AOMedia AV1 Codec
speed_features.h
Go to the documentation of this file.
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 #ifndef AOM_AV1_ENCODER_SPEED_FEATURES_H_
13 #define AOM_AV1_ENCODER_SPEED_FEATURES_H_
14 
15 #include "av1/common/enums.h"
16 #include "av1/encoder/enc_enums.h"
17 #include "av1/encoder/mcomp.h"
18 #include "av1/encoder/encodemb.h"
19 
20 #ifdef __cplusplus
21 extern "C" {
22 #endif
23 
27 #define MAX_MESH_STEP 4
28 
29 typedef struct MESH_PATTERN {
30  int range;
31  int interval;
32 } MESH_PATTERN;
33 
34 enum {
35  GM_FULL_SEARCH,
36  GM_REDUCED_REF_SEARCH_SKIP_L2_L3,
37  GM_REDUCED_REF_SEARCH_SKIP_L2_L3_ARF2,
38 
39  // Same as GM_REDUCED_REF_SEARCH_SKIP_L2_L3_ARF2 but with extra filtering
40  // to keep at most two ref frames
41  GM_SEARCH_CLOSEST_REFS_ONLY,
42 
43  GM_DISABLE_SEARCH
44 } UENUM1BYTE(GM_SEARCH_TYPE);
45 
46 enum {
47  DIST_WTD_COMP_ENABLED,
48  DIST_WTD_COMP_SKIP_MV_SEARCH,
49  DIST_WTD_COMP_DISABLED,
50 } UENUM1BYTE(DIST_WTD_COMP_FLAG);
51 
52 enum {
53  INTRA_ALL = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED) | (1 << D45_PRED) |
54  (1 << D135_PRED) | (1 << D113_PRED) | (1 << D157_PRED) |
55  (1 << D203_PRED) | (1 << D67_PRED) | (1 << SMOOTH_PRED) |
56  (1 << SMOOTH_V_PRED) | (1 << SMOOTH_H_PRED) | (1 << PAETH_PRED),
57  UV_INTRA_ALL =
58  (1 << UV_DC_PRED) | (1 << UV_V_PRED) | (1 << UV_H_PRED) |
59  (1 << UV_D45_PRED) | (1 << UV_D135_PRED) | (1 << UV_D113_PRED) |
60  (1 << UV_D157_PRED) | (1 << UV_D203_PRED) | (1 << UV_D67_PRED) |
61  (1 << UV_SMOOTH_PRED) | (1 << UV_SMOOTH_V_PRED) |
62  (1 << UV_SMOOTH_H_PRED) | (1 << UV_PAETH_PRED) | (1 << UV_CFL_PRED),
63  UV_INTRA_DC = (1 << UV_DC_PRED),
64  UV_INTRA_DC_CFL = (1 << UV_DC_PRED) | (1 << UV_CFL_PRED),
65  UV_INTRA_DC_TM = (1 << UV_DC_PRED) | (1 << UV_PAETH_PRED),
66  UV_INTRA_DC_PAETH_CFL =
67  (1 << UV_DC_PRED) | (1 << UV_PAETH_PRED) | (1 << UV_CFL_PRED),
68  UV_INTRA_DC_H_V = (1 << UV_DC_PRED) | (1 << UV_V_PRED) | (1 << UV_H_PRED),
69  UV_INTRA_DC_H_V_CFL = (1 << UV_DC_PRED) | (1 << UV_V_PRED) |
70  (1 << UV_H_PRED) | (1 << UV_CFL_PRED),
71  UV_INTRA_DC_PAETH_H_V = (1 << UV_DC_PRED) | (1 << UV_PAETH_PRED) |
72  (1 << UV_V_PRED) | (1 << UV_H_PRED),
73  UV_INTRA_DC_PAETH_H_V_CFL = (1 << UV_DC_PRED) | (1 << UV_PAETH_PRED) |
74  (1 << UV_V_PRED) | (1 << UV_H_PRED) |
75  (1 << UV_CFL_PRED),
76  INTRA_DC = (1 << DC_PRED),
77  INTRA_DC_TM = (1 << DC_PRED) | (1 << PAETH_PRED),
78  INTRA_DC_H_V = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED),
79  INTRA_DC_H_V_SMOOTH =
80  (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED) | (1 << SMOOTH_PRED),
81  INTRA_DC_PAETH_H_V =
82  (1 << DC_PRED) | (1 << PAETH_PRED) | (1 << V_PRED) | (1 << H_PRED)
83 };
84 
85 enum {
86  INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << GLOBALMV) |
87  (1 << NEWMV) | (1 << NEAREST_NEARESTMV) | (1 << NEAR_NEARMV) |
88  (1 << NEW_NEWMV) | (1 << NEAREST_NEWMV) | (1 << NEAR_NEWMV) |
89  (1 << NEW_NEARMV) | (1 << NEW_NEARESTMV) | (1 << GLOBAL_GLOBALMV),
90  INTER_NEAREST_NEAR_ZERO = (1 << NEARESTMV) | (1 << NEARMV) | (1 << GLOBALMV) |
91  (1 << NEAREST_NEARESTMV) | (1 << GLOBAL_GLOBALMV) |
92  (1 << NEAREST_NEWMV) | (1 << NEW_NEARESTMV) |
93  (1 << NEW_NEARMV) | (1 << NEAR_NEWMV) |
94  (1 << NEAR_NEARMV),
95  INTER_SINGLE_ALL =
96  (1 << NEARESTMV) | (1 << NEARMV) | (1 << GLOBALMV) | (1 << NEWMV),
97 };
98 
99 enum {
100  DISABLE_ALL_INTER_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
101  (1 << THR_ALTR) | (1 << THR_GOLD) | (1 << THR_LAST),
102 
103  DISABLE_ALL_SPLIT = (1 << THR_INTRA) | DISABLE_ALL_INTER_SPLIT,
104 
105  DISABLE_COMPOUND_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA),
106 
107  LAST_AND_INTRA_SPLIT_ONLY = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
108  (1 << THR_ALTR) | (1 << THR_GOLD)
109 };
110 
111 enum {
112  TXFM_CODING_SF = 1,
113  INTER_PRED_SF = 2,
114  INTRA_PRED_SF = 4,
115  PARTITION_SF = 8,
116  LOOP_FILTER_SF = 16,
117  RD_SKIP_SF = 32,
118  RESERVE_2_SF = 64,
119  RESERVE_3_SF = 128,
120 } UENUM1BYTE(DEV_SPEED_FEATURES);
121 
122 /* This enumeration defines when the rate control recode loop will be
123  * enabled.
124  */
125 enum {
126  /*
127  * No recodes allowed
128  */
129  DISALLOW_RECODE = 0,
130  /*
131  * Allow recode only for KF/ARF/GF frames
132  */
133  ALLOW_RECODE_KFARFGF = 1,
134  /*
135  * Allow recode for all frame types based on bitrate constraints.
136  */
137  ALLOW_RECODE = 2,
138 } UENUM1BYTE(RECODE_LOOP_TYPE);
139 
140 enum {
141  SUBPEL_TREE = 0,
142  SUBPEL_TREE_PRUNED = 1, // Prunes 1/2-pel searches
143  SUBPEL_TREE_PRUNED_MORE = 2, // Prunes 1/2-pel searches more aggressively
144  SUBPEL_SEARCH_METHODS
145 } UENUM1BYTE(SUBPEL_SEARCH_METHOD);
146 
147 enum {
148  // Try the full image with different values.
149  LPF_PICK_FROM_FULL_IMAGE,
150  // Try the full image filter search with non-dual filter only.
151  LPF_PICK_FROM_FULL_IMAGE_NON_DUAL,
152  // Try a small portion of the image with different values.
153  LPF_PICK_FROM_SUBIMAGE,
154  // Estimate the level based on quantizer and frame type
155  LPF_PICK_FROM_Q,
156  // Pick 0 to disable LPF if LPF was enabled last frame
157  LPF_PICK_MINIMAL_LPF
158 } UENUM1BYTE(LPF_PICK_METHOD);
164 typedef enum {
173  CDEF_PICK_METHODS
175 
177 enum {
178  // Terminate search early based on distortion so far compared to
179  // qp step, distortion in the neighborhood of the frame, etc.
180  FLAG_EARLY_TERMINATE = 1 << 0,
181 
182  // Skips comp inter modes if the best so far is an intra mode.
183  FLAG_SKIP_COMP_BESTINTRA = 1 << 1,
184 
185  // Skips oblique intra modes if the best so far is an inter mode.
186  FLAG_SKIP_INTRA_BESTINTER = 1 << 3,
187 
188  // Skips oblique intra modes at angles 27, 63, 117, 153 if the best
189  // intra so far is not one of the neighboring directions.
190  FLAG_SKIP_INTRA_DIRMISMATCH = 1 << 4,
191 
192  // Skips intra modes other than DC_PRED if the source variance is small
193  FLAG_SKIP_INTRA_LOWVAR = 1 << 5,
194 } UENUM1BYTE(MODE_SEARCH_SKIP_LOGIC);
195 
196 enum {
197  // No tx type pruning
198  TX_TYPE_PRUNE_0 = 0,
199  // adaptively prunes the least perspective tx types out of all 16
200  // (tuned to provide negligible quality loss)
201  TX_TYPE_PRUNE_1 = 1,
202  // similar, but applies much more aggressive pruning to get better speed-up
203  TX_TYPE_PRUNE_2 = 2,
204  TX_TYPE_PRUNE_3 = 3,
205  // More aggressive pruning based on tx type score and allowed tx count
206  TX_TYPE_PRUNE_4 = 4,
207  TX_TYPE_PRUNE_5 = 5,
208 } UENUM1BYTE(TX_TYPE_PRUNE_MODE);
209 
210 enum {
211  // No reaction to rate control on a detected slide/scene change.
212  NO_DETECTION = 0,
213 
214  // Set to larger Q based only on the detected slide/scene change and
215  // current/past Q.
216  FAST_DETECTION_MAXQ = 1,
217 } UENUM1BYTE(OVERSHOOT_DETECTION_CBR);
218 
219 enum {
220  // Turns off multi-winner mode. So we will do txfm search on either all modes
221  // if winner mode is off, or we will only on txfm search on a single winner
222  // mode.
223  MULTI_WINNER_MODE_OFF = 0,
224 
225  // Limits the number of winner modes to at most 2
226  MULTI_WINNER_MODE_FAST = 1,
227 
228  // Uses the default number of winner modes, which is 3 for intra mode, and 1
229  // for inter mode.
230  MULTI_WINNER_MODE_DEFAULT = 2,
231 
232  // Maximum number of winner modes allowed.
