT64_MAX;
2826         const int i = idy * 2 + idx;
2827         PREDICTION_MODE this_mode;
2828         RD_COST this_rdc;
2829         unsigned int var_y, sse_y;
2830 
2831         struct macroblock_plane *p = &x->plane[0];
2832         struct macroblockd_plane *pd = &xd->plane[0];
2833 
2834         const struct buf_2d orig_src = p->src;
2835         const struct buf_2d orig_dst = pd->dst;
2836         struct buf_2d orig_pre[2];
2837         memcpy(orig_pre, xd->plane[0].pre, sizeof(orig_pre));
2838 
2839         // set buffer pointers for sub8x8 motion search.
2840         p->src.buf =
2841             &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
2842         pd->dst.buf =
2843             &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
2844         pd->pre[0].buf =
2845             &pd->pre[0]
2846                  .buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
2847 
2848         b_mv[ZEROMV].as_int = 0;
2849         b_mv[NEWMV].as_int = INVALID_MV;
2850         vp9_append_sub8x8_mvs_for_idx(cm, xd, i, 0, mi_row, mi_col,
2851                                       &b_mv[NEARESTMV], &b_mv[NEARMV],
2852                                       mbmi_ext->mode_context);
2853 
2854         for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2855           int b_rate = 0;
2856           xd->mi[0]->bmi[i].as_mv[0].as_int = b_mv[this_mode].as_int;
2857 
2858           if (this_mode == NEWMV) {
2859             const int step_param = cpi->sf.mv.fullpel_search_step_param;
2860             MV mvp_full;
2861             MV tmp_mv;
2862             int cost_list[5];
2863             const MvLimits tmp_mv_limits = x->mv_limits;
2864             uint32_t dummy_dist;
2865 
2866             if (i == 0) {
2867               mvp_full.row = b_mv[NEARESTMV].as_mv.row >> 3;
2868               mvp_full.col = b_mv[NEARESTMV].as_mv.col >> 3;
2869             } else {
2870               mvp_full.row = xd->mi[0]->bmi[0].as_mv[0].as_mv.row >> 3;
2871               mvp_full.col = xd->mi[0]->bmi[0].as_mv[0].as_mv.col >> 3;
2872             }
2873 
2874             vp9_set_mv_search_range(&x->mv_limits,
2875                                     &mbmi_ext->ref_mvs[ref_frame][0].as_mv);
2876 
2877             vp9_full_pixel_search(
2878                 cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method,
2879                 x->sadperbit4, cond_cost_list(cpi, cost_list),
2880                 &mbmi_ext->ref_mvs[ref_frame][0].as_mv, &tmp_mv, INT_MAX, 0);
2881 
2882             x->mv_limits = tmp_mv_limits;
2883 
2884             // calculate the bit cost on motion vector
2885             mvp_full.row = tmp_mv.row * 8;
2886             mvp_full.col = tmp_mv.col * 8;
2887 
2888             b_rate += vp9_mv_bit_cost(
2889                 &mvp_full, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2890                 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2891 
2892             b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2893                                           [INTER_OFFSET(NEWMV)];
2894             if (RDCOST(x->rdmult, x->rddiv, b_rate, 0) > b_best_rd) continue;
2895 
2896             cpi->find_fractional_mv_step(
2897                 x, &tmp_mv, &mbmi_ext->ref_mvs[ref_frame][0].as_mv,
2898                 cpi->common.allow_high_precision_mv, x->errorperbit,
2899                 &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
2900                 cpi->sf.mv.subpel_search_level, cond_cost_list(cpi, cost_list),
2901                 x->nmvjointcost, x->mvcost, &dummy_dist,
2902                 &x->pred_sse[ref_frame], NULL, 0, 0,
