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
25
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);
160
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);
330
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
838 // Decoder side speed feature to add penalty for use of smaller partitions.
839 // Takes values 0 - 2, 0 indicating no penalty and higher level indicating
840 // increased penalty.
841 int split_partition_penalty_level;
842} PARTITION_SPEED_FEATURES;
843
844typedef struct MV_SPEED_FEATURES {
845 // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
846 SEARCH_METHODS search_method;
847
848 // Enable the use of faster, less accurate mv search method
849 // 0: disable, 1: if bsize >= BLOCK_32X32, 2: based on bsize, SAD and qp
850 // TODO(chiyotsai@google.com): Take the clip's resolution and mv activity into
851 // account.
852 int use_bsize_dependent_search_method;
853
854 // If this is set to 1, we limit the motion search range to 2 times the
855 // largest motion vector found in the last frame.
856 int auto_mv_step_size;
857
858 // Subpel_search_method can only be subpel_tree which does a subpixel
859 // logarithmic search that keeps stepping at 1/2 pixel units until
860 // you stop getting a gain, and then goes on to 1/4 and repeats
861 // the same process. Along the way it skips many diagonals.
862 SUBPEL_SEARCH_METHOD subpel_search_method;
863
864 // Maximum number of steps in logarithmic subpel search before giving up.
865 int subpel_iters_per_step;
866
867 // When to stop subpel search.
868 SUBPEL_FORCE_STOP subpel_force_stop;
869
870 // When to stop subpel search in simple motion search.
871 SUBPEL_FORCE_STOP simple_motion_subpel_force_stop;
872
873 // If true, sub-pixel search uses the exact convolve function used for final
874 // encoding and decoding; otherwise, it uses bilinear interpolation.
875 SUBPEL_SEARCH_TYPE use_accurate_subpel_search;
876
877 // Threshold for allowing exhaustive motion search.
878 int exhaustive_searches_thresh;
879
880 // Pattern to be used for any exhaustive mesh searches (except intraBC ME).
881 MESH_PATTERN mesh_patterns[MAX_MESH_STEP];
882
883 // Pattern to be used for exhaustive mesh searches of intraBC ME.
884 MESH_PATTERN intrabc_mesh_patterns[MAX_MESH_STEP];
885
886 // Reduce single motion search range based on MV result of prior ref_mv_idx.
887 int reduce_search_range;
888
889 // Prune mesh search.
890 PRUNE_MESH_SEARCH_LEVEL prune_mesh_search;
891
892 // Use the rd cost around the best FULLPEL_MV to speed up subpel search
893 int use_fullpel_costlist;
894
895 // Set the full pixel search level of obmc
896 // 0: obmc_full_pixel_diamond
897 // 1: obmc_refining_search_sad (faster)
898 int obmc_full_pixel_search_level;
899
900 // Accurate full pixel motion search based on TPL stats.
901 int full_pixel_search_level;
902
903 // Allow intrabc motion search
904 int use_intrabc;
905
906 // Prune intrabc candidate block hash search
907 // 0: check every block hash candidate
908 // 1: check the first 64 block hash candidates only
909 int prune_intrabc_candidate_block_hash_search;
910
911 // Intrabc search level
912 // 0: top + left search
913 // 1: top search only
914 int intrabc_search_level;
915
916 // Whether the maximum intrabc block size to hash is 8x8
917 // 0: Hash from 4x4 up to superblock size
918 // 1: Hash 4x4 and 8x8 only
919 int hash_max_8x8_intrabc_blocks;
920
921 // Whether to downsample the rows in sad calculation during motion search.
922 // This is only active when there are at least 16 rows. When this sf is
923 // active, if there is a large discrepancy in the SAD values for the final
924 // motion vector between skipping vs not skipping, motion search is redone
925 // with skip row features off.
926 // 0: Disabled (do not downsample rows)
927 // 1: Skip SAD calculation of odd rows if the SAD deviation of the even and
928 // odd rows for the starting MV is small. Redo motion search with sf off
929 // when SAD deviation is high for the final motion vector.
930 // 2: Skip SAD calculation of odd rows. SAD deviation is not tested for the
931 // start MV and tested only for the final MV.
932 int use_downsampled_sad;
933
934 // Enable/disable extensive joint motion search.
935 int disable_extensive_joint_motion_search;
936
937 // Enable second best mv check in joint mv search.
938 // 0: allow second MV (use rd cost as the metric)
939 // 1: use var as the metric
940 // 2: disable second MV
941 int disable_second_mv;
942
943 // Skips full pixel search based on start mv of prior ref_mv_idx.
944 // 0: Disabled
945 // 1: Skips the full pixel search upto 4 neighbor full-pel MV positions.
946 // 2: Skips the full pixel search upto 8 neighbor full-pel MV positions.
947 int skip_fullpel_search_using_startmv;
948
949 // Method to use for refining WARPED_CAUSAL motion vectors
950 // TODO(rachelbarker): Can this be unified with OBMC in some way?
951 WARP_SEARCH_METHOD warp_search_method;
952
953 // Maximum number of iterations in WARPED_CAUSAL refinement search
954 int warp_search_iters;
955} MV_SPEED_FEATURES;
956
957typedef struct INTER_MODE_SPEED_FEATURES {
958 // 2-pass inter mode model estimation where the preliminary pass skips
959 // transform search and uses a model to estimate rd, while the final pass
960 // computes the full transform search. Two types of models are supported:
961 // 0: not used
962 // 1: used with online dynamic rd model
963 // 2: used with static rd model
964 int inter_mode_rd_model_estimation;
965
966 // Bypass transform search based on skip rd at following stages
967 // i. Compound type mode search
968 // ii. Motion mode search (mode evaluation and winner motion mode stage)
969 // iii. Transform search for best inter candidates
970 int txfm_rd_gate_level[TX_SEARCH_CASES];
971
972 // Limit the inter mode tested in the RD loop
973 int reduce_inter_modes;
974
975 // This variable is used to cap the maximum number of times we skip testing a
976 // mode to be evaluated. A high value means we will be faster.
977 int adaptive_rd_thresh;
978
979 // Aggressively prune inter modes when best mode is skippable.
980 int prune_inter_modes_if_skippable;
981
982 // Drop less likely to be picked reference frames in the RD search.
983 // Has seven levels for now: 0, 1, 2, 3, 4, 5 and 6 where higher levels prune
984 // more aggressively than lower ones. (0 means no pruning).
985 int selective_ref_frame;
986
987 // Prune reference frames for rectangular partitions.
988 // 0 implies no pruning
989 // 1 implies prune for extended partition
990 // 2 implies prune horiz, vert and extended partition
991 int prune_ref_frame_for_rect_partitions;
992
993 // Prune inter modes w.r.t past reference frames
994 // 0 no pruning
995 // 1 prune inter modes w.r.t ALTREF2 and ALTREF reference frames
996 // 2 prune inter modes w.r.t BWDREF, ALTREF2 and ALTREF reference frames
997 int alt_ref_search_fp;
998
999 // Prune reference frames for single prediction modes based on temporal
1000 // distance and pred MV SAD. Feasible values are 0, 1, 2. The feature is
1001 // disabled for 0. An increasing value indicates more aggressive pruning
1002 // threshold.
