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 GM_DISABLE_SEARCH
39} UENUM1BYTE(GM_SEARCH_TYPE);
40
41enum {
42 DIST_WTD_COMP_ENABLED,
43 DIST_WTD_COMP_SKIP_MV_SEARCH,
44 DIST_WTD_COMP_DISABLED,
45} UENUM1BYTE(DIST_WTD_COMP_FLAG);
46
47enum {
48 INTRA_ALL = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED) | (1 << D45_PRED) |
49 (1 << D135_PRED) | (1 << D113_PRED) | (1 << D157_PRED) |
50 (1 << D203_PRED) | (1 << D67_PRED) | (1 << SMOOTH_PRED) |
51 (1 << SMOOTH_V_PRED) | (1 << SMOOTH_H_PRED) | (1 << PAETH_PRED),
52 UV_INTRA_ALL =
53 (1 << UV_DC_PRED) | (1 << UV_V_PRED) | (1 << UV_H_PRED) |
54 (1 << UV_D45_PRED) | (1 << UV_D135_PRED) | (1 << UV_D113_PRED) |
55 (1 << UV_D157_PRED) | (1 << UV_D203_PRED) | (1 << UV_D67_PRED) |
56 (1 << UV_SMOOTH_PRED) | (1 << UV_SMOOTH_V_PRED) |
57 (1 << UV_SMOOTH_H_PRED) | (1 << UV_PAETH_PRED) | (1 << UV_CFL_PRED),
58 UV_INTRA_DC = (1 << UV_DC_PRED),
59 UV_INTRA_DC_CFL = (1 << UV_DC_PRED) | (1 << UV_CFL_PRED),
60 UV_INTRA_DC_TM = (1 << UV_DC_PRED) | (1 << UV_PAETH_PRED),
61 UV_INTRA_DC_PAETH_CFL =
62 (1 << UV_DC_PRED) | (1 << UV_PAETH_PRED) | (1 << UV_CFL_PRED),
63 UV_INTRA_DC_H_V = (1 << UV_DC_PRED) | (1 << UV_V_PRED) | (1 << UV_H_PRED),
64 UV_INTRA_DC_H_V_CFL = (1 << UV_DC_PRED) | (1 << UV_V_PRED) |
65 (1 << UV_H_PRED) | (1 << UV_CFL_PRED),
66 UV_INTRA_DC_PAETH_H_V = (1 << UV_DC_PRED) | (1 << UV_PAETH_PRED) |
67 (1 << UV_V_PRED) | (1 << UV_H_PRED),
68 UV_INTRA_DC_PAETH_H_V_CFL = (1 << UV_DC_PRED) | (1 << UV_PAETH_PRED) |
69 (1 << UV_V_PRED) | (1 << UV_H_PRED) |
70 (1 << UV_CFL_PRED),
71 INTRA_DC = (1 << DC_PRED),
72 INTRA_DC_TM = (1 << DC_PRED) | (1 << PAETH_PRED),
73 INTRA_DC_H_V = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED),
74 INTRA_DC_H_V_SMOOTH =
75 (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED) | (1 << SMOOTH_PRED),
76 INTRA_DC_PAETH_H_V =
77 (1 << DC_PRED) | (1 << PAETH_PRED) | (1 << V_PRED) | (1 << H_PRED)
78};
79
80enum {
81 INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << GLOBALMV) |
82 (1 << NEWMV) | (1 << NEAREST_NEARESTMV) | (1 << NEAR_NEARMV) |
83 (1 << NEW_NEWMV) | (1 << NEAREST_NEWMV) | (1 << NEAR_NEWMV) |
84 (1 << NEW_NEARMV) | (1 << NEW_NEARESTMV) | (1 << GLOBAL_GLOBALMV),
85 INTER_NEAREST_NEAR_ZERO = (1 << NEARESTMV) | (1 << NEARMV) | (1 << GLOBALMV) |
86 (1 << NEAREST_NEARESTMV) | (1 << GLOBAL_GLOBALMV) |
87 (1 << NEAREST_NEWMV) | (1 << NEW_NEARESTMV) |
88 (1 << NEW_NEARMV) | (1 << NEAR_NEWMV) |
89 (1 << NEAR_NEARMV),
90};
91
92enum {
93 DISABLE_ALL_INTER_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
94 (1 << THR_ALTR) | (1 << THR_GOLD) | (1 << THR_LAST),
95
96 DISABLE_ALL_SPLIT = (1 << THR_INTRA) | DISABLE_ALL_INTER_SPLIT,
97
98 DISABLE_COMPOUND_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA),
99
100 LAST_AND_INTRA_SPLIT_ONLY = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
101 (1 << THR_ALTR) | (1 << THR_GOLD)
102};
103
104enum {
105 TXFM_CODING_SF = 1,
106 INTER_PRED_SF = 2,
107 INTRA_PRED_SF = 4,
108 PARTITION_SF = 8,
109 LOOP_FILTER_SF = 16,
110 RD_SKIP_SF = 32,
111 RESERVE_2_SF = 64,
112 RESERVE_3_SF = 128,
113} UENUM1BYTE(DEV_SPEED_FEATURES);
114
115/* This enumeration defines when the rate control recode loop will be
116 * enabled.
117 */
118enum {
119 /*
120 * No recodes allowed
121 */
122 DISALLOW_RECODE = 0,
123 /*
124 * Allow recode only for KF/ARF/GF frames
125 */
126 ALLOW_RECODE_KFARFGF = 1,
127 /*
128 * Allow recode for all frame types based on bitrate constraints.
129 */
130 ALLOW_RECODE = 2,
131} UENUM1BYTE(RECODE_LOOP_TYPE);
132
133enum {
134 SUBPEL_TREE = 0,
135 SUBPEL_TREE_PRUNED = 1, // Prunes 1/2-pel searches
136 SUBPEL_TREE_PRUNED_MORE = 2, // Prunes 1/2-pel searches more aggressively
137} UENUM1BYTE(SUBPEL_SEARCH_METHODS);
138
139enum {
140 // Try the full image with different values.
141 LPF_PICK_FROM_FULL_IMAGE,
142 // Try the full image filter search with non-dual filter only.
143 LPF_PICK_FROM_FULL_IMAGE_NON_DUAL,
144 // Try a small portion of the image with different values.
145 LPF_PICK_FROM_SUBIMAGE,
146 // Estimate the level based on quantizer and frame type
147 LPF_PICK_FROM_Q,
148 // Pick 0 to disable LPF if LPF was enabled last frame
149 LPF_PICK_MINIMAL_LPF
150} UENUM1BYTE(LPF_PICK_METHOD);
167
169enum {
170 // Terminate search early based on distortion so far compared to
171 // qp step, distortion in the neighborhood of the frame, etc.
172 FLAG_EARLY_TERMINATE = 1 << 0,
173
174 // Skips comp inter modes if the best so far is an intra mode.
175 FLAG_SKIP_COMP_BESTINTRA = 1 << 1,
176
177 // Skips oblique intra modes if the best so far is an inter mode.
178 FLAG_SKIP_INTRA_BESTINTER = 1 << 3,
179
180 // Skips oblique intra modes at angles 27, 63, 117, 153 if the best
181 // intra so far is not one of the neighboring directions.
182 FLAG_SKIP_INTRA_DIRMISMATCH = 1 << 4,
183
184 // Skips intra modes other than DC_PRED if the source variance is small
185 FLAG_SKIP_INTRA_LOWVAR = 1 << 5,
186} UENUM1BYTE(MODE_SEARCH_SKIP_LOGIC);
187
188enum {
189 // No tx type pruning
190 TX_TYPE_PRUNE_0 = 0,
191 // adaptively prunes the least perspective tx types out of all 16
192 // (tuned to provide negligible quality loss)
193 TX_TYPE_PRUNE_1 = 1,
194 // similar, but applies much more aggressive pruning to get better speed-up
195 TX_TYPE_PRUNE_2 = 2,
196 TX_TYPE_PRUNE_3 = 3,
197 // More aggressive pruning based on tx type score and allowed tx count
198 TX_TYPE_PRUNE_4 = 4,
199 TX_TYPE_PRUNE_5 = 5,
200} UENUM1BYTE(TX_TYPE_PRUNE_MODE);
201
202enum {
203 // No reaction to rate control on a detected slide/scene change.
204 NO_DETECTION = 0,
205
206 // Set to larger Q based only on the detected slide/scene change and
207 // current/past Q.
208 FAST_DETECTION_MAXQ = 1,
209} UENUM1BYTE(OVERSHOOT_DETECTION_CBR);
210
211enum {
212 // Turns off multi-winner mode. So we will do txfm search on either all modes
213 // if winner mode is off, or we will only on txfm search on a single winner
214 // mode.
