xref: /aosp_15_r20/external/libaom/av1/encoder/speed_features.h (revision 77c1e3ccc04c968bd2bc212e87364f250e820521)
1 /*
2  * Copyright (c) 2016, Alliance for Open Media. All rights reserved.
3  *
4  * This source code is subject to the terms of the BSD 2 Clause License and
5  * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6  * was not distributed with this source code in the LICENSE file, you can
7  * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8  * Media Patent License 1.0 was not distributed with this source code in the
9  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10  */
11 
12 #ifndef AOM_AV1_ENCODER_SPEED_FEATURES_H_
13 #define AOM_AV1_ENCODER_SPEED_FEATURES_H_
14 
15 #include "av1/common/enums.h"
16 #include "av1/encoder/enc_enums.h"
17 #include "av1/encoder/mcomp.h"
18 #include "av1/encoder/encodemb.h"
19 
20 #ifdef __cplusplus
21 extern "C" {
22 #endif
23 
24 /*! @file */
25 
26 /*!\cond */
27 #define MAX_MESH_STEP 4
28 
29 typedef struct MESH_PATTERN {
30   int range;
31   int interval;
32 } MESH_PATTERN;
33 
34 enum {
35   GM_FULL_SEARCH,
36   GM_REDUCED_REF_SEARCH_SKIP_L2_L3,
37   GM_REDUCED_REF_SEARCH_SKIP_L2_L3_ARF2,
38 
39   // Same as GM_REDUCED_REF_SEARCH_SKIP_L2_L3_ARF2 but with extra filtering
40   // to keep at most two ref frames
41   GM_SEARCH_CLOSEST_REFS_ONLY,
42 
43   GM_DISABLE_SEARCH
44 } UENUM1BYTE(GM_SEARCH_TYPE);
45 
46 enum {
47   DIST_WTD_COMP_ENABLED,
48   DIST_WTD_COMP_SKIP_MV_SEARCH,
49   DIST_WTD_COMP_DISABLED,
50 } UENUM1BYTE(DIST_WTD_COMP_FLAG);
51 
52 enum {
53   INTRA_ALL = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED) | (1 << D45_PRED) |
54               (1 << D135_PRED) | (1 << D113_PRED) | (1 << D157_PRED) |
55               (1 << D203_PRED) | (1 << D67_PRED) | (1 << SMOOTH_PRED) |
56               (1 << SMOOTH_V_PRED) | (1 << SMOOTH_H_PRED) | (1 << PAETH_PRED),
57   UV_INTRA_ALL =
58       (1 << UV_DC_PRED) | (1 << UV_V_PRED) | (1 << UV_H_PRED) |
59       (1 << UV_D45_PRED) | (1 << UV_D135_PRED) | (1 << UV_D113_PRED) |
60       (1 << UV_D157_PRED) | (1 << UV_D203_PRED) | (1 << UV_D67_PRED) |
61       (1 << UV_SMOOTH_PRED) | (1 << UV_SMOOTH_V_PRED) |
62       (1 << UV_SMOOTH_H_PRED) | (1 << UV_PAETH_PRED) | (1 << UV_CFL_PRED),
63   UV_INTRA_DC = (1 << UV_DC_PRED),
64   UV_INTRA_DC_CFL = (1 << UV_DC_PRED) | (1 << UV_CFL_PRED),
65   UV_INTRA_DC_TM = (1 << UV_DC_PRED) | (1 << UV_PAETH_PRED),
66   UV_INTRA_DC_PAETH_CFL =
67       (1 << UV_DC_PRED) | (1 << UV_PAETH_PRED) | (1 << UV_CFL_PRED),
68   UV_INTRA_DC_H_V = (1 << UV_DC_PRED) | (1 << UV_V_PRED) | (1 << UV_H_PRED),
69   UV_INTRA_DC_H_V_CFL = (1 << UV_DC_PRED) | (1 << UV_V_PRED) |
70                         (1 << UV_H_PRED) | (1 << UV_CFL_PRED),
71   UV_INTRA_DC_PAETH_H_V = (1 << UV_DC_PRED) | (1 << UV_PAETH_PRED) |
72                           (1 << UV_V_PRED) | (1 << UV_H_PRED),
73   UV_INTRA_DC_PAETH_H_V_CFL = (1 << UV_DC_PRED) | (1 << UV_PAETH_PRED) |
74                               (1 << UV_V_PRED) | (1 << UV_H_PRED) |
75                               (1 << UV_CFL_PRED),
76   INTRA_DC = (1 << DC_PRED),
77   INTRA_DC_TM = (1 << DC_PRED) | (1 << PAETH_PRED),
78   INTRA_DC_H_V = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED),
79   INTRA_DC_H_V_SMOOTH =
80       (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED) | (1 << SMOOTH_PRED),
81   INTRA_DC_PAETH_H_V =
82       (1 << DC_PRED) | (1 << PAETH_PRED) | (1 << V_PRED) | (1 << H_PRED)
83 };
84 
85 enum {
86   INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << GLOBALMV) |
87               (1 << NEWMV) | (1 << NEAREST_NEARESTMV) | (1 << NEAR_NEARMV) |
88               (1 << NEW_NEWMV) | (1 << NEAREST_NEWMV) | (1 << NEAR_NEWMV) |
89               (1 << NEW_NEARMV) | (1 << NEW_NEARESTMV) | (1 << GLOBAL_GLOBALMV),
90   INTER_NEAREST_NEAR_ZERO = (1 << NEARESTMV) | (1 << NEARMV) | (1 << GLOBALMV) |
91                             (1 << NEAREST_NEARESTMV) | (1 << GLOBAL_GLOBALMV) |
92                             (1 << NEAREST_NEWMV) | (1 << NEW_NEARESTMV) |
93                             (1 << NEW_NEARMV) | (1 << NEAR_NEWMV) |
94                             (1 << NEAR_NEARMV),
95   INTER_SINGLE_ALL =
96       (1 << NEARESTMV) | (1 << NEARMV) | (1 << GLOBALMV) | (1 << NEWMV),
97 };
98 
99 enum {
100   DISABLE_ALL_INTER_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
101                             (1 << THR_ALTR) | (1 << THR_GOLD) | (1 << THR_LAST),
102 
103   DISABLE_ALL_SPLIT = (1 << THR_INTRA) | DISABLE_ALL_INTER_SPLIT,
104 
105   DISABLE_COMPOUND_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA),
106 
107   LAST_AND_INTRA_SPLIT_ONLY = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
108                               (1 << THR_ALTR) | (1 << THR_GOLD)
109 };
110 
111 enum {
112   TXFM_CODING_SF = 1,
113   INTER_PRED_SF = 2,
114   INTRA_PRED_SF = 4,
115   PARTITION_SF = 8,
116   LOOP_FILTER_SF = 16,
117   RD_SKIP_SF = 32,
118   RESERVE_2_SF = 64,
119   RESERVE_3_SF = 128,
120 } UENUM1BYTE(DEV_SPEED_FEATURES);
121 
122 /* This enumeration defines when the rate control recode loop will be
123  * enabled.
124  */
125 enum {
126   /*
127    * No recodes allowed
128    */
129   DISALLOW_RECODE = 0,
130   /*
131    * Allow recode only for KF/ARF/GF frames
132    */
133   ALLOW_RECODE_KFARFGF = 1,
134   /*
135    * Allow recode for all frame types based on bitrate constraints.
136    */
137   ALLOW_RECODE = 2,
138 } UENUM1BYTE(RECODE_LOOP_TYPE);
139 
140 enum {
141   SUBPEL_TREE = 0,
142   SUBPEL_TREE_PRUNED = 1,       // Prunes 1/2-pel searches
143   SUBPEL_TREE_PRUNED_MORE = 2,  // Prunes 1/2-pel searches more aggressively
144   SUBPEL_SEARCH_METHODS
145 } UENUM1BYTE(SUBPEL_SEARCH_METHOD);
146 
147 enum {
148   // Try the full image with different values.
149   LPF_PICK_FROM_FULL_IMAGE,
150   // Try the full image filter search with non-dual filter only.
151   LPF_PICK_FROM_FULL_IMAGE_NON_DUAL,
152   // Try a small portion of the image with different values.
153   LPF_PICK_FROM_SUBIMAGE,
154   // Estimate the level based on quantizer and frame type
155   LPF_PICK_FROM_Q,
156   // Pick 0 to disable LPF if LPF was enabled last frame
157   LPF_PICK_MINIMAL_LPF
158 } UENUM1BYTE(LPF_PICK_METHOD);
159 /*!\endcond */
160 
161 /*!\enum CDEF_PICK_METHOD
162  * \brief This enumeration defines a variety of CDEF pick methods
163  */
164 typedef enum {
165   CDEF_FULL_SEARCH,      /**< Full search */
166   CDEF_FAST_SEARCH_LVL1, /**< Search among a subset of all possible filters. */
167   CDEF_FAST_SEARCH_LVL2, /**< Search reduced subset of filters than Level 1. */
168   CDEF_FAST_SEARCH_LVL3, /**< Search reduced subset of secondary filters than
169                               Level 2. */
170   CDEF_FAST_SEARCH_LVL4, /**< Search reduced subset of filters than Level 3. */
171   CDEF_FAST_SEARCH_LVL5, /**< Search reduced subset of filters than Level 4. */
172   CDEF_PICK_FROM_Q,      /**< Estimate filter strength based on quantizer. */
173   CDEF_PICK_METHODS
174 } CDEF_PICK_METHOD;
175 
176 /*!\cond */
177 enum {
178   // Terminate search early based on distortion so far compared to
179   // qp step, distortion in the neighborhood of the frame, etc.
180   FLAG_EARLY_TERMINATE = 1 << 0,
181 
182   // Skips comp inter modes if the best so far is an intra mode.
183   FLAG_SKIP_COMP_BESTINTRA = 1 << 1,
184 
185   // Skips oblique intra modes if the best so far is an inter mode.
186   FLAG_SKIP_INTRA_BESTINTER = 1 << 3,
187 
188   // Skips oblique intra modes  at angles 27, 63, 117, 153 if the best
189   // intra so far is not one of the neighboring directions.
190   FLAG_SKIP_INTRA_DIRMISMATCH = 1 << 4,
191 
192   // Skips intra modes other than DC_PRED if the source variance is small
193   FLAG_SKIP_INTRA_LOWVAR = 1 << 5,
194 } UENUM1BYTE(MODE_SEARCH_SKIP_LOGIC);
195 
196 enum {
197   // No tx type pruning
198   TX_TYPE_PRUNE_0 = 0,
199   // adaptively prunes the least perspective tx types out of all 16
200   // (tuned to provide negligible quality loss)
201   TX_TYPE_PRUNE_1 = 1,
202   // similar, but applies much more aggressive pruning to get better speed-up
203   TX_TYPE_PRUNE_2 = 2,
204   TX_TYPE_PRUNE_3 = 3,
205   // More aggressive pruning based on tx type score and allowed tx count
206   TX_TYPE_PRUNE_4 = 4,
207   TX_TYPE_PRUNE_5 = 5,
208 } UENUM1BYTE(TX_TYPE_PRUNE_MODE);
209 
210 enum {
211   // No reaction to rate control on a detected slide/scene change.
212   NO_DETECTION = 0,
213 
214   // Set to larger Q based only on the detected slide/scene change and
215   // current/past Q.
216   FAST_DETECTION_MAXQ = 1,
217 } UENUM1BYTE(OVERSHOOT_DETECTION_CBR);
218 
219 enum {
220   // Turns off multi-winner mode. So we will do txfm search on either all modes
221   // if winner mode is off, or we will only on txfm search on a single winner
222   // mode.
223   MULTI_WINNER_MODE_OFF = 0,
224 
225   // Limits the number of winner modes to at most 2
226   MULTI_WINNER_MODE_FAST = 1,
227 
228   // Uses the default number of winner modes, which is 3 for intra mode, and 1
229   // for inter mode.
