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 #include <limits.h>
12
13 #include "vp9/encoder/vp9_encoder.h"
14 #include "vp9/encoder/vp9_speed_features.h"
15 #include "vp9/encoder/vp9_rdopt.h"
16 #include "vpx_dsp/vpx_dsp_common.h"
17
18 // Mesh search patters for various speed settings
19 static MESH_PATTERN best_quality_mesh_pattern[MAX_MESH_STEP] = {
20 { 64, 4 }, { 28, 2 }, { 15, 1 }, { 7, 1 }
21 };
22
23 // Define 3 mesh density levels to control the number of searches.
24 #define MESH_DENSITY_LEVELS 3
25 static MESH_PATTERN
26 good_quality_mesh_patterns[MESH_DENSITY_LEVELS][MAX_MESH_STEP] = {
27 { { 64, 8 }, { 28, 4 }, { 15, 1 }, { 7, 1 } },
28 { { 64, 8 }, { 14, 2 }, { 7, 1 }, { 7, 1 } },
29 { { 64, 16 }, { 24, 8 }, { 12, 4 }, { 7, 1 } },
30 };
31
32 // Intra only frames, golden frames (except alt ref overlays) and
33 // alt ref frames tend to be coded at a higher than ambient quality
frame_is_boosted(const VP9_COMP * cpi)34 static int frame_is_boosted(const VP9_COMP *cpi) {
35 return frame_is_kf_gf_arf(cpi) || vp9_is_upper_layer_key_frame(cpi);
36 }
37
38 // Sets a partition size down to which the auto partition code will always
39 // search (can go lower), based on the image dimensions. The logic here
40 // is that the extent to which ringing artefacts are offensive, depends
41 // partly on the screen area that over which they propogate. Propogation is
42 // limited by transform block size but the screen area take up by a given block
43 // size will be larger for a small image format stretched to full screen.
set_partition_min_limit(VP9_COMMON * const cm)44 static BLOCK_SIZE set_partition_min_limit(VP9_COMMON *const cm) {
45 unsigned int screen_area = (cm->width * cm->height);
46
47 // Select block size based on image format size.
48 if (screen_area < 1280 * 720) {
49 // Formats smaller in area than 720P
50 return BLOCK_4X4;
51 } else if (screen_area < 1920 * 1080) {
52 // Format >= 720P and < 1080P
53 return BLOCK_8X8;
54 } else {
55 // Formats 1080P and up
56 return BLOCK_16X16;
57 }
58 }
59
set_good_speed_feature_framesize_dependent(VP9_COMP * cpi,SPEED_FEATURES * sf,int speed)60 static void set_good_speed_feature_framesize_dependent(VP9_COMP *cpi,
61 SPEED_FEATURES *sf,
62 int speed) {
63 VP9_COMMON *const cm = &cpi->common;
64
65 // speed 0 features
66 sf->partition_search_breakout_thr.dist = (1 << 20);
67 sf->partition_search_breakout_thr.rate = 80;
68
69 // Currently, the machine-learning based partition search early termination
70 // is only used while VPXMIN(cm->width, cm->height) >= 480 and speed = 0.
71 if (VPXMIN(cm->width, cm->height) >= 480) {
72 sf->ml_partition_search_early_termination = 1;
73 }
74
75 if (speed >= 1) {
76 sf->ml_partition_search_early_termination = 0;
77
78 if (VPXMIN(cm->width, cm->height) >= 720) {
79 sf->disable_split_mask =
80 cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT;
81 sf->partition_search_breakout_thr.dist = (1 << 23);
82 } else {
83 sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
84 sf->partition_search_breakout_thr.dist = (1 << 21);
85 }
86 }
87
88 if (speed >= 2) {
89 if (VPXMIN(cm->width, cm->height) >= 720) {
90 sf->disable_split_mask =
91 cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT;
92 sf->adaptive_pred_interp_filter = 0;
93 sf->partition_search_breakout_thr.dist = (1 << 24);
94 sf->partition_search_breakout_thr.rate = 120;
95 } else {
96 sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY;
97 sf->partition_search_breakout_thr.dist = (1 << 22);
98 sf->partition_search_breakout_thr.rate = 100;
99 }
100 sf->rd_auto_partition_min_limit = set_partition_min_limit(cm);
101
102 // Use a set of speed features for 4k videos.