215 MULTI_WINNER_MODE_OFF = 0,
216
217 // Limits the number of winner modes to at most 2
218 MULTI_WINNER_MODE_FAST = 1,
219
220 // Uses the default number of winner modes, which is 3 for intra mode, and 1
221 // for inter mode.
222 MULTI_WINNER_MODE_DEFAULT = 2,
223
224 // Maximum number of winner modes allowed.
225 MULTI_WINNER_MODE_LEVELS,
226} UENUM1BYTE(MULTI_WINNER_MODE_TYPE);
227
228enum {
229 PRUNE_NEARMV_OFF = 0, // Turn off nearmv pruning
230 PRUNE_NEARMV_LEVEL1 = 1, // Prune nearmv for qindex (0-85)
231 PRUNE_NEARMV_LEVEL2 = 2, // Prune nearmv for qindex (0-170)
232 PRUNE_NEARMV_LEVEL3 = 3, // Prune nearmv more aggressively for qindex (0-170)
233 PRUNE_NEARMV_MAX = PRUNE_NEARMV_LEVEL3,
234} UENUM1BYTE(PRUNE_NEARMV_LEVEL);
235
236typedef struct {
237 TX_TYPE_PRUNE_MODE prune_2d_txfm_mode;
238 int fast_intra_tx_type_search;
239
240 // INT_MAX: Disable fast search.
241 // 1 - 1024: Probability threshold used for conditionally forcing tx type,
242 // during mode search.
243 // 0: Force tx type to be DCT_DCT unconditionally, during
244 // mode search.
245 int fast_inter_tx_type_prob_thresh;
246
247 // Prune less likely chosen transforms for each intra mode. The speed
248 // feature ranges from 0 to 2, for different speed / compression trade offs.
249 int use_reduced_intra_txset;
250
251 // Use a skip flag prediction model to detect blocks with skip = 1 early
252 // and avoid doing full TX type search for such blocks.
253 int use_skip_flag_prediction;
254
255 // Threshold used by the ML based method to predict TX block split decisions.
256 int ml_tx_split_thresh;
257
258 // skip remaining transform type search when we found the rdcost of skip is
259 // better than applying transform
260 int skip_tx_search;
261
262 // Prune tx type search using previous frame stats.
263 int prune_tx_type_using_stats;
264 // Prune tx type search using estimated RDcost
265 int prune_tx_type_est_rd;
266
267 // Flag used to control the winner mode processing for tx type pruning for
268 // inter blocks. It enables further tx type mode pruning based on ML model for
269 // mode evaluation and disables tx type mode pruning for winner mode
270 // processing.
271 int winner_mode_tx_type_pruning;
272} TX_TYPE_SEARCH;
273
274enum {
275 // Search partitions using RD criterion
276 SEARCH_PARTITION,
277
278 // Always use a fixed size partition
279 FIXED_PARTITION,
280
281 // Partition using source variance
282 VAR_BASED_PARTITION,
283
284#if CONFIG_RT_ML_PARTITIONING
285 // Partition using ML model
286 ML_BASED_PARTITION
287#endif
288} UENUM1BYTE(PARTITION_SEARCH_TYPE);
289
290enum {
291 NOT_IN_USE,
292 DIRECT_PRED,
293 RELAXED_PRED,
294 ADAPT_PRED
295} UENUM1BYTE(MAX_PART_PRED_MODE);
296
297enum {
298 LAST_MV_DATA,
299 CURRENT_Q,
300 QTR_ONLY,
301} UENUM1BYTE(MV_PREC_LOGIC);
302
303enum {
304 SUPERRES_AUTO_ALL, // Tries all possible superres ratios
305 SUPERRES_AUTO_DUAL, // Tries no superres and q-based superres ratios
306 SUPERRES_AUTO_SOLO, // Only apply the q-based superres ratio
307} UENUM1BYTE(SUPERRES_AUTO_SEARCH_TYPE);
332
355
364
382
439
467
469typedef struct TPL_SPEED_FEATURES {
470 // GOP length adaptive decision.
471 // If set to 0, tpl model decides whether a shorter gf interval is better.
472 // If set to 1, tpl stats of ARFs from base layer, (base+1) layer and
473 // (base+2) layer decide whether a shorter gf interval is better.
474 // If set to 2, tpl stats of ARFs from base layer, (base+1) layer and GF boost
475 // decide whether a shorter gf interval is better.
476 // If set to 3, gop length adaptive decision is disabled.
477 int gop_length_decision_method;
478 // Prune the intra modes search by tpl.
479 // If set to 0, we will search all intra modes from DC_PRED to PAETH_PRED.
480 // If set to 1, we only search DC_PRED, V_PRED, and H_PRED.
481 int prune_intra_modes;
482 // This parameter controls which step in the n-step process we start at.
483 int reduce_first_step_size;
484 // Skip motion estimation based on the precision of center MVs and the
485 // difference between center MVs.
486 // If set to 0, motion estimation is skipped for duplicate center MVs
487 // (default). If set to 1, motion estimation is skipped for duplicate
488 // full-pixel center MVs. If set to 2, motion estimation is skipped if the
489 // difference between center MVs is less than the threshold.
490 int skip_alike_starting_mv;
491
492 // When to stop subpel search.
493 SUBPEL_FORCE_STOP subpel_force_stop;
494
495 // Which search method to use.
496 SEARCH_METHODS search_method;
497
498 // Prune starting mvs in TPL based on sad scores.
499 int prune_starting_mv;
500
501 // Not run TPL for filtered Key frame.
502 int disable_filtered_key_tpl;
503
504 // Prune reference frames in TPL.
505 int prune_ref_frames_in_tpl;
506
507 // Support compound predictions.
508 int allow_compound_pred;
509
510 // Calculate rate and distortion based on Y plane only.
511 int use_y_only_rate_distortion;
512} TPL_SPEED_FEATURES;
513
514typedef struct GLOBAL_MOTION_SPEED_FEATURES {
515 GM_SEARCH_TYPE gm_search_type;
516
517 // During global motion estimation, prune remaining reference frames in a
518 // given direction(past/future), if the evaluated ref_frame in that direction
519 // yields gm_type as INVALID/TRANSLATION/IDENTITY
520 int prune_ref_frame_for_gm_search;
521
522 // When the current GM type is set to ZEROMV, prune ZEROMV if its performance
523 // is worse than NEWMV under SSE metric.
524 // 0 : no pruning
525 // 1 : conservative pruning
526 // 2 : aggressive pruning
528
529 // Disable global motion estimation based on stats of previous frames in the
530 // GF group
531 int disable_gm_search_based_on_stats;
532} GLOBAL_MOTION_SPEED_FEATURES;
533
534typedef struct PARTITION_SPEED_FEATURES {
535 PARTITION_SEARCH_TYPE partition_search_type;
536
537 // Used if partition_search_type = FIXED_PARTITION
538 BLOCK_SIZE fixed_partition_size;
539
540 // Prune extended partition types search
541 // Can take values 0 - 2, 0 referring to no pruning, and 1 - 2 increasing
542 // aggressiveness of pruning in order.
543 int prune_ext_partition_types_search_level;
544
545 // Prune part4 based on block size
546 int prune_part4_search;
547
548 // Use a ML model to prune rectangular, ab and 4-way horz
549 // and vert partitions
550 int ml_prune_partition;
551
552 // Use a ML model to adaptively terminate partition search after trying
553 // PARTITION_SPLIT. Can take values 0 - 2, 0 meaning not being enabled, and
554 // 1 - 2 increasing aggressiveness in order.
555 int ml_early_term_after_part_split_level;
556
557 // Skip rectangular partition test when partition type none gives better
558 // rd than partition type split. Can take values 0 - 2, 0 referring to no
559 // skipping, and 1 - 2 increasing aggressiveness of skipping in order.
560 int less_rectangular_check_level;
561
562 // Use square partition only beyond this block size.
563 BLOCK_SIZE use_square_partition_only_threshold;
564
565 // Sets max square partition levels for this superblock based on
566 // motion vector and prediction error distribution produced from 16x16
567 // simple motion search
568 MAX_PART_PRED_MODE auto_max_partition_based_on_simple_motion;
569
570 // Min and max square partition size we enable (block_size) as per auto
571 // min max, but also used by adjust partitioning, and pick_partitioning.
572 BLOCK_SIZE default_min_partition_size;
573 BLOCK_SIZE default_max_partition_size;
574
575 // Sets level of adjustment of variance-based partitioning during
576 // rd_use_partition 0 - no partition adjustment, 1 - try to merge partitions
577 // for small blocks and high QP, 2 - try to merge partitions, 3 - try to merge
578 // and split leaf partitions and 0 - 3 decreasing aggressiveness in order.
