• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2016, Alliance for Open Media. All rights reserved
3  *
4  * This source code is subject to the terms of the BSD 2 Clause License and
5  * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6  * was not distributed with this source code in the LICENSE file, you can
7  * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8  * Media Patent License 1.0 was not distributed with this source code in the
9  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 
11  */
12 
13 #include <assert.h>
14 #include <limits.h>
15 #include <math.h>
16 #include <stdio.h>
17 
18 #include "config/aom_dsp_rtcd.h"
19 #include "config/av1_rtcd.h"
20 
21 #include "aom_dsp/aom_dsp_common.h"
22 #include "aom_dsp/blend.h"
23 #include "aom_mem/aom_mem.h"
24 #include "aom_ports/aom_timer.h"
25 #include "aom_ports/mem.h"
26 #include "aom_ports/system_state.h"
27 
28 #include "av1/encoder/model_rd.h"
29 #include "av1/common/mvref_common.h"
30 #include "av1/common/pred_common.h"
31 #include "av1/common/reconinter.h"
32 #include "av1/common/reconintra.h"
33 
34 #include "av1/encoder/encodemv.h"
35 #include "av1/encoder/rdopt.h"
36 #include "av1/encoder/reconinter_enc.h"
37 
38 extern int g_pick_inter_mode_cnt;
39 typedef struct {
40   uint8_t *data;
41   int stride;
42   int in_use;
43 } PRED_BUFFER;
44 
45 typedef struct {
46   PRED_BUFFER *best_pred;
47   PREDICTION_MODE best_mode;
48   TX_SIZE best_tx_size;
49   TX_SIZE best_intra_tx_size;
50   MV_REFERENCE_FRAME best_ref_frame;
51   MV_REFERENCE_FRAME best_second_ref_frame;
52   uint8_t best_mode_skip_txfm;
53   int_interpfilters best_pred_filter;
54 } BEST_PICKMODE;
55 
56 typedef struct {
57   MV_REFERENCE_FRAME ref_frame;
58   PREDICTION_MODE pred_mode;
59 } REF_MODE;
60 
61 static const int pos_shift_16x16[4][4] = {
62   { 9, 10, 13, 14 }, { 11, 12, 15, 16 }, { 17, 18, 21, 22 }, { 19, 20, 23, 24 }
63 };
64 
65 #define RT_INTER_MODES 9
66 static const REF_MODE ref_mode_set[RT_INTER_MODES] = {
67   { LAST_FRAME, NEARESTMV },   { LAST_FRAME, NEARMV },
68   { LAST_FRAME, NEWMV },       { GOLDEN_FRAME, NEARESTMV },
69   { GOLDEN_FRAME, NEARMV },    { GOLDEN_FRAME, NEWMV },
70   { ALTREF_FRAME, NEARESTMV }, { ALTREF_FRAME, NEARMV },
71   { ALTREF_FRAME, NEWMV }
72 };
73 
74 static const THR_MODES mode_idx[REF_FRAMES][4] = {
75   { THR_DC, THR_V_PRED, THR_H_PRED, THR_SMOOTH },
76   { THR_NEARESTMV, THR_NEARMV, THR_GLOBALMV, THR_NEWMV },
77   { THR_NEARESTL2, THR_NEARL2, THR_GLOBALL2, THR_NEWL2 },
78   { THR_NEARESTL3, THR_NEARL3, THR_GLOBALL3, THR_NEWL3 },
79   { THR_NEARESTG, THR_NEARG, THR_GLOBALMV, THR_NEWG },
80 };
81 
82 static const PREDICTION_MODE intra_mode_list[] = { DC_PRED, V_PRED, H_PRED,
83                                                    SMOOTH_PRED };
84 
mode_offset(const PREDICTION_MODE mode)85 static INLINE int mode_offset(const PREDICTION_MODE mode) {
86   if (mode >= NEARESTMV) {
87     return INTER_OFFSET(mode);
88   } else {
89     switch (mode) {
90       case DC_PRED: return 0;
91       case V_PRED: return 1;
92       case H_PRED: return 2;
93       case SMOOTH_PRED: return 3;
94       default: assert(0); return -1;
95     }
96   }
97 }
98 
99 enum {
100   //  INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << NEWMV),
101   INTER_NEAREST = (1 << NEARESTMV),
102   INTER_NEAREST_NEW = (1 << NEARESTMV) | (1 << NEWMV),
103   INTER_NEAREST_NEAR = (1 << NEARESTMV) | (1 << NEARMV),
104   INTER_NEAR_NEW = (1 << NEARMV) | (1 << NEWMV),
105 };
106 
init_best_pickmode(BEST_PICKMODE * bp)107 static INLINE void init_best_pickmode(BEST_PICKMODE *bp) {
108   bp->best_mode = NEARESTMV;
109   bp->best_ref_frame = LAST_FRAME;
110   bp->best_tx_size = TX_8X8;
111   bp->best_intra_tx_size = TX_8X8;
112   bp->best_pred_filter = av1_broadcast_interp_filter(EIGHTTAP_REGULAR);
113   bp->best_mode_skip_txfm = 0;
114   bp->best_second_ref_frame = NONE_FRAME;
115   bp->best_pred = NULL;
116 }
117 
combined_motion_search(AV1_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)118 static int combined_motion_search(AV1_COMP *cpi, MACROBLOCK *x,
119                                   BLOCK_SIZE bsize, int mi_row, int mi_col,
120                                   int_mv *tmp_mv, int *rate_mv,
121                                   int64_t best_rd_sofar, int use_base_mv) {
122   MACROBLOCKD *xd = &x->e_mbd;
123   const AV1_COMMON *cm = &cpi->common;
124   const int num_planes = av1_num_planes(cm);
125   MB_MODE_INFO *mi = xd->mi[0];
126   struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0, 0, 0, 0 } };
127   int step_param = cpi->mv_search_params.mv_step_param;
128   FULLPEL_MV start_mv;
129   const int ref = mi->ref_frame[0];
130   const MV ref_mv = av1_get_ref_mv(x, mi->ref_mv_idx).as_mv;
131   MV center_mv;
132   int dis;
133   int rv = 0;
134   int cost_list[5];
135   int search_subpel = 1;
136   const YV12_BUFFER_CONFIG *scaled_ref_frame =
137       av1_get_scaled_ref_frame(cpi, ref);
138 
139   if (scaled_ref_frame) {
140     int i;
141     // Swap out the reference frame for a version that's been scaled to
142     // match the resolution of the current frame, allowing the existing
143     // motion search code to be used without additional modifications.
144     for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
145     av1_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL,
146                          num_planes);
147   }
148 
149   start_mv = get_fullmv_from_mv(&ref_mv);
150 
151   if (!use_base_mv)
152     center_mv = ref_mv;
153   else
154     center_mv = tmp_mv->as_mv;
155 
156   const search_site_config *src_search_sites =
157       &cpi->mv_search_params.ss_cfg[SS_CFG_SRC];
158   FULLPEL_MOTION_SEARCH_PARAMS full_ms_params;
159   av1_make_default_fullpel_ms_params(&full_ms_params, cpi, x, bsize, &center_mv,
160                                      src_search_sites);
161 
162   av1_full_pixel_search(start_mv, &full_ms_params, step_param,
163                         cond_cost_list(cpi, cost_list), &tmp_mv->as_fullmv,
164                         NULL);
165 
166   // calculate the bit cost on motion vector
167   MV mvp_full = get_mv_from_fullmv(&tmp_mv->as_fullmv);
168 
169   *rate_mv = av1_mv_bit_cost(&mvp_full, &ref_mv, x->nmv_vec_cost,
170                              x->mv_cost_stack, MV_COST_WEIGHT);
171 
172   // TODO(kyslov) Account for Rate Mode!
173   rv = !(RDCOST(x->rdmult, (*rate_mv), 0) > best_rd_sofar);
174 
175   if (rv && search_subpel) {
176     SUBPEL_MOTION_SEARCH_PARAMS ms_params;
177     av1_make_default_subpel_ms_params(&ms_params, cpi, x, bsize, &ref_mv,
178                                       cost_list);
179     MV subpel_start_mv = get_mv_from_fullmv(&tmp_mv->as_fullmv);
180     cpi->mv_search_params.find_fractional_mv_step(
181         xd, cm, &ms_params, subpel_start_mv, &tmp_mv->as_mv, &dis,
182         &x->pred_sse[ref], NULL);
183 
184     *rate_mv = av1_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmv_vec_cost,
185                                x->mv_cost_stack, MV_COST_WEIGHT);
186   }
187 
188   if (scaled_ref_frame) {
189     int i;
190     for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
191   }
192   return rv;
193 }
194 
search_new_mv(AV1_COMP * cpi,MACROBLOCK * x,int_mv frame_mv[][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,RD_STATS * best_rdc)195 static int search_new_mv(AV1_COMP *cpi, MACROBLOCK *x,
196                          int_mv frame_mv[][REF_FRAMES],
197                          MV_REFERENCE_FRAME ref_frame, int gf_temporal_ref,
198                          BLOCK_SIZE bsize, int mi_row, int mi_col,
199                          int best_pred_sad, int *rate_mv, RD_STATS *best_rdc) {
200   MACROBLOCKD *const xd = &x->e_mbd;
201   MB_MODE_INFO *const mi = xd->mi[0];
202   AV1_COMMON *cm = &cpi->common;
203   if (ref_frame > LAST_FRAME && gf_temporal_ref &&
204       cpi->oxcf.rc_mode == AOM_CBR) {
205     int tmp_sad;
206     int dis;
207     int cost_list[5] = { INT_MAX, INT_MAX, INT_MAX, INT_MAX, INT_MAX };
208 
209     if (bsize < BLOCK_16X16) return -1;
210 
211     tmp_sad = av1_int_pro_motion_estimation(
212         cpi, x, bsize, mi_row, mi_col,
213         &x->mbmi_ext->ref_mv_stack[ref_frame][0].this_mv.as_mv);
214 
215     if (tmp_sad > x->pred_mv_sad[LAST_FRAME]) return -1;
216     if (tmp_sad + (num_pels_log2_lookup[bsize] << 4) > best_pred_sad) return -1;
217 
218     frame_mv[NEWMV][ref_frame].as_int = mi->mv[0].as_int;
219     int_mv best_mv = mi->mv[0];
220     best_mv.as_mv.row >>= 3;
221     best_mv.as_mv.col >>= 3;
222     MV ref_mv = av1_get_ref_mv(x, 0).as_mv;
223 
224     *rate_mv =
225         av1_mv_bit_cost(&frame_mv[NEWMV][ref_frame].as_mv, &ref_mv,
226                         x->nmv_vec_cost, x->mv_cost_stack, MV_COST_WEIGHT);
227     frame_mv[NEWMV][ref_frame].as_mv.row >>= 3;
228     frame_mv[NEWMV][ref_frame].as_mv.col >>= 3;
229 
230     SUBPEL_MOTION_SEARCH_PARAMS ms_params;
231     av1_make_default_subpel_ms_params(&ms_params, cpi, x, bsize, &ref_mv,
232                                       cost_list);
233     MV start_mv = get_mv_from_fullmv(&best_mv.as_fullmv);
234     cpi->mv_search_params.find_fractional_mv_step(
235         xd, cm, &ms_params, start_mv, &best_mv.as_mv, &dis,
236         &x->pred_sse[ref_frame], NULL);
237     frame_mv[NEWMV][ref_frame].as_int = best_mv.as_int;
238   } else if (!combined_motion_search(cpi, x, bsize, mi_row, mi_col,
239                                      &frame_mv[NEWMV][ref_frame], rate_mv,
240                                      best_rdc->rdcost, 0)) {
241     return -1;
242   }
243 
244   return 0;
245 }
246 
find_predictors(AV1_COMP * cpi,MACROBLOCK * x,MV_REFERENCE_FRAME ref_frame,int_mv frame_mv[MB_MODE_COUNT][REF_FRAMES],int * ref_frame_skip_mask,const int flag_list[4],TileDataEnc * tile_data,struct buf_2d yv12_mb[8][MAX_MB_PLANE],BLOCK_SIZE bsize,int force_skip_low_temp_var)247 static INLINE void find_predictors(
248     AV1_COMP *cpi, MACROBLOCK *x, MV_REFERENCE_FRAME ref_frame,
249     int_mv frame_mv[MB_MODE_COUNT][REF_FRAMES], int *ref_frame_skip_mask,
250     const int flag_list[4], TileDataEnc *tile_data,
251     struct buf_2d yv12_mb[8][MAX_MB_PLANE], BLOCK_SIZE bsize,
252     int force_skip_low_temp_var) {
253   AV1_COMMON *const cm = &cpi->common;
254   MACROBLOCKD *const xd = &x->e_mbd;
255   MB_MODE_INFO *const mbmi = xd->mi[0];
256   MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
257   const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_yv12_buf(cm, ref_frame);
258   const int num_planes = av1_num_planes(cm);
259   (void)tile_data;
260 
261   x->pred_mv_sad[ref_frame] = INT_MAX;
262   frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
263   // TODO(kyslov) this needs various further optimizations. to be continued..
264   if ((cpi->ref_frame_flags & flag_list[ref_frame]) && (yv12 != NULL)) {
265     const struct scale_factors *const sf =
266         get_ref_scale_factors_const(cm, ref_frame);
267     av1_setup_pred_block(xd, yv12_mb[ref_frame], yv12, sf, sf, num_planes);
268     av1_find_mv_refs(cm, xd, mbmi, ref_frame, mbmi_ext->ref_mv_count,
269                      xd->ref_mv_stack, xd->weight, NULL, mbmi_ext->global_mvs,
270                      mbmi_ext->mode_context);
271     // TODO(Ravi): Populate mbmi_ext->ref_mv_stack[ref_frame][4] and
272     // mbmi_ext->weight[ref_frame][4] inside av1_find_mv_refs.
273     av1_copy_usable_ref_mv_stack_and_weight(xd, mbmi_ext, ref_frame);
274     av1_find_best_ref_mvs_from_stack(
275         cm->features.allow_high_precision_mv, mbmi_ext, ref_frame,
276         &frame_mv[NEARESTMV][ref_frame], &frame_mv[NEARMV][ref_frame], 0);
277     // Early exit for non-LAST frame if force_skip_low_temp_var is set.