2903                 cpi->sf.use_accurate_subpel_search);
2904 
2905             xd->mi[0]->bmi[i].as_mv[0].as_mv = tmp_mv;
2906           } else {
2907             b_rate += cpi->inter_mode_cost[x->mbmi_ext->mode_context[ref_frame]]
2908                                           [INTER_OFFSET(this_mode)];
2909           }
2910 
2911 #if CONFIG_VP9_HIGHBITDEPTH
2912           if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2913             vp9_highbd_build_inter_predictor(
2914                 CONVERT_TO_SHORTPTR(pd->pre[0].buf), pd->pre[0].stride,
2915                 CONVERT_TO_SHORTPTR(pd->dst.buf), pd->dst.stride,
2916                 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2917                 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2918                 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2919                 mi_col * MI_SIZE + 4 * (i & 0x01),
2920                 mi_row * MI_SIZE + 4 * (i >> 1), xd->bd);
2921           } else {
2922 #endif
2923             vp9_build_inter_predictor(
2924                 pd->pre[0].buf, pd->pre[0].stride, pd->dst.buf, pd->dst.stride,
2925                 &xd->mi[0]->bmi[i].as_mv[0].as_mv, &xd->block_refs[0]->sf,
2926                 4 * num_4x4_blocks_wide, 4 * num_4x4_blocks_high, 0,
2927                 vp9_filter_kernels[mi->interp_filter], MV_PRECISION_Q3,
2928                 mi_col * MI_SIZE + 4 * (i & 0x01),
2929                 mi_row * MI_SIZE + 4 * (i >> 1));
2930 
2931 #if CONFIG_VP9_HIGHBITDEPTH
2932           }
2933 #endif
2934 
2935           model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
2936                             &var_y, &sse_y, 0);
2937 
2938           this_rdc.rate += b_rate;
2939           this_rdc.rdcost =
2940               RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
2941           if (this_rdc.rdcost < b_best_rd) {
2942             b_best_rd = this_rdc.rdcost;
2943             bsi[ref_frame][i].as_mode = this_mode;
2944             bsi[ref_frame][i].as_mv[0].as_mv = xd->mi[0]->bmi[i].as_mv[0].as_mv;
2945           }
2946         }  // mode search
2947 
2948         // restore source and prediction buffer pointers.
2949         p->src = orig_src;
2950         pd->pre[0] = orig_pre[0];
2951         pd->dst = orig_dst;
2952         this_rd += b_best_rd;
2953 
2954         xd->mi[0]->bmi[i] = bsi[ref_frame][i];
2955         if (num_4x4_blocks_wide > 1) xd->mi[0]->bmi[i + 1] = xd->mi[0]->bmi[i];
2956         if (num_4x4_blocks_high > 1) xd->mi[0]->bmi[i + 2] = xd->mi[0]->bmi[i];
2957       }
2958     }  // loop through sub8x8 blocks
2959 
2960     if (this_rd < best_rd) {
2961       best_rd = this_rd;
2962       best_ref_frame = ref_frame;
2963     }
2964   }  // reference frames
2965 
2966   mi->tx_size = TX_4X4;
2967   mi->ref_frame[0] = best_ref_frame;
2968   for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2969     for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2970       const int block = idy * 2 + idx;
2971       xd->mi[0]->bmi[block] = bsi[best_ref_frame][block];
2972       if (num_4x4_blocks_wide > 1)
2973         xd->mi[0]->bmi[block + 1] = bsi[best_ref_frame][block];
2974       if (num_4x4_blocks_high > 1)
2975         xd->mi[0]->bmi[block + 2] = bsi[best_ref_frame][block];
2976     }
2977   }
2978   mi->mode = xd->mi[0]->bmi[3].as_mode;
2979   ctx->mic = *(xd->mi[0]);
2980   ctx->mbmi_ext = *x->mbmi_ext;
2981   ctx->skip_txfm[0] = SKIP_TXFM_NONE;
2982   ctx->skip = 0;
2983   // Dummy assignment for speed -5. No effect in speed -6.
2984   rd_cost->rdcost = best_rd;
2985 }
2986