1003 int prune_single_ref;
1004
1005 // Prune compound reference frames
1006 // 0 no pruning
1007 // 1 prune compound references which do not satisfy the two conditions:
1008 // a) The references are at a nearest distance from the current frame in
1009 // both past and future direction.
1010 // b) The references have minimum pred_mv_sad in both past and future
1011 // direction.
1012 // 2 prune compound references except the one with nearest distance from the
1013 // current frame in both past and future direction.
1014 int prune_comp_ref_frames;
1015
1016 // Skip the current ref_mv in NEW_MV mode based on mv, rate cost, etc.
1017 // This speed feature equaling 0 means no skipping.
1018 // If the speed feature equals 1 or 2, skip the current ref_mv in NEW_MV mode
1019 // if we have already encountered ref_mv in the drl such that:
1020 // 1. The other drl has the same mv during the SIMPLE_TRANSLATION search
1021 // process as the current mv.
1022 // 2. The rate needed to encode the current mv is larger than that for the
1023 // other ref_mv.
1024 // The speed feature equaling 1 means using subpel mv in the comparison.
1025 // The speed feature equaling 2 means using fullpel mv in the comparison.
1026 // If the speed feature >= 3, skip the current ref_mv in NEW_MV mode based on
1027 // known full_mv bestsme and drl cost.
1028 int skip_newmv_in_drl;
1029
1030 // This speed feature checks duplicate ref MVs among NEARESTMV, NEARMV,
1031 // GLOBALMV and skips NEARMV or GLOBALMV (in order) if a duplicate is found
1032 // TODO(any): Instead of skipping repeated ref mv, use the recalculated
1033 // rd-cost based on mode rate and skip the mode evaluation
1034 int skip_repeated_ref_mv;
1035
1036 // Flag used to control the ref_best_rd based gating for chroma
1037 int perform_best_rd_based_gating_for_chroma;
1038
1039 // Reuse the inter_intra_mode search result from NEARESTMV mode to other
1040 // single ref modes
1041 int reuse_inter_intra_mode;
1042
1043 // prune wedge and compound segment approximate rd evaluation based on
1044 // compound average modeled rd
1045 int prune_comp_type_by_model_rd;
1046
1047 // prune wedge and compound segment approximate rd evaluation based on
1048 // compound average rd/ref_best_rd
1049 int prune_comp_type_by_comp_avg;
1050
1051 // Skip some ref frames in compound motion search by single motion search
1052 // result. Has three levels for now: 0 referring to no skipping, and 1 - 3
1053 // increasing aggressiveness of skipping in order.
1054 // Note: The search order might affect the result. It assumes that the single
1055 // reference modes are searched before compound modes. It is better to search
1056 // same single inter mode as a group.
1057 int prune_comp_search_by_single_result;
1058
1059 // Instead of performing a full MV search, do a simple translation first
1060 // and only perform a full MV search on the motion vectors that performed
1061 // well.
1062 int prune_mode_search_simple_translation;
1063
1064 // Only search compound modes with at least one "good" reference frame.
1065 // A reference frame is good if, after looking at its performance among
1066 // the single reference modes, it is one of the two best performers.
1067 int prune_compound_using_single_ref;
1068
1069 // Skip extended compound mode (NEAREST_NEWMV, NEW_NEARESTMV, NEAR_NEWMV,
1070 // NEW_NEARMV) using ref frames of above and left neighbor
1071 // blocks.
1072 // 0 : no pruning
1073 // 1 : prune ext compound modes using neighbor blocks (less aggressiveness)
1074 // 2 : prune ext compound modes using neighbor blocks (high aggressiveness)
1075 // 3 : prune ext compound modes unconditionally (highest aggressiveness)
1076 int prune_ext_comp_using_neighbors;
1077
1078 // Skip NEW_NEARMV and NEAR_NEWMV extended compound modes
1079 int skip_ext_comp_nearmv_mode;
1080
1081 // Skip extended compound mode when ref frame corresponding to NEWMV does not
1082 // have NEWMV as single mode winner.
1083 // 0 : no pruning
1084 // 1 : prune extended compound mode (less aggressiveness)
1085 // 2 : prune extended compound mode (high aggressiveness)
1086 int prune_comp_using_best_single_mode_ref;
1087
1088 // Skip NEARESTMV and NEARMV using weight computed in ref mv list population
1089 //
1090 // Pruning is enabled only when both the top and left neighbor blocks are
1091 // available and when the current block already has a valid inter prediction.
1092 int prune_nearest_near_mv_using_refmv_weight;
1093
1094 // Based on previous ref_mv_idx search result, prune the following search.
1095 int prune_ref_mv_idx_search;
1096
1097 // Disable one sided compound modes.
1098 int disable_onesided_comp;
1099
1100 // Prune obmc search using previous frame stats.
1101 // INT_MAX : disable obmc search
1102 int prune_obmc_prob_thresh;
1103
1104 // Prune warped motion search using previous frame stats.
1105 int prune_warped_prob_thresh;
1106
1107 // Variance threshold to enable/disable Interintra wedge search
1108 unsigned int disable_interintra_wedge_var_thresh;
1109
1110 // Variance threshold to enable/disable Interinter wedge search
1111 unsigned int disable_interinter_wedge_var_thresh;
1112
1113 // De-couple wedge and mode search during interintra RDO.
1114 int fast_interintra_wedge_search;
1115
1116 // Whether fast wedge sign estimate is used
1117 int fast_wedge_sign_estimate;
1118
1119 // Enable/disable ME for interinter wedge search.
1120 int disable_interinter_wedge_newmv_search;
1121
1122 // Decide when and how to use joint_comp.
1123 DIST_WTD_COMP_FLAG use_dist_wtd_comp_flag;
1124
1125 // Clip the frequency of updating the mv cost.
1126 INTERNAL_COST_UPDATE_TYPE mv_cost_upd_level;
1127
1128 // Clip the frequency of updating the coeff cost.
1129 INTERNAL_COST_UPDATE_TYPE coeff_cost_upd_level;
1130
1131 // Clip the frequency of updating the mode cost.
1132 INTERNAL_COST_UPDATE_TYPE mode_cost_upd_level;
1133
1134 // Prune inter modes based on tpl stats
1135 // 0 : no pruning
1136 // 1 - 3 indicate increasing aggressiveness in order.
1137 int prune_inter_modes_based_on_tpl;
1138
1139 // Skip NEARMV and NEAR_NEARMV modes using ref frames of above and left
1140 // neighbor blocks and qindex.
1141 PRUNE_NEARMV_LEVEL prune_nearmv_using_neighbors;
1142
1143 // Model based breakout after interpolation filter search
1144 // 0: no breakout
1145 // 1: use model based rd breakout
1146 int model_based_post_interp_filter_breakout;
1147
1148 // Reuse compound type rd decision when exact match is found
1149 // 0: No reuse
1150 // 1: Reuse the compound type decision
1151 int reuse_compound_type_decision;
1152
1153 // Enable/disable masked compound.