579 int adjust_var_based_rd_partitioning;
580
581 // Partition search early breakout thresholds.
582 int64_t partition_search_breakout_dist_thr;
583 int partition_search_breakout_rate_thr;
584
585 // Thresholds for ML based partition search breakout.
586 int ml_partition_search_breakout_thresh[PARTITION_BLOCK_SIZES];
587
588 // Aggressiveness levels for pruning split and rectangular partitions based on
589 // simple_motion_search. SIMPLE_AGG_LVL0 to SIMPLE_AGG_LVL3 correspond to
590 // simple motion search based pruning. QIDX_BASED_AGG_LVL1 corresponds to
591 // qindex based and simple motion search based pruning.
592 int simple_motion_search_prune_agg;
593
594 // Perform simple_motion_search on each possible subblock and use it to prune
595 // PARTITION_HORZ and PARTITION_VERT.
596 int simple_motion_search_prune_rect;
597
598 // Perform simple motion search before none_partition to decide if we
599 // want to remove all partitions other than PARTITION_SPLIT. If set to 0, this
600 // model is disabled. If set to 1, the model attempts to perform
601 // PARTITION_SPLIT only. If set to 2, the model also attempts to prune
602 // PARTITION_SPLIT.
603 int simple_motion_search_split;
604
605 // Use features from simple_motion_search to terminate prediction block
606 // partition after PARTITION_NONE
607 int simple_motion_search_early_term_none;
608
609 // Controls whether to reduce the number of motion search steps. If this is 0,
610 // then simple_motion_search has the same number of steps as
611 // single_motion_search (assuming no other speed features). Otherwise, reduce
612 // the number of steps by the value contained in this variable.
613 int simple_motion_search_reduce_search_steps;
614
615 // This variable controls the maximum block size where intra blocks can be
616 // used in inter frames.
617 // TODO(aconverse): Fold this into one of the other many mode skips
618 BLOCK_SIZE max_intra_bsize;
619
620 // Use CNN with luma pixels on source frame on each of the 64x64 subblock to
621 // perform partition pruning in intra frames.
622 // 0: No Pruning
623 // 1: Prune split and rectangular partitions only
624 // 2: Prune none, split and rectangular partitions
625 int intra_cnn_based_part_prune_level;
626
627 // Disable extended partition search for lower block sizes.
628 int ext_partition_eval_thresh;
629
630 // Disable rectangular partitions for larger block sizes.
631 int rect_partition_eval_thresh;
632
633 // prune extended partition search
634 // 0 : no pruning
635 // 1 : prune 1:4 partition search using winner info from split partitions
636 // 2 : prune 1:4 and AB partition search using split and HORZ/VERT info
637 int prune_ext_part_using_split_info;
638
639 // Prunt rectangular, AB and 4-way partition based on q index and block size
640 // 0 : no pruning
641 // 1 : prune sub_8x8 at very low quantizers
642 // 2 : prune all block size based on qindex
643 int prune_rectangular_split_based_on_qidx;
644
645 // Terminate partition search for child partition,
646 // when NONE and SPLIT partition rd_costs are INT64_MAX.
647 int early_term_after_none_split;
648
649 // Level used to adjust threshold for av1_ml_predict_breakout(). At lower
650 // levels, more conservative threshold is used, and value of 0 indicates
651 // av1_ml_predict_breakout() is disabled. Value of 3 corresponds to default
652 // case with no adjustment to lbd thresholds.
653 int ml_predict_breakout_level;
654
655 // Prune sub_8x8 (BLOCK_4X4, BLOCK_4X8 and BLOCK_8X4) partitions.
656 // 0 : no pruning
657 // 1 : pruning based on neighbour block information
658 // 2 : prune always
659 int prune_sub_8x8_partition_level;
660
661 // Prune rectangular split based on simple motion search split/no_split score.
662 // 0: disable pruning, 1: enable pruning
663 int simple_motion_search_rect_split;
664
665 // The current encoder adopts a DFS search for block partitions.
666 // Therefore the mode selection and associated rdcost is ready for smaller
667 // blocks before the mode selection for some partition types.
668 // AB partition could use previous rd information and skip mode search.
669 // An example is:
670 //
671 // current block
672 // +---+---+
673 // | |
674 // + +
675 // | |
676 // +-------+
677 //
678 // SPLIT partition has been searched first before trying HORZ_A
679 // +---+---+
680 // | R | R |
681 // +---+---+
682 // | R | R |
683 // +---+---+
684 //
685 // HORZ_A
686 // +---+---+
687 // | | |
688 // +---+---+
689 // | |
690 // +-------+
691 //
692 // With this speed feature, the top two sub blocks can directly use rdcost
693 // searched in split partition, and the mode info is also copied from
694 // saved info. Similarly, the bottom rectangular block can also use
695 // the available information from previous rectangular search.
696 int reuse_prev_rd_results_for_part_ab;
697
698 // Reuse the best prediction modes found in PARTITION_SPLIT and PARTITION_RECT
699 // when encoding PARTITION_AB.
700 int reuse_best_prediction_for_part_ab;
701
702 // The current partition search records the best rdcost so far and uses it
703 // in mode search and transform search to early skip when some criteria is
704 // met. For example, when the current rdcost is larger than the best rdcost,
705 // or the model rdcost is larger than the best rdcost times some thresholds.
706 // By default, this feature is turned on to speed up the encoder partition
707 // search.
708 // If disabling it, at speed 0, 30 frames, we could get
709 // about -0.25% quality gain (psnr, ssim, vmaf), with about 13% slowdown.
710 int use_best_rd_for_pruning;
711
712 // Skip evaluation of non-square partitions based on the corresponding NONE
713 // partition.
714 // 0: no pruning
715 // 1: prune extended partitions if NONE is skippable
716 // 2: on top of 1, prune rectangular partitions if NONE is inter, not a newmv
717 // mode and skippable
718 int skip_non_sq_part_based_on_none;
719
720 // Disables 8x8 and below partitions for low quantizers.
721 int disable_8x8_part_based_on_qidx;
722} PARTITION_SPEED_FEATURES;
723
724typedef struct MV_SPEED_FEATURES {
725 // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
726 SEARCH_METHODS search_method;
727
728 // Enable the use of faster, less accurate mv search method on bsize >=
729 // BLOCK_32X32.
730 // TODO(chiyotsai@google.com): Take the clip's resolution and mv activity into
731 // account.
732 int use_bsize_dependent_search_method;
733
734 // If this is set to 1, we limit the motion search range to 2 times the
735 // largest motion vector found in the last frame.
736 int auto_mv_step_size;
737
738 // Subpel_search_method can only be subpel_tree which does a subpixel
739 // logarithmic search that keeps stepping at 1/2 pixel units until
740 // you stop getting a gain, and then goes on to 1/4 and repeats
741 // the same process. Along the way it skips many diagonals.
742 SUBPEL_SEARCH_METHODS subpel_search_method;
743
744 // Maximum number of steps in logarithmic subpel search before giving up.
745 int subpel_iters_per_step;
746
747 // When to stop subpel search.
748 SUBPEL_FORCE_STOP subpel_force_stop;
749
750 // When to stop subpel search in simple motion search.
751 SUBPEL_FORCE_STOP simple_motion_subpel_force_stop;
752
753 // If true, sub-pixel search uses the exact convolve function used for final
754 // encoding and decoding; otherwise, it uses bilinear interpolation.
755 SUBPEL_SEARCH_TYPE use_accurate_subpel_search;
756
757 // Threshold for allowing exhaustive motion search.
758 int exhaustive_searches_thresh;
759
760 // Pattern to be used for any exhaustive mesh searches (except intraBC ME).
761 MESH_PATTERN mesh_patterns[MAX_MESH_STEP];
762
763 // Pattern to be used for exhaustive mesh searches of intraBC ME.
764 MESH_PATTERN intrabc_mesh_patterns[MAX_MESH_STEP];
765
766 // Reduce single motion search range based on MV result of prior ref_mv_idx.
767 int reduce_search_range;
768
769 // Prune mesh search.
770 PRUNE_MESH_SEARCH_LEVEL prune_mesh_search;
771
772 // Use the rd cost around the best FULLPEL_MV to speed up subpel search
773 int use_fullpel_costlist;
774
775 // Set the full pixel search level of obmc
776 // 0: obmc_full_pixel_diamond
777 // 1: obmc_refining_search_sad (faster)
778 int obmc_full_pixel_search_level;
779
780 // Accurate full pixel motion search based on TPL stats.
781 int full_pixel_search_level;
782
783 // Whether to downsample the rows in sad calculation during motion search.
784 // This is only active when there are at least 16 rows.
