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) |
22 (1 << V_PRED) | (1 << H_PRED) |
23 (1 << D45_PRED) | (1 << D135_PRED) |
24 (1 << D117_PRED) | (1 << D153_PRED) |
25 (1 << D207_PRED) | (1 << D63_PRED) |
27 INTRA_DC = (1 << DC_PRED),
28 INTRA_DC_TM = (1 << DC_PRED) | (1 << TM_PRED),
29 INTRA_DC_H_V = (1 << DC_PRED) | (1 << V_PRED) | (1 << H_PRED),
30 INTRA_DC_TM_H_V = (1 << DC_PRED) | (1 << TM_PRED) | (1 << V_PRED) |
35 INTER_ALL = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV) | (1 << NEWMV),
36 INTER_NEAREST = (1 << NEARESTMV),
37 INTER_NEAREST_NEW = (1 << NEARESTMV) | (1 << NEWMV),
38 INTER_NEAREST_ZERO = (1 << NEARESTMV) | (1 << ZEROMV),
39 INTER_NEAREST_NEW_ZERO = (1 << NEARESTMV) | (1 << ZEROMV) | (1 << NEWMV),
40 INTER_NEAREST_NEAR_NEW = (1 << NEARESTMV) | (1 << NEARMV) | (1 << NEWMV),
41 INTER_NEAREST_NEAR_ZERO = (1 << NEARESTMV) | (1 << NEARMV) | (1 << ZEROMV),
45 DISABLE_ALL_INTER_SPLIT = (1 << THR_COMP_GA) |
51 DISABLE_ALL_SPLIT = (1 << THR_INTRA) | DISABLE_ALL_INTER_SPLIT,
53 DISABLE_COMPOUND_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA),
55 LAST_AND_INTRA_SPLIT_ONLY = (1 << THR_COMP_GA) |
74 // Allow recode for KF and exceeding maximum frame bandwidth.
75 ALLOW_RECODE_KFMAXBW = 1,
76 // Allow recode only for KF/ARF/GF frames.
77 ALLOW_RECODE_KFARFGF = 2,
78 // Allow recode for all frames based on bitrate constraints.
84 SUBPEL_TREE_PRUNED = 1, // Prunes 1/2-pel searches
85 SUBPEL_TREE_PRUNED_MORE = 2, // Prunes 1/2-pel searches more aggressively
86 SUBPEL_TREE_PRUNED_EVENMORE = 3, // Prunes 1/2- and 1/4-pel searches
87 // Other methods to come
88 } SUBPEL_SEARCH_METHODS;
91 NO_MOTION_THRESHOLD = 0,
92 LOW_MOTION_THRESHOLD = 7
99 } TX_SIZE_SEARCH_METHOD;
103 RELAXED_NEIGHBORING_MIN_MAX = 1,
104 STRICT_NEIGHBORING_MIN_MAX = 2
108 // Try the full image with different values.
109 LPF_PICK_FROM_FULL_IMAGE,
110 // Try a small portion of the image with different values.
111 LPF_PICK_FROM_SUBIMAGE,
112 // Estimate the level based on quantizer and frame type
114 // Pick 0 to disable LPF if LPF was enabled last frame
119 // Terminate search early based on distortion so far compared to
120 // qp step, distortion in the neighborhood of the frame, etc.
121 FLAG_EARLY_TERMINATE = 1 << 0,
123 // Skips comp inter modes if the best so far is an intra mode.
124 FLAG_SKIP_COMP_BESTINTRA = 1 << 1,
126 // Skips oblique intra modes if the best so far is an inter mode.
127 FLAG_SKIP_INTRA_BESTINTER = 1 << 3,
129 // Skips oblique intra modes at angles 27, 63, 117, 153 if the best
130 // intra so far is not one of the neighboring directions.
131 FLAG_SKIP_INTRA_DIRMISMATCH = 1 << 4,
133 // Skips intra modes other than DC_PRED if the source variance is small
134 FLAG_SKIP_INTRA_LOWVAR = 1 << 5,
135 } MODE_SEARCH_SKIP_LOGIC;
138 FLAG_SKIP_EIGHTTAP = 1 << EIGHTTAP,
139 FLAG_SKIP_EIGHTTAP_SMOOTH = 1 << EIGHTTAP_SMOOTH,
140 FLAG_SKIP_EIGHTTAP_SHARP = 1 << EIGHTTAP_SHARP,
141 } INTERP_FILTER_MASK;
144 // Search partitions using RD/NONRD criterion
147 // Always use a fixed size partition
152 // Use an arbitrary partitioning scheme based on source variance within
156 // Use non-fixed partitions based on source variance
157 SOURCE_VAR_BASED_PARTITION
158 } PARTITION_SEARCH_TYPE;
161 // Does a dry run to see if any of the contexts need to be updated or not,
162 // before the final run.
