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);
156typedef enum {
165 CDEF_PICK_METHODS
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} UENUM1BYTE(MULTI_WINNER_MODE_TYPE);
224
225enum {
226 PRUNE_NEARMV_OFF = 0, // Turn off nearmv pruning
227 PRUNE_NEARMV_LEVEL1 = 1, // Prune nearmv for qindex (0-85)
228 PRUNE_NEARMV_LEVEL2 = 2, // Prune nearmv for qindex (0-170)
229 PRUNE_NEARMV_LEVEL3 = 3, // Prune nearmv more aggressively for qindex (0-170)
230 PRUNE_NEARMV_MAX = PRUNE_NEARMV_LEVEL3,
231} UENUM1BYTE(PRUNE_NEARMV_LEVEL);
232
233typedef struct {
234 TX_TYPE_PRUNE_MODE prune_2d_txfm_mode;
235 int fast_intra_tx_type_search;
236
237 // INT_MAX: Disable fast search.
238 // 1 - 1024: Probability threshold used for conditionally forcing tx type,
239 // during mode search.
240 // 0: Force tx type to be DCT_DCT unconditionally, during
241 // mode search.
242 int fast_inter_tx_type_prob_thresh;
243
244 // Prune less likely chosen transforms for each intra mode. The speed
245 // feature ranges from 0 to 2, for different speed / compression trade offs.
246 int use_reduced_intra_txset;
247
248 // Use a skip flag prediction model to detect blocks with skip = 1 early
249 // and avoid doing full TX type search for such blocks.
250 int use_skip_flag_prediction;
251
252 // Threshold used by the ML based method to predict TX block split decisions.
253 int ml_tx_split_thresh;
254
255 // skip remaining transform type search when we found the rdcost of skip is
256 // better than applying transform
257 int skip_tx_search;
258
259 // Prune tx type search using previous frame stats.
260 int prune_tx_type_using_stats;
261 // Prune tx type search using estimated RDcost
262 int prune_tx_type_est_rd;
263
264 // Flag used to control the winner mode processing for tx type pruning for
265 // inter blocks. It enables further tx type mode pruning based on ML model for
266 // mode evaluation and disables tx type mode pruning for winner mode
267 // processing.
268 int winner_mode_tx_type_pruning;
269} TX_TYPE_SEARCH;
270
271enum {
272 // Search partitions using RD criterion
273 SEARCH_PARTITION,
274
275 // Always use a fixed size partition
276 FIXED_PARTITION,
277
278 // Partition using source variance
279 VAR_BASED_PARTITION,
280
281#if CONFIG_RT_ML_PARTITIONING
282 // Partition using ML model
283 ML_BASED_PARTITION
284#endif
285} UENUM1BYTE(PARTITION_SEARCH_TYPE);
286
287enum {
288 NOT_IN_USE,
289 DIRECT_PRED,
290 RELAXED_PRED,
291 ADAPT_PRED
292} UENUM1BYTE(MAX_PART_PRED_MODE);
293
294enum {
295 LAST_MV_DATA,
296 CURRENT_Q,
297 QTR_ONLY,
298} UENUM1BYTE(MV_PREC_LOGIC);
299
300enum {
301 SUPERRES_AUTO_ALL, // Tries all possible superres ratios
302 SUPERRES_AUTO_DUAL, // Tries no superres and q-based superres ratios
303 SUPERRES_AUTO_SOLO, // Only apply the q-based superres ratio
304} UENUM1BYTE(SUPERRES_AUTO_SEARCH_TYPE);
315typedef enum {
321
326typedef enum {
327 NO_PRUNING = -1,
344
351
355 RECODE_LOOP_TYPE recode_loop;
356
362
370
379
383 SUPERRES_AUTO_SEARCH_TYPE superres_auto_search_type;
384
389
395
407
412
417
423
425typedef struct TPL_SPEED_FEATURES {
426 // GOP length adaptive decision.
427 // If set to 0, tpl model decides whether a shorter gf interval is better.
428 // If set to 1, tpl stats of ARFs from base layer, (base+1) layer and
429 // (base+2) layer decide whether a shorter gf interval is better.
430 // If set to 2, tpl stats of ARFs from base layer, (base+1) layer and GF boost
431 // decide whether a shorter gf interval is better.
432 // If set to 3, gop length adaptive decision is disabled.
433 int gop_length_decision_method;
434 // Prune the intra modes search by tpl.
435 // If set to 0, we will search all intra modes from DC_PRED to PAETH_PRED.
436 // If set to 1, we only search DC_PRED, V_PRED, and H_PRED.
437 int prune_intra_modes;
438 // This parameter controls which step in the n-step process we start at.
439 int reduce_first_step_size;
440 // Skip motion estimation based on the precision of center MVs and the
441 // difference between center MVs.
442 // If set to 0, motion estimation is skipped for duplicate center MVs
443 // (default). If set to 1, motion estimation is skipped for duplicate
444 // full-pixel center MVs. If set to 2, motion estimation is skipped if the
445 // difference between center MVs is less than the threshold.
446 int skip_alike_starting_mv;
447
448 // When to stop subpel search.
449 SUBPEL_FORCE_STOP subpel_force_stop;
450
451 // Which search method to use.
452 SEARCH_METHODS search_method;
453
454 // Prune starting mvs in TPL based on sad scores.
455 int prune_starting_mv;
456
457 // Not run TPL for filtered Key frame.
458 int disable_filtered_key_tpl;
459
460 // Prune reference frames in TPL.
461 int prune_ref_frames_in_tpl;
462
463 // Support compound predictions.
464 int allow_compound_pred;
465
466 // Calculate rate and distortion based on Y plane only.
467 int use_y_only_rate_distortion;
468} TPL_SPEED_FEATURES;
469
470typedef struct GLOBAL_MOTION_SPEED_FEATURES {
471 GM_SEARCH_TYPE gm_search_type;
472
473 // During global motion estimation, prune remaining reference frames in a
474 // given direction(past/future), if the evaluated ref_frame in that direction
475 // yields gm_type as INVALID/TRANSLATION/IDENTITY
476 int prune_ref_frame_for_gm_search;
477
478 // When the current GM type is set to ZEROMV, prune ZEROMV if its performance
479 // is worse than NEWMV under SSE metric.
