• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
3  *
4  *  Use of this source code is governed by a BSD-style license
5  *  that can be found in the LICENSE file in the root of the source
6  *  tree. An additional intellectual property rights grant can be found
7  *  in the file PATENTS.  All contributing project authors may
8  *  be found in the AUTHORS file in the root of the source tree.
9  */
10 
11 #ifndef VP9_ENCODER_VP9_SPEED_FEATURES_H_
12 #define VP9_ENCODER_VP9_SPEED_FEATURES_H_
13 
14 #include "vp9/common/vp9_enums.h"
15 
16 #ifdef __cplusplus
17 extern "C" {
18 #endif
19 
20 enum {
21   INTRA_ALL = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED) | (1 << D45_PRED) |
22               (1 << D135_PRED) | (1 << D117_PRED) | (1 << D153_PRED) |
23               (1 << D207_PRED) | (1 << D63_PRED) | (1 << TM_PRED),
24   INTRA_DC = (1 << DC_PRED),
25   INTRA_DC_TM = (1 << DC_PRED) | (1 << TM_PRED),
26   INTRA_DC_H_V = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED),
27   INTRA_DC_TM_H_V =
28       (1 << DC_PRED) | (1 << TM_PRED) | (1 << V_PRED) | (1 << H_PRED)
29 };
30 
31 enum {
32   INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV) | (1 << NEWMV),
33   INTER_NEAREST = (1 << NEARESTMV),
34   INTER_NEAREST_NEW = (1 << NEARESTMV) | (1 << NEWMV),
35   INTER_NEAREST_ZERO = (1 << NEARESTMV) | (1 << ZEROMV),
36   INTER_NEAREST_NEW_ZERO = (1 << NEARESTMV) | (1 << ZEROMV) | (1 << NEWMV),
37   INTER_NEAREST_NEAR_NEW = (1 << NEARESTMV) | (1 << NEARMV) | (1 << NEWMV),
38   INTER_NEAREST_NEAR_ZERO = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV),
39 };
40 
41 enum {
42   DISABLE_ALL_INTER_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
43                             (1 << THR_ALTR) | (1 << THR_GOLD) | (1 << THR_LAST),
44 
45   DISABLE_ALL_SPLIT = (1 << THR_INTRA) | DISABLE_ALL_INTER_SPLIT,
46 
47   DISABLE_COMPOUND_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA),
48 
49   LAST_AND_INTRA_SPLIT_ONLY = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
50                               (1 << THR_ALTR) | (1 << THR_GOLD)
51 };
52 
53 typedef enum {
54   DIAMOND = 0,
55   NSTEP = 1,
56   HEX = 2,
57   BIGDIA = 3,
58   SQUARE = 4,
59   FAST_HEX = 5,
60   FAST_DIAMOND = 6
61 } SEARCH_METHODS;
62 
63 typedef enum {
64   // No recode.
65   DISALLOW_RECODE = 0,
66   // Allow recode for KF and exceeding maximum frame bandwidth.
67   ALLOW_RECODE_KFMAXBW = 1,
68   // Allow recode only for KF/ARF/GF frames.
69   ALLOW_RECODE_KFARFGF = 2,
70   // Allow recode for ARF/GF/KF and first normal frame in each group.
71   ALLOW_RECODE_FIRST = 3,
72   // Allow recode for all frames based on bitrate constraints.
73   ALLOW_RECODE = 4,
74 } RECODE_LOOP_TYPE;
75 
76 typedef enum {
77   SUBPEL_TREE = 0,
78   SUBPEL_TREE_PRUNED = 1,           // Prunes 1/2-pel searches
79   SUBPEL_TREE_PRUNED_MORE = 2,      // Prunes 1/2-pel searches more aggressively
80   SUBPEL_TREE_PRUNED_EVENMORE = 3,  // Prunes 1/2- and 1/4-pel searches
81   // Other methods to come
82 } SUBPEL_SEARCH_METHODS;
83 
84 typedef enum {
85   NO_MOTION_THRESHOLD = 0,
86   LOW_MOTION_THRESHOLD = 7
87 } MOTION_THRESHOLD;
88 
89 typedef enum {
90   USE_FULL_RD = 0,
91   USE_LARGESTALL,
92   USE_TX_8X8
93 } TX_SIZE_SEARCH_METHOD;
94 
95 typedef enum {
96   NOT_IN_USE = 0,
97   RELAXED_NEIGHBORING_MIN_MAX = 1,
98   STRICT_NEIGHBORING_MIN_MAX = 2
99 } AUTO_MIN_MAX_MODE;
100 
101 typedef enum {
102   // Try the full image with different values.
