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
21extern "C" {
22#endif
23
27#define MAX_MESH_STEP 4
28
29typedef struct MESH_PATTERN {
30 int range;
31 int interval;
32} MESH_PATTERN;
33
34enum {
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
46enum {
47 DIST_WTD_COMP_ENABLED,
48 DIST_WTD_COMP_SKIP_MV_SEARCH,
49 DIST_WTD_COMP_DISABLED,
50} UENUM1BYTE(DIST_WTD_COMP_FLAG);
51
52enum {
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
85enum {
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
99enum {
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
111enum {
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 */
125enum {
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
140enum {
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
147enum {
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);
175
177enum {
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
196enum {
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
210enum {
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
219enum {
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
236enum {
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
244enum {
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
257typedef 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
295enum {
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
311enum {
312 NOT_IN_USE,
313 DIRECT_PRED,
314 RELAXED_PRED,
315 ADAPT_PRED
316} UENUM1BYTE(MAX_PART_PRED_MODE);
317
318enum {
319 LAST_MV_DATA,
320 CURRENT_Q,
321 QTR_ONLY,
322} UENUM1BYTE(MV_PREC_LOGIC);
323
324enum {
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);
353
376
385
403
494
522
524typedef struct TPL_SPEED_FEATURES {
525 // GOP length adaptive decision.
526 // If set to 0, tpl model decides whether a shorter gf interval is better.
527 // If set to 1, tpl stats of ARFs from base layer, (base+1) layer and
528 // (base+2) layer decide whether a shorter gf interval is better.
529 // If set to 2, tpl stats of ARFs from base layer, (base+1) layer and GF boost
530 // decide whether a shorter gf interval is better.
531 // If set to 3, gop length adaptive decision is disabled.
532 int gop_length_decision_method;
533 // Prune the intra modes search by tpl.
534 // If set to 0, we will search all intra modes from DC_PRED to PAETH_PRED.
535 // If set to 1, we only search DC_PRED, V_PRED, and H_PRED.
536 int prune_intra_modes;
537 // This parameter controls which step in the n-step process we start at.
538 int reduce_first_step_size;
539 // Skip motion estimation based on the precision of center MVs and the
540 // difference between center MVs.
541 // If set to 0, motion estimation is skipped for duplicate center MVs
542 // (default). If set to 1, motion estimation is skipped for duplicate
543 // full-pixel center MVs. If set to 2, motion estimation is skipped if the
544 // difference between center MVs is less than the threshold.
545 int skip_alike_starting_mv;
546
547 // When to stop subpel search.
548 SUBPEL_FORCE_STOP subpel_force_stop;
549
550 // Which search method to use.
551 SEARCH_METHODS search_method;
552
553 // Prune starting mvs in TPL based on sad scores.
554 int prune_starting_mv;
555
556 // Prune reference frames in TPL.
557 int prune_ref_frames_in_tpl;
558
559 // Support compound predictions.
560 int allow_compound_pred;
561
562 // Calculate rate and distortion based on Y plane only.
563 int use_y_only_rate_distortion;
564
565 // Use SAD instead of SATD during intra/inter mode search.
566 // If set to 0, use SATD always.
567 // If set to 1, use SAD during intra/inter mode search for frames in the
568 // higher temporal layers of the hierarchical prediction structure.
569 // If set to 2, use SAD during intra/inter mode search for all frames.
570 // This sf is disabled for the first GF group of the key-frame interval,
571 // i.e., SATD is used during intra/inter mode search of the first GF group.
572 int use_sad_for_mode_decision;
573
574 // Skip tpl processing for frames of type LF_UPDATE.
575 // This sf is disabled for the first GF group of the key-frame interval.
576 int reduce_num_frames;
577} TPL_SPEED_FEATURES;
578
579typedef struct GLOBAL_MOTION_SPEED_FEATURES {
580 GM_SEARCH_TYPE gm_search_type;
581
582 // During global motion estimation, prune remaining reference frames in a
583 // given direction(past/future), if the evaluated ref_frame in that direction
584 // yields gm_type as INVALID/TRANSLATION/IDENTITY
585 int prune_ref_frame_for_gm_search;
586
587 // When the current GM type is set to ZEROMV, prune ZEROMV if its performance
588 // is worse than NEWMV under SSE metric.
589 // 0 : no pruning
590 // 1 : conservative pruning
591 // 2 : aggressive pruning
593
594 // Disable global motion estimation based on stats of previous frames in the
595 // GF group
596 int disable_gm_search_based_on_stats;
597
598 // Downsampling pyramid level to use for global motion estimation
599 int downsample_level;
600
601 // Number of refinement steps to apply after initial model generation
602 int num_refinement_steps;
603
604 // Error advantage threshold level used to determine whether global motion
605 // compensation should be enabled
606 int gm_erroradv_tr_level;
607} GLOBAL_MOTION_SPEED_FEATURES;
608
609typedef struct PARTITION_SPEED_FEATURES {
610 PARTITION_SEARCH_TYPE partition_search_type;
611
612 // Used if partition_search_type = FIXED_PARTITION
613 BLOCK_SIZE fixed_partition_size;
614
615 // Prune extended partition types search based on the current best partition
616 // and the combined rdcost of the subblocks estimated from previous
617 // partitions. Can take values 0 - 2, 0 referring to no pruning, and 1 - 2
618 // increasing aggressiveness of pruning in order.
619 int prune_ext_partition_types_search_level;
620
621 // Prune part4 based on block size
622 int prune_part4_search;
623
624 // Use a ML model to prune rectangular, ab and 4-way horz
625 // and vert partitions
626 int ml_prune_partition;
627
628 // Use a ML model to adaptively terminate partition search after trying
629 // PARTITION_SPLIT. Can take values 0 - 2, 0 meaning not being enabled, and
630 // 1 - 2 increasing aggressiveness in order.
631 int ml_early_term_after_part_split_level;
632
633 // Skip rectangular partition test when partition type none gives better
634 // rd than partition type split. Can take values 0 - 2, 0 referring to no
635 // skipping, and 1 - 2 increasing aggressiveness of skipping in order.
636 int less_rectangular_check_level;
637
638 // Use square partition only beyond this block size.
639 BLOCK_SIZE use_square_partition_only_threshold;
640
641 // Sets max square partition levels for this superblock based on
642 // motion vector and prediction error distribution produced from 16x16
643 // simple motion search
644 MAX_PART_PRED_MODE auto_max_partition_based_on_simple_motion;
645
646 // Min and max square partition size we enable (block_size) as per auto
647 // min max, but also used by adjust partitioning, and pick_partitioning.
648 BLOCK_SIZE default_min_partition_size;
649 BLOCK_SIZE default_max_partition_size;
650
651 // Sets level of adjustment of variance-based partitioning during
652 // rd_use_partition 0 - no partition adjustment, 1 - try to merge partitions
653 // for small blocks and high QP, 2 - try to merge partitions, 3 - try to merge
654 // and split leaf partitions and 0 - 3 decreasing aggressiveness in order.
655 int adjust_var_based_rd_partitioning;
656
657 // Partition search early breakout thresholds.
658 int64_t partition_search_breakout_dist_thr;
659 int partition_search_breakout_rate_thr;
660
661 // Thresholds for ML based partition search breakout.
662 int ml_partition_search_breakout_thresh[PARTITION_BLOCK_SIZES];
663
664 // Aggressiveness levels for pruning split and rectangular partitions based on
665 // simple_motion_search. SIMPLE_AGG_LVL0 to SIMPLE_AGG_LVL3 correspond to
666 // simple motion search based pruning. QIDX_BASED_AGG_LVL1 corresponds to
667 // qindex based and simple motion search based pruning.
668 int simple_motion_search_prune_agg;
669
670 // Perform simple_motion_search on each possible subblock and use it to prune
671 // PARTITION_HORZ and PARTITION_VERT.
672 int simple_motion_search_prune_rect;
673
674 // Perform simple motion search before none_partition to decide if we
675 // want to remove all partitions other than PARTITION_SPLIT. If set to 0, this
676 // model is disabled. If set to 1, the model attempts to perform
677 // PARTITION_SPLIT only. If set to 2, the model also attempts to prune
678 // PARTITION_SPLIT.
679 int simple_motion_search_split;
680
681 // Use features from simple_motion_search to terminate prediction block
682 // partition after PARTITION_NONE
683 int simple_motion_search_early_term_none;
684
685 // Controls whether to reduce the number of motion search steps. If this is 0,
686 // then simple_motion_search has the same number of steps as
687 // single_motion_search (assuming no other speed features). Otherwise, reduce
688 // the number of steps by the value contained in this variable.
689 int simple_motion_search_reduce_search_steps;
690
691 // This variable controls the maximum block size where intra blocks can be
692 // used in inter frames.