480 // 0 : no pruning
481 // 1 : conservative pruning
482 // 2 : aggressive pruning
484
485 // Disable global motion estimation based on stats of previous frames in the
486 // GF group
487 int disable_gm_search_based_on_stats;
488} GLOBAL_MOTION_SPEED_FEATURES;
489
490typedef struct PARTITION_SPEED_FEATURES {
491 PARTITION_SEARCH_TYPE partition_search_type;
492
493 // Used if partition_search_type = FIXED_PARTITION
494 BLOCK_SIZE fixed_partition_size;
495
496 // Prune extended partition types search
497 // Can take values 0 - 2, 0 referring to no pruning, and 1 - 2 increasing
498 // aggressiveness of pruning in order.
499 int prune_ext_partition_types_search_level;
500
501 // Prune part4 based on block size
502 int prune_part4_search;
503
504 // Use a ML model to prune rectangular, ab and 4-way horz
505 // and vert partitions
506 int ml_prune_partition;
507
508 // Use a ML model to adaptively terminate partition search after trying
509 // PARTITION_SPLIT. Can take values 0 - 2, 0 meaning not being enabled, and
510 // 1 - 2 increasing aggressiveness in order.
511 int ml_early_term_after_part_split_level;
512
513 // Skip rectangular partition test when partition type none gives better
514 // rd than partition type split. Can take values 0 - 2, 0 referring to no
515 // skipping, and 1 - 2 increasing aggressiveness of skipping in order.
516 int less_rectangular_check_level;
517
518 // Use square partition only beyond this block size.
519 BLOCK_SIZE use_square_partition_only_threshold;
520
521 // Sets max square partition levels for this superblock based on
522 // motion vector and prediction error distribution produced from 16x16
523 // simple motion search
524 MAX_PART_PRED_MODE auto_max_partition_based_on_simple_motion;
525
526 // Min and max square partition size we enable (block_size) as per auto
527 // min max, but also used by adjust partitioning, and pick_partitioning.
528 BLOCK_SIZE default_min_partition_size;
529 BLOCK_SIZE default_max_partition_size;
530
531 // Sets level of adjustment of variance-based partitioning during
532 // rd_use_partition 0 - no partition adjustment, 1 - try to merge partitions
533 // for small blocks and high QP, 2 - try to merge partitions, 3 - always try
534 // to merge leaf partitions for small blocks, 4 - try to merge and split leaf
535 // partitions and 0 - 4 decreasing aggressiveness in order.
536 int adjust_var_based_rd_partitioning;
537
538 // Partition search early breakout thresholds.
539 int64_t partition_search_breakout_dist_thr;
540 int partition_search_breakout_rate_thr;
541
542 // Thresholds for ML based partition search breakout.
543 int ml_partition_search_breakout_thresh[PARTITION_BLOCK_SIZES];
544
545 // Aggressiveness levels for pruning split and rectangular partitions based on
546 // simple_motion_search. SIMPLE_AGG_LVL0 to SIMPLE_AGG_LVL3 correspond to
547 // simple motion search based pruning. QIDX_BASED_AGG_LVL1 corresponds to
548 // qindex based and simple motion search based pruning.
549 int simple_motion_search_prune_agg;
550
551 // Perform simple_motion_search on each possible subblock and use it to prune
552 // PARTITION_HORZ and PARTITION_VERT.
553 int simple_motion_search_prune_rect;
554
555 // Perform simple motion search before none_partition to decide if we
556 // want to remove all partitions other than PARTITION_SPLIT. If set to 0, this
557 // model is disabled. If set to 1, the model attempts to perform
558 // PARTITION_SPLIT only. If set to 2, the model also attempts to prune
559 // PARTITION_SPLIT.
560 int simple_motion_search_split;
561
562 // Use features from simple_motion_search to terminate prediction block
563 // partition after PARTITION_NONE
564 int simple_motion_search_early_term_none;
565
566 // Controls whether to reduce the number of motion search steps. If this is 0,
567 // then simple_motion_search has the same number of steps as
568 // single_motion_search (assuming no other speed features). Otherwise, reduce
569 // the number of steps by the value contained in this variable.
570 int simple_motion_search_reduce_search_steps;
571
572 // This variable controls the maximum block size where intra blocks can be
573 // used in inter frames.
574 // TODO(aconverse): Fold this into one of the other many mode skips
575 BLOCK_SIZE max_intra_bsize;
576
577 // Use CNN with luma pixels on source frame on each of the 64x64 subblock to
578 // perform partition pruning in intra frames.
579 // 0: No Pruning
580 // 1: Prune split and rectangular partitions only
581 // 2: Prune none, split and rectangular partitions
582 int intra_cnn_based_part_prune_level;
583
584 // Disable extended partition search for lower block sizes.
585 int ext_partition_eval_thresh;
586
587 // Disable rectangular partitions for larger block sizes.
588 int rect_partition_eval_thresh;
589
590 // prune extended partition search
591 // 0 : no pruning
592 // 1 : prune 1:4 partition search using winner info from split partitions
593 // 2 : prune 1:4 and AB partition search using split and HORZ/VERT info
594 int prune_ext_part_using_split_info;
595
596 // Prunt rectangular, AB and 4-way partition based on q index and block size
597 // 0 : no pruning
598 // 1 : prune sub_8x8 at very low quantizers
599 // 2 : prune all block size based on qindex
600 int prune_rectangular_split_based_on_qidx;
601
602 // Terminate partition search for child partition,
603 // when NONE and SPLIT partition rd_costs are INT64_MAX.
604 int early_term_after_none_split;
605
606 // Level used to adjust threshold for av1_ml_predict_breakout(). At lower
607 // levels, more conservative threshold is used, and value of 0 indicates
608 // av1_ml_predict_breakout() is disabled. Value of 3 corresponds to default
609 // case with no adjustment to lbd thresholds.
610 int ml_predict_breakout_level;
611
612 // Prune sub_8x8 (BLOCK_4X4, BLOCK_4X8 and BLOCK_8X4) partitions.
613 // 0 : no pruning
614 // 1 : pruning based on neighbour block information
615 // 2 : prune always
616 int prune_sub_8x8_partition_level;
617
618 // Prune rectangular split based on simple motion search split/no_split score.
619 // 0: disable pruning, 1: enable pruning
620 int simple_motion_search_rect_split;
621
622 // The current encoder adopts a DFS search for block partitions.
623 // Therefore the mode selection and associated rdcost is ready for smaller
624 // blocks before the mode selection for some partition types.