103   LPF_PICK_FROM_FULL_IMAGE,
104   // Try a small portion of the image with different values.
105   LPF_PICK_FROM_SUBIMAGE,
106   // Estimate the level based on quantizer and frame type
107   LPF_PICK_FROM_Q,
108   // Pick 0 to disable LPF if LPF was enabled last frame
109   LPF_PICK_MINIMAL_LPF
110 } LPF_PICK_METHOD;
111 
112 typedef enum {
113   // Terminate search early based on distortion so far compared to
114   // qp step, distortion in the neighborhood of the frame, etc.
115   FLAG_EARLY_TERMINATE = 1 << 0,
116 
117   // Skips comp inter modes if the best so far is an intra mode.
118   FLAG_SKIP_COMP_BESTINTRA = 1 << 1,
119 
120   // Skips oblique intra modes if the best so far is an inter mode.
121   FLAG_SKIP_INTRA_BESTINTER = 1 << 3,
122 
123   // Skips oblique intra modes  at angles 27, 63, 117, 153 if the best
124   // intra so far is not one of the neighboring directions.
125   FLAG_SKIP_INTRA_DIRMISMATCH = 1 << 4,
126 
127   // Skips intra modes other than DC_PRED if the source variance is small
128   FLAG_SKIP_INTRA_LOWVAR = 1 << 5,
129 } MODE_SEARCH_SKIP_LOGIC;
130 
131 typedef enum {
132   FLAG_SKIP_EIGHTTAP = 1 << EIGHTTAP,
133   FLAG_SKIP_EIGHTTAP_SMOOTH = 1 << EIGHTTAP_SMOOTH,
134   FLAG_SKIP_EIGHTTAP_SHARP = 1 << EIGHTTAP_SHARP,
135 } INTERP_FILTER_MASK;
136 
137 typedef enum {
138   // Search partitions using RD/NONRD criterion
139   SEARCH_PARTITION,
140 
141   // Always use a fixed size partition
142   FIXED_PARTITION,
143 
144   REFERENCE_PARTITION,
145 
146   // Use an arbitrary partitioning scheme based on source variance within
147   // a 64X64 SB
148   VAR_BASED_PARTITION,
149 
150   // Use non-fixed partitions based on source variance
151   SOURCE_VAR_BASED_PARTITION
152 } PARTITION_SEARCH_TYPE;
153 
154 typedef enum {
155   // Does a dry run to see if any of the contexts need to be updated or not,
156   // before the final run.
157   TWO_LOOP = 0,
158 
159   // No dry run, also only half the coef contexts and bands are updated.
160   // The rest are not updated at all.
161   ONE_LOOP_REDUCED = 1
162 } FAST_COEFF_UPDATE;
163 
164 typedef struct MV_SPEED_FEATURES {
165   // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
166   SEARCH_METHODS search_method;
167 
168   // This parameter controls which step in the n-step process we start at.
169   // It's changed adaptively based on circumstances.
170   int reduce_first_step_size;
171 
172   // If this is set to 1, we limit the motion search range to 2 times the
173   // largest motion vector found in the last frame.
174   int auto_mv_step_size;
175 
176   // Subpel_search_method can only be subpel_tree which does a subpixel
177   // logarithmic search that keeps stepping at 1/2 pixel units until
178   // you stop getting a gain, and then goes on to 1/4 and repeats
179   // the same process. Along the way it skips many diagonals.
180   SUBPEL_SEARCH_METHODS subpel_search_method;
181 
182   // Maximum number of steps in logarithmic subpel search before giving up.
183   int subpel_iters_per_step;
184 
185   // Control when to stop subpel search:
186   // 0: Full subpel search.
187   // 1: Stop at quarter pixel.
188   // 2: Stop at half pixel.
189   // 3: Stop at full pixel.