165 // No dry run, also only half the coef contexts and bands are updated.
166 // The rest are not updated at all.
170 typedef struct MV_SPEED_FEATURES {
171 // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
172 SEARCH_METHODS search_method;
174 // This parameter controls which step in the n-step process we start at.
175 // It's changed adaptively based on circumstances.
176 int reduce_first_step_size;
178 // If this is set to 1, we limit the motion search range to 2 times the
179 // largest motion vector found in the last frame.
180 int auto_mv_step_size;
182 // Subpel_search_method can only be subpel_tree which does a subpixel
183 // logarithmic search that keeps stepping at 1/2 pixel units until
184 // you stop getting a gain, and then goes on to 1/4 and repeats
185 // the same process. Along the way it skips many diagonals.
186 SUBPEL_SEARCH_METHODS subpel_search_method;
188 // Maximum number of steps in logarithmic subpel search before giving up.
189 int subpel_iters_per_step;
191 // Control when to stop subpel search
192 int subpel_force_stop;
194 // This variable sets the step_param used in full pel motion search.
195 int fullpel_search_step_param;
198 typedef struct SPEED_FEATURES {
199 MV_SPEED_FEATURES mv;
201 // Frame level coding parameter update
202 int frame_parameter_update;
204 RECODE_LOOP_TYPE recode_loop;
206 // Trellis (dynamic programming) optimization of quantized values (+1, 0).
207 int optimize_coefficients;
209 // Always set to 0. If on it enables 0 cost background transmission
210 // (except for the initial transmission of the segmentation). The feature is
211 // disabled because the addition of very large block sizes make the
212 // backgrounds very to cheap to encode, and the segmentation we have
214 int static_segmentation;
216 // If 1 we iterate finding a best reference for 2 ref frames together - via
217 // a log search that iterates 4 times (check around mv for last for best
218 // error of combined predictor then check around mv for alt). If 0 we
219 // we just use the best motion vector found for each frame by itself.
220 BLOCK_SIZE comp_inter_joint_search_thresh;
222 // This variable is used to cap the maximum number of times we skip testing a
223 // mode to be evaluated. A high value means we will be faster.
224 int adaptive_rd_thresh;
226 // Enables skipping the reconstruction step (idct, recon) in the
227 // intermediate steps assuming the last frame didn't have too many intra
228 // blocks and the q is less than a threshold.
230 int skip_encode_frame;
231 // Speed feature to allow or disallow skipping of recode at block
232 // level within a frame.
233 int allow_skip_recode;
235 // Coefficient probability model approximation step size
236 int coeff_prob_appx_step;
238 // The threshold is to determine how slow the motino is, it is used when
239 // use_lastframe_partitioning is set to LAST_FRAME_PARTITION_LOW_MOTION
240 MOTION_THRESHOLD lf_motion_threshold;
242 // Determine which method we use to determine transform size. We can choose
243 // between options like full rd, largest for prediction size, largest
244 // for intra and model coefs for the rest.
245 TX_SIZE_SEARCH_METHOD tx_size_search_method;
247 // Low precision 32x32 fdct keeps everything in 16 bits and thus is less
248 // precise but significantly faster than the non lp version.
251 // After looking at the first set of modes (set by index here), skip
252 // checking modes for reference frames that don't match the reference frame
253 // of the best so far.
256 // TODO(JBB): Remove this.
257 int reference_masking;
259 PARTITION_SEARCH_TYPE partition_search_type;
261 // Used if partition_search_type = FIXED_SIZE_PARTITION
262 BLOCK_SIZE always_this_block_size;
264 // Skip rectangular partition test when partition type none gives better
265 // rd than partition type split.
266 int less_rectangular_check;
268 // Disable testing non square partitions. (eg 16x32)
269 int use_square_partition_only;
271 // Sets min and max partition sizes for this 64x64 region based on the
272 // same 64x64 in last encoded frame, and the left and above neighbor.
273 AUTO_MIN_MAX_MODE auto_min_max_partition_size;
274 // Ensures the rd based auto partition search will always
275 // go down at least to the specified level.
276 BLOCK_SIZE rd_auto_partition_min_limit;
278 // Min and max partition size we enable (block_size) as per auto
279 // min max, but also used by adjust partitioning, and pick_partitioning.
280 BLOCK_SIZE default_min_partition_size;
281 BLOCK_SIZE default_max_partition_size;
283 // Whether or not we allow partitions one smaller or one greater than the last
284 // frame's partitioning. Only used if use_lastframe_partitioning is set.
285 int adjust_partitioning_from_last_frame;
287 // How frequently we re do the partitioning from scratch. Only used if
288 // use_lastframe_partitioning is set.
289 int last_partitioning_redo_frequency;
291 // Disables sub 8x8 blocksizes in different scenarios: Choices are to disable
292 // it always, to allow it for only Last frame and Intra, disable it for all
293 // inter modes or to enable it always.