103 if (VPXMIN(cm->width, cm->height) >= 2160) {
104 sf->use_square_partition_only = 1;
105 sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
106 sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC;
107 sf->alt_ref_search_fp = 1;
108 sf->cb_pred_filter_search = 1;
109 sf->adaptive_interp_filter_search = 1;
110 sf->disable_split_mask = DISABLE_ALL_SPLIT;
111 }
112 }
113
114 if (speed >= 3) {
115 if (VPXMIN(cm->width, cm->height) >= 720) {
116 sf->disable_split_mask = DISABLE_ALL_SPLIT;
117 sf->schedule_mode_search = cm->base_qindex < 220 ? 1 : 0;
118 sf->partition_search_breakout_thr.dist = (1 << 25);
119 sf->partition_search_breakout_thr.rate = 200;
120 } else {
121 sf->max_intra_bsize = BLOCK_32X32;
122 sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT;
123 sf->schedule_mode_search = cm->base_qindex < 175 ? 1 : 0;
124 sf->partition_search_breakout_thr.dist = (1 << 23);
125 sf->partition_search_breakout_thr.rate = 120;
126 }
127 }
128
129 // If this is a two pass clip that fits the criteria for animated or
130 // graphics content then reset disable_split_mask for speeds 1-4.
131 // Also if the image edge is internal to the coded area.
132 if ((speed >= 1) && (cpi->oxcf.pass == 2) &&
133 ((cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ||
134 (vp9_internal_image_edge(cpi)))) {
135 sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
136 }
137
138 if (speed >= 4) {
139 sf->partition_search_breakout_thr.rate = 300;
140 if (VPXMIN(cm->width, cm->height) >= 720) {
141 sf->partition_search_breakout_thr.dist = (1 << 26);
142 } else {
143 sf->partition_search_breakout_thr.dist = (1 << 24);
144 }
145 sf->disable_split_mask = DISABLE_ALL_SPLIT;
146 }
147
148 if (speed >= 5) {
149 sf->partition_search_breakout_thr.rate = 500;
150 }
151 }
152
153 static double tx_dom_thresholds[6] = { 99.0, 14.0, 12.0, 8.0, 4.0, 0.0 };
154 static double qopt_thresholds[6] = { 99.0, 12.0, 10.0, 4.0, 2.0, 0.0 };
155
set_good_speed_feature_framesize_independent(VP9_COMP * cpi,VP9_COMMON * cm,SPEED_FEATURES * sf,int speed)156 static void set_good_speed_feature_framesize_independent(VP9_COMP *cpi,
157 VP9_COMMON *cm,
158 SPEED_FEATURES *sf,
159 int speed) {
160 const int boosted = frame_is_boosted(cpi);
161 int i;
162
163 sf->tx_size_search_breakout = 1;
164 sf->adaptive_rd_thresh = 1;
165 sf->adaptive_rd_thresh_row_mt = 0;
166 sf->allow_skip_recode = 1;
167 sf->less_rectangular_check = 1;
168 sf->use_square_partition_only = !frame_is_boosted(cpi);
169 sf->use_square_only_threshold = BLOCK_16X16;
170
171 if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) {
172 sf->exhaustive_searches_thresh = (1 << 22);
173 for (i = 0; i < MAX_MESH_STEP; ++i) {
174 int mesh_density_level = 0;
175 sf->mesh_patterns[i].range =
176 good_quality_mesh_patterns[mesh_density_level][i].range;
177 sf->mesh_patterns[i].interval =
178 good_quality_mesh_patterns[mesh_density_level][i].interval;
179 }
180 } else {
181 sf->exhaustive_searches_thresh = INT_MAX;
182 }
183
184 if (speed >= 1) {
185 if (cpi->oxcf.pass == 2) {
186 TWO_PASS *const twopass = &cpi->twopass;
187 if ((twopass->fr_content_type == FC_GRAPHICS_ANIMATION) ||
188 vp9_internal_image_edge(cpi)) {
189 sf->use_square_partition_only = !frame_is_boosted(cpi);
190 } else {
191 sf->use_square_partition_only = !frame_is_intra_only(cm);
192 }
193 } else {
194 sf->use_square_partition_only = !frame_is_intra_only(cm);
195 }
196
197 sf->allow_txfm_domain_distortion = 1;
198 sf->tx_domain_thresh = tx_dom_thresholds[(speed < 6) ? speed : 5];
199 sf->allow_quant_coeff_opt = sf->optimize_coefficients;
200 sf->quant_opt_thresh = qopt_thresholds[(speed < 6) ? speed : 5];
201
202 sf->use_square_only_threshold = BLOCK_4X4;
203 sf->less_rectangular_check = 1;
204
205 sf->use_rd_breakout = 1;
206 sf->adaptive_motion_search = 1;
207 sf->mv.auto_mv_step_size = 1;
208 sf->adaptive_rd_thresh = 2;
209 sf->mv.subpel_iters_per_step = 1;
210 sf->mode_skip_start = 10;
211 sf->adaptive_pred_interp_filter = 1;
212 sf->allow_acl = 0;
213
214 sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
215 sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
216 sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
217 sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
218
219 sf->recode_tolerance_low = 15;
220 sf->recode_tolerance_high = 30;
221
222 sf->exhaustive_searches_thresh =
223 (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ? (1 << 23)
224 : INT_MAX;
225 }
226
227 if (speed >= 2) {
228 sf->recode_loop = ALLOW_RECODE_KFARFGF;
229 sf->tx_size_search_method =
230 frame_is_boosted(cpi) ? USE_FULL_RD : USE_LARGESTALL;
231
232 // Reference masking is not supported in dynamic scaling mode.