190   int subpel_force_stop;
191 
192   // This variable sets the step_param used in full pel motion search.
193   int fullpel_search_step_param;
194 } MV_SPEED_FEATURES;
195 
196 typedef struct PARTITION_SEARCH_BREAKOUT_THR {
197   int64_t dist;
198   int rate;
199 } PARTITION_SEARCH_BREAKOUT_THR;
200 
201 #define MAX_MESH_STEP 4
202 
203 typedef struct MESH_PATTERN {
204   int range;
205   int interval;
206 } MESH_PATTERN;
207 
208 typedef struct SPEED_FEATURES {
209   MV_SPEED_FEATURES mv;
210 
211   // Frame level coding parameter update
212   int frame_parameter_update;
213 
214   RECODE_LOOP_TYPE recode_loop;
215 
216   // Trellis (dynamic programming) optimization of quantized values (+1, 0).
217   int optimize_coefficients;
218 
219   // Always set to 0. If on it enables 0 cost background transmission
220   // (except for the initial transmission of the segmentation). The feature is
221   // disabled because the addition of very large block sizes make the
222   // backgrounds very to cheap to encode, and the segmentation we have
223   // adds overhead.
224   int static_segmentation;
225 
226   // If 1 we iterate finding a best reference for 2 ref frames together - via
227   // a log search that iterates 4 times (check around mv for last for best
228   // error of combined predictor then check around mv for alt). If 0 we
229   // we just use the best motion vector found for each frame by itself.
230   BLOCK_SIZE comp_inter_joint_search_thresh;
231 
232   // This variable is used to cap the maximum number of times we skip testing a
233   // mode to be evaluated. A high value means we will be faster.
234   // Turned off when (row_mt_bit_exact == 1 && adaptive_rd_thresh_row_mt == 0).
235   int adaptive_rd_thresh;
236 
237   // Flag to use adaptive_rd_thresh when row-mt it enabled, only for non-rd
238   // pickmode.
239   int adaptive_rd_thresh_row_mt;
240 
241   // Enables skipping the reconstruction step (idct, recon) in the
242   // intermediate steps assuming the last frame didn't have too many intra
243   // blocks and the q is less than a threshold.
244   int skip_encode_sb;
245   int skip_encode_frame;
246   // Speed feature to allow or disallow skipping of recode at block
247   // level within a frame.
248   int allow_skip_recode;
249 
250   // Coefficient probability model approximation step size
251   int coeff_prob_appx_step;
252 
253   // Enable uniform quantizer followed by trellis coefficient optimization
254   int allow_quant_coeff_opt;
255   double quant_opt_thresh;
256 
257   // Enable asymptotic closed-loop encoding decision for key frame and
258   // alternate reference frames.
259   int allow_acl;
260 
261   // Use transform domain distortion. Use pixel domain distortion in speed 0
262   // and certain situations in higher speed to improve the RD model precision.
263   int allow_txfm_domain_distortion;
264   double tx_domain_thresh;
265 
266   // The threshold is to determine how slow the motino is, it is used when
267   // use_lastframe_partitioning is set to LAST_FRAME_PARTITION_LOW_MOTION
268   MOTION_THRESHOLD lf_motion_threshold;
269 
270   // Determine which method we use to determine transform size. We can choose
271   // between options like full rd, largest for prediction size, largest
272   // for intra and model coefs for the rest.
273   TX_SIZE_SEARCH_METHOD tx_size_search_method;
274 
275   // Low precision 32x32 fdct keeps everything in 16 bits and thus is less
276   // precise but significantly faster than the non lp version.
277   int use_lp32x32fdct;
278 
279   // After looking at the first set of modes (set by index here), skip
280   // checking modes for reference frames that don't match the reference frame
281   // of the best so far.
282   int mode_skip_start;
283 
284   // TODO(JBB): Remove this.
285   int reference_masking;
286 
287   PARTITION_SEARCH_TYPE partition_search_type;
288 
289   // Used if partition_search_type = FIXED_SIZE_PARTITION
290   BLOCK_SIZE always_this_block_size;
291 
292   // Skip rectangular partition test when partition type none gives better
293   // rd than partition type split.