785 int use_downsampled_sad;
786
787 // Enable/disable extensive joint motion search.
788 int disable_extensive_joint_motion_search;
789
790 // Enable second best mv check in joint mv search.
791 // 0: allow second MV (use rd cost as the metric)
792 // 1: use var as the metric
793 // 2: disable second MV
794 int disable_second_mv;
795
796 // Skips full pixel search based on start mv of prior ref_mv_idx.
797 int skip_fullpel_search_using_startmv;
798} MV_SPEED_FEATURES;
799
800typedef struct INTER_MODE_SPEED_FEATURES {
801 // 2-pass inter mode model estimation where the preliminary pass skips
802 // transform search and uses a model to estimate rd, while the final pass
803 // computes the full transform search. Two types of models are supported:
804 // 0: not used
805 // 1: used with online dynamic rd model
806 // 2: used with static rd model
807 int inter_mode_rd_model_estimation;
808
809 // Bypass transform search based on skip rd
810 int txfm_rd_gate_level;
811
812 // Limit the inter mode tested in the RD loop
813 int reduce_inter_modes;
814
815 // This variable is used to cap the maximum number of times we skip testing a
816 // mode to be evaluated. A high value means we will be faster.
817 int adaptive_rd_thresh;
818
819 // Aggressively prune inter modes when best mode is skippable.
820 int prune_inter_modes_if_skippable;
821
822 // Drop less likely to be picked reference frames in the RD search.
823 // Has seven levels for now: 0, 1, 2, 3, 4, 5 and 6 where higher levels prune
824 // more aggressively than lower ones. (0 means no pruning).
825 int selective_ref_frame;
826
827 // Prune reference frames for rectangular partitions.
828 // 0 implies no pruning
829 // 1 implies prune for extended partition
830 // 2 implies prune horiz, vert and extended partition
831 int prune_ref_frame_for_rect_partitions;
832
833 // Prune inter modes w.r.t past reference frames
834 // 0 no pruning
835 // 1 prune inter modes w.r.t ALTREF2 and ALTREF reference frames
836 // 2 prune inter modes w.r.t BWDREF, ALTREF2 and ALTREF reference frames
837 int alt_ref_search_fp;
838
839 // Prune compound reference frames
840 // 0 no pruning
841 // 1 prune compound references which do not satisfy the two conditions:
842 // a) The references are at a nearest distance from the current frame in
843 // both past and future direction.
844 // b) The references have minimum pred_mv_sad in both past and future
845 // direction.
846 // 2 prune compound references except the one with nearest distance from the
847 // current frame in both past and future direction.
848 int prune_comp_ref_frames;
849
850 // Skip the current ref_mv in NEW_MV mode based on mv, rate cost, etc.
851 // This speed feature equaling 0 means no skipping.
852 // If the speed feature equals 1 or 2, skip the current ref_mv in NEW_MV mode
853 // if we have already encountered ref_mv in the drl such that:
854 // 1. The other drl has the same mv during the SIMPLE_TRANSLATION search
855 // process as the current mv.
856 // 2. The rate needed to encode the current mv is larger than that for the
857 // other ref_mv.
858 // The speed feature equaling 1 means using subpel mv in the comparison.
859 // The speed feature equaling 2 means using fullpel mv in the comparison.
860 // If the speed feature >= 3, skip the current ref_mv in NEW_MV mode based on
861 // known full_mv bestsme and drl cost.
862 int skip_newmv_in_drl;
863
864 // This speed feature checks duplicate ref MVs among NEARESTMV, NEARMV,
865 // GLOBALMV and skips NEARMV or GLOBALMV (in order) if a duplicate is found
866 // TODO(any): Instead of skipping repeated ref mv, use the recalculated
867 // rd-cost based on mode rate and skip the mode evaluation
868 int skip_repeated_ref_mv;
869
870 // Flag used to control the ref_best_rd based gating for chroma
871 int perform_best_rd_based_gating_for_chroma;
872
873 // Reuse the inter_intra_mode search result from NEARESTMV mode to other
874 // single ref modes
875 int reuse_inter_intra_mode;
876
877 // prune wedge and compound segment approximate rd evaluation based on
878 // compound average modeled rd
879 int prune_comp_type_by_model_rd;
880
881 // prune wedge and compound segment approximate rd evaluation based on
882 // compound average rd/ref_best_rd
883 int prune_comp_type_by_comp_avg;
884
885 // Skip some ref frames in compound motion search by single motion search
886 // result. Has three levels for now: 0 referring to no skipping, and 1 - 3
887 // increasing aggressiveness of skipping in order.
888 // Note: The search order might affect the result. It assumes that the single
889 // reference modes are searched before compound modes. It is better to search
890 // same single inter mode as a group.
891 int prune_comp_search_by_single_result;
892
893 // Instead of performing a full MV search, do a simple translation first
894 // and only perform a full MV search on the motion vectors that performed
895 // well.
896 int prune_mode_search_simple_translation;
897
898 // Only search compound modes with at least one "good" reference frame.
899 // A reference frame is good if, after looking at its performance among
900 // the single reference modes, it is one of the two best performers.
901 int prune_compound_using_single_ref;
902
903 // Skip extended compound mode (NEAREST_NEWMV, NEW_NEARESTMV, NEAR_NEWMV,
904 // NEW_NEARMV) using ref frames of above and left neighbor
905 // blocks.
906 // 0 : no pruning
907 // 1 : prune ext compound modes using neighbor blocks (less aggressiveness)
908 // 2 : prune ext compound modes using neighbor blocks (high aggressiveness)
909 // 3 : prune ext compound modes unconditionally (highest aggressiveness)
910 int prune_ext_comp_using_neighbors;
911
912 // Skip NEW_NEARMV and NEAR_NEWMV extended compound modes
913 int skip_ext_comp_nearmv_mode;
914
915 // Skip extended compound mode when ref frame corresponding to NEWMV does not
916 // have NEWMV as single mode winner.
917 // 0 : no pruning
918 // 1 : prune extended compound mode (less aggressiveness)
919 // 2 : prune extended compound mode (high aggressiveness)
920 int prune_comp_using_best_single_mode_ref;
921
922 // Skip NEARESTMV and NEARMV using weight computed in ref mv list population
923 int prune_nearest_near_mv_using_refmv_weight;
924
925 // Based on previous ref_mv_idx search result, prune the following search.
926 int prune_ref_mv_idx_search;
927
928 // Disable one sided compound modes.
929 int disable_onesided_comp;
930
931 // Prune obmc search using previous frame stats.
932 // INT_MAX : disable obmc search
933 int prune_obmc_prob_thresh;
934
935 // Prune warped motion search using previous frame stats.
936 int prune_warped_prob_thresh;
937
938 // Variance threshold to enable/disable Interintra wedge search
939 unsigned int disable_interintra_wedge_var_thresh;
940
941 // Variance threshold to enable/disable Interinter wedge search
942 unsigned int disable_interinter_wedge_var_thresh;
943
944 // De-couple wedge and mode search during interintra RDO.
945 int fast_interintra_wedge_search;
946
947 // Whether fast wedge sign estimate is used
948 int fast_wedge_sign_estimate;
949
950 // Enable/disable ME for interinter wedge search.
951 int disable_interinter_wedge_newmv_search;
952
953 // Decide when and how to use joint_comp.
954 DIST_WTD_COMP_FLAG use_dist_wtd_comp_flag;
955
956 // Clip the frequency of updating the mv cost.
957 INTERNAL_COST_UPDATE_TYPE mv_cost_upd_level;
958
959 // Clip the frequency of updating the coeff cost.
960 INTERNAL_COST_UPDATE_TYPE coeff_cost_upd_level;
961
962 // Clip the frequency of updating the mode cost.
963 INTERNAL_COST_UPDATE_TYPE mode_cost_upd_level;
964
965 // Prune inter modes based on tpl stats
966 // 0 : no pruning
967 // 1 - 3 indicate increasing aggressiveness in order.
968 int prune_inter_modes_based_on_tpl;
969
970 // Skip NEARMV and NEAR_NEARMV modes using ref frames of above and left
971 // neighbor blocks and qindex.
972 PRUNE_NEARMV_LEVEL prune_nearmv_using_neighbors;
973
974 // Model based breakout after interpolation filter search
975 // 0: no breakout
976 // 1: use model based rd breakout
977 int model_based_post_interp_filter_breakout;
978
979 // Reuse compound type rd decision when exact match is found
980 // 0: No reuse
981 // 1: Reuse the compound type decision
982 int reuse_compound_type_decision;
983
984 // Enable/disable masked compound.
985 int disable_masked_comp;
986
987 // Enable/disable the fast compound mode search.