233 sf->reference_masking = cpi->oxcf.resize_mode != RESIZE_DYNAMIC ? 1 : 0;
234
235 sf->mode_search_skip_flags =
236 (cm->frame_type == KEY_FRAME)
237 ? 0
238 : FLAG_SKIP_INTRA_DIRMISMATCH | FLAG_SKIP_INTRA_BESTINTER |
239 FLAG_SKIP_COMP_BESTINTRA | FLAG_SKIP_INTRA_LOWVAR;
240 sf->disable_filter_search_var_thresh = 100;
241 sf->comp_inter_joint_search_thresh = BLOCK_SIZES;
242 sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX;
243 sf->allow_partition_search_skip = 1;
244 sf->recode_tolerance_low = 15;
245 sf->recode_tolerance_high = 45;
246
247 if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) {
248 for (i = 0; i < MAX_MESH_STEP; ++i) {
249 int mesh_density_level = 1;
250 sf->mesh_patterns[i].range =
251 good_quality_mesh_patterns[mesh_density_level][i].range;
252 sf->mesh_patterns[i].interval =
253 good_quality_mesh_patterns[mesh_density_level][i].interval;
254 }
255 }
256 }
257
258 if (speed >= 3) {
259 sf->use_square_partition_only = !frame_is_intra_only(cm);
260 sf->tx_size_search_method =
261 frame_is_intra_only(cm) ? USE_FULL_RD : USE_LARGESTALL;
262 sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED;
263 sf->adaptive_pred_interp_filter = 0;
264 sf->adaptive_mode_search = 1;
265 sf->cb_partition_search = !boosted;
266 sf->cb_pred_filter_search = 1;
267 sf->alt_ref_search_fp = 1;
268 sf->recode_loop = ALLOW_RECODE_KFMAXBW;
269 sf->adaptive_rd_thresh = 3;
270 sf->mode_skip_start = 6;
271 sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
272 sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC;
273 sf->adaptive_interp_filter_search = 1;
274
275 if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) {
276 for (i = 0; i < MAX_MESH_STEP; ++i) {
277 int mesh_density_level = 2;
278 sf->mesh_patterns[i].range =
279 good_quality_mesh_patterns[mesh_density_level][i].range;
280 sf->mesh_patterns[i].interval =
281 good_quality_mesh_patterns[mesh_density_level][i].interval;
282 }
283 }
284 }
285
286 if (speed >= 4) {
287 sf->use_square_partition_only = 1;
288 sf->tx_size_search_method = USE_LARGESTALL;
289 sf->mv.search_method = BIGDIA;
290 sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED_MORE;
291 sf->adaptive_rd_thresh = 4;
292 if (cm->frame_type != KEY_FRAME)
293 sf->mode_search_skip_flags |= FLAG_EARLY_TERMINATE;
294 sf->disable_filter_search_var_thresh = 200;
295 sf->use_lp32x32fdct = 1;
296 sf->use_fast_coef_updates = ONE_LOOP_REDUCED;
297 sf->use_fast_coef_costing = 1;
298 sf->motion_field_mode_search = !boosted;
299 }
300
301 if (speed >= 5) {
302 int i;
303 sf->optimize_coefficients = 0;
304 sf->mv.search_method = HEX;
305 sf->disable_filter_search_var_thresh = 500;
306 for (i = 0; i < TX_SIZES; ++i) {
307 sf->intra_y_mode_mask[i] = INTRA_DC;
308 sf->intra_uv_mode_mask[i] = INTRA_DC;
309 }
310 sf->mv.reduce_first_step_size = 1;
311 sf->simple_model_rd_from_var = 1;
312 }
313 }
314
set_rt_speed_feature_framesize_dependent(VP9_COMP * cpi,SPEED_FEATURES * sf,int speed)315 static void set_rt_speed_feature_framesize_dependent(VP9_COMP *cpi,