233  MULTI_WINNER_MODE_LEVELS,
234 } UENUM1BYTE(MULTI_WINNER_MODE_TYPE);
235 
236 enum {
237  PRUNE_NEARMV_OFF = 0, // Turn off nearmv pruning
238  PRUNE_NEARMV_LEVEL1 = 1, // Prune nearmv for qindex (0-85)
239  PRUNE_NEARMV_LEVEL2 = 2, // Prune nearmv for qindex (0-170)
240  PRUNE_NEARMV_LEVEL3 = 3, // Prune nearmv more aggressively for qindex (0-170)
241  PRUNE_NEARMV_MAX = PRUNE_NEARMV_LEVEL3,
242 } UENUM1BYTE(PRUNE_NEARMV_LEVEL);
243 
244 enum {
245  // Default transform search used in evaluation of best inter candidates
246  // (MODE_EVAL stage) and motion mode winner processing (WINNER_MODE_EVAL
247  // stage).
248  TX_SEARCH_DEFAULT = 0,
249  // Transform search in motion mode rd during MODE_EVAL stage.
250  TX_SEARCH_MOTION_MODE,
251  // Transform search in compound type mode rd during MODE_EVAL stage.
252  TX_SEARCH_COMP_TYPE_MODE,
253  // All transform search cases
254  TX_SEARCH_CASES
255 } UENUM1BYTE(TX_SEARCH_CASE);
256 
257 typedef struct {
258  TX_TYPE_PRUNE_MODE prune_2d_txfm_mode;
259  int fast_intra_tx_type_search;
260 
261  // INT_MAX: Disable fast search.
262  // 1 - 1024: Probability threshold used for conditionally forcing tx type,
263  // during mode search.
264  // 0: Force tx type to be DCT_DCT unconditionally, during
265  // mode search.
266  int fast_inter_tx_type_prob_thresh;
267 
268  // Prune less likely chosen transforms for each intra mode. The speed
269  // feature ranges from 0 to 2, for different speed / compression trade offs.
270  int use_reduced_intra_txset;
271 
272  // Use a skip flag prediction model to detect blocks with skip = 1 early
273  // and avoid doing full TX type search for such blocks.
274  int use_skip_flag_prediction;
275 
276  // Threshold used by the ML based method to predict TX block split decisions.
277  int ml_tx_split_thresh;
278 
279  // skip remaining transform type search when we found the rdcost of skip is
280  // better than applying transform
281  int skip_tx_search;
282 
283  // Prune tx type search using previous frame stats.
284  int prune_tx_type_using_stats;
285  // Prune tx type search using estimated RDcost
286  int prune_tx_type_est_rd;
287 
288  // Flag used to control the winner mode processing for tx type pruning for
289  // inter blocks. It enables further tx type mode pruning based on ML model for
290  // mode evaluation and disables tx type mode pruning for winner mode
291  // processing.
292  int winner_mode_tx_type_pruning;
293 } TX_TYPE_SEARCH;
294 
295 enum {
296  // Search partitions using RD criterion
297  SEARCH_PARTITION,
298 
299  // Always use a fixed size partition
300  FIXED_PARTITION,
301 
302  // Partition using source variance
303  VAR_BASED_PARTITION,
304 
305 #if CONFIG_RT_ML_PARTITIONING
306  // Partition using ML model
307  ML_BASED_PARTITION
308 #endif
309 } UENUM1BYTE(PARTITION_SEARCH_TYPE);
310 
311 enum {
312  NOT_IN_USE,
313  DIRECT_PRED,
314  RELAXED_PRED,
315  ADAPT_PRED
316 } UENUM1BYTE(MAX_PART_PRED_MODE);
317 
318 enum {
319  LAST_MV_DATA,
320  CURRENT_Q,
321  QTR_ONLY,
322 } UENUM1BYTE(MV_PREC_LOGIC);
323 
324 enum {
325  SUPERRES_AUTO_ALL, // Tries all possible superres ratios
326  SUPERRES_AUTO_DUAL, // Tries no superres and q-based superres ratios
327  SUPERRES_AUTO_SOLO, // Only apply the q-based superres ratio
328 } UENUM1BYTE(SUPERRES_AUTO_SEARCH_TYPE);
346 typedef enum {
353 
358 typedef enum {
359  NO_PRUNING = -1,
376 
380 typedef enum {
385 
390 typedef enum {
392  0,
394  1,
396  2,
398  3,
400  4,
403 
410 
414  RECODE_LOOP_TYPE recode_loop;
415 
421 
428  MV_PREC_LOGIC high_precision_mv_usage;
429 
438 
442  SUPERRES_AUTO_SEARCH_TYPE superres_auto_search_type;
443 
448 
453 
463 
470 
477 
486 
498 
503 
508 
514 
516 typedef struct TPL_SPEED_FEATURES {
517  // GOP length adaptive decision.
518  // If set to 0, tpl model decides whether a shorter gf interval is better.
519  // If set to 1, tpl stats of ARFs from base layer, (base+1) layer and
520  // (base+2) layer decide whether a shorter gf interval is better.
521  // If set to 2, tpl stats of ARFs from base layer, (base+1) layer and GF boost
522  // decide whether a shorter gf interval is better.
523  // If set to 3, gop length adaptive decision is disabled.
524  int gop_length_decision_method;
525  // Prune the intra modes search by tpl.
526  // If set to 0, we will search all intra modes from DC_PRED to PAETH_PRED.
527  // If set to 1, we only search DC_PRED, V_PRED, and H_PRED.
528  int prune_intra_modes;
529  // This parameter controls which step in the n-step process we start at.
530  int reduce_first_step_size;
531  // Skip motion estimation based on the precision of center MVs and the
532  // difference between center MVs.
533  // If set to 0, motion estimation is skipped for duplicate center MVs
534  // (default). If set to 1, motion estimation is skipped for duplicate
535  // full-pixel center MVs. If set to 2, motion estimation is skipped if the
536  // difference between center MVs is less than the threshold.
537  int skip_alike_starting_mv;
538 
539  // When to stop subpel search.
540  SUBPEL_FORCE_STOP subpel_force_stop;
541 
542  // Which search method to use.
543  SEARCH_METHODS search_method;
544 
545  // Prune starting mvs in TPL based on sad scores.
546  int prune_starting_mv;
547 
548  // Prune reference frames in TPL.
549  int prune_ref_frames_in_tpl;
550 
551  // Support compound predictions.
552  int allow_compound_pred;
553 
554  // Calculate rate and distortion based on Y plane only.
555  int use_y_only_rate_distortion;
556 
557  // Use SAD instead of SATD during intra/inter mode search.
558  // If set to 0, use SATD always.
559  // If set to 1, use SAD during intra/inter mode search for frames in the
560  // higher temporal layers of the hierarchical prediction structure.
561  // If set to 2, use SAD during intra/inter mode search for all frames.
562  // This sf is disabled for the first GF group of the key-frame interval,
563  // i.e., SATD is used during intra/inter mode search of the first GF group.
564  int use_sad_for_mode_decision;
565 
566  // Skip tpl processing for frames of type LF_UPDATE.
567  // This sf is disabled for the first GF group of the key-frame interval.
568  int reduce_num_frames;
569 } TPL_SPEED_FEATURES;
570 
571 typedef struct GLOBAL_MOTION_SPEED_FEATURES {
572  GM_SEARCH_TYPE gm_search_type;
573 
574  // During global motion estimation, prune remaining reference frames in a
575  // given direction(past/future), if the evaluated ref_frame in that direction
576  // yields gm_type as INVALID/TRANSLATION/IDENTITY
577  int prune_ref_frame_for_gm_search;
578 
579  // When the current GM type is set to ZEROMV, prune ZEROMV if its performance
580  // is worse than NEWMV under SSE metric.
581  // 0 : no pruning
582  // 1 : conservative pruning
583  // 2 : aggressive pruning
585 
586  // Disable global motion estimation based on stats of previous frames in the
587  // GF group
588  int disable_gm_search_based_on_stats;
589 
590  // Number of refinement steps to apply after initial model generation
591  int num_refinement_steps;
592 } GLOBAL_MOTION_SPEED_FEATURES;
593 
594 typedef struct PARTITION_SPEED_FEATURES {
595  PARTITION_SEARCH_TYPE partition_search_type;
596 
597  // Used if partition_search_type = FIXED_PARTITION
598  BLOCK_SIZE fixed_partition_size;
599 
600  // Prune extended partition types search based on the current best partition
601  // and the combined rdcost of the subblocks estimated from previous
602  // partitions. Can take values 0 - 2, 0 referring to no pruning, and 1 - 2
603  // increasing aggressiveness of pruning in order.
604  int prune_ext_partition_types_search_level;
605 
606  // Prune part4 based on block size
607  int prune_part4_search;
608 
609  // Use a ML model to prune rectangular, ab and 4-way horz
610  // and vert partitions
611  int ml_prune_partition;
612 
613  // Use a ML model to adaptively terminate partition search after trying
614  // PARTITION_SPLIT. Can take values 0 - 2, 0 meaning not being enabled, and
615  // 1 - 2 increasing aggressiveness in order.
616  int ml_early_term_after_part_split_level;
617 
618  // Skip rectangular partition test when partition type none gives better
619  // rd than partition type split. Can take values 0 - 2, 0 referring to no
620  // skipping, and 1 - 2 increasing aggressiveness of skipping in order.
621  int less_rectangular_check_level;
622 
623  // Use square partition only beyond this block size.
624  BLOCK_SIZE use_square_partition_only_threshold;
625 
626  // Sets max square partition levels for this superblock based on
627  // motion vector and prediction error distribution produced from 16x16
628  // simple motion search
629  MAX_PART_PRED_MODE auto_max_partition_based_on_simple_motion;
630 
631  // Min and max square partition size we enable (block_size) as per auto
632  // min max, but also used by adjust partitioning, and pick_partitioning.
633  BLOCK_SIZE default_min_partition_size;
634  BLOCK_SIZE default_max_partition_size;
635 
636  // Sets level of adjustment of variance-based partitioning during
637  // rd_use_partition 0 - no partition adjustment, 1 - try to merge partitions
638  // for small blocks and high QP, 2 - try to merge partitions, 3 - try to merge
639  // and split leaf partitions and 0 - 3 decreasing aggressiveness in order.
640  int adjust_var_based_rd_partitioning;
641 
642  // Partition search early breakout thresholds.
643  int64_t partition_search_breakout_dist_thr;
644  int partition_search_breakout_rate_thr;
645 
646  // Thresholds for ML based partition search breakout.
647  int ml_partition_search_breakout_thresh[PARTITION_BLOCK_SIZES];
648 
649  // Aggressiveness levels for pruning split and rectangular partitions based on
650  // simple_motion_search. SIMPLE_AGG_LVL0 to SIMPLE_AGG_LVL3 correspond to
651  // simple motion search based pruning. QIDX_BASED_AGG_LVL1 corresponds to
652  // qindex based and simple motion search based pruning.