278     if (!av1_is_scaled(sf) && bsize >= BLOCK_8X8 &&
279         !(force_skip_low_temp_var && ref_frame != LAST_FRAME)) {
280       av1_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
281                   bsize);
282     }
283   } else {
284     *ref_frame_skip_mask |= (1 << ref_frame);
285   }
286   av1_count_overlappable_neighbors(cm, xd);
287   mbmi->num_proj_ref = 1;
288 }
289 
estimate_single_ref_frame_costs(const AV1_COMMON * cm,const MACROBLOCKD * xd,const MACROBLOCK * x,int segment_id,unsigned int * ref_costs_single)290 static void estimate_single_ref_frame_costs(const AV1_COMMON *cm,
291                                             const MACROBLOCKD *xd,
292                                             const MACROBLOCK *x, int segment_id,
293                                             unsigned int *ref_costs_single) {
294   int seg_ref_active =
295       segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME);
296   if (seg_ref_active) {
297     memset(ref_costs_single, 0, REF_FRAMES * sizeof(*ref_costs_single));
298   } else {
299     int intra_inter_ctx = av1_get_intra_inter_context(xd);
300     ref_costs_single[INTRA_FRAME] = x->intra_inter_cost[intra_inter_ctx][0];
301     unsigned int base_cost = x->intra_inter_cost[intra_inter_ctx][1];
302 
303     for (int i = LAST_FRAME; i <= ALTREF_FRAME; ++i)
304       ref_costs_single[i] = base_cost;
305 
306     const int ctx_p1 = av1_get_pred_context_single_ref_p1(xd);
307     const int ctx_p2 = av1_get_pred_context_single_ref_p2(xd);
308     const int ctx_p3 = av1_get_pred_context_single_ref_p3(xd);
309     const int ctx_p4 = av1_get_pred_context_single_ref_p4(xd);
310     const int ctx_p5 = av1_get_pred_context_single_ref_p5(xd);
311     const int ctx_p6 = av1_get_pred_context_single_ref_p6(xd);
312 
313     // Determine cost of a single ref frame, where frame types are represented
314     // by a tree:
315     // Level 0: add cost whether this ref is a forward or backward ref
316     ref_costs_single[LAST_FRAME] += x->single_ref_cost[ctx_p1][0][0];
317     ref_costs_single[LAST2_FRAME] += x->single_ref_cost[ctx_p1][0][0];
318     ref_costs_single[LAST3_FRAME] += x->single_ref_cost[ctx_p1][0][0];
319     ref_costs_single[GOLDEN_FRAME] += x->single_ref_cost[ctx_p1][0][0];
320     ref_costs_single[BWDREF_FRAME] += x->single_ref_cost[ctx_p1][0][1];
321     ref_costs_single[ALTREF2_FRAME] += x->single_ref_cost[ctx_p1][0][1];
322     ref_costs_single[ALTREF_FRAME] += x->single_ref_cost[ctx_p1][0][1];
323 
324     // Level 1: if this ref is forward ref,
325     // add cost whether it is last/last2 or last3/golden
326     ref_costs_single[LAST_FRAME] += x->single_ref_cost[ctx_p3][2][0];
327     ref_costs_single[LAST2_FRAME] += x->single_ref_cost[ctx_p3][2][0];
328     ref_costs_single[LAST3_FRAME] += x->single_ref_cost[ctx_p3][2][1];
329     ref_costs_single[GOLDEN_FRAME] += x->single_ref_cost[ctx_p3][2][1];
330 
331     // Level 1: if this ref is backward ref
332     // then add cost whether this ref is altref or backward ref
333     ref_costs_single[BWDREF_FRAME] += x->single_ref_cost[ctx_p2][1][0];
334     ref_costs_single[ALTREF2_FRAME] += x->single_ref_cost[ctx_p2][1][0];
335     ref_costs_single[ALTREF_FRAME] += x->single_ref_cost[ctx_p2][1][1];
336 
337     // Level 2: further add cost whether this ref is last or last2
338     ref_costs_single[LAST_FRAME] += x->single_ref_cost[ctx_p4][3][0];
339     ref_costs_single[LAST2_FRAME] += x->single_ref_cost[ctx_p4][3][1];
340 
341     // Level 2: last3 or golden
342     ref_costs_single[LAST3_FRAME] += x->single_ref_cost[ctx_p5][4][0];
343     ref_costs_single[GOLDEN_FRAME] += x->single_ref_cost[ctx_p5][4][1];
344 
345     // Level 2: bwdref or altref2
346     ref_costs_single[BWDREF_FRAME] += x->single_ref_cost[ctx_p6][5][0];
347     ref_costs_single[ALTREF2_FRAME] += x->single_ref_cost[ctx_p6][5][1];
348   }
349 }
350 
estimate_comp_ref_frame_costs(const AV1_COMMON * cm,const MACROBLOCKD * xd,const MACROBLOCK * x,int segment_id,unsigned int (* ref_costs_comp)[REF_FRAMES])351 static void estimate_comp_ref_frame_costs(
352     const AV1_COMMON *cm, const MACROBLOCKD *xd, const MACROBLOCK *x,
353     int segment_id, unsigned int (*ref_costs_comp)[REF_FRAMES]) {
354   if (segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
355     for (int ref_frame = 0; ref_frame < REF_FRAMES; ++ref_frame)
356       memset(ref_costs_comp[ref_frame], 0,
357              REF_FRAMES * sizeof((*ref_costs_comp)[0]));
358   } else {
359     int intra_inter_ctx = av1_get_intra_inter_context(xd);
360     unsigned int base_cost = x->intra_inter_cost[intra_inter_ctx][1];
361 
362     if (cm->current_frame.reference_mode != SINGLE_REFERENCE) {
363       // Similar to single ref, determine cost of compound ref frames.
364       // cost_compound_refs = cost_first_ref + cost_second_ref
365       const int bwdref_comp_ctx_p = av1_get_pred_context_comp_bwdref_p(xd);
366       const int bwdref_comp_ctx_p1 = av1_get_pred_context_comp_bwdref_p1(xd);
367       const int ref_comp_ctx_p = av1_get_pred_context_comp_ref_p(xd);
368       const int ref_comp_ctx_p1 = av1_get_pred_context_comp_ref_p1(xd);
369       const int ref_comp_ctx_p2 = av1_get_pred_context_comp_ref_p2(xd);
370 
371       const int comp_ref_type_ctx = av1_get_comp_reference_type_context(xd);
372       unsigned int ref_bicomp_costs[REF_FRAMES] = { 0 };
373 
374       ref_bicomp_costs[LAST_FRAME] = ref_bicomp_costs[LAST2_FRAME] =
375           ref_bicomp_costs[LAST3_FRAME] = ref_bicomp_costs[GOLDEN_FRAME] =
376               base_cost + x->comp_ref_type_cost[comp_ref_type_ctx][1];
377       ref_bicomp_costs[BWDREF_FRAME] = ref_bicomp_costs[ALTREF2_FRAME] = 0;
378       ref_bicomp_costs[ALTREF_FRAME] = 0;
379 
380       // cost of first ref frame
381       ref_bicomp_costs[LAST_FRAME] += x->comp_ref_cost[ref_comp_ctx_p][0][0];
382       ref_bicomp_costs[LAST2_FRAME] += x->comp_ref_cost[ref_comp_ctx_p][0][0];
383       ref_bicomp_costs[LAST3_FRAME] += x->comp_ref_cost[ref_comp_ctx_p][0][1];
384       ref_bicomp_costs[GOLDEN_FRAME] += x->comp_ref_cost[ref_comp_ctx_p][0][1];
385 
386       ref_bicomp_costs[LAST_FRAME] += x->comp_ref_cost[ref_comp_ctx_p1][1][0];
387       ref_bicomp_costs[LAST2_FRAME] += x->comp_ref_cost[ref_comp_ctx_p1][1][1];
388 
389       ref_bicomp_costs[LAST3_FRAME] += x->comp_ref_cost[ref_comp_ctx_p2][2][0];
390       ref_bicomp_costs[GOLDEN_FRAME] += x->comp_ref_cost[ref_comp_ctx_p2][2][1];
391 
392       // cost of second ref frame
393       ref_bicomp_costs[BWDREF_FRAME] +=
394           x->comp_bwdref_cost[bwdref_comp_ctx_p][0][0];
395       ref_bicomp_costs[ALTREF2_FRAME] +=
396           x->comp_bwdref_cost[bwdref_comp_ctx_p][0][0];
397       ref_bicomp_costs[ALTREF_FRAME] +=
398           x->comp_bwdref_cost[bwdref_comp_ctx_p][0][1];
399 
400       ref_bicomp_costs[BWDREF_FRAME] +=
401           x->comp_bwdref_cost[bwdref_comp_ctx_p1][1][0];
402       ref_bicomp_costs[ALTREF2_FRAME] +=
403           x->comp_bwdref_cost[bwdref_comp_ctx_p1][1][1];
404 
405       // cost: if one ref frame is forward ref, the other ref is backward ref
406       for (int ref0 = LAST_FRAME; ref0 <= GOLDEN_FRAME; ++ref0) {
407         for (int ref1 = BWDREF_FRAME; ref1 <= ALTREF_FRAME; ++ref1) {
408           ref_costs_comp[ref0][ref1] =
409               ref_bicomp_costs[ref0] + ref_bicomp_costs[ref1];
410         }
411       }
412 
413       // cost: if both ref frames are the same side.
414       const int uni_comp_ref_ctx_p = av1_get_pred_context_uni_comp_ref_p(xd);
415       const int uni_comp_ref_ctx_p1 = av1_get_pred_context_uni_comp_ref_p1(xd);
416       const int uni_comp_ref_ctx_p2 = av1_get_pred_context_uni_comp_ref_p2(xd);
417       ref_costs_comp[LAST_FRAME][LAST2_FRAME] =
418           base_cost + x->comp_ref_type_cost[comp_ref_type_ctx][0] +
419           x->uni_comp_ref_cost[uni_comp_ref_ctx_p][0][0] +
420           x->uni_comp_ref_cost[uni_comp_ref_ctx_p1][1][0];
421       ref_costs_comp[LAST_FRAME][LAST3_FRAME] =
422           base_cost + x->comp_ref_type_cost[comp_ref_type_ctx][0] +
423           x->uni_comp_ref_cost[uni_comp_ref_ctx_p][0][0] +
424           x->uni_comp_ref_cost[uni_comp_ref_ctx_p1][1][1] +
425           x->uni_comp_ref_cost[uni_comp_ref_ctx_p2][2][0];
426       ref_costs_comp[LAST_FRAME][GOLDEN_FRAME] =
427           base_cost + x->comp_ref_type_cost[comp_ref_type_ctx][0] +
428           x->uni_comp_ref_cost[uni_comp_ref_ctx_p][0][0] +
429           x->uni_comp_ref_cost[uni_comp_ref_ctx_p1][1][1] +
430           x->uni_comp_ref_cost[uni_comp_ref_ctx_p2][2][1];
431       ref_costs_comp[BWDREF_FRAME][ALTREF_FRAME] =
432           base_cost + x->comp_ref_type_cost[comp_ref_type_ctx][0] +
433           x->uni_comp_ref_cost[uni_comp_ref_ctx_p][0][1];
434     } else {
435       for (int ref0 = LAST_FRAME; ref0 <= GOLDEN_FRAME; ++ref0) {
436         for (int ref1 = BWDREF_FRAME; ref1 <= ALTREF_FRAME; ++ref1)
437           ref_costs_comp[ref0][ref1] = 512;
438       }
439       ref_costs_comp[LAST_FRAME][LAST2_FRAME] = 512;
440       ref_costs_comp[LAST_FRAME][LAST3_FRAME] = 512;
441       ref_costs_comp[LAST_FRAME][GOLDEN_FRAME] = 512;
442       ref_costs_comp[BWDREF_FRAME][ALTREF_FRAME] = 512;
443     }
444   }
445 }
446 
calculate_tx_size(const AV1_COMP * const cpi,BLOCK_SIZE bsize,MACROBLOCK * const x,unsigned int var,unsigned int sse)447 static TX_SIZE calculate_tx_size(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
448                                  MACROBLOCK *const x, unsigned int var,
449                                  unsigned int sse) {
450   MACROBLOCKD *const xd = &x->e_mbd;
451   TX_SIZE tx_size;
452   if (x->tx_mode_search_type == TX_MODE_SELECT) {
453     if (sse > (var << 2))
454       tx_size = AOMMIN(max_txsize_lookup[bsize],
455                        tx_mode_to_biggest_tx_size[x->tx_mode_search_type]);
456     else
457       tx_size = TX_8X8;
458 
459     if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ &&
460         cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id))
461       tx_size = TX_8X8;
462     else if (tx_size > TX_16X16)
463       tx_size = TX_16X16;
464   } else {
465     tx_size = AOMMIN(max_txsize_lookup[bsize],
466                      tx_mode_to_biggest_tx_size[x->tx_mode_search_type]);
467   }
468 
469   if (x->tx_mode_search_type != ONLY_4X4 && bsize > BLOCK_32X32)
470     tx_size = TX_16X16;
471 
472   return AOMMIN(tx_size, TX_16X16);
473 }
474 
475 static const uint8_t b_width_log2_lookup[BLOCK_SIZES] = { 0, 0, 1, 1, 1, 2,
476                                                           2, 2, 3, 3, 3, 4,
477                                                           4, 4, 5, 5 };
478 static const uint8_t b_height_log2_lookup[BLOCK_SIZES] = { 0, 1, 0, 1, 2, 1,
479                                                            2, 3, 2, 3, 4, 3,
480                                                            4, 5, 4, 5 };
481 
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,uint32_t * sse8x8,int * sum8x8,uint32_t * var8x8)482 static void block_variance(const uint8_t *src, int src_stride,
483                            const uint8_t *ref, int ref_stride, int w, int h,
484                            unsigned int *sse, int *sum, int block_size,
485                            uint32_t *sse8x8, int *sum8x8, uint32_t *var8x8) {
486   int i, j, k = 0;
487 
488   *sse = 0;
489   *sum = 0;
490 
491   for (i = 0; i < h; i += block_size) {
492     for (j = 0; j < w; j += block_size) {
493       aom_get8x8var(src + src_stride * i + j, src_stride,
494                     ref + ref_stride * i + j, ref_stride, &sse8x8[k],
495                     &sum8x8[k]);
496       *sse += sse8x8[k];
497       *sum += sum8x8[k];
498       var8x8[k] = sse8x8[k] - (uint32_t)(((int64_t)sum8x8[k] * sum8x8[k]) >> 6);
499       k++;
500     }
501   }
502 }
503 
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)504 static void calculate_variance(int bw, int bh, TX_SIZE tx_size,
505                                unsigned int *sse_i, int *sum_i,
506                                unsigned int *var_o, unsigned int *sse_o,
507                                int *sum_o) {
508   const BLOCK_SIZE unit_size = txsize_to_bsize[tx_size];
509   const int nw = 1 << (bw - b_width_log2_lookup[unit_size]);
510   const int nh = 1 << (bh - b_height_log2_lookup[unit_size]);
511   int i, j, k = 0;
512 
513   for (i = 0; i < nh; i += 2) {
514     for (j = 0; j < nw; j += 2) {
515       sse_o[k] = sse_i[i * nw + j] + sse_i[i * nw + j + 1] +
516                  sse_i[(i + 1) * nw + j] + sse_i[(i + 1) * nw + j + 1];
517       sum_o[k] = sum_i[i * nw + j] + sum_i[i * nw + j + 1] +
518                  sum_i[(i + 1) * nw + j] + sum_i[(i + 1) * nw + j + 1];
519       var_o[k] = sse_o[k] - (uint32_t)(((int64_t)sum_o[k] * sum_o[k]) >>
520                                        (b_width_log2_lookup[unit_size] +
521                                         b_height_log2_lookup[unit_size] + 6));
522       k++;
523     }
524   }
525 }
526 
527 // 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)528 static int ac_thr_factor(const int speed, const int width, const int height,
529                          const int norm_sum) {
530   if (speed >= 8 && norm_sum < 5) {
531     if (width <= 640 && height <= 480)
532       return 4;
533     else
534       return 2;
535   }
536   return 1;
537 }
538 
model_skip_for_sb_y_large(AV1_COMP * cpi,BLOCK_SIZE bsize,int mi_row,int mi_col,MACROBLOCK * x,MACROBLOCKD * xd,int * out_rate,int64_t * out_dist,unsigned int * var_y,unsigned int * sse_y,int * early_term,int calculate_rd)539 static void model_skip_for_sb_y_large(AV1_COMP *cpi, BLOCK_SIZE bsize,
540                                       int mi_row, int mi_col, MACROBLOCK *x,
541                                       MACROBLOCKD *xd, int *out_rate,
542                                       int64_t *out_dist, unsigned int *var_y,
543                                       unsigned int *sse_y, int *early_term,
544                                       int calculate_rd) {
545   // Note our transform coeffs are 8 times an orthogonal transform.