316 SPEED_FEATURES *sf,
317 int speed) {
318 VP9_COMMON *const cm = &cpi->common;
319
320 if (speed >= 1) {
321 if (VPXMIN(cm->width, cm->height) >= 720) {
322 sf->disable_split_mask =
323 cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT;
324 } else {
325 sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
326 }
327 }
328
329 if (speed >= 2) {
330 if (VPXMIN(cm->width, cm->height) >= 720) {
331 sf->disable_split_mask =
332 cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT;
333 } else {
334 sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY;
335 }
336 }
337
338 if (speed >= 5) {
339 sf->partition_search_breakout_thr.rate = 200;
340 if (VPXMIN(cm->width, cm->height) >= 720) {
341 sf->partition_search_breakout_thr.dist = (1 << 25);
342 } else {
343 sf->partition_search_breakout_thr.dist = (1 << 23);
344 }
345 }
346
347 if (speed >= 7) {
348 sf->encode_breakout_thresh =
349 (VPXMIN(cm->width, cm->height) >= 720) ? 800 : 300;
350 }
351 }
352
set_rt_speed_feature_framesize_independent(VP9_COMP * cpi,SPEED_FEATURES * sf,int speed,vp9e_tune_content content)353 static void set_rt_speed_feature_framesize_independent(
354 VP9_COMP *cpi, SPEED_FEATURES *sf, int speed, vp9e_tune_content content) {
355 VP9_COMMON *const cm = &cpi->common;
356 const int is_keyframe = cm->frame_type == KEY_FRAME;
357 const int frames_since_key = is_keyframe ? 0 : cpi->rc.frames_since_key;
358 sf->static_segmentation = 0;
359 sf->adaptive_rd_thresh = 1;
360 sf->adaptive_rd_thresh_row_mt = 0;
361 sf->use_fast_coef_costing = 1;
362 sf->exhaustive_searches_thresh = INT_MAX;
363 sf->allow_acl = 0;
364 sf->copy_partition_flag = 0;
365 sf->use_source_sad = 0;
366 sf->use_simple_block_yrd = 0;
367
368 if (speed >= 1) {
369 sf->allow_txfm_domain_distortion = 1;
370 sf->tx_domain_thresh = 0.0;
371 sf->allow_quant_coeff_opt = 0;
372 sf->quant_opt_thresh = 0.0;
373 sf->use_square_partition_only = !frame_is_intra_only(cm);
374 sf->less_rectangular_check = 1;
375 sf->tx_size_search_method =
376 frame_is_intra_only(cm) ? USE_FULL_RD : USE_LARGESTALL;
377
378 sf->use_rd_breakout = 1;
379
380 sf->adaptive_motion_search = 1;
381 sf->adaptive_pred_interp_filter = 1;
382 sf->mv.auto_mv_step_size = 1;
383 sf->adaptive_rd_thresh = 2;
384 sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
385 sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
386 sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
387 }
388
389 if (speed >= 2) {
390 sf->mode_search_skip_flags =
391 (cm->frame_type == KEY_FRAME)
392 ? 0
393 : FLAG_SKIP_INTRA_DIRMISMATCH | FLAG_SKIP_INTRA_BESTINTER |
394 FLAG_SKIP_COMP_BESTINTRA | FLAG_SKIP_INTRA_LOWVAR;
395 sf->adaptive_pred_interp_filter = 2;
396
397 // Reference masking only enabled for 1 spatial layer, and if none of the
398 // references have been scaled. The latter condition needs to be checked
399 // for external or internal dynamic resize.