653  int simple_motion_search_prune_agg;
654 
655  // Perform simple_motion_search on each possible subblock and use it to prune
656  // PARTITION_HORZ and PARTITION_VERT.
657  int simple_motion_search_prune_rect;
658 
659  // Perform simple motion search before none_partition to decide if we
660  // want to remove all partitions other than PARTITION_SPLIT. If set to 0, this
661  // model is disabled. If set to 1, the model attempts to perform
662  // PARTITION_SPLIT only. If set to 2, the model also attempts to prune
663  // PARTITION_SPLIT.
664  int simple_motion_search_split;
665 
666  // Use features from simple_motion_search to terminate prediction block
667  // partition after PARTITION_NONE
668  int simple_motion_search_early_term_none;
669 
670  // Controls whether to reduce the number of motion search steps. If this is 0,
671  // then simple_motion_search has the same number of steps as
672  // single_motion_search (assuming no other speed features). Otherwise, reduce
673  // the number of steps by the value contained in this variable.
674  int simple_motion_search_reduce_search_steps;
675 
676  // This variable controls the maximum block size where intra blocks can be
677  // used in inter frames.
678  // TODO(aconverse): Fold this into one of the other many mode skips
679  BLOCK_SIZE max_intra_bsize;
680 
681  // Use CNN with luma pixels on source frame on each of the 64x64 subblock to
682  // perform partition pruning in intra frames.
683  // 0: No Pruning
684  // 1: Prune split and rectangular partitions only
685  // 2: Prune none, split and rectangular partitions
686  int intra_cnn_based_part_prune_level;
687 
688  // Disable extended partition search if the current bsize is greater than the
689  // threshold. Must be a square block size BLOCK_8X8 or higher.
690  BLOCK_SIZE ext_partition_eval_thresh;
691 
692  // Use best partition decision so far to tune 'ext_partition_eval_thresh'
693  int ext_part_eval_based_on_cur_best;
694 
695  // Disable rectangular partitions for larger block sizes.
696  int rect_partition_eval_thresh;
697 
698  // Prune extended partition search based on whether the split/rect partitions
699  // provided an improvement in the previous search.
700  // 0 : no pruning
701  // 1 : prune 1:4 partition search using winner info from split partitions
702  // 2 : prune 1:4 and AB partition search using split and HORZ/VERT info
703  int prune_ext_part_using_split_info;
704 
705  // Prunt rectangular, AB and 4-way partition based on q index and block size
706  // 0 : no pruning
707  // 1 : prune sub_8x8 at very low quantizers
708  // 2 : prune all block size based on qindex
709  int prune_rectangular_split_based_on_qidx;
710 
711  // Prune rectangular partitions based on 4x4 sub-block variance
712  // false : no pruning
713  // true : prune rectangular partitions based on 4x4 sub-block variance
714  // deviation
715  //
716  // For allintra encode, this speed feature reduces instruction count by 6.4%
717  // for speed=6 with coding performance change less than 0.24%. For AVIF image
718  // encode, this speed feature reduces encode time by 8.14% for speed 6 on a
719  // typical image dataset with coding performance change less than 0.16%. This
720  // speed feature is not applicable to speed >= 7.
721  bool prune_rect_part_using_4x4_var_deviation;
722 
723  // Terminate partition search for child partition,
724  // when NONE and SPLIT partition rd_costs are INT64_MAX.
725  int early_term_after_none_split;
726 
727  // Level used to adjust threshold for av1_ml_predict_breakout(). At lower
728  // levels, more conservative threshold is used, and value of 0 indicates
729  // av1_ml_predict_breakout() is disabled. Value of 3 corresponds to default
730  // case with no adjustment to lbd thresholds.
731  int ml_predict_breakout_level;
732 
733  // Prune sub_8x8 (BLOCK_4X4, BLOCK_4X8 and BLOCK_8X4) partitions.
734  // 0 : no pruning
735  // 1 : pruning based on neighbour block information
736  // 2 : prune always
737  int prune_sub_8x8_partition_level;
738 
739  // Prune rectangular split based on simple motion search split/no_split score.
740  // 0: disable pruning, 1: enable pruning
741  int simple_motion_search_rect_split;
742 
743  // The current encoder adopts a DFS search for block partitions.
744  // Therefore the mode selection and associated rdcost is ready for smaller
745  // blocks before the mode selection for some partition types.
746  // AB partition could use previous rd information and skip mode search.
747  // An example is:
748  //
749  // current block
750  // +---+---+
751  // | |
752  // + +
753  // | |
754  // +-------+
755  //
756  // SPLIT partition has been searched first before trying HORZ_A
757  // +---+---+
758  // | R | R |
759  // +---+---+
760  // | R | R |
761  // +---+---+
762  //
763  // HORZ_A
764  // +---+---+
765  // | | |
766  // +---+---+
767  // | |
768  // +-------+
769  //
770  // With this speed feature, the top two sub blocks can directly use rdcost
771  // searched in split partition, and the mode info is also copied from
772  // saved info. Similarly, the bottom rectangular block can also use
773  // the available information from previous rectangular search.
774  int reuse_prev_rd_results_for_part_ab;
775 
776  // Reuse the best prediction modes found in PARTITION_SPLIT and PARTITION_RECT
777  // when encoding PARTITION_AB.
778  int reuse_best_prediction_for_part_ab;
779 
780  // The current partition search records the best rdcost so far and uses it
781  // in mode search and transform search to early skip when some criteria is
782  // met. For example, when the current rdcost is larger than the best rdcost,
783  // or the model rdcost is larger than the best rdcost times some thresholds.
784  // By default, this feature is turned on to speed up the encoder partition
785  // search.
786  // If disabling it, at speed 0, 30 frames, we could get
787  // about -0.25% quality gain (psnr, ssim, vmaf), with about 13% slowdown.
788  int use_best_rd_for_pruning;
789 
790  // Skip evaluation of non-square partitions based on the corresponding NONE
791  // partition.
792  // 0: no pruning
793  // 1: prune extended partitions if NONE is skippable
794  // 2: on top of 1, prune rectangular partitions if NONE is inter, not a newmv
795  // mode and skippable
796  int skip_non_sq_part_based_on_none;
797 
798  // Disables 8x8 and below partitions for low quantizers.
799  int disable_8x8_part_based_on_qidx;
800 } PARTITION_SPEED_FEATURES;
801 
802 typedef struct MV_SPEED_FEATURES {
803  // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
804  SEARCH_METHODS search_method;
805 
806  // Enable the use of faster, less accurate mv search method
807  // 0: disable, 1: if bsize >= BLOCK_32X32, 2: based on bsize, SAD and qp
808  // TODO(chiyotsai@google.com): Take the clip's resolution and mv activity into
809  // account.
810  int use_bsize_dependent_search_method;
811 
812  // If this is set to 1, we limit the motion search range to 2 times the
813  // largest motion vector found in the last frame.
814  int auto_mv_step_size;
815 
816  // Subpel_search_method can only be subpel_tree which does a subpixel
817  // logarithmic search that keeps stepping at 1/2 pixel units until
818  // you stop getting a gain, and then goes on to 1/4 and repeats
819  // the same process. Along the way it skips many diagonals.
820  SUBPEL_SEARCH_METHOD subpel_search_method;
821 
822  // Maximum number of steps in logarithmic subpel search before giving up.
823  int subpel_iters_per_step;
824 
825  // When to stop subpel search.
826  SUBPEL_FORCE_STOP subpel_force_stop;
827 
828  // When to stop subpel search in simple motion search.
829  SUBPEL_FORCE_STOP simple_motion_subpel_force_stop;
830 
831  // If true, sub-pixel search uses the exact convolve function used for final
832  // encoding and decoding; otherwise, it uses bilinear interpolation.
833  SUBPEL_SEARCH_TYPE use_accurate_subpel_search;
834 
835  // Threshold for allowing exhaustive motion search.
836  int exhaustive_searches_thresh;
837 
838  // Pattern to be used for any exhaustive mesh searches (except intraBC ME).
839  MESH_PATTERN mesh_patterns[MAX_MESH_STEP];
840 
841  // Pattern to be used for exhaustive mesh searches of intraBC ME.
842  MESH_PATTERN intrabc_mesh_patterns[MAX_MESH_STEP];
843 
844  // Reduce single motion search range based on MV result of prior ref_mv_idx.
845  int reduce_search_range;
846 
847  // Prune mesh search.
848  PRUNE_MESH_SEARCH_LEVEL prune_mesh_search;
849 
850  // Use the rd cost around the best FULLPEL_MV to speed up subpel search
851  int use_fullpel_costlist;
852 
853  // Set the full pixel search level of obmc
854  // 0: obmc_full_pixel_diamond
855  // 1: obmc_refining_search_sad (faster)
856  int obmc_full_pixel_search_level;
857 
858  // Accurate full pixel motion search based on TPL stats.
859  int full_pixel_search_level;
860 
861  // Allow intrabc motion search
862  int use_intrabc;
863 
864  // Whether to downsample the rows in sad calculation during motion search.
865  // This is only active when there are at least 16 rows. When this sf is
866  // active, if there is a large discrepancy in the SAD values for the final
867  // motion vector between skipping vs not skipping, motion search is redone
868  // with skip row features off.
869  // 0: Disabled (do not downsample rows)
870  // 1: Skip SAD calculation of odd rows if the SAD deviation of the even and
871  // odd rows for the starting MV is small. Redo motion search with sf off
872  // when SAD deviation is high for the final motion vector.
873  // 2: Skip SAD calculation of odd rows. SAD deviation is not tested for the
874  // start MV and tested only for the final MV.
875  int use_downsampled_sad;
876 
877  // Enable/disable extensive joint motion search.
878  int disable_extensive_joint_motion_search;
879 
880  // Enable second best mv check in joint mv search.
881  // 0: allow second MV (use rd cost as the metric)
882  // 1: use var as the metric
883  // 2: disable second MV
884  int disable_second_mv;
885 
886  // Skips full pixel search based on start mv of prior ref_mv_idx.
887  // 0: Disabled
888  // 1: Skips the full pixel search upto 4 neighbor full-pel MV positions.
889  // 2: Skips the full pixel search upto 8 neighbor full-pel MV positions.
890  int skip_fullpel_search_using_startmv;
891 
892  // Method to use for refining WARPED_CAUSAL motion vectors
893  // TODO(rachelbarker): Can this be unified with OBMC in some way?