1154 int disable_masked_comp;
1155
1156 // Enable/disable MV refinement for compound modes corresponds to compound
1157 // types COMPOUND_AVERAGE, COMPOUND_DISTWTD (currently, this compound type
1158 // is disabled for speeds >= 2 using the sf 'use_dist_wtd_comp_flag') and
1159 // COMPOUND_DIFFWTD based on the availability. Levels 0 to 3 indicate
1160 // increasing order of aggressiveness to disable MV refinement.
1161 // 0: MV Refinement is enabled and for NEW_NEWMV mode used two iterations of
1162 // refinement in av1_joint_motion_search().
1163 // 1: MV Refinement is disabled for COMPOUND_DIFFWTD and enabled for
1164 // COMPOUND_AVERAGE & COMPOUND_DISTWTD.
1165 // 2: MV Refinement is enabled for COMPOUND_AVERAGE & COMPOUND_DISTWTD for
1166 // NEW_NEWMV mode with one iteration of refinement in
1167 // av1_joint_motion_search() and MV Refinement is disabled for other compound
1168 // type modes.
1169 // 3: MV Refinement is disabled.
1170 int enable_fast_compound_mode_search;
1171
1172 // Reuse masked compound type search results
1173 int reuse_mask_search_results;
1174
1175 // Enable/disable fast search for wedge masks
1176 int enable_fast_wedge_mask_search;
1177
1178 // Early breakout from transform search of inter modes
1179 int inter_mode_txfm_breakout;
1180
1181 // Limit number of inter modes for txfm search if a newmv mode gets
1182 // evaluated among the top modes.
1183 // 0: no pruning
1184 // 1 to 3 indicate increasing order of aggressiveness
1185 int limit_inter_mode_cands;
1186
1187 // Cap the no. of txfm searches for a given prediction mode.
1188 // 0: no cap, 1: cap beyond first 4 searches, 2: cap beyond first 3 searches.
1189 int limit_txfm_eval_per_mode;
1190
1191 // Prune warped motion search based on block size.
1192 int extra_prune_warped;
1193
1194 // Do not search compound modes for ARF.
1195 // The intuition is that ARF is predicted by frames far away from it,
1196 // whose temporal correlations with the ARF are likely low.
1197 // It is therefore likely that compound modes do not work as well for ARF
1198 // as other inter frames.
1199 // Speed/quality impact:
1200 // Speed 1: 12% faster, 0.1% psnr loss.
1201 // Speed 2: 2% faster, 0.05% psnr loss.
1202 // No change for speed 3 and up, because |disable_onesided_comp| is true.
1203 int skip_arf_compound;
1204
1205 // Percentage of scaling used to increase the rd cost of warp mode so that
1206 // encoder decisions are biased against local warp, favoring low complexity
1207 // modes.
1208 int bias_warp_mode_rd_scale_pct;
1209} INTER_MODE_SPEED_FEATURES;
1210
1211typedef struct INTERP_FILTER_SPEED_FEATURES {
1212 // Do limited interpolation filter search for dual filters, since best choice
1213 // usually includes EIGHTTAP_REGULAR.
1214 int use_fast_interpolation_filter_search;
1215
1216 // Disable dual filter
1217 int disable_dual_filter;
1218
1219 // Save results of av1_interpolation_filter_search for a block
1220 // Check mv and ref_frames before search, if they are very close with previous
1221 // saved results, filter search can be skipped.
1222 int use_interp_filter;
1223
1224 // skip sharp_filter evaluation based on regular and smooth filter rd for
1225 // dual_filter=0 case
1226 int skip_sharp_interp_filter_search;
1227
1228 // skip interpolation filter search for a block in chessboard pattern
1229 int cb_pred_filter_search;
1230
1231 // adaptive interp_filter search to allow skip of certain filter types.
1232 int adaptive_interp_filter_search;
1233
1234 // Forces interpolation filter to EIGHTTAP_REGULAR and skips interpolation
1235 // filter search.
1236 int skip_interp_filter_search;
1237} INTERP_FILTER_SPEED_FEATURES;
1238
1239typedef struct INTRA_MODE_SPEED_FEATURES {
1240 // These bit masks allow you to enable or disable intra modes for each
1241 // transform size separately.
1242 int intra_y_mode_mask[TX_SIZES];
1243 int intra_uv_mode_mask[TX_SIZES];
1244
1245 // flag to allow skipping intra mode for inter frame prediction
1246 int skip_intra_in_interframe;
1247
1248 // Prune intra mode candidates based on source block histogram of gradient.
1249 // Applies to luma plane only.
1250 // Feasible values are 0..4. The feature is disabled for 0. An increasing
1251 // value indicates more aggressive pruning threshold.
1252 int intra_pruning_with_hog;
1253
1254 // Prune intra mode candidates based on source block histogram of gradient.
1255 // Applies to chroma plane only.
1256 // Feasible values are 0..4. The feature is disabled for 0. An increasing
1257 // value indicates more aggressive pruning threshold.
1258 int chroma_intra_pruning_with_hog;
1259
1260 // Enable/disable smooth intra modes.
1261 int disable_smooth_intra;
1262
1263 // Prune UV_SMOOTH_PRED mode for chroma based on chroma source variance.
1264 // false : No pruning
1265 // true : Prune UV_SMOOTH_PRED mode based on chroma source variance
1266 //
1267 // For allintra encode, this speed feature reduces instruction count
1268 // by 1.90%, 2.21% and 1.97% for speed 6, 7 and 8 with coding performance
1269 // change less than 0.04%. For AVIF image encode, this speed feature reduces
1270 // encode time by 1.56%, 2.14% and 0.90% for speed 6, 7 and 8 on a typical
1271 // image dataset with coding performance change less than 0.05%.
1272 bool prune_smooth_intra_mode_for_chroma;
1273
1274 // Prune filter intra modes in intra frames.
1275 // 0 : No pruning
1276 // 1 : Evaluate applicable filter intra modes based on best intra mode so far
1277 // 2 : Do not evaluate filter intra modes
1278 int prune_filter_intra_level;
1279
1280 // prune palette search
1281 // 0: No pruning
1282 // 1: Perform coarse search to prune the palette colors. For winner colors,
1283 // neighbors are also evaluated using a finer search.
1284 // 2: Perform 2 way palette search from max colors to min colors (and min
1285 // colors to remaining colors) and terminate the search if current number of
1286 // palette colors is not the winner.
1287 int prune_palette_search_level;
1288
1289 // Terminate early in luma palette_size search. Speed feature values indicate
1290 // increasing level of pruning.
1291 // 0: No early termination
1292 // 1: Terminate early for higher luma palette_size, if header rd cost of lower
1293 // palette_size is more than 2 * best_rd. This level of pruning is more
1294 // conservative when compared to sf level 2 as the cases which will get pruned
1295 // with sf level 1 is a subset of the cases which will get pruned with sf
1296 // level 2.
1297 // 2: Terminate early for higher luma palette_size, if header rd cost of lower
1298 // palette_size is more than best_rd.