546   // Hence quantizer step is also 8 times. To get effective quantizer
547   // we need to divide by 8 before sending to modeling function.
548   unsigned int sse;
549   struct macroblock_plane *const p = &x->plane[0];
550   struct macroblockd_plane *const pd = &xd->plane[0];
551   const uint32_t dc_quant = p->dequant_QTX[0];
552   const uint32_t ac_quant = p->dequant_QTX[1];
553   const int64_t dc_thr = dc_quant * dc_quant >> 6;
554   int64_t ac_thr = ac_quant * ac_quant >> 6;
555   unsigned int var;
556   int sum;
557 
558   const int bw = b_width_log2_lookup[bsize];
559   const int bh = b_height_log2_lookup[bsize];
560   const int num8x8 = 1 << (bw + bh - 2);
561   unsigned int sse8x8[256] = { 0 };
562   int sum8x8[256] = { 0 };
563   unsigned int var8x8[256] = { 0 };
564   TX_SIZE tx_size;
565   int k;
566   // Calculate variance for whole partition, and also save 8x8 blocks' variance
567   // to be used in following transform skipping test.
568   block_variance(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
569                  4 << bw, 4 << bh, &sse, &sum, 8, sse8x8, sum8x8, var8x8);
570   var = sse - (unsigned int)(((int64_t)sum * sum) >> (bw + bh + 4));
571 
572   *var_y = var;
573   *sse_y = sse;
574 
575   ac_thr *= ac_thr_factor(cpi->oxcf.speed, cpi->common.width,
576                           cpi->common.height, abs(sum) >> (bw + bh));
577 
578   tx_size = calculate_tx_size(cpi, bsize, x, var, sse);
579   // The code below for setting skip flag assumes tranform size of at least 8x8,
580   // so force this lower limit on transform.
581   if (tx_size < TX_8X8) tx_size = TX_8X8;
582   xd->mi[0]->tx_size = tx_size;
583 
584   // Evaluate if the partition block is a skippable block in Y plane.
585   {
586     unsigned int sse16x16[64] = { 0 };
587     int sum16x16[64] = { 0 };
588     unsigned int var16x16[64] = { 0 };
589     const int num16x16 = num8x8 >> 2;
590 
591     unsigned int sse32x32[16] = { 0 };
592     int sum32x32[16] = { 0 };
593     unsigned int var32x32[16] = { 0 };
594     const int num32x32 = num8x8 >> 4;
595 
596     int ac_test = 1;
597     int dc_test = 1;
598     const int num = (tx_size == TX_8X8)
599                         ? num8x8
600                         : ((tx_size == TX_16X16) ? num16x16 : num32x32);
601     const unsigned int *sse_tx =
602         (tx_size == TX_8X8) ? sse8x8
603                             : ((tx_size == TX_16X16) ? sse16x16 : sse32x32);
604     const unsigned int *var_tx =
605         (tx_size == TX_8X8) ? var8x8
606                             : ((tx_size == TX_16X16) ? var16x16 : var32x32);
607 
608     // Calculate variance if tx_size > TX_8X8
609     if (tx_size >= TX_16X16)
610       calculate_variance(bw, bh, TX_8X8, sse8x8, sum8x8, var16x16, sse16x16,
611                          sum16x16);
612     if (tx_size == TX_32X32)
613       calculate_variance(bw, bh, TX_16X16, sse16x16, sum16x16, var32x32,
614                          sse32x32, sum32x32);
615 
616     // Skipping test
617     *early_term = 0;
618     for (k = 0; k < num; k++)
619       // Check if all ac coefficients can be quantized to zero.
620       if (!(var_tx[k] < ac_thr || var == 0)) {
621         ac_test = 0;
622         break;
623       }
624 
625     for (k = 0; k < num; k++)
626       // Check if dc coefficient can be quantized to zero.
627       if (!(sse_tx[k] - var_tx[k] < dc_thr || sse == var)) {
628         dc_test = 0;
629         break;
630       }
631 
632     if (ac_test && dc_test) {
633       int skip_uv[2] = { 0 };
634       unsigned int var_uv[2];
635       unsigned int sse_uv[2];
636       AV1_COMMON *const cm = &cpi->common;
637       // Transform skipping test in UV planes.
638       for (int i = 1; i <= 2; i++) {
639         int j = i - 1;
640         skip_uv[j] = 1;
641         if (x->color_sensitivity[j]) {
642           skip_uv[j] = 0;
643           struct macroblock_plane *const puv = &x->plane[i];
644           struct macroblockd_plane *const puvd = &xd->plane[i];
645           const BLOCK_SIZE uv_bsize = get_plane_block_size(
646               bsize, puvd->subsampling_x, puvd->subsampling_y);
647           // Adjust these thresholds for UV.
648           const int64_t uv_dc_thr =
649               (puv->dequant_QTX[0] * puv->dequant_QTX[0]) >> 3;
650           const int64_t uv_ac_thr =
651               (puv->dequant_QTX[1] * puv->dequant_QTX[1]) >> 3;
652           av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize, i,
653                                         i);
654           var_uv[j] = cpi->fn_ptr[uv_bsize].vf(puv->src.buf, puv->src.stride,
655                                                puvd->dst.buf, puvd->dst.stride,
656                                                &sse_uv[j]);
657           if ((var_uv[j] < uv_ac_thr || var_uv[j] == 0) &&
658               (sse_uv[j] - var_uv[j] < uv_dc_thr || sse_uv[j] == var_uv[j]))
659             skip_uv[j] = 1;
660           else
661             break;
662         }
663       }
664       if (skip_uv[0] & skip_uv[1]) {
665         *early_term = 1;
666       }
667     }
668   }
669   if (calculate_rd && out_dist != NULL && out_rate != NULL) {
670     if (!*early_term) {
671       const int bwide = block_size_wide[bsize];
672       const int bhigh = block_size_high[bsize];
673 
674       model_rd_with_curvfit(cpi, x, bsize, AOM_PLANE_Y, sse, bwide * bhigh,
675                             out_rate, out_dist);
676     }
677 
678     if (*early_term) {
679       *out_rate = 0;
680       *out_dist = sse << 4;
681     }
682   }
683 }
684 
model_rd_for_sb_y(const AV1_COMP * const cpi,BLOCK_SIZE bsize,MACROBLOCK * x,MACROBLOCKD * xd,int * out_rate_sum,int64_t * out_dist_sum,int * skip_txfm_sb,int64_t * skip_sse_sb,unsigned int * var_y,unsigned int * sse_y,int calculate_rd)685 static void model_rd_for_sb_y(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
686                               MACROBLOCK *x, MACROBLOCKD *xd, int *out_rate_sum,
687                               int64_t *out_dist_sum, int *skip_txfm_sb,
688                               int64_t *skip_sse_sb, unsigned int *var_y,
689                               unsigned int *sse_y, int calculate_rd) {
690   // Note our transform coeffs are 8 times an orthogonal transform.
691   // Hence quantizer step is also 8 times. To get effective quantizer
692   // we need to divide by 8 before sending to modeling function.
693   const int ref = xd->mi[0]->ref_frame[0];
694 
695   assert(bsize < BLOCK_SIZES_ALL);
696 
697   struct macroblock_plane *const p = &x->plane[0];
698   struct macroblockd_plane *const pd = &xd->plane[0];
699   unsigned int sse;
700   int rate;
701   int64_t dist;
702 
703   unsigned int var = cpi->fn_ptr[bsize].vf(p->src.buf, p->src.stride,
704                                            pd->dst.buf, pd->dst.stride, &sse);
705   xd->mi[0]->tx_size = calculate_tx_size(cpi, bsize, x, var, sse);
706 
707   if (calculate_rd) {
708     const int bwide = block_size_wide[bsize];
709     const int bhigh = block_size_high[bsize];
710     model_rd_with_curvfit(cpi, x, bsize, AOM_PLANE_Y, sse, bwide * bhigh, &rate,
711                           &dist);
712   } else {
713     rate = INT_MAX;  // this will be overwritten later with block_yrd
714     dist = INT_MAX;
715   }
716   *var_y = var;
717   *sse_y = sse;
718   x->pred_sse[ref] = (unsigned int)AOMMIN(sse, UINT_MAX);
719 
720   assert(rate >= 0);
721 
722   if (skip_txfm_sb) *skip_txfm_sb = rate == 0;
723   if (skip_sse_sb) *skip_sse_sb = sse << 4;
724   rate = AOMMIN(rate, INT_MAX);
725   *out_rate_sum = (int)rate;
726   *out_dist_sum = dist;
727 }
728 
block_yrd(AV1_COMP * cpi,MACROBLOCK * x,int mi_row,int mi_col,RD_STATS * this_rdc,int * skippable,int64_t * sse,BLOCK_SIZE bsize,TX_SIZE tx_size)729 static void block_yrd(AV1_COMP *cpi, MACROBLOCK *x, int mi_row, int mi_col,
730                       RD_STATS *this_rdc, int *skippable, int64_t *sse,
731                       BLOCK_SIZE bsize, TX_SIZE tx_size) {
732   MACROBLOCKD *xd = &x->e_mbd;
733   const struct macroblockd_plane *pd = &xd->plane[0];
734   struct macroblock_plane *const p = &x->plane[0];
735   const int num_4x4_w = mi_size_wide[bsize];
736   const int num_4x4_h = mi_size_high[bsize];
737   const int step = 1 << (tx_size << 1);
738   const int block_step = (1 << tx_size);
739   int block = 0;
740   const int max_blocks_wide =
741       num_4x4_w + (xd->mb_to_right_edge >= 0 ? 0 : xd->mb_to_right_edge >> 5);
742   const int max_blocks_high =
743       num_4x4_h + (xd->mb_to_bottom_edge >= 0 ? 0 : xd->mb_to_bottom_edge >> 5);
744   int eob_cost = 0;
745   const int bw = 4 * num_4x4_w;
746   const int bh = 4 * num_4x4_h;
747 
748   (void)mi_row;
749   (void)mi_col;
750   (void)cpi;
751 
752 #if CONFIG_AV1_HIGHBITDEPTH
753   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
754     aom_highbd_subtract_block(bh, bw, p->src_diff, bw, p->src.buf,
755                               p->src.stride, pd->dst.buf, pd->dst.stride,
756                               x->e_mbd.bd);
757   } else {
758     aom_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
759                        pd->dst.buf, pd->dst.stride);
760   }
761 #else
762   aom_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
763                      pd->dst.buf, pd->dst.stride);
764 #endif
765 
766   *skippable = 1;
767   // Keep track of the row and column of the blocks we use so that we know
768   // if we are in the unrestricted motion border.