400 sf->reference_masking = (cpi->svc.number_spatial_layers == 1);
401 if (sf->reference_masking == 1 &&
402 (cpi->external_resize == 1 ||
403 cpi->oxcf.resize_mode == RESIZE_DYNAMIC)) {
404 MV_REFERENCE_FRAME ref_frame;
405 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
406 VP9_ALT_FLAG };
407 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
408 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
409 if (yv12 != NULL && (cpi->ref_frame_flags & flag_list[ref_frame])) {
410 const struct scale_factors *const scale_fac =
411 &cm->frame_refs[ref_frame - 1].sf;
412 if (vp9_is_scaled(scale_fac)) sf->reference_masking = 0;
413 }
414 }
415 }
416
417 sf->disable_filter_search_var_thresh = 50;
418 sf->comp_inter_joint_search_thresh = BLOCK_SIZES;
419 sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX;
420 sf->lf_motion_threshold = LOW_MOTION_THRESHOLD;
421 sf->adjust_partitioning_from_last_frame = 1;
422 sf->last_partitioning_redo_frequency = 3;
423 sf->use_lp32x32fdct = 1;
424 sf->mode_skip_start = 11;
425 sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
426 }
427
428 if (speed >= 3) {
429 sf->use_square_partition_only = 1;
430 sf->disable_filter_search_var_thresh = 100;
431 sf->use_uv_intra_rd_estimate = 1;
432 sf->skip_encode_sb = 1;
433 sf->mv.subpel_iters_per_step = 1;
434 sf->adaptive_rd_thresh = 4;
435 sf->mode_skip_start = 6;
436 sf->allow_skip_recode = 0;
437 sf->optimize_coefficients = 0;
438 sf->disable_split_mask = DISABLE_ALL_SPLIT;
439 sf->lpf_pick = LPF_PICK_FROM_Q;
440 }
441
442 if (speed >= 4) {
443 int i;
444 sf->last_partitioning_redo_frequency = 4;
445 sf->adaptive_rd_thresh = 5;
446 sf->use_fast_coef_costing = 0;
447 sf->auto_min_max_partition_size = STRICT_NEIGHBORING_MIN_MAX;
448 sf->adjust_partitioning_from_last_frame =
449 cm->last_frame_type != cm->frame_type ||
450 (0 == (frames_since_key + 1) % sf->last_partitioning_redo_frequency);
451 sf->mv.subpel_force_stop = 1;
452 for (i = 0; i < TX_SIZES; i++) {
453 sf->intra_y_mode_mask[i] = INTRA_DC_H_V;
454 sf->intra_uv_mode_mask[i] = INTRA_DC;
455 }
456 sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
457 sf->frame_parameter_update = 0;
458 sf->mv.search_method = FAST_HEX;
459
460 sf->inter_mode_mask[BLOCK_32X32] = INTER_NEAREST_NEAR_NEW;
461 sf->inter_mode_mask[BLOCK_32X64] = INTER_NEAREST;
462 sf->inter_mode_mask[BLOCK_64X32] = INTER_NEAREST;
463 sf->inter_mode_mask[BLOCK_64X64] = INTER_NEAREST;
464 sf->max_intra_bsize = BLOCK_32X32;
465 sf->allow_skip_recode = 1;
466 }
467
468 if (speed >= 5) {
469 sf->use_quant_fp = !is_keyframe;
470 sf->auto_min_max_partition_size =
471 is_keyframe ? RELAXED_NEIGHBORING_MIN_MAX : STRICT_NEIGHBORING_MIN_MAX;
472 sf->default_max_partition_size = BLOCK_32X32;
473 sf->default_min_partition_size = BLOCK_8X8;
474 sf->force_frame_boost =
475 is_keyframe ||
476 (frames_since_key % (sf->last_partitioning_redo_frequency << 1) == 1);
477 sf->max_delta_qindex = is_keyframe ? 20 : 15;
478 sf->partition_search_type = REFERENCE_PARTITION;
479 if (cpi->oxcf.rc_mode == VPX_VBR && cpi->oxcf.lag_in_frames > 0 &&
480 cpi->rc.is_src_frame_alt_ref) {
481 sf->partition_search_type = VAR_BASED_PARTITION;
482 }
483 sf->use_nonrd_pick_mode = 1;
484 sf->allow_skip_recode = 0;
485 sf->inter_mode_mask[BLOCK_32X32] = INTER_NEAREST_NEW_ZERO;
486 sf->inter_mode_mask[BLOCK_32X64] = INTER_NEAREST_NEW_ZERO;
487 sf->inter_mode_mask[BLOCK_64X32] = INTER_NEAREST_NEW_ZERO;
488 sf->inter_mode_mask[BLOCK_64X64] = INTER_NEAREST_NEW_ZERO;
489 sf->adaptive_rd_thresh = 2;
490 // This feature is only enabled when partition search is disabled.