625 // AB partition could use previous rd information and skip mode search.
626 // An example is:
627 //
628 // current block
629 // +---+---+
630 // | |
631 // + +
632 // | |
633 // +-------+
634 //
635 // SPLIT partition has been searched first before trying HORZ_A
636 // +---+---+
637 // | R | R |
638 // +---+---+
639 // | R | R |
640 // +---+---+
641 //
642 // HORZ_A
643 // +---+---+
644 // | | |
645 // +---+---+
646 // | |
647 // +-------+
648 //
649 // With this speed feature, the top two sub blocks can directly use rdcost
650 // searched in split partition, and the mode info is also copied from
651 // saved info. Similarly, the bottom rectangular block can also use
652 // the available information from previous rectangular search.
653 int reuse_prev_rd_results_for_part_ab;
654
655 // Reuse the best prediction modes found in PARTITION_SPLIT and PARTITION_RECT
656 // when encoding PARTITION_AB.
657 int reuse_best_prediction_for_part_ab;
658
659 // The current partition search records the best rdcost so far and uses it
660 // in mode search and transform search to early skip when some criteria is
661 // met. For example, when the current rdcost is larger than the best rdcost,
662 // or the model rdcost is larger than the best rdcost times some thresholds.
663 // By default, this feature is turned on to speed up the encoder partition
664 // search.
665 // If disabling it, at speed 0, 30 frames, we could get
666 // about -0.25% quality gain (psnr, ssim, vmaf), with about 13% slowdown.
667 int use_best_rd_for_pruning;
668
669 // Skip evaluation of non-square partitions based on the corresponding NONE
670 // partition.
671 // 0: no pruning
672 // 1: prune extended partitions if NONE is skippable
673 // 2: on top of 1, prune rectangular partitions if NONE is inter, not a newmv
674 // mode and skippable
675 int skip_non_sq_part_based_on_none;
676} PARTITION_SPEED_FEATURES;
677
678typedef struct MV_SPEED_FEATURES {
679 // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
680 SEARCH_METHODS search_method;
681
682 // Enable the use of faster, less accurate mv search method on bsize >=
683 // BLOCK_32X32.
684 // TODO(chiyotsai@google.com): Take the clip's resolution and mv activity into
685 // account.
686 int use_bsize_dependent_search_method;
687
688 // If this is set to 1, we limit the motion search range to 2 times the
689 // largest motion vector found in the last frame.
690 int auto_mv_step_size;
691
692 // Subpel_search_method can only be subpel_tree which does a subpixel
693 // logarithmic search that keeps stepping at 1/2 pixel units until
694 // you stop getting a gain, and then goes on to 1/4 and repeats
695 // the same process. Along the way it skips many diagonals.
696 SUBPEL_SEARCH_METHODS subpel_search_method;
697
698 // Maximum number of steps in logarithmic subpel search before giving up.
699 int subpel_iters_per_step;
700
701 // When to stop subpel search.
702 SUBPEL_FORCE_STOP subpel_force_stop;
703
704 // When to stop subpel search in simple motion search.
705 SUBPEL_FORCE_STOP simple_motion_subpel_force_stop;
706
707 // If true, sub-pixel search uses the exact convolve function used for final
708 // encoding and decoding; otherwise, it uses bilinear interpolation.
709 SUBPEL_SEARCH_TYPE use_accurate_subpel_search;
710
711 // Threshold for allowing exhaustive motion search.
712 int exhaustive_searches_thresh;
713
714 // Pattern to be used for any exhaustive mesh searches (except intraBC ME).
715 MESH_PATTERN mesh_patterns[MAX_MESH_STEP];
716
717 // Pattern to be used for exhaustive mesh searches of intraBC ME.
718 MESH_PATTERN intrabc_mesh_patterns[MAX_MESH_STEP];
719
720 // Reduce single motion search range based on MV result of prior ref_mv_idx.
721 int reduce_search_range;
722
723 // Prune mesh search.
724 int prune_mesh_search;
725
726 // Use the rd cost around the best FULLPEL_MV to speed up subpel search
727 int use_fullpel_costlist;
728
729 // Set the full pixel search level of obmc
730 // 0: obmc_full_pixel_diamond
731 // 1: obmc_refining_search_sad (faster)
732 int obmc_full_pixel_search_level;
733
734 // Accurate full pixel motion search based on TPL stats.
735 int full_pixel_search_level;
736
737 // Whether to downsample the rows in sad calculation during motion search.
738 // This is only active when there are at least 16 rows.
739 int use_downsampled_sad;
740
741 // Enable/disable extensive joint motion search.
742 int disable_extensive_joint_motion_search;
743
744 // Enable second best mv check in joint mv search.
745 // 0: allow second MV (use rd cost as the metric)
746 // 1: use var as the metric
747 // 2: disable second MV
748 int disable_second_mv;
749} MV_SPEED_FEATURES;
750
751typedef struct INTER_MODE_SPEED_FEATURES {
752 // 2-pass inter mode model estimation where the preliminary pass skips
753 // transform search and uses a model to estimate rd, while the final pass
754 // computes the full transform search. Two types of models are supported:
755 // 0: not used
756 // 1: used with online dynamic rd model
757 // 2: used with static rd model
758 int inter_mode_rd_model_estimation;
759
760 // Bypass transform search based on skip rd
761 int txfm_rd_gate_level;
762
763 // Limit the inter mode tested in the RD loop
764 int reduce_inter_modes;
765
766 // This variable is used to cap the maximum number of times we skip testing a
767 // mode to be evaluated. A high value means we will be faster.
768 int adaptive_rd_thresh;
769
770 // Aggressively prune inter modes when best mode is skippable.
771 int prune_inter_modes_if_skippable;
772
773 // Drop less likely to be picked reference frames in the RD search.
774 // Has seven levels for now: 0, 1, 2, 3, 4, 5 and 6 where higher levels prune
775 // more aggressively than lower ones. (0 means no pruning).
776 int selective_ref_frame;
777
778 // Prune reference frames for rectangular partitions.
779 // 0 implies no pruning
780 // 1 implies prune for extended partition
781 // 2 implies prune horiz, vert and extended partition
782 int prune_ref_frame_for_rect_partitions;
783
784 // Prune inter modes w.r.t past reference frames
785 // 0 no pruning
786 // 1 prune inter modes w.r.t ALTREF2 and ALTREF reference frames
787 // 2 prune inter modes w.r.t BWDREF, ALTREF2 and ALTREF reference frames
788 int alt_ref_search_fp;
789
790 // Skip the current ref_mv in NEW_MV mode based on mv, rate cost, etc.
