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_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) |
48 DISABLE_ALL_SPLIT = (1 << THR_INTRA) | DISABLE_ALL_INTER_SPLIT,
50 DISABLE_COMPOUND_SPLIT = (1 << THR_COMP_GA) | (1 << THR_COMP_LA),
52 LAST_AND_INTRA_SPLIT_ONLY = (1 << THR_COMP_GA) |
71 // Allow recode for KF and exceeding maximum frame bandwidth.
72 ALLOW_RECODE_KFMAXBW = 1,
73 // Allow recode only for KF/ARF/GF frames.
74 ALLOW_RECODE_KFARFGF = 2,
75 // Allow recode for all frames based on bitrate constraints.
81 SUBPEL_TREE_PRUNED = 1,
82 // Other methods to come
83 } SUBPEL_SEARCH_METHODS;
86 NO_MOTION_THRESHOLD = 0,
87 LOW_MOTION_THRESHOLD = 7
91 LAST_FRAME_PARTITION_OFF = 0,
92 LAST_FRAME_PARTITION_LOW_MOTION = 1,
93 LAST_FRAME_PARTITION_ALL = 2
94 } LAST_FRAME_PARTITION_METHOD;
100 } TX_SIZE_SEARCH_METHOD;
104 RELAXED_NEIGHBORING_MIN_MAX = 1,
105 CONSTRAIN_NEIGHBORING_MIN_MAX = 2,
106 STRICT_NEIGHBORING_MIN_MAX = 3
110 // Try the full image with different values.
111 LPF_PICK_FROM_FULL_IMAGE,
112 // Try a small portion of the image with different values.
113 LPF_PICK_FROM_SUBIMAGE,
114 // Estimate the level based on quantizer and frame type
116 // Pick 0 to disable LPF if LPF was enabled last frame
121 // Terminate search early based on distortion so far compared to
122 // qp step, distortion in the neighborhood of the frame, etc.
123 FLAG_EARLY_TERMINATE = 1 << 0,
125 // Skips comp inter modes if the best so far is an intra mode.
126 FLAG_SKIP_COMP_BESTINTRA = 1 << 1,
128 // Skips oblique intra modes if the best so far is an inter mode.
129 FLAG_SKIP_INTRA_BESTINTER = 1 << 3,
131 // Skips oblique intra modes at angles 27, 63, 117, 153 if the best
132 // intra so far is not one of the neighboring directions.
133 FLAG_SKIP_INTRA_DIRMISMATCH = 1 << 4,
135 // Skips intra modes other than DC_PRED if the source variance is small
136 FLAG_SKIP_INTRA_LOWVAR = 1 << 5,
137 } MODE_SEARCH_SKIP_LOGIC;
140 FLAG_SKIP_EIGHTTAP = 1 << EIGHTTAP,
141 FLAG_SKIP_EIGHTTAP_SMOOTH = 1 << EIGHTTAP_SMOOTH,
142 FLAG_SKIP_EIGHTTAP_SHARP = 1 << EIGHTTAP_SHARP,
143 } INTERP_FILTER_MASK;
146 // Search partitions using RD/NONRD criterion
147 SEARCH_PARTITION = 0,
149 // Always use a fixed size partition
152 // Use a fixed size partition in every 64X64 SB, where the size is
153 // determined based on source variance
154 VAR_BASED_FIXED_PARTITION = 2,
156 REFERENCE_PARTITION = 3,
158 // Use an arbitrary partitioning scheme based on source variance within
162 // Use non-fixed partitions based on source variance
163 SOURCE_VAR_BASED_PARTITION
164 } PARTITION_SEARCH_TYPE;
167 // Does a dry run to see if any of the contexts need to be updated or not,
168 // before the final run.
171 // No dry run conducted.
174 // No dry run, also only half the coef contexts and bands are updated.
175 // The rest are not updated at all.
179 typedef struct MV_SPEED_FEATURES {
180 // Motion search method (Diamond, NSTEP, Hex, Big Diamond, Square, etc).
181 SEARCH_METHODS search_method;
183 // This parameter controls which step in the n-step process we start at.
184 // It's changed adaptively based on circumstances.
185 int reduce_first_step_size;
187 // If this is set to 1, we limit the motion search range to 2 times the
188 // largest motion vector found in the last frame.
189 int auto_mv_step_size;
191 // Subpel_search_method can only be subpel_tree which does a subpixel
192 // logarithmic search that keeps stepping at 1/2 pixel units until
193 // you stop getting a gain, and then goes on to 1/4 and repeats
194 // the same process. Along the way it skips many diagonals.