894  WARP_SEARCH_METHOD warp_search_method;
895 
896  // Maximum number of iterations in WARPED_CAUSAL refinement search
897  int warp_search_iters;
898 } MV_SPEED_FEATURES;
899 
900 typedef struct INTER_MODE_SPEED_FEATURES {
901  // 2-pass inter mode model estimation where the preliminary pass skips
902  // transform search and uses a model to estimate rd, while the final pass
903  // computes the full transform search. Two types of models are supported:
904  // 0: not used
905  // 1: used with online dynamic rd model
906  // 2: used with static rd model
907  int inter_mode_rd_model_estimation;
908 
909  // Bypass transform search based on skip rd at following stages
910  // i. Compound type mode search
911  // ii. Motion mode search (mode evaluation and winner motion mode stage)
912  // iii. Transform search for best inter candidates
913  int txfm_rd_gate_level[TX_SEARCH_CASES];
914 
915  // Limit the inter mode tested in the RD loop
916  int reduce_inter_modes;
917 
918  // This variable is used to cap the maximum number of times we skip testing a
919  // mode to be evaluated. A high value means we will be faster.
920  int adaptive_rd_thresh;
921 
922  // Aggressively prune inter modes when best mode is skippable.
923  int prune_inter_modes_if_skippable;
924 
925  // Drop less likely to be picked reference frames in the RD search.
926  // Has seven levels for now: 0, 1, 2, 3, 4, 5 and 6 where higher levels prune
927  // more aggressively than lower ones. (0 means no pruning).
928  int selective_ref_frame;
929 
930  // Prune reference frames for rectangular partitions.
931  // 0 implies no pruning
932  // 1 implies prune for extended partition
933  // 2 implies prune horiz, vert and extended partition
934  int prune_ref_frame_for_rect_partitions;
935 
936  // Prune inter modes w.r.t past reference frames
937  // 0 no pruning
938  // 1 prune inter modes w.r.t ALTREF2 and ALTREF reference frames
939  // 2 prune inter modes w.r.t BWDREF, ALTREF2 and ALTREF reference frames
940  int alt_ref_search_fp;
941 
942  // Prune reference frames for single prediction modes based on temporal
943  // distance and pred MV SAD. Feasible values are 0, 1, 2. The feature is
944  // disabled for 0. An increasing value indicates more aggressive pruning
945  // threshold.
946  int prune_single_ref;
947 
948  // Prune compound reference frames
949  // 0 no pruning
950  // 1 prune compound references which do not satisfy the two conditions:
951  // a) The references are at a nearest distance from the current frame in
952  // both past and future direction.
953  // b) The references have minimum pred_mv_sad in both past and future
954  // direction.
955  // 2 prune compound references except the one with nearest distance from the
956  // current frame in both past and future direction.
957  int prune_comp_ref_frames;
958 
959  // Skip the current ref_mv in NEW_MV mode based on mv, rate cost, etc.
960  // This speed feature equaling 0 means no skipping.
961  // If the speed feature equals 1 or 2, skip the current ref_mv in NEW_MV mode
962  // if we have already encountered ref_mv in the drl such that:
963  // 1. The other drl has the same mv during the SIMPLE_TRANSLATION search
964  // process as the current mv.
965  // 2. The rate needed to encode the current mv is larger than that for the
966  // other ref_mv.
967  // The speed feature equaling 1 means using subpel mv in the comparison.
968  // The speed feature equaling 2 means using fullpel mv in the comparison.
969  // If the speed feature >= 3, skip the current ref_mv in NEW_MV mode based on
970  // known full_mv bestsme and drl cost.
971  int skip_newmv_in_drl;
972 
973  // This speed feature checks duplicate ref MVs among NEARESTMV, NEARMV,
974  // GLOBALMV and skips NEARMV or GLOBALMV (in order) if a duplicate is found
975  // TODO(any): Instead of skipping repeated ref mv, use the recalculated
976  // rd-cost based on mode rate and skip the mode evaluation
977  int skip_repeated_ref_mv;
978 
979  // Flag used to control the ref_best_rd based gating for chroma
980  int perform_best_rd_based_gating_for_chroma;
981 
982  // Reuse the inter_intra_mode search result from NEARESTMV mode to other
983  // single ref modes
984  int reuse_inter_intra_mode;
985 
986  // prune wedge and compound segment approximate rd evaluation based on
987  // compound average modeled rd
988  int prune_comp_type_by_model_rd;
989 
990  // prune wedge and compound segment approximate rd evaluation based on
991  // compound average rd/ref_best_rd
992  int prune_comp_type_by_comp_avg;
993 
994  // Skip some ref frames in compound motion search by single motion search
995  // result. Has three levels for now: 0 referring to no skipping, and 1 - 3
996  // increasing aggressiveness of skipping in order.
997  // Note: The search order might affect the result. It assumes that the single
998  // reference modes are searched before compound modes. It is better to search
999  // same single inter mode as a group.
1000  int prune_comp_search_by_single_result;
1001 
1002  // Instead of performing a full MV search, do a simple translation first
1003  // and only perform a full MV search on the motion vectors that performed
1004  // well.
1005  int prune_mode_search_simple_translation;
1006 
1007  // Only search compound modes with at least one "good" reference frame.
1008  // A reference frame is good if, after looking at its performance among
1009  // the single reference modes, it is one of the two best performers.
1010  int prune_compound_using_single_ref;
1011 
1012  // Skip extended compound mode (NEAREST_NEWMV, NEW_NEARESTMV, NEAR_NEWMV,
1013  // NEW_NEARMV) using ref frames of above and left neighbor
1014  // blocks.
1015  // 0 : no pruning
1016  // 1 : prune ext compound modes using neighbor blocks (less aggressiveness)
1017  // 2 : prune ext compound modes using neighbor blocks (high aggressiveness)
1018  // 3 : prune ext compound modes unconditionally (highest aggressiveness)
1019  int prune_ext_comp_using_neighbors;
1020 
1021  // Skip NEW_NEARMV and NEAR_NEWMV extended compound modes
1022  int skip_ext_comp_nearmv_mode;
1023 
1024  // Skip extended compound mode when ref frame corresponding to NEWMV does not
1025  // have NEWMV as single mode winner.
1026  // 0 : no pruning
1027  // 1 : prune extended compound mode (less aggressiveness)
1028  // 2 : prune extended compound mode (high aggressiveness)
1029  int prune_comp_using_best_single_mode_ref;
1030 
1031  // Skip NEARESTMV and NEARMV using weight computed in ref mv list population
1032  //
1033  // Pruning is enabled only when both the top and left neighbor blocks are
1034  // available and when the current block already has a valid inter prediction.
1035  int prune_nearest_near_mv_using_refmv_weight;
1036 
1037  // Based on previous ref_mv_idx search result, prune the following search.
1038  int prune_ref_mv_idx_search;
1039 
1040  // Disable one sided compound modes.
1041  int disable_onesided_comp;
1042 
1043  // Prune obmc search using previous frame stats.
1044  // INT_MAX : disable obmc search
1045  int prune_obmc_prob_thresh;
1046 
1047  // Prune warped motion search using previous frame stats.
1048  int prune_warped_prob_thresh;
1049 
1050  // Variance threshold to enable/disable Interintra wedge search
1051  unsigned int disable_interintra_wedge_var_thresh;
1052 
1053  // Variance threshold to enable/disable Interinter wedge search
1054  unsigned int disable_interinter_wedge_var_thresh;
1055 
1056  // De-couple wedge and mode search during interintra RDO.
1057  int fast_interintra_wedge_search;
1058 
1059  // Whether fast wedge sign estimate is used
1060  int fast_wedge_sign_estimate;
1061 
1062  // Enable/disable ME for interinter wedge search.
1063  int disable_interinter_wedge_newmv_search;
1064 
1065  // Decide when and how to use joint_comp.
1066  DIST_WTD_COMP_FLAG use_dist_wtd_comp_flag;
1067 
1068  // Clip the frequency of updating the mv cost.
1069  INTERNAL_COST_UPDATE_TYPE mv_cost_upd_level;
1070 
1071  // Clip the frequency of updating the coeff cost.
1072  INTERNAL_COST_UPDATE_TYPE coeff_cost_upd_level;
1073 
1074  // Clip the frequency of updating the mode cost.
1075  INTERNAL_COST_UPDATE_TYPE mode_cost_upd_level;
1076 
1077  // Prune inter modes based on tpl stats
1078  // 0 : no pruning
1079  // 1 - 3 indicate increasing aggressiveness in order.
1080  int prune_inter_modes_based_on_tpl;
1081 
1082  // Skip NEARMV and NEAR_NEARMV modes using ref frames of above and left
1083  // neighbor blocks and qindex.
1084  PRUNE_NEARMV_LEVEL prune_nearmv_using_neighbors;
1085 
1086  // Model based breakout after interpolation filter search
1087  // 0: no breakout
1088  // 1: use model based rd breakout
1089  int model_based_post_interp_filter_breakout;
1090 
1091  // Reuse compound type rd decision when exact match is found
1092  // 0: No reuse
1093  // 1: Reuse the compound type decision
1094  int reuse_compound_type_decision;
1095 
1096  // Enable/disable masked compound.
1097  int disable_masked_comp;
1098 
1099  // Enable/disable MV refinement for compound modes corresponds to compound
1100  // types COMPOUND_AVERAGE, COMPOUND_DISTWTD (currently, this compound type
1101  // is disabled for speeds >= 2 using the sf 'use_dist_wtd_comp_flag') and
1102  // COMPOUND_DIFFWTD based on the availability. Levels 0 to 3 indicate
1103  // increasing order of aggressiveness to disable MV refinement.
1104  // 0: MV Refinement is enabled and for NEW_NEWMV mode used two iterations of
1105  // refinement in av1_joint_motion_search().
1106  // 1: MV Refinement is disabled for COMPOUND_DIFFWTD and enabled for
1107  // COMPOUND_AVERAGE & COMPOUND_DISTWTD.
1108  // 2: MV Refinement is enabled for COMPOUND_AVERAGE & COMPOUND_DISTWTD for
1109  // NEW_NEWMV mode with one iteration of refinement in
1110  // av1_joint_motion_search() and MV Refinement is disabled for other compound
1111  // type modes.
1112  // 3: MV Refinement is disabled.
1113  int enable_fast_compound_mode_search;
1114 
1115  // Reuse masked compound type search results
1116  int reuse_mask_search_results;
1117 
1118  // Enable/disable fast search for wedge masks
1119  int enable_fast_wedge_mask_search;
1120 
1121  // Early breakout from transform search of inter modes
1122  int inter_mode_txfm_breakout;
1123 
1124  // Limit number of inter modes for txfm search if a newmv mode gets
1125  // evaluated among the top modes.
1126  // 0: no pruning
1127  // 1 to 3 indicate increasing order of aggressiveness
1128  int limit_inter_mode_cands;
1129 
1130  // Cap the no. of txfm searches for a given prediction mode.
1131  // 0: no cap, 1: cap beyond first 4 searches, 2: cap beyond first 3 searches.
1132  int limit_txfm_eval_per_mode;
1133 
1134  // Prune warped motion search based on block size.
1135  int extra_prune_warped;
1136 
1137  // Do not search compound modes for ARF.
1138  // The intuition is that ARF is predicted by frames far away from it,
1139  // whose temporal correlations with the ARF are likely low.