1299 // For allintra encode, this sf reduces instruction count by 2.49%, 1.07%,
1300 // 2.76%, 2.30%, 1.84%, 2.69%, 2.04%, 2.05% and 1.44% for speed 0, 1, 2, 3, 4,
1301 // 5, 6, 7 and 8 on screen content set with coding performance change less
1302 // than 0.01% for speed <= 2 and less than 0.03% for speed >= 3. For AVIF
1303 // image encode, this sf reduces instruction count by 1.94%, 1.13%, 1.29%,
1304 // 0.93%, 0.89%, 1.03%, 1.07%, 1.20% and 0.18% for speed 0, 1, 2, 3, 4, 5, 6,
1305 // 7 and 8 on a typical image dataset with coding performance change less than
1306 // 0.01%.
1307 int prune_luma_palette_size_search_level;
1308
1309 // Prune chroma intra modes based on luma intra mode winner.
1310 // 0: No pruning
1311 // 1: Prune chroma intra modes other than UV_DC_PRED, UV_SMOOTH_PRED,
1312 // UV_CFL_PRED and the mode that corresponds to luma intra mode winner.
1313 int prune_chroma_modes_using_luma_winner;
1314
1315 // Clip the frequency of updating the mv cost for intrabc.
1316 INTERNAL_COST_UPDATE_TYPE dv_cost_upd_level;
1317
1318 // We use DCT_DCT transform followed by computing SATD (Sum of Absolute
1319 // Transformed Differences) as an estimation of RD score to quickly find the
1320 // best possible Chroma from Luma (CFL) parameter. Then we do a full RD search
1321 // near the best possible parameter. The search range is set here.
1322 // The range of cfl_searh_range should be [1, 33], and the following are the
1323 // recommended values.
1324 // 1: Fastest mode.
1325 // 3: Default mode that provides good speedup without losing compression
1326 // performance at speed 0.
1327 // 33: Exhaustive rd search (33 == CFL_MAGS_SIZE). This mode should only
1328 // be used for debugging purpose.
1329 int cfl_search_range;
1330
1331 // TOP_INTRA_MODEL_COUNT is 4 that is the number of top model rd to store in
1332 // intra mode decision. Here, add a speed feature to reduce this number for
1333 // higher speeds.
1334 int top_intra_model_count_allowed;
1335
1336 // Adapt top_intra_model_count_allowed locally to prune luma intra modes using
1337 // neighbor block and quantizer information.
1338 int adapt_top_model_rd_count_using_neighbors;
1339
1340 // Prune the evaluation of odd delta angles of directional luma intra modes by
1341 // using the rdcosts of neighbouring delta angles.
1342 // For allintra encode, this speed feature reduces instruction count
1343 // by 4.461%, 3.699% and 3.536% for speed 6, 7 and 8 on a typical video
1344 // dataset with coding performance change less than 0.26%. For AVIF image
1345 // encode, this speed feature reduces encode time by 2.849%, 2.471%,
1346 // and 2.051% for speed 6, 7 and 8 on a typical image dataset with coding
1347 // performance change less than 0.27%.
1348 int prune_luma_odd_delta_angles_in_intra;
1349
1350 // Terminate early in chroma palette_size search.
1351 // 0: No early termination
1352 // 1: Terminate early for higher palette_size, if header rd cost of lower
1353 // palette_size is more than best_rd.
1354 // For allintra encode, this sf reduces instruction count by 0.45%,
1355 // 0.62%, 1.73%, 2.50%, 2.89%, 3.09% and 3.86% for speed 0 to 6 on screen
1356 // content set with coding performance change less than 0.01%.
1357 // For AVIF image encode, this sf reduces instruction count by 0.45%, 0.81%,
1358 // 0.85%, 1.05%, 1.45%, 1.66% and 1.95% for speed 0 to 6 on a typical image
1359 // dataset with no quality drop.
1360 int early_term_chroma_palette_size_search;
1361
1362 // Skips the evaluation of filter intra modes in inter frames if rd evaluation
1363 // of luma intra dc mode results in invalid rd stats.
1364 int skip_filter_intra_in_inter_frames;
1365} INTRA_MODE_SPEED_FEATURES;
1366
1367typedef struct TX_SPEED_FEATURES {
1368 // Init search depth for square and rectangular transform partitions.
1369 // Values:
1370 // 0 - search full tree, 1: search 1 level, 2: search the highest level only
1371 int inter_tx_size_search_init_depth_sqr;
1372 int inter_tx_size_search_init_depth_rect;
1373 int intra_tx_size_search_init_depth_sqr;
1374 int intra_tx_size_search_init_depth_rect;
1375
1376 // If any dimension of a coding block size above 64, always search the
1377 // largest transform only, since the largest transform block size is 64x64.
1378 int tx_size_search_lgr_block;
1379
1380 TX_TYPE_SEARCH tx_type_search;
1381
1382 // Skip split transform block partition when the collocated bigger block
1383 // is selected as all zero coefficients.
1384 int txb_split_cap;
1385
1386 // Shortcut the transform block partition and type search when the target
1387 // rdcost is relatively lower.
1388 // Values are 0 (not used) , or 1 - 2 with progressively increasing
1389 // aggressiveness
1390 int adaptive_txb_search_level;
1391
1392 // Prune level for tx_size_type search for inter based on rd model
1393 // 0: no pruning
1394 // 1-2: progressively increasing aggressiveness of pruning
1395 int model_based_prune_tx_search_level;
1396
1397 // Refine TX type after fast TX search.
1398 int refine_fast_tx_search_results;
1399
1400 // Prune transform split/no_split eval based on residual properties. A value
1401 // of 0 indicates no pruning, and the aggressiveness of pruning progressively
1402 // increases from levels 1 to 3.
1403 int prune_tx_size_level;
1404
1405 // Prune the evaluation of transform depths as decided by the NN model.
1406 // false: No pruning.
1407 // true : Avoid the evaluation of specific transform depths using NN model.
1408 //
1409 // For allintra encode, this speed feature reduces instruction count
1410 // by 4.76%, 8.92% and 11.28% for speed 6, 7 and 8 with coding performance
1411 // change less than 0.32%. For AVIF image encode, this speed feature reduces
1412 // encode time by 4.65%, 9.16% and 10.45% for speed 6, 7 and 8 on a typical
1413 // image dataset with coding performance change less than 0.19%.
1414 bool prune_intra_tx_depths_using_nn;
1415
1416 // Enable/disable early breakout during transform search of intra modes, by
1417 // using the minimum rd cost possible. By using this approach, the rd
1418 // evaluation of applicable transform blocks (in the current block) can be
1419 // avoided as
1420 // 1) best_rd evolves during the search in choose_tx_size_type_from_rd()
1421 // 2) appropriate ref_best_rd is passed in intra_block_yrd()
1422 //
1423 // For allintra encode, this speed feature reduces instruction count
1424 // by 1.11%, 1.08%, 1.02% and 0.93% for speed 3, 6, 7 and 8 with coding
1425 // performance change less than 0.02%. For AVIF image encode, this speed
1426 // feature reduces encode time by 0.93%, 1.46%, 1.07%, 0.84%, 0.99% and 0.73%
1427 // for speed 3, 4, 5, 6, 7 and 8 on a typical image dataset with coding
1428 // performance change less than 0.004%.