230   MULTI_WINNER_MODE_DEFAULT = 2,
231 
232   // Maximum number of winner modes allowed.
233   MULTI_WINNER_MODE_LEVELS,
234 } UENUM1BYTE(MULTI_WINNER_MODE_TYPE);
235 
236 enum {
237   PRUNE_NEARMV_OFF = 0,     // Turn off nearmv pruning
238   PRUNE_NEARMV_LEVEL1 = 1,  // Prune nearmv for qindex (0-85)
239   PRUNE_NEARMV_LEVEL2 = 2,  // Prune nearmv for qindex (0-170)
240   PRUNE_NEARMV_LEVEL3 = 3,  // Prune nearmv more aggressively for qindex (0-170)
241   PRUNE_NEARMV_MAX = PRUNE_NEARMV_LEVEL3,
242 } UENUM1BYTE(PRUNE_NEARMV_LEVEL);
243 
244 enum {
245   // Default transform search used in evaluation of best inter candidates
246   // (MODE_EVAL stage) and motion mode winner processing (WINNER_MODE_EVAL
247   // stage).
248   TX_SEARCH_DEFAULT = 0,
249   // Transform search in motion mode rd during MODE_EVAL stage.
250   TX_SEARCH_MOTION_MODE,
251   // Transform search in compound type mode rd during MODE_EVAL stage.
252   TX_SEARCH_COMP_TYPE_MODE,
253   // All transform search cases
254   TX_SEARCH_CASES
255 } UENUM1BYTE(TX_SEARCH_CASE);
256 
257 typedef struct {
258   TX_TYPE_PRUNE_MODE prune_2d_txfm_mode;
259   int fast_intra_tx_type_search;
260 
261   // INT_MAX: Disable fast search.
262   // 1 - 1024: Probability threshold used for conditionally forcing tx type,
263   // during mode search.
264   // 0: Force tx type to be DCT_DCT unconditionally, during
265   // mode search.
266   int fast_inter_tx_type_prob_thresh;
267 
268   // Prune less likely chosen transforms for each intra mode. The speed
269   // feature ranges from 0 to 2, for different speed / compression trade offs.
270   int use_reduced_intra_txset;
271 
272   // Use a skip flag prediction model to detect blocks with skip = 1 early
273   // and avoid doing full TX type search for such blocks.
274   int use_skip_flag_prediction;
275 
276   // Threshold used by the ML based method to predict TX block split decisions.
277   int ml_tx_split_thresh;
278 
279   // skip remaining transform type search when we found the rdcost of skip is
280   // better than applying transform
281   int skip_tx_search;
282 
283   // Prune tx type search using previous frame stats.
284   int prune_tx_type_using_stats;
285   // Prune tx type search using estimated RDcost
286   int prune_tx_type_est_rd;
287 
288   // Flag used to control the winner mode processing for tx type pruning for
289   // inter blocks. It enables further tx type mode pruning based on ML model for
290   // mode evaluation and disables tx type mode pruning for winner mode
291   // processing.
292   int winner_mode_tx_type_pruning;
293 } TX_TYPE_SEARCH;
294 
295 enum {
296   // Search partitions using RD criterion
297   SEARCH_PARTITION,
298 
299   // Always use a fixed size partition
300   FIXED_PARTITION,
301 
302   // Partition using source variance
303   VAR_BASED_PARTITION,
304 
305 #if CONFIG_RT_ML_PARTITIONING
306   // Partition using ML model
307   ML_BASED_PARTITION
308 #endif
309 } UENUM1BYTE(PARTITION_SEARCH_TYPE);
310 
311 enum {
312   NOT_IN_USE,
313   DIRECT_PRED,
314   RELAXED_PRED,
315   ADAPT_PRED
316 } UENUM1BYTE(MAX_PART_PRED_MODE);
317 
318 enum {
319   LAST_MV_DATA,
320   CURRENT_Q,
321   QTR_ONLY,
322 } UENUM1BYTE(MV_PREC_LOGIC);
323 
324 enum {
325   SUPERRES_AUTO_ALL,   // Tries all possible superres ratios
326   SUPERRES_AUTO_DUAL,  // Tries no superres and q-based superres ratios
327   SUPERRES_AUTO_SOLO,  // Only apply the q-based superres ratio
328 } UENUM1BYTE(SUPERRES_AUTO_SEARCH_TYPE);
329 /*!\endcond */
330 
331 /*!\enum INTERNAL_COST_UPDATE_TYPE
332  * \brief This enum decides internally how often to update the entropy costs
333  *
334  * INTERNAL_COST_UPD_TYPE is similar to \ref COST_UPDATE_TYPE but has slightly
335  * more flexibility in update frequency. This enum is separate from \ref
336  * COST_UPDATE_TYPE because although \ref COST_UPDATE_TYPE is not exposed, its
337  * values are public so it cannot be modified without breaking public API.
338  * Due to the use of AOMMIN() in populate_unified_cost_update_freq() to
339  * compute the unified cost update frequencies (out of COST_UPDATE_TYPE and
340  * INTERNAL_COST_UPDATE_TYPE), the values of this enum type must be listed in
341  * the order of increasing frequencies.
342  *
343  * \warning  In case of any updates/modifications to the enum COST_UPDATE_TYPE,
344  * update the enum INTERNAL_COST_UPDATE_TYPE as well.
345  */
346 typedef enum {
347   INTERNAL_COST_UPD_OFF,       /*!< Turn off cost updates. */
348   INTERNAL_COST_UPD_TILE,      /*!< Update every tile. */
349   INTERNAL_COST_UPD_SBROW_SET, /*!< Update every row_set of height 256 pixs. */
350   INTERNAL_COST_UPD_SBROW,     /*!< Update every sb rows inside a tile. */
351   INTERNAL_COST_UPD_SB,        /*!< Update every sb. */
352 } INTERNAL_COST_UPDATE_TYPE;
353 
354 /*!\enum SIMPLE_MOTION_SEARCH_PRUNE_LEVEL
355  * \brief This enumeration defines a variety of simple motion search based
356  * partition prune levels
357  */
358 typedef enum {
359   NO_PRUNING = -1,
360   SIMPLE_AGG_LVL0,     /*!< Simple prune aggressiveness level 0. */
361   SIMPLE_AGG_LVL1,     /*!< Simple prune aggressiveness level 1. */
362   SIMPLE_AGG_LVL2,     /*!< Simple prune aggressiveness level 2. */
363   SIMPLE_AGG_LVL3,     /*!< Simple prune aggressiveness level 3. */
364   QIDX_BASED_AGG_LVL1, /*!< Qindex based prune aggressiveness level, aggressive
365                           level maps to simple agg level 1 or 2 based on qindex.
366                         */
367   TOTAL_SIMPLE_AGG_LVLS = QIDX_BASED_AGG_LVL1, /*!< Total number of simple prune
368                                                   aggressiveness levels. */
369   TOTAL_QINDEX_BASED_AGG_LVLS =
370       QIDX_BASED_AGG_LVL1 -
371       SIMPLE_AGG_LVL3, /*!< Total number of qindex based simple prune
372                           aggressiveness levels. */
373   TOTAL_AGG_LVLS = TOTAL_SIMPLE_AGG_LVLS +
374                    TOTAL_QINDEX_BASED_AGG_LVLS, /*!< Total number of levels. */
375 } SIMPLE_MOTION_SEARCH_PRUNE_LEVEL;
376 
377 /*!\enum PRUNE_MESH_SEARCH_LEVEL
378  * \brief This enumeration defines a variety of mesh search prune levels.
379  */
380 typedef enum {
381   PRUNE_MESH_SEARCH_DISABLED = 0, /*!< Prune mesh search level 0. */
382   PRUNE_MESH_SEARCH_LVL_1 = 1,    /*!< Prune mesh search level 1. */
383   PRUNE_MESH_SEARCH_LVL_2 = 2,    /*!< Prune mesh search level 2. */
384 } PRUNE_MESH_SEARCH_LEVEL;
385 
386 /*!\enum INTER_SEARCH_EARLY_TERM_IDX
387  * \brief This enumeration defines inter search early termination index in
388  * non-rd path based on sse value.
389  */
390 typedef enum {
391   EARLY_TERM_DISABLED =
392       0, /*!< Early terminate inter mode search based on sse disabled. */
393   EARLY_TERM_IDX_1 =
394       1, /*!< Early terminate inter mode search based on sse, index 1. */
395   EARLY_TERM_IDX_2 =
396       2, /*!< Early terminate inter mode search based on sse, index 2. */
397   EARLY_TERM_IDX_3 =
398       3, /*!< Early terminate inter mode search based on sse, index 3. */
399   EARLY_TERM_IDX_4 =
400       4, /*!< Early terminate inter mode search based on sse, index 4. */
401   EARLY_TERM_INDICES, /*!< Total number of early terminate indices */
402 } INTER_SEARCH_EARLY_TERM_IDX;
403 
404 /*!
405  * \brief Sequence/frame level speed vs quality features
406  */
407 typedef struct HIGH_LEVEL_SPEED_FEATURES {
408   /*! Frame level coding parameter update. */
409   int frame_parameter_update;
410 
411   /*!
412    * Cases and frame types for which the recode loop is enabled.
413    */
414   RECODE_LOOP_TYPE recode_loop;
415 
416   /*!
417    * Controls the tolerance vs target rate used in deciding whether to
418    * recode a frame. It has no meaning if recode is disabled.
419    */
420   int recode_tolerance;
421 
422   /*!
423    * Determine how motion vector precision is chosen. The possibilities are:
424    * LAST_MV_DATA: use the mv data from the last coded frame
425    * CURRENT_Q: use the current q as a threshold
426    * QTR_ONLY: use quarter pel precision only.
427    */
428   MV_PREC_LOGIC high_precision_mv_usage;
429 
430   /*!
431    * Always set to 0. If on it enables 0 cost background transmission
432    * (except for the initial transmission of the segmentation). The feature is
433    * disabled because the addition of very large block sizes make the
434    * backgrounds very to cheap to encode, and the segmentation we have
435    * adds overhead.
436    */
437   int static_segmentation;
438 
439   /*!
440    * Superres-auto mode search type:
441    */
442   SUPERRES_AUTO_SEARCH_TYPE superres_auto_search_type;
443 
444   /*!
445    * Enable/disable extra screen content test by encoding key frame twice.
446    */
447   int disable_extra_sc_testing;
448 
449   /*!
450    * Enable/disable second_alt_ref temporal filtering.
451    */
452   int second_alt_ref_filtering;
453 
454   /*!
455    * The number of frames to be used during temporal filtering of an ARF frame
456    * is adjusted based on noise level of the current frame. The sf has three
457    * levels to decide number of frames to be considered for filtering:
458    * 0       : Use default number of frames
459    * 1 and 2 : Reduce the number of frames based on noise level with varied
460    * aggressiveness
461    */
462   int adjust_num_frames_for_arf_filtering;
463 
464   /*!
465    * Decide the bit estimation approach used in qindex decision.
466    * 0: estimate bits based on a constant value;
467    * 1: estimate bits more accurately based on the frame complexity.
468    */
469   int accurate_bit_estimate;
470 
471   /*!
472    * Decide the approach for weight calculation during temporal filtering.
473    * 0: Calculate weight using exp()
474    * 1: Calculate weight using a lookup table that approximates exp().
475    */
476   int weight_calc_level_in_tf;
477 
478   /*!
479    * Decide whether to perform motion estimation at split block (i.e. 16x16)
480    * level or not.
481    * 0: Always allow motion estimation.
482    * 1: Conditionally allow motion estimation based on 4x4 sub-blocks variance.
483    */
484   int allow_sub_blk_me_in_tf;
485 } HIGH_LEVEL_SPEED_FEATURES;
486 
487 /*!
488  * Speed features for the first pass.