1140  // It is therefore likely that compound modes do not work as well for ARF
1141  // as other inter frames.
1142  // Speed/quality impact:
1143  // Speed 1: 12% faster, 0.1% psnr loss.
1144  // Speed 2: 2% faster, 0.05% psnr loss.
1145  // No change for speed 3 and up, because |disable_onesided_comp| is true.
1146  int skip_arf_compound;
1147 } INTER_MODE_SPEED_FEATURES;
1148 
1149 typedef struct INTERP_FILTER_SPEED_FEATURES {
1150  // Do limited interpolation filter search for dual filters, since best choice
1151  // usually includes EIGHTTAP_REGULAR.
1152  int use_fast_interpolation_filter_search;
1153 
1154  // Disable dual filter
1155  int disable_dual_filter;
1156 
1157  // Save results of av1_interpolation_filter_search for a block
1158  // Check mv and ref_frames before search, if they are very close with previous
1159  // saved results, filter search can be skipped.
1160  int use_interp_filter;
1161 
1162  // skip sharp_filter evaluation based on regular and smooth filter rd for
1163  // dual_filter=0 case
1164  int skip_sharp_interp_filter_search;
1165 
1166  // skip interpolation filter search for a block in chessboard pattern
1167  int cb_pred_filter_search;
1168 
1169  // adaptive interp_filter search to allow skip of certain filter types.
1170  int adaptive_interp_filter_search;
1171 
1172  // Forces interpolation filter to EIGHTTAP_REGULAR and skips interpolation
1173  // filter search.
1174  int skip_interp_filter_search;
1175 } INTERP_FILTER_SPEED_FEATURES;
1176 
1177 typedef struct INTRA_MODE_SPEED_FEATURES {
1178  // These bit masks allow you to enable or disable intra modes for each
1179  // transform size separately.
1180  int intra_y_mode_mask[TX_SIZES];
1181  int intra_uv_mode_mask[TX_SIZES];
1182 
1183  // flag to allow skipping intra mode for inter frame prediction
1184  int skip_intra_in_interframe;
1185 
1186  // Prune intra mode candidates based on source block histogram of gradient.
1187  // Applies to luma plane only.
1188  // Feasible values are 0..4. The feature is disabled for 0. An increasing
1189  // value indicates more aggressive pruning threshold.
1190  int intra_pruning_with_hog;
1191 
1192  // Prune intra mode candidates based on source block histogram of gradient.
1193  // Applies to chroma plane only.
1194  // Feasible values are 0..4. The feature is disabled for 0. An increasing
1195  // value indicates more aggressive pruning threshold.
1196  int chroma_intra_pruning_with_hog;
1197 
1198  // Enable/disable smooth intra modes.
1199  int disable_smooth_intra;
1200 
1201  // Prune UV_SMOOTH_PRED mode for chroma based on chroma source variance.
1202  // false : No pruning
1203  // true : Prune UV_SMOOTH_PRED mode based on chroma source variance
1204  //
1205  // For allintra encode, this speed feature reduces instruction count
1206  // by 1.90%, 2.21% and 1.97% for speed 6, 7 and 8 with coding performance
1207  // change less than 0.04%. For AVIF image encode, this speed feature reduces
1208  // encode time by 1.56%, 2.14% and 0.90% for speed 6, 7 and 8 on a typical
1209  // image dataset with coding performance change less than 0.05%.
1210  bool prune_smooth_intra_mode_for_chroma;
1211 
1212  // Prune filter intra modes in intra frames.
1213  // 0 : No pruning
1214  // 1 : Evaluate applicable filter intra modes based on best intra mode so far
1215  // 2 : Do not evaluate filter intra modes
1216  int prune_filter_intra_level;
1217 
1218  // prune palette search
1219  // 0: No pruning
1220  // 1: Perform coarse search to prune the palette colors. For winner colors,
1221  // neighbors are also evaluated using a finer search.
1222  // 2: Perform 2 way palette search from max colors to min colors (and min
1223  // colors to remaining colors) and terminate the search if current number of
1224  // palette colors is not the winner.
1225  int prune_palette_search_level;
1226 
1227  // Terminate early in luma palette_size search. Speed feature values indicate
1228  // increasing level of pruning.
1229  // 0: No early termination
1230  // 1: Terminate early for higher luma palette_size, if header rd cost of lower
1231  // palette_size is more than 2 * best_rd. This level of pruning is more
1232  // conservative when compared to sf level 2 as the cases which will get pruned
1233  // with sf level 1 is a subset of the cases which will get pruned with sf
1234  // level 2.
1235  // 2: Terminate early for higher luma palette_size, if header rd cost of lower
1236  // palette_size is more than best_rd.
1237  // For allintra encode, this sf reduces instruction count by 2.49%, 1.07%,
1238  // 2.76%, 2.30%, 1.84%, 2.69%, 2.04%, 2.05% and 1.44% for speed 0, 1, 2, 3, 4,
1239  // 5, 6, 7 and 8 on screen content set with coding performance change less
1240  // than 0.01% for speed <= 2 and less than 0.03% for speed >= 3. For AVIF
1241  // image encode, this sf reduces instruction count by 1.94%, 1.13%, 1.29%,
1242  // 0.93%, 0.89%, 1.03%, 1.07%, 1.20% and 0.18% for speed 0, 1, 2, 3, 4, 5, 6,
1243  // 7 and 8 on a typical image dataset with coding performance change less than
1244  // 0.01%.
1245  int prune_luma_palette_size_search_level;
1246 
1247  // Prune chroma intra modes based on luma intra mode winner.
1248  // 0: No pruning
1249  // 1: Prune chroma intra modes other than UV_DC_PRED, UV_SMOOTH_PRED,
1250  // UV_CFL_PRED and the mode that corresponds to luma intra mode winner.
1251  int prune_chroma_modes_using_luma_winner;
1252 
1253  // Clip the frequency of updating the mv cost for intrabc.
1254  INTERNAL_COST_UPDATE_TYPE dv_cost_upd_level;
1255 
1256  // We use DCT_DCT transform followed by computing SATD (Sum of Absolute
1257  // Transformed Differences) as an estimation of RD score to quickly find the
1258  // best possible Chroma from Luma (CFL) parameter. Then we do a full RD search
1259  // near the best possible parameter. The search range is set here.
1260  // The range of cfl_searh_range should be [1, 33], and the following are the
1261  // recommended values.
1262  // 1: Fastest mode.
1263  // 3: Default mode that provides good speedup without losing compression
1264  // performance at speed 0.
1265  // 33: Exhaustive rd search (33 == CFL_MAGS_SIZE). This mode should only
1266  // be used for debugging purpose.
1267  int cfl_search_range;
1268 
1269  // TOP_INTRA_MODEL_COUNT is 4 that is the number of top model rd to store in
1270  // intra mode decision. Here, add a speed feature to reduce this number for
1271  // higher speeds.
1272  int top_intra_model_count_allowed;
1273 
1274  // Adapt top_intra_model_count_allowed locally to prune luma intra modes using
1275  // neighbor block and quantizer information.
1276  int adapt_top_model_rd_count_using_neighbors;
1277 
1278  // Prune the evaluation of odd delta angles of directional luma intra modes by
1279  // using the rdcosts of neighbouring delta angles.
1280  // For allintra encode, this speed feature reduces instruction count
1281  // by 4.461%, 3.699% and 3.536% for speed 6, 7 and 8 on a typical video
1282  // dataset with coding performance change less than 0.26%. For AVIF image
1283  // encode, this speed feature reduces encode time by 2.849%, 2.471%,
1284  // and 2.051% for speed 6, 7 and 8 on a typical image dataset with coding
1285  // performance change less than 0.27%.
1286  int prune_luma_odd_delta_angles_in_intra;
1287 
1288  // Terminate early in chroma palette_size search.
1289  // 0: No early termination
1290  // 1: Terminate early for higher palette_size, if header rd cost of lower
1291  // palette_size is more than best_rd.
1292  // For allintra encode, this sf reduces instruction count by 0.45%,
1293  // 0.62%, 1.73%, 2.50%, 2.89%, 3.09% and 3.86% for speed 0 to 6 on screen
1294  // content set with coding performance change less than 0.01%.
1295  // For AVIF image encode, this sf reduces instruction count by 0.45%, 0.81%,
1296  // 0.85%, 1.05%, 1.45%, 1.66% and 1.95% for speed 0 to 6 on a typical image
1297  // dataset with no quality drop.
1298  int early_term_chroma_palette_size_search;
1299 
1300  // Skips the evaluation of filter intra modes in inter frames if rd evaluation
1301  // of luma intra dc mode results in invalid rd stats.
1302  int skip_filter_intra_in_inter_frames;
1303 } INTRA_MODE_SPEED_FEATURES;
1304 
1305 typedef struct TX_SPEED_FEATURES {
1306  // Init search depth for square and rectangular transform partitions.
1307  // Values:
1308  // 0 - search full tree, 1: search 1 level, 2: search the highest level only
1309  int inter_tx_size_search_init_depth_sqr;
1310  int inter_tx_size_search_init_depth_rect;
1311  int intra_tx_size_search_init_depth_sqr;
1312  int intra_tx_size_search_init_depth_rect;
1313 
1314  // If any dimension of a coding block size above 64, always search the
1315  // largest transform only, since the largest transform block size is 64x64.
1316  int tx_size_search_lgr_block;
1317 
1318  TX_TYPE_SEARCH tx_type_search;
1319 
1320  // Skip split transform block partition when the collocated bigger block
1321  // is selected as all zero coefficients.
1322  int txb_split_cap;
1323 
1324  // Shortcut the transform block partition and type search when the target
1325  // rdcost is relatively lower.
1326  // Values are 0 (not used) , or 1 - 2 with progressively increasing
1327  // aggressiveness
1328  int adaptive_txb_search_level;
1329 
1330  // Prune level for tx_size_type search for inter based on rd model
1331  // 0: no pruning
1332  // 1-2: progressively increasing aggressiveness of pruning
1333  int model_based_prune_tx_search_level;
1334 
1335  // Refine TX type after fast TX search.
1336  int refine_fast_tx_search_results;
1337 
1338  // Prune transform split/no_split eval based on residual properties. A value
1339  // of 0 indicates no pruning, and the aggressiveness of pruning progressively
1340  // increases from levels 1 to 3.
1341  int prune_tx_size_level;
1342 
1343  // Prune the evaluation of transform depths as decided by the NN model.
1344  // false: No pruning.
1345  // true : Avoid the evaluation of specific transform depths using NN model.
1346  //
1347  // For allintra encode, this speed feature reduces instruction count
1348  // by 4.76%, 8.92% and 11.28% for speed 6, 7 and 8 with coding performance
1349  // change less than 0.32%. For AVIF image encode, this speed feature reduces
1350  // encode time by 4.65%, 9.16% and 10.45% for speed 6, 7 and 8 on a typical
1351  // image dataset with coding performance change less than 0.19%.