693 // TODO(aconverse): Fold this into one of the other many mode skips
694 BLOCK_SIZE max_intra_bsize;
695
696 // Use CNN with luma pixels on source frame on each of the 64x64 subblock to
697 // perform partition pruning in intra frames.
698 // 0: No Pruning
699 // 1: Prune split and rectangular partitions only
700 // 2: Prune none, split and rectangular partitions
701 int intra_cnn_based_part_prune_level;
702
703 // Disable extended partition search if the current bsize is greater than the
704 // threshold. Must be a square block size BLOCK_8X8 or higher.
705 BLOCK_SIZE ext_partition_eval_thresh;
706
707 // Use best partition decision so far to tune 'ext_partition_eval_thresh'
708 int ext_part_eval_based_on_cur_best;
709
710 // Disable rectangular partitions for larger block sizes.
711 int rect_partition_eval_thresh;
712
713 // Prune extended partition search based on whether the split/rect partitions
714 // provided an improvement in the previous search.
715 // 0 : no pruning
716 // 1 : prune 1:4 partition search using winner info from split partitions
717 // 2 : prune 1:4 and AB partition search using split and HORZ/VERT info
718 int prune_ext_part_using_split_info;
719
720 // Prunt rectangular, AB and 4-way partition based on q index and block size
721 // 0 : no pruning
722 // 1 : prune sub_8x8 at very low quantizers
723 // 2 : prune all block size based on qindex
724 int prune_rectangular_split_based_on_qidx;
725
726 // Prune rectangular partitions based on 4x4 sub-block variance
727 // false : no pruning
728 // true : prune rectangular partitions based on 4x4 sub-block variance
729 // deviation
730 //
731 // For allintra encode, this speed feature reduces instruction count by 6.4%
732 // for speed=6 with coding performance change less than 0.24%. For AVIF image
733 // encode, this speed feature reduces encode time by 8.14% for speed 6 on a
734 // typical image dataset with coding performance change less than 0.16%. This
735 // speed feature is not applicable to speed >= 7.
736 bool prune_rect_part_using_4x4_var_deviation;
737
738 // Prune rectangular partitions based on prediction mode chosen by NONE
739 // partition.
740 // false : no pruning
741 // true : prunes rectangular partition as described below
742 // If prediction mode chosen by NONE partition is
743 // DC_PRED or SMOOTH_PRED: Prunes both horizontal and vertical partitions if
744 // at least one of the left and top neighbor blocks is larger than the
745 // current block.
746 // Directional Mode: Prunes either of the horizontal and vertical partition
747 // based on center angle of the prediction mode chosen by NONE partition. For
748 // example, vertical partition is pruned if center angle of the prediction
749 // mode chosen by NONE partition is close to 180 degrees (i.e. horizontal
750 // direction) and vice versa.
751 // For allintra encode, this speed feature reduces instruction count by 5.1%
752 // for speed=6 with coding performance change less than 0.22%. For AVIF image
753 // encode, this speed feature reduces encode time by 4.44% for speed 6 on a
754 // typical image dataset with coding performance change less than 0.15%.
755 // For speed >= 7, variance-based logic is used to determine the partition
756 // structure instead of recursive partition search. Therefore, this speed
757 // feature is not applicable in such cases.
758 bool prune_rect_part_using_none_pred_mode;
759
760 // Terminate partition search for child partition,
761 // when NONE and SPLIT partition rd_costs are INT64_MAX.
762 int early_term_after_none_split;
763
764 // Level used to adjust threshold for av1_ml_predict_breakout(). At lower
765 // levels, more conservative threshold is used, and value of 0 indicates
766 // av1_ml_predict_breakout() is disabled. Value of 3 corresponds to default
767 // case with no adjustment to lbd thresholds.
768 int ml_predict_breakout_level;
769
770 // Prune sub_8x8 (BLOCK_4X4, BLOCK_4X8 and BLOCK_8X4) partitions.
771 // 0 : no pruning
772 // 1 : pruning based on neighbour block information
773 // 2 : prune always
774 int prune_sub_8x8_partition_level;
775
776 // Prune rectangular split based on simple motion search split/no_split score.
777 // 0: disable pruning, 1: enable pruning
778 int simple_motion_search_rect_split;
779
780 // The current encoder adopts a DFS search for block partitions.
781 // Therefore the mode selection and associated rdcost is ready for smaller
782 // blocks before the mode selection for some partition types.
783 // AB partition could use previous rd information and skip mode search.
784 // An example is:
785 //
786 // current block
787 // +---+---+
788 // | |
789 // + +
790 // | |
791 // +-------+
792 //
793 // SPLIT partition has been searched first before trying HORZ_A
794 // +---+---+
795 // | R | R |
796 // +---+---+
797 // | R | R |
798 // +---+---+
799 //
800 // HORZ_A
801 // +---+---+
802 // | | |
803 // +---+---+
804 // | |
805 // +-------+
806 //
807 // With this speed feature, the top two sub blocks can directly use rdcost
808 // searched in split partition, and the mode info is also copied from
809 // saved info. Similarly, the bottom rectangular block can also use
810 // the available information from previous rectangular search.
811 int reuse_prev_rd_results_for_part_ab;
812
813 // Reuse the best prediction modes found in PARTITION_SPLIT and PARTITION_RECT
814 // when encoding PARTITION_AB.
815 int reuse_best_prediction_for_part_ab;
816
817 // The current partition search records the best rdcost so far and uses it
818 // in mode search and transform search to early skip when some criteria is
819 // met. For example, when the current rdcost is larger than the best rdcost,
820 // or the model rdcost is larger than the best rdcost times some thresholds.
821 // By default, this feature is turned on to speed up the encoder partition
822 // search.
823 // If disabling it, at speed 0, 30 frames, we could get
824 // about -0.25% quality gain (psnr, ssim, vmaf), with about 13% slowdown.
825 int use_best_rd_for_pruning;
826
827 // Skip evaluation of non-square partitions based on the corresponding NONE
828 // partition.
829 // 0: no pruning
830 // 1: prune extended partitions if NONE is skippable
831 // 2: on top of 1, prune rectangular partitions if NONE is inter, not a newmv
832 // mode and skippable
833 int skip_non_sq_part_based_on_none;
834
835 // Disables 8x8 and below partitions for low quantizers.
836 int disable_8x8_part_based_on_qidx;
837} PARTITION_SPEED_FEATURES;
838
839typedef struct MV_SPEED_FEATURES {
840 // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
841 SEARCH_METHODS search_method;
842
843 // Enable the use of faster, less accurate mv search method
844 // 0: disable, 1: if bsize >= BLOCK_32X32, 2: based on bsize, SAD and qp
845 // TODO(chiyotsai@google.com): Take the clip's resolution and mv activity into
846 // account.
847 int use_bsize_dependent_search_method;
848
849 // If this is set to 1, we limit the motion search range to 2 times the
850 // largest motion vector found in the last frame.
851 int auto_mv_step_size;
852
853 // Subpel_search_method can only be subpel_tree which does a subpixel
854 // logarithmic search that keeps stepping at 1/2 pixel units until
855 // you stop getting a gain, and then goes on to 1/4 and repeats
856 // the same process. Along the way it skips many diagonals.
857 SUBPEL_SEARCH_METHOD subpel_search_method;
858
859 // Maximum number of steps in logarithmic subpel search before giving up.
860 int subpel_iters_per_step;
861
862 // When to stop subpel search.
863 SUBPEL_FORCE_STOP subpel_force_stop;
864
865 // When to stop subpel search in simple motion search.
866 SUBPEL_FORCE_STOP simple_motion_subpel_force_stop;
867
868 // If true, sub-pixel search uses the exact convolve function used for final
869 // encoding and decoding; otherwise, it uses bilinear interpolation.
870 SUBPEL_SEARCH_TYPE use_accurate_subpel_search;
871
872 // Threshold for allowing exhaustive motion search.
873 int exhaustive_searches_thresh;
874
875 // Pattern to be used for any exhaustive mesh searches (except intraBC ME).
876 MESH_PATTERN mesh_patterns[MAX_MESH_STEP];
877
878 // Pattern to be used for exhaustive mesh searches of intraBC ME.
879 MESH_PATTERN intrabc_mesh_patterns[MAX_MESH_STEP];
880
881 // Reduce single motion search range based on MV result of prior ref_mv_idx.
882 int reduce_search_range;
883
884 // Prune mesh search.
885 PRUNE_MESH_SEARCH_LEVEL prune_mesh_search;
886
887 // Use the rd cost around the best FULLPEL_MV to speed up subpel search
888 int use_fullpel_costlist;
889
890 // Set the full pixel search level of obmc
891 // 0: obmc_full_pixel_diamond
892 // 1: obmc_refining_search_sad (faster)
893 int obmc_full_pixel_search_level;
894
895 // Accurate full pixel motion search based on TPL stats.