491 sf->reuse_inter_pred_sby = 1;
492 sf->coeff_prob_appx_step = 4;
493 sf->use_fast_coef_updates = is_keyframe ? TWO_LOOP : ONE_LOOP_REDUCED;
494 sf->mode_search_skip_flags = FLAG_SKIP_INTRA_DIRMISMATCH;
495 sf->tx_size_search_method = is_keyframe ? USE_LARGESTALL : USE_TX_8X8;
496 sf->simple_model_rd_from_var = 1;
497 if (cpi->oxcf.rc_mode == VPX_VBR) sf->mv.search_method = NSTEP;
498
499 if (!is_keyframe) {
500 int i;
501 if (content == VP9E_CONTENT_SCREEN) {
502 for (i = 0; i < BLOCK_SIZES; ++i)
503 sf->intra_y_mode_bsize_mask[i] = INTRA_DC_TM_H_V;
504 } else {
505 for (i = 0; i < BLOCK_SIZES; ++i)
506 if (i > BLOCK_16X16)
507 sf->intra_y_mode_bsize_mask[i] = INTRA_DC;
508 else
509 // Use H and V intra mode for block sizes <= 16X16.
510 sf->intra_y_mode_bsize_mask[i] = INTRA_DC_H_V;
511 }
512 }
513 if (content == VP9E_CONTENT_SCREEN) {
514 sf->short_circuit_flat_blocks = 1;
515 }
516 if (cpi->oxcf.rc_mode == VPX_CBR &&
517 cpi->oxcf.content != VP9E_CONTENT_SCREEN) {
518 sf->limit_newmv_early_exit = 1;
519 if (!cpi->use_svc) sf->bias_golden = 1;
520 }
521 }
522
523 if (speed >= 6) {
524 sf->partition_search_type = VAR_BASED_PARTITION;
525 // Turn on this to use non-RD key frame coding mode.
526 sf->use_nonrd_pick_mode = 1;
527 sf->mv.search_method = NSTEP;
528 sf->mv.reduce_first_step_size = 1;
529 sf->skip_encode_sb = 0;
530 if (cpi->oxcf.rc_mode == VPX_CBR && content != VP9E_CONTENT_SCREEN) {
531 // Enable short circuit for low temporal variance.
532 sf->short_circuit_low_temp_var = 1;
533 }
534 if (cpi->svc.temporal_layer_id > 0) {
535 sf->adaptive_rd_thresh = 4;
536 sf->limit_newmv_early_exit = 0;
537 sf->mv.subpel_force_stop = (cpi->svc.temporal_layer_id == 1) ? 1 : 2;
538 sf->base_mv_aggressive =
539 (cpi->svc.temporal_layer_id == cpi->svc.number_temporal_layers - 1)
540 ? 1
541 : 0;
542 }
543 }
544
545 if (speed >= 7) {
546 sf->adaptive_rd_thresh = 3;
547 sf->mv.search_method = FAST_DIAMOND;
548 sf->mv.fullpel_search_step_param = 10;
549 if (cpi->svc.number_temporal_layers > 2 &&
550 cpi->svc.temporal_layer_id == 0) {
551 sf->mv.search_method = NSTEP;
552 sf->mv.fullpel_search_step_param = 6;
553 }
554 if (!cpi->external_resize) sf->use_source_sad = 1;
555 if (sf->use_source_sad) {
556 if (cpi->content_state_sb_fd == NULL &&
557 (!cpi->use_svc ||
558 cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1)) {
559 cpi->content_state_sb_fd = (uint8_t *)vpx_calloc(
560 (cm->mi_stride >> 3) * ((cm->mi_rows >> 3) + 1), sizeof(uint8_t));
561 }
562 }
563 }
564
565 if (speed >= 8) {
566 sf->adaptive_rd_thresh = 4;
567 // Enable partition copy. For SVC, only enabled for top resolution layer,
568 if (!cpi->last_frame_dropped && cpi->resize_state == ORIG &&
569 !cpi->external_resize &&
570 (!cpi->use_svc ||
571 cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1)) {
572 sf->copy_partition_flag = 1;
573 cpi->max_copied_frame = 4;
574 }
575
576 if (cpi->row_mt && cpi->oxcf.max_threads > 1)
577 sf->adaptive_rd_thresh_row_mt = 1;
578
579 if (content == VP9E_CONTENT_SCREEN)
580 sf->mv.subpel_force_stop = 3;
581 else if (cm->width * cm->height > 352 * 288)
582 sf->mv.subpel_force_stop = 2;
583
584 if (content == VP9E_CONTENT_SCREEN) sf->lpf_pick = LPF_PICK_MINIMAL_LPF;
585 // Only keep INTRA_DC mode for speed 8.
586 if (!is_keyframe) {
587 int i = 0;
588 for (i = 0; i < BLOCK_SIZES; ++i)
589 sf->intra_y_mode_bsize_mask[i] = INTRA_DC;
590 }
591 if (!cpi->use_svc && cpi->oxcf.rc_mode == VPX_CBR &&
592 content != VP9E_CONTENT_SCREEN) {
593 // More aggressive short circuit for speed 8.