1352  bool prune_intra_tx_depths_using_nn;
1353 
1354  // Enable/disable early breakout during transform search of intra modes, by
1355  // using the minimum rd cost possible. By using this approach, the rd
1356  // evaluation of applicable transform blocks (in the current block) can be
1357  // avoided as
1358  // 1) best_rd evolves during the search in choose_tx_size_type_from_rd()
1359  // 2) appropriate ref_best_rd is passed in intra_block_yrd()
1360  //
1361  // For allintra encode, this speed feature reduces instruction count
1362  // by 1.11%, 1.08%, 1.02% and 0.93% for speed 3, 6, 7 and 8 with coding
1363  // performance change less than 0.02%. For AVIF image encode, this speed
1364  // feature reduces encode time by 0.93%, 1.46%, 1.07%, 0.84%, 0.99% and 0.73%
1365  // for speed 3, 4, 5, 6, 7 and 8 on a typical image dataset with coding
1366  // performance change less than 0.004%.
1367  bool use_rd_based_breakout_for_intra_tx_search;
1368 } TX_SPEED_FEATURES;
1369 
1370 typedef struct RD_CALC_SPEED_FEATURES {
1371  // Fast approximation of av1_model_rd_from_var_lapndz
1372  int simple_model_rd_from_var;
1373 
1374  // Perform faster distortion computation during the R-D evaluation by trying
1375  // to approximate the prediction error with transform coefficients (faster but
1376  // less accurate) rather than computing distortion in the pixel domain (slower
1377  // but more accurate). The following methods are used for distortion
1378  // computation:
1379  // Method 0: Always compute distortion in the pixel domain
1380  // Method 1: Based on block error, try using transform domain distortion for
1381  // tx_type search and compute distortion in pixel domain for final RD_STATS
1382  // Method 2: Based on block error, try to compute distortion in transform
1383  // domain
1384  // Methods 1 and 2 may fallback to computing distortion in the pixel domain in
1385  // case the block error is less than the threshold, which is controlled by the
1386  // speed feature tx_domain_dist_thres_level.
1387  //
1388  // The speed feature tx_domain_dist_level decides which of the above methods
1389  // needs to be used across different mode evaluation stages as described
1390  // below:
1391  // Eval type: Default Mode Winner
1392  // Level 0 : Method 0 Method 2 Method 0
1393  // Level 1 : Method 1 Method 2 Method 0
1394  // Level 2 : Method 2 Method 2 Method 0
1395  // Level 3 : Method 2 Method 2 Method 2
1396  int tx_domain_dist_level;
1397 
1398  // Transform domain distortion threshold level
1399  int tx_domain_dist_thres_level;
1400 
1401  // Trellis (dynamic programming) optimization of quantized values
1402  TRELLIS_OPT_TYPE optimize_coefficients;
1403 
1404  // Use hash table to store macroblock RD search results
1405  // to avoid repeated search on the same residue signal.
1406  int use_mb_rd_hash;
1407 
1408  // Flag used to control the extent of coeff R-D optimization
1409  int perform_coeff_opt;
1410 } RD_CALC_SPEED_FEATURES;
1411 
1412 typedef struct WINNER_MODE_SPEED_FEATURES {
1413  // Flag used to control the winner mode processing for better R-D optimization
1414  // of quantized coeffs
1415  int enable_winner_mode_for_coeff_opt;
1416 
1417  // Flag used to control the winner mode processing for transform size
1418  // search method
1419  int enable_winner_mode_for_tx_size_srch;
1420 
1421  // Control transform size search level
1422  // Eval type: Default Mode Winner
1423  // Level 0 : FULL RD LARGEST ALL FULL RD
1424  // Level 1 : FAST RD LARGEST ALL FULL RD
1425  // Level 2 : LARGEST ALL LARGEST ALL FULL RD
1426  // Level 3 : LARGEST ALL LARGEST ALL LARGEST ALL
1427  int tx_size_search_level;
1428 
1429  // Flag used to control the winner mode processing for use transform
1430  // domain distortion
1431  int enable_winner_mode_for_use_tx_domain_dist;
1432 
1433  // Flag used to enable processing of multiple winner modes
1434  MULTI_WINNER_MODE_TYPE multi_winner_mode_type;
1435 
1436  // Motion mode for winner candidates:
1437  // 0: speed feature OFF
1438  // 1 / 2 : Use configured number of winner candidates
1439  int motion_mode_for_winner_cand;
1440 
1441  // Controls the prediction of transform skip block or DC only block.
1442  //
1443  // Different speed feature values (0 to 3) decide the aggressiveness of
1444  // prediction (refer to predict_dc_levels[][] in speed_features.c) to be used
1445  // during different mode evaluation stages.
1446  int dc_blk_pred_level;
1447 
1448  // If on, disables interpolation filter search in handle_inter_mode loop, and
1449  // performs it during winner mode processing by \ref
1450  // tx_search_best_inter_candidates.
1451  int winner_mode_ifs;
1452 
1453  // Controls the disabling of winner mode processing. Speed feature levels
1454  // are ordered in increasing aggressiveness of pruning. The method considered
1455  // for disabling, depends on the sf level value and it is described as below.
1456  // 0: Do not disable
1457  // 1: Disable for blocks with low source variance.
1458  // 2: Disable for blocks which turn out to be transform skip (skipped based on
1459  // eob) during MODE_EVAL stage except NEWMV mode.
1460  // 3: Disable for blocks which turn out to be transform skip during MODE_EVAL
1461  // stage except NEWMV mode. For high quantizers, prune conservatively based on
1462  // transform skip (skipped based on eob) except for NEWMV mode.
1463  // 4: Disable for blocks which turn out to be transform skip during MODE_EVAL
1464  // stage.
1465  int prune_winner_mode_eval_level;
1466 } WINNER_MODE_SPEED_FEATURES;
1467 
1468 typedef struct LOOP_FILTER_SPEED_FEATURES {
1469  // This feature controls how the loop filter level is determined.
1470  LPF_PICK_METHOD lpf_pick;
1471 
1472  // Skip some final iterations in the determination of the best loop filter
1473  // level.
1474  int use_coarse_filter_level_search;
1475 
1476  // Control how the CDEF strength is determined.
1477  CDEF_PICK_METHOD cdef_pick_method;
1478 
1479  // Decoder side speed feature to add penalty for use of dual-sgr filters.
1480  // Takes values 0 - 10, 0 indicating no penalty and each additional level
1481  // adding a penalty of 1%
1482  int dual_sgr_penalty_level;
1483 
1484  // prune sgr ep using binary search like mechanism
1485  int enable_sgr_ep_pruning;
1486 
1487  // Disable loop restoration for Chroma plane
1488  int disable_loop_restoration_chroma;
1489 
1490  // Disable loop restoration for luma plane
1491  int disable_loop_restoration_luma;
1492 
1493  // Range of loop restoration unit sizes to search
1494  // The minimum size is clamped against the superblock size in
1495  // av1_pick_filter_restoration, so that the code which sets this value does
1496  // not need to know the superblock size ahead of time.
1497  int min_lr_unit_size;
1498  int max_lr_unit_size;
1499 
1500  // Prune RESTORE_WIENER evaluation based on source variance
1501  // 0 : no pruning
1502  // 1 : conservative pruning
1503  // 2 : aggressive pruning
1504  int prune_wiener_based_on_src_var;
1505 
1506  // Prune self-guided loop restoration based on wiener search results
1507  // 0 : no pruning
1508  // 1 : pruning based on rdcost ratio of RESTORE_WIENER and RESTORE_NONE
1509  // 2 : pruning based on winner restoration type among RESTORE_WIENER and
1510  // RESTORE_NONE
1511  int prune_sgr_based_on_wiener;
1512 
1513  // Reduce the wiener filter win size for luma
1514  int reduce_wiener_window_size;
1515 
1516  // Flag to disable Wiener Loop restoration filter.
1517  bool disable_wiener_filter;
1518 
1519  // Flag to disable Self-guided Loop restoration filter.
1520  bool disable_sgr_filter;
1521 
1522  // Disable the refinement search around the wiener filter coefficients.
1523  bool disable_wiener_coeff_refine_search;
1524 
1525  // Whether to downsample the rows in computation of wiener stats.
1526  int use_downsampled_wiener_stats;
1527 } LOOP_FILTER_SPEED_FEATURES;
1528 
1529 typedef struct REAL_TIME_SPEED_FEATURES {
1530  // check intra prediction for non-RD mode.
1531  int check_intra_pred_nonrd;
1532 
1533  // Skip checking intra prediction.
1534  // 0 - don't skip
1535  // 1 - skip if TX is skipped and best mode is not NEWMV
1536  // 2 - skip if TX is skipped
1537  // Skipping aggressiveness increases from level 1 to 2.
1538  int skip_intra_pred;
1539 
1540  // Estimate motion before calculating variance in variance-based partition
1541  // 0 - Only use zero MV
1542  // 1 - perform coarse ME
1543  // 2 - perform coarse ME, and also use neighbours' MVs
1544  // 3 - use neighbours' MVs without performing coarse ME
1545  int estimate_motion_for_var_based_partition;
1546 
1547  // For nonrd_use_partition: mode of extra check of leaf partition
1548  // 0 - don't check merge
1549  // 1 - always check merge
1550  // 2 - check merge and prune checking final split
1551  // 3 - check merge and prune checking final split based on bsize and qindex
1552  int nonrd_check_partition_merge_mode;
1553 
1554  // For nonrd_use_partition: check of leaf partition extra split
1555  int nonrd_check_partition_split;
1556 
1557  // Implements various heuristics to skip searching modes
1558  // The heuristics selected are based on flags
1559  // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
1560  unsigned int mode_search_skip_flags;
1561 
1562  // For nonrd: Reduces ref frame search.
1563  // 0 - low level of search prune in non last frames
1564  // 1 - pruned search in non last frames
1565  // 2 - more pruned search in non last frames
1566  int nonrd_prune_ref_frame_search;
1567 
1568  // This flag controls the use of non-RD mode decision.
1569  int use_nonrd_pick_mode;
1570 
1571  // Use ALTREF frame in non-RD mode decision.
1572  int use_nonrd_altref_frame;
1573 
1574  // Use compound reference for non-RD mode.
1575  int use_comp_ref_nonrd;
1576 
1577  // Reference frames for compound prediction for nonrd pickmode:
1578  // LAST_GOLDEN (0), LAST_LAST2 (1), or LAST_ALTREF (2).
1579  int ref_frame_comp_nonrd[3];
1580 
1581  // use reduced ref set for real-time mode
1582  int use_real_time_ref_set;
1583 
1584  // Skip a number of expensive mode evaluations for blocks with very low
1585  // temporal variance.