896 int full_pixel_search_level;
897
898 // Allow intrabc motion search
899 int use_intrabc;
900
901 // Whether to downsample the rows in sad calculation during motion search.
902 // This is only active when there are at least 16 rows. When this sf is
903 // active, if there is a large discrepancy in the SAD values for the final
904 // motion vector between skipping vs not skipping, motion search is redone
905 // with skip row features off.
906 // 0: Disabled (do not downsample rows)
907 // 1: Skip SAD calculation of odd rows if the SAD deviation of the even and
908 // odd rows for the starting MV is small. Redo motion search with sf off
909 // when SAD deviation is high for the final motion vector.
910 // 2: Skip SAD calculation of odd rows. SAD deviation is not tested for the
911 // start MV and tested only for the final MV.
912 int use_downsampled_sad;
913
914 // Enable/disable extensive joint motion search.
915 int disable_extensive_joint_motion_search;
916
917 // Enable second best mv check in joint mv search.
918 // 0: allow second MV (use rd cost as the metric)
919 // 1: use var as the metric
920 // 2: disable second MV
921 int disable_second_mv;
922
923 // Skips full pixel search based on start mv of prior ref_mv_idx.
924 // 0: Disabled
925 // 1: Skips the full pixel search upto 4 neighbor full-pel MV positions.
926 // 2: Skips the full pixel search upto 8 neighbor full-pel MV positions.
927 int skip_fullpel_search_using_startmv;
928
929 // Method to use for refining WARPED_CAUSAL motion vectors
930 // TODO(rachelbarker): Can this be unified with OBMC in some way?
931 WARP_SEARCH_METHOD warp_search_method;
932
933 // Maximum number of iterations in WARPED_CAUSAL refinement search
934 int warp_search_iters;
935} MV_SPEED_FEATURES;
936
937typedef struct INTER_MODE_SPEED_FEATURES {
938 // 2-pass inter mode model estimation where the preliminary pass skips
939 // transform search and uses a model to estimate rd, while the final pass
940 // computes the full transform search. Two types of models are supported:
941 // 0: not used
942 // 1: used with online dynamic rd model
943 // 2: used with static rd model
944 int inter_mode_rd_model_estimation;
945
946 // Bypass transform search based on skip rd at following stages
947 // i. Compound type mode search
948 // ii. Motion mode search (mode evaluation and winner motion mode stage)
949 // iii. Transform search for best inter candidates
950 int txfm_rd_gate_level[TX_SEARCH_CASES];
951
952 // Limit the inter mode tested in the RD loop
953 int reduce_inter_modes;
954
955 // This variable is used to cap the maximum number of times we skip testing a
956 // mode to be evaluated. A high value means we will be faster.
957 int adaptive_rd_thresh;
958
959 // Aggressively prune inter modes when best mode is skippable.
960 int prune_inter_modes_if_skippable;
961
962 // Drop less likely to be picked reference frames in the RD search.
963 // Has seven levels for now: 0, 1, 2, 3, 4, 5 and 6 where higher levels prune
964 // more aggressively than lower ones. (0 means no pruning).
965 int selective_ref_frame;
966
967 // Prune reference frames for rectangular partitions.
968 // 0 implies no pruning
969 // 1 implies prune for extended partition
970 // 2 implies prune horiz, vert and extended partition
971 int prune_ref_frame_for_rect_partitions;
972
973 // Prune inter modes w.r.t past reference frames
974 // 0 no pruning
975 // 1 prune inter modes w.r.t ALTREF2 and ALTREF reference frames
976 // 2 prune inter modes w.r.t BWDREF, ALTREF2 and ALTREF reference frames
977 int alt_ref_search_fp;
978
979 // Prune reference frames for single prediction modes based on temporal
980 // distance and pred MV SAD. Feasible values are 0, 1, 2. The feature is
981 // disabled for 0. An increasing value indicates more aggressive pruning
982 // threshold.
983 int prune_single_ref;
984
985 // Prune compound reference frames
986 // 0 no pruning
987 // 1 prune compound references which do not satisfy the two conditions:
988 // a) The references are at a nearest distance from the current frame in
989 // both past and future direction.
990 // b) The references have minimum pred_mv_sad in both past and future
991 // direction.
992 // 2 prune compound references except the one with nearest distance from the
993 // current frame in both past and future direction.
994 int prune_comp_ref_frames;
995
996 // Skip the current ref_mv in NEW_MV mode based on mv, rate cost, etc.
997 // This speed feature equaling 0 means no skipping.
998 // If the speed feature equals 1 or 2, skip the current ref_mv in NEW_MV mode
999 // if we have already encountered ref_mv in the drl such that:
1000 // 1. The other drl has the same mv during the SIMPLE_TRANSLATION search
1001 // process as the current mv.
1002 // 2. The rate needed to encode the current mv is larger than that for the
1003 // other ref_mv.
1004 // The speed feature equaling 1 means using subpel mv in the comparison.
1005 // The speed feature equaling 2 means using fullpel mv in the comparison.
1006 // If the speed feature >= 3, skip the current ref_mv in NEW_MV mode based on
1007 // known full_mv bestsme and drl cost.
1008 int skip_newmv_in_drl;
1009
1010 // This speed feature checks duplicate ref MVs among NEARESTMV, NEARMV,
1011 // GLOBALMV and skips NEARMV or GLOBALMV (in order) if a duplicate is found
1012 // TODO(any): Instead of skipping repeated ref mv, use the recalculated
1013 // rd-cost based on mode rate and skip the mode evaluation
1014 int skip_repeated_ref_mv;
1015
1016 // Flag used to control the ref_best_rd based gating for chroma
1017 int perform_best_rd_based_gating_for_chroma;
1018
1019 // Reuse the inter_intra_mode search result from NEARESTMV mode to other
1020 // single ref modes
1021 int reuse_inter_intra_mode;
1022
1023 // prune wedge and compound segment approximate rd evaluation based on
1024 // compound average modeled rd
1025 int prune_comp_type_by_model_rd;
1026
1027 // prune wedge and compound segment approximate rd evaluation based on
1028 // compound average rd/ref_best_rd
1029 int prune_comp_type_by_comp_avg;
1030
1031 // Skip some ref frames in compound motion search by single motion search
1032 // result. Has three levels for now: 0 referring to no skipping, and 1 - 3
1033 // increasing aggressiveness of skipping in order.
1034 // Note: The search order might affect the result. It assumes that the single
1035 // reference modes are searched before compound modes. It is better to search
1036 // same single inter mode as a group.
1037 int prune_comp_search_by_single_result;
1038
1039 // Instead of performing a full MV search, do a simple translation first
1040 // and only perform a full MV search on the motion vectors that performed
1041 // well.
1042 int prune_mode_search_simple_translation;
1043
1044 // Only search compound modes with at least one "good" reference frame.
1045 // A reference frame is good if, after looking at its performance among
1046 // the single reference modes, it is one of the two best performers.
1047 int prune_compound_using_single_ref;
1048
1049 // Skip extended compound mode (NEAREST_NEWMV, NEW_NEARESTMV, NEAR_NEWMV,
1050 // NEW_NEARMV) using ref frames of above and left neighbor
1051 // blocks.
1052 // 0 : no pruning
1053 // 1 : prune ext compound modes using neighbor blocks (less aggressiveness)
1054 // 2 : prune ext compound modes using neighbor blocks (high aggressiveness)
1055 // 3 : prune ext compound modes unconditionally (highest aggressiveness)
1056 int prune_ext_comp_using_neighbors;
1057
1058 // Skip NEW_NEARMV and NEAR_NEWMV extended compound modes
1059 int skip_ext_comp_nearmv_mode;
1060
1061 // Skip extended compound mode when ref frame corresponding to NEWMV does not
1062 // have NEWMV as single mode winner.
1063 // 0 : no pruning
1064 // 1 : prune extended compound mode (less aggressiveness)
1065 // 2 : prune extended compound mode (high aggressiveness)
1066 int prune_comp_using_best_single_mode_ref;
1067
1068 // Skip NEARESTMV and NEARMV using weight computed in ref mv list population
1069 //
1070 // Pruning is enabled only when both the top and left neighbor blocks are
1071 // available and when the current block already has a valid inter prediction.
1072 int prune_nearest_near_mv_using_refmv_weight;
1073
1074 // Based on previous ref_mv_idx search result, prune the following search.
1075 int prune_ref_mv_idx_search;
1076
1077 // Disable one sided compound modes.
1078 int disable_onesided_comp;
1079
1080 // Prune obmc search using previous frame stats.
1081 // INT_MAX : disable obmc search
1082 int prune_obmc_prob_thresh;
1083
1084 // Prune warped motion search using previous frame stats.
1085 int prune_warped_prob_thresh;
1086
1087 // Variance threshold to enable/disable Interintra wedge search
1088 unsigned int disable_interintra_wedge_var_thresh;
1089
1090 // Variance threshold to enable/disable Interinter wedge search
1091 unsigned int disable_interinter_wedge_var_thresh;
1092
1093 // De-couple wedge and mode search during interintra RDO.