489  */
490 typedef struct FIRST_PASS_SPEED_FEATURES {
491   /*!
492    * \brief Reduces the mv search window.
493    * By default, the initial search window is around
494    * MIN(MIN(dims), MAX_FULL_PEL_VAL) = MIN(MIN(dims), 1023).
495    * Each step reduction decrease the window size by about a factor of 2.
496    */
497   int reduce_mv_step_param;
498 
499   /*!
500    * \brief Skips the motion search when the zero mv has small sse.
501    */
502   int skip_motion_search_threshold;
503 
504   /*!
505    * \brief Skips reconstruction by using source buffers for prediction
506    */
507   int disable_recon;
508 
509   /*!
510    * \brief Skips the motion search centered on 0,0 mv.
511    */
512   int skip_zeromv_motion_search;
513 } FIRST_PASS_SPEED_FEATURES;
514 
515 /*!\cond */
516 typedef struct TPL_SPEED_FEATURES {
517   // GOP length adaptive decision.
518   // If set to 0, tpl model decides whether a shorter gf interval is better.
519   // If set to 1, tpl stats of ARFs from base layer, (base+1) layer and
520   // (base+2) layer decide whether a shorter gf interval is better.
521   // If set to 2, tpl stats of ARFs from base layer, (base+1) layer and GF boost
522   // decide whether a shorter gf interval is better.
523   // If set to 3, gop length adaptive decision is disabled.
524   int gop_length_decision_method;
525   // Prune the intra modes search by tpl.
526   // If set to 0, we will search all intra modes from DC_PRED to PAETH_PRED.
527   // If set to 1, we only search DC_PRED, V_PRED, and H_PRED.
528   int prune_intra_modes;
529   // This parameter controls which step in the n-step process we start at.
530   int reduce_first_step_size;
531   // Skip motion estimation based on the precision of center MVs and the
532   // difference between center MVs.
533   // If set to 0, motion estimation is skipped for duplicate center MVs
534   // (default). If set to 1, motion estimation is skipped for duplicate
535   // full-pixel center MVs. If set to 2, motion estimation is skipped if the
536   // difference between center MVs is less than the threshold.
537   int skip_alike_starting_mv;
538 
539   // When to stop subpel search.
540   SUBPEL_FORCE_STOP subpel_force_stop;
541 
542   // Which search method to use.
543   SEARCH_METHODS search_method;
544 
545   // Prune starting mvs in TPL based on sad scores.
546   int prune_starting_mv;
547 
548   // Prune reference frames in TPL.
549   int prune_ref_frames_in_tpl;
550 
551   // Support compound predictions.
552   int allow_compound_pred;
553 
554   // Calculate rate and distortion based on Y plane only.
555   int use_y_only_rate_distortion;
556 
557   // Use SAD instead of SATD during intra/inter mode search.
558   // If set to 0, use SATD always.
559   // If set to 1, use SAD during intra/inter mode search for frames in the
560   // higher temporal layers of the hierarchical prediction structure.
561   // If set to 2, use SAD during intra/inter mode search for all frames.
562   // This sf is disabled for the first GF group of the key-frame interval,
563   // i.e., SATD is used during intra/inter mode search of the first GF group.
564   int use_sad_for_mode_decision;
565 
566   // Skip tpl processing for frames of type LF_UPDATE.
567   // This sf is disabled for the first GF group of the key-frame interval.
568   int reduce_num_frames;
569 } TPL_SPEED_FEATURES;
570 
571 typedef struct GLOBAL_MOTION_SPEED_FEATURES {
572   GM_SEARCH_TYPE gm_search_type;
573 
574   // During global motion estimation, prune remaining reference frames in a
575   // given direction(past/future), if the evaluated ref_frame in that direction
576   // yields gm_type as INVALID/TRANSLATION/IDENTITY
577   int prune_ref_frame_for_gm_search;
578 
579   // When the current GM type is set to ZEROMV, prune ZEROMV if its performance
580   // is worse than NEWMV under SSE metric.
581   // 0 : no pruning
582   // 1 : conservative pruning
583   // 2 : aggressive pruning
584   int prune_zero_mv_with_sse;
585 
586   // Disable global motion estimation based on stats of previous frames in the
587   // GF group
588   int disable_gm_search_based_on_stats;
589 
590   // Downsampling pyramid level to use for global motion estimation
591   int downsample_level;
592 
593   // Number of refinement steps to apply after initial model generation
594   int num_refinement_steps;
595 } GLOBAL_MOTION_SPEED_FEATURES;
596 
597 typedef struct PARTITION_SPEED_FEATURES {
598   PARTITION_SEARCH_TYPE partition_search_type;
599 
600   // Used if partition_search_type = FIXED_PARTITION
601   BLOCK_SIZE fixed_partition_size;
602 
603   // Prune extended partition types search based on the current best partition
604   // and the combined rdcost of the subblocks estimated from previous
605   // partitions. Can take values 0 - 2, 0 referring to no pruning, and 1 - 2
606   // increasing aggressiveness of pruning in order.
607   int prune_ext_partition_types_search_level;
608 
609   // Prune part4 based on block size
610   int prune_part4_search;
611 
612   // Use a ML model to prune rectangular, ab and 4-way horz
613   // and vert partitions
614   int ml_prune_partition;
615 
616   // Use a ML model to adaptively terminate partition search after trying
617   // PARTITION_SPLIT. Can take values 0 - 2, 0 meaning not being enabled, and
618   // 1 - 2 increasing aggressiveness in order.
619   int ml_early_term_after_part_split_level;
620 
621   // Skip rectangular partition test when partition type none gives better
622   // rd than partition type split. Can take values 0 - 2, 0 referring to no
623   // skipping, and 1 - 2 increasing aggressiveness of skipping in order.
624   int less_rectangular_check_level;
625 
626   // Use square partition only beyond this block size.
627   BLOCK_SIZE use_square_partition_only_threshold;
628 
629   // Sets max square partition levels for this superblock based on
630   // motion vector and prediction error distribution produced from 16x16
631   // simple motion search
632   MAX_PART_PRED_MODE auto_max_partition_based_on_simple_motion;
633 
634   // Min and max square partition size we enable (block_size) as per auto
635   // min max, but also used by adjust partitioning, and pick_partitioning.
636   BLOCK_SIZE default_min_partition_size;
637   BLOCK_SIZE default_max_partition_size;
638 
639   // Sets level of adjustment of variance-based partitioning during
640   // rd_use_partition 0 - no partition adjustment, 1 - try to merge partitions
641   // for small blocks and high QP, 2 - try to merge partitions, 3 - try to merge
642   // and split leaf partitions and 0 - 3 decreasing aggressiveness in order.
643   int adjust_var_based_rd_partitioning;
644 
645   // Partition search early breakout thresholds.
646   int64_t partition_search_breakout_dist_thr;
647   int partition_search_breakout_rate_thr;
648 
649   // Thresholds for ML based partition search breakout.
650   int ml_partition_search_breakout_thresh[PARTITION_BLOCK_SIZES];
651 
652   // Aggressiveness levels for pruning split and rectangular partitions based on
653   // simple_motion_search. SIMPLE_AGG_LVL0 to SIMPLE_AGG_LVL3 correspond to
654   // simple motion search based pruning. QIDX_BASED_AGG_LVL1 corresponds to
655   // qindex based and simple motion search based pruning.
656   int simple_motion_search_prune_agg;
657 
658   // Perform simple_motion_search on each possible subblock and use it to prune
659   // PARTITION_HORZ and PARTITION_VERT.
660   int simple_motion_search_prune_rect;
661 
662   // Perform simple motion search before none_partition to decide if we
663   // want to remove all partitions other than PARTITION_SPLIT. If set to 0, this
664   // model is disabled. If set to 1, the model attempts to perform
665   // PARTITION_SPLIT only. If set to 2, the model also attempts to prune
666   // PARTITION_SPLIT.
667   int simple_motion_search_split;
668 
669   // Use features from simple_motion_search to terminate prediction block
670   // partition after PARTITION_NONE
671   int simple_motion_search_early_term_none;
672 
673   // Controls whether to reduce the number of motion search steps. If this is 0,
674   // then simple_motion_search has the same number of steps as
675   // single_motion_search (assuming no other speed features). Otherwise, reduce
676   // the number of steps by the value contained in this variable.
677   int simple_motion_search_reduce_search_steps;
678 
679   // This variable controls the maximum block size where intra blocks can be
680   // used in inter frames.
681   // TODO(aconverse): Fold this into one of the other many mode skips
682   BLOCK_SIZE max_intra_bsize;
683 
684   // Use CNN with luma pixels on source frame on each of the 64x64 subblock to
685   // perform partition pruning in intra frames.
686   // 0: No Pruning
687   // 1: Prune split and rectangular partitions only
688   // 2: Prune none, split and rectangular partitions
689   int intra_cnn_based_part_prune_level;
690 
691   // Disable extended partition search if the current bsize is greater than the
692   // threshold. Must be a square block size BLOCK_8X8 or higher.
693   BLOCK_SIZE ext_partition_eval_thresh;
694 
695   // Use best partition decision so far to tune 'ext_partition_eval_thresh'
696   int ext_part_eval_based_on_cur_best;
697 
698   // Disable rectangular partitions for larger block sizes.
699   int rect_partition_eval_thresh;
700 
701   // Prune extended partition search based on whether the split/rect partitions
702   // provided an improvement in the previous search.
703   // 0 : no pruning
704   // 1 : prune 1:4 partition search using winner info from split partitions
705   // 2 : prune 1:4 and AB partition search using split and HORZ/VERT info
706   int prune_ext_part_using_split_info;
707 
708   // Prunt rectangular, AB and 4-way partition based on q index and block size
709   // 0 : no pruning
710   // 1 : prune sub_8x8 at very low quantizers
711   // 2 : prune all block size based on qindex
712   int prune_rectangular_split_based_on_qidx;
713 
714   // Prune rectangular partitions based on 4x4 sub-block variance
715   // false : no pruning
716   // true : prune rectangular partitions based on 4x4 sub-block variance
717   // deviation
718   //
719   // For allintra encode, this speed feature reduces instruction count by 6.4%
720   // for speed=6 with coding performance change less than 0.24%. For AVIF image
721   // encode, this speed feature reduces encode time by 8.14% for speed 6 on a
722   // typical image dataset with coding performance change less than 0.16%. This
723   // speed feature is not applicable to speed >= 7.
724   bool prune_rect_part_using_4x4_var_deviation;
725 
726   // Prune rectangular partitions based on prediction mode chosen by NONE
727   // partition.
728   // false : no pruning
729   // true : prunes rectangular partition as described below
730   // If prediction mode chosen by NONE partition is
731   // DC_PRED or SMOOTH_PRED: Prunes both horizontal and vertical partitions if
732   // at least one of the left and top neighbor blocks is larger than the
733   // current block.
734   // Directional Mode: Prunes either of the horizontal and vertical partition
735   // based on center angle of the prediction mode chosen by NONE partition. For
736   // example, vertical partition is pruned if center angle of the prediction
737   // mode chosen by NONE partition is close to 180 degrees (i.e. horizontal
738   // direction) and vice versa.
739   // For allintra encode, this speed feature reduces instruction count by 5.1%
740   // for speed=6 with coding performance change less than 0.22%. For AVIF image
741   // encode, this speed feature reduces encode time by 4.44% for speed 6 on a
742   // typical image dataset with coding performance change less than 0.15%.
743   // For speed >= 7, variance-based logic is used to determine the partition
744   // structure instead of recursive partition search. Therefore, this speed
745   // feature is not applicable in such cases.
746   bool prune_rect_part_using_none_pred_mode;
747 
748   // Terminate partition search for child partition,
749   // when NONE and SPLIT partition rd_costs are INT64_MAX.