1586  int short_circuit_low_temp_var;
1587 
1588  // Reuse inter prediction in fast non-rd mode.
1589  int reuse_inter_pred_nonrd;
1590 
1591  // Number of best inter modes to search transform. INT_MAX - search all.
1592  int num_inter_modes_for_tx_search;
1593 
1594  // Use interpolation filter search in non-RD mode decision.
1595  int use_nonrd_filter_search;
1596 
1597  // Use simplified RD model for interpolation search and Intra
1598  int use_simple_rd_model;
1599 
1600  // For nonrd mode: use hybrid intra mode search for intra only frames based on
1601  // block properties.
1602  // 0 : use nonrd pick intra for all blocks
1603  // 1 : use rd for bsize < 16x16, nonrd otherwise
1604  // 2 : use rd for bsize < 16x16 and src var >= 101, nonrd otherwise
1605  int hybrid_intra_pickmode;
1606 
1607  // Compute variance/sse on source difference, prior to encoding superblock.
1608  int source_metrics_sb_nonrd;
1609 
1610  // Flag to indicate process for handling overshoot on slide/scene change,
1611  // for real-time CBR mode.
1612  OVERSHOOT_DETECTION_CBR overshoot_detection_cbr;
1613 
1614  // Check for scene/content change detection on every frame before encoding.
1615  int check_scene_detection;
1616 
1617  // For nonrd mode: Prefer larger partition blks in variance based partitioning
1618  // 0: disabled, 1-3: increasing aggressiveness
1619  int prefer_large_partition_blocks;
1620 
1621  // uses results of temporal noise estimate
1622  int use_temporal_noise_estimate;
1623 
1624  // Parameter indicating initial search window to be used in full-pixel search
1625  // for nonrd_pickmode. Range [0, MAX_MVSEARCH_STEPS - 1]. Lower value
1626  // indicates larger window. If set to 0, step_param is set based on internal
1627  // logic in set_mv_search_params().
1628  int fullpel_search_step_param;
1629 
1630  // Bit mask to enable or disable intra modes for each prediction block size
1631  // separately, for nonrd_pickmode. Currently, the sf is not respected when
1632  // 'force_intra_check' is true in 'av1_estimate_intra_mode()' function. Also,
1633  // H and V pred modes allowed through this sf can be further pruned when
1634  //'prune_hv_pred_modes_using_src_sad' sf is true.
1635  int intra_y_mode_bsize_mask_nrd[BLOCK_SIZES];
1636 
1637  // Prune H and V intra predition modes evalution in inter frame.
1638  // The sf does not have any impact.
1639  // i. when frame_source_sad is 1.1 times greater than avg_source_sad
1640  // ii. when cyclic_refresh_segment_id_boosted is enabled
1641  // iii. when SB level source sad is greater than kMedSad
1642  // iv. when color sensitivity is non zero for both the chroma channels
1643  bool prune_hv_pred_modes_using_src_sad;
1644 
1645  // Skips mode checks more aggressively in nonRD mode
1646  int nonrd_aggressive_skip;
1647 
1648  // Skip cdef on 64x64 blocks/
1649  // 0: disabled
1650  // 1: skip when NEWMV or INTRA is not picked or color sensitivity is off.
1651  // When color sensitivity is on for a superblock, all 64x64 blocks within
1652  // will not skip.
1653  // 2: more aggressive mode where skip is done for all frames where
1654  // rc->high_source_sad = 0 (non slide-changes), and color sensitivity off.
1655  int skip_cdef_sb;
1656 
1657  // Forces larger partition blocks in variance based partitioning for intra
1658  // frames
1659  int force_large_partition_blocks_intra;
1660 
1661  // Skip evaluation of no split in tx size selection for merge partition
1662  int skip_tx_no_split_var_based_partition;
1663 
1664  // Intermediate termination of newMV mode evaluation based on so far best mode
1665  // sse
1666  int skip_newmv_mode_based_on_sse;
1667 
1668  // Define gf length multiplier.
1669  // Level 0: use large multiplier, level 1: use medium multiplier.
1670  int gf_length_lvl;
1671 
1672  // Prune inter modes with golden frame as reference for NEARMV and NEWMV modes
1673  int prune_inter_modes_with_golden_ref;
1674 
1675  // Prune inter modes w.r.t golden or alt-ref frame based on sad
1676  int prune_inter_modes_wrt_gf_arf_based_on_sad;
1677 
1678  // Prune inter mode search in rd path based on current block's temporal
1679  // variance wrt LAST reference.
1680  int prune_inter_modes_using_temp_var;
1681 
1682  // Reduce MV precision to halfpel for higher int MV value & frame-level motion
1683  // 0: disabled
1684  // 1-2: Reduce precision to halfpel, fullpel based on conservative
1685  // thresholds, aggressiveness increases with increase in level
1686  // 3: Reduce precision to halfpel using more aggressive thresholds
1687  int reduce_mv_pel_precision_highmotion;
1688 
1689  // Reduce MV precision for low complexity blocks
1690  // 0: disabled
1691  // 1: Reduce the mv resolution for zero mv if the variance is low
1692  // 2: Switch to halfpel, fullpel based on low block spatial-temporal
1693  // complexity.
1694  int reduce_mv_pel_precision_lowcomplex;
1695 
1696  // Prune intra mode evaluation in inter frames based on mv range.
1697  BLOCK_SIZE prune_intra_mode_based_on_mv_range;
1698  // The number of times to left shift the splitting thresholds in variance
1699  // based partitioning. The minimum values should be 7 to avoid left shifting
1700  // by a negative number.
1701  int var_part_split_threshold_shift;
1702 
1703  // Qindex based variance partition threshold index, which determines
1704  // the aggressiveness of partition pruning
1705  // 0: disabled for speeds 9,10
1706  // 1,2: (rd-path) lowers qindex thresholds conditionally (for low SAD sb)
1707  // 3,4: (non-rd path) uses pre-tuned qindex thresholds
1708  int var_part_based_on_qidx;
1709 
1710  // Enable GF refresh based on Q value.
1711  int gf_refresh_based_on_qp;
1712 
1713  // Temporal filtering
1714  // The value can be 1 or 2, which indicates the threshold to use.
1715  // Must be off for lossless mode.
1716  int use_rtc_tf;
1717 
1718  // Prune the use of the identity transform in nonrd_pickmode,
1719  // used for screen content mode: only for smaller blocks
1720  // and higher spatial variance, and when skip_txfm is not
1721  // already set.
1722  int prune_idtx_nonrd;
1723 
1724  // Prune the use of paletter mode in nonrd pickmode.
1725  int prune_palette_nonrd;
1726 
1727  // Force to only use dct for palette search in nonrd pickmode.
1728  int dct_only_palette_nonrd;
1729 
1730  // Skip loopfilter, for static content after slide change
1731  // or key frame, once quality has ramped up.
1732  // 0: disabled
1733  // 1: skip only after quality is ramped up.
1734  // 2: aggrssive mode, where skip is done for all frames that
1735  // where rc->high_source_sad = 0 (no slide-changes).
1736  int skip_lf_screen;
1737 
1738  // For nonrd: early exit out of variance partition that sets the
1739  // block size to superblock size, and sets mode to zeromv-last skip.
1740  // 0: disabled
1741  // 1: zeromv-skip is enabled at SB level only
1742  // 2: zeromv-skip is enabled at SB level and coding block level
1743  int part_early_exit_zeromv;
1744 
1745  // Early terminate inter mode search based on sse in non-rd path.
1746  INTER_SEARCH_EARLY_TERM_IDX sse_early_term_inter_search;
1747 
1748  // SAD based adaptive altref selection
1749  int sad_based_adp_altref_lag;
1750 
1751  // Enable/disable partition direct merging.
1752  int partition_direct_merging;
1753 
1754  // Level of aggressiveness for obtaining tx size based on qstep
1755  int tx_size_level_based_on_qstep;
1756 
1757  // Avoid the partitioning of a 16x16 block in variance based partitioning
1758  // (VBP) by making use of minimum and maximum sub-block variances.
1759  // For allintra encode, this speed feature reduces instruction count by 5.39%
1760  // for speed 9 on a typical video dataset with coding performance gain
1761  // of 1.44%.
1762  // For AVIF image encode, this speed feature reduces encode time
1763  // by 8.44% for speed 9 on a typical image dataset with coding performance
1764  // gain of 0.78%.
1765  bool vbp_prune_16x16_split_using_min_max_sub_blk_var;
1766 
1767  // A qindex threshold that determines whether to use qindex based CDEF filter
1768  // strength estimation for screen content types. The strength estimation model
1769  // used for screen contents prefers to allow cdef filtering for more frames.
1770  // This sf is used to limit the frames which go through cdef filtering and
1771  // following explains the setting of the same.
1772  // MAXQ (255): This disables the usage of this sf. Here, frame does not use a
1773  // screen content model thus reduces the number of frames that go through cdef
1774  // filtering.
1775  // MINQ (0): Frames always use screen content model thus increasing the number
1776  // of frames that go through cdef filtering.
1777  // This speed feature has a substantial gain on coding metrics, with moderate
1778  // increase encoding time. Select threshold based on speed vs quality
1779  // trade-off.
1780  int screen_content_cdef_filter_qindex_thresh;
1781 
1782  // Prune compound mode if its variance is higher than the variance of single
1783  // modes.
1784  bool prune_compoundmode_with_singlecompound_var;
1785 
1786  // Allow mode cost update at frame level every couple frames. This
1787  // overrides the command line setting --mode-cost-upd-freq=3 (never update
1788  // except on key frame and first delta).
1789  bool frame_level_mode_cost_update;
1790 
1791  // Prune H_PRED during intra mode evaluation in the nonrd path based on best
1792  // mode so far.
1793  //
1794  // For allintra encode, this speed feature reduces instruction count by 1.10%
1795  // for speed 9 with coding performance change less than 0.04%.
1796  // For AVIF image encode, this speed feature reduces encode time by 1.03% for
1797  // speed 9 on a typical image dataset with coding performance change less than
1798  // 0.08%.
1799  bool prune_h_pred_using_best_mode_so_far;
1800 
1801  // Enable pruning of intra mode evaluations in nonrd path based on source
1802  // variance and best mode so far. The pruning logic is enabled only if the
1803  // mode is not a winner mode of both the neighboring blocks (left/top).
1804  //
1805  // For allintra encode, this speed feature reduces instruction count by 3.96%
1806  // for speed 9 with coding performance change less than 0.38%.
1807  // For AVIF image encode, this speed feature reduces encode time by 3.46% for
1808  // speed 9 on a typical image dataset with coding performance change less than
1809  // -0.06%.
1810  bool enable_intra_mode_pruning_using_neighbors;
1811 
1812  // Prune intra mode evaluations in nonrd path based on best sad so far.