769   for (int r = 0; r < max_blocks_high; r += block_step) {
770     for (int c = 0; c < num_4x4_w; c += block_step) {
771       if (c < max_blocks_wide) {
772         const SCAN_ORDER *const scan_order = &av1_default_scan_orders[tx_size];
773         const int block_offset = BLOCK_OFFSET(block);
774 #if CONFIG_AV1_HIGHBITDEPTH
775         tran_low_t *const coeff = p->coeff + block_offset;
776         tran_low_t *const qcoeff = p->qcoeff + block_offset;
777         tran_low_t *const dqcoeff = pd->dqcoeff + block_offset;
778 #else
779         int16_t *const low_coeff = (int16_t *)p->coeff + block_offset;
780         int16_t *const low_qcoeff = (int16_t *)p->qcoeff + block_offset;
781         int16_t *const low_dqcoeff = (int16_t *)pd->dqcoeff + block_offset;
782 #endif
783         uint16_t *const eob = &p->eobs[block];
784         const int diff_stride = bw;
785         const int16_t *src_diff;
786         src_diff = &p->src_diff[(r * diff_stride + c) << 2];
787 
788         switch (tx_size) {
789           case TX_64X64:
790             assert(0);  // Not implemented
791             break;
792           case TX_32X32:
793             assert(0);  // Not used
794             break;
795 #if CONFIG_AV1_HIGHBITDEPTH
796           case TX_16X16:
797             aom_hadamard_16x16(src_diff, diff_stride, coeff);
798             av1_quantize_fp(coeff, 16 * 16, p->zbin_QTX, p->round_fp_QTX,
799                             p->quant_fp_QTX, p->quant_shift_QTX, qcoeff,
800                             dqcoeff, p->dequant_QTX, eob, scan_order->scan,
801                             scan_order->iscan);
802             break;
803           case TX_8X8:
804             aom_hadamard_8x8(src_diff, diff_stride, coeff);
805             av1_quantize_fp(coeff, 8 * 8, p->zbin_QTX, p->round_fp_QTX,
806                             p->quant_fp_QTX, p->quant_shift_QTX, qcoeff,
807                             dqcoeff, p->dequant_QTX, eob, scan_order->scan,
808                             scan_order->iscan);
809             break;
810 #else
811           case TX_16X16:
812             aom_hadamard_lp_16x16(src_diff, diff_stride, low_coeff);
813             av1_quantize_lp(low_coeff, 16 * 16, p->round_fp_QTX,
814                             p->quant_fp_QTX, low_qcoeff, low_dqcoeff,
815                             p->dequant_QTX, eob, scan_order->scan);
816             break;
817           case TX_8X8:
818             aom_hadamard_lp_8x8(src_diff, diff_stride, low_coeff);
819             av1_quantize_lp(low_coeff, 8 * 8, p->round_fp_QTX, p->quant_fp_QTX,
820                             low_qcoeff, low_dqcoeff, p->dequant_QTX, eob,
821                             scan_order->scan);
822             break;
823           default:
824             assert(tx_size == TX_4X4);
825             x->fwd_txfm4x4(src_diff, low_coeff, diff_stride);
826             av1_quantize_lp(low_coeff, 4 * 4, p->round_fp_QTX, p->quant_fp_QTX,
827                             low_qcoeff, low_dqcoeff, p->dequant_QTX, eob,
828                             scan_order->scan);
829             break;
830 #endif
831         }
832         *skippable &= (*eob == 0);
833         eob_cost += 1;
834       }
835       block += step;
836     }
837   }
838   this_rdc->skip = *skippable;
839   this_rdc->rate = 0;
840   if (*sse < INT64_MAX) {
841     *sse = (*sse << 6) >> 2;
842     if (*skippable) {
843       this_rdc->dist = *sse;
844       return;
845     }
846   }
847 
848   block = 0;
849   this_rdc->dist = 0;
850   for (int r = 0; r < max_blocks_high; r += block_step) {
851     for (int c = 0; c < num_4x4_w; c += block_step) {
852       if (c < max_blocks_wide) {
853         const int block_offset = BLOCK_OFFSET(block);
854         uint16_t *const eob = &p->eobs[block];
855 #if CONFIG_AV1_HIGHBITDEPTH
856         int64_t dummy;
857         tran_low_t *const coeff = p->coeff + block_offset;
858         tran_low_t *const qcoeff = p->qcoeff + block_offset;
859         tran_low_t *const dqcoeff = pd->dqcoeff + block_offset;
860 
861         if (*eob == 1)
862           this_rdc->rate += (int)abs(qcoeff[0]);
863         else if (*eob > 1)
864           this_rdc->rate += aom_satd(qcoeff, step << 4);
865 
866         this_rdc->dist +=
867             av1_block_error(coeff, dqcoeff, step << 4, &dummy) >> 2;
868 #else
869         int16_t *const low_coeff = (int16_t *)p->coeff + block_offset;
870         int16_t *const low_qcoeff = (int16_t *)p->qcoeff + block_offset;
871         int16_t *const low_dqcoeff = (int16_t *)pd->dqcoeff + block_offset;
872 
873         if (*eob == 1)
874           this_rdc->rate += (int)abs(low_qcoeff[0]);
875         else if (*eob > 1)
876           this_rdc->rate += aom_satd_lp(low_qcoeff, step << 4);
877 
878         this_rdc->dist +=
879             av1_block_error_lp(low_coeff, low_dqcoeff, step << 4) >> 2;
880 #endif
881       }
882       block += step;
883     }
884   }
885 
886   // If skippable is set, rate gets clobbered later.
887   this_rdc->rate <<= (2 + AV1_PROB_COST_SHIFT);
888   this_rdc->rate += (eob_cost << AV1_PROB_COST_SHIFT);
889 }
890 
init_mbmi(MB_MODE_INFO * mbmi,PREDICTION_MODE pred_mode,MV_REFERENCE_FRAME ref_frame0,MV_REFERENCE_FRAME ref_frame1,const AV1_COMMON * cm)891 static INLINE void init_mbmi(MB_MODE_INFO *mbmi, PREDICTION_MODE pred_mode,
892                              MV_REFERENCE_FRAME ref_frame0,
893                              MV_REFERENCE_FRAME ref_frame1,
894                              const AV1_COMMON *cm) {
895   PALETTE_MODE_INFO *const pmi = &mbmi->palette_mode_info;
896   mbmi->ref_mv_idx = 0;
897   mbmi->mode = pred_mode;
898   mbmi->uv_mode = UV_DC_PRED;
899   mbmi->ref_frame[0] = ref_frame0;
900   mbmi->ref_frame[1] = ref_frame1;
901   pmi->palette_size[0] = 0;
902   pmi->palette_size[1] = 0;
903   mbmi->filter_intra_mode_info.use_filter_intra = 0;
904   mbmi->mv[0].as_int = mbmi->mv[1].as_int = 0;
905   mbmi->motion_mode = SIMPLE_TRANSLATION;
906   mbmi->num_proj_ref = 1;
907   mbmi->interintra_mode = 0;
908   set_default_interp_filters(mbmi, cm->features.interp_filter);
909 }
910 
911 #if CONFIG_INTERNAL_STATS
store_coding_context(MACROBLOCK * x,PICK_MODE_CONTEXT * ctx,int mode_index)912 static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx,
913                                  int mode_index) {
914 #else
915 static void store_coding_context(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
916 #endif  // CONFIG_INTERNAL_STATS
917   MACROBLOCKD *const xd = &x->e_mbd;
918 
919   // Take a snapshot of the coding context so it can be
920   // restored if we decide to encode this way
921   ctx->rd_stats.skip = x->force_skip;
922   memset(ctx->blk_skip, 0, sizeof(ctx->blk_skip[0]) * ctx->num_4x4_blk);
923   memset(ctx->tx_type_map, DCT_DCT,
924          sizeof(ctx->tx_type_map[0]) * ctx->num_4x4_blk);
925   ctx->skippable = x->force_skip;
926 #if CONFIG_INTERNAL_STATS
927   ctx->best_mode_index = mode_index;
928 #endif  // CONFIG_INTERNAL_STATS
929   ctx->mic = *xd->mi[0];
930   ctx->skippable = x->force_skip;
931   av1_copy_mbmi_ext_to_mbmi_ext_frame(&ctx->mbmi_ext_best, x->mbmi_ext,
932                                       av1_ref_frame_type(xd->mi[0]->ref_frame));
933   ctx->comp_pred_diff = 0;
934   ctx->hybrid_pred_diff = 0;
935   ctx->single_pred_diff = 0;
936 }
937 
938 static int get_pred_buffer(PRED_BUFFER *p, int len) {
939   for (int i = 0; i < len; i++) {
940     if (!p[i].in_use) {
941       p[i].in_use = 1;
942       return i;
943     }
944   }
945   return -1;
946 }
947 
948 static void free_pred_buffer(PRED_BUFFER *p) {
949   if (p != NULL) p->in_use = 0;
950 }
951 
952 static int cost_mv_ref(const MACROBLOCK *const x, PREDICTION_MODE mode,
953                        int16_t mode_context) {
954   if (is_inter_compound_mode(mode)) {
955     return x
956         ->inter_compound_mode_cost[mode_context][INTER_COMPOUND_OFFSET(mode)];
957   }
958 
959   int mode_cost = 0;
960   int16_t mode_ctx = mode_context & NEWMV_CTX_MASK;
961 
962   assert(is_inter_mode(mode));
963 
964   if (mode == NEWMV) {
965     mode_cost = x->newmv_mode_cost[mode_ctx][0];
966     return mode_cost;
967   } else {
968     mode_cost = x->newmv_mode_cost[mode_ctx][1];
969     mode_ctx = (mode_context >> GLOBALMV_OFFSET) & GLOBALMV_CTX_MASK;
970 
971     if (mode == GLOBALMV) {
972       mode_cost += x->zeromv_mode_cost[mode_ctx][0];
973       return mode_cost;
974     } else {
975       mode_cost += x->zeromv_mode_cost[mode_ctx][1];
976       mode_ctx = (mode_context >> REFMV_OFFSET) & REFMV_CTX_MASK;
977       mode_cost += x->refmv_mode_cost[mode_ctx][mode != NEARESTMV];
978       return mode_cost;
979     }
980   }
981 }
982 
983 static void newmv_diff_bias(MACROBLOCKD *xd, PREDICTION_MODE this_mode,
984                             RD_STATS *this_rdc, BLOCK_SIZE bsize, int mv_row,
985                             int mv_col, int speed, uint32_t spatial_variance) {
986   // Bias against MVs associated with NEWMV mode that are very different from
987   // top/left neighbors.
988   if (this_mode == NEWMV) {
989     int al_mv_average_row;
990     int al_mv_average_col;
991     int left_row, left_col;
992     int row_diff, col_diff;
993     int above_mv_valid = 0;
994     int left_mv_valid = 0;
995     int above_row = 0;
996     int above_col = 0;
997 
998     if (xd->above_mbmi) {
999       above_mv_valid = xd->above_mbmi->mv[0].as_int != INVALID_MV;
1000       above_row = xd->above_mbmi->mv[0].as_mv.row;
1001       above_col = xd->above_mbmi->mv[0].as_mv.col;
1002     }
1003     if (xd->left_mbmi) {
1004       left_mv_valid = xd->left_mbmi->mv[0].as_int != INVALID_MV;
1005       left_row = xd->left_mbmi->mv[0].as_mv.row;
1006       left_col = xd->left_mbmi->mv[0].as_mv.col;
1007     }
1008     if (above_mv_valid && left_mv_valid) {
1009       al_mv_average_row = (above_row + left_row + 1) >> 1;
1010       al_mv_average_col = (above_col + left_col + 1) >> 1;
1011     } else if (above_mv_valid) {
1012       al_mv_average_row = above_row;
1013       al_mv_average_col = above_col;
1014     } else if (left_mv_valid) {
1015       al_mv_average_row = left_row;
1016       al_mv_average_col = left_col;
1017     } else {
1018       al_mv_average_row = al_mv_average_col = 0;
1019     }
1020     row_diff = al_mv_average_row - mv_row;
1021     col_diff = al_mv_average_col - mv_col;
1022     if (row_diff > 80 || row_diff < -80 || col_diff > 80 || col_diff < -80) {
1023       if (bsize >= BLOCK_32X32)
1024         this_rdc->rdcost = this_rdc->rdcost << 1;
1025       else
1026         this_rdc->rdcost = 5 * this_rdc->rdcost >> 2;
1027     }
1028   } else {
1029     // Bias for speed >= 8 for low spatial variance.
1030     if (speed >= 8 && spatial_variance < 150 &&
1031         (mv_row > 64 || mv_row < -64 || mv_col > 64 || mv_col < -64))
1032       this_rdc->rdcost = 5 * this_rdc->rdcost >> 2;
1033   }
1034 }
1035 
1036 static void model_rd_for_sb_uv(AV1_COMP *cpi, BLOCK_SIZE plane_bsize,
1037                                MACROBLOCK *x, MACROBLOCKD *xd,
1038                                RD_STATS *this_rdc, unsigned int *var_y,
1039                                unsigned int *sse_y, int start_plane,
1040                                int stop_plane) {
1041   // Note our transform coeffs are 8 times an orthogonal transform.
1042   // Hence quantizer step is also 8 times. To get effective quantizer
1043   // we need to divide by 8 before sending to modeling function.