750   int early_term_after_none_split;
751 
752   // Level used to adjust threshold for av1_ml_predict_breakout(). At lower
753   // levels, more conservative threshold is used, and value of 0 indicates
754   // av1_ml_predict_breakout() is disabled. Value of 3 corresponds to default
755   // case with no adjustment to lbd thresholds.
756   int ml_predict_breakout_level;
757 
758   // Prune sub_8x8 (BLOCK_4X4, BLOCK_4X8 and BLOCK_8X4) partitions.
759   // 0 : no pruning
760   // 1 : pruning based on neighbour block information
761   // 2 : prune always
762   int prune_sub_8x8_partition_level;
763 
764   // Prune rectangular split based on simple motion search split/no_split score.
765   // 0: disable pruning, 1: enable pruning
766   int simple_motion_search_rect_split;
767 
768   // The current encoder adopts a DFS search for block partitions.
769   // Therefore the mode selection and associated rdcost is ready for smaller
770   // blocks before the mode selection for some partition types.
771   // AB partition could use previous rd information and skip mode search.
772   // An example is:
773   //
774   //  current block
775   //  +---+---+
776   //  |       |
777   //  +       +
778   //  |       |
779   //  +-------+
780   //
781   //  SPLIT partition has been searched first before trying HORZ_A
782   //  +---+---+
783   //  | R | R |
784   //  +---+---+
785   //  | R | R |
786   //  +---+---+
787   //
788   //  HORZ_A
789   //  +---+---+
790   //  |   |   |
791   //  +---+---+
792   //  |       |
793   //  +-------+
794   //
795   //  With this speed feature, the top two sub blocks can directly use rdcost
796   //  searched in split partition, and the mode info is also copied from
797   //  saved info. Similarly, the bottom rectangular block can also use
798   //  the available information from previous rectangular search.
799   int reuse_prev_rd_results_for_part_ab;
800 
801   // Reuse the best prediction modes found in PARTITION_SPLIT and PARTITION_RECT
802   // when encoding PARTITION_AB.
803   int reuse_best_prediction_for_part_ab;
804 
805   // The current partition search records the best rdcost so far and uses it
806   // in mode search and transform search to early skip when some criteria is
807   // met. For example, when the current rdcost is larger than the best rdcost,
808   // or the model rdcost is larger than the best rdcost times some thresholds.
809   // By default, this feature is turned on to speed up the encoder partition
810   // search.
811   // If disabling it, at speed 0, 30 frames, we could get
812   // about -0.25% quality gain (psnr, ssim, vmaf), with about 13% slowdown.
813   int use_best_rd_for_pruning;
814 
815   // Skip evaluation of non-square partitions based on the corresponding NONE
816   // partition.
817   // 0: no pruning
818   // 1: prune extended partitions if NONE is skippable
819   // 2: on top of 1, prune rectangular partitions if NONE is inter, not a newmv
820   // mode and skippable
821   int skip_non_sq_part_based_on_none;
822 
823   // Disables 8x8 and below partitions for low quantizers.
824   int disable_8x8_part_based_on_qidx;
825 } PARTITION_SPEED_FEATURES;
826 
827 typedef struct MV_SPEED_FEATURES {
828   // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
829   SEARCH_METHODS search_method;
830 
831   // Enable the use of faster, less accurate mv search method
832   // 0: disable, 1: if bsize >= BLOCK_32X32, 2: based on bsize, SAD and qp
833   // TODO([email protected]): Take the clip's resolution and mv activity into
834   // account.
835   int use_bsize_dependent_search_method;
836 
837   // If this is set to 1, we limit the motion search range to 2 times the
838   // largest motion vector found in the last frame.
839   int auto_mv_step_size;
840 
841   // Subpel_search_method can only be subpel_tree which does a subpixel
842   // logarithmic search that keeps stepping at 1/2 pixel units until
843   // you stop getting a gain, and then goes on to 1/4 and repeats
844   // the same process. Along the way it skips many diagonals.
845   SUBPEL_SEARCH_METHOD subpel_search_method;
846 
847   // Maximum number of steps in logarithmic subpel search before giving up.
848   int subpel_iters_per_step;
849 
850   // When to stop subpel search.
851   SUBPEL_FORCE_STOP subpel_force_stop;
852 
853   // When to stop subpel search in simple motion search.
854   SUBPEL_FORCE_STOP simple_motion_subpel_force_stop;
855 
856   // If true, sub-pixel search uses the exact convolve function used for final
857   // encoding and decoding; otherwise, it uses bilinear interpolation.
858   SUBPEL_SEARCH_TYPE use_accurate_subpel_search;
859 
860   // Threshold for allowing exhaustive motion search.
861   int exhaustive_searches_thresh;
862 
863   // Pattern to be used for any exhaustive mesh searches (except intraBC ME).
864   MESH_PATTERN mesh_patterns[MAX_MESH_STEP];
865 
866   // Pattern to be used for exhaustive mesh searches of intraBC ME.
867   MESH_PATTERN intrabc_mesh_patterns[MAX_MESH_STEP];
868 
869   // Reduce single motion search range based on MV result of prior ref_mv_idx.
870   int reduce_search_range;
871 
872   // Prune mesh search.
873   PRUNE_MESH_SEARCH_LEVEL prune_mesh_search;
874 
875   // Use the rd cost around the best FULLPEL_MV to speed up subpel search
876   int use_fullpel_costlist;
877 
878   // Set the full pixel search level of obmc
879   // 0: obmc_full_pixel_diamond
880   // 1: obmc_refining_search_sad (faster)
881   int obmc_full_pixel_search_level;
882 
883   // Accurate full pixel motion search based on TPL stats.
884   int full_pixel_search_level;
885 
886   // Allow intrabc motion search
887   int use_intrabc;
888 
889   // Whether to downsample the rows in sad calculation during motion search.
890   // This is only active when there are at least 16 rows. When this sf is
891   // active, if there is a large discrepancy in the SAD values for the final
892   // motion vector between skipping vs not skipping, motion search is redone
893   // with skip row features off.
894   // 0: Disabled (do not downsample rows)
895   // 1: Skip SAD calculation of odd rows if the SAD deviation of the even and
896   //    odd rows for the starting MV is small. Redo motion search with sf off
897   //    when SAD deviation is high for the final motion vector.
898   // 2: Skip SAD calculation of odd rows. SAD deviation is not tested for the
899   //    start MV and tested only for the final MV.
900   int use_downsampled_sad;
901 
902   // Enable/disable extensive joint motion search.
903   int disable_extensive_joint_motion_search;
904 
905   // Enable second best mv check in joint mv search.
906   // 0: allow second MV (use rd cost as the metric)
907   // 1: use var as the metric
908   // 2: disable second MV
909   int disable_second_mv;
910 
911   // Skips full pixel search based on start mv of prior ref_mv_idx.
912   // 0: Disabled
913   // 1: Skips the full pixel search upto 4 neighbor full-pel MV positions.
914   // 2: Skips the full pixel search upto 8 neighbor full-pel MV positions.
915   int skip_fullpel_search_using_startmv;
916 
917   // Method to use for refining WARPED_CAUSAL motion vectors
918   // TODO(rachelbarker): Can this be unified with OBMC in some way?
919   WARP_SEARCH_METHOD warp_search_method;
920 
921   // Maximum number of iterations in WARPED_CAUSAL refinement search
922   int warp_search_iters;
923 } MV_SPEED_FEATURES;
924 
925 typedef struct INTER_MODE_SPEED_FEATURES {
926   // 2-pass inter mode model estimation where the preliminary pass skips
927   // transform search and uses a model to estimate rd, while the final pass
928   // computes the full transform search. Two types of models are supported:
929   // 0: not used
930   // 1: used with online dynamic rd model
931   // 2: used with static rd model
932   int inter_mode_rd_model_estimation;
933 
934   // Bypass transform search based on skip rd at following stages
935   //   i. Compound type mode search
936   //  ii. Motion mode search (mode evaluation and winner motion mode stage)
937   // iii. Transform search for best inter candidates
938   int txfm_rd_gate_level[TX_SEARCH_CASES];
939 
940   // Limit the inter mode tested in the RD loop
941   int reduce_inter_modes;
942 
943   // This variable is used to cap the maximum number of times we skip testing a
944   // mode to be evaluated. A high value means we will be faster.
945   int adaptive_rd_thresh;
946 
947   // Aggressively prune inter modes when best mode is skippable.
948   int prune_inter_modes_if_skippable;
949 
950   // Drop less likely to be picked reference frames in the RD search.
951   // Has seven levels for now: 0, 1, 2, 3, 4, 5 and 6 where higher levels prune
952   // more aggressively than lower ones. (0 means no pruning).
953   int selective_ref_frame;
954 
955   // Prune reference frames for rectangular partitions.
956   // 0 implies no pruning
957   // 1 implies prune for extended partition
958   // 2 implies prune horiz, vert and extended partition
959   int prune_ref_frame_for_rect_partitions;
960 
961   // Prune inter modes w.r.t past reference frames
962   // 0 no pruning
963   // 1 prune inter modes w.r.t ALTREF2 and ALTREF reference frames
964   // 2 prune inter modes w.r.t BWDREF, ALTREF2 and ALTREF reference frames
965   int alt_ref_search_fp;
966 
967   // Prune reference frames for single prediction modes based on temporal
968   // distance and pred MV SAD. Feasible values are 0, 1, 2. The feature is
969   // disabled for 0. An increasing value indicates more aggressive pruning
970   // threshold.
971   int prune_single_ref;
972 
973   // Prune compound reference frames
974   // 0 no pruning
975   // 1 prune compound references which do not satisfy the two conditions:
976   //   a) The references are at a nearest distance from the current frame in
977   //   both past and future direction.
978   //   b) The references have minimum pred_mv_sad in both past and future
979   //   direction.
980   // 2 prune compound references except the one with nearest distance from the
981   //   current frame in both past and future direction.
982   int prune_comp_ref_frames;
983 
984   // Skip the current ref_mv in NEW_MV mode based on mv, rate cost, etc.
985   // This speed feature equaling 0 means no skipping.
986   // If the speed feature equals 1 or 2, skip the current ref_mv in NEW_MV mode
987   // if we have already encountered ref_mv in the drl such that:
988   //  1. The other drl has the same mv during the SIMPLE_TRANSLATION search
989   //     process as the current mv.
990   //  2. The rate needed to encode the current mv is larger than that for the
991   //     other ref_mv.
992   // The speed feature equaling 1 means using subpel mv in the comparison.
993   // The speed feature equaling 2 means using fullpel mv in the comparison.
994   // If the speed feature >= 3, skip the current ref_mv in NEW_MV mode based on
995   // known full_mv bestsme and drl cost.
996   int skip_newmv_in_drl;
997 
998   // This speed feature checks duplicate ref MVs among NEARESTMV, NEARMV,
999   // GLOBALMV and skips NEARMV or GLOBALMV (in order) if a duplicate is found
1000   // TODO(any): Instead of skipping repeated ref mv, use the recalculated
1001   // rd-cost based on mode rate and skip the mode evaluation
1002   int skip_repeated_ref_mv;
1003 
1004   // Flag used to control the ref_best_rd based gating for chroma
1005   int perform_best_rd_based_gating_for_chroma;
1006 
1007   // Reuse the inter_intra_mode search result from NEARESTMV mode to other
1008   // single ref modes
1009   int reuse_inter_intra_mode;
1010 
1011   // prune wedge and compound segment approximate rd evaluation based on
1012   // compound average modeled rd
1013   int prune_comp_type_by_model_rd;
1014 
1015   // prune wedge and compound segment approximate rd evaluation based on
1016   // compound average rd/ref_best_rd
1017   int prune_comp_type_by_comp_avg;
1018 
1019   // Skip some ref frames in compound motion search by single motion search
1020   // result. Has three levels for now: 0 referring to no skipping, and 1 - 3
1021   // increasing aggressiveness of skipping in order.
1022   // Note: The search order might affect the result. It assumes that the single
1023   // reference modes are searched before compound modes. It is better to search
1024   // same single inter mode as a group.