988 int enable_fast_compound_mode_search;
989
990 // Reuse masked compound type search results
991 int reuse_mask_search_results;
992
993 // Enable/disable fast search for wedge masks
994 int enable_fast_wedge_mask_search;
995
996 // Early breakout from transform search of inter modes
997 int inter_mode_txfm_breakout;
998
999 // Limit number of inter modes for txfm search if a newmv mode gets
1000 // evaluated among the top modes.
1001 // 0: no pruning
1002 // 1 to 3 indicate increasing order of aggressiveness
1003 int limit_inter_mode_cands;
1004
1005 // Cap the no. of txfm searches for a given prediction mode.
1006 // 0: no cap, 1: cap beyond first 4 searches, 2: cap beyond first 3 searches.
1007 int limit_txfm_eval_per_mode;
1008
1009 // Prune warped motion search based on block size.
1010 int extra_prune_warped;
1011
1012 // Do not search compound modes for ARF.
1013 // The intuition is that ARF is predicted by frames far away from it,
1014 // whose temporal correlations with the ARF are likely low.
1015 // It is therefore likely that compound modes do not work as well for ARF
1016 // as other inter frames.
1017 // Speed/quality impact:
1018 // Speed 1: 12% faster, 0.1% psnr loss.
1019 // Speed 2: 2% faster, 0.05% psnr loss.
1020 // No change for speed 3 and up, because |disable_onesided_comp| is true.
1021 int skip_arf_compound;
1022} INTER_MODE_SPEED_FEATURES;
1023
1024typedef struct INTERP_FILTER_SPEED_FEATURES {
1025 // Do limited interpolation filter search for dual filters, since best choice
1026 // usually includes EIGHTTAP_REGULAR.
1027 int use_fast_interpolation_filter_search;
1028
1029 // Disable dual filter
1030 int disable_dual_filter;
1031
1032 // Save results of av1_interpolation_filter_search for a block
1033 // Check mv and ref_frames before search, if they are very close with previous
1034 // saved results, filter search can be skipped.
1035 int use_interp_filter;
1036
1037 // skip sharp_filter evaluation based on regular and smooth filter rd for
1038 // dual_filter=0 case
1039 int skip_sharp_interp_filter_search;
1040
1041 int cb_pred_filter_search;
1042
1043 // adaptive interp_filter search to allow skip of certain filter types.
1044 int adaptive_interp_filter_search;
1045} INTERP_FILTER_SPEED_FEATURES;
1046
1047typedef struct INTRA_MODE_SPEED_FEATURES {
1048 // These bit masks allow you to enable or disable intra modes for each
1049 // transform size separately.
1050 int intra_y_mode_mask[TX_SIZES];
1051 int intra_uv_mode_mask[TX_SIZES];
1052
1053 // flag to allow skipping intra mode for inter frame prediction
1054 int skip_intra_in_interframe;
1055
1056 // Prune intra mode candidates based on source block histogram of gradient.
1057 // Applies to luma plane only.
1058 // Feasible values are 0..4. The feature is disabled for 0. An increasing
1059 // value indicates more aggressive pruning threshold.
1060 int intra_pruning_with_hog;
1061
1062 // Prune intra mode candidates based on source block histogram of gradient.
1063 // Applies to chroma plane only.
1064 // Feasible values are 0..4. The feature is disabled for 0. An increasing
1065 // value indicates more aggressive pruning threshold.
1066 int chroma_intra_pruning_with_hog;
1067
1068 // Enable/disable smooth intra modes.
1069 int disable_smooth_intra;
1070
1071 // Prune UV_SMOOTH_PRED mode for chroma based on chroma source variance.
1072 // false : No pruning
1073 // true : Prune UV_SMOOTH_PRED mode based on chroma source variance
1074 //
1075 // For allintra encode, this speed feature reduces instruction count
1076 // by 1.90%, 2.21% and 1.97% for speed 6, 7 and 8 with coding performance
1077 // change less than 0.04%. For AVIF image encode, this speed feature reduces
1078 // encode time by 1.56%, 2.14% and 0.90% for speed 6, 7 and 8 on a typical
1079 // image dataset with coding performance change less than 0.05%.
1080 bool prune_smooth_intra_mode_for_chroma;
1081
1082 // Prune filter intra modes in intra frames.
1083 // 0 : No pruning
1084 // 1 : Evaluate applicable filter intra modes based on best intra mode so far
1085 // 2 : Do not evaluate filter intra modes
1086 int prune_filter_intra_level;
1087
1088 // prune palette search
1089 // 0: No pruning
1090 // 1: Perform coarse search to prune the palette colors. For winner colors,
1091 // neighbors are also evaluated using a finer search.
1092 // 2: Perform 2 way palette search from max colors to min colors (and min
1093 // colors to remaining colors) and terminate the search if current number of
1094 // palette colors is not the winner.
1095 int prune_palette_search_level;
1096
1097 // Terminate early in luma palette_size search. Speed feature values indicate
1098 // increasing level of pruning.
1099 // 0: No early termination
1100 // 1: Terminate early for higher luma palette_size, if header rd cost of lower
1101 // palette_size is more than 2 * best_rd. This level of pruning is more
1102 // conservative when compared to sf level 2 as the cases which will get pruned
1103 // with sf level 1 is a subset of the cases which will get pruned with sf
1104 // level 2.
1105 // 2: Terminate early for higher luma palette_size, if header rd cost of lower
1106 // palette_size is more than best_rd.
1107 // For allintra encode, this sf reduces instruction count by 2.49%, 1.07%,
1108 // 2.76%, 2.30%, 1.84%, 2.69%, 2.04%, 2.05% and 1.44% for speed 0, 1, 2, 3, 4,
1109 // 5, 6, 7 and 8 on screen content set with coding performance change less
1110 // than 0.01% for speed <= 2 and less than 0.03% for speed >= 3. For AVIF
1111 // image encode, this sf reduces instruction count by 1.94%, 1.13%, 1.29%,
1112 // 0.93%, 0.89%, 1.03%, 1.07%, 1.20% and 0.18% for speed 0, 1, 2, 3, 4, 5, 6,
1113 // 7 and 8 on a typical image dataset with coding performance change less than
1114 // 0.01%.
1115 int prune_luma_palette_size_search_level;
1116
1117 // Prune chroma intra modes based on luma intra mode winner.
1118 // 0: No pruning
1119 // 1: Prune chroma intra modes other than UV_DC_PRED, UV_SMOOTH_PRED,
1120 // UV_CFL_PRED and the mode that corresponds to luma intra mode winner.
1121 int prune_chroma_modes_using_luma_winner;
1122
1123 // Clip the frequency of updating the mv cost for intrabc.
1124 INTERNAL_COST_UPDATE_TYPE dv_cost_upd_level;
1125
1126 // We use DCT_DCT transform followed by computing SATD (Sum of Absolute
1127 // Transformed Differences) as an estimation of RD score to quickly find the
1128 // best possible Chroma from Luma (CFL) parameter. Then we do a full RD search
1129 // near the best possible parameter. The search range is set here.
1130 // The range of cfl_searh_range should be [1, 33], and the following are the
1131 // recommended values.
1132 // 1: Fastest mode.
1133 // 3: Default mode that provides good speedup without losing compression
1134 // performance at speed 0.
1135 // 33: Exhaustive rd search (33 == CFL_MAGS_SIZE). This mode should only
1136 // be used for debugging purpose.
1137 int cfl_search_range;
1138
1139 // TOP_INTRA_MODEL_COUNT is 4 that is the number of top model rd to store in
1140 // intra mode decision. Here, add a speed feature to reduce this number for
1141 // higher speeds.
1142 int top_intra_model_count_allowed;
1143
1144 // Adapt top_intra_model_count_allowed locally to prune luma intra modes using
1145 // neighbor block and quantizer information.
1146 int adapt_top_model_rd_count_using_neighbors;
1147
1148 // Prune the evaluation of odd delta angles of directional luma intra modes by
1149 // using the rdcosts of neighbouring delta angles.
1150 // For allintra encode, this speed feature reduces instruction count
1151 // by 4.461%, 3.699% and 3.536% for speed 6, 7 and 8 on a typical video
1152 // dataset with coding performance change less than 0.26%. For AVIF image
1153 // encode, this speed feature reduces encode time by 2.849%, 2.471%,
1154 // and 2.051% for speed 6, 7 and 8 on a typical image dataset with coding
1155 // performance change less than 0.27%.
1156 int prune_luma_odd_delta_angles_in_intra;
1157
1158 // Terminate early in chroma palette_size search.
1159 // 0: No early termination
1160 // 1: Terminate early for higher palette_size, if header rd cost of lower
1161 // palette_size is more than best_rd.