1044   unsigned int sse;
1045   int rate;
1046   int64_t dist;
1047   int i;
1048   uint32_t tot_var = *var_y;
1049   uint32_t tot_sse = *sse_y;
1050 
1051   this_rdc->rate = 0;
1052   this_rdc->dist = 0;
1053   this_rdc->skip = 0;
1054 
1055   for (i = start_plane; i <= stop_plane; ++i) {
1056     struct macroblock_plane *const p = &x->plane[i];
1057     struct macroblockd_plane *const pd = &xd->plane[i];
1058     const uint32_t dc_quant = p->dequant_QTX[0];
1059     const uint32_t ac_quant = p->dequant_QTX[1];
1060     const BLOCK_SIZE bs = plane_bsize;
1061     unsigned int var;
1062     if (!x->color_sensitivity[i - 1]) continue;
1063 
1064     var = cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf,
1065                              pd->dst.stride, &sse);
1066     assert(sse >= var);
1067     tot_var += var;
1068     tot_sse += sse;
1069 
1070     av1_model_rd_from_var_lapndz(sse - var, num_pels_log2_lookup[bs],
1071                                  dc_quant >> 3, &rate, &dist);
1072 
1073     this_rdc->rate += rate >> 1;
1074     this_rdc->dist += dist << 3;
1075 
1076     av1_model_rd_from_var_lapndz(var, num_pels_log2_lookup[bs], ac_quant >> 3,
1077                                  &rate, &dist);
1078 
1079     this_rdc->rate += rate;
1080     this_rdc->dist += dist << 4;
1081   }
1082 
1083   if (this_rdc->rate == 0) {
1084     this_rdc->skip = 1;
1085   }
1086 
1087   if (RDCOST(x->rdmult, this_rdc->rate, this_rdc->dist) >=
1088       RDCOST(x->rdmult, 0, ((int64_t)tot_sse) << 4)) {
1089     this_rdc->rate = 0;
1090     this_rdc->dist = tot_sse << 4;
1091     this_rdc->skip = 1;
1092   }
1093 
1094   *var_y = tot_var;
1095   *sse_y = tot_sse;
1096 }
1097 
1098 struct estimate_block_intra_args {
1099   AV1_COMP *cpi;
1100   MACROBLOCK *x;
1101   PREDICTION_MODE mode;
1102   int skippable;
1103   RD_STATS *rdc;
1104 };
1105 
1106 static void estimate_block_intra(int plane, int block, int row, int col,
1107                                  BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
1108                                  void *arg) {
1109   struct estimate_block_intra_args *const args = arg;
1110   AV1_COMP *const cpi = args->cpi;
1111   AV1_COMMON *const cm = &cpi->common;
1112   MACROBLOCK *const x = args->x;
1113   MACROBLOCKD *const xd = &x->e_mbd;
1114   struct macroblock_plane *const p = &x->plane[plane];
1115   struct macroblockd_plane *const pd = &xd->plane[plane];
1116   const BLOCK_SIZE bsize_tx = txsize_to_bsize[tx_size];
1117   uint8_t *const src_buf_base = p->src.buf;
1118   uint8_t *const dst_buf_base = pd->dst.buf;
1119   const int64_t src_stride = p->src.stride;
1120   const int64_t dst_stride = pd->dst.stride;
1121   RD_STATS this_rdc;
1122 
1123   (void)block;
1124 
1125   p->src.buf = &src_buf_base[4 * (row * src_stride + col)];
1126   pd->dst.buf = &dst_buf_base[4 * (row * dst_stride + col)];
1127 
1128   av1_predict_intra_block_facade(cm, xd, plane, col, row, tx_size);
1129 
1130   if (plane == 0) {
1131     int64_t this_sse = INT64_MAX;
1132     block_yrd(cpi, x, 0, 0, &this_rdc, &args->skippable, &this_sse, bsize_tx,
1133               AOMMIN(tx_size, TX_16X16));
1134   } else {
1135     unsigned int var = 0;
1136     unsigned int sse = 0;
1137     model_rd_for_sb_uv(cpi, plane_bsize, x, xd, &this_rdc, &var, &sse, plane,
1138                        plane);
1139   }
1140 
1141   p->src.buf = src_buf_base;
1142   pd->dst.buf = dst_buf_base;
1143   args->rdc->rate += this_rdc.rate;
1144   args->rdc->dist += this_rdc.dist;
1145 }
1146 
1147 static INLINE void update_thresh_freq_fact(AV1_COMP *cpi, MACROBLOCK *x,
1148                                            BLOCK_SIZE bsize,
1149                                            MV_REFERENCE_FRAME ref_frame,
1150                                            THR_MODES best_mode_idx,
1151                                            PREDICTION_MODE mode) {
1152   THR_MODES thr_mode_idx = mode_idx[ref_frame][mode_offset(mode)];
1153   int *freq_fact = &x->thresh_freq_fact[bsize][thr_mode_idx];
1154   if (thr_mode_idx == best_mode_idx) {
1155     *freq_fact -= (*freq_fact >> 4);
1156   } else {
1157     *freq_fact =
1158         AOMMIN(*freq_fact + RD_THRESH_INC,
1159                cpi->sf.inter_sf.adaptive_rd_thresh * RD_THRESH_MAX_FACT);
1160   }
1161 }
1162 
1163 static INLINE int get_force_skip_low_temp_var_small_sb(uint8_t *variance_low,
1164                                                        int mi_row, int mi_col,
1165                                                        BLOCK_SIZE bsize) {
1166   // Relative indices of MB inside the superblock.
1167   const int mi_x = mi_row & 0xF;
1168   const int mi_y = mi_col & 0xF;
1169   // Relative indices of 16x16 block inside the superblock.
1170   const int i = mi_x >> 2;
1171   const int j = mi_y >> 2;
1172   int force_skip_low_temp_var = 0;
1173   // Set force_skip_low_temp_var based on the block size and block offset.
1174   switch (bsize) {
1175     case BLOCK_64X64: force_skip_low_temp_var = variance_low[0]; break;
1176     case BLOCK_64X32:
1177       if (!mi_y && !mi_x) {
1178         force_skip_low_temp_var = variance_low[1];
1179       } else if (!mi_y && mi_x) {
1180         force_skip_low_temp_var = variance_low[2];
1181       }
1182       break;
1183     case BLOCK_32X64:
1184       if (!mi_y && !mi_x) {
1185         force_skip_low_temp_var = variance_low[3];
1186       } else if (mi_y && !mi_x) {
1187         force_skip_low_temp_var = variance_low[4];
1188       }
1189       break;
1190     case BLOCK_32X32:
1191       if (!mi_y && !mi_x) {
1192         force_skip_low_temp_var = variance_low[5];
1193       } else if (mi_y && !mi_x) {
1194         force_skip_low_temp_var = variance_low[6];
1195       } else if (!mi_y && mi_x) {
1196         force_skip_low_temp_var = variance_low[7];
1197       } else if (mi_y && mi_x) {
1198         force_skip_low_temp_var = variance_low[8];
1199       }
1200       break;
1201     case BLOCK_32X16:
1202     case BLOCK_16X32:
1203     case BLOCK_16X16:
1204       force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]];
1205       break;
1206     default: break;
1207   }
1208 
1209   return force_skip_low_temp_var;
1210 }
1211 
1212 static INLINE int get_force_skip_low_temp_var(uint8_t *variance_low, int mi_row,
1213                                               int mi_col, BLOCK_SIZE bsize) {
1214   int force_skip_low_temp_var = 0;
1215   int x, y;
1216   x = (mi_col & 0x1F) >> 4;
1217   // y = (mi_row & 0x1F) >> 4;
1218   // const int idx64 = (y << 1) + x;
1219   y = (mi_row & 0x17) >> 3;
1220   const int idx64 = y + x;
1221 
1222   x = (mi_col & 0xF) >> 3;
1223   // y = (mi_row & 0xF) >> 3;
1224   // const int idx32 = (y << 1) + x;
1225   y = (mi_row & 0xB) >> 2;
1226   const int idx32 = y + x;
1227 
1228   x = (mi_col & 0x7) >> 2;
1229   // y = (mi_row & 0x7) >> 2;
1230   // const int idx16 = (y << 1) + x;
1231   y = (mi_row & 0x5) >> 1;
1232   const int idx16 = y + x;
1233   // Set force_skip_low_temp_var based on the block size and block offset.
1234   switch (bsize) {
1235     case BLOCK_128X128: force_skip_low_temp_var = variance_low[0]; break;
1236     case BLOCK_128X64:
1237       assert((mi_col & 0x1F) == 0);
1238       force_skip_low_temp_var = variance_low[1 + ((mi_row & 0x1F) != 0)];
1239       break;
1240     case BLOCK_64X128:
1241       assert((mi_row & 0x1F) == 0);
1242       force_skip_low_temp_var = variance_low[3 + ((mi_col & 0x1F) != 0)];
1243       break;
1244     case BLOCK_64X64:
1245       // Location of this 64x64 block inside the 128x128 superblock
1246       force_skip_low_temp_var = variance_low[5 + idx64];
1247       break;
1248     case BLOCK_64X32:
1249       x = (mi_col & 0x1F) >> 4;
1250       y = (mi_row & 0x1F) >> 3;
1251       /*
1252       .---------------.---------------.
1253       | x=0,y=0,idx=0 | x=0,y=0,idx=2 |
1254       :---------------+---------------:
1255       | x=0,y=1,idx=1 | x=1,y=1,idx=3 |
1256       :---------------+---------------:
1257       | x=0,y=2,idx=4 | x=1,y=2,idx=6 |
1258       :---------------+---------------:
1259       | x=0,y=3,idx=5 | x=1,y=3,idx=7 |
1260       '---------------'---------------'
1261       */
1262       const int idx64x32 = (x << 1) + (y % 2) + ((y >> 1) << 2);
1263       force_skip_low_temp_var = variance_low[9 + idx64x32];
1264       break;
1265     case BLOCK_32X64:
1266       x = (mi_col & 0x1F) >> 3;
1267       y = (mi_row & 0x1F) >> 4;
1268       const int idx32x64 = (y << 2) + x;
1269       force_skip_low_temp_var = variance_low[17 + idx32x64];
1270       break;
1271     case BLOCK_32X32:
1272       force_skip_low_temp_var = variance_low[25 + (idx64 << 2) + idx32];
1273       break;
1274     case BLOCK_32X16:
1275     case BLOCK_16X32:
1276     case BLOCK_16X16:
1277       force_skip_low_temp_var =
1278           variance_low[41 + (idx64 << 4) + (idx32 << 2) + idx16];
1279       break;
1280     default: break;
1281   }
1282   return force_skip_low_temp_var;
1283 }
1284 
1285 #define FILTER_SEARCH_SIZE 2
1286 static void search_filter_ref(AV1_COMP *cpi, MACROBLOCK *x, RD_STATS *this_rdc,
1287                               int mi_row, int mi_col, PRED_BUFFER *tmp,
1288                               BLOCK_SIZE bsize, int reuse_inter_pred,
1289                               PRED_BUFFER **this_mode_pred, unsigned int *var_y,
1290                               unsigned int *sse_y, int *this_early_term,
1291                               int use_model_yrd_large, int64_t *sse_block_yrd) {
1292   AV1_COMMON *const cm = &cpi->common;
1293   MACROBLOCKD *const xd = &x->e_mbd;
1294   struct macroblockd_plane *const pd = &xd->plane[0];
1295   MB_MODE_INFO *const mi = xd->mi[0];
1296   const int bw = block_size_wide[bsize];
1297   int pf_rate[FILTER_SEARCH_SIZE] = { 0 };
1298   int64_t pf_dist[FILTER_SEARCH_SIZE] = { 0 };
1299   unsigned int pf_var[FILTER_SEARCH_SIZE] = { 0 };
1300   unsigned int pf_sse[FILTER_SEARCH_SIZE] = { 0 };
1301   int64_t pf_sse_block_yrd[FILTER_SEARCH_SIZE] = { 0 };
1302   TX_SIZE pf_tx_size[FILTER_SEARCH_SIZE] = { 0 };
1303   PRED_BUFFER *current_pred = *this_mode_pred;
1304   int skip_txfm[FILTER_SEARCH_SIZE] = { 0 };
1305   int best_skip = 0;
1306   int best_early_term = 0;
1307   int64_t best_cost = INT64_MAX;
1308   int best_filter_index = -1;
1309   InterpFilter filters[FILTER_SEARCH_SIZE] = { EIGHTTAP_REGULAR,
1310                                                EIGHTTAP_SMOOTH };
1311   int i;
1312   for (i = 0; i < FILTER_SEARCH_SIZE; ++i) {
1313     int64_t cost;
1314     InterpFilter filter = filters[i];
1315     mi->interp_filters = av1_broadcast_interp_filter(filter);
1316     av1_enc_build_inter_predictor_y(xd, mi_row, mi_col);
1317     if (use_model_yrd_large)
1318       model_skip_for_sb_y_large(cpi, bsize, mi_row, mi_col, x, xd, &pf_rate[i],
1319                                 &pf_dist[i], &pf_var[i], &pf_sse[i],
1320                                 this_early_term, 1);
1321     else
1322       model_rd_for_sb_y(cpi, bsize, x, xd, &pf_rate[i], &pf_dist[i],
1323                         &skip_txfm[i], NULL, &pf_var[i], &pf_sse[i], 1);
1324     pf_rate[i] += av1_get_switchable_rate(x, xd, cm->features.interp_filter);
1325     cost = RDCOST(x->rdmult, pf_rate[i], pf_dist[i]);
1326     pf_tx_size[i] = mi->tx_size;
1327     if (cost < best_cost) {
1328       best_filter_index = i;
1329       best_cost = cost;
1330       best_skip = skip_txfm[i];
1331       best_early_term = *this_early_term;
1332       if (reuse_inter_pred) {
1333         if (*this_mode_pred != current_pred) {
1334           free_pred_buffer(*this_mode_pred);
1335           *this_mode_pred = current_pred;
1336         }
1337         current_pred = &tmp[get_pred_buffer(tmp, 3)];
1338         pd->dst.buf = current_pred->data;
1339         pd->dst.stride = bw;
1340       }
1341     }
1342   }
1343   assert(best_filter_index >= 0 && best_filter_index < FILTER_SEARCH_SIZE);
1344   if (reuse_inter_pred && *this_mode_pred != current_pred)
1345     free_pred_buffer(current_pred);
1346 
1347   mi->interp_filters = av1_broadcast_interp_filter(filters[best_filter_index]);
1348   mi->tx_size = pf_tx_size[best_filter_index];
1349   this_rdc->rate = pf_rate[best_filter_index];
1350   this_rdc->dist = pf_dist[best_filter_index];
1351   *var_y = pf_var[best_filter_index];
1352   *sse_y = pf_sse[best_filter_index];
1353   *sse_block_yrd = pf_sse_block_yrd[best_filter_index];
1354   this_rdc->skip = (best_skip || best_early_term);
1355   *this_early_term = best_early_term;
1356   if (reuse_inter_pred) {
1357     pd->dst.buf = (*this_mode_pred)->data;
1358     pd->dst.stride = (*this_mode_pred)->stride;
1359   } else if (best_filter_index < FILTER_SEARCH_SIZE - 1) {
1360     av1_enc_build_inter_predictor_y(xd, mi_row, mi_col);
1361   }
1362 }
1363 
1364 #define COLLECT_PICK_MODE_STAT 0
1365 
1366 #if COLLECT_PICK_MODE_STAT
1367 typedef struct _mode_search_stat {
1368   int32_t num_blocks[BLOCK_SIZES];
1369   int64_t avg_block_times[BLOCK_SIZES];
1370   int32_t num_searches[BLOCK_SIZES][MB_MODE_COUNT];
1371   int32_t num_nonskipped_searches[BLOCK_SIZES][MB_MODE_COUNT];
1372   int64_t search_times[BLOCK_SIZES][MB_MODE_COUNT];
1373   int64_t nonskipped_search_times[BLOCK_SIZES][MB_MODE_COUNT];
1374   struct aom_usec_timer timer1;
1375   struct aom_usec_timer timer2;
1376 } mode_search_stat;
1377 #endif  // COLLECT_PICK_MODE_STAT
1378 
1379 static void compute_intra_yprediction(const AV1_COMMON *cm,
1380                                       PREDICTION_MODE mode, BLOCK_SIZE bsize,
1381                                       MACROBLOCK *x, MACROBLOCKD *xd) {
1382   struct macroblockd_plane *const pd = &xd->plane[0];
1383   struct macroblock_plane *const p = &x->plane[0];
1384   uint8_t *const src_buf_base = p->src.buf;
1385   uint8_t *const dst_buf_base = pd->dst.buf;
1386   const int src_stride = p->src.stride;
1387   const int dst_stride = pd->dst.stride;
1388   int plane = 0;
1389   int row, col;
1390   // block and transform sizes, in number of 4x4 blocks log 2 ("*_b")
1391   // 4x4=0, 8x8=2, 16x16=4, 32x32=6, 64x64=8
1392   // transform size varies per plane, look it up in a common way.