791 // This speed feature equaling 0 means no skipping.
792 // If the speed feature equals 1 or 2, skip the current ref_mv in NEW_MV mode
793 // if we have already encountered ref_mv in the drl such that:
794 // 1. The other drl has the same mv during the SIMPLE_TRANSLATION search
795 // process as the current mv.
796 // 2. The rate needed to encode the current mv is larger than that for the
797 // other ref_mv.
798 // The speed feature equaling 1 means using subpel mv in the comparison.
799 // The speed feature equaling 2 means using fullpel mv in the comparison.
800 // If the speed feature >= 3, skip the current ref_mv in NEW_MV mode based on
801 // known full_mv bestsme and drl cost.
802 int skip_newmv_in_drl;
803
804 // This speed feature checks duplicate ref MVs among NEARESTMV, NEARMV,
805 // GLOBALMV and skips NEARMV or GLOBALMV (in order) if a duplicate is found
806 // TODO(any): Instead of skipping repeated ref mv, use the recalculated
807 // rd-cost based on mode rate and skip the mode evaluation
808 int skip_repeated_ref_mv;
809
810 // Flag used to control the ref_best_rd based gating for chroma
811 int perform_best_rd_based_gating_for_chroma;
812
813 // Reuse the inter_intra_mode search result from NEARESTMV mode to other
814 // single ref modes
815 int reuse_inter_intra_mode;
816
817 // prune wedge and compound segment approximate rd evaluation based on
818 // compound average modeled rd
819 int prune_comp_type_by_model_rd;
820
821 // prune wedge and compound segment approximate rd evaluation based on
822 // compound average rd/ref_best_rd
823 int prune_comp_type_by_comp_avg;
824
825 // Skip some ref frames in compound motion search by single motion search
826 // result. Has three levels for now: 0 referring to no skipping, and 1 - 3
827 // increasing aggressiveness of skipping in order.
828 // Note: The search order might affect the result. It assumes that the single
829 // reference modes are searched before compound modes. It is better to search
830 // same single inter mode as a group.
831 int prune_comp_search_by_single_result;
832
833 // Instead of performing a full MV search, do a simple translation first
834 // and only perform a full MV search on the motion vectors that performed
835 // well.
836 int prune_mode_search_simple_translation;
837
838 // Only search compound modes with at least one "good" reference frame.
839 // A reference frame is good if, after looking at its performance among
840 // the single reference modes, it is one of the two best performers.
841 int prune_compound_using_single_ref;
842
843 // Skip extended compound mode (NEAREST_NEWMV, NEW_NEARESTMV, NEAR_NEWMV,
844 // NEW_NEARMV) using ref frames of above and left neighbor
845 // blocks.
846 // 0 : no pruning
847 // 1 : prune ext compound modes using neighbor blocks (less aggressiveness)
848 // 2 : prune ext compound modes using neighbor blocks (high aggressiveness)
849 // 3 : prune ext compound modes unconditionally (highest aggressiveness)
850 int prune_ext_comp_using_neighbors;
851
852 // Skip NEW_NEARMV and NEAR_NEWMV extended compound modes
853 int skip_ext_comp_nearmv_mode;
854
855 // Skip extended compound mode when ref frame corresponding to NEWMV does not
856 // have NEWMV as single mode winner.
857 // 0 : no pruning
858 // 1 : prune extended compound mode (less aggressiveness)
859 // 2 : prune extended compound mode (high aggressiveness)
860 int prune_comp_using_best_single_mode_ref;
861
862 // Skip NEARESTMV and NEARMV using weight computed in ref mv list population
863 int prune_nearest_near_mv_using_refmv_weight;
864
865 // Based on previous ref_mv_idx search result, prune the following search.
866 int prune_ref_mv_idx_search;
867
868 // Disable one sided compound modes.
869 int disable_onesided_comp;
870
871 // Prune obmc search using previous frame stats.
872 // INT_MAX : disable obmc search
873 int prune_obmc_prob_thresh;
874
875 // Prune warped motion search using previous frame stats.
876 int prune_warped_prob_thresh;
877
878 // Variance threshold to enable/disable Interintra wedge search
879 unsigned int disable_interintra_wedge_var_thresh;
880
881 // Variance threshold to enable/disable Interinter wedge search
882 unsigned int disable_interinter_wedge_var_thresh;
883
884 // De-couple wedge and mode search during interintra RDO.
885 int fast_interintra_wedge_search;
886
887 // Whether fast wedge sign estimate is used
888 int fast_wedge_sign_estimate;
889
890 // Enable/disable ME for interinter wedge search.
891 int disable_interinter_wedge_newmv_search;
892
893 // Decide when and how to use joint_comp.
894 DIST_WTD_COMP_FLAG use_dist_wtd_comp_flag;
895
896 // Clip the frequency of updating the mv cost.
897 INTERNAL_COST_UPDATE_TYPE mv_cost_upd_level;
898
899 // Clip the frequency of updating the coeff cost.
900 INTERNAL_COST_UPDATE_TYPE coeff_cost_upd_level;
901
902 // Clip the frequency of updating the mode cost.
903 INTERNAL_COST_UPDATE_TYPE mode_cost_upd_level;
904
905 // Prune inter modes based on tpl stats
906 // 0 : no pruning
907 // 1 - 3 indicate increasing aggressiveness in order.
908 int prune_inter_modes_based_on_tpl;
909
910 // Skip NEARMV and NEAR_NEARMV modes using ref frames of above and left
911 // neighbor blocks and qindex.
912 PRUNE_NEARMV_LEVEL prune_nearmv_using_neighbors;
913
914 // Model based breakout after interpolation filter search
915 // 0: no breakout
916 // 1: use model based rd breakout
917 int model_based_post_interp_filter_breakout;
918
919 // Reuse compound type rd decision when exact match is found
920 // 0: No reuse
921 // 1: Reuse the compound type decision
922 int reuse_compound_type_decision;
923
924 // Enable/disable masked compound.
925 int disable_masked_comp;
926
927 // Enable/disable the fast compound mode search.