1162 // For allintra encode, this sf reduces instruction count by 0.45%,
1163 // 0.62%, 1.73%, 2.50%, 2.89%, 3.09% and 3.86% for speed 0 to 6 on screen
1164 // content set with coding performance change less than 0.01%.
1165 // For AVIF image encode, this sf reduces instruction count by 0.45%, 0.81%,
1166 // 0.85%, 1.05%, 1.45%, 1.66% and 1.95% for speed 0 to 6 on a typical image
1167 // dataset with no quality drop.
1168 int early_term_chroma_palette_size_search;
1169
1170 // Skips the evaluation of filter intra modes in inter frames if rd evaluation
1171 // of luma intra dc mode results in invalid rd stats.
1172 int skip_filter_intra_in_inter_frames;
1173} INTRA_MODE_SPEED_FEATURES;
1174
1175typedef struct TX_SPEED_FEATURES {
1176 // Init search depth for square and rectangular transform partitions.
1177 // Values:
1178 // 0 - search full tree, 1: search 1 level, 2: search the highest level only
1179 int inter_tx_size_search_init_depth_sqr;
1180 int inter_tx_size_search_init_depth_rect;
1181 int intra_tx_size_search_init_depth_sqr;
1182 int intra_tx_size_search_init_depth_rect;
1183
1184 // If any dimension of a coding block size above 64, always search the
1185 // largest transform only, since the largest transform block size is 64x64.
1186 int tx_size_search_lgr_block;
1187
1188 TX_TYPE_SEARCH tx_type_search;
1189
1190 // Skip split transform block partition when the collocated bigger block
1191 // is selected as all zero coefficients.
1192 int txb_split_cap;
1193
1194 // Shortcut the transform block partition and type search when the target
1195 // rdcost is relatively lower.
1196 // Values are 0 (not used) , or 1 - 2 with progressively increasing
1197 // aggressiveness
1198 int adaptive_txb_search_level;
1199
1200 // Prune level for tx_size_type search for inter based on rd model
1201 // 0: no pruning
1202 // 1-2: progressively increasing aggressiveness of pruning
1203 int model_based_prune_tx_search_level;
1204
1205 // Refine TX type after fast TX search.
1206 int refine_fast_tx_search_results;
1207
1208 // Prune transform split/no_split eval based on residual properties. A value
1209 // of 0 indicates no pruning, and the aggressiveness of pruning progressively
1210 // increases from levels 1 to 3.
1211 int prune_tx_size_level;
1212
1213 // Prune the evaluation of transform depths as decided by the NN model.
1214 // false: No pruning.
1215 // true : Avoid the evaluation of specific transform depths using NN model.
1216 //
1217 // For allintra encode, this speed feature reduces instruction count
1218 // by 4.76%, 8.92% and 11.28% for speed 6, 7 and 8 with coding performance
1219 // change less than 0.32%. For AVIF image encode, this speed feature reduces
1220 // encode time by 4.65%, 9.16% and 10.45% for speed 6, 7 and 8 on a typical
1221 // image dataset with coding performance change less than 0.19%.
1222 bool prune_intra_tx_depths_using_nn;
1223} TX_SPEED_FEATURES;
1224
1225typedef struct RD_CALC_SPEED_FEATURES {
1226 // Fast approximation of av1_model_rd_from_var_lapndz
1227 int simple_model_rd_from_var;
1228
1229 // Perform faster distortion computation during the R-D evaluation by trying
1230 // to approximate the prediction error with transform coefficients (faster but
1231 // less accurate) rather than computing distortion in the pixel domain (slower
1232 // but more accurate). The following methods are used for distortion
1233 // computation:
1234 // Method 0: Always compute distortion in the pixel domain
1235 // Method 1: Based on block error, try using transform domain distortion for
1236 // tx_type search and compute distortion in pixel domain for final RD_STATS
1237 // Method 2: Based on block error, try to compute distortion in transform
1238 // domain
1239 // Methods 1 and 2 may fallback to computing distortion in the pixel domain in
1240 // case the block error is less than the threshold, which is controlled by the
1241 // speed feature tx_domain_dist_thres_level.
1242 //
1243 // The speed feature tx_domain_dist_level decides which of the above methods
1244 // needs to be used across different mode evaluation stages as described
1245 // below:
1246 // Eval type: Default Mode Winner
1247 // Level 0 : Method 0 Method 2 Method 0
1248 // Level 1 : Method 1 Method 2 Method 0
1249 // Level 2 : Method 2 Method 2 Method 0
1250 // Level 3 : Method 2 Method 2 Method 2
1251 int tx_domain_dist_level;
1252
1253 // Transform domain distortion threshold level
1254 int tx_domain_dist_thres_level;
1255
1256 // Trellis (dynamic programming) optimization of quantized values
1257 TRELLIS_OPT_TYPE optimize_coefficients;
1258
1259 // Use hash table to store macroblock RD search results
1260 // to avoid repeated search on the same residue signal.
1261 int use_mb_rd_hash;
1262
1263 // Flag used to control the extent of coeff R-D optimization
1264 int perform_coeff_opt;
1265} RD_CALC_SPEED_FEATURES;
1266
1267typedef struct WINNER_MODE_SPEED_FEATURES {
1268 // Flag used to control the winner mode processing for better R-D optimization
1269 // of quantized coeffs
1270 int enable_winner_mode_for_coeff_opt;
1271
1272 // Flag used to control the winner mode processing for transform size
1273 // search method
1274 int enable_winner_mode_for_tx_size_srch;
1275
1276 // Control transform size search level
1277 // Eval type: Default Mode Winner
1278 // Level 0 : FULL RD LARGEST ALL FULL RD
1279 // Level 1 : FAST RD LARGEST ALL FULL RD
1280 // Level 2 : LARGEST ALL LARGEST ALL FULL RD
1281 // Level 3 : LARGEST ALL LARGEST ALL LARGEST ALL
1282 int tx_size_search_level;
1283
1284 // Flag used to control the winner mode processing for use transform
1285 // domain distortion
1286 int enable_winner_mode_for_use_tx_domain_dist;
1287
1288 // Flag used to enable processing of multiple winner modes
1289 MULTI_WINNER_MODE_TYPE multi_winner_mode_type;
1290
1291 // Motion mode for winner candidates:
1292 // 0: speed feature OFF
1293 // 1 / 2 : Use configured number of winner candidates
1294 int motion_mode_for_winner_cand;
1295
1296 // Controls the prediction of transform skip block or DC only block.
1297 //
1298 // Different speed feature values (0 to 3) decide the aggressiveness of
1299 // prediction (refer to predict_dc_levels[][] in speed_features.c) to be used
1300 // during different mode evaluation stages.
1301 int dc_blk_pred_level;
1302
1303 // If on, disables interpolation filter search in handle_inter_mode loop, and
1304 // performs it during winner mode processing by \ref
1305 // tx_search_best_inter_candidates.
1306 int winner_mode_ifs;
1307
1308 // Controls the disabling of winner mode processing. Speed feature levels
1309 // are ordered in increasing aggressiveness of pruning. The method considered
1310 // for disabling, depends on the sf level value and it is described as below.
1311 // 0: Do not disable
1312 // 1: Disable for blocks with low source variance.
1313 // 2: Disable for blocks which turn out to be transform skip (skipped based on
1314 // eob) during MODE_EVAL stage except NEWMV mode.
1315 // 3: Disable for blocks which turn out to be transform skip during MODE_EVAL
1316 // stage except NEWMV mode. For high quantizers, prune conservatively based on
1317 // transform skip (skipped based on eob) except for NEWMV mode.
1318 // 4: Disable for blocks which turn out to be transform skip during MODE_EVAL
1319 // stage.
1320 int prune_winner_mode_eval_level;
1321} WINNER_MODE_SPEED_FEATURES;
1322
1323typedef struct LOOP_FILTER_SPEED_FEATURES {
1324 // This feature controls how the loop filter level is determined.
1325 LPF_PICK_METHOD lpf_pick;
1326
1327 // Skip some final iterations in the determination of the best loop filter
1328 // level.
1329 int use_coarse_filter_level_search;
1330
1331 // Control how the CDEF strength is determined.
1332 CDEF_PICK_METHOD cdef_pick_method;
1333
1334 // Decoder side speed feature to add penalty for use of dual-sgr filters.