195 SUBPEL_SEARCH_METHODS subpel_search_method;
197 // Maximum number of steps in logarithmic subpel search before giving up.
198 int subpel_iters_per_step;
200 // Control when to stop subpel search
201 int subpel_force_stop;
203 // This variable sets the step_param used in full pel motion search.
204 int fullpel_search_step_param;
207 typedef struct SPEED_FEATURES {
208 MV_SPEED_FEATURES mv;
210 // Frame level coding parameter update
211 int frame_parameter_update;
213 RECODE_LOOP_TYPE recode_loop;
215 // Trellis (dynamic programming) optimization of quantized values (+1, 0).
216 int optimize_coefficients;
218 // Always set to 0. If on it enables 0 cost background transmission
219 // (except for the initial transmission of the segmentation). The feature is
220 // disabled because the addition of very large block sizes make the
221 // backgrounds very to cheap to encode, and the segmentation we have
223 int static_segmentation;
225 // If 1 we iterate finding a best reference for 2 ref frames together - via
226 // a log search that iterates 4 times (check around mv for last for best
227 // error of combined predictor then check around mv for alt). If 0 we
228 // we just use the best motion vector found for each frame by itself.
229 BLOCK_SIZE comp_inter_joint_search_thresh;
231 // This variable is used to cap the maximum number of times we skip testing a
232 // mode to be evaluated. A high value means we will be faster.
233 int adaptive_rd_thresh;
235 // Enables skipping the reconstruction step (idct, recon) in the
236 // intermediate steps assuming the last frame didn't have too many intra
237 // blocks and the q is less than a threshold.
239 int skip_encode_frame;
240 // Speed feature to allow or disallow skipping of recode at block
241 // level within a frame.
242 int allow_skip_recode;
244 // This variable allows us to reuse the last frames partition choices
245 // (64x64 v 32x32 etc) for this frame. It can be set to only use the last
246 // frame as a starting point in low motion scenes or always use it. If set
247 // we use last partitioning_redo frequency to determine how often to redo
248 // the partitioning from scratch. Adjust_partitioning_from_last_frame
249 // enables us to adjust up or down one partitioning from the last frames
251 LAST_FRAME_PARTITION_METHOD use_lastframe_partitioning;
253 // The threshold is to determine how slow the motino is, it is used when
254 // use_lastframe_partitioning is set to LAST_FRAME_PARTITION_LOW_MOTION
255 MOTION_THRESHOLD lf_motion_threshold;
257 // Determine which method we use to determine transform size. We can choose
258 // between options like full rd, largest for prediction size, largest
259 // for intra and model coefs for the rest.
260 TX_SIZE_SEARCH_METHOD tx_size_search_method;
262 // Low precision 32x32 fdct keeps everything in 16 bits and thus is less
263 // precise but significantly faster than the non lp version.
266 // TODO(JBB): remove this as its no longer used.
268 // After looking at the first set of modes (set by index here), skip
269 // checking modes for reference frames that don't match the reference frame
270 // of the best so far.
273 // TODO(JBB): Remove this.
274 int reference_masking;
276 PARTITION_SEARCH_TYPE partition_search_type;
278 // Used if partition_search_type = FIXED_SIZE_PARTITION
279 BLOCK_SIZE always_this_block_size;
281 // Skip rectangular partition test when partition type none gives better
282 // rd than partition type split.
283 int less_rectangular_check;
285 // Disable testing non square partitions. (eg 16x32)
286 int use_square_partition_only;
288 // Sets min and max partition sizes for this 64x64 region based on the
289 // same 64x64 in last encoded frame, and the left and above neighbor.
290 AUTO_MIN_MAX_MODE auto_min_max_partition_size;
292 // Min and max partition size we enable (block_size) as per auto
293 // min max, but also used by adjust partitioning, and pick_partitioning.
294 BLOCK_SIZE min_partition_size;
295 BLOCK_SIZE max_partition_size;
297 // Whether or not we allow partitions one smaller or one greater than the last
298 // frame's partitioning. Only used if use_lastframe_partitioning is set.
299 int adjust_partitioning_from_last_frame;
301 // How frequently we re do the partitioning from scratch. Only used if
302 // use_lastframe_partitioning is set.
303 int last_partitioning_redo_frequency;
305 // This enables constrained copy partitioning, which, given an input block
306 // size bsize, will copy previous partition for partitions less than bsize,
307 // otherwise bsize partition is used. bsize is currently set to 16x16.
308 // Used for the case where motion is detected in superblock.
309 int constrain_copy_partition;
311 // Disables sub 8x8 blocksizes in different scenarios: Choices are to disable
312 // it always, to allow it for only Last frame and Intra, disable it for all
313 // inter modes or to enable it always.
