2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
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.
11 #ifndef VP9_ENCODER_VP9_SPEED_FEATURES_H_
12 #define VP9_ENCODER_VP9_SPEED_FEATURES_H_
14 #include "vp9/common/vp9_enums.h"
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),
28 (1 << DC_PRED) | (1 << TM_PRED) | (1 << V_PRED) | (1 << H_PRED)
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),
42 DISABLE_ALL_INTER_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
43 (1 << THR_ALTR) | (1 << THR_GOLD) | (1 << THR_LAST),
45 DISABLE_ALL_SPLIT = (1 << THR_INTRA) | DISABLE_ALL_INTER_SPLIT,
47 DISABLE_COMPOUND_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA),
49 LAST_AND_INTRA_SPLIT_ONLY = (1 << THR_COMP_GA) | (1 << THR_COMP_LA) |
50 (1 << THR_ALTR) | (1 << THR_GOLD)
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.
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;
85 NO_MOTION_THRESHOLD = 0,
86 LOW_MOTION_THRESHOLD = 7
93 } TX_SIZE_SEARCH_METHOD;
97 RELAXED_NEIGHBORING_MIN_MAX = 1,
98 STRICT_NEIGHBORING_MIN_MAX = 2
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
108 // Pick 0 to disable LPF if LPF was enabled last frame
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,
117 // Skips comp inter modes if the best so far is an intra mode.
118 FLAG_SKIP_COMP_BESTINTRA = 1 << 1,
120 // Skips oblique intra modes if the best so far is an inter mode.
121 FLAG_SKIP_INTRA_BESTINTER = 1 << 3,
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,
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;
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;
138 // Search partitions using RD/NONRD criterion
141 // Always use a fixed size partition
146 // Use an arbitrary partitioning scheme based on source variance within
150 // Use non-fixed partitions based on source variance
151 SOURCE_VAR_BASED_PARTITION
152 } PARTITION_SEARCH_TYPE;
155 // Does a dry run to see if any of the contexts need to be updated or not,
156 // before the final run.
159 // No dry run, also only half the coef contexts and bands are updated.
160 // The rest are not updated at all.
164 typedef struct MV_SPEED_FEATURES {
165 // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
166 SEARCH_METHODS search_method;
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;
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;
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;
182 // Maximum number of steps in logarithmic subpel search before giving up.
183 int subpel_iters_per_step;
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;
192 // This variable sets the step_param used in full pel motion search.
193 int fullpel_search_step_param;
196 typedef struct PARTITION_SEARCH_BREAKOUT_THR {
199 } PARTITION_SEARCH_BREAKOUT_THR;
201 #define MAX_MESH_STEP 4
203 typedef struct MESH_PATTERN {
208 typedef struct SPEED_FEATURES {
209 MV_SPEED_FEATURES mv;
211 // Frame level coding parameter update
212 int frame_parameter_update;
214 RECODE_LOOP_TYPE recode_loop;
216 // Trellis (dynamic programming) optimization of quantized values (+1, 0).
217 int optimize_coefficients;
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
224 int static_segmentation;
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;
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 int adaptive_rd_thresh;
236 // Flag to use adaptive_rd_thresh when row-mt it enabled.
237 int adaptive_rd_thresh_row_mt;
239 // Enables skipping the reconstruction step (idct, recon) in the
240 // intermediate steps assuming the last frame didn't have too many intra
241 // blocks and the q is less than a threshold.
243 int skip_encode_frame;
244 // Speed feature to allow or disallow skipping of recode at block
245 // level within a frame.
246 int allow_skip_recode;
248 // Coefficient probability model approximation step size
249 int coeff_prob_appx_step;
251 // Enable uniform quantizer followed by trellis coefficient optimization
252 int allow_quant_coeff_opt;
253 double quant_opt_thresh;
255 // Enable asymptotic closed-loop encoding decision for key frame and
256 // alternate reference frames.
259 // Use transform domain distortion. Use pixel domain distortion in speed 0
260 // and certain situations in higher speed to improve the RD model precision.
261 int allow_txfm_domain_distortion;
262 double tx_domain_thresh;
264 // The threshold is to determine how slow the motino is, it is used when
265 // use_lastframe_partitioning is set to LAST_FRAME_PARTITION_LOW_MOTION
266 MOTION_THRESHOLD lf_motion_threshold;
268 // Determine which method we use to determine transform size. We can choose
269 // between options like full rd, largest for prediction size, largest
270 // for intra and model coefs for the rest.