928 int enable_fast_compound_mode_search;
929
930 // Reuse masked compound type search results
931 int reuse_mask_search_results;
932
933 // Enable/disable fast search for wedge masks
934 int enable_fast_wedge_mask_search;
935
936 // Early breakout from transform search of inter modes
937 int inter_mode_txfm_breakout;
938
939 // Limit number of inter modes for txfm search if a newmv mode gets
940 // evaluated among the top modes.
941 // 0: no pruning
942 // 1 to 3 indicate increasing order of aggressiveness
943 int limit_inter_mode_cands;
944
945 // Cap the no. of txfm searches for a given prediction mode.
946 // 0: no cap, 1: cap beyond first 4 searches, 2: cap beyond first 3 searches.
947 int limit_txfm_eval_per_mode;
948
949 // Prune warped motion search based on block size.
950 int extra_prune_warped;
951
952 // Do not search compound modes for ARF.
953 // The intuition is that ARF is predicted by frames far away from it,
954 // whose temporal correlations with the ARF are likely low.
955 // It is therefore likely that compound modes do not work as well for ARF
956 // as other inter frames.
957 // Speed/quality impact:
958 // Speed 1: 12% faster, 0.1% psnr loss.
959 // Speed 2: 2% faster, 0.05% psnr loss.
960 // No change for speed 3 and up, because |disable_onesided_comp| is true.
961 int skip_arf_compound;
962} INTER_MODE_SPEED_FEATURES;
963
964typedef struct INTERP_FILTER_SPEED_FEATURES {
965 // Do limited interpolation filter search for dual filters, since best choice
966 // usually includes EIGHTTAP_REGULAR.
967 int use_fast_interpolation_filter_search;
968
969 // Disable dual filter
970 int disable_dual_filter;
971
972 // Save results of av1_interpolation_filter_search for a block
973 // Check mv and ref_frames before search, if they are very close with previous
974 // saved results, filter search can be skipped.
975 int use_interp_filter;
976
977 // skip sharp_filter evaluation based on regular and smooth filter rd for
978 // dual_filter=0 case
979 int skip_sharp_interp_filter_search;
980
981 int cb_pred_filter_search;
982
983 // adaptive interp_filter search to allow skip of certain filter types.
984 int adaptive_interp_filter_search;
985} INTERP_FILTER_SPEED_FEATURES;
986
987typedef struct INTRA_MODE_SPEED_FEATURES {
988 // These bit masks allow you to enable or disable intra modes for each
989 // transform size separately.
990 int intra_y_mode_mask[TX_SIZES];
991 int intra_uv_mode_mask[TX_SIZES];
992
993 // flag to allow skipping intra mode for inter frame prediction
994 int skip_intra_in_interframe;
995
996 // Prune intra mode candidates based on source block histogram of gradient.
997 // Applies to luma plane only.
998 // Feasible values are 0..4. The feature is disabled for 0. An increasing
999 // value indicates more aggressive pruning threshold.
1000 int intra_pruning_with_hog;
1001
1002 // Prune intra mode candidates based on source block histogram of gradient.
1003 // Applies to chroma plane only.
1004 // Feasible values are 0..4. The feature is disabled for 0. An increasing
1005 // value indicates more aggressive pruning threshold.
1006 int chroma_intra_pruning_with_hog;
1007
1008 // Enable/disable smooth intra modes.
1009 int disable_smooth_intra;
1010
1011 // Prune filter intra modes in intra frames.
1012 // 0 : No pruning
1013 // 1 : Evaluate applicable filter intra modes based on best intra mode so far
1014 // 2 : Do not evaluate filter intra modes
1015 int prune_filter_intra_level;
1016
1017 // prune palette search
1018 // 0: No pruning
1019 // 1: Perform coarse search to prune the palette colors. For winner colors,
1020 // neighbors are also evaluated using a finer search.
1021 // 2: Perform 2 way palette search from max colors to min colors (and min
1022 // colors to remaining colors) and terminate the search if current number of
1023 // palette colors is not the winner.
1024 int prune_palette_search_level;
1025
1026 // Terminate early in luma palette_size search. Speed feature values indicate
1027 // increasing level of pruning.
1028 // 0: No early termination
1029 // 1: Terminate early for higher luma palette_size, if header rd cost of lower
1030 // palette_size is more than 2 * best_rd. This level of pruning is more
1031 // conservative when compared to sf level 2 as the cases which will get pruned
1032 // with sf level 1 is a subset of the cases which will get pruned with sf
1033 // level 2.
1034 // 2: Terminate early for higher luma palette_size, if header rd cost of lower
1035 // palette_size is more than best_rd.
1036 // For allintra encode, this sf reduces instruction count by 2.49%, 1.07%,
1037 // 2.76%, 2.30%, 1.84%, 2.69%, 2.04%, 2.05% and 1.44% for speed 0, 1, 2, 3, 4,
1038 // 5, 6, 7 and 8 on screen content set with coding performance change less
1039 // than 0.01% for speed <= 2 and less than 0.03% for speed >= 3. For AVIF
1040 // image encode, this sf reduces instruction count by 1.94%, 1.13%, 1.29%,
1041 // 0.93%, 0.89%, 1.03%, 1.07%, 1.20% and 0.18% for speed 0, 1, 2, 3, 4, 5, 6,
1042 // 7 and 8 on a typical image dataset with coding performance change less than
1043 // 0.01%.
1044 int prune_luma_palette_size_search_level;
1045
1046 // Prune chroma intra modes based on luma intra mode winner.
1047 // 0: No pruning
1048 // 1: Prune chroma intra modes other than UV_DC_PRED, UV_SMOOTH_PRED,
1049 // UV_CFL_PRED and the mode that corresponds to luma intra mode winner.
1050 int prune_chroma_modes_using_luma_winner;
1051
1052 // Clip the frequency of updating the mv cost for intrabc.
1053 INTERNAL_COST_UPDATE_TYPE dv_cost_upd_level;
1054
1055 // We use DCT_DCT transform followed by computing SATD (Sum of Absolute
1056 // Transformed Differences) as an estimation of RD score to quickly find the
1057 // best possible Chroma from Luma (CFL) parameter. Then we do a full RD search
1058 // near the best possible parameter. The search range is set here.
1059 // The range of cfl_searh_range should be [1, 33], and the following are the
1060 // recommended values.
1061 // 1: Fastest mode.
1062 // 3: Default mode that provides good speedup without losing compression
1063 // performance at speed 0.