1429 bool use_rd_based_breakout_for_intra_tx_search;
1430} TX_SPEED_FEATURES;
1431
1432typedef struct RD_CALC_SPEED_FEATURES {
1433 // Fast approximation of av1_model_rd_from_var_lapndz
1434 int simple_model_rd_from_var;
1435
1436 // Perform faster distortion computation during the R-D evaluation by trying
1437 // to approximate the prediction error with transform coefficients (faster but
1438 // less accurate) rather than computing distortion in the pixel domain (slower
1439 // but more accurate). The following methods are used for distortion
1440 // computation:
1441 // Method 0: Always compute distortion in the pixel domain
1442 // Method 1: Based on block error, try using transform domain distortion for
1443 // tx_type search and compute distortion in pixel domain for final RD_STATS
1444 // Method 2: Based on block error, try to compute distortion in transform
1445 // domain
1446 // Methods 1 and 2 may fallback to computing distortion in the pixel domain in
1447 // case the block error is less than the threshold, which is controlled by the
1448 // speed feature tx_domain_dist_thres_level.
1449 //
1450 // The speed feature tx_domain_dist_level decides which of the above methods
1451 // needs to be used across different mode evaluation stages as described
1452 // below:
1453 // Eval type: Default Mode Winner
1454 // Level 0 : Method 0 Method 2 Method 0
1455 // Level 1 : Method 1 Method 2 Method 0
1456 // Level 2 : Method 2 Method 2 Method 0
1457 // Level 3 : Method 2 Method 2 Method 2
1458 int tx_domain_dist_level;
1459
1460 // Transform domain distortion threshold level
1461 int tx_domain_dist_thres_level;
1462
1463 // Trellis (dynamic programming) optimization of quantized values
1464 TRELLIS_OPT_TYPE optimize_coefficients;
1465
1466 // Use hash table to store macroblock RD search results
1467 // to avoid repeated search on the same residue signal.
1468 int use_mb_rd_hash;
1469
1470 // Flag used to control the extent of coeff R-D optimization
1471 int perform_coeff_opt;
1472} RD_CALC_SPEED_FEATURES;
1473
1474typedef struct WINNER_MODE_SPEED_FEATURES {
1475 // Flag used to control the winner mode processing for better R-D optimization
1476 // of quantized coeffs
1477 int enable_winner_mode_for_coeff_opt;
1478
1479 // Flag used to control the winner mode processing for transform size
1480 // search method
1481 int enable_winner_mode_for_tx_size_srch;
1482
1483 // Control transform size search level
1484 // Eval type: Default Mode Winner
1485 // Level 0 : FULL RD LARGEST ALL FULL RD
1486 // Level 1 : FAST RD LARGEST ALL FULL RD
1487 // Level 2 : LARGEST ALL LARGEST ALL FULL RD
1488 // Level 3 : LARGEST ALL LARGEST ALL LARGEST ALL
1489 int tx_size_search_level;
1490
1491 // Flag used to control the winner mode processing for use transform
1492 // domain distortion
1493 int enable_winner_mode_for_use_tx_domain_dist;
1494
1495 // Flag used to enable processing of multiple winner modes
1496 MULTI_WINNER_MODE_TYPE multi_winner_mode_type;
1497
1498 // Motion mode for winner candidates:
1499 // 0: speed feature OFF
1500 // 1 / 2 : Use configured number of winner candidates
1501 int motion_mode_for_winner_cand;
1502
1503 // Controls the prediction of transform skip block or DC only block.
1504 //
1505 // Different speed feature values (0 to 3) decide the aggressiveness of
1506 // prediction (refer to predict_dc_levels[][] in speed_features.c) to be used
1507 // during different mode evaluation stages.
1508 int dc_blk_pred_level;
1509
1510 // If on, disables interpolation filter search in handle_inter_mode loop, and
1511 // performs it during winner mode processing by \ref
1512 // tx_search_best_inter_candidates.
1513 int winner_mode_ifs;
1514
1515 // Controls the disabling of winner mode processing. Speed feature levels
1516 // are ordered in increasing aggressiveness of pruning. The method considered
1517 // for disabling, depends on the sf level value and it is described as below.
1518 // 0: Do not disable
1519 // 1: Disable for blocks with low source variance.
1520 // 2: Disable for blocks which turn out to be transform skip (skipped based on
1521 // eob) during MODE_EVAL stage except NEWMV mode.
1522 // 3: Disable for blocks which turn out to be transform skip during MODE_EVAL
1523 // stage except NEWMV mode. For high quantizers, prune conservatively based on
1524 // transform skip (skipped based on eob) except for NEWMV mode.
1525 // 4: Disable for blocks which turn out to be transform skip during MODE_EVAL
1526 // stage.
1527 int prune_winner_mode_eval_level;
1528} WINNER_MODE_SPEED_FEATURES;
1529
1530typedef struct LOOP_FILTER_SPEED_FEATURES {
1531 // This feature controls how the loop filter level is determined.
1532 LPF_PICK_METHOD lpf_pick;
1533
1534 // Skip some final iterations in the determination of the best loop filter
1535 // level.
1536 int use_coarse_filter_level_search;
1537
1538 // Reset luma filter levels to zero based on minimum filter levels of
1539 // reference frames and current frame's pyramid level.
1540 int adaptive_luma_loop_filter_skip;
1541
1542 // Reset luma filter levels to zero when the percentage of SSE difference
1543 // between the unfiltered and filtered versions of the current frame is below
1544 // a threshold.
1545 int skip_loop_filter_using_filt_error;
1546
1547 // Control how the CDEF strength is determined.
1548 CDEF_PICK_METHOD cdef_pick_method;
1549
1550 // Decoder side speed feature to add penalty for use of dual-sgr filters.
1551 // Takes values 0 - 10, 0 indicating no penalty and each additional level
1552 // adding a penalty of 1%
1553 int dual_sgr_penalty_level;
1554
1555 // Restricts loop restoration to RESTORE_SWITCHABLE by skipping RD cost
1556 // comparisons for RESTORE_WIENER and RESTORE_SGRPROJ. Also applies a bias
1557 // during switchable restoration search: each level adds a 0.5% penalty to
1558 // Wiener and SGR selection.
1559 // 0 : No restriction or bias (all restoration types allowed)
1560 // 1+: Skip WIENER/SGRPROJ and apply (level x 0.5%) penalty in
1561 // search_switchable()
1562 int switchable_lr_with_bias_level;
1563
1564 // prune sgr ep using binary search like mechanism
1565 int enable_sgr_ep_pruning;
1566
1567 // Disable loop restoration for Chroma plane
1568 int disable_loop_restoration_chroma;
1569
1570 // Disable loop restoration for luma plane
1571 int disable_loop_restoration_luma;
1572
1573 // Range of loop restoration unit sizes to search
1574 // The minimum size is clamped against the superblock size in
1575 // av1_pick_filter_restoration, so that the code which sets this value does
1576 // not need to know the superblock size ahead of time.