1335 // Takes values 0 - 10, 0 indicating no penalty and each additional level
1336 // adding a penalty of 1%
1337 int dual_sgr_penalty_level;
1338
1339 // prune sgr ep using binary search like mechanism
1340 int enable_sgr_ep_pruning;
1341
1342 // Disable loop restoration for Chroma plane
1343 int disable_loop_restoration_chroma;
1344
1345 // Disable loop restoration for luma plane
1346 int disable_loop_restoration_luma;
1347
1348 // Prune RESTORE_WIENER evaluation based on source variance
1349 // 0 : no pruning
1350 // 1 : conservative pruning
1351 // 2 : aggressive pruning
1352 int prune_wiener_based_on_src_var;
1353
1354 // Prune self-guided loop restoration based on wiener search results
1355 // 0 : no pruning
1356 // 1 : pruning based on rdcost ratio of RESTORE_WIENER and RESTORE_NONE
1357 // 2 : pruning based on winner restoration type among RESTORE_WIENER and
1358 // RESTORE_NONE
1359 int prune_sgr_based_on_wiener;
1360
1361 // Reduce the wiener filter win size for luma
1362 int reduce_wiener_window_size;
1363
1364 // Disable loop restoration filter
1365 int disable_lr_filter;
1366
1367 // Whether to downsample the rows in computation of wiener stats.
1368 int use_downsampled_wiener_stats;
1369} LOOP_FILTER_SPEED_FEATURES;
1370
1371typedef struct REAL_TIME_SPEED_FEATURES {
1372 // check intra prediction for non-RD mode.
1373 int check_intra_pred_nonrd;
1374
1375 // Skip checking intra prediction.
1376 // 0 - don't skip
1377 // 1 - skip if TX is skipped and best mode is not NEWMV
1378 // 2 - skip if TX is skipped
1379 // Skipping aggressiveness increases from level 1 to 2.
1380 int skip_intra_pred;
1381
1382 // Perform coarse ME before calculating variance in variance-based partition
1383 int estimate_motion_for_var_based_partition;
1384
1385 // For nonrd_use_partition: mode of extra check of leaf partition
1386 // 0 - don't check merge
1387 // 1 - always check merge
1388 // 2 - check merge and prune checking final split
1389 // 3 - check merge and prune checking final split based on bsize and qindex
1390 int nonrd_check_partition_merge_mode;
1391
1392 // For nonrd_use_partition: check of leaf partition extra split
1393 int nonrd_check_partition_split;
1394
1395 // Implements various heuristics to skip searching modes
1396 // The heuristics selected are based on flags
1397 // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
1398 unsigned int mode_search_skip_flags;
1399
1400 // For nonrd: Reduces ref frame search.
1401 // 0 - low level of search prune in non last frames
1402 // 1 - pruned search in non last frames
1403 // 2 - more pruned search in non last frames
1404 int nonrd_prune_ref_frame_search;
1405
1406 // This flag controls the use of non-RD mode decision.
1407 int use_nonrd_pick_mode;
1408
1409 // Use ALTREF frame in non-RD mode decision.
1410 int use_nonrd_altref_frame;
1411
1412 // Use compound reference for non-RD mode.
1413 int use_comp_ref_nonrd;
1414
1415 // Reference frames for compound prediction for nonrd pickmode:
1416 // LAST_GOLDEN (0), LAST_LAST2 (1), or LAST_ALTREF (2).
1417 int ref_frame_comp_nonrd[3];
1418
1419 // use reduced ref set for real-time mode
1420 int use_real_time_ref_set;
1421
1422 // Skip a number of expensive mode evaluations for blocks with very low
1423 // temporal variance.
1424 int short_circuit_low_temp_var;
1425
1426 // Use modeled (currently CurvFit model) RDCost for fast non-RD mode
1427 int use_modeled_non_rd_cost;
1428
1429 // Reuse inter prediction in fast non-rd mode.
1430 int reuse_inter_pred_nonrd;
1431
1432 // Number of best inter modes to search transform. INT_MAX - search all.
1433 int num_inter_modes_for_tx_search;
1434
1435 // Use interpolation filter search in non-RD mode decision.
1436 int use_nonrd_filter_search;
1437
1438 // Use simplified RD model for interpolation search and Intra
1439 int use_simple_rd_model;
1440
1441 // If set forces interpolation filter to EIGHTTAP_REGULAR
1442 int skip_interp_filter_search;
1443
1444 // For nonrd mode: use hybrid intra mode search for intra only frames based on
1445 // block properties.
1446 // 0 : use nonrd pick intra for all blocks
1447 // 1 : use rd for bsize < 16x16, nonrd otherwise
1448 // 2 : use rd for bsize < 16x16 and src var >= 101, nonrd otherwise
1449 int hybrid_intra_pickmode;
1450
1451 // Compute variance/sse on source difference, prior to encoding superblock.
1452 int source_metrics_sb_nonrd;
1453
1454 // Flag to indicate process for handling overshoot on slide/scene change,
1455 // for real-time CBR mode.
1456 OVERSHOOT_DETECTION_CBR overshoot_detection_cbr;
1457
1458 // Check for scene/content change detection on every frame before encoding.
1459 int check_scene_detection;
1460
1461 // For nonrd mode: Prefer larger partition blks in variance based partitioning
1462 // 0: disabled, 1-4: increasing aggressiveness
1463 int prefer_large_partition_blocks;
1464
1465 // uses results of temporal noise estimate
1466 int use_temporal_noise_estimate;
1467
1468 // Parameter indicating initial search window to be used in full-pixel search
1469 // for nonrd_pickmode. Range [0, MAX_MVSEARCH_STEPS - 1]. Lower value
1470 // indicates larger window. If set to 0, step_param is set based on internal
1471 // logic in set_mv_search_params().
1472 int fullpel_search_step_param;
1473
1474 // Bit mask to enable or disable intra modes for each prediction block size
1475 // separately, for nonrd pickmode.
1476 int intra_y_mode_bsize_mask_nrd[BLOCK_SIZES];
1477
1478 // Skips mode checks more agressively in nonRD mode
1479 int nonrd_agressive_skip;
1480
1481 // Skip cdef on 64x64 blocks when NEWMV or INTRA is not picked or color
1482 // sensitivity is off. When color sensitivity is on for a superblock, all
1483 // 64x64 blocks within will not skip.
1484 int skip_cdef_sb;
1485
1486 // Forces larger partition blocks in variance based partitioning for intra
1487 // frames
1488 int force_large_partition_blocks_intra;
1489
1490 // Skip evaluation of no split in tx size selection for merge partition
1491 int skip_tx_no_split_var_based_partition;
1492
1493 // Intermediate termination of newMV mode evaluation based on so far best mode
1494 // sse
1495 int skip_newmv_mode_based_on_sse;
1496
1497 // Define gf length multiplier.
1498 // Level 0: use large multiplier, level 1: use medium multiplier.
1499 int gf_length_lvl;
1500
1501 // Prune inter modes with golden frame as reference for NEARMV and NEWMV modes
1502 int prune_inter_modes_with_golden_ref;
1503
1504 // Prune inter modes w.r.t golden or alt-ref frame based on sad
1505 int prune_inter_modes_wrt_gf_arf_based_on_sad;
1506
1507 // Prune inter mode search in rd path based on current block's temporal
1508 // variance wrt LAST reference.
1509 int prune_inter_modes_using_temp_var;
1510
1511 // Force half_pel at block level.
1512 int force_half_pel_block;
1513
1514 // Prune intra mode evaluation in inter frames based on mv range.
1515 BLOCK_SIZE prune_intra_mode_based_on_mv_range;
1516 // The number of times to left shift the splitting thresholds in variance
1517 // based partitioning. The minimum values should be 7 to avoid left shifting
1518 // by a negative number.
1519 int var_part_split_threshold_shift;
1520
1521 // Qindex based variance partition threshold index, which determines
1522 // the aggressiveness of partition pruning
1523 // 0: disabled for speeds 9,10
1524 // 1,2: (rd-path) lowers qindex thresholds conditionally (for low SAD sb)
1525 // 3,4: (non-rd path) uses pre-tuned qindex thresholds
1526 int var_part_based_on_qidx;
1527
1528 // Enable GF refresh based on Q value.
1529 int gf_refresh_based_on_qp;
1530
1531 // Temporal filtering
1532 int use_rtc_tf;
1533
1534 // Prune the use of the identity transform in nonrd_pickmode,
1535 // used for screen content mode: only for smaller blocks
1536 // and higher spatial variance, and when skip_txfm is not
1537 // already set.
1538 int prune_idtx_nonrd;
1539
1540 // Skip loopfilter, for static content after slide change
1541 // or key frame, once quality has ramped up.
1542 int skip_lf_screen;
1543
1544 // For nonrd: early exit out of variance partition that sets the
1545 // block size to superblock size, and sets mode to zeromv-last skip.
1546 int part_early_exit_zeromv;
1547
1548 // Early terminate inter mode search based on sse in non-rd path.