1393   const TX_SIZE tx_size = max_txsize_lookup[bsize];
1394   const BLOCK_SIZE plane_bsize =
1395       get_plane_block_size(bsize, pd->subsampling_x, pd->subsampling_y);
1396   // If mb_to_right_edge is < 0 we are in a situation in which
1397   // the current block size extends into the UMV and we won't
1398   // visit the sub blocks that are wholly within the UMV.
1399   const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);
1400   const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
1401   // Keep track of the row and column of the blocks we use so that we know
1402   // if we are in the unrestricted motion border.
1403   for (row = 0; row < max_blocks_high; row += (1 << tx_size)) {
1404     // Skip visiting the sub blocks that are wholly within the UMV.
1405     for (col = 0; col < max_blocks_wide; col += (1 << tx_size)) {
1406       p->src.buf = &src_buf_base[4 * (row * (int64_t)src_stride + col)];
1407       pd->dst.buf = &dst_buf_base[4 * (row * (int64_t)dst_stride + col)];
1408       av1_predict_intra_block(cm, xd, block_size_wide[bsize],
1409                               block_size_high[bsize], tx_size, mode, 0, 0,
1410                               FILTER_INTRA_MODES, pd->dst.buf, dst_stride,
1411                               pd->dst.buf, dst_stride, 0, 0, plane);
1412     }
1413   }
1414   p->src.buf = src_buf_base;
1415   pd->dst.buf = dst_buf_base;
1416 }
1417 
1418 void av1_pick_intra_mode(AV1_COMP *cpi, MACROBLOCK *x, RD_STATS *rd_cost,
1419                          BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
1420   AV1_COMMON *const cm = &cpi->common;
1421   MACROBLOCKD *const xd = &x->e_mbd;
1422   MB_MODE_INFO *const mi = xd->mi[0];
1423   RD_STATS this_rdc, best_rdc;
1424   struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
1425   const TX_SIZE intra_tx_size =
1426       AOMMIN(max_txsize_lookup[bsize],
1427              tx_mode_to_biggest_tx_size[x->tx_mode_search_type]);
1428   int *bmode_costs;
1429   const MB_MODE_INFO *above_mi = xd->above_mbmi;
1430   const MB_MODE_INFO *left_mi = xd->left_mbmi;
1431   const PREDICTION_MODE A = av1_above_block_mode(above_mi);
1432   const PREDICTION_MODE L = av1_left_block_mode(left_mi);
1433   bmode_costs = x->y_mode_costs[A][L];
1434 
1435   av1_invalid_rd_stats(&best_rdc);
1436   av1_invalid_rd_stats(&this_rdc);
1437 
1438   init_mbmi(mi, DC_PRED, INTRA_FRAME, NONE_FRAME, cm);
1439   mi->mv[0].as_int = mi->mv[1].as_int = INVALID_MV;
1440 
1441   // Change the limit of this loop to add other intra prediction
1442   // mode tests.
1443   for (int i = 0; i < 4; ++i) {
1444     PREDICTION_MODE this_mode = intra_mode_list[i];
1445     this_rdc.dist = this_rdc.rate = 0;
1446     args.mode = this_mode;
1447     args.skippable = 1;
1448     args.rdc = &this_rdc;
1449     mi->tx_size = intra_tx_size;
1450     av1_foreach_transformed_block_in_plane(xd, bsize, 0, estimate_block_intra,
1451                                            &args);
1452     if (args.skippable) {
1453       this_rdc.rate = av1_cost_symbol(av1_get_skip_cdf(xd)[1]);
1454     } else {
1455       this_rdc.rate += av1_cost_symbol(av1_get_skip_cdf(xd)[0]);
1456     }
1457     this_rdc.rate += bmode_costs[this_mode];
1458     this_rdc.rdcost = RDCOST(x->rdmult, this_rdc.rate, this_rdc.dist);
1459 
1460     if (this_rdc.rdcost < best_rdc.rdcost) {
1461       best_rdc = this_rdc;
1462       mi->mode = this_mode;
1463     }
1464   }
1465 
1466   *rd_cost = best_rdc;
1467 
1468 #if CONFIG_INTERNAL_STATS
1469   store_coding_context(x, ctx, mi->mode);
1470 #else
1471   store_coding_context(x, ctx);
1472 #endif  // CONFIG_INTERNAL_STATS
1473 }
1474 
1475 void av1_nonrd_pick_inter_mode_sb(AV1_COMP *cpi, TileDataEnc *tile_data,
1476                                   MACROBLOCK *x, RD_STATS *rd_cost,
1477                                   BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
1478                                   int64_t best_rd_so_far) {
1479   AV1_COMMON *const cm = &cpi->common;
1480   MACROBLOCKD *const xd = &x->e_mbd;
1481   MB_MODE_INFO *const mi = xd->mi[0];
1482   struct macroblockd_plane *const pd = &xd->plane[0];
1483 
1484   BEST_PICKMODE best_pickmode;
1485   int inter_mode_mask[BLOCK_SIZES];
1486 #if COLLECT_PICK_MODE_STAT
1487   static mode_search_stat ms_stat;
1488 #endif
1489   MV_REFERENCE_FRAME ref_frame;
1490   MV_REFERENCE_FRAME usable_ref_frame, second_ref_frame;
1491   int_mv frame_mv[MB_MODE_COUNT][REF_FRAMES];
1492   uint8_t mode_checked[MB_MODE_COUNT][REF_FRAMES];
1493   struct buf_2d yv12_mb[8][MAX_MB_PLANE];
1494   static const int flag_list[8] = { 0, AOM_LAST_FLAG, 0, 0, AOM_GOLD_FLAG, 0,
1495                                     0, AOM_ALT_FLAG };
1496   RD_STATS this_rdc, best_rdc;
1497   // var_y and sse_y are saved to be used in skipping checking
1498   unsigned int sse_y = UINT_MAX;
1499   unsigned int var_y = UINT_MAX;
1500   const int *const rd_threshes = cpi->rd.threshes[mi->segment_id][bsize];
1501   const int *const rd_thresh_freq_fact = x->thresh_freq_fact[bsize];
1502   InterpFilter filter_ref;
1503   int ref_frame_skip_mask = 0;
1504   int best_pred_sad = INT_MAX;
1505   int best_early_term = 0;
1506   unsigned int ref_costs_single[REF_FRAMES],
1507       ref_costs_comp[REF_FRAMES][REF_FRAMES];
1508   int force_skip_low_temp_var = 0;
1509   int skip_ref_find_pred[8] = { 0 };
1510   unsigned int sse_zeromv_norm = UINT_MAX;
1511   const unsigned int thresh_skip_golden = 500;
1512   int gf_temporal_ref = 0;
1513   const struct segmentation *const seg = &cm->seg;
1514   int num_inter_modes = RT_INTER_MODES;
1515   unsigned char segment_id = mi->segment_id;
1516   PRED_BUFFER tmp[4];
1517   DECLARE_ALIGNED(16, uint8_t, pred_buf[3 * 128 * 128]);
1518   PRED_BUFFER *this_mode_pred = NULL;
1519   const int reuse_inter_pred =
1520       cpi->sf.rt_sf.reuse_inter_pred_nonrd && cm->seq_params.bit_depth == 8;
1521   const int bh = block_size_high[bsize];
1522   const int bw = block_size_wide[bsize];
1523   const int pixels_in_block = bh * bw;
1524   struct buf_2d orig_dst = pd->dst;
1525   const CommonQuantParams *quant_params = &cm->quant_params;
1526 #if COLLECT_PICK_MODE_STAT
1527   aom_usec_timer_start(&ms_stat.timer2);
1528 #endif
1529   int intra_cost_penalty = av1_get_intra_cost_penalty(
1530       quant_params->base_qindex, quant_params->y_dc_delta_q,
1531       cm->seq_params.bit_depth);
1532   int64_t inter_mode_thresh = RDCOST(x->rdmult, intra_cost_penalty, 0);
1533   const int perform_intra_pred = cpi->sf.rt_sf.check_intra_pred_nonrd;
1534   int use_modeled_non_rd_cost = 0;
1535   int enable_filter_search = 0;
1536   InterpFilter default_interp_filter = EIGHTTAP_REGULAR;
1537   int64_t thresh_sad_pred = INT64_MAX;
1538 
1539   (void)best_rd_so_far;
1540 
1541   init_best_pickmode(&best_pickmode);
1542 
1543   for (int i = 0; i < BLOCK_SIZES; ++i) inter_mode_mask[i] = INTER_ALL;
1544 
1545   // TODO(kyslov) Move this to Speed Features
1546   inter_mode_mask[BLOCK_128X128] = INTER_NEAREST_NEAR;
1547 
1548   struct scale_factors *const sf_last = get_ref_scale_factors(cm, LAST_FRAME);
1549   struct scale_factors *const sf_golden =
1550       get_ref_scale_factors(cm, GOLDEN_FRAME);
1551   gf_temporal_ref = 1;
1552   // For temporal long term prediction, check that the golden reference
1553   // is same scale as last reference, otherwise disable.
1554   if ((sf_last->x_scale_fp != sf_golden->x_scale_fp) ||
1555       (sf_last->y_scale_fp != sf_golden->y_scale_fp)) {
1556     gf_temporal_ref = 0;
1557   }
1558 
1559   av1_collect_neighbors_ref_counts(xd);
1560 
1561   estimate_single_ref_frame_costs(cm, xd, x, segment_id, ref_costs_single);
1562   if (cpi->sf.rt_sf.use_comp_ref_nonrd)
1563     estimate_comp_ref_frame_costs(cm, xd, x, segment_id, ref_costs_comp);
1564 
1565   memset(&mode_checked[0][0], 0, MB_MODE_COUNT * REF_FRAMES);
1566   if (reuse_inter_pred) {
1567     for (int i = 0; i < 3; i++) {
1568       tmp[i].data = &pred_buf[pixels_in_block * i];
1569       tmp[i].stride = bw;
1570       tmp[i].in_use = 0;
1571     }
1572     tmp[3].data = pd->dst.buf;
1573     tmp[3].stride = pd->dst.stride;
1574     tmp[3].in_use = 0;
1575   }
1576 
1577   x->force_skip = 0;
1578 
1579   // Instead of using av1_get_pred_context_switchable_interp(xd) to assign
1580   // filter_ref, we use a less strict condition on assigning filter_ref.
1581   // This is to reduce the probabily of entering the flow of not assigning
1582   // filter_ref and then skip filter search.
1583   filter_ref = cm->features.interp_filter;
1584 
1585   // initialize mode decisions
1586   av1_invalid_rd_stats(&best_rdc);
1587   av1_invalid_rd_stats(&this_rdc);
1588   av1_invalid_rd_stats(rd_cost);
1589   mi->sb_type = bsize;
1590   mi->ref_frame[0] = NONE_FRAME;
1591   mi->ref_frame[1] = NONE_FRAME;
1592 
1593   usable_ref_frame =
1594       cpi->sf.rt_sf.use_nonrd_altref_frame ? ALTREF_FRAME : GOLDEN_FRAME;
1595 
1596   if (cpi->rc.frames_since_golden == 0 && gf_temporal_ref) {
1597     skip_ref_find_pred[GOLDEN_FRAME] = 1;
1598     if (!cpi->sf.rt_sf.use_nonrd_altref_frame) usable_ref_frame = LAST_FRAME;
1599   }
1600 
1601   const int mi_row = xd->mi_row;
1602   const int mi_col = xd->mi_col;
1603   const int is_small_sb = (cm->seq_params.sb_size == BLOCK_64X64);
1604   if (cpi->sf.rt_sf.short_circuit_low_temp_var &&
1605       x->nonrd_prune_ref_frame_search) {
1606     if (is_small_sb)
1607       force_skip_low_temp_var = get_force_skip_low_temp_var_small_sb(
1608           &x->variance_low[0], mi_row, mi_col, bsize);
1609     else
1610       force_skip_low_temp_var = get_force_skip_low_temp_var(
1611           &x->variance_low[0], mi_row, mi_col, bsize);
1612     // If force_skip_low_temp_var is set, skip golden reference.
1613     if (force_skip_low_temp_var) {
1614       usable_ref_frame = LAST_FRAME;
1615     }
1616   }
1617 
1618   // If the segment reference frame feature is enabled and it's set to GOLDEN
1619   // reference, then make sure we don't skip checking GOLDEN, this is to
1620   // prevent possibility of not picking any mode.
1621   if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME) &&
1622       get_segdata(seg, mi->segment_id, SEG_LVL_REF_FRAME) == GOLDEN_FRAME) {
1623     usable_ref_frame = GOLDEN_FRAME;
1624     skip_ref_find_pred[GOLDEN_FRAME] = 0;
1625   }
1626 
1627   for (MV_REFERENCE_FRAME ref_frame_iter = LAST_FRAME;
1628        ref_frame_iter <= usable_ref_frame; ++ref_frame_iter) {
1629     // Skip find_predictor if the reference frame is not in the
1630     // ref_frame_flags (i.e., not used as a reference for this frame).