1577 int min_lr_unit_size;
1578 int max_lr_unit_size;
1579
1580 // Prune RESTORE_WIENER evaluation based on source variance
1581 // 0 : no pruning
1582 // 1 : conservative pruning
1583 // 2 : aggressive pruning
1584 int prune_wiener_based_on_src_var;
1585
1586 // Prune self-guided loop restoration based on wiener search results
1587 // 0 : no pruning
1588 // 1 : pruning based on rdcost ratio of RESTORE_WIENER and RESTORE_NONE
1589 // 2 : pruning based on winner restoration type among RESTORE_WIENER and
1590 // RESTORE_NONE
1591 int prune_sgr_based_on_wiener;
1592
1593 // Reduce the wiener filter win size for luma
1594 int reduce_wiener_window_size;
1595
1596 // Flag to disable Wiener Loop restoration filter.
1597 bool disable_wiener_filter;
1598
1599 // Flag to disable Self-guided Loop restoration filter.
1600 bool disable_sgr_filter;
1601
1602 // Disable the refinement search around the wiener filter coefficients.
1603 bool disable_wiener_coeff_refine_search;
1604
1605 // Whether to downsample the rows in computation of wiener stats.
1606 int use_downsampled_wiener_stats;
1607} LOOP_FILTER_SPEED_FEATURES;
1608
1609typedef struct REAL_TIME_SPEED_FEATURES {
1610 // check intra prediction for non-RD mode.
1611 int check_intra_pred_nonrd;
1612
1613 // Skip checking intra prediction.
1614 // 0 - don't skip
1615 // 1 - skip if TX is skipped and best mode is not NEWMV
1616 // 2 - skip if TX is skipped
1617 // Skipping aggressiveness increases from level 1 to 2.
1618 int skip_intra_pred;
1619
1620 // Estimate motion before calculating variance in variance-based partition
1621 // 0 - Only use zero MV
1622 // 1 - perform coarse ME
1623 // 2 - perform coarse ME, and also use neighbours' MVs
1624 // 3 - use neighbours' MVs without performing coarse ME
1625 int estimate_motion_for_var_based_partition;
1626
1627 // For nonrd_use_partition: mode of extra check of leaf partition
1628 // 0 - don't check merge
1629 // 1 - always check merge
1630 // 2 - check merge and prune checking final split
1631 // 3 - check merge and prune checking final split based on bsize and qindex
1632 int nonrd_check_partition_merge_mode;
1633
1634 // For nonrd_use_partition: check of leaf partition extra split
1635 int nonrd_check_partition_split;
1636
1637 // Implements various heuristics to skip searching modes
1638 // The heuristics selected are based on flags
1639 // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
1640 unsigned int mode_search_skip_flags;
1641
1642 // For nonrd: Reduces ref frame search.
1643 // 0 - low level of search prune in non last frames
1644 // 1 - pruned search in non last frames
1645 // 2 - more pruned search in non last frames
1646 int nonrd_prune_ref_frame_search;
1647
1648 // This flag controls the use of non-RD mode decision.
1649 int use_nonrd_pick_mode;
1650
1651 // Flag that controls discounting for color map cost during palette search.
1652 // This saves about 5% of CPU and in non-RD speeds delivers better results
1653 // across rtc_screen set (on speed 10 overall BDRate growth is 13%)
1654 int discount_color_cost;
1655
1656 // Use ALTREF frame in non-RD mode decision.
1657 int use_nonrd_altref_frame;
1658
1659 // Use compound reference for non-RD mode.
1660 int use_comp_ref_nonrd;
1661
1662 // Reference frames for compound prediction for nonrd pickmode:
1663 // LAST_GOLDEN (0), LAST_LAST2 (1), or LAST_ALTREF (2).
1664 int ref_frame_comp_nonrd[3];
1665
1666 // use reduced ref set for real-time mode
1667 int use_real_time_ref_set;
1668
1669 // Skip a number of expensive mode evaluations for blocks with very low
1670 // temporal variance.
1671 int short_circuit_low_temp_var;
1672
1673 // Reuse inter prediction in fast non-rd mode.
1674 int reuse_inter_pred_nonrd;
1675
1676 // Number of best inter modes to search transform. INT_MAX - search all.
1677 int num_inter_modes_for_tx_search;
1678
1679 // Use interpolation filter search in non-RD mode decision.
1680 int use_nonrd_filter_search;
1681
1682 // Use simplified RD model for interpolation search and Intra
1683 int use_simple_rd_model;
1684
1685 // For nonrd mode: use hybrid intra mode search for intra only frames based on
1686 // block properties.
1687 // 0 : use nonrd pick intra for all blocks
1688 // 1 : use rd for bsize < 16x16, nonrd otherwise
1689 // 2 : use rd for bsize < 16x16 and src var >= 101, nonrd otherwise
1690 int hybrid_intra_pickmode;
1691
1692 // Filter blocks by certain criteria such as SAD, source variance, such that
1693 // fewer blocks will go through the palette search.
1694 // For nonrd encoding path, enable this feature reduces encoding time when
1695 // palette mode is used. Disabling it leads to better compression efficiency.
1696 // 0: off
1697 // 1: less aggressive pruning mode
1698 // 2, 3: more aggressive pruning mode
1699 int prune_palette_search_nonrd;
1700
1701 // Compute variance/sse on source difference, prior to encoding superblock.
1702 int source_metrics_sb_nonrd;
1703
1704 // Flag to indicate process for handling overshoot on slide/scene change,
1705 // for real-time CBR mode.
1706 OVERSHOOT_DETECTION_CBR overshoot_detection_cbr;
1707
1708 // Check for scene/content change detection on every frame before encoding.
1709 int check_scene_detection;
1710
1711 // For keyframes in rtc: adjust the rc_bits_per_mb, to reduce overshoot.
1712 int rc_adjust_keyframe;
1713
1714 // On scene change: compute spatial variance.
1715 int rc_compute_spatial_var_sc;
1716
1717 // For nonrd mode: Prefer larger partition blks in variance based partitioning
1718 // 0: disabled, 1-3: increasing aggressiveness
1719 int prefer_large_partition_blocks;
1720
1721 // uses results of temporal noise estimate
1722 int use_temporal_noise_estimate;
1723
1724 // Parameter indicating initial search window to be used in full-pixel search
1725 // for nonrd_pickmode. Range [0, MAX_MVSEARCH_STEPS - 1]. Lower value
1726 // indicates larger window. If set to 0, step_param is set based on internal
1727 // logic in set_mv_search_params().
1728 int fullpel_search_step_param;
1729
1730 // Bit mask to enable or disable intra modes for each prediction block size
1731 // separately, for nonrd_pickmode. Currently, the sf is not respected when
1732 // 'force_intra_check' is true in 'av1_estimate_intra_mode()' function. Also,
1733 // H and V pred modes allowed through this sf can be further pruned when
1734 //'prune_hv_pred_modes_using_src_sad' sf is true.
1735 int intra_y_mode_bsize_mask_nrd[BLOCK_SIZES];
1736
1737 // Prune H and V intra predition modes evalution in inter frame.
1738 // The sf does not have any impact.