1094 int fast_interintra_wedge_search;
1095
1096 // Whether fast wedge sign estimate is used
1097 int fast_wedge_sign_estimate;
1098
1099 // Enable/disable ME for interinter wedge search.
1100 int disable_interinter_wedge_newmv_search;
1101
1102 // Decide when and how to use joint_comp.
1103 DIST_WTD_COMP_FLAG use_dist_wtd_comp_flag;
1104
1105 // Clip the frequency of updating the mv cost.
1106 INTERNAL_COST_UPDATE_TYPE mv_cost_upd_level;
1107
1108 // Clip the frequency of updating the coeff cost.
1109 INTERNAL_COST_UPDATE_TYPE coeff_cost_upd_level;
1110
1111 // Clip the frequency of updating the mode cost.
1112 INTERNAL_COST_UPDATE_TYPE mode_cost_upd_level;
1113
1114 // Prune inter modes based on tpl stats
1115 // 0 : no pruning
1116 // 1 - 3 indicate increasing aggressiveness in order.
1117 int prune_inter_modes_based_on_tpl;
1118
1119 // Skip NEARMV and NEAR_NEARMV modes using ref frames of above and left
1120 // neighbor blocks and qindex.
1121 PRUNE_NEARMV_LEVEL prune_nearmv_using_neighbors;
1122
1123 // Model based breakout after interpolation filter search
1124 // 0: no breakout
1125 // 1: use model based rd breakout
1126 int model_based_post_interp_filter_breakout;
1127
1128 // Reuse compound type rd decision when exact match is found
1129 // 0: No reuse
1130 // 1: Reuse the compound type decision
1131 int reuse_compound_type_decision;
1132
1133 // Enable/disable masked compound.
1134 int disable_masked_comp;
1135
1136 // Enable/disable MV refinement for compound modes corresponds to compound
1137 // types COMPOUND_AVERAGE, COMPOUND_DISTWTD (currently, this compound type
1138 // is disabled for speeds >= 2 using the sf 'use_dist_wtd_comp_flag') and
1139 // COMPOUND_DIFFWTD based on the availability. Levels 0 to 3 indicate
1140 // increasing order of aggressiveness to disable MV refinement.
1141 // 0: MV Refinement is enabled and for NEW_NEWMV mode used two iterations of
1142 // refinement in av1_joint_motion_search().
1143 // 1: MV Refinement is disabled for COMPOUND_DIFFWTD and enabled for
1144 // COMPOUND_AVERAGE & COMPOUND_DISTWTD.
1145 // 2: MV Refinement is enabled for COMPOUND_AVERAGE & COMPOUND_DISTWTD for
1146 // NEW_NEWMV mode with one iteration of refinement in
1147 // av1_joint_motion_search() and MV Refinement is disabled for other compound
1148 // type modes.
1149 // 3: MV Refinement is disabled.
1150 int enable_fast_compound_mode_search;
1151
1152 // Reuse masked compound type search results
1153 int reuse_mask_search_results;
1154
1155 // Enable/disable fast search for wedge masks
1156 int enable_fast_wedge_mask_search;
1157
1158 // Early breakout from transform search of inter modes
1159 int inter_mode_txfm_breakout;
1160
1161 // Limit number of inter modes for txfm search if a newmv mode gets
1162 // evaluated among the top modes.
1163 // 0: no pruning
1164 // 1 to 3 indicate increasing order of aggressiveness
1165 int limit_inter_mode_cands;
1166
1167 // Cap the no. of txfm searches for a given prediction mode.
1168 // 0: no cap, 1: cap beyond first 4 searches, 2: cap beyond first 3 searches.
1169 int limit_txfm_eval_per_mode;
1170
1171 // Prune warped motion search based on block size.
1172 int extra_prune_warped;
1173
1174 // Do not search compound modes for ARF.
1175 // The intuition is that ARF is predicted by frames far away from it,
1176 // whose temporal correlations with the ARF are likely low.
1177 // It is therefore likely that compound modes do not work as well for ARF
1178 // as other inter frames.
1179 // Speed/quality impact:
1180 // Speed 1: 12% faster, 0.1% psnr loss.
1181 // Speed 2: 2% faster, 0.05% psnr loss.
1182 // No change for speed 3 and up, because |disable_onesided_comp| is true.
1183 int skip_arf_compound;
1184
1185 // Percentage of scaling used to increase the rd cost of warp mode so that
1186 // encoder decisions are biased against local warp, favoring low complexity
1187 // modes.
1188 int bias_warp_mode_rd_scale_pct;
1189} INTER_MODE_SPEED_FEATURES;
1190
1191typedef struct INTERP_FILTER_SPEED_FEATURES {
1192 // Do limited interpolation filter search for dual filters, since best choice
1193 // usually includes EIGHTTAP_REGULAR.
1194 int use_fast_interpolation_filter_search;
1195
1196 // Disable dual filter
1197 int disable_dual_filter;
1198
1199 // Save results of av1_interpolation_filter_search for a block
1200 // Check mv and ref_frames before search, if they are very close with previous
1201 // saved results, filter search can be skipped.
1202 int use_interp_filter;
1203
1204 // skip sharp_filter evaluation based on regular and smooth filter rd for
1205 // dual_filter=0 case
1206 int skip_sharp_interp_filter_search;
1207
1208 // skip interpolation filter search for a block in chessboard pattern
1209 int cb_pred_filter_search;
1210
1211 // adaptive interp_filter search to allow skip of certain filter types.
1212 int adaptive_interp_filter_search;
1213
1214 // Forces interpolation filter to EIGHTTAP_REGULAR and skips interpolation
1215 // filter search.
1216 int skip_interp_filter_search;
1217} INTERP_FILTER_SPEED_FEATURES;
1218
1219typedef struct INTRA_MODE_SPEED_FEATURES {
1220 // These bit masks allow you to enable or disable intra modes for each
1221 // transform size separately.
1222 int intra_y_mode_mask[TX_SIZES];
1223 int intra_uv_mode_mask[TX_SIZES];
1224
1225 // flag to allow skipping intra mode for inter frame prediction
1226 int skip_intra_in_interframe;
1227
1228 // Prune intra mode candidates based on source block histogram of gradient.
1229 // Applies to luma plane only.
1230 // Feasible values are 0..4. The feature is disabled for 0. An increasing
1231 // value indicates more aggressive pruning threshold.
1232 int intra_pruning_with_hog;
1233
1234 // Prune intra mode candidates based on source block histogram of gradient.
1235 // Applies to chroma plane only.
1236 // Feasible values are 0..4. The feature is disabled for 0. An increasing
1237 // value indicates more aggressive pruning threshold.
1238 int chroma_intra_pruning_with_hog;
1239
1240 // Enable/disable smooth intra modes.
1241 int disable_smooth_intra;
1242
1243 // Prune UV_SMOOTH_PRED mode for chroma based on chroma source variance.
1244 // false : No pruning
1245 // true : Prune UV_SMOOTH_PRED mode based on chroma source variance
1246 //
1247 // For allintra encode, this speed feature reduces instruction count
1248 // by 1.90%, 2.21% and 1.97% for speed 6, 7 and 8 with coding performance
1249 // change less than 0.04%. For AVIF image encode, this speed feature reduces
1250 // encode time by 1.56%, 2.14% and 0.90% for speed 6, 7 and 8 on a typical
1251 // image dataset with coding performance change less than 0.05%.
1252 bool prune_smooth_intra_mode_for_chroma;
1253
1254 // Prune filter intra modes in intra frames.
1255 // 0 : No pruning
1256 // 1 : Evaluate applicable filter intra modes based on best intra mode so far
1257 // 2 : Do not evaluate filter intra modes
1258 int prune_filter_intra_level;
1259
1260 // prune palette search
1261 // 0: No pruning
1262 // 1: Perform coarse search to prune the palette colors. For winner colors,
1263 // neighbors are also evaluated using a finer search.
1264 // 2: Perform 2 way palette search from max colors to min colors (and min
1265 // colors to remaining colors) and terminate the search if current number of
1266 // palette colors is not the winner.
1267 int prune_palette_search_level;
1268
1269 // Terminate early in luma palette_size search. Speed feature values indicate
1270 // increasing level of pruning.
1271 // 0: No early termination
1272 // 1: Terminate early for higher luma palette_size, if header rd cost of lower
1273 // palette_size is more than 2 * best_rd. This level of pruning is more
1274 // conservative when compared to sf level 2 as the cases which will get pruned
1275 // with sf level 1 is a subset of the cases which will get pruned with sf
1276 // level 2.
1277 // 2: Terminate early for higher luma palette_size, if header rd cost of lower
1278 // palette_size is more than best_rd.