1631     skip_ref_find_pred[ref_frame_iter] =
1632         !(cpi->ref_frame_flags & flag_list[ref_frame_iter]);
1633     if (!skip_ref_find_pred[ref_frame_iter]) {
1634       find_predictors(cpi, x, ref_frame_iter, frame_mv, &ref_frame_skip_mask,
1635                       flag_list, tile_data, yv12_mb, bsize,
1636                       force_skip_low_temp_var);
1637     }
1638   }
1639 
1640   thresh_sad_pred = ((int64_t)x->pred_mv_sad[LAST_FRAME]) << 1;
1641   // Increase threshold for less agressive pruning.
1642   if (cpi->sf.rt_sf.nonrd_prune_ref_frame_search == 1)
1643     thresh_sad_pred += (x->pred_mv_sad[LAST_FRAME] >> 2);
1644 
1645   const int large_block = bsize >= BLOCK_32X32;
1646   const int use_model_yrd_large =
1647       cpi->oxcf.rc_mode == AOM_CBR && large_block &&
1648       !cyclic_refresh_segment_id_boosted(xd->mi[0]->segment_id) &&
1649       quant_params->base_qindex && cm->seq_params.bit_depth == 8;
1650 
1651 #if COLLECT_PICK_MODE_STAT
1652   ms_stat.num_blocks[bsize]++;
1653 #endif
1654   init_mbmi(mi, DC_PRED, NONE_FRAME, NONE_FRAME, cm);
1655   mi->tx_size =
1656       AOMMIN(AOMMIN(max_txsize_lookup[bsize],
1657                     tx_mode_to_biggest_tx_size[x->tx_mode_search_type]),
1658              TX_16X16);
1659 
1660   // TODO(marpan): Look into reducing these conditions. For now constrain
1661   // it to avoid significant bdrate loss.
1662   if (cpi->sf.rt_sf.use_modeled_non_rd_cost &&
1663       quant_params->base_qindex > 120 && x->source_variance > 100 &&
1664       bsize <= BLOCK_16X16 && x->content_state_sb != kLowVarHighSumdiff &&
1665       x->content_state_sb != kHighSad)
1666     use_modeled_non_rd_cost = 1;
1667 
1668   if (cpi->sf.rt_sf.use_nonrd_filter_search) {
1669     enable_filter_search = 1;
1670     if (cpi->sf.interp_sf.cb_pred_filter_search) {
1671       const int bsl = mi_size_wide_log2[bsize];
1672       enable_filter_search =
1673           (((mi_row + mi_col) >> bsl) +
1674            get_chessboard_index(cm->current_frame.frame_number)) &
1675           0x1;
1676     }
1677     if (x->source_variance <=
1678         cpi->sf.interp_sf.disable_filter_search_var_thresh)
1679       enable_filter_search = 0;
1680   }
1681 
1682   for (int idx = 0; idx < num_inter_modes; ++idx) {
1683     int rate_mv = 0;
1684     int mode_rd_thresh;
1685     int mode_index;
1686     int64_t this_sse;
1687     int is_skippable;
1688     int this_early_term = 0;
1689     int skip_this_mv = 0;
1690     int comp_pred = 0;
1691     int force_mv_inter_layer = 0;
1692     PREDICTION_MODE this_mode;
1693     MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1694     second_ref_frame = NONE_FRAME;
1695 
1696     this_mode = ref_mode_set[idx].pred_mode;
1697     ref_frame = ref_mode_set[idx].ref_frame;
1698 
1699 #if COLLECT_PICK_MODE_STAT
1700     aom_usec_timer_start(&ms_stat.timer1);
1701     ms_stat.num_searches[bsize][this_mode]++;
1702 #endif
1703     mi->mode = this_mode;
1704     mi->ref_frame[0] = ref_frame;
1705 
1706     if (ref_frame > usable_ref_frame) continue;
1707     if (skip_ref_find_pred[ref_frame]) continue;
1708 
1709     // Skip non-zero motion for SVC if skip_nonzeromv_ref is set.
1710     if (cpi->use_svc && frame_mv[this_mode][ref_frame].as_int != 0) {
1711       if (ref_frame == LAST_FRAME && cpi->svc.skip_nonzeromv_last)
1712         continue;
1713       else if (ref_frame == GOLDEN_FRAME && cpi->svc.skip_nonzeromv_gf)
1714         continue;
1715     }
1716 
1717     // If the segment reference frame feature is enabled then do nothing if the
1718     // current ref frame is not allowed.
1719     if (segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME) &&
1720         get_segdata(seg, mi->segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame)
1721       continue;
1722 
1723     if (ref_frame != LAST_FRAME && cpi->oxcf.rc_mode == AOM_CBR &&
1724         sse_zeromv_norm < thresh_skip_golden && this_mode == NEWMV)
1725       continue;
1726 
1727     if (!(cpi->ref_frame_flags & flag_list[ref_frame])) continue;
1728 
1729     if (!(inter_mode_mask[bsize] & (1 << this_mode))) continue;
1730 
1731     // Skip testing non-LAST if this flag is set.
1732     if (x->nonrd_prune_ref_frame_search) {
1733       if (x->nonrd_prune_ref_frame_search > 1 && ref_frame != LAST_FRAME &&
1734           (bsize > BLOCK_64X64 || (bsize > BLOCK_16X16 && this_mode == NEWMV)))
1735         continue;
1736 
1737       if (ref_frame != LAST_FRAME && this_mode == NEARMV) continue;
1738     }
1739 
1740     // Skip non-zeromv mode search for non-LAST frame if force_skip_low_temp_var
1741     // is set. If nearestmv for golden frame is 0, zeromv mode will be skipped
1742     // later.
1743     if (!force_mv_inter_layer && force_skip_low_temp_var &&
1744         ref_frame != LAST_FRAME && frame_mv[this_mode][ref_frame].as_int != 0) {
1745       continue;
1746     }
1747 
1748 #if 0
1749         if (x->content_state_sb != kVeryHighSad &&
1750         (cpi->sf.short_circuit_low_temp_var >= 2 ||
1751         (cpi->sf.short_circuit_low_temp_var == 1 && bsize == BLOCK_64X64))
1752         && force_skip_low_temp_var && ref_frame == LAST_FRAME && this_mode ==
1753             NEWMV)  {
1754           continue;
1755         }
1756 #endif
1757 
1758     // Disable this drop out case if the ref frame segment level feature is
1759     // enabled for this segment. This is to prevent the possibility that we
1760     // end up unable to pick any mode.
1761     if (!segfeature_active(seg, mi->segment_id, SEG_LVL_REF_FRAME)) {
1762       // Check for skipping GOLDEN and ALTREF based pred_mv_sad.
1763       if (cpi->sf.rt_sf.nonrd_prune_ref_frame_search > 0 &&
1764           x->pred_mv_sad[ref_frame] != INT_MAX && ref_frame != LAST_FRAME) {
1765         if ((int64_t)(x->pred_mv_sad[ref_frame]) > thresh_sad_pred)
1766           ref_frame_skip_mask |= (1 << ref_frame);
1767       }
1768       if (ref_frame_skip_mask & (1 << ref_frame)) continue;
1769     }
1770 
1771     // Select prediction reference frames.
1772     for (int i = 0; i < MAX_MB_PLANE; i++) {
1773       xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
1774     }
1775 
1776     mi->ref_frame[0] = ref_frame;
1777     mi->ref_frame[1] = second_ref_frame;
1778     set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
1779 
1780     mode_index = mode_idx[ref_frame][INTER_OFFSET(this_mode)];
1781     mode_rd_thresh = best_pickmode.best_mode_skip_txfm
1782                          ? rd_threshes[mode_index] << 1
1783                          : rd_threshes[mode_index];
1784 
1785     // Increase mode_rd_thresh value for non-LAST for improved encoding
1786     // speed
1787     if (ref_frame != LAST_FRAME) {
1788       mode_rd_thresh = mode_rd_thresh << 1;
1789       if (ref_frame == GOLDEN_FRAME && cpi->rc.frames_since_golden > 4)
1790         mode_rd_thresh = mode_rd_thresh << 1;
1791     }
1792 
1793     if (rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
1794                             rd_thresh_freq_fact[mode_index]))
1795       if (frame_mv[this_mode][ref_frame].as_int != 0) continue;
1796 
1797     if (this_mode == NEWMV && !force_mv_inter_layer) {
1798       if (search_new_mv(cpi, x, frame_mv, ref_frame, gf_temporal_ref, bsize,
1799                         mi_row, mi_col, best_pred_sad, &rate_mv, &best_rdc))
1800         continue;
1801     }
1802 
1803     for (PREDICTION_MODE inter_mv_mode = NEARESTMV; inter_mv_mode <= NEWMV;
1804          inter_mv_mode++) {
1805       if (inter_mv_mode == this_mode || comp_pred) continue;
1806       if (mode_checked[inter_mv_mode][ref_frame] &&
1807           frame_mv[this_mode][ref_frame].as_int ==
1808               frame_mv[inter_mv_mode][ref_frame].as_int) {
1809         skip_this_mv = 1;
1810         break;
1811       }
1812     }
1813 
1814     if (skip_this_mv) continue;
1815 
1816     mi->mode = this_mode;
1817     mi->mv[0].as_int = frame_mv[this_mode][ref_frame].as_int;
1818     mi->mv[1].as_int = 0;
1819     if (reuse_inter_pred) {
1820       if (!this_mode_pred) {
1821         this_mode_pred = &tmp[3];
1822       } else {
1823         this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
1824         pd->dst.buf = this_mode_pred->data;
1825         pd->dst.stride = bw;
1826       }
1827     }
1828 #if COLLECT_PICK_MODE_STAT
1829     ms_stat.num_nonskipped_searches[bsize][this_mode]++;
1830 #endif
1831     if (enable_filter_search &&
1832         ((mi->mv[0].as_mv.row & 0x07) || (mi->mv[0].as_mv.col & 0x07)) &&
1833         (ref_frame == LAST_FRAME || !x->nonrd_prune_ref_frame_search)) {
1834       search_filter_ref(cpi, x, &this_rdc, mi_row, mi_col, tmp, bsize,
1835                         reuse_inter_pred, &this_mode_pred, &var_y, &sse_y,
1836                         &this_early_term, use_model_yrd_large, &this_sse);
1837     } else {
1838       mi->interp_filters =
1839           (filter_ref == SWITCHABLE)
1840               ? av1_broadcast_interp_filter(default_interp_filter)
1841               : av1_broadcast_interp_filter(filter_ref);
1842       av1_enc_build_inter_predictor_y(xd, mi_row, mi_col);
1843       if (use_model_yrd_large) {
1844         model_skip_for_sb_y_large(cpi, bsize, mi_row, mi_col, x, xd, NULL, NULL,
1845                                   &var_y, &sse_y, &this_early_term,
1846                                   use_modeled_non_rd_cost);
1847       } else {
1848         model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
1849                           &this_rdc.skip, NULL, &var_y, &sse_y,
1850                           use_modeled_non_rd_cost);
1851       }
1852     }
1853 
1854     if (ref_frame == LAST_FRAME && frame_mv[this_mode][ref_frame].as_int == 0) {
1855       sse_zeromv_norm =
1856           sse_y >> (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
1857     }
1858 
1859     const int skip_ctx = av1_get_skip_context(xd);
1860     const int skip_cost = x->skip_cost[skip_ctx][1];
1861     const int no_skip_cost = x->skip_cost[skip_ctx][0];
1862     if (!this_early_term) {
1863       if (use_modeled_non_rd_cost) {
1864         if (this_rdc.skip) {
1865           this_rdc.rate = skip_cost;
1866         } else {
1867           this_rdc.rate += no_skip_cost;
1868         }
1869       } else {
1870         this_sse = (int64_t)sse_y;
1871         block_yrd(cpi, x, mi_row, mi_col, &this_rdc, &is_skippable, &this_sse,
1872                   bsize, mi->tx_size);
1873         if (this_rdc.skip) {
1874           this_rdc.rate = skip_cost;
1875         } else {
1876           if (RDCOST(x->rdmult, this_rdc.rate, this_rdc.dist) >=
1877               RDCOST(x->rdmult, 0,
1878                      this_sse)) {  // this_sse already multiplied by 16 in
1879                                    // block_yrd
1880             this_rdc.skip = 1;
1881             this_rdc.rate = skip_cost;
1882             this_rdc.dist = this_sse;
1883           } else {
1884             this_rdc.rate += no_skip_cost;
1885           }
1886         }
1887       }
1888     } else {
1889       this_rdc.skip = 1;
1890       this_rdc.rate = skip_cost;
1891       this_rdc.dist = sse_y << 4;
1892     }
1893 
1894     if (!this_early_term &&
1895         (x->color_sensitivity[0] || x->color_sensitivity[1])) {
1896       RD_STATS rdc_uv;
1897       const BLOCK_SIZE uv_bsize = get_plane_block_size(
1898           bsize, xd->plane[1].subsampling_x, xd->plane[1].subsampling_y);
1899       if (x->color_sensitivity[0]) {
1900         av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
1901                                       AOM_PLANE_U, AOM_PLANE_U);
1902       }
1903       if (x->color_sensitivity[1]) {
1904         av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
1905                                       AOM_PLANE_V, AOM_PLANE_V);
1906       }
1907       model_rd_for_sb_uv(cpi, uv_bsize, x, xd, &rdc_uv, &var_y, &sse_y, 1, 2);
1908       this_rdc.rate += rdc_uv.rate;
1909       this_rdc.dist += rdc_uv.dist;
1910       this_rdc.skip = this_rdc.skip && rdc_uv.skip;
1911     }
1912 
1913     // TODO(kyslov) account for UV prediction cost
1914     this_rdc.rate += rate_mv;
1915     const int16_t mode_ctx =
1916         av1_mode_context_analyzer(mbmi_ext->mode_context, mi->ref_frame);
1917     this_rdc.rate += cost_mv_ref(x, this_mode, mode_ctx);
1918 
1919     this_rdc.rate += ref_costs_single[ref_frame];
1920 
1921     this_rdc.rdcost = RDCOST(x->rdmult, this_rdc.rate, this_rdc.dist);
1922     if (cpi->oxcf.rc_mode == AOM_CBR) {
1923       newmv_diff_bias(xd, this_mode, &this_rdc, bsize,
1924                       frame_mv[this_mode][ref_frame].as_mv.row,
1925                       frame_mv[this_mode][ref_frame].as_mv.col, cpi->speed,
1926                       x->source_variance);
1927     }
1928 
1929     mode_checked[this_mode][ref_frame] = 1;
1930 #if COLLECT_PICK_MODE_STAT
1931     aom_usec_timer_mark(&ms_stat.timer1);
1932     ms_stat.nonskipped_search_times[bsize][this_mode] +=
1933         aom_usec_timer_elapsed(&ms_stat.timer1);
1934 #endif
1935     if (this_rdc.rdcost < best_rdc.rdcost) {
1936       best_rdc = this_rdc;
1937       best_early_term = this_early_term;
1938       best_pickmode.best_mode = this_mode;
1939       best_pickmode.best_pred_filter = mi->interp_filters;
1940       best_pickmode.best_tx_size = mi->tx_size;
1941       best_pickmode.best_ref_frame = ref_frame;
1942       best_pickmode.best_mode_skip_txfm = this_rdc.skip;
1943       best_pickmode.best_second_ref_frame = second_ref_frame;
1944       if (reuse_inter_pred) {
1945         free_pred_buffer(best_pickmode.best_pred);
1946         best_pickmode.best_pred = this_mode_pred;
1947       }
1948     } else {
1949       if (reuse_inter_pred) free_pred_buffer(this_mode_pred);
1950     }
1951     if (best_early_term && idx > 0) {
1952       x->force_skip = 1;
1953       break;
1954     }
1955   }
1956 
1957   mi->mode = best_pickmode.best_mode;
1958   mi->interp_filters = best_pickmode.best_pred_filter;
1959   mi->tx_size = best_pickmode.best_tx_size;
1960   memset(mi->inter_tx_size, mi->tx_size, sizeof(mi->inter_tx_size));
1961   mi->ref_frame[0] = best_pickmode.best_ref_frame;
1962   mi->mv[0].as_int =
1963       frame_mv[best_pickmode.best_mode][best_pickmode.best_ref_frame].as_int;
1964   mi->ref_frame[1] = best_pickmode.best_second_ref_frame;
1965   x->force_skip = best_rdc.skip;
1966 
1967   // Perform intra prediction search, if the best SAD is above a certain
1968   // threshold.