1739 // i. when frame_source_sad is 1.1 times greater than avg_source_sad
1740 // ii. when cyclic_refresh_segment_id_boosted is enabled
1741 // iii. when SB level source sad is greater than kMedSad
1742 // iv. when color sensitivity is non zero for both the chroma channels
1743 bool prune_hv_pred_modes_using_src_sad;
1744
1745 // Skips mode checks more aggressively in nonRD mode
1746 int nonrd_aggressive_skip;
1747
1748 // Skip cdef on 64x64 blocks/
1749 // 0: disabled
1750 // 1: skip when NEWMV or INTRA is not picked or color sensitivity is off.
1751 // When color sensitivity is on for a superblock, all 64x64 blocks within
1752 // will not skip.
1753 // 2: more aggressive mode where skip is done for all frames where
1754 // rc->high_source_sad = 0 (non slide-changes), and color sensitivity off.
1755 int skip_cdef_sb;
1756
1757 // Force selective cdf update.
1758 int selective_cdf_update;
1759
1760 // Force only single reference (LAST) for prediction.
1761 int force_only_last_ref;
1762
1763 // Forces larger partition blocks in variance based partitioning for intra
1764 // frames
1765 int force_large_partition_blocks_intra;
1766
1767 // Use fixed partition for superblocks based on source_sad.
1768 // 0: disabled
1769 // 1: enabled
1770 int use_fast_fixed_part;
1771
1772 // Increase source_sad thresholds in nonrd pickmode.
1773 int increase_source_sad_thresh;
1774
1775 // Skip evaluation of no split in tx size selection for merge partition
1776 int skip_tx_no_split_var_based_partition;
1777
1778 // Intermediate termination of newMV mode evaluation based on so far best mode
1779 // sse
1780 int skip_newmv_mode_based_on_sse;
1781
1782 // Define gf length multiplier.
1783 // Level 0: use large multiplier, level 1: use medium multiplier.
1784 int gf_length_lvl;
1785
1786 // Prune inter modes with golden frame as reference for NEARMV and NEWMV modes
1787 int prune_inter_modes_with_golden_ref;
1788
1789 // Prune inter modes w.r.t golden or alt-ref frame based on sad
1790 int prune_inter_modes_wrt_gf_arf_based_on_sad;
1791
1792 // Prune inter mode search in rd path based on current block's temporal
1793 // variance wrt LAST reference.
1794 int prune_inter_modes_using_temp_var;
1795
1796 // Reduce MV precision to halfpel for higher int MV value & frame-level motion
1797 // 0: disabled
1798 // 1-2: Reduce precision to halfpel, fullpel based on conservative
1799 // thresholds, aggressiveness increases with increase in level
1800 // 3: Reduce precision to halfpel using more aggressive thresholds
1801 int reduce_mv_pel_precision_highmotion;
1802
1803 // Reduce MV precision for low complexity blocks
1804 // 0: disabled
1805 // 1: Reduce the mv resolution for zero mv if the variance is low
1806 // 2: Switch to halfpel, fullpel based on low block spatial-temporal
1807 // complexity.
1808 int reduce_mv_pel_precision_lowcomplex;
1809
1810 // Prune intra mode evaluation in inter frames based on mv range.
1811 BLOCK_SIZE prune_intra_mode_based_on_mv_range;
1812 // The number of times to left shift the splitting thresholds in variance
1813 // based partitioning. The minimum values should be 7 to avoid left shifting
1814 // by a negative number.
1815 int var_part_split_threshold_shift;
1816
1817 // Qindex based variance partition threshold index, which determines
1818 // the aggressiveness of partition pruning
1819 // 0: disabled for speeds 9,10
1820 // 1,2: (rd-path) lowers qindex thresholds conditionally (for low SAD sb)
1821 // 3,4: (non-rd path) uses pre-tuned qindex thresholds
1822 int var_part_based_on_qidx;
1823
1824 // Enable GF refresh based on Q value.
1825 int gf_refresh_based_on_qp;
1826
1827 // Temporal filtering
1828 // The value can be 1 or 2, which indicates the threshold to use.
1829 // Must be off for lossless mode.
1830 int use_rtc_tf;
1831
1832 // Use of the identity transform in nonrd_pickmode,
1833 int use_idtx_nonrd;
1834
1835 // Prune the use of the identity transform in nonrd_pickmode:
1836 // only for smaller blocks and higher spatial variance, and when skip_txfm
1837 // is not already set.
1838 int prune_idtx_nonrd;
1839
1840 // Force to only use dct for palette search in nonrd pickmode.
1841 int dct_only_palette_nonrd;
1842
1843 // Skip loopfilter, for static content after slide change
1844 // or key frame, once quality has ramped up.
1845 // 0: disabled
1846 // 1: skip only after quality is ramped up.
1847 // 2: aggrssive mode, where skip is done for all frames that
1848 // where rc->high_source_sad = 0 (no slide-changes).
1849 int skip_lf_screen;
1850
1851 // Threshold on the active/inactive region percent to disable
1852 // the loopfilter and cdef. Setting to 100 disables this feature.
1853 int thresh_active_maps_skip_lf_cdef;
1854
1855 // For nonrd: early exit out of variance partition that sets the
1856 // block size to superblock size, and sets mode to zeromv-last skip.
1857 // 0: disabled
1858 // 1: zeromv-skip is enabled at SB level only
1859 // 2: zeromv-skip is enabled at SB level and coding block level
1860 int part_early_exit_zeromv;
1861
1862 // Early terminate inter mode search based on sse in non-rd path.
1863 INTER_SEARCH_EARLY_TERM_IDX sse_early_term_inter_search;
1864
1865 // SAD based adaptive altref selection
1866 int sad_based_adp_altref_lag;
1867
1868 // Enable/disable partition direct merging.
1869 int partition_direct_merging;
1870
1871 // Level of aggressiveness for obtaining tx size based on qstep
1872 int tx_size_level_based_on_qstep;
1873
1874 // Avoid the partitioning of a 16x16 block in variance based partitioning
1875 // (VBP) by making use of minimum and maximum sub-block variances.
1876 // For allintra encode, this speed feature reduces instruction count by 5.39%
1877 // for speed 9 on a typical video dataset with coding performance gain
1878 // of 1.44%.
1879 // For AVIF image encode, this speed feature reduces encode time
1880 // by 8.44% for speed 9 on a typical image dataset with coding performance
1881 // gain of 0.78%.
1882 bool vbp_prune_16x16_split_using_min_max_sub_blk_var;
1883
1884 // A qindex threshold that determines whether to use qindex based CDEF filter
1885 // strength estimation for screen content types. The strength estimation model
1886 // used for screen contents prefers to allow cdef filtering for more frames.
1887 // This sf is used to limit the frames which go through cdef filtering and
1888 // following explains the setting of the same.
1889 // MAXQ (255): This disables the usage of this sf. Here, frame does not use a
1890 // screen content model thus reduces the number of frames that go through cdef
1891 // filtering.
1892 // MINQ (0): Frames always use screen content model thus increasing the number
1893 // of frames that go through cdef filtering.
1894 // This speed feature has a substantial gain on coding metrics, with moderate
1895 // increase encoding time. Select threshold based on speed vs quality
1896 // trade-off.
1897 int screen_content_cdef_filter_qindex_thresh;
1898
1899 // Prune compound mode if its variance is higher than the variance of single
1900 // modes.