1025   int prune_comp_search_by_single_result;
1026 
1027   // Instead of performing a full MV search, do a simple translation first
1028   // and only perform a full MV search on the motion vectors that performed
1029   // well.
1030   int prune_mode_search_simple_translation;
1031 
1032   // Only search compound modes with at least one "good" reference frame.
1033   // A reference frame is good if, after looking at its performance among
1034   // the single reference modes, it is one of the two best performers.
1035   int prune_compound_using_single_ref;
1036 
1037   // Skip extended compound mode (NEAREST_NEWMV, NEW_NEARESTMV, NEAR_NEWMV,
1038   // NEW_NEARMV) using ref frames of above and left neighbor
1039   // blocks.
1040   // 0 : no pruning
1041   // 1 : prune ext compound modes using neighbor blocks (less aggressiveness)
1042   // 2 : prune ext compound modes using neighbor blocks (high aggressiveness)
1043   // 3 : prune ext compound modes unconditionally (highest aggressiveness)
1044   int prune_ext_comp_using_neighbors;
1045 
1046   // Skip NEW_NEARMV and NEAR_NEWMV extended compound modes
1047   int skip_ext_comp_nearmv_mode;
1048 
1049   // Skip extended compound mode when ref frame corresponding to NEWMV does not
1050   // have NEWMV as single mode winner.
1051   // 0 : no pruning
1052   // 1 : prune extended compound mode (less aggressiveness)
1053   // 2 : prune extended compound mode (high aggressiveness)
1054   int prune_comp_using_best_single_mode_ref;
1055 
1056   // Skip NEARESTMV and NEARMV using weight computed in ref mv list population
1057   //
1058   // Pruning is enabled only when both the top and left neighbor blocks are
1059   // available and when the current block already has a valid inter prediction.
1060   int prune_nearest_near_mv_using_refmv_weight;
1061 
1062   // Based on previous ref_mv_idx search result, prune the following search.
1063   int prune_ref_mv_idx_search;
1064 
1065   // Disable one sided compound modes.
1066   int disable_onesided_comp;
1067 
1068   // Prune obmc search using previous frame stats.
1069   // INT_MAX : disable obmc search
1070   int prune_obmc_prob_thresh;
1071 
1072   // Prune warped motion search using previous frame stats.
1073   int prune_warped_prob_thresh;
1074 
1075   // Variance threshold to enable/disable Interintra wedge search
1076   unsigned int disable_interintra_wedge_var_thresh;
1077 
1078   // Variance threshold to enable/disable Interinter wedge search
1079   unsigned int disable_interinter_wedge_var_thresh;
1080 
1081   // De-couple wedge and mode search during interintra RDO.
1082   int fast_interintra_wedge_search;
1083 
1084   // Whether fast wedge sign estimate is used
1085   int fast_wedge_sign_estimate;
1086 
1087   // Enable/disable ME for interinter wedge search.
1088   int disable_interinter_wedge_newmv_search;
1089 
1090   // Decide when and how to use joint_comp.
1091   DIST_WTD_COMP_FLAG use_dist_wtd_comp_flag;
1092 
1093   // Clip the frequency of updating the mv cost.
1094   INTERNAL_COST_UPDATE_TYPE mv_cost_upd_level;
1095 
1096   // Clip the frequency of updating the coeff cost.
1097   INTERNAL_COST_UPDATE_TYPE coeff_cost_upd_level;
1098 
1099   // Clip the frequency of updating the mode cost.
1100   INTERNAL_COST_UPDATE_TYPE mode_cost_upd_level;
1101 
1102   // Prune inter modes based on tpl stats
1103   // 0 : no pruning
1104   // 1 - 3 indicate increasing aggressiveness in order.
1105   int prune_inter_modes_based_on_tpl;
1106 
1107   // Skip NEARMV and NEAR_NEARMV modes using ref frames of above and left
1108   // neighbor blocks and qindex.
1109   PRUNE_NEARMV_LEVEL prune_nearmv_using_neighbors;
1110 
1111   // Model based breakout after interpolation filter search
1112   // 0: no breakout
1113   // 1: use model based rd breakout
1114   int model_based_post_interp_filter_breakout;
1115 
1116   // Reuse compound type rd decision when exact match is found
1117   // 0: No reuse
1118   // 1: Reuse the compound type decision
1119   int reuse_compound_type_decision;
1120 
1121   // Enable/disable masked compound.
1122   int disable_masked_comp;
1123 
1124   // Enable/disable MV refinement for compound modes corresponds to compound
1125   // types COMPOUND_AVERAGE, COMPOUND_DISTWTD (currently, this compound type
1126   // is disabled for speeds >= 2 using the sf 'use_dist_wtd_comp_flag') and
1127   // COMPOUND_DIFFWTD based on the availability. Levels 0 to 3 indicate
1128   // increasing order of aggressiveness to disable MV refinement.
1129   // 0: MV Refinement is enabled and for NEW_NEWMV mode used two iterations of
1130   // refinement in av1_joint_motion_search().
1131   // 1: MV Refinement is disabled for COMPOUND_DIFFWTD and enabled for
1132   // COMPOUND_AVERAGE & COMPOUND_DISTWTD.
1133   // 2: MV Refinement is enabled for COMPOUND_AVERAGE & COMPOUND_DISTWTD for
1134   // NEW_NEWMV mode with one iteration of refinement in
1135   // av1_joint_motion_search() and MV Refinement is disabled for other compound
1136   // type modes.
1137   // 3: MV Refinement is disabled.
1138   int enable_fast_compound_mode_search;
1139 
1140   // Reuse masked compound type search results
1141   int reuse_mask_search_results;
1142 
1143   // Enable/disable fast search for wedge masks
1144   int enable_fast_wedge_mask_search;
1145 
1146   // Early breakout from transform search of inter modes
1147   int inter_mode_txfm_breakout;
1148 
1149   // Limit number of inter modes for txfm search if a newmv mode gets
1150   // evaluated among the top modes.
1151   // 0: no pruning
1152   // 1 to 3 indicate increasing order of aggressiveness
1153   int limit_inter_mode_cands;
1154 
1155   // Cap the no. of txfm searches for a given prediction mode.
1156   // 0: no cap, 1: cap beyond first 4 searches, 2: cap beyond first 3 searches.
1157   int limit_txfm_eval_per_mode;
1158 
1159   // Prune warped motion search based on block size.
1160   int extra_prune_warped;
1161 
1162   // Do not search compound modes for ARF.
1163   // The intuition is that ARF is predicted by frames far away from it,
1164   // whose temporal correlations with the ARF are likely low.
1165   // It is therefore likely that compound modes do not work as well for ARF
1166   // as other inter frames.
1167   // Speed/quality impact:
1168   // Speed 1: 12% faster, 0.1% psnr loss.
1169   // Speed 2: 2%  faster, 0.05% psnr loss.
1170   // No change for speed 3 and up, because |disable_onesided_comp| is true.
1171   int skip_arf_compound;
1172 } INTER_MODE_SPEED_FEATURES;
1173 
1174 typedef struct INTERP_FILTER_SPEED_FEATURES {
1175   // Do limited interpolation filter search for dual filters, since best choice
1176   // usually includes EIGHTTAP_REGULAR.
1177   int use_fast_interpolation_filter_search;
1178 
1179   // Disable dual filter
1180   int disable_dual_filter;
1181 
1182   // Save results of av1_interpolation_filter_search for a block
1183   // Check mv and ref_frames before search, if they are very close with previous
1184   // saved results, filter search can be skipped.
1185   int use_interp_filter;
1186 
1187   // skip sharp_filter evaluation based on regular and smooth filter rd for
1188   // dual_filter=0 case
1189   int skip_sharp_interp_filter_search;
1190 
1191   // skip interpolation filter search for a block in chessboard pattern
1192   int cb_pred_filter_search;
1193 
1194   // adaptive interp_filter search to allow skip of certain filter types.
1195   int adaptive_interp_filter_search;
1196 
1197   // Forces interpolation filter to EIGHTTAP_REGULAR and skips interpolation
1198   // filter search.
1199   int skip_interp_filter_search;
1200 } INTERP_FILTER_SPEED_FEATURES;
1201 
1202 typedef struct INTRA_MODE_SPEED_FEATURES {
1203   // These bit masks allow you to enable or disable intra modes for each
1204   // transform size separately.
1205   int intra_y_mode_mask[TX_SIZES];
1206   int intra_uv_mode_mask[TX_SIZES];
1207 
1208   // flag to allow skipping intra mode for inter frame prediction
1209   int skip_intra_in_interframe;
1210 
1211   // Prune intra mode candidates based on source block histogram of gradient.
1212   // Applies to luma plane only.
1213   // Feasible values are 0..4. The feature is disabled for 0. An increasing
1214   // value indicates more aggressive pruning threshold.
1215   int intra_pruning_with_hog;
1216 
1217   // Prune intra mode candidates based on source block histogram of gradient.
1218   // Applies to chroma plane only.
1219   // Feasible values are 0..4. The feature is disabled for 0. An increasing
1220   // value indicates more aggressive pruning threshold.
1221   int chroma_intra_pruning_with_hog;
1222 
1223   // Enable/disable smooth intra modes.
1224   int disable_smooth_intra;
1225 
1226   // Prune UV_SMOOTH_PRED mode for chroma based on chroma source variance.
1227   // false : No pruning
1228   // true  : Prune UV_SMOOTH_PRED mode based on chroma source variance
1229   //
1230   // For allintra encode, this speed feature reduces instruction count
1231   // by 1.90%, 2.21% and 1.97% for speed 6, 7 and 8 with coding performance
1232   // change less than 0.04%. For AVIF image encode, this speed feature reduces
1233   // encode time by 1.56%, 2.14% and 0.90% for speed 6, 7 and 8 on a typical
1234   // image dataset with coding performance change less than 0.05%.
1235   bool prune_smooth_intra_mode_for_chroma;
1236 
1237   // Prune filter intra modes in intra frames.
1238   // 0 : No pruning
1239   // 1 : Evaluate applicable filter intra modes based on best intra mode so far
1240   // 2 : Do not evaluate filter intra modes
1241   int prune_filter_intra_level;
1242 
1243   // prune palette search
1244   // 0: No pruning
1245   // 1: Perform coarse search to prune the palette colors. For winner colors,
1246   // neighbors are also evaluated using a finer search.
1247   // 2: Perform 2 way palette search from max colors to min colors (and min
1248   // colors to remaining colors) and terminate the search if current number of
1249   // palette colors is not the winner.
1250   int prune_palette_search_level;
1251 
1252   // Terminate early in luma palette_size search. Speed feature values indicate
1253   // increasing level of pruning.
1254   // 0: No early termination
1255   // 1: Terminate early for higher luma palette_size, if header rd cost of lower
1256   // palette_size is more than 2 * best_rd. This level of pruning is more
1257   // conservative when compared to sf level 2 as the cases which will get pruned
1258   // with sf level 1 is a subset of the cases which will get pruned with sf
1259   // level 2.
1260   // 2: Terminate early for higher luma palette_size, if header rd cost of lower
1261   // palette_size is more than best_rd.
1262   // For allintra encode, this sf reduces instruction count by 2.49%, 1.07%,
1263   // 2.76%, 2.30%, 1.84%, 2.69%, 2.04%, 2.05% and 1.44% for speed 0, 1, 2, 3, 4,
1264   // 5, 6, 7 and 8 on screen content set with coding performance change less
1265   // than 0.01% for speed <= 2 and less than 0.03% for speed >= 3. For AVIF
1266   // image encode, this sf reduces instruction count by 1.94%, 1.13%, 1.29%,
1267   // 0.93%, 0.89%, 1.03%, 1.07%, 1.20% and 0.18% for speed 0, 1, 2, 3, 4, 5, 6,
1268   // 7 and 8 on a typical image dataset with coding performance change less than
1269   // 0.01%.