1279 // For allintra encode, this sf reduces instruction count by 2.49%, 1.07%,
1280 // 2.76%, 2.30%, 1.84%, 2.69%, 2.04%, 2.05% and 1.44% for speed 0, 1, 2, 3, 4,
1281 // 5, 6, 7 and 8 on screen content set with coding performance change less
1282 // than 0.01% for speed <= 2 and less than 0.03% for speed >= 3. For AVIF
1283 // image encode, this sf reduces instruction count by 1.94%, 1.13%, 1.29%,
1284 // 0.93%, 0.89%, 1.03%, 1.07%, 1.20% and 0.18% for speed 0, 1, 2, 3, 4, 5, 6,
1285 // 7 and 8 on a typical image dataset with coding performance change less than
1286 // 0.01%.
1287 int prune_luma_palette_size_search_level;
1288
1289 // Prune chroma intra modes based on luma intra mode winner.
1290 // 0: No pruning
1291 // 1: Prune chroma intra modes other than UV_DC_PRED, UV_SMOOTH_PRED,
1292 // UV_CFL_PRED and the mode that corresponds to luma intra mode winner.
1293 int prune_chroma_modes_using_luma_winner;
1294
1295 // Clip the frequency of updating the mv cost for intrabc.
1296 INTERNAL_COST_UPDATE_TYPE dv_cost_upd_level;
1297
1298 // We use DCT_DCT transform followed by computing SATD (Sum of Absolute
1299 // Transformed Differences) as an estimation of RD score to quickly find the
1300 // best possible Chroma from Luma (CFL) parameter. Then we do a full RD search
1301 // near the best possible parameter. The search range is set here.
1302 // The range of cfl_searh_range should be [1, 33], and the following are the
1303 // recommended values.
1304 // 1: Fastest mode.
1305 // 3: Default mode that provides good speedup without losing compression
1306 // performance at speed 0.
1307 // 33: Exhaustive rd search (33 == CFL_MAGS_SIZE). This mode should only
1308 // be used for debugging purpose.
1309 int cfl_search_range;
1310
1311 // TOP_INTRA_MODEL_COUNT is 4 that is the number of top model rd to store in
1312 // intra mode decision. Here, add a speed feature to reduce this number for
1313 // higher speeds.
1314 int top_intra_model_count_allowed;
1315
1316 // Adapt top_intra_model_count_allowed locally to prune luma intra modes using
1317 // neighbor block and quantizer information.
1318 int adapt_top_model_rd_count_using_neighbors;
1319
1320 // Prune the evaluation of odd delta angles of directional luma intra modes by
1321 // using the rdcosts of neighbouring delta angles.
1322 // For allintra encode, this speed feature reduces instruction count
1323 // by 4.461%, 3.699% and 3.536% for speed 6, 7 and 8 on a typical video
1324 // dataset with coding performance change less than 0.26%. For AVIF image
1325 // encode, this speed feature reduces encode time by 2.849%, 2.471%,
1326 // and 2.051% for speed 6, 7 and 8 on a typical image dataset with coding
1327 // performance change less than 0.27%.
1328 int prune_luma_odd_delta_angles_in_intra;
1329
1330 // Terminate early in chroma palette_size search.
1331 // 0: No early termination
1332 // 1: Terminate early for higher palette_size, if header rd cost of lower
1333 // palette_size is more than best_rd.
1334 // For allintra encode, this sf reduces instruction count by 0.45%,
1335 // 0.62%, 1.73%, 2.50%, 2.89%, 3.09% and 3.86% for speed 0 to 6 on screen
1336 // content set with coding performance change less than 0.01%.
1337 // For AVIF image encode, this sf reduces instruction count by 0.45%, 0.81%,
1338 // 0.85%, 1.05%, 1.45%, 1.66% and 1.95% for speed 0 to 6 on a typical image
1339 // dataset with no quality drop.
1340 int early_term_chroma_palette_size_search;
1341
1342 // Skips the evaluation of filter intra modes in inter frames if rd evaluation
1343 // of luma intra dc mode results in invalid rd stats.
1344 int skip_filter_intra_in_inter_frames;
1345} INTRA_MODE_SPEED_FEATURES;
1346
1347typedef struct TX_SPEED_FEATURES {
1348 // Init search depth for square and rectangular transform partitions.
1349 // Values:
1350 // 0 - search full tree, 1: search 1 level, 2: search the highest level only
1351 int inter_tx_size_search_init_depth_sqr;
1352 int inter_tx_size_search_init_depth_rect;
1353 int intra_tx_size_search_init_depth_sqr;
1354 int intra_tx_size_search_init_depth_rect;
1355
1356 // If any dimension of a coding block size above 64, always search the
1357 // largest transform only, since the largest transform block size is 64x64.
1358 int tx_size_search_lgr_block;
1359
1360 TX_TYPE_SEARCH tx_type_search;
1361
1362 // Skip split transform block partition when the collocated bigger block
1363 // is selected as all zero coefficients.
1364 int txb_split_cap;
1365
1366 // Shortcut the transform block partition and type search when the target
1367 // rdcost is relatively lower.
1368 // Values are 0 (not used) , or 1 - 2 with progressively increasing
1369 // aggressiveness
1370 int adaptive_txb_search_level;
1371
1372 // Prune level for tx_size_type search for inter based on rd model
1373 // 0: no pruning
1374 // 1-2: progressively increasing aggressiveness of pruning
1375 int model_based_prune_tx_search_level;
1376
1377 // Refine TX type after fast TX search.
1378 int refine_fast_tx_search_results;
1379
1380 // Prune transform split/no_split eval based on residual properties. A value
1381 // of 0 indicates no pruning, and the aggressiveness of pruning progressively
1382 // increases from levels 1 to 3.
1383 int prune_tx_size_level;
1384
1385 // Prune the evaluation of transform depths as decided by the NN model.
1386 // false: No pruning.
1387 // true : Avoid the evaluation of specific transform depths using NN model.
1388 //
1389 // For allintra encode, this speed feature reduces instruction count
1390 // by 4.76%, 8.92% and 11.28% for speed 6, 7 and 8 with coding performance
1391 // change less than 0.32%. For AVIF image encode, this speed feature reduces
1392 // encode time by 4.65%, 9.16% and 10.45% for speed 6, 7 and 8 on a typical
1393 // image dataset with coding performance change less than 0.19%.
1394 bool prune_intra_tx_depths_using_nn;
1395
1396 // Enable/disable early breakout during transform search of intra modes, by
1397 // using the minimum rd cost possible. By using this approach, the rd
1398 // evaluation of applicable transform blocks (in the current block) can be
1399 // avoided as
1400 // 1) best_rd evolves during the search in choose_tx_size_type_from_rd()
1401 // 2) appropriate ref_best_rd is passed in intra_block_yrd()
1402 //
1403 // For allintra encode, this speed feature reduces instruction count
1404 // by 1.11%, 1.08%, 1.02% and 0.93% for speed 3, 6, 7 and 8 with coding
1405 // performance change less than 0.02%. For AVIF image encode, this speed
1406 // feature reduces encode time by 0.93%, 1.46%, 1.07%, 0.84%, 0.99% and 0.73%
1407 // for speed 3, 4, 5, 6, 7 and 8 on a typical image dataset with coding
1408 // performance change less than 0.004%.
1409 bool use_rd_based_breakout_for_intra_tx_search;
1410} TX_SPEED_FEATURES;
1411
1412typedef struct RD_CALC_SPEED_FEATURES {
1413 // Fast approximation of av1_model_rd_from_var_lapndz
1414 int simple_model_rd_from_var;
1415
1416 // Perform faster distortion computation during the R-D evaluation by trying
1417 // to approximate the prediction error with transform coefficients (faster but
1418 // less accurate) rather than computing distortion in the pixel domain (slower
1419 // but more accurate). The following methods are used for distortion
1420 // computation:
1421 // Method 0: Always compute distortion in the pixel domain
1422 // Method 1: Based on block error, try using transform domain distortion for
1423 // tx_type search and compute distortion in pixel domain for final RD_STATS
1424 // Method 2: Based on block error, try to compute distortion in transform
1425 // domain
1426 // Methods 1 and 2 may fallback to computing distortion in the pixel domain in
1427 // case the block error is less than the threshold, which is controlled by the
1428 // speed feature tx_domain_dist_thres_level.
1429 //
1430 // The speed feature tx_domain_dist_level decides which of the above methods
1431 // needs to be used across different mode evaluation stages as described
1432 // below:
1433 // Eval type: Default Mode Winner
1434 // Level 0 : Method 0 Method 2 Method 0
1435 // Level 1 : Method 1 Method 2 Method 0
1436 // Level 2 : Method 2 Method 2 Method 0
1437 // Level 3 : Method 2 Method 2 Method 2
1438 int tx_domain_dist_level;
1439
1440 // Transform domain distortion threshold level
1441 int tx_domain_dist_thres_level;
1442
1443 // Trellis (dynamic programming) optimization of quantized values
1444 TRELLIS_OPT_TYPE optimize_coefficients;
1445
1446 // Use hash table to store macroblock RD search results
1447 // to avoid repeated search on the same residue signal.