594 sf->short_circuit_low_temp_var = 3;
595 // Use level 2 for noisey cases as there is a regression in some
596 // noisy clips with level 3.
597 if (cpi->noise_estimate.enabled && cm->width >= 1280 &&
598 cm->height >= 720) {
599 NOISE_LEVEL noise_level =
600 vp9_noise_estimate_extract_level(&cpi->noise_estimate);
601 if (noise_level >= kMedium) sf->short_circuit_low_temp_var = 2;
602 }
603 // Since the short_circuit_low_temp_var is used, reduce the
604 // adaptive_rd_thresh level.
605 if (cm->width * cm->height > 352 * 288)
606 sf->adaptive_rd_thresh = 1;
607 else
608 sf->adaptive_rd_thresh = 2;
609 }
610 sf->limit_newmv_early_exit = 0;
611 sf->use_simple_block_yrd = 1;
612 }
613 }
614
vp9_set_speed_features_framesize_dependent(VP9_COMP * cpi)615 void vp9_set_speed_features_framesize_dependent(VP9_COMP *cpi) {
616 SPEED_FEATURES *const sf = &cpi->sf;
617 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
618 RD_OPT *const rd = &cpi->rd;
619 int i;
620
621 // best quality defaults
622 // Some speed-up features even for best quality as minimal impact on quality.
623 sf->partition_search_breakout_thr.dist = (1 << 19);
624 sf->partition_search_breakout_thr.rate = 80;
625 sf->ml_partition_search_early_termination = 0;
626
627 if (oxcf->mode == REALTIME) {
628 set_rt_speed_feature_framesize_dependent(cpi, sf, oxcf->speed);
629 } else if (oxcf->mode == GOOD) {
630 set_good_speed_feature_framesize_dependent(cpi, sf, oxcf->speed);
631 }
632
633 if (sf->disable_split_mask == DISABLE_ALL_SPLIT) {
634 sf->adaptive_pred_interp_filter = 0;
635 }
636
637 if (cpi->encode_breakout && oxcf->mode == REALTIME &&
638 sf->encode_breakout_thresh > cpi->encode_breakout) {
639 cpi->encode_breakout = sf->encode_breakout_thresh;
640 }
641
642 // Check for masked out split cases.
643 for (i = 0; i < MAX_REFS; ++i) {
644 if (sf->disable_split_mask & (1 << i)) {
645 rd->thresh_mult_sub8x8[i] = INT_MAX;
646 }
647 }
648
649 // With row based multi-threading, the following speed features
650 // have to be disabled to guarantee that bitstreams encoded with single thread
651 // and multiple threads match.
652 // It can be used in realtime when adaptive_rd_thresh_row_mt is enabled since
653 // adaptive_rd_thresh is defined per-row for non-rd pickmode.
654 if (!sf->adaptive_rd_thresh_row_mt && cpi->row_mt_bit_exact)
655 sf->adaptive_rd_thresh = 0;
656
657 // This is only used in motion vector unit test.
658 if (cpi->oxcf.motion_vector_unit_test == 1)
659 cpi->find_fractional_mv_step = vp9_return_max_sub_pixel_mv;
660 else if (cpi->oxcf.motion_vector_unit_test == 2)
661 cpi->find_fractional_mv_step = vp9_return_min_sub_pixel_mv;
662 }
663
vp9_set_speed_features_framesize_independent(VP9_COMP * cpi)664 void vp9_set_speed_features_framesize_independent(VP9_COMP *cpi) {
665 SPEED_FEATURES *const sf = &cpi->sf;
666 VP9_COMMON *const cm = &cpi->common;
667 MACROBLOCK *const x = &cpi->td.mb;
668 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
669 int i;
670
671 // best quality defaults
672 sf->frame_parameter_update = 1;
673 sf->mv.search_method = NSTEP;
674 sf->recode_loop = ALLOW_RECODE_FIRST;
675 sf->mv.subpel_search_method = SUBPEL_TREE;
676 sf->mv.subpel_iters_per_step = 2;
677 sf->mv.subpel_force_stop = 0;
678 sf->optimize_coefficients = !is_lossless_requested(&cpi->oxcf);
679 sf->mv.reduce_first_step_size = 0;
680 sf->coeff_prob_appx_step = 1;
681 sf->mv.auto_mv_step_size = 0;