1270   int prune_luma_palette_size_search_level;
1271 
1272   // Prune chroma intra modes based on luma intra mode winner.
1273   // 0: No pruning
1274   // 1: Prune chroma intra modes other than UV_DC_PRED, UV_SMOOTH_PRED,
1275   // UV_CFL_PRED and the mode that corresponds to luma intra mode winner.
1276   int prune_chroma_modes_using_luma_winner;
1277 
1278   // Clip the frequency of updating the mv cost for intrabc.
1279   INTERNAL_COST_UPDATE_TYPE dv_cost_upd_level;
1280 
1281   // We use DCT_DCT transform followed by computing SATD (Sum of Absolute
1282   // Transformed Differences) as an estimation of RD score to quickly find the
1283   // best possible Chroma from Luma (CFL) parameter. Then we do a full RD search
1284   // near the best possible parameter. The search range is set here.
1285   // The range of cfl_searh_range should be [1, 33], and the following are the
1286   // recommended values.
1287   // 1: Fastest mode.
1288   // 3: Default mode that provides good speedup without losing compression
1289   // performance at speed 0.
1290   // 33: Exhaustive rd search (33 == CFL_MAGS_SIZE). This mode should only
1291   // be used for debugging purpose.
1292   int cfl_search_range;
1293 
1294   // TOP_INTRA_MODEL_COUNT is 4 that is the number of top model rd to store in
1295   // intra mode decision. Here, add a speed feature to reduce this number for
1296   // higher speeds.
1297   int top_intra_model_count_allowed;
1298 
1299   // Adapt top_intra_model_count_allowed locally to prune luma intra modes using
1300   // neighbor block and quantizer information.
1301   int adapt_top_model_rd_count_using_neighbors;
1302 
1303   // Prune the evaluation of odd delta angles of directional luma intra modes by
1304   // using the rdcosts of neighbouring delta angles.
1305   // For allintra encode, this speed feature reduces instruction count
1306   // by 4.461%, 3.699% and 3.536% for speed 6, 7 and 8 on a typical video
1307   // dataset with coding performance change less than 0.26%. For AVIF image
1308   // encode, this speed feature reduces encode time by 2.849%, 2.471%,
1309   // and 2.051% for speed 6, 7 and 8 on a typical image dataset with coding
1310   // performance change less than 0.27%.
1311   int prune_luma_odd_delta_angles_in_intra;
1312 
1313   // Terminate early in chroma palette_size search.
1314   // 0: No early termination
1315   // 1: Terminate early for higher palette_size, if header rd cost of lower
1316   // palette_size is more than best_rd.
1317   // For allintra encode, this sf reduces instruction count by 0.45%,
1318   // 0.62%, 1.73%, 2.50%, 2.89%, 3.09% and 3.86% for speed 0 to 6 on screen
1319   // content set with coding performance change less than 0.01%.
1320   // For AVIF image encode, this sf reduces instruction count by 0.45%, 0.81%,
1321   // 0.85%, 1.05%, 1.45%, 1.66% and 1.95% for speed 0 to 6 on a typical image
1322   // dataset with no quality drop.
1323   int early_term_chroma_palette_size_search;
1324 
1325   // Skips the evaluation of filter intra modes in inter frames if rd evaluation
1326   // of luma intra dc mode results in invalid rd stats.
1327   int skip_filter_intra_in_inter_frames;
1328 } INTRA_MODE_SPEED_FEATURES;
1329 
1330 typedef struct TX_SPEED_FEATURES {
1331   // Init search depth for square and rectangular transform partitions.
1332   // Values:
1333   // 0 - search full tree, 1: search 1 level, 2: search the highest level only
1334   int inter_tx_size_search_init_depth_sqr;
1335   int inter_tx_size_search_init_depth_rect;
1336   int intra_tx_size_search_init_depth_sqr;
1337   int intra_tx_size_search_init_depth_rect;
1338 
1339   // If any dimension of a coding block size above 64, always search the
1340   // largest transform only, since the largest transform block size is 64x64.
1341   int tx_size_search_lgr_block;
1342 
1343   TX_TYPE_SEARCH tx_type_search;
1344 
1345   // Skip split transform block partition when the collocated bigger block
1346   // is selected as all zero coefficients.
1347   int txb_split_cap;
1348 
1349   // Shortcut the transform block partition and type search when the target
1350   // rdcost is relatively lower.
1351   // Values are 0 (not used) , or 1 - 2 with progressively increasing
1352   // aggressiveness
1353   int adaptive_txb_search_level;
1354 
1355   // Prune level for tx_size_type search for inter based on rd model
1356   // 0: no pruning
1357   // 1-2: progressively increasing aggressiveness of pruning
1358   int model_based_prune_tx_search_level;
1359 
1360   // Refine TX type after fast TX search.
1361   int refine_fast_tx_search_results;
1362 
1363   // Prune transform split/no_split eval based on residual properties. A value
1364   // of 0 indicates no pruning, and the aggressiveness of pruning progressively
1365   // increases from levels 1 to 3.
1366   int prune_tx_size_level;
1367 
1368   // Prune the evaluation of transform depths as decided by the NN model.
1369   // false: No pruning.
1370   // true : Avoid the evaluation of specific transform depths using NN model.
1371   //
1372   // For allintra encode, this speed feature reduces instruction count
1373   // by 4.76%, 8.92% and 11.28% for speed 6, 7 and 8 with coding performance
1374   // change less than 0.32%. For AVIF image encode, this speed feature reduces
1375   // encode time by 4.65%, 9.16% and 10.45% for speed 6, 7 and 8 on a typical
1376   // image dataset with coding performance change less than 0.19%.
1377   bool prune_intra_tx_depths_using_nn;
1378 
1379   // Enable/disable early breakout during transform search of intra modes, by
1380   // using the minimum rd cost possible. By using this approach, the rd
1381   // evaluation of applicable transform blocks (in the current block) can be
1382   // avoided as
1383   // 1) best_rd evolves during the search in choose_tx_size_type_from_rd()
1384   // 2) appropriate ref_best_rd is passed in intra_block_yrd()
1385   //
1386   // For allintra encode, this speed feature reduces instruction count
1387   // by 1.11%, 1.08%, 1.02% and 0.93% for speed 3, 6, 7 and 8 with coding
1388   // performance change less than 0.02%. For AVIF image encode, this speed
1389   // feature reduces encode time by 0.93%, 1.46%, 1.07%, 0.84%, 0.99% and 0.73%
1390   // for speed 3, 4, 5, 6, 7 and 8 on a typical image dataset with coding
1391   // performance change less than 0.004%.
1392   bool use_rd_based_breakout_for_intra_tx_search;
1393 } TX_SPEED_FEATURES;
1394 
1395 typedef struct RD_CALC_SPEED_FEATURES {
1396   // Fast approximation of av1_model_rd_from_var_lapndz
1397   int simple_model_rd_from_var;
1398 
1399   // Perform faster distortion computation during the R-D evaluation by trying
1400   // to approximate the prediction error with transform coefficients (faster but
1401   // less accurate) rather than computing distortion in the pixel domain (slower
1402   // but more accurate). The following methods are used for distortion
1403   // computation:
1404   // Method 0: Always compute distortion in the pixel domain
1405   // Method 1: Based on block error, try using transform domain distortion for
1406   // tx_type search and compute distortion in pixel domain for final RD_STATS
1407   // Method 2: Based on block error, try to compute distortion in transform
1408   // domain
1409   // Methods 1 and 2 may fallback to computing distortion in the pixel domain in
1410   // case the block error is less than the threshold, which is controlled by the
1411   // speed feature tx_domain_dist_thres_level.
1412   //
1413   // The speed feature tx_domain_dist_level decides which of the above methods
1414   // needs to be used across different mode evaluation stages as described
1415   // below:
1416   // Eval type:    Default      Mode        Winner
1417   // Level 0  :    Method 0    Method 2    Method 0
1418   // Level 1  :    Method 1    Method 2    Method 0
1419   // Level 2  :    Method 2    Method 2    Method 0
1420   // Level 3  :    Method 2    Method 2    Method 2
1421   int tx_domain_dist_level;
1422 
1423   // Transform domain distortion threshold level
1424   int tx_domain_dist_thres_level;
1425 
1426   // Trellis (dynamic programming) optimization of quantized values
1427   TRELLIS_OPT_TYPE optimize_coefficients;
1428 
1429   // Use hash table to store macroblock RD search results
1430   // to avoid repeated search on the same residue signal.
1431   int use_mb_rd_hash;
1432 
1433   // Flag used to control the extent of coeff R-D optimization
1434   int perform_coeff_opt;
1435 } RD_CALC_SPEED_FEATURES;
1436 
1437 typedef struct WINNER_MODE_SPEED_FEATURES {
1438   // Flag used to control the winner mode processing for better R-D optimization
1439   // of quantized coeffs
1440   int enable_winner_mode_for_coeff_opt;
1441 
1442   // Flag used to control the winner mode processing for transform size
1443   // search method
1444   int enable_winner_mode_for_tx_size_srch;
1445 
1446   // Control transform size search level
1447   // Eval type: Default       Mode        Winner
1448   // Level 0  : FULL RD     LARGEST ALL   FULL RD
1449   // Level 1  : FAST RD     LARGEST ALL   FULL RD
1450   // Level 2  : LARGEST ALL LARGEST ALL   FULL RD
1451   // Level 3 :  LARGEST ALL LARGEST ALL   LARGEST ALL
1452   int tx_size_search_level;
1453 
1454   // Flag used to control the winner mode processing for use transform
1455   // domain distortion
1456   int enable_winner_mode_for_use_tx_domain_dist;
1457 
1458   // Flag used to enable processing of multiple winner modes
1459   MULTI_WINNER_MODE_TYPE multi_winner_mode_type;
1460 
1461   // Motion mode for winner candidates:
1462   // 0: speed feature OFF
1463   // 1 / 2 : Use configured number of winner candidates
1464   int motion_mode_for_winner_cand;
1465 
1466   // Controls the prediction of transform skip block or DC only block.
1467   //
1468   // Different speed feature values (0 to 3) decide the aggressiveness of
1469   // prediction (refer to predict_dc_levels[][] in speed_features.c) to be used
1470   // during different mode evaluation stages.
1471   int dc_blk_pred_level;
1472 
1473   // If on, disables interpolation filter search in handle_inter_mode loop, and
1474   // performs it during winner mode processing by \ref
1475   // tx_search_best_inter_candidates.
1476   int winner_mode_ifs;
1477 
1478   // Controls the disabling of winner mode processing. Speed feature levels
1479   // are ordered in increasing aggressiveness of pruning. The method considered
1480   // for disabling, depends on the sf level value and it is described as below.
1481   // 0: Do not disable
1482   // 1: Disable for blocks with low source variance.
1483   // 2: Disable for blocks which turn out to be transform skip (skipped based on
1484   // eob) during MODE_EVAL stage except NEWMV mode.
1485   // 3: Disable for blocks which turn out to be transform skip during MODE_EVAL
1486   // stage except NEWMV mode. For high quantizers, prune conservatively based on
1487   // transform skip (skipped based on eob) except for NEWMV mode.
1488   // 4: Disable for blocks which turn out to be transform skip during MODE_EVAL
1489   // stage.
1490   int prune_winner_mode_eval_level;
1491 } WINNER_MODE_SPEED_FEATURES;
1492 
1493 typedef struct LOOP_FILTER_SPEED_FEATURES {
1494   // This feature controls how the loop filter level is determined.
1495   LPF_PICK_METHOD lpf_pick;
1496 
1497   // Skip some final iterations in the determination of the best loop filter
1498   // level.
1499   int use_coarse_filter_level_search;
1500 
1501   // Control how the CDEF strength is determined.
1502   CDEF_PICK_METHOD cdef_pick_method;
1503 
1504   // Decoder side speed feature to add penalty for use of dual-sgr filters.