1064 // 33: Exhaustive rd search (33 == CFL_MAGS_SIZE). This mode should only
1065 // be used for debugging purpose.
1066 int cfl_search_range;
1067
1068 // TOP_INTRA_MODEL_COUNT is 4 that is the number of top model rd to store in
1069 // intra mode decision. Here, add a speed feature to reduce this number for
1070 // higher speeds.
1071 int top_intra_model_count_allowed;
1072
1073 // Adapt top_intra_model_count_allowed locally to prune luma intra modes using
1074 // neighbor block and quantizer information.
1075 int adapt_top_model_rd_count_using_neighbors;
1076
1077 // Terminate early in chroma palette_size search.
1078 // 0: No early termination
1079 // 1: Terminate early for higher palette_size, if header rd cost of lower
1080 // palette_size is more than best_rd.
1081 // For allintra encode, this sf reduces instruction count by 0.45%,
1082 // 0.62%, 1.73%, 2.50%, 2.89%, 3.09% and 3.86% for speed 0 to 6 on screen
1083 // content set with coding performance change less than 0.01%.
1084 // For AVIF image encode, this sf reduces instruction count by 0.45%, 0.81%,
1085 // 0.85%, 1.05%, 1.45%, 1.66% and 1.95% for speed 0 to 6 on a typical image
1086 // dataset with no quality drop.
1087 int early_term_chroma_palette_size_search;
1088
1089 // Skips the evaluation of filter intra modes in inter frames if rd evaluation
1090 // of luma intra dc mode results in invalid rd stats.
1091 int skip_filter_intra_in_inter_frames;
1092} INTRA_MODE_SPEED_FEATURES;
1093
1094typedef struct TX_SPEED_FEATURES {
1095 // Init search depth for square and rectangular transform partitions.
1096 // Values:
1097 // 0 - search full tree, 1: search 1 level, 2: search the highest level only
1098 int inter_tx_size_search_init_depth_sqr;
1099 int inter_tx_size_search_init_depth_rect;
1100 int intra_tx_size_search_init_depth_sqr;
1101 int intra_tx_size_search_init_depth_rect;
1102
1103 // If any dimension of a coding block size above 64, always search the
1104 // largest transform only, since the largest transform block size is 64x64.
1105 int tx_size_search_lgr_block;
1106
1107 TX_TYPE_SEARCH tx_type_search;
1108
1109 // Skip split transform block partition when the collocated bigger block
1110 // is selected as all zero coefficients.
1111 int txb_split_cap;
1112
1113 // Shortcut the transform block partition and type search when the target
1114 // rdcost is relatively lower.
1115 // Values are 0 (not used) , or 1 - 2 with progressively increasing
1116 // aggressiveness
1117 int adaptive_txb_search_level;
1118
1119 // Prune level for tx_size_type search for inter based on rd model
1120 // 0: no pruning
1121 // 1-2: progressively increasing aggressiveness of pruning
1122 int model_based_prune_tx_search_level;
1123
1124 // Refine TX type after fast TX search.
1125 int refine_fast_tx_search_results;
1126
1127 // Prune transform split/no_split eval based on residual properties. A value
1128 // of 0 indicates no pruning, and the aggressiveness of pruning progressively
1129 // increases from levels 1 to 3.
1130 int prune_tx_size_level;
1131} TX_SPEED_FEATURES;
1132
1133typedef struct RD_CALC_SPEED_FEATURES {
1134 // Fast approximation of av1_model_rd_from_var_lapndz
1135 int simple_model_rd_from_var;
1136
1137 // Whether to compute distortion in the image domain (slower but
1138 // more accurate), or in the transform domain (faster but less acurate).
1139 // 0: use image domain
1140 // 1: use transform domain in tx_type search, and use image domain for
1141 // RD_STATS
1142 // 2: use transform domain
1143 int tx_domain_dist_level;
1144
1145 // Transform domain distortion threshold level
1146 int tx_domain_dist_thres_level;
1147
1148 // Trellis (dynamic programming) optimization of quantized values
1149 TRELLIS_OPT_TYPE optimize_coefficients;
1150
1151 // Use hash table to store macroblock RD search results
1152 // to avoid repeated search on the same residue signal.
1153 int use_mb_rd_hash;
1154
1155 // Flag used to control the extent of coeff R-D optimization
1156 int perform_coeff_opt;
1157} RD_CALC_SPEED_FEATURES;
1158
1159typedef struct WINNER_MODE_SPEED_FEATURES {
1160 // Flag used to control the winner mode processing for better R-D optimization
1161 // of quantized coeffs
1162 int enable_winner_mode_for_coeff_opt;
1163
1164 // Flag used to control the winner mode processing for transform size
1165 // search method
1166 int enable_winner_mode_for_tx_size_srch;
1167
1168 // Control transform size search level
1169 // Eval type: Default Mode Winner
1170 // Level 0 : FULL RD LARGEST ALL FULL RD
1171 // Level 1 : FAST RD LARGEST ALL FULL RD
1172 // Level 2 : LARGEST ALL LARGEST ALL FULL RD
1173 // Level 3 : LARGEST ALL LARGEST ALL LARGEST ALL
1174 int tx_size_search_level;
1175
1176 // Flag used to control the winner mode processing for use transform
1177 // domain distortion
1178 int enable_winner_mode_for_use_tx_domain_dist;
1179
1180 // Flag used to enable processing of multiple winner modes
1181 MULTI_WINNER_MODE_TYPE multi_winner_mode_type;
1182
1183 // Motion mode for winner candidates:
1184 // 0: speed feature OFF
1185 // 1 / 2 : Use configured number of winner candidates
1186 int motion_mode_for_winner_cand;
1187
1188 // Early DC only txfm block prediction
1189 // 0: speed feature OFF
1190 // 1 / 2 : Use the configured level for different modes
1191 int dc_blk_pred_level;
1192
1193 // If on, disables interpolation filter search in handle_inter_mode loop, and
1194 // performs it during winner mode processing by \ref
1195 // tx_search_best_inter_candidates.
1196 int winner_mode_ifs;
1197} WINNER_MODE_SPEED_FEATURES;
1198
1199typedef struct LOOP_FILTER_SPEED_FEATURES {
1200 // This feature controls how the loop filter level is determined.
1201 LPF_PICK_METHOD lpf_pick;
1202
1203 // Skip some final iterations in the determination of the best loop filter
1204 // level.