1813  //
1814  // For allintra encode, this speed feature reduces instruction count by 3.05%
1815  // for speed 9 with coding performance change less than 0.24%.
1816  // For AVIF image encode, this speed feature reduces encode time by 1.87% for
1817  // speed 9 on a typical image dataset with coding performance change less than
1818  // 0.16%.
1819  bool prune_intra_mode_using_best_sad_so_far;
1820 
1821  // If compound is enabled, and the current block size is \geq BLOCK_16X16,
1822  // limit the compound modes to GLOBAL_GLOBALMV. This does not apply to the
1823  // base layer of svc.
1824  bool check_only_zero_zeromv_on_large_blocks;
1825 
1826  // Allow for disabling cdf update for non reference frames in svc mode.
1827  bool disable_cdf_update_non_reference_frame;
1828 
1829  // Prune compound modes if the single modes variances do not perform well.
1830  bool prune_compoundmode_with_singlemode_var;
1831 
1832  // Skip searching all compound mode if the variance of single_mode residue is
1833  // sufficiently low.
1834  bool skip_compound_based_on_var;
1835 
1836  // Sets force_zeromv_skip based on the source sad available. Aggressiveness
1837  // increases with increase in the level set for speed feature.
1838  // 0: No setting
1839  // 1: If source sad is kZeroSad
1840  // 2: If source sad <= kVeryLowSad
1841  int set_zeromv_skip_based_on_source_sad;
1842 
1843  // Downgrades the block-level subpel motion search to
1844  // av1_find_best_sub_pixel_tree_pruned_more for higher QP and when fullpel
1845  // search performed well, zeromv has low sad or low source_var
1846  bool use_adaptive_subpel_search;
1847 
1848  // A flag used in RTC case to control frame_refs_short_signaling. Note that
1849  // the final decision is made in check_frame_refs_short_signaling(). The flag
1850  // can only be turned on when res < 360p and speed >= 9, in which case only
1851  // LAST and GOLDEN ref frames are used now.
1852  bool enable_ref_short_signaling;
1853 
1854  // A flag that controls if we check or bypass GLOBALMV in rtc single ref frame
1855  // case.
1856  bool check_globalmv_on_single_ref;
1857 
1858  // Allows for increasing the color_threshold for palette prediction.
1859  // This generally leads to better coding efficiency but with some speed loss.
1860  // Only used for screen content and for nonrd_pickmode.
1861  bool increase_color_thresh_palette;
1862 } REAL_TIME_SPEED_FEATURES;
1863 
1869 typedef struct SPEED_FEATURES {
1874 
1879 
1883  TPL_SPEED_FEATURES tpl_sf;
1884 
1888  GLOBAL_MOTION_SPEED_FEATURES gm_sf;
1889 
1893  PARTITION_SPEED_FEATURES part_sf;
1894 
1898  MV_SPEED_FEATURES mv_sf;
1899 
1903  INTER_MODE_SPEED_FEATURES inter_sf;
1904 
1908  INTERP_FILTER_SPEED_FEATURES interp_sf;
1909 
1913  INTRA_MODE_SPEED_FEATURES intra_sf;
1914 
1918  TX_SPEED_FEATURES tx_sf;
1919 
1923  RD_CALC_SPEED_FEATURES rd_sf;
1924 
1928  WINNER_MODE_SPEED_FEATURES winner_mode_sf;
1929 
1933  LOOP_FILTER_SPEED_FEATURES lpf_sf;
1934 
1938  REAL_TIME_SPEED_FEATURES rt_sf;
1942 struct AV1_COMP;
1943 
1957  int speed);
1958 
1971  int speed);
1984 
1985 #ifdef __cplusplus
1986 } // extern "C"
1987 #endif
1988 
1989 #endif // AOM_AV1_ENCODER_SPEED_FEATURES_H_
static int prune_zero_mv_with_sse(const aom_variance_fn_ptr_t *fn_ptr, const MACROBLOCK *x, BLOCK_SIZE bsize, const HandleInterModeArgs *args, int prune_zero_mv_with_sse)
Prunes ZeroMV Search Using Best NEWMV's SSE.
Definition: rdopt.c:2503
void av1_set_speed_features_framesize_independent(struct AV1_COMP *cpi, int speed)
Frame size independent speed vs quality trade off flags.
void av1_set_speed_features_qindex_dependent(struct AV1_COMP *cpi, int speed)
Q index dependent speed vs quality trade off flags.
void av1_set_speed_features_framesize_dependent(struct AV1_COMP *cpi, int speed)
Frame size dependent speed vs quality trade off flags.
INTERNAL_COST_UPDATE_TYPE
This enum decides internally how often to update the entropy costs.
Definition: speed_features.h:346
@ INTERNAL_COST_UPD_OFF
Definition: speed_features.h:347
@ INTERNAL_COST_UPD_SBROW_SET
Definition: speed_features.h:349
@ INTERNAL_COST_UPD_SBROW
Definition: speed_features.h:350
@ INTERNAL_COST_UPD_SB
Definition: speed_features.h:351
@ INTERNAL_COST_UPD_TILE
Definition: speed_features.h:348
INTER_SEARCH_EARLY_TERM_IDX
This enumeration defines inter search early termination index in non-rd path based on sse value.
Definition: speed_features.h:390
@ EARLY_TERM_INDICES
Definition: speed_features.h:401
@ EARLY_TERM_IDX_1
Definition: speed_features.h:393
@ EARLY_TERM_IDX_4
Definition: speed_features.h:399
@ EARLY_TERM_IDX_2
Definition: speed_features.h:395
@ EARLY_TERM_DISABLED
Definition: speed_features.h:391
@ EARLY_TERM_IDX_3
Definition: speed_features.h:397
SIMPLE_MOTION_SEARCH_PRUNE_LEVEL
This enumeration defines a variety of simple motion search based partition prune levels.
Definition: speed_features.h:358
@ QIDX_BASED_AGG_LVL1
Definition: speed_features.h:364
@ SIMPLE_AGG_LVL3
Definition: speed_features.h:363
@ TOTAL_SIMPLE_AGG_LVLS
Definition: speed_features.h:367
@ SIMPLE_AGG_LVL1
Definition: speed_features.h:361
@ SIMPLE_AGG_LVL0
Definition: speed_features.h:360
@ TOTAL_AGG_LVLS
Definition: speed_features.h:373
@ SIMPLE_AGG_LVL2
Definition: speed_features.h:362
@ TOTAL_QINDEX_BASED_AGG_LVLS
Definition: speed_features.h:369
struct SPEED_FEATURES SPEED_FEATURES
Top level speed vs quality trade off data struture.
PRUNE_MESH_SEARCH_LEVEL
This enumeration defines a variety of mesh search prune levels.
Definition: speed_features.h:380
@ PRUNE_MESH_SEARCH_LVL_1
Definition: speed_features.h:382
@ PRUNE_MESH_SEARCH_LVL_2
Definition: speed_features.h:383
@ PRUNE_MESH_SEARCH_DISABLED
Definition: speed_features.h:381
struct HIGH_LEVEL_SPEED_FEATURES HIGH_LEVEL_SPEED_FEATURES
Sequence/frame level speed vs quality features.
struct FIRST_PASS_SPEED_FEATURES FIRST_PASS_SPEED_FEATURES
CDEF_PICK_METHOD
This enumeration defines a variety of CDEF pick methods.
Definition: speed_features.h:164
@ CDEF_FAST_SEARCH_LVL2
Definition: speed_features.h:167
@ CDEF_FAST_SEARCH_LVL5
Definition: speed_features.h:171
@ CDEF_FAST_SEARCH_LVL1
Definition: speed_features.h:166
@ CDEF_FULL_SEARCH
Definition: speed_features.h:165
@ CDEF_PICK_FROM_Q
Definition: speed_features.h:172
@ CDEF_FAST_SEARCH_LVL4
Definition: speed_features.h:170
@ CDEF_FAST_SEARCH_LVL3
Definition: speed_features.h:168
Top level encoder structure.
Definition: encoder.h:2872
int speed
Definition: encoder.h:3094
Definition: speed_features.h:490
int disable_recon
Skips reconstruction by using source buffers for prediction.
Definition: speed_features.h:507
int reduce_mv_step_param
Reduces the mv search window. By default, the initial search window is around MIN(MIN(dims),...
Definition: speed_features.h:497
int skip_zeromv_motion_search
Skips the motion search centered on 0,0 mv.
Definition: speed_features.h:512
int skip_motion_search_threshold
Skips the motion search when the zero mv has small sse.
Definition: speed_features.h:502
Sequence/frame level speed vs quality features.
Definition: speed_features.h:407
int adjust_num_frames_for_arf_filtering
Definition: speed_features.h:462
int allow_sub_blk_me_in_tf
Definition: speed_features.h:484
int second_alt_ref_filtering
Definition: speed_features.h:452
int frame_parameter_update
Definition: speed_features.h:409
int weight_calc_level_in_tf
Definition: speed_features.h:476
MV_PREC_LOGIC high_precision_mv_usage
Definition: speed_features.h:428
int disable_extra_sc_testing
Definition: speed_features.h:447
int recode_tolerance
Definition: speed_features.h:420
SUPERRES_AUTO_SEARCH_TYPE superres_auto_search_type
Definition: speed_features.h:442
int static_segmentation
Definition: speed_features.h:437
int accurate_bit_estimate
Definition: speed_features.h:469
RECODE_LOOP_TYPE recode_loop
Definition: speed_features.h:414
Top level speed vs quality trade off data struture.
Definition: speed_features.h:1869
MV_SPEED_FEATURES mv_sf
Definition: speed_features.h:1898
TPL_SPEED_FEATURES tpl_sf
Definition: speed_features.h:1883
LOOP_FILTER_SPEED_FEATURES lpf_sf
Definition: speed_features.h:1933
TX_SPEED_FEATURES tx_sf
Definition: speed_features.h:1918
INTER_MODE_SPEED_FEATURES inter_sf
Definition: speed_features.h:1903
RD_CALC_SPEED_FEATURES rd_sf
Definition: speed_features.h:1923
PARTITION_SPEED_FEATURES part_sf
Definition: speed_features.h:1893
GLOBAL_MOTION_SPEED_FEATURES gm_sf
Definition: speed_features.h:1888
INTERP_FILTER_SPEED_FEATURES interp_sf
Definition: speed_features.h:1908
FIRST_PASS_SPEED_FEATURES fp_sf
Definition: speed_features.h:1878
INTRA_MODE_SPEED_FEATURES intra_sf
Definition: speed_features.h:1913
WINNER_MODE_SPEED_FEATURES winner_mode_sf
Definition: speed_features.h:1928
REAL_TIME_SPEED_FEATURES rt_sf
Definition: speed_features.h:1938
HIGH_LEVEL_SPEED_FEATURES hl_sf
Definition: speed_features.h:1873