294 int disable_split_mask;
296 // TODO(jingning): combine the related motion search speed features
297 // This allows us to use motion search at other sizes as a starting
298 // point for this motion search and limits the search range around it.
299 int adaptive_motion_search;
301 int schedule_mode_search;
303 // Allows sub 8x8 modes to use the prediction filter that was determined
304 // best for 8x8 mode. If set to 0 we always re check all the filters for
305 // sizes less than 8x8, 1 means we check all filter modes if no 8x8 filter
306 // was selected, and 2 means we use 8 tap if no 8x8 filter mode was selected.
307 int adaptive_pred_interp_filter;
309 // Adaptive prediction mode search
310 int adaptive_mode_search;
312 // Chessboard pattern prediction filter type search
313 int cb_pred_filter_search;
315 int cb_partition_search;
317 int motion_field_mode_search;
319 int alt_ref_search_fp;
321 // Fast quantization process path
324 // Use finer quantizer in every other few frames that run variable block
325 // partition type search.
326 int force_frame_boost;
328 // Maximally allowed base quantization index fluctuation.
329 int max_delta_qindex;
331 // Implements various heuristics to skip searching modes
332 // The heuristics selected are based on flags
333 // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
334 unsigned int mode_search_skip_flags;
336 // A source variance threshold below which filter search is disabled
337 // Choose a very large value (UINT_MAX) to use 8-tap always
338 unsigned int disable_filter_search_var_thresh;
340 // These bit masks allow you to enable or disable intra modes for each
341 // transform size separately.
342 int intra_y_mode_mask[TX_SIZES];
343 int intra_uv_mode_mask[TX_SIZES];
345 // These bit masks allow you to enable or disable intra modes for each
346 // prediction block size separately.
347 int intra_y_mode_bsize_mask[BLOCK_SIZES];
349 // This variable enables an early break out of mode testing if the model for
350 // rd built from the prediction signal indicates a value that's much
351 // higher than the best rd we've seen so far.
354 // This enables us to use an estimate for intra rd based on dc mode rather
355 // than choosing an actual uv mode in the stage of encoding before the actual
357 int use_uv_intra_rd_estimate;
359 // This feature controls how the loop filter level is determined.
360 LPF_PICK_METHOD lpf_pick;
362 // This feature limits the number of coefficients updates we actually do
363 // by only looking at counts from 1/2 the bands.
364 FAST_COEFF_UPDATE use_fast_coef_updates;
366 // This flag controls the use of non-RD mode decision.
367 int use_nonrd_pick_mode;
369 // A binary mask indicating if NEARESTMV, NEARMV, ZEROMV, NEWMV
370 // modes are used in order from LSB to MSB for each BLOCK_SIZE.
371 int inter_mode_mask[BLOCK_SIZES];
373 // This feature controls whether we do the expensive context update and
374 // calculation in the rd coefficient costing loop.
375 int use_fast_coef_costing;
377 // This feature controls the tolerence vs target used in deciding whether to
378 // recode a frame. It has no meaning if recode is disabled.
379 int recode_tolerance;
381 // This variable controls the maximum block size where intra blocks can be
382 // used in inter frames.
383 // TODO(aconverse): Fold this into one of the other many mode skips
384 BLOCK_SIZE max_intra_bsize;
386 // The frequency that we check if SOURCE_VAR_BASED_PARTITION or
387 // FIXED_PARTITION search type should be used.
388 int search_type_check_frequency;
390 // When partition is pre-set, the inter prediction result from pick_inter_mode
391 // can be reused in final block encoding process. It is enabled only for real-
392 // time mode speed 6.
393 int reuse_inter_pred_sby;
395 // This variable sets the encode_breakout threshold. Currently, it is only
396 // enabled in real time mode.
397 int encode_breakout_thresh;
399 // default interp filter choice
400 INTERP_FILTER default_interp_filter;
402 // Early termination in transform size search, which only applies while
403 // tx_size_search_method is USE_FULL_RD.
404 int tx_size_search_breakout;
406 // adaptive interp_filter search to allow skip of certain filter types.
407 int adaptive_interp_filter_search;
409 // mask for skip evaluation of certain interp_filter type.
410 INTERP_FILTER_MASK interp_filter_search_mask;
412 // Partition search early breakout thresholds.
413 int64_t partition_search_breakout_dist_thr;
414 int partition_search_breakout_rate_thr;
416 // Allow skipping partition search for still image frame
417 int allow_partition_search_skip;
422 void vp9_set_speed_features_framesize_independent(struct VP9_COMP *cpi);
423 void vp9_set_speed_features_framesize_dependent(struct VP9_COMP *cpi);
429 #endif // VP9_ENCODER_VP9_SPEED_FEATURES_H_