271 TX_SIZE_SEARCH_METHOD tx_size_search_method;
273 // Low precision 32x32 fdct keeps everything in 16 bits and thus is less
274 // precise but significantly faster than the non lp version.
277 // After looking at the first set of modes (set by index here), skip
278 // checking modes for reference frames that don't match the reference frame
279 // of the best so far.
282 // TODO(JBB): Remove this.
283 int reference_masking;
285 PARTITION_SEARCH_TYPE partition_search_type;
287 // Used if partition_search_type = FIXED_SIZE_PARTITION
288 BLOCK_SIZE always_this_block_size;
290 // Skip rectangular partition test when partition type none gives better
291 // rd than partition type split.
292 int less_rectangular_check;
294 // Disable testing non square partitions. (eg 16x32)
295 int use_square_partition_only;
296 BLOCK_SIZE use_square_only_threshold;
298 // Sets min and max partition sizes for this 64x64 region based on the
299 // same 64x64 in last encoded frame, and the left and above neighbor.
300 AUTO_MIN_MAX_MODE auto_min_max_partition_size;
301 // Ensures the rd based auto partition search will always
302 // go down at least to the specified level.
303 BLOCK_SIZE rd_auto_partition_min_limit;
305 // Min and max partition size we enable (block_size) as per auto
306 // min max, but also used by adjust partitioning, and pick_partitioning.
307 BLOCK_SIZE default_min_partition_size;
308 BLOCK_SIZE default_max_partition_size;
310 // Whether or not we allow partitions one smaller or one greater than the last
311 // frame's partitioning. Only used if use_lastframe_partitioning is set.
312 int adjust_partitioning_from_last_frame;
314 // How frequently we re do the partitioning from scratch. Only used if
315 // use_lastframe_partitioning is set.
316 int last_partitioning_redo_frequency;
318 // Disables sub 8x8 blocksizes in different scenarios: Choices are to disable
319 // it always, to allow it for only Last frame and Intra, disable it for all
320 // inter modes or to enable it always.
321 int disable_split_mask;
323 // TODO(jingning): combine the related motion search speed features
324 // This allows us to use motion search at other sizes as a starting
325 // point for this motion search and limits the search range around it.
326 int adaptive_motion_search;
328 // Flag for allowing some use of exhaustive searches;
329 int allow_exhaustive_searches;
331 // Threshold for allowing exhaistive motion search.
332 int exhaustive_searches_thresh;
334 // Maximum number of exhaustive searches for a frame.
335 int max_exaustive_pct;
337 // Pattern to be used for any exhaustive mesh searches.
338 MESH_PATTERN mesh_patterns[MAX_MESH_STEP];
340 int schedule_mode_search;
342 // Allows sub 8x8 modes to use the prediction filter that was determined
343 // best for 8x8 mode. If set to 0 we always re check all the filters for
344 // sizes less than 8x8, 1 means we check all filter modes if no 8x8 filter
345 // was selected, and 2 means we use 8 tap if no 8x8 filter mode was selected.
346 int adaptive_pred_interp_filter;
348 // Adaptive prediction mode search
349 int adaptive_mode_search;
351 // Chessboard pattern prediction filter type search
352 int cb_pred_filter_search;
354 int cb_partition_search;
356 int motion_field_mode_search;
358 int alt_ref_search_fp;
360 // Fast quantization process path
363 // Use finer quantizer in every other few frames that run variable block
364 // partition type search.
365 int force_frame_boost;
367 // Maximally allowed base quantization index fluctuation.
368 int max_delta_qindex;
370 // Implements various heuristics to skip searching modes
371 // The heuristics selected are based on flags
372 // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
373 unsigned int mode_search_skip_flags;
375 // A source variance threshold below which filter search is disabled
376 // Choose a very large value (UINT_MAX) to use 8-tap always
377 unsigned int disable_filter_search_var_thresh;
379 // These bit masks allow you to enable or disable intra modes for each
380 // transform size separately.
381 int intra_y_mode_mask[TX_SIZES];
382 int intra_uv_mode_mask[TX_SIZES];
384 // These bit masks allow you to enable or disable intra modes for each
385 // prediction block size separately.