1448 int use_mb_rd_hash;
1449
1450 // Flag used to control the extent of coeff R-D optimization
1451 int perform_coeff_opt;
1452} RD_CALC_SPEED_FEATURES;
1453
1454typedef struct WINNER_MODE_SPEED_FEATURES {
1455 // Flag used to control the winner mode processing for better R-D optimization
1456 // of quantized coeffs
1457 int enable_winner_mode_for_coeff_opt;
1458
1459 // Flag used to control the winner mode processing for transform size
1460 // search method
1461 int enable_winner_mode_for_tx_size_srch;
1462
1463 // Control transform size search level
1464 // Eval type: Default Mode Winner
1465 // Level 0 : FULL RD LARGEST ALL FULL RD
1466 // Level 1 : FAST RD LARGEST ALL FULL RD
1467 // Level 2 : LARGEST ALL LARGEST ALL FULL RD
1468 // Level 3 : LARGEST ALL LARGEST ALL LARGEST ALL
1469 int tx_size_search_level;
1470
1471 // Flag used to control the winner mode processing for use transform
1472 // domain distortion
1473 int enable_winner_mode_for_use_tx_domain_dist;
1474
1475 // Flag used to enable processing of multiple winner modes
1476 MULTI_WINNER_MODE_TYPE multi_winner_mode_type;
1477
1478 // Motion mode for winner candidates:
1479 // 0: speed feature OFF
1480 // 1 / 2 : Use configured number of winner candidates
1481 int motion_mode_for_winner_cand;
1482
1483 // Controls the prediction of transform skip block or DC only block.
1484 //
1485 // Different speed feature values (0 to 3) decide the aggressiveness of
1486 // prediction (refer to predict_dc_levels[][] in speed_features.c) to be used
1487 // during different mode evaluation stages.
1488 int dc_blk_pred_level;
1489
1490 // If on, disables interpolation filter search in handle_inter_mode loop, and
1491 // performs it during winner mode processing by \ref
1492 // tx_search_best_inter_candidates.
1493 int winner_mode_ifs;
1494
1495 // Controls the disabling of winner mode processing. Speed feature levels
1496 // are ordered in increasing aggressiveness of pruning. The method considered
1497 // for disabling, depends on the sf level value and it is described as below.
1498 // 0: Do not disable
1499 // 1: Disable for blocks with low source variance.
1500 // 2: Disable for blocks which turn out to be transform skip (skipped based on
1501 // eob) during MODE_EVAL stage except NEWMV mode.
1502 // 3: Disable for blocks which turn out to be transform skip during MODE_EVAL
1503 // stage except NEWMV mode. For high quantizers, prune conservatively based on
1504 // transform skip (skipped based on eob) except for NEWMV mode.
1505 // 4: Disable for blocks which turn out to be transform skip during MODE_EVAL
1506 // stage.
1507 int prune_winner_mode_eval_level;
1508} WINNER_MODE_SPEED_FEATURES;
1509
1510typedef struct LOOP_FILTER_SPEED_FEATURES {
1511 // This feature controls how the loop filter level is determined.
1512 LPF_PICK_METHOD lpf_pick;
1513
1514 // Skip some final iterations in the determination of the best loop filter
1515 // level.
1516 int use_coarse_filter_level_search;
1517
1518 // Reset luma filter levels to zero based on minimum filter levels of
1519 // reference frames and current frame's pyramid level.
1520 int adaptive_luma_loop_filter_skip;
1521
1522 // Reset luma filter levels to zero when the percentage of SSE difference
1523 // between the unfiltered and filtered versions of the current frame is below
1524 // a threshold.
1525 int skip_loop_filter_using_filt_error;
1526
1527 // Control how the CDEF strength is determined.
1528 CDEF_PICK_METHOD cdef_pick_method;
1529
1530 // Decoder side speed feature to add penalty for use of dual-sgr filters.
1531 // Takes values 0 - 10, 0 indicating no penalty and each additional level
1532 // adding a penalty of 1%
1533 int dual_sgr_penalty_level;
1534
1535 // prune sgr ep using binary search like mechanism
1536 int enable_sgr_ep_pruning;
1537
1538 // Disable loop restoration for Chroma plane
1539 int disable_loop_restoration_chroma;
1540
1541 // Disable loop restoration for luma plane
1542 int disable_loop_restoration_luma;
1543
1544 // Range of loop restoration unit sizes to search
1545 // The minimum size is clamped against the superblock size in
1546 // av1_pick_filter_restoration, so that the code which sets this value does
1547 // not need to know the superblock size ahead of time.
1548 int min_lr_unit_size;
1549 int max_lr_unit_size;
1550
1551 // Prune RESTORE_WIENER evaluation based on source variance
1552 // 0 : no pruning
1553 // 1 : conservative pruning
1554 // 2 : aggressive pruning
1555 int prune_wiener_based_on_src_var;
1556
1557 // Prune self-guided loop restoration based on wiener search results
1558 // 0 : no pruning
1559 // 1 : pruning based on rdcost ratio of RESTORE_WIENER and RESTORE_NONE
1560 // 2 : pruning based on winner restoration type among RESTORE_WIENER and
1561 // RESTORE_NONE
1562 int prune_sgr_based_on_wiener;
1563
1564 // Reduce the wiener filter win size for luma
1565 int reduce_wiener_window_size;
1566
1567 // Flag to disable Wiener Loop restoration filter.
1568 bool disable_wiener_filter;
1569
1570 // Flag to disable Self-guided Loop restoration filter.
1571 bool disable_sgr_filter;
1572
1573 // Disable the refinement search around the wiener filter coefficients.
1574 bool disable_wiener_coeff_refine_search;
1575
1576 // Whether to downsample the rows in computation of wiener stats.
1577 int use_downsampled_wiener_stats;
1578} LOOP_FILTER_SPEED_FEATURES;
1579
1580typedef struct REAL_TIME_SPEED_FEATURES {
1581 // check intra prediction for non-RD mode.
1582 int check_intra_pred_nonrd;
1583
1584 // Skip checking intra prediction.
1585 // 0 - don't skip
1586 // 1 - skip if TX is skipped and best mode is not NEWMV
1587 // 2 - skip if TX is skipped
1588 // Skipping aggressiveness increases from level 1 to 2.
1589 int skip_intra_pred;
1590
1591 // Estimate motion before calculating variance in variance-based partition
1592 // 0 - Only use zero MV
1593 // 1 - perform coarse ME
1594 // 2 - perform coarse ME, and also use neighbours' MVs
1595 // 3 - use neighbours' MVs without performing coarse ME
1596 int estimate_motion_for_var_based_partition;
1597
1598 // For nonrd_use_partition: mode of extra check of leaf partition
1599 // 0 - don't check merge
1600 // 1 - always check merge
1601 // 2 - check merge and prune checking final split
1602 // 3 - check merge and prune checking final split based on bsize and qindex
1603 int nonrd_check_partition_merge_mode;
1604
1605 // For nonrd_use_partition: check of leaf partition extra split
1606 int nonrd_check_partition_split;
1607
1608 // Implements various heuristics to skip searching modes
1609 // The heuristics selected are based on flags
1610 // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
1611 unsigned int mode_search_skip_flags;
1612
1613 // For nonrd: Reduces ref frame search.
1614 // 0 - low level of search prune in non last frames
1615 // 1 - pruned search in non last frames
1616 // 2 - more pruned search in non last frames
1617 int nonrd_prune_ref_frame_search;
1618
1619 // This flag controls the use of non-RD mode decision.
1620 int use_nonrd_pick_mode;
1621
1622 // Use ALTREF frame in non-RD mode decision.
1623 int use_nonrd_altref_frame;
1624
1625 // Use compound reference for non-RD mode.
1626 int use_comp_ref_nonrd;
1627
1628 // Reference frames for compound prediction for nonrd pickmode:
1629 // LAST_GOLDEN (0), LAST_LAST2 (1), or LAST_ALTREF (2).
1630 int ref_frame_comp_nonrd[3];
1631
1632 // use reduced ref set for real-time mode
1633 int use_real_time_ref_set;
1634
1635 // Skip a number of expensive mode evaluations for blocks with very low
1636 // temporal variance.
1637 int short_circuit_low_temp_var;
1638
1639 // Reuse inter prediction in fast non-rd mode.
1640 int reuse_inter_pred_nonrd;
1641
1642 // Number of best inter modes to search transform. INT_MAX - search all.
1643 int num_inter_modes_for_tx_search;
1644
1645 // Use interpolation filter search in non-RD mode decision.