314 int disable_split_mask;
316 // TODO(jingning): combine the related motion search speed features
317 // This allows us to use motion search at other sizes as a starting
318 // point for this motion search and limits the search range around it.
319 int adaptive_motion_search;
321 int schedule_mode_search;
323 // Allows sub 8x8 modes to use the prediction filter that was determined
324 // best for 8x8 mode. If set to 0 we always re check all the filters for
325 // sizes less than 8x8, 1 means we check all filter modes if no 8x8 filter
326 // was selected, and 2 means we use 8 tap if no 8x8 filter mode was selected.
327 int adaptive_pred_interp_filter;
329 // Adaptive prediction mode search
330 int adaptive_mode_search;
332 // Chessboard pattern prediction filter type search
333 int cb_pred_filter_search;
335 int cb_partition_search;
337 int motion_field_mode_search;
339 int alt_ref_search_fp;
341 // Fast quantization process path
344 // Search through variable block partition types in non-RD mode decision
345 // encoding process for RTC.
348 // Use finer quantizer in every other few frames that run variable block
349 // partition type search.
350 int force_frame_boost;
352 // Maximally allowed base quantization index fluctuation.
353 int max_delta_qindex;
355 // Implements various heuristics to skip searching modes
356 // The heuristics selected are based on flags
357 // defined in the MODE_SEARCH_SKIP_HEURISTICS enum
358 unsigned int mode_search_skip_flags;
360 // A source variance threshold below which filter search is disabled
361 // Choose a very large value (UINT_MAX) to use 8-tap always
362 unsigned int disable_filter_search_var_thresh;
364 // These bit masks allow you to enable or disable intra modes for each
365 // transform size separately.
366 int intra_y_mode_mask[TX_SIZES];
367 int intra_uv_mode_mask[TX_SIZES];
369 // This variable enables an early break out of mode testing if the model for
370 // rd built from the prediction signal indicates a value that's much
371 // higher than the best rd we've seen so far.
374 // This enables us to use an estimate for intra rd based on dc mode rather
375 // than choosing an actual uv mode in the stage of encoding before the actual
377 int use_uv_intra_rd_estimate;
379 // This feature controls how the loop filter level is determined.
380 LPF_PICK_METHOD lpf_pick;
382 // This feature limits the number of coefficients updates we actually do
383 // by only looking at counts from 1/2 the bands.
384 FAST_COEFF_UPDATE use_fast_coef_updates;
386 // This flag controls the use of non-RD mode decision.
387 int use_nonrd_pick_mode;
389 // A binary mask indicating if NEARESTMV, NEARMV, ZEROMV, NEWMV
390 // modes are used in order from LSB to MSB for each BLOCK_SIZE.
391 int inter_mode_mask[BLOCK_SIZES];
393 // This feature controls whether we do the expensive context update and
394 // calculation in the rd coefficient costing loop.
395 int use_fast_coef_costing;
397 // This feature controls the tolerence vs target used in deciding whether to
398 // recode a frame. It has no meaning if recode is disabled.
399 int recode_tolerance;
401 // This variable controls the maximum block size where intra blocks can be
402 // used in inter frames.
403 // TODO(aconverse): Fold this into one of the other many mode skips
404 BLOCK_SIZE max_intra_bsize;
406 // The frequency that we check if SOURCE_VAR_BASED_PARTITION or
407 // FIXED_PARTITION search type should be used.
408 int search_type_check_frequency;
410 // When partition is pre-set, the inter prediction result from pick_inter_mode
411 // can be reused in final block encoding process. It is enabled only for real-
412 // time mode speed 6.
413 int reuse_inter_pred_sby;
415 // This variable sets the encode_breakout threshold. Currently, it is only
416 // enabled in real time mode.
417 int encode_breakout_thresh;
419 // In real time encoding, increase the threshold for NEWMV.
420 int elevate_newmv_thresh;
422 // default interp filter choice
423 INTERP_FILTER default_interp_filter;
425 // Early termination in transform size search, which only applies while
426 // tx_size_search_method is USE_FULL_RD.
427 int tx_size_search_breakout;
429 // adaptive interp_filter search to allow skip of certain filter types.
430 int adaptive_interp_filter_search;
432 // mask for skip evaluation of certain interp_filter type.
433 INTERP_FILTER_MASK interp_filter_search_mask;
435 // Partition search early breakout thresholds.
436 int64_t partition_search_breakout_dist_thr;
437 int partition_search_breakout_rate_thr;
442 void vp9_set_speed_features(struct VP9_COMP *cpi);
448 #endif // VP9_ENCODER_VP9_SPEED_FEATURES_H_