294   int less_rectangular_check;
295 
296   // Disable testing non square partitions. (eg 16x32)
297   int use_square_partition_only;
298   BLOCK_SIZE use_square_only_threshold;
299 
300   // Sets min and max partition sizes for this 64x64 region based on the
301   // same 64x64 in last encoded frame, and the left and above neighbor.
302   AUTO_MIN_MAX_MODE auto_min_max_partition_size;
303   // Ensures the rd based auto partition search will always
304   // go down at least to the specified level.
305   BLOCK_SIZE rd_auto_partition_min_limit;
306 
307   // Min and max partition size we enable (block_size) as per auto
308   // min max, but also used by adjust partitioning, and pick_partitioning.
309   BLOCK_SIZE default_min_partition_size;
310   BLOCK_SIZE default_max_partition_size;
311 
312   // Whether or not we allow partitions one smaller or one greater than the last
313   // frame's partitioning. Only used if use_lastframe_partitioning is set.
314   int adjust_partitioning_from_last_frame;
315 
316   // How frequently we re do the partitioning from scratch. Only used if
317   // use_lastframe_partitioning is set.
318   int last_partitioning_redo_frequency;
319 
320   // Disables sub 8x8 blocksizes in different scenarios: Choices are to disable
321   // it always, to allow it for only Last frame and Intra, disable it for all
322   // inter modes or to enable it always.
323   int disable_split_mask;
324 
325   // TODO(jingning): combine the related motion search speed features
326   // This allows us to use motion search at other sizes as a starting
327   // point for this motion search and limits the search range around it.
328   int adaptive_motion_search;
329 
330   // Threshold for allowing exhaistive motion search.
331   int exhaustive_searches_thresh;
332 
333   // Pattern to be used for any exhaustive mesh searches.
334   MESH_PATTERN mesh_patterns[MAX_MESH_STEP];
335 
336   int schedule_mode_search;
337 
338   // Allows sub 8x8 modes to use the prediction filter that was determined
339   // best for 8x8 mode. If set to 0 we always re check all the filters for
340   // sizes less than 8x8, 1 means we check all filter modes if no 8x8 filter
341   // was selected, and 2 means we use 8 tap if no 8x8 filter mode was selected.
342   int adaptive_pred_interp_filter;
343 
344   // Adaptive prediction mode search
345   int adaptive_mode_search;
346 
347   // Chessboard pattern prediction filter type search
348   int cb_pred_filter_search;
349 
350   int cb_partition_search;
351 
352   int motion_field_mode_search;
353 
354   int alt_ref_search_fp;
355 
356   // Fast quantization process path
357   int use_quant_fp;
358 
359   // Use finer quantizer in every other few frames that run variable block
360   // partition type search.
361   int force_frame_boost;
362 
363   // Maximally allowed base quantization index fluctuation.
364   int max_delta_qindex;
365 
366   // Implements various heuristics to skip searching modes
367   // The heuristics selected are based on  flags
368   // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
369   unsigned int mode_search_skip_flags;
370 
371   // A source variance threshold below which filter search is disabled
372   // Choose a very large value (UINT_MAX) to use 8-tap always
373   unsigned int disable_filter_search_var_thresh;
374 
375   // These bit masks allow you to enable or disable intra modes for each
376   // transform size separately.
377   int intra_y_mode_mask[TX_SIZES];
378   int intra_uv_mode_mask[TX_SIZES];
379 
380   // These bit masks allow you to enable or disable intra modes for each
381   // prediction block size separately.
382   int intra_y_mode_bsize_mask[BLOCK_SIZES];
383 
384   // This variable enables an early break out of mode testing if the model for
385   // rd built from the prediction signal indicates a value that's much
386   // higher than the best rd we've seen so far.
387   int use_rd_breakout;
388 
389   // This enables us to use an estimate for intra rd based on dc mode rather
390   // than choosing an actual uv mode in the stage of encoding before the actual
391   // final encode.
392   int use_uv_intra_rd_estimate;
393 
394   // This feature controls how the loop filter level is determined.
395   LPF_PICK_METHOD lpf_pick;
396 
397   // This feature limits the number of coefficients updates we actually do
398   // by only looking at counts from 1/2 the bands.
399   FAST_COEFF_UPDATE use_fast_coef_updates;
400 
401   // This flag controls the use of non-RD mode decision.