1205 int use_coarse_filter_level_search;
1206
1207 // Control how the CDEF strength is determined.
1208 CDEF_PICK_METHOD cdef_pick_method;
1209
1210 // Decoder side speed feature to add penalty for use of dual-sgr filters.
1211 // Takes values 0 - 10, 0 indicating no penalty and each additional level
1212 // adding a penalty of 1%
1213 int dual_sgr_penalty_level;
1214
1215 // prune sgr ep using binary search like mechanism
1216 int enable_sgr_ep_pruning;
1217
1218 // Disable loop restoration for Chroma plane
1219 int disable_loop_restoration_chroma;
1220
1221 // Disable loop restoration for luma plane
1222 int disable_loop_restoration_luma;
1223
1224 // Prune RESTORE_WIENER evaluation based on source variance
1225 // 0 : no pruning
1226 // 1 : conservative pruning
1227 // 2 : aggressive pruning
1228 int prune_wiener_based_on_src_var;
1229
1230 // Prune self-guided loop restoration based on wiener search results
1231 // 0 : no pruning
1232 // 1 : pruning based on rdcost ratio of RESTORE_WIENER and RESTORE_NONE
1233 // 2 : pruning based on winner restoration type among RESTORE_WIENER and
1234 // RESTORE_NONE
1235 int prune_sgr_based_on_wiener;
1236
1237 // Reduce the wiener filter win size for luma
1238 int reduce_wiener_window_size;
1239
1240 // Disable loop restoration filter
1241 int disable_lr_filter;
1242} LOOP_FILTER_SPEED_FEATURES;
1243
1244typedef struct REAL_TIME_SPEED_FEATURES {
1245 // check intra prediction for non-RD mode.
1246 int check_intra_pred_nonrd;
1247
1248 // skip checking intra prediction if TX is skipped
1249 int skip_intra_pred_if_tx_skip;
1250
1251 // Perform coarse ME before calculating variance in variance-based partition
1252 int estimate_motion_for_var_based_partition;
1253
1254 // For nonrd_use_partition: mode of extra check of leaf partition
1255 // 0 - don't check merge
1256 // 1 - always check merge
1257 // 2 - check merge and prune checking final split
1258 int nonrd_check_partition_merge_mode;
1259
1260 // For nonrd_use_partition: check of leaf partition extra split
1261 int nonrd_check_partition_split;
1262
1263 // Implements various heuristics to skip searching modes
1264 // The heuristics selected are based on flags
1265 // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
1266 unsigned int mode_search_skip_flags;
1267
1268 // For nonrd: Reduces ref frame search.
1269 // 0 - low level of search prune in non last frames
1270 // 1 - pruned search in non last frames
1271 // 2 - more pruned search in non last frames
1272 int nonrd_prune_ref_frame_search;
1273
1274 // This flag controls the use of non-RD mode decision.
1275 int use_nonrd_pick_mode;
1276
1277 // Use ALTREF frame in non-RD mode decision.
1278 int use_nonrd_altref_frame;
1279
1280 // Use compound reference for non-RD mode.
1281 int use_comp_ref_nonrd;
1282
1283 // Reference frames for compound prediction for nonrd pickmode:
1284 // LAST_GOLDEN (0), LAST_LAST2 (1), or LAST_ALTREF (2).
1285 int ref_frame_comp_nonrd[3];
1286
1287 // use reduced ref set for real-time mode
1288 int use_real_time_ref_set;
1289
1290 // Skip a number of expensive mode evaluations for blocks with very low
1291 // temporal variance.
1292 int short_circuit_low_temp_var;
1293
1294 // Use modeled (currently CurvFit model) RDCost for fast non-RD mode
1295 int use_modeled_non_rd_cost;
1296
1297 // Reuse inter prediction in fast non-rd mode.
1298 int reuse_inter_pred_nonrd;
1299
1300 // Number of best inter modes to search transform. INT_MAX - search all.
1301 int num_inter_modes_for_tx_search;
1302
1303 // Forces TX search off for RDCost calulation.
1304 int force_tx_search_off;
1305
1306 // Use interpolation filter search in non-RD mode decision.
1307 int use_nonrd_filter_search;
1308
1309 // Use simplified RD model for interpolation search and Intra
1310 int use_simple_rd_model;
1311
1312 // If set forces interpolation filter to EIGHTTAP_REGULAR
1313 int skip_interp_filter_search;
1314
1315 // For nonrd mode: use hybrid (rd for bsize < 16x16, otherwise nonrd)
1316 // intra mode search for intra only frames. If set to 0 then nonrd pick
1317 // intra is used for all blocks.
1318 int hybrid_intra_pickmode;
1319
1320 // Compute variance/sse on source difference, prior to encoding superblock.
1321 int source_metrics_sb_nonrd;
1322
1323 // Flag to indicate process for handling overshoot on slide/scene change,
1324 // for real-time CBR mode.
1325 OVERSHOOT_DETECTION_CBR overshoot_detection_cbr;
1326
1327 // Check for scene/content change detection on every frame before encoding.
1328 int check_scene_detection;
1329
1330 // Forces larger partition blocks in variance based partitioning
1331 int force_large_partition_blocks;
1332
1333 // uses results of temporal noise estimate
1334 int use_temporal_noise_estimate;
1335
1336 // Parameter indicating initial search window to be used in full-pixel search
1337 // for nonrd_pickmode. Range [0, MAX_MVSEARCH_STEPS - 1]. Lower value
1338 // indicates larger window. If set to 0, step_param is set based on internal
1339 // logic in set_mv_search_params().
1340 int fullpel_search_step_param;
1341
1342 // Bit mask to enable or disable intra modes for each prediction block size
1343 // separately, for nonrd pickmode.
1344 int intra_y_mode_bsize_mask_nrd[BLOCK_SIZES];
1345
1346 // Skips mode checks more agressively in nonRD mode
1347 int nonrd_agressive_skip;
1348
1349 // Skip cdef on 64x64 blocks when NEWMV or INTRA is not picked or color
1350 // sensitivity is off. When color sensitivity is on for a superblock, all
1351 // 64x64 blocks within will not skip.