1901 bool prune_compoundmode_with_singlecompound_var;
1902
1903 // Allow mode cost update at frame level every couple frames. This
1904 // overrides the command line setting --mode-cost-upd-freq=3 (never update
1905 // except on key frame and first delta).
1906 bool frame_level_mode_cost_update;
1907
1908 // Prune H_PRED during intra mode evaluation in the nonrd path based on best
1909 // mode so far.
1910 //
1911 // For allintra encode, this speed feature reduces instruction count by 1.10%
1912 // for speed 9 with coding performance change less than 0.04%.
1913 // For AVIF image encode, this speed feature reduces encode time by 1.03% for
1914 // speed 9 on a typical image dataset with coding performance change less than
1915 // 0.08%.
1916 bool prune_h_pred_using_best_mode_so_far;
1917
1918 // Enable pruning of intra mode evaluations in nonrd path based on source
1919 // variance and best mode so far. The pruning logic is enabled only if the
1920 // mode is not a winner mode of both the neighboring blocks (left/top).
1921 //
1922 // For allintra encode, this speed feature reduces instruction count by 3.96%
1923 // for speed 9 with coding performance change less than 0.38%.
1924 // For AVIF image encode, this speed feature reduces encode time by 3.46% for
1925 // speed 9 on a typical image dataset with coding performance change less than
1926 // -0.06%.
1927 bool enable_intra_mode_pruning_using_neighbors;
1928
1929 // Prune intra mode evaluations in nonrd path based on best sad so far.
1930 //
1931 // For allintra encode, this speed feature reduces instruction count by 3.05%
1932 // for speed 9 with coding performance change less than 0.24%.
1933 // For AVIF image encode, this speed feature reduces encode time by 1.87% for
1934 // speed 9 on a typical image dataset with coding performance change less than
1935 // 0.16%.
1936 bool prune_intra_mode_using_best_sad_so_far;
1937
1938 // If compound is enabled, and the current block size is \geq BLOCK_16X16,
1939 // limit the compound modes to GLOBAL_GLOBALMV. This does not apply to the
1940 // base layer of svc.
1941 bool check_only_zero_zeromv_on_large_blocks;
1942
1943 // Allow for disabling cdf update for non reference frames in svc mode.
1944 bool disable_cdf_update_non_reference_frame;
1945
1946 // Prune compound modes if the single modes variances do not perform well.
1947 bool prune_compoundmode_with_singlemode_var;
1948
1949 // Skip searching all compound mode if the variance of single_mode residue is
1950 // sufficiently low.
1951 bool skip_compound_based_on_var;
1952
1953 // Sets force_zeromv_skip based on the source sad available. Aggressiveness
1954 // increases with increase in the level set for speed feature.
1955 // 0: No setting
1956 // 1: If source sad is kZeroSad
1957 // 2: If source sad <= kVeryLowSad
1958 int set_zeromv_skip_based_on_source_sad;
1959
1960 // Downgrades the block-level subpel motion search to
1961 // av1_find_best_sub_pixel_tree_pruned_more for higher QP and when fullpel
1962 // search performed well, zeromv has low sad or low source_var
1963 bool use_adaptive_subpel_search;
1964
1965 // A flag used in RTC case to control frame_refs_short_signaling. Note that
1966 // the final decision is made in check_frame_refs_short_signaling(). The flag
1967 // can only be turned on when res < 360p and speed >= 9, in which case only
1968 // LAST and GOLDEN ref frames are used now.
1969 bool enable_ref_short_signaling;
1970
1971 // A flag that controls if we check or bypass GLOBALMV in rtc single ref frame
1972 // case.
1973 bool check_globalmv_on_single_ref;
1974
1975 // Allows for increasing the color_threshold for palette prediction.
1976 // This generally leads to better coding efficiency but with some speed loss.
1977 // Only used for screen content and for nonrd_pickmode.
1978 bool increase_color_thresh_palette;
1979
1980 // Flag to indicate selecting of higher threshold for scenee change detection.
1981 int higher_thresh_scene_detection;
1982
1983 // FLag to indicate skip testing of NEWMV for flat blocks.
1984 int skip_newmv_flat_blocks_screen;
1985
1986 // Flag to force skip encoding for non_reference_frame on slide/scene changes.
1987 int skip_encoding_non_reference_slide_change;
1988
1989 // Flag to indicate more aggressive QP downward adjustment for screen static
1990 // content, to make convergence to min_qp faster.
1991 int rc_faster_convergence_static;
1992
1993 // Skip NEWMV mode evaluation based on sad for screen content.
1994 int skip_newmv_mode_sad_screen;
1995} REAL_TIME_SPEED_FEATURES;
1996
1998
2002typedef struct SPEED_FEATURES {
2007
2012
2016 TPL_SPEED_FEATURES tpl_sf;
2017
2021 GLOBAL_MOTION_SPEED_FEATURES gm_sf;
2022
2026 PARTITION_SPEED_FEATURES part_sf;
2027
2031 MV_SPEED_FEATURES mv_sf;
2032
2036 INTER_MODE_SPEED_FEATURES inter_sf;
2037
2041 INTERP_FILTER_SPEED_FEATURES interp_sf;
2042
2046 INTRA_MODE_SPEED_FEATURES intra_sf;
2047
2051 TX_SPEED_FEATURES tx_sf;
2052
2056 RD_CALC_SPEED_FEATURES rd_sf;
2057
2061 WINNER_MODE_SPEED_FEATURES winner_mode_sf;
2062
2066 LOOP_FILTER_SPEED_FEATURES lpf_sf;
2067
2071 REAL_TIME_SPEED_FEATURES rt_sf;
2074
2075struct AV1_COMP;
2076
2090 int speed);
2091
2104 int speed);
2117
2118#ifdef __cplusplus
2119} // extern "C"
2120#endif
2121
2122#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
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
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:2890
int speed
Definition encoder.h:3112
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:2002
MV_SPEED_FEATURES mv_sf
Definition speed_features.h:2031
TPL_SPEED_FEATURES tpl_sf
Definition speed_features.h:2016
LOOP_FILTER_SPEED_FEATURES lpf_sf
Definition speed_features.h:2066
TX_SPEED_FEATURES tx_sf
Definition speed_features.h:2051
INTER_MODE_SPEED_FEATURES inter_sf
Definition speed_features.h:2036
RD_CALC_SPEED_FEATURES rd_sf
Definition speed_features.h:2056
PARTITION_SPEED_FEATURES part_sf
Definition speed_features.h:2026
GLOBAL_MOTION_SPEED_FEATURES gm_sf
Definition speed_features.h:2021
INTERP_FILTER_SPEED_FEATURES interp_sf
Definition speed_features.h:2041
FIRST_PASS_SPEED_FEATURES fp_sf
Definition speed_features.h:2011
INTRA_MODE_SPEED_FEATURES intra_sf
Definition speed_features.h:2046
WINNER_MODE_SPEED_FEATURES winner_mode_sf
Definition speed_features.h:2061
REAL_TIME_SPEED_FEATURES rt_sf
Definition speed_features.h:2071
HIGH_LEVEL_SPEED_FEATURES hl_sf
Definition speed_features.h:2006