1646 int use_nonrd_filter_search;
1647
1648 // Use simplified RD model for interpolation search and Intra
1649 int use_simple_rd_model;
1650
1651 // For nonrd mode: use hybrid intra mode search for intra only frames based on
1652 // block properties.
1653 // 0 : use nonrd pick intra for all blocks
1654 // 1 : use rd for bsize < 16x16, nonrd otherwise
1655 // 2 : use rd for bsize < 16x16 and src var >= 101, nonrd otherwise
1656 int hybrid_intra_pickmode;
1657
1658 // Filter blocks by certain criteria such as SAD, source variance, such that
1659 // fewer blocks will go through the palette search.
1660 // For nonrd encoding path, enable this feature reduces encoding time when
1661 // palette mode is used. Disabling it leads to better compression efficiency.
1662 // 0: off
1663 // 1: less aggressive pruning mode
1664 // 2, 3: more aggressive pruning mode
1665 int prune_palette_search_nonrd;
1666
1667 // Compute variance/sse on source difference, prior to encoding superblock.
1668 int source_metrics_sb_nonrd;
1669
1670 // Flag to indicate process for handling overshoot on slide/scene change,
1671 // for real-time CBR mode.
1672 OVERSHOOT_DETECTION_CBR overshoot_detection_cbr;
1673
1674 // Check for scene/content change detection on every frame before encoding.
1675 int check_scene_detection;
1676
1677 // For keyframes in rtc: adjust the rc_bits_per_mb, to reduce overshoot.
1678 int rc_adjust_keyframe;
1679
1680 // On scene change: compute spatial variance.
1681 int rc_compute_spatial_var_sc;
1682
1683 // For nonrd mode: Prefer larger partition blks in variance based partitioning
1684 // 0: disabled, 1-3: increasing aggressiveness
1685 int prefer_large_partition_blocks;
1686
1687 // uses results of temporal noise estimate
1688 int use_temporal_noise_estimate;
1689
1690 // Parameter indicating initial search window to be used in full-pixel search
1691 // for nonrd_pickmode. Range [0, MAX_MVSEARCH_STEPS - 1]. Lower value
1692 // indicates larger window. If set to 0, step_param is set based on internal
1693 // logic in set_mv_search_params().
1694 int fullpel_search_step_param;
1695
1696 // Bit mask to enable or disable intra modes for each prediction block size
1697 // separately, for nonrd_pickmode. Currently, the sf is not respected when
1698 // 'force_intra_check' is true in 'av1_estimate_intra_mode()' function. Also,
1699 // H and V pred modes allowed through this sf can be further pruned when
1700 //'prune_hv_pred_modes_using_src_sad' sf is true.
1701 int intra_y_mode_bsize_mask_nrd[BLOCK_SIZES];
1702
1703 // Prune H and V intra predition modes evalution in inter frame.
1704 // The sf does not have any impact.
1705 // i. when frame_source_sad is 1.1 times greater than avg_source_sad
1706 // ii. when cyclic_refresh_segment_id_boosted is enabled
1707 // iii. when SB level source sad is greater than kMedSad
1708 // iv. when color sensitivity is non zero for both the chroma channels
1709 bool prune_hv_pred_modes_using_src_sad;
1710
1711 // Skips mode checks more aggressively in nonRD mode
1712 int nonrd_aggressive_skip;
1713
1714 // Skip cdef on 64x64 blocks/
1715 // 0: disabled
1716 // 1: skip when NEWMV or INTRA is not picked or color sensitivity is off.
1717 // When color sensitivity is on for a superblock, all 64x64 blocks within
1718 // will not skip.
1719 // 2: more aggressive mode where skip is done for all frames where
1720 // rc->high_source_sad = 0 (non slide-changes), and color sensitivity off.
1721 int skip_cdef_sb;
1722
1723 // Force selective cdf update.
1724 int selective_cdf_update;
1725
1726 // Force only single reference (LAST) for prediction.
1727 int force_only_last_ref;
1728
1729 // Forces larger partition blocks in variance based partitioning for intra
1730 // frames
1731 int force_large_partition_blocks_intra;
1732
1733 // Use fixed partition for superblocks based on source_sad.
1734 // 0: disabled
1735 // 1: enabled
1736 int use_fast_fixed_part;
1737
1738 // Increase source_sad thresholds in nonrd pickmode.
1739 int increase_source_sad_thresh;
1740
1741 // Skip evaluation of no split in tx size selection for merge partition
1742 int skip_tx_no_split_var_based_partition;
1743
1744 // Intermediate termination of newMV mode evaluation based on so far best mode
1745 // sse
1746 int skip_newmv_mode_based_on_sse;
1747
1748 // Define gf length multiplier.
1749 // Level 0: use large multiplier, level 1: use medium multiplier.
1750 int gf_length_lvl;
1751
1752 // Prune inter modes with golden frame as reference for NEARMV and NEWMV modes
1753 int prune_inter_modes_with_golden_ref;
1754
1755 // Prune inter modes w.r.t golden or alt-ref frame based on sad
1756 int prune_inter_modes_wrt_gf_arf_based_on_sad;
1757
1758 // Prune inter mode search in rd path based on current block's temporal
1759 // variance wrt LAST reference.
1760 int prune_inter_modes_using_temp_var;
1761
1762 // Reduce MV precision to halfpel for higher int MV value & frame-level motion
1763 // 0: disabled
1764 // 1-2: Reduce precision to halfpel, fullpel based on conservative
1765 // thresholds, aggressiveness increases with increase in level
1766 // 3: Reduce precision to halfpel using more aggressive thresholds
1767 int reduce_mv_pel_precision_highmotion;
1768
1769 // Reduce MV precision for low complexity blocks
1770 // 0: disabled
1771 // 1: Reduce the mv resolution for zero mv if the variance is low
1772 // 2: Switch to halfpel, fullpel based on low block spatial-temporal
1773 // complexity.
1774 int reduce_mv_pel_precision_lowcomplex;
1775
1776 // Prune intra mode evaluation in inter frames based on mv range.
1777 BLOCK_SIZE prune_intra_mode_based_on_mv_range;
1778 // The number of times to left shift the splitting thresholds in variance
1779 // based partitioning. The minimum values should be 7 to avoid left shifting
1780 // by a negative number.
1781 int var_part_split_threshold_shift;
1782
1783 // Qindex based variance partition threshold index, which determines
1784 // the aggressiveness of partition pruning
1785 // 0: disabled for speeds 9,10
1786 // 1,2: (rd-path) lowers qindex thresholds conditionally (for low SAD sb)
1787 // 3,4: (non-rd path) uses pre-tuned qindex thresholds
1788 int var_part_based_on_qidx;
1789
1790 // Enable GF refresh based on Q value.
1791 int gf_refresh_based_on_qp;
1792
1793 // Temporal filtering
1794 // The value can be 1 or 2, which indicates the threshold to use.
1795 // Must be off for lossless mode.
1796 int use_rtc_tf;
1797
1798 // Use of the identity transform in nonrd_pickmode,
1799 int use_idtx_nonrd;
1800
1801 // Prune the use of the identity transform in nonrd_pickmode:
1802 // only for smaller blocks and higher spatial variance, and when skip_txfm
1803 // is not already set.
1804 int prune_idtx_nonrd;
1805
1806 // Force to only use dct for palette search in nonrd pickmode.
1807 int dct_only_palette_nonrd;
1808
1809 // Skip loopfilter, for static content after slide change
1810 // or key frame, once quality has ramped up.
1811 // 0: disabled
1812 // 1: skip only after quality is ramped up.
1813 // 2: aggrssive mode, where skip is done for all frames that
1814 // where rc->high_source_sad = 0 (no slide-changes).
1815 int skip_lf_screen;
1816
1817 // Threshold on the active/inactive region percent to disable
1818 // the loopfilter and cdef. Setting to 100 disables this feature.
1819 int thresh_active_maps_skip_lf_cdef;
1820
1821 // For nonrd: early exit out of variance partition that sets the
1822 // block size to superblock size, and sets mode to zeromv-last skip.
1823 // 0: disabled
1824 // 1: zeromv-skip is enabled at SB level only
1825 // 2: zeromv-skip is enabled at SB level and coding block level
1826 int part_early_exit_zeromv;
1827
1828 // Early terminate inter mode search based on sse in non-rd path.
1829 INTER_SEARCH_EARLY_TERM_IDX sse_early_term_inter_search;
1830
1831 // SAD based adaptive altref selection
1832 int sad_based_adp_altref_lag;
1833
1834 // Enable/disable partition direct merging.
1835 int partition_direct_merging;
1836
1837 // Level of aggressiveness for obtaining tx size based on qstep
1838 int tx_size_level_based_on_qstep;
1839
1840 // Avoid the partitioning of a 16x16 block in variance based partitioning
1841 // (VBP) by making use of minimum and maximum sub-block variances.