1505   // Takes values 0 - 10, 0 indicating no penalty and each additional level
1506   // adding a penalty of 1%
1507   int dual_sgr_penalty_level;
1508 
1509   // prune sgr ep using binary search like mechanism
1510   int enable_sgr_ep_pruning;
1511 
1512   // Disable loop restoration for Chroma plane
1513   int disable_loop_restoration_chroma;
1514 
1515   // Disable loop restoration for luma plane
1516   int disable_loop_restoration_luma;
1517 
1518   // Range of loop restoration unit sizes to search
1519   // The minimum size is clamped against the superblock size in
1520   // av1_pick_filter_restoration, so that the code which sets this value does
1521   // not need to know the superblock size ahead of time.
1522   int min_lr_unit_size;
1523   int max_lr_unit_size;
1524 
1525   // Prune RESTORE_WIENER evaluation based on source variance
1526   // 0 : no pruning
1527   // 1 : conservative pruning
1528   // 2 : aggressive pruning
1529   int prune_wiener_based_on_src_var;
1530 
1531   // Prune self-guided loop restoration based on wiener search results
1532   // 0 : no pruning
1533   // 1 : pruning based on rdcost ratio of RESTORE_WIENER and RESTORE_NONE
1534   // 2 : pruning based on winner restoration type among RESTORE_WIENER and
1535   // RESTORE_NONE
1536   int prune_sgr_based_on_wiener;
1537 
1538   // Reduce the wiener filter win size for luma
1539   int reduce_wiener_window_size;
1540 
1541   // Flag to disable Wiener Loop restoration filter.
1542   bool disable_wiener_filter;
1543 
1544   // Flag to disable Self-guided Loop restoration filter.
1545   bool disable_sgr_filter;
1546 
1547   // Disable the refinement search around the wiener filter coefficients.
1548   bool disable_wiener_coeff_refine_search;
1549 
1550   // Whether to downsample the rows in computation of wiener stats.
1551   int use_downsampled_wiener_stats;
1552 } LOOP_FILTER_SPEED_FEATURES;
1553 
1554 typedef struct REAL_TIME_SPEED_FEATURES {
1555   // check intra prediction for non-RD mode.
1556   int check_intra_pred_nonrd;
1557 
1558   // Skip checking intra prediction.
1559   // 0 - don't skip
1560   // 1 - skip if TX is skipped and best mode is not NEWMV
1561   // 2 - skip if TX is skipped
1562   // Skipping aggressiveness increases from level 1 to 2.
1563   int skip_intra_pred;
1564 
1565   // Estimate motion before calculating variance in variance-based partition
1566   // 0 - Only use zero MV
1567   // 1 - perform coarse ME
1568   // 2 - perform coarse ME, and also use neighbours' MVs
1569   // 3 - use neighbours' MVs without performing coarse ME
1570   int estimate_motion_for_var_based_partition;
1571 
1572   // For nonrd_use_partition: mode of extra check of leaf partition
1573   // 0 - don't check merge
1574   // 1 - always check merge
1575   // 2 - check merge and prune checking final split
1576   // 3 - check merge and prune checking final split based on bsize and qindex
1577   int nonrd_check_partition_merge_mode;
1578 
1579   // For nonrd_use_partition: check of leaf partition extra split
1580   int nonrd_check_partition_split;
1581 
1582   // Implements various heuristics to skip searching modes
1583   // The heuristics selected are based on  flags
1584   // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
1585   unsigned int mode_search_skip_flags;
1586 
1587   // For nonrd: Reduces ref frame search.
1588   // 0 - low level of search prune in non last frames
1589   // 1 - pruned search in non last frames
1590   // 2 - more pruned search in non last frames
1591   int nonrd_prune_ref_frame_search;
1592 
1593   // This flag controls the use of non-RD mode decision.
1594   int use_nonrd_pick_mode;
1595 
1596   // Use ALTREF frame in non-RD mode decision.
1597   int use_nonrd_altref_frame;
1598 
1599   // Use compound reference for non-RD mode.
1600   int use_comp_ref_nonrd;
1601 
1602   // Reference frames for compound prediction for nonrd pickmode:
1603   // LAST_GOLDEN (0), LAST_LAST2 (1), or LAST_ALTREF (2).
1604   int ref_frame_comp_nonrd[3];
1605 
1606   // use reduced ref set for real-time mode
1607   int use_real_time_ref_set;
1608 
1609   // Skip a number of expensive mode evaluations for blocks with very low
1610   // temporal variance.
1611   int short_circuit_low_temp_var;
1612 
1613   // Reuse inter prediction in fast non-rd mode.
1614   int reuse_inter_pred_nonrd;
1615 
1616   // Number of best inter modes to search transform. INT_MAX - search all.
1617   int num_inter_modes_for_tx_search;
1618 
1619   // Use interpolation filter search in non-RD mode decision.
1620   int use_nonrd_filter_search;
1621 
1622   // Use simplified RD model for interpolation search and Intra
1623   int use_simple_rd_model;
1624 
1625   // For nonrd mode: use hybrid intra mode search for intra only frames based on
1626   // block properties.
1627   // 0 : use nonrd pick intra for all blocks
1628   // 1 : use rd for bsize < 16x16, nonrd otherwise
1629   // 2 : use rd for bsize < 16x16 and src var >= 101, nonrd otherwise
1630   int hybrid_intra_pickmode;
1631 
1632   // Filter blocks by certain criteria such as SAD, source variance, such that
1633   // fewer blocks will go through the palette search.
1634   // For nonrd encoding path, enable this feature reduces encoding time when
1635   // palette mode is used. Disabling it leads to better compression efficiency.
1636   // 0: off
1637   // 1: less aggressive pruning mode
1638   // 2: more aggressive pruning mode
1639   int prune_palette_search_nonrd;
1640 
1641   // Compute variance/sse on source difference, prior to encoding superblock.
1642   int source_metrics_sb_nonrd;
1643 
1644   // Flag to indicate process for handling overshoot on slide/scene change,
1645   // for real-time CBR mode.
1646   OVERSHOOT_DETECTION_CBR overshoot_detection_cbr;
1647 
1648   // Check for scene/content change detection on every frame before encoding.
1649   int check_scene_detection;
1650 
1651   // For keyframes in rtc: adjust the rc_bits_per_mb, to reduce overshoot.
1652   int rc_adjust_keyframe;
1653 
1654   // On scene change: compute spatial variance.
1655   int rc_compute_spatial_var_sc;
1656 
1657   // For nonrd mode: Prefer larger partition blks in variance based partitioning
1658   // 0: disabled, 1-3: increasing aggressiveness
1659   int prefer_large_partition_blocks;
1660 
1661   // uses results of temporal noise estimate
1662   int use_temporal_noise_estimate;
1663 
1664   // Parameter indicating initial search window to be used in full-pixel search
1665   // for nonrd_pickmode. Range [0, MAX_MVSEARCH_STEPS - 1]. Lower value
1666   // indicates larger window. If set to 0, step_param is set based on internal
1667   // logic in set_mv_search_params().
1668   int fullpel_search_step_param;
1669 
1670   // Bit mask to enable or disable intra modes for each prediction block size
1671   // separately, for nonrd_pickmode.  Currently, the sf is not respected when
1672   // 'force_intra_check' is true in 'av1_estimate_intra_mode()' function. Also,
1673   // H and V pred modes allowed through this sf can be further pruned when
1674   //'prune_hv_pred_modes_using_src_sad' sf is true.
1675   int intra_y_mode_bsize_mask_nrd[BLOCK_SIZES];
1676 
1677   // Prune H and V intra predition modes evalution in inter frame.
1678   // The sf does not have any impact.
1679   // i. when frame_source_sad is 1.1 times greater than avg_source_sad
1680   // ii. when cyclic_refresh_segment_id_boosted is enabled
1681   // iii. when SB level source sad is greater than kMedSad
1682   // iv. when color sensitivity is non zero for both the chroma channels
1683   bool prune_hv_pred_modes_using_src_sad;
1684 
1685   // Skips mode checks more aggressively in nonRD mode
1686   int nonrd_aggressive_skip;
1687 
1688   // Skip cdef on 64x64 blocks/
1689   // 0: disabled
1690   // 1: skip when NEWMV or INTRA is not picked or color sensitivity is off.
1691   // When color sensitivity is on for a superblock, all 64x64 blocks within
1692   // will not skip.
1693   // 2: more aggressive mode where skip is done for all frames where
1694   // rc->high_source_sad = 0 (non slide-changes), and color sensitivity off.
1695   int skip_cdef_sb;
1696 
1697   // Force selective cdf update.
1698   int selective_cdf_update;
1699 
1700   // Force only single reference (LAST) for prediction.
1701   int force_only_last_ref;
1702 
1703   // Forces larger partition blocks in variance based partitioning for intra
1704   // frames
1705   int force_large_partition_blocks_intra;
1706 
1707   // Use fixed partition for superblocks based on source_sad.
1708   // 0: disabled
1709   // 1: enabled
1710   int use_fast_fixed_part;
1711 
1712   // Increase source_sad thresholds in nonrd pickmode.
1713   int increase_source_sad_thresh;
1714 
1715   // Skip evaluation of no split in tx size selection for merge partition
1716   int skip_tx_no_split_var_based_partition;
1717 
1718   // Intermediate termination of newMV mode evaluation based on so far best mode
1719   // sse
1720   int skip_newmv_mode_based_on_sse;
1721 
1722   // Define gf length multiplier.
1723   // Level 0: use large multiplier, level 1: use medium multiplier.
1724   int gf_length_lvl;
1725 
1726   // Prune inter modes with golden frame as reference for NEARMV and NEWMV modes
1727   int prune_inter_modes_with_golden_ref;
1728 
1729   // Prune inter modes w.r.t golden or alt-ref frame based on sad
1730   int prune_inter_modes_wrt_gf_arf_based_on_sad;
1731 
1732   // Prune inter mode search in rd path based on current block's temporal
1733   // variance wrt LAST reference.
1734   int prune_inter_modes_using_temp_var;
1735 
1736   // Reduce MV precision to halfpel for higher int MV value & frame-level motion
1737   // 0: disabled
1738   // 1-2: Reduce precision to halfpel, fullpel based on conservative
1739   // thresholds, aggressiveness increases with increase in level
1740   // 3: Reduce precision to halfpel using more aggressive thresholds
1741   int reduce_mv_pel_precision_highmotion;
1742 
1743   // Reduce MV precision for low complexity blocks
1744   // 0: disabled
1745   // 1: Reduce the mv resolution for zero mv if the variance is low
1746   // 2: Switch to halfpel, fullpel based on low block spatial-temporal
1747   // complexity.
1748   int reduce_mv_pel_precision_lowcomplex;
1749 
1750   // Prune intra mode evaluation in inter frames based on mv range.
1751   BLOCK_SIZE prune_intra_mode_based_on_mv_range;
1752   // The number of times to left shift the splitting thresholds in variance
1753   // based partitioning. The minimum values should be 7 to avoid left shifting
1754   // by a negative number.
1755   int var_part_split_threshold_shift;
1756 
1757   // Qindex based variance partition threshold index, which determines
1758   // the aggressiveness of partition pruning
1759   // 0: disabled for speeds 9,10
1760   // 1,2: (rd-path) lowers qindex thresholds conditionally (for low SAD sb)
1761   // 3,4: (non-rd path) uses pre-tuned qindex thresholds
1762   int var_part_based_on_qidx;
1763 
1764   // Enable GF refresh based on Q value.
1765   int gf_refresh_based_on_qp;
1766 
1767   // Temporal filtering
1768   // The value can be 1 or 2, which indicates the threshold to use.
1769   // Must be off for lossless mode.
1770   int use_rtc_tf;
1771 
1772   // Use of the identity transform in nonrd_pickmode,
1773   int use_idtx_nonrd;
1774 
1775   // Prune the use of the identity transform in nonrd_pickmode:
1776   // only for smaller blocks and higher spatial variance, and when skip_txfm
1777   // is not already set.