1352 int skip_cdef_sb;
1353
1354 // Forces larger partition blocks in variance based partitioning for intra
1355 // frames
1356 int force_large_partition_blocks_intra;
1357
1358 // Skip evaluation of no split in tx size selection for merge partition
1359 int skip_tx_no_split_var_based_partition;
1360
1361 // Intermediate termination of newMV mode evaluation based on so far best mode
1362 // sse
1363 int skip_newmv_mode_based_on_sse;
1364
1365 // Define gf length multiplier.
1366 // Level 0: use large multiplier, level 1: use medium multiplier.
1367 int gf_length_lvl;
1368
1369 // Prune inter modes with golden frame as reference for NEARMV and NEWMV modes
1370 int prune_inter_modes_with_golden_ref;
1371
1372 // Prune inter modes w.r.t golden or alt-ref frame based on sad
1373 int prune_inter_modes_wrt_gf_arf_based_on_sad;
1374
1375 // Prune inter mode search in rd path based on current block's temporal
1376 // variance wrt LAST reference.
1377 int prune_inter_modes_using_temp_var;
1378
1379 // Force half_pel at block level.
1380 int force_half_pel_block;
1381} REAL_TIME_SPEED_FEATURES;
1382
1388typedef struct SPEED_FEATURES {
1393
1398
1402 TPL_SPEED_FEATURES tpl_sf;
1403
1407 GLOBAL_MOTION_SPEED_FEATURES gm_sf;
1408
1412 PARTITION_SPEED_FEATURES part_sf;
1413
1417 MV_SPEED_FEATURES mv_sf;
1418
1422 INTER_MODE_SPEED_FEATURES inter_sf;
1423
1427 INTERP_FILTER_SPEED_FEATURES interp_sf;
1428
1432 INTRA_MODE_SPEED_FEATURES intra_sf;
1433
1437 TX_SPEED_FEATURES tx_sf;
1438
1442 RD_CALC_SPEED_FEATURES rd_sf;
1443
1447 WINNER_MODE_SPEED_FEATURES winner_mode_sf;
1448
1452 LOOP_FILTER_SPEED_FEATURES lpf_sf;
1453
1457 REAL_TIME_SPEED_FEATURES rt_sf;
1461struct AV1_COMP;
1462
1476 int speed);
1477
1490 int speed);
1502void av1_set_speed_features_qindex_dependent(struct AV1_COMP *cpi, int speed);
1503
1504#ifdef __cplusplus
1505} // extern "C"
1506#endif
1507
1508#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:2475
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:315
@ INTERNAL_COST_UPD_OFF
Definition: speed_features.h:316
@ INTERNAL_COST_UPD_SBROW_SET
Definition: speed_features.h:317
@ INTERNAL_COST_UPD_SBROW
Definition: speed_features.h:318
@ INTERNAL_COST_UPD_SB
Definition: speed_features.h:319
SIMPLE_MOTION_SEARCH_PRUNE_LEVEL
This enumeration defines a variety of simple motion search based partition prune levels.
Definition: speed_features.h:326
@ QIDX_BASED_AGG_LVL1
Definition: speed_features.h:332
@ SIMPLE_AGG_LVL3
Definition: speed_features.h:331
@ TOTAL_SIMPLE_AGG_LVLS
Definition: speed_features.h:335
@ SIMPLE_AGG_LVL1
Definition: speed_features.h:329
@ SIMPLE_AGG_LVL0
Definition: speed_features.h:328
@ TOTAL_AGG_LVLS
Definition: speed_features.h:341
@ SIMPLE_AGG_LVL2
Definition: speed_features.h:330
@ TOTAL_QINDEX_BASED_AGG_LVLS
Definition: speed_features.h:337
struct SPEED_FEATURES SPEED_FEATURES
Top level speed vs quality trade off data struture.
struct HIGH_LEVEL_SPEED_FEATURES HIGH_LEVEL_SPEED_FEATURES
Sequence/frame level speed vs quality features.
struct FIRST_PASS_SPEED_FEATURES FIRST_PASS_SPEED_FEATURES
CDEF_PICK_METHOD
This enumeration defines a variety of CDEF pick methods.
Definition: speed_features.h: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:399
int disable_recon
Skips reconstruction by using source buffers for prediction.
Definition: speed_features.h:416
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:406
int skip_zeromv_motion_search
Skips the motion search centered on 0,0 mv.
Definition: speed_features.h:421
int skip_motion_search_threshold
Skips the motion search when the zero mv has small sse.
Definition: speed_features.h:411
Sequence/frame level speed vs quality features.
Definition: speed_features.h:348
int second_alt_ref_filtering
Definition: speed_features.h:393
int frame_parameter_update
Definition: speed_features.h:350
MV_PREC_LOGIC high_precision_mv_usage
Definition: speed_features.h:369
int disable_extra_sc_testing
Definition: speed_features.h:388
int recode_tolerance
Definition: speed_features.h:361
SUPERRES_AUTO_SEARCH_TYPE superres_auto_search_type
Definition: speed_features.h:383
int static_segmentation
Definition: speed_features.h:378
RECODE_LOOP_TYPE recode_loop
Definition: speed_features.h:355
Top level speed vs quality trade off data struture.
Definition: speed_features.h:1388
MV_SPEED_FEATURES mv_sf
Definition: speed_features.h:1417
TPL_SPEED_FEATURES tpl_sf
Definition: speed_features.h:1402
LOOP_FILTER_SPEED_FEATURES lpf_sf
Definition: speed_features.h:1452
TX_SPEED_FEATURES tx_sf
Definition: speed_features.h:1437
INTER_MODE_SPEED_FEATURES inter_sf
Definition: speed_features.h:1422
RD_CALC_SPEED_FEATURES rd_sf
Definition: speed_features.h:1442
PARTITION_SPEED_FEATURES part_sf
Definition: speed_features.h:1412
GLOBAL_MOTION_SPEED_FEATURES gm_sf
Definition: speed_features.h:1407
INTERP_FILTER_SPEED_FEATURES interp_sf
Definition: speed_features.h:1427
FIRST_PASS_SPEED_FEATURES fp_sf
Definition: speed_features.h:1397
INTRA_MODE_SPEED_FEATURES intra_sf
Definition: speed_features.h:1432
WINNER_MODE_SPEED_FEATURES winner_mode_sf
Definition: speed_features.h:1447
REAL_TIME_SPEED_FEATURES rt_sf
Definition: speed_features.h:1457
HIGH_LEVEL_SPEED_FEATURES hl_sf
Definition: speed_features.h:1392