386 int intra_y_mode_bsize_mask[BLOCK_SIZES];
388 // This variable enables an early break out of mode testing if the model for
389 // rd built from the prediction signal indicates a value that's much
390 // higher than the best rd we've seen so far.
393 // This enables us to use an estimate for intra rd based on dc mode rather
394 // than choosing an actual uv mode in the stage of encoding before the actual
396 int use_uv_intra_rd_estimate;
398 // This feature controls how the loop filter level is determined.
399 LPF_PICK_METHOD lpf_pick;
401 // This feature limits the number of coefficients updates we actually do
402 // by only looking at counts from 1/2 the bands.
403 FAST_COEFF_UPDATE use_fast_coef_updates;
405 // This flag controls the use of non-RD mode decision.
406 int use_nonrd_pick_mode;
408 // A binary mask indicating if NEARESTMV, NEARMV, ZEROMV, NEWMV
409 // modes are used in order from LSB to MSB for each BLOCK_SIZE.
410 int inter_mode_mask[BLOCK_SIZES];
412 // This feature controls whether we do the expensive context update and
413 // calculation in the rd coefficient costing loop.
414 int use_fast_coef_costing;
416 // This feature controls the tolerence vs target used in deciding whether to
417 // recode a frame. It has no meaning if recode is disabled.
418 int recode_tolerance_low;
419 int recode_tolerance_high;
421 // This variable controls the maximum block size where intra blocks can be
422 // used in inter frames.
423 // TODO(aconverse): Fold this into one of the other many mode skips
424 BLOCK_SIZE max_intra_bsize;
426 // The frequency that we check if SOURCE_VAR_BASED_PARTITION or
427 // FIXED_PARTITION search type should be used.
428 int search_type_check_frequency;
430 // When partition is pre-set, the inter prediction result from pick_inter_mode
431 // can be reused in final block encoding process. It is enabled only for real-
432 // time mode speed 6.
433 int reuse_inter_pred_sby;
435 // This variable sets the encode_breakout threshold. Currently, it is only
436 // enabled in real time mode.
437 int encode_breakout_thresh;
439 // default interp filter choice
440 INTERP_FILTER default_interp_filter;
442 // Early termination in transform size search, which only applies while
443 // tx_size_search_method is USE_FULL_RD.
444 int tx_size_search_breakout;
446 // adaptive interp_filter search to allow skip of certain filter types.
447 int adaptive_interp_filter_search;
449 // mask for skip evaluation of certain interp_filter type.
450 INTERP_FILTER_MASK interp_filter_search_mask;
452 // Partition search early breakout thresholds.
453 PARTITION_SEARCH_BREAKOUT_THR partition_search_breakout_thr;
455 // Machine-learning based partition search early termination
456 int ml_partition_search_early_termination;
458 // Allow skipping partition search for still image frame
459 int allow_partition_search_skip;
461 // Fast approximation of vp9_model_rd_from_var_lapndz
462 int simple_model_rd_from_var;
464 // Skip a number of expensive mode evaluations for blocks with zero source
466 int short_circuit_flat_blocks;
468 // Skip a number of expensive mode evaluations for blocks with very low
469 // temporal variance. If the low temporal variance flag is set for a block,
471 // 1: Skip all golden modes and ALL INTRA for bsize >= 32x32.
472 // 2: Skip golden non-zeromv and newmv-last for bsize >= 16x16, skip ALL
473 // INTRA for bsize >= 32x32 and vert/horz INTRA for bsize 16x16, 16x32 and
475 // 3: Same as (2), but also skip golden zeromv.
476 int short_circuit_low_temp_var;
478 // Limits the rd-threshold update for early exit for the newmv-last mode,
480 int limit_newmv_early_exit;
482 // Adds a bias against golden reference, for non-rd mode.
485 // Bias to use base mv and skip 1/4 subpel search when use base mv in
486 // enhancement layer.
487 int base_mv_aggressive;
489 // Global flag to enable partition copy from the previous frame.
490 int copy_partition_flag;
492 // Compute the source sad for every superblock of the frame,
493 // prior to encoding the frame, to be used to bypass some encoder decisions.
496 int use_simple_block_yrd;
501 void vp9_set_speed_features_framesize_independent(struct VP9_COMP *cpi);
502 void vp9_set_speed_features_framesize_dependent(struct VP9_COMP *cpi);
508 #endif // VP9_ENCODER_VP9_SPEED_FEATURES_H_