1549 INTER_SEARCH_EARLY_TERM_IDX sse_early_term_inter_search;
1550
1551 // SAD based adaptive altref selection
1552 int sad_based_adp_altref_lag;
1553
1554 // Enable/disable partition direct merging.
1555 int partition_direct_merging;
1556
1557 // SAD based compound mode pruning
1558 int sad_based_comp_prune;
1559
1560 // Level of aggressiveness for obtaining tx size based on qstep
1561 int tx_size_level_based_on_qstep;
1562
1563 // Reduce the mv resolution for zero mv if the variance is low.
1564 bool reduce_zeromv_mvres;
1565
1566 // Avoid the partitioning of a 16x16 block in variance based partitioning
1567 // (VBP) by making use of minimum and maximum sub-block variances.
1568 // For allintra encode, this speed feature reduces instruction count by 5.39%
1569 // for speed 9 on a typical video dataset with coding performance gain
1570 // of 1.44%.
1571 // For AVIF image encode, this speed feature reduces encode time
1572 // by 8.44% for speed 9 on a typical image dataset with coding performance
1573 // gain of 0.78%.
1574 bool vbp_prune_16x16_split_using_min_max_sub_blk_var;
1575
1576 // A qindex threshold that determines whether to use qindex based
1577 // CDEF filter strength estimation for screen content types.
1578 // This speed feature has a substantial gain on coding metrics,
1579 // with moderate increased encoding time.
1580 // Set to zero to turn off this speed feature.
1581 int screen_content_cdef_filter_qindex_thresh;
1582
1583 // Prunes global_globalmv search if its variance is \gt the globalmv's
1584 // variance.
1585 bool prune_global_globalmv_with_zeromv;
1586
1587 // Allow mode cost update at frame level every couple frames. This
1588 // overrides the command line setting --mode-cost-upd-freq=3 (never update
1589 // except on key frame and first delta).
1590 bool frame_level_mode_cost_update;
1591
1592 // If compound is enabled, and the current block size is \geq BLOCK_16X16,
1593 // limit the compound modes to GLOBAL_GLOBALMV. This does not apply to the
1594 // base layer of svc.
1595 bool check_only_zero_zeromv_on_large_blocks;
1596
1597 // Allow for disabling cdf update for non reference frames in svc mode.
1598 bool disable_cdf_update_non_reference_frame;
1599} REAL_TIME_SPEED_FEATURES;
1600
1606typedef struct SPEED_FEATURES {
1611
1616
1620 TPL_SPEED_FEATURES tpl_sf;
1621
1625 GLOBAL_MOTION_SPEED_FEATURES gm_sf;
1626
1630 PARTITION_SPEED_FEATURES part_sf;
1631
1635 MV_SPEED_FEATURES mv_sf;
1636
1640 INTER_MODE_SPEED_FEATURES inter_sf;
1641
1645 INTERP_FILTER_SPEED_FEATURES interp_sf;
1646
1650 INTRA_MODE_SPEED_FEATURES intra_sf;
1651
1655 TX_SPEED_FEATURES tx_sf;
1656
1660 RD_CALC_SPEED_FEATURES rd_sf;
1661
1665 WINNER_MODE_SPEED_FEATURES winner_mode_sf;
1666
1670 LOOP_FILTER_SPEED_FEATURES lpf_sf;
1671
1675 REAL_TIME_SPEED_FEATURES rt_sf;
1679struct AV1_COMP;
1680
1694 int speed);
1695
1708 int speed);
1720void av1_set_speed_features_qindex_dependent(struct AV1_COMP *cpi, int speed);
1721
1722#ifdef __cplusplus
1723} // extern "C"
1724#endif
1725
1726#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:2476
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:325
@ INTERNAL_COST_UPD_OFF
Definition speed_features.h:326
@ INTERNAL_COST_UPD_SBROW_SET
Definition speed_features.h:328
@ INTERNAL_COST_UPD_SBROW
Definition speed_features.h:329
@ INTERNAL_COST_UPD_SB
Definition speed_features.h:330
@ INTERNAL_COST_UPD_TILE
Definition speed_features.h:327
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:369
@ EARLY_TERM_INDICES
Definition speed_features.h:380
@ EARLY_TERM_IDX_1
Definition speed_features.h:372
@ EARLY_TERM_IDX_4
Definition speed_features.h:378
@ EARLY_TERM_IDX_2
Definition speed_features.h:374
@ EARLY_TERM_DISABLED
Definition speed_features.h:370
@ EARLY_TERM_IDX_3
Definition speed_features.h:376
SIMPLE_MOTION_SEARCH_PRUNE_LEVEL
This enumeration defines a variety of simple motion search based partition prune levels.
Definition speed_features.h:337
@ QIDX_BASED_AGG_LVL1
Definition speed_features.h:343
@ SIMPLE_AGG_LVL3
Definition speed_features.h:342
@ TOTAL_SIMPLE_AGG_LVLS
Definition speed_features.h:346
@ SIMPLE_AGG_LVL1
Definition speed_features.h:340
@ SIMPLE_AGG_LVL0
Definition speed_features.h:339
@ TOTAL_AGG_LVLS
Definition speed_features.h:352
@ SIMPLE_AGG_LVL2
Definition speed_features.h:341
@ TOTAL_QINDEX_BASED_AGG_LVLS
Definition speed_features.h:348
PRUNE_MESH_SEARCH_LEVEL
This enumeration defines a variety of mesh search prune levels.
Definition speed_features.h:359
@ PRUNE_MESH_SEARCH_LVL_1
Definition speed_features.h:361
@ PRUNE_MESH_SEARCH_LVL_2
Definition speed_features.h:362
@ PRUNE_MESH_SEARCH_DISABLED
Definition speed_features.h:360
CDEF_PICK_METHOD
This enumeration defines a variety of CDEF pick methods.
Definition speed_features.h:156
@ CDEF_FAST_SEARCH_LVL2
Definition speed_features.h:159
@ CDEF_FAST_SEARCH_LVL5
Definition speed_features.h:163
@ CDEF_FAST_SEARCH_LVL1
Definition speed_features.h:158
@ CDEF_FULL_SEARCH
Definition speed_features.h:157
@ CDEF_PICK_FROM_Q
Definition speed_features.h:164
@ CDEF_FAST_SEARCH_LVL4
Definition speed_features.h:162
@ CDEF_FAST_SEARCH_LVL3
Definition speed_features.h:160
Definition speed_features.h:443
int disable_recon
Skips reconstruction by using source buffers for prediction.
Definition speed_features.h:460
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:450
int skip_zeromv_motion_search
Skips the motion search centered on 0,0 mv.
Definition speed_features.h:465
int skip_motion_search_threshold
Skips the motion search when the zero mv has small sse.
Definition speed_features.h:455
Sequence/frame level speed vs quality features.
Definition speed_features.h:386
int second_alt_ref_filtering
Definition speed_features.h:431
int frame_parameter_update
Definition speed_features.h:388
MV_PREC_LOGIC high_precision_mv_usage
Definition speed_features.h:407
int disable_extra_sc_testing
Definition speed_features.h:426
int recode_tolerance
Definition speed_features.h:399
SUPERRES_AUTO_SEARCH_TYPE superres_auto_search_type
Definition speed_features.h:421
int static_segmentation
Definition speed_features.h:416
int num_frames_used_in_tf
Definition speed_features.h:437
RECODE_LOOP_TYPE recode_loop
Definition speed_features.h:393
Top level speed vs quality trade off data struture.
Definition speed_features.h:1606
MV_SPEED_FEATURES mv_sf
Definition speed_features.h:1635
TPL_SPEED_FEATURES tpl_sf
Definition speed_features.h:1620
LOOP_FILTER_SPEED_FEATURES lpf_sf
Definition speed_features.h:1670
TX_SPEED_FEATURES tx_sf
Definition speed_features.h:1655
INTER_MODE_SPEED_FEATURES inter_sf
Definition speed_features.h:1640
RD_CALC_SPEED_FEATURES rd_sf
Definition speed_features.h:1660
PARTITION_SPEED_FEATURES part_sf
Definition speed_features.h:1630
GLOBAL_MOTION_SPEED_FEATURES gm_sf
Definition speed_features.h:1625
INTERP_FILTER_SPEED_FEATURES interp_sf
Definition speed_features.h:1645
FIRST_PASS_SPEED_FEATURES fp_sf
Definition speed_features.h:1615
INTRA_MODE_SPEED_FEATURES intra_sf
Definition speed_features.h:1650
WINNER_MODE_SPEED_FEATURES winner_mode_sf
Definition speed_features.h:1665
REAL_TIME_SPEED_FEATURES rt_sf
Definition speed_features.h:1675
HIGH_LEVEL_SPEED_FEATURES hl_sf
Definition speed_features.h:1610