402   int use_nonrd_pick_mode;
403 
404   // A binary mask indicating if NEARESTMV, NEARMV, ZEROMV, NEWMV
405   // modes are used in order from LSB to MSB for each BLOCK_SIZE.
406   int inter_mode_mask[BLOCK_SIZES];
407 
408   // This feature controls whether we do the expensive context update and
409   // calculation in the rd coefficient costing loop.
410   int use_fast_coef_costing;
411 
412   // This feature controls the tolerence vs target used in deciding whether to
413   // recode a frame. It has no meaning if recode is disabled.
414   int recode_tolerance_low;
415   int recode_tolerance_high;
416 
417   // This variable controls the maximum block size where intra blocks can be
418   // used in inter frames.
419   // TODO(aconverse): Fold this into one of the other many mode skips
420   BLOCK_SIZE max_intra_bsize;
421 
422   // The frequency that we check if SOURCE_VAR_BASED_PARTITION or
423   // FIXED_PARTITION search type should be used.
424   int search_type_check_frequency;
425 
426   // When partition is pre-set, the inter prediction result from pick_inter_mode
427   // can be reused in final block encoding process. It is enabled only for real-
428   // time mode speed 6.
429   int reuse_inter_pred_sby;
430 
431   // This variable sets the encode_breakout threshold. Currently, it is only
432   // enabled in real time mode.
433   int encode_breakout_thresh;
434 
435   // default interp filter choice
436   INTERP_FILTER default_interp_filter;
437 
438   // Early termination in transform size search, which only applies while
439   // tx_size_search_method is USE_FULL_RD.
440   int tx_size_search_breakout;
441 
442   // adaptive interp_filter search to allow skip of certain filter types.
443   int adaptive_interp_filter_search;
444 
445   // mask for skip evaluation of certain interp_filter type.
446   INTERP_FILTER_MASK interp_filter_search_mask;
447 
448   // Partition search early breakout thresholds.
449   PARTITION_SEARCH_BREAKOUT_THR partition_search_breakout_thr;
450 
451   // Machine-learning based partition search early termination
452   int ml_partition_search_early_termination;
453 
454   // Allow skipping partition search for still image frame
455   int allow_partition_search_skip;
456 
457   // Fast approximation of vp9_model_rd_from_var_lapndz
458   int simple_model_rd_from_var;
459 
460   // Skip a number of expensive mode evaluations for blocks with zero source
461   // variance.
462   int short_circuit_flat_blocks;
463 
464   // Skip a number of expensive mode evaluations for blocks with very low
465   // temporal variance. If the low temporal variance flag is set for a block,
466   // do the following:
467   // 1: Skip all golden modes and ALL INTRA for bsize >= 32x32.
468   // 2: Skip golden non-zeromv and newmv-last for bsize >= 16x16, skip ALL
469   // INTRA for bsize >= 32x32 and vert/horz INTRA for bsize 16x16, 16x32 and
470   // 32x16.
471   // 3: Same as (2), but also skip golden zeromv.
472   int short_circuit_low_temp_var;
473 
474   // Limits the rd-threshold update for early exit for the newmv-last mode,
475   // for non-rd mode.
476   int limit_newmv_early_exit;
477 
478   // Adds a bias against golden reference, for non-rd mode.
479   int bias_golden;
480 
481   // Bias to use base mv and skip 1/4 subpel search when use base mv in
482   // enhancement layer.
483   int base_mv_aggressive;
484 
485   // Global flag to enable partition copy from the previous frame.
486   int copy_partition_flag;
487 
488   // Compute the source sad for every superblock of the frame,
489   // prior to encoding the frame, to be used to bypass some encoder decisions.
490   int use_source_sad;
491 
492   int use_simple_block_yrd;
493 } SPEED_FEATURES;
494 
495 struct VP9_COMP;
496 
497 void vp9_set_speed_features_framesize_independent(struct VP9_COMP *cpi);
498 void vp9_set_speed_features_framesize_dependent(struct VP9_COMP *cpi);
499 
500 #ifdef __cplusplus
501 }  // extern "C"
502 #endif
503 
504 #endif  // VP9_ENCODER_VP9_SPEED_FEATURES_H_
505