1969   mi->angle_delta[PLANE_TYPE_Y] = 0;
1970   mi->angle_delta[PLANE_TYPE_UV] = 0;
1971   mi->filter_intra_mode_info.use_filter_intra = 0;
1972 
1973   uint32_t spatial_var_thresh = 50;
1974   int motion_thresh = 32;
1975   // Adjust thresholds to make intra mode likely tested if the other
1976   // references (golden, alt) are skipped/not checked.
1977   if (cpi->sf.rt_sf.use_nonrd_altref_frame == 0 &&
1978       cpi->sf.rt_sf.nonrd_prune_ref_frame_search > 0) {
1979     spatial_var_thresh = 150;
1980     motion_thresh = 0;
1981   }
1982   int do_early_exit_rdthresh = 1;
1983   // Some adjustments to checking intra mode based on source variance.
1984   if (x->source_variance < spatial_var_thresh) {
1985     // If the best inter mode is large motion or non-LAST ref reduce intra cost
1986     // penalty, so intra mode is more likely tested.
1987     if (best_pickmode.best_ref_frame != LAST_FRAME ||
1988         abs(mi->mv[0].as_mv.row) >= motion_thresh ||
1989         abs(mi->mv[0].as_mv.col) >= motion_thresh) {
1990       intra_cost_penalty = intra_cost_penalty >> 2;
1991       inter_mode_thresh = RDCOST(x->rdmult, intra_cost_penalty, 0);
1992       do_early_exit_rdthresh = 0;
1993     }
1994     // For big blocks worth checking intra (since only DC will be checked),
1995     // even if best_early_term is set.
1996     if (bsize >= BLOCK_32X32) best_early_term = 0;
1997   }
1998 
1999   if (best_rdc.rdcost == INT64_MAX ||
2000       (perform_intra_pred && !best_early_term &&
2001        best_rdc.rdcost > inter_mode_thresh &&
2002        bsize <= cpi->sf.part_sf.max_intra_bsize)) {
2003     int64_t this_sse = INT64_MAX;
2004     struct estimate_block_intra_args args = { cpi, x, DC_PRED, 1, 0 };
2005     PRED_BUFFER *const best_pred = best_pickmode.best_pred;
2006     TX_SIZE intra_tx_size =
2007         AOMMIN(AOMMIN(max_txsize_lookup[bsize],
2008                       tx_mode_to_biggest_tx_size[x->tx_mode_search_type]),
2009                TX_16X16);
2010 
2011     if (reuse_inter_pred && best_pred != NULL) {
2012       if (best_pred->data == orig_dst.buf) {
2013         this_mode_pred = &tmp[get_pred_buffer(tmp, 3)];
2014         aom_convolve_copy(best_pred->data, best_pred->stride,
2015                           this_mode_pred->data, this_mode_pred->stride, 0, 0, 0,
2016                           0, bw, bh);
2017         best_pickmode.best_pred = this_mode_pred;
2018       }
2019     }
2020     pd->dst = orig_dst;
2021 
2022     for (int i = 0; i < 4; ++i) {
2023       const PREDICTION_MODE this_mode = intra_mode_list[i];
2024       const THR_MODES mode_index =
2025           mode_idx[INTRA_FRAME][mode_offset(this_mode)];
2026       const int mode_rd_thresh = rd_threshes[mode_index];
2027 
2028       // Only check DC for blocks >= 32X32.
2029       if (this_mode > 0 && bsize >= BLOCK_32X32) continue;
2030 
2031       if (rd_less_than_thresh(best_rdc.rdcost, mode_rd_thresh,
2032                               rd_thresh_freq_fact[mode_index]) &&
2033           (do_early_exit_rdthresh || this_mode == SMOOTH_PRED)) {
2034         continue;
2035       }
2036       const BLOCK_SIZE uv_bsize = get_plane_block_size(
2037           bsize, xd->plane[1].subsampling_x, xd->plane[1].subsampling_y);
2038 
2039       mi->mode = this_mode;
2040       mi->ref_frame[0] = INTRA_FRAME;
2041       mi->ref_frame[1] = NONE_FRAME;
2042 
2043       this_rdc.dist = this_rdc.rate = 0;
2044       args.mode = this_mode;
2045       args.skippable = 1;
2046       args.rdc = &this_rdc;
2047       mi->tx_size = intra_tx_size;
2048       compute_intra_yprediction(cm, this_mode, bsize, x, xd);
2049       // Look into selecting tx_size here, based on prediction residual.
2050       if (use_modeled_non_rd_cost)
2051         model_rd_for_sb_y(cpi, bsize, x, xd, &this_rdc.rate, &this_rdc.dist,
2052                           &this_rdc.skip, NULL, &var_y, &sse_y, 1);
2053       else
2054         block_yrd(cpi, x, mi_row, mi_col, &this_rdc, &args.skippable, &this_sse,
2055                   bsize, mi->tx_size);
2056       // TODO(kyslov@) Need to account for skippable
2057       if (x->color_sensitivity[0]) {
2058         av1_foreach_transformed_block_in_plane(xd, uv_bsize, 1,
2059                                                estimate_block_intra, &args);
2060       }
2061       if (x->color_sensitivity[1]) {
2062         av1_foreach_transformed_block_in_plane(xd, uv_bsize, 2,
2063                                                estimate_block_intra, &args);
2064       }
2065 
2066       int mode_cost = 0;
2067       if (av1_is_directional_mode(this_mode) && av1_use_angle_delta(bsize)) {
2068         mode_cost += x->angle_delta_cost[this_mode - V_PRED]
2069                                         [MAX_ANGLE_DELTA +
2070                                          mi->angle_delta[PLANE_TYPE_Y]];
2071       }
2072       if (this_mode == DC_PRED && av1_filter_intra_allowed_bsize(cm, bsize)) {
2073         mode_cost += x->filter_intra_cost[bsize][0];
2074       }
2075       this_rdc.rate += ref_costs_single[INTRA_FRAME];
2076       this_rdc.rate += intra_cost_penalty;
2077       this_rdc.rate += mode_cost;
2078       this_rdc.rdcost = RDCOST(x->rdmult, this_rdc.rate, this_rdc.dist);
2079 
2080       if (this_rdc.rdcost < best_rdc.rdcost) {
2081         best_rdc = this_rdc;
2082         best_pickmode.best_mode = this_mode;
2083         best_pickmode.best_intra_tx_size = mi->tx_size;
2084         best_pickmode.best_ref_frame = INTRA_FRAME;
2085         best_pickmode.best_second_ref_frame = NONE_FRAME;
2086         mi->uv_mode = this_mode;
2087         mi->mv[0].as_int = INVALID_MV;
2088         mi->mv[1].as_int = INVALID_MV;
2089       }
2090     }
2091 
2092     // Reset mb_mode_info to the best inter mode.
2093     if (best_pickmode.best_ref_frame != INTRA_FRAME) {
2094       mi->tx_size = best_pickmode.best_tx_size;
2095     } else {
2096       mi->tx_size = best_pickmode.best_intra_tx_size;
2097     }
2098   }
2099 
2100   pd->dst = orig_dst;
2101   mi->mode = best_pickmode.best_mode;
2102   mi->ref_frame[0] = best_pickmode.best_ref_frame;
2103   mi->ref_frame[1] = best_pickmode.best_second_ref_frame;
2104 
2105   if (!is_inter_block(mi)) {
2106     mi->interp_filters = av1_broadcast_interp_filter(SWITCHABLE_FILTERS);
2107   }
2108 
2109   if (reuse_inter_pred && best_pickmode.best_pred != NULL) {
2110     PRED_BUFFER *const best_pred = best_pickmode.best_pred;
2111     if (best_pred->data != orig_dst.buf && is_inter_mode(mi->mode)) {
2112       aom_convolve_copy(best_pred->data, best_pred->stride, pd->dst.buf,
2113                         pd->dst.stride, 0, 0, 0, 0, bw, bh);
2114     }
2115   }
2116   if (cpi->sf.inter_sf.adaptive_rd_thresh) {
2117     THR_MODES best_mode_idx =
2118         mode_idx[best_pickmode.best_ref_frame][mode_offset(mi->mode)];
2119     if (best_pickmode.best_ref_frame == INTRA_FRAME) {
2120       // Only consider the modes that are included in the intra_mode_list.
2121       int intra_modes = sizeof(intra_mode_list) / sizeof(PREDICTION_MODE);
2122       for (int i = 0; i < intra_modes; i++) {
2123         update_thresh_freq_fact(cpi, x, bsize, INTRA_FRAME, best_mode_idx,
2124                                 intra_mode_list[i]);
2125       }
2126     } else {
2127       for (ref_frame = LAST_FRAME; ref_frame <= usable_ref_frame; ++ref_frame) {
2128         PREDICTION_MODE this_mode;
2129         if (best_pickmode.best_ref_frame != ref_frame) continue;
2130         for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2131           update_thresh_freq_fact(cpi, x, bsize, ref_frame, best_mode_idx,
2132                                   this_mode);
2133         }
2134       }
2135     }
2136   }
2137 
2138 #if CONFIG_INTERNAL_STATS
2139   store_coding_context(x, ctx, mi->mode);
2140 #else
2141   store_coding_context(x, ctx);
2142 #endif  // CONFIG_INTERNAL_STATS
2143 #if COLLECT_PICK_MODE_STAT
2144   aom_usec_timer_mark(&ms_stat.timer2);
2145   ms_stat.avg_block_times[bsize] += aom_usec_timer_elapsed(&ms_stat.timer2);
2146   //
2147   if ((mi_row + mi_size_high[bsize] >= (cpi->common.mi_params.mi_rows)) &&
2148       (mi_col + mi_size_wide[bsize] >= (cpi->common.mi_params.mi_cols))) {
2149     int i, j;
2150     PREDICTION_MODE used_modes[3] = { NEARESTMV, NEARMV, NEWMV };
2151     BLOCK_SIZE bss[5] = { BLOCK_8X8, BLOCK_16X16, BLOCK_32X32, BLOCK_64X64,
2152                           BLOCK_128X128 };
2153     int64_t total_time = 0l;
2154     int32_t total_blocks = 0;
2155 
2156     printf("\n");
2157     for (i = 0; i < 5; i++) {
2158       printf("BS(%d) Num %d, Avg_time %f: ", bss[i], ms_stat.num_blocks[bss[i]],
2159              ms_stat.num_blocks[bss[i]] > 0
2160                  ? (float)ms_stat.avg_block_times[bss[i]] /
2161                        ms_stat.num_blocks[bss[i]]
2162                  : 0);
2163       total_time += ms_stat.avg_block_times[bss[i]];
2164       total_blocks += ms_stat.num_blocks[bss[i]];
2165       for (j = 0; j < 3; j++) {
2166         printf("Mode %d, %d/%d tps %f ", used_modes[j],
2167                ms_stat.num_nonskipped_searches[bss[i]][used_modes[j]],
2168                ms_stat.num_searches[bss[i]][used_modes[j]],
2169                ms_stat.num_nonskipped_searches[bss[i]][used_modes[j]] > 0
2170                    ? (float)ms_stat
2171                              .nonskipped_search_times[bss[i]][used_modes[j]] /
2172                          ms_stat.num_nonskipped_searches[bss[i]][used_modes[j]]
2173                    : 0l);
2174       }
2175       printf("\n");
2176     }
2177     printf("Total time = %ld. Total blocks = %d\n", total_time, total_blocks);
2178   }
2179   //
2180 #endif  // COLLECT_PICK_MODE_STAT
2181   *rd_cost = best_rdc;
2182 }
2183