1842 // For allintra encode, this speed feature reduces instruction count by 5.39%
1843 // for speed 9 on a typical video dataset with coding performance gain
1844 // of 1.44%.
1845 // For AVIF image encode, this speed feature reduces encode time
1846 // by 8.44% for speed 9 on a typical image dataset with coding performance
1847 // gain of 0.78%.
1848 bool vbp_prune_16x16_split_using_min_max_sub_blk_var;
1849
1850 // A qindex threshold that determines whether to use qindex based CDEF filter
1851 // strength estimation for screen content types. The strength estimation model
1852 // used for screen contents prefers to allow cdef filtering for more frames.
1853 // This sf is used to limit the frames which go through cdef filtering and
1854 // following explains the setting of the same.
1855 // MAXQ (255): This disables the usage of this sf. Here, frame does not use a
1856 // screen content model thus reduces the number of frames that go through cdef
1857 // filtering.
1858 // MINQ (0): Frames always use screen content model thus increasing the number
1859 // of frames that go through cdef filtering.
1860 // This speed feature has a substantial gain on coding metrics, with moderate
1861 // increase encoding time. Select threshold based on speed vs quality
1862 // trade-off.
1863 int screen_content_cdef_filter_qindex_thresh;
1864
1865 // Prune compound mode if its variance is higher than the variance of single
1866 // modes.
1867 bool prune_compoundmode_with_singlecompound_var;
1868
1869 // Allow mode cost update at frame level every couple frames. This
1870 // overrides the command line setting --mode-cost-upd-freq=3 (never update
1871 // except on key frame and first delta).
1872 bool frame_level_mode_cost_update;
1873
1874 // Prune H_PRED during intra mode evaluation in the nonrd path based on best
1875 // mode so far.
1876 //
1877 // For allintra encode, this speed feature reduces instruction count by 1.10%
1878 // for speed 9 with coding performance change less than 0.04%.
1879 // For AVIF image encode, this speed feature reduces encode time by 1.03% for
1880 // speed 9 on a typical image dataset with coding performance change less than
1881 // 0.08%.
1882 bool prune_h_pred_using_best_mode_so_far;
1883
1884 // Enable pruning of intra mode evaluations in nonrd path based on source
1885 // variance and best mode so far. The pruning logic is enabled only if the
1886 // mode is not a winner mode of both the neighboring blocks (left/top).
1887 //
1888 // For allintra encode, this speed feature reduces instruction count by 3.96%
1889 // for speed 9 with coding performance change less than 0.38%.
1890 // For AVIF image encode, this speed feature reduces encode time by 3.46% for
1891 // speed 9 on a typical image dataset with coding performance change less than
1892 // -0.06%.
1893 bool enable_intra_mode_pruning_using_neighbors;
1894
1895 // Prune intra mode evaluations in nonrd path based on best sad so far.
1896 //
1897 // For allintra encode, this speed feature reduces instruction count by 3.05%
1898 // for speed 9 with coding performance change less than 0.24%.
1899 // For AVIF image encode, this speed feature reduces encode time by 1.87% for
1900 // speed 9 on a typical image dataset with coding performance change less than
1901 // 0.16%.
1902 bool prune_intra_mode_using_best_sad_so_far;
1903
1904 // If compound is enabled, and the current block size is \geq BLOCK_16X16,
1905 // limit the compound modes to GLOBAL_GLOBALMV. This does not apply to the
1906 // base layer of svc.
1907 bool check_only_zero_zeromv_on_large_blocks;
1908
1909 // Allow for disabling cdf update for non reference frames in svc mode.
1910 bool disable_cdf_update_non_reference_frame;
1911
1912 // Prune compound modes if the single modes variances do not perform well.
1913 bool prune_compoundmode_with_singlemode_var;
1914
1915 // Skip searching all compound mode if the variance of single_mode residue is
1916 // sufficiently low.
1917 bool skip_compound_based_on_var;
1918
1919 // Sets force_zeromv_skip based on the source sad available. Aggressiveness
1920 // increases with increase in the level set for speed feature.
1921 // 0: No setting
1922 // 1: If source sad is kZeroSad
1923 // 2: If source sad <= kVeryLowSad
1924 int set_zeromv_skip_based_on_source_sad;
1925
1926 // Downgrades the block-level subpel motion search to
1927 // av1_find_best_sub_pixel_tree_pruned_more for higher QP and when fullpel
1928 // search performed well, zeromv has low sad or low source_var
1929 bool use_adaptive_subpel_search;
1930
1931 // A flag used in RTC case to control frame_refs_short_signaling. Note that
1932 // the final decision is made in check_frame_refs_short_signaling(). The flag
1933 // can only be turned on when res < 360p and speed >= 9, in which case only
1934 // LAST and GOLDEN ref frames are used now.
1935 bool enable_ref_short_signaling;
1936
1937 // A flag that controls if we check or bypass GLOBALMV in rtc single ref frame
1938 // case.
1939 bool check_globalmv_on_single_ref;
1940
1941 // Allows for increasing the color_threshold for palette prediction.
1942 // This generally leads to better coding efficiency but with some speed loss.
1943 // Only used for screen content and for nonrd_pickmode.
1944 bool increase_color_thresh_palette;
1945
1946 // Flag to indicate selecting of higher threshold for scenee change detection.
1947 int higher_thresh_scene_detection;
1948
1949 // FLag to indicate skip testing of NEWMV for flat blocks.
1950 int skip_newmv_flat_blocks_screen;
1951
1952 // Flag to force skip encoding for non_reference_frame on slide/scene changes.
1953 int skip_encoding_non_reference_slide_change;
1954
1955 // Flag to indicate more aggressive QP downward adjustment for screen static
1956 // content, to make convergence to min_qp faster.
1957 int rc_faster_convergence_static;
1958
1959 // Skip NEWMV mode evaluation based on sad for screen content.
1960 int skip_newmv_mode_sad_screen;
1961} REAL_TIME_SPEED_FEATURES;
1962
1968typedef struct SPEED_FEATURES {
1973
1978
1982 TPL_SPEED_FEATURES tpl_sf;
1983
1987 GLOBAL_MOTION_SPEED_FEATURES gm_sf;
1988
1992 PARTITION_SPEED_FEATURES part_sf;
1993
1997 MV_SPEED_FEATURES mv_sf;
1998
2002 INTER_MODE_SPEED_FEATURES inter_sf;
2003
2007 INTERP_FILTER_SPEED_FEATURES interp_sf;
2008
2012 INTRA_MODE_SPEED_FEATURES intra_sf;
2013
2017 TX_SPEED_FEATURES tx_sf;
2018
2022 RD_CALC_SPEED_FEATURES rd_sf;
2023
2027 WINNER_MODE_SPEED_FEATURES winner_mode_sf;
2028
2032 LOOP_FILTER_SPEED_FEATURES lpf_sf;
2033
2037 REAL_TIME_SPEED_FEATURES rt_sf;
2041struct AV1_COMP;
2042
2056 int speed);
2057
2070 int speed);
2083
2084#ifdef __cplusplus
2085} // extern "C"
2086#endif
2087
2088#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:2761
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:2885
int speed
Definition encoder.h:3107
Definition speed_features.h:498
int disable_recon
Skips reconstruction by using source buffers for prediction.
Definition speed_features.h:515
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:505
int skip_zeromv_motion_search
Skips the motion search centered on 0,0 mv.
Definition speed_features.h:520
int skip_motion_search_threshold
Skips the motion search when the zero mv has small sse.
Definition speed_features.h:510
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 ref_frame_mvs_lvl
Definition speed_features.h:492
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:1968
MV_SPEED_FEATURES mv_sf
Definition speed_features.h:1997
TPL_SPEED_FEATURES tpl_sf
Definition speed_features.h:1982
LOOP_FILTER_SPEED_FEATURES lpf_sf
Definition speed_features.h:2032
TX_SPEED_FEATURES tx_sf
Definition speed_features.h:2017
INTER_MODE_SPEED_FEATURES inter_sf
Definition speed_features.h:2002
RD_CALC_SPEED_FEATURES rd_sf
Definition speed_features.h:2022
PARTITION_SPEED_FEATURES part_sf
Definition speed_features.h:1992
GLOBAL_MOTION_SPEED_FEATURES gm_sf
Definition speed_features.h:1987
INTERP_FILTER_SPEED_FEATURES interp_sf
Definition speed_features.h:2007
FIRST_PASS_SPEED_FEATURES fp_sf
Definition speed_features.h:1977
INTRA_MODE_SPEED_FEATURES intra_sf
Definition speed_features.h:2012
WINNER_MODE_SPEED_FEATURES winner_mode_sf
Definition speed_features.h:2027
REAL_TIME_SPEED_FEATURES rt_sf
Definition speed_features.h:2037
HIGH_LEVEL_SPEED_FEATURES hl_sf
Definition speed_features.h:1972