1778   int prune_idtx_nonrd;
1779 
1780   // Force to only use dct for palette search in nonrd pickmode.
1781   int dct_only_palette_nonrd;
1782 
1783   // Skip loopfilter, for static content after slide change
1784   // or key frame, once quality has ramped up.
1785   // 0: disabled
1786   // 1: skip only after quality is ramped up.
1787   // 2: aggrssive mode, where skip is done for all frames that
1788   // where rc->high_source_sad = 0 (no slide-changes).
1789   int skip_lf_screen;
1790 
1791   // Threshold on the active/inactive region percent to disable
1792   // the loopfilter and cdef. Setting to 100 disables this feature.
1793   int thresh_active_maps_skip_lf_cdef;
1794 
1795   // For nonrd: early exit out of variance partition that sets the
1796   // block size to superblock size, and sets mode to zeromv-last skip.
1797   // 0: disabled
1798   // 1: zeromv-skip is enabled at SB level only
1799   // 2: zeromv-skip is enabled at SB level and coding block level
1800   int part_early_exit_zeromv;
1801 
1802   // Early terminate inter mode search based on sse in non-rd path.
1803   INTER_SEARCH_EARLY_TERM_IDX sse_early_term_inter_search;
1804 
1805   // SAD based adaptive altref selection
1806   int sad_based_adp_altref_lag;
1807 
1808   // Enable/disable partition direct merging.
1809   int partition_direct_merging;
1810 
1811   // Level of aggressiveness for obtaining tx size based on qstep
1812   int tx_size_level_based_on_qstep;
1813 
1814   // Avoid the partitioning of a 16x16 block in variance based partitioning
1815   // (VBP) by making use of minimum and maximum sub-block variances.
1816   // For allintra encode, this speed feature reduces instruction count by 5.39%
1817   // for speed 9 on a typical video dataset with coding performance gain
1818   // of 1.44%.
1819   // For AVIF image encode, this speed feature reduces encode time
1820   // by 8.44% for speed 9 on a typical image dataset with coding performance
1821   // gain of 0.78%.
1822   bool vbp_prune_16x16_split_using_min_max_sub_blk_var;
1823 
1824   // A qindex threshold that determines whether to use qindex based CDEF filter
1825   // strength estimation for screen content types. The strength estimation model
1826   // used for screen contents prefers to allow cdef filtering for more frames.
1827   // This sf is used to limit the frames which go through cdef filtering and
1828   // following explains the setting of the same.
1829   // MAXQ (255): This disables the usage of this sf. Here, frame does not use a
1830   // screen content model thus reduces the number of frames that go through cdef
1831   // filtering.
1832   // MINQ (0): Frames always use screen content model thus increasing the number
1833   // of frames that go through cdef filtering.
1834   // This speed feature has a substantial gain on coding metrics, with moderate
1835   // increase encoding time. Select threshold based on speed vs quality
1836   // trade-off.
1837   int screen_content_cdef_filter_qindex_thresh;
1838 
1839   // Prune compound mode if its variance is higher than the variance of single
1840   // modes.
1841   bool prune_compoundmode_with_singlecompound_var;
1842 
1843   // Allow mode cost update at frame level every couple frames. This
1844   // overrides the command line setting --mode-cost-upd-freq=3 (never update
1845   // except on key frame and first delta).
1846   bool frame_level_mode_cost_update;
1847 
1848   // Prune H_PRED during intra mode evaluation in the nonrd path based on best
1849   // mode so far.
1850   //
1851   // For allintra encode, this speed feature reduces instruction count by 1.10%
1852   // for speed 9 with coding performance change less than 0.04%.
1853   // For AVIF image encode, this speed feature reduces encode time by 1.03% for
1854   // speed 9 on a typical image dataset with coding performance change less than
1855   // 0.08%.
1856   bool prune_h_pred_using_best_mode_so_far;
1857 
1858   // Enable pruning of intra mode evaluations in nonrd path based on source
1859   // variance and best mode so far. The pruning logic is enabled only if the
1860   // mode is not a winner mode of both the neighboring blocks (left/top).
1861   //
1862   // For allintra encode, this speed feature reduces instruction count by 3.96%
1863   // for speed 9 with coding performance change less than 0.38%.
1864   // For AVIF image encode, this speed feature reduces encode time by 3.46% for
1865   // speed 9 on a typical image dataset with coding performance change less than
1866   // -0.06%.
1867   bool enable_intra_mode_pruning_using_neighbors;
1868 
1869   // Prune intra mode evaluations in nonrd path based on best sad so far.
1870   //
1871   // For allintra encode, this speed feature reduces instruction count by 3.05%
1872   // for speed 9 with coding performance change less than 0.24%.
1873   // For AVIF image encode, this speed feature reduces encode time by 1.87% for
1874   // speed 9 on a typical image dataset with coding performance change less than
1875   // 0.16%.
1876   bool prune_intra_mode_using_best_sad_so_far;
1877 
1878   // If compound is enabled, and the current block size is \geq BLOCK_16X16,
1879   // limit the compound modes to GLOBAL_GLOBALMV. This does not apply to the
1880   // base layer of svc.
1881   bool check_only_zero_zeromv_on_large_blocks;
1882 
1883   // Allow for disabling cdf update for non reference frames in svc mode.
1884   bool disable_cdf_update_non_reference_frame;
1885 
1886   // Prune compound modes if the single modes variances do not perform well.
1887   bool prune_compoundmode_with_singlemode_var;
1888 
1889   // Skip searching all compound mode if the variance of single_mode residue is
1890   // sufficiently low.
1891   bool skip_compound_based_on_var;
1892 
1893   // Sets force_zeromv_skip based on the source sad available. Aggressiveness
1894   // increases with increase in the level set for speed feature.
1895   // 0: No setting
1896   // 1: If source sad is kZeroSad
1897   // 2: If source sad <= kVeryLowSad
1898   int set_zeromv_skip_based_on_source_sad;
1899 
1900   // Downgrades the block-level subpel motion search to
1901   // av1_find_best_sub_pixel_tree_pruned_more for higher QP and when fullpel
1902   // search performed well, zeromv has low sad or low source_var
1903   bool use_adaptive_subpel_search;
1904 
1905   // A flag used in RTC case to control frame_refs_short_signaling. Note that
1906   // the final decision is made in check_frame_refs_short_signaling(). The flag
1907   // can only be turned on when res < 360p and speed >= 9, in which case only
1908   // LAST and GOLDEN ref frames are used now.
1909   bool enable_ref_short_signaling;
1910 
1911   // A flag that controls if we check or bypass GLOBALMV in rtc single ref frame
1912   // case.
1913   bool check_globalmv_on_single_ref;
1914 
1915   // Allows for increasing the color_threshold for palette prediction.
1916   // This generally leads to better coding efficiency but with some speed loss.
1917   // Only used for screen content and for nonrd_pickmode.
1918   bool increase_color_thresh_palette;
1919 
1920   // Flag to indicate selecting of higher threshold for scenee change detection.
1921   int higher_thresh_scene_detection;
1922 
1923   // FLag to indicate skip testing of NEWMV for flat blocks.
1924   int skip_newmv_flat_blocks_screen;
1925 
1926   // Flag to force skip encoding for non_reference_frame on slide/scene changes.
1927   int skip_encoding_non_reference_slide_change;
1928 
1929   // Flag to indicate more aggressive QP downward adjustment for screen static
1930   // content, to make convergence to min_qp faster.
1931   int rc_faster_convergence_static;
1932 } REAL_TIME_SPEED_FEATURES;
1933 
1934 /*!\endcond */
1935 
1936 /*!
1937  * \brief Top level speed vs quality trade off data struture.
1938  */
1939 typedef struct SPEED_FEATURES {
1940   /*!
1941    * Sequence/frame level speed features:
1942    */
1943   HIGH_LEVEL_SPEED_FEATURES hl_sf;
1944 
1945   /*!
1946    * Speed features for the first pass.
1947    */
1948   FIRST_PASS_SPEED_FEATURES fp_sf;
1949 
1950   /*!
1951    * Speed features related to how tpl's searches are done.
1952    */
1953   TPL_SPEED_FEATURES tpl_sf;
1954 
1955   /*!
1956    * Global motion speed features:
1957    */
1958   GLOBAL_MOTION_SPEED_FEATURES gm_sf;
1959 
1960   /*!
1961    * Partition search speed features:
1962    */
1963   PARTITION_SPEED_FEATURES part_sf;
1964 
1965   /*!
1966    * Motion search speed features:
1967    */
1968   MV_SPEED_FEATURES mv_sf;
1969 
1970   /*!
1971    * Inter mode search speed features:
1972    */
1973   INTER_MODE_SPEED_FEATURES inter_sf;
1974 
1975   /*!
1976    * Interpolation filter search speed features:
1977    */
1978   INTERP_FILTER_SPEED_FEATURES interp_sf;
1979 
1980   /*!
1981    * Intra mode search speed features:
1982    */
1983   INTRA_MODE_SPEED_FEATURES intra_sf;
1984 
1985   /*!
1986    * Transform size/type search speed features:
1987    */
1988   TX_SPEED_FEATURES tx_sf;
1989 
1990   /*!
1991    * RD calculation speed features:
1992    */
1993   RD_CALC_SPEED_FEATURES rd_sf;
1994 
1995   /*!
1996    * Two-pass mode evaluation features:
1997    */
1998   WINNER_MODE_SPEED_FEATURES winner_mode_sf;
1999 
2000   /*!
2001    * In-loop filter speed features:
2002    */
2003   LOOP_FILTER_SPEED_FEATURES lpf_sf;
2004 
2005   /*!
2006    * Real-time mode speed features:
2007    */
2008   REAL_TIME_SPEED_FEATURES rt_sf;
2009 } SPEED_FEATURES;
2010 /*!\cond */
2011 
2012 struct AV1_COMP;
2013 
2014 /*!\endcond */
2015 /*!\brief Frame size independent speed vs quality trade off flags
2016  *
2017  *\ingroup speed_features
2018  *
2019  * \param[in]    cpi     Top - level encoder instance structure
2020  * \param[in]    speed   Speed setting passed in from the command  line
2021  *
2022  * \remark No return value but configures the various speed trade off flags
2023  *         based on the passed in speed setting. (Higher speed gives lower
2024  *         quality)
2025  */
2026 void av1_set_speed_features_framesize_independent(struct AV1_COMP *cpi,
2027                                                   int speed);
2028 
2029 /*!\brief Frame size dependent speed vs quality trade off flags
2030  *
2031  *\ingroup speed_features
2032  *
2033  * \param[in]    cpi     Top - level encoder instance structure
2034  * \param[in]    speed   Speed setting passed in from the command  line
2035  *
2036  * \remark No return value but configures the various speed trade off flags
2037  *         based on the passed in speed setting and frame size. (Higher speed
2038  *         corresponds to lower quality)
2039  */
2040 void av1_set_speed_features_framesize_dependent(struct AV1_COMP *cpi,
2041                                                 int speed);
2042 /*!\brief Q index dependent speed vs quality trade off flags
2043  *
2044  *\ingroup speed_features
2045  *
2046  * \param[in]    cpi     Top - level encoder instance structure
2047  * \param[in]    speed   Speed setting passed in from the command  line
2048  *
2049  * \remark No return value but configures the various speed trade off flags
2050  *         based on the passed in speed setting and current frame's Q index.
2051  *         (Higher speed corresponds to lower quality)
2052  */
2053 void av1_set_speed_features_qindex_dependent(struct AV1_COMP *cpi, int speed);
2054 
2055 #ifdef __cplusplus
2056 }  // extern "C"
2057 #endif
2058 
2059 #endif  // AOM_AV1_ENCODER_SPEED_FEATURES_H_
2060