682 sf->mv.fullpel_search_step_param = 6;
683 sf->comp_inter_joint_search_thresh = BLOCK_4X4;
684 sf->tx_size_search_method = USE_FULL_RD;
685 sf->use_lp32x32fdct = 0;
686 sf->adaptive_motion_search = 0;
687 sf->adaptive_pred_interp_filter = 0;
688 sf->adaptive_mode_search = 0;
689 sf->cb_pred_filter_search = 0;
690 sf->cb_partition_search = 0;
691 sf->motion_field_mode_search = 0;
692 sf->alt_ref_search_fp = 0;
693 sf->use_quant_fp = 0;
694 sf->reference_masking = 0;
695 sf->partition_search_type = SEARCH_PARTITION;
696 sf->less_rectangular_check = 0;
697 sf->use_square_partition_only = 0;
698 sf->use_square_only_threshold = BLOCK_SIZES;
699 sf->auto_min_max_partition_size = NOT_IN_USE;
700 sf->rd_auto_partition_min_limit = BLOCK_4X4;
701 sf->default_max_partition_size = BLOCK_64X64;
702 sf->default_min_partition_size = BLOCK_4X4;
703 sf->adjust_partitioning_from_last_frame = 0;
704 sf->last_partitioning_redo_frequency = 4;
705 sf->disable_split_mask = 0;
706 sf->mode_search_skip_flags = 0;
707 sf->force_frame_boost = 0;
708 sf->max_delta_qindex = 0;
709 sf->disable_filter_search_var_thresh = 0;
710 sf->adaptive_interp_filter_search = 0;
711 sf->allow_partition_search_skip = 0;
712 sf->allow_txfm_domain_distortion = 0;
713 sf->tx_domain_thresh = 99.0;
714 sf->allow_quant_coeff_opt = sf->optimize_coefficients;
715 sf->quant_opt_thresh = 99.0;
716 sf->allow_acl = 1;
717
718 for (i = 0; i < TX_SIZES; i++) {
719 sf->intra_y_mode_mask[i] = INTRA_ALL;
720 sf->intra_uv_mode_mask[i] = INTRA_ALL;
721 }
722 sf->use_rd_breakout = 0;
723 sf->skip_encode_sb = 0;
724 sf->use_uv_intra_rd_estimate = 0;
725 sf->allow_skip_recode = 0;
726 sf->lpf_pick = LPF_PICK_FROM_FULL_IMAGE;
727 sf->use_fast_coef_updates = TWO_LOOP;
728 sf->use_fast_coef_costing = 0;
729 sf->mode_skip_start = MAX_MODES; // Mode index at which mode skip mask set
730 sf->schedule_mode_search = 0;
731 sf->use_nonrd_pick_mode = 0;
732 for (i = 0; i < BLOCK_SIZES; ++i) sf->inter_mode_mask[i] = INTER_ALL;
733 sf->max_intra_bsize = BLOCK_64X64;
734 sf->reuse_inter_pred_sby = 0;
735 // This setting only takes effect when partition_search_type is set
736 // to FIXED_PARTITION.
737 sf->always_this_block_size = BLOCK_16X16;
738 sf->search_type_check_frequency = 50;
739 sf->encode_breakout_thresh = 0;
740 // Recode loop tolerance %.
741 sf->recode_tolerance_low = 12;
742 sf->recode_tolerance_high = 25;
743 sf->default_interp_filter = SWITCHABLE;
744 sf->simple_model_rd_from_var = 0;
745 sf->short_circuit_flat_blocks = 0;
746 sf->short_circuit_low_temp_var = 0;
747 sf->limit_newmv_early_exit = 0;
748 sf->bias_golden = 0;
749 sf->base_mv_aggressive = 0;
750
751 // Some speed-up features even for best quality as minimal impact on quality.
752 sf->adaptive_rd_thresh = 1;
753 sf->tx_size_search_breakout = 1;
754
755 sf->exhaustive_searches_thresh =
756 (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ? (1 << 20)
757 : INT_MAX;
758 if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) {
759 for (i = 0; i < MAX_MESH_STEP; ++i) {
760 sf->mesh_patterns[i].range = best_quality_mesh_pattern[i].range;
761 sf->mesh_patterns[i].interval = best_quality_mesh_pattern[i].interval;
762 }
763 }
764
765 if (oxcf->mode == REALTIME)
766 set_rt_speed_feature_framesize_independent(cpi, sf, oxcf->speed,
767 oxcf->content);
768 else if (oxcf->mode == GOOD)
769 set_good_speed_feature_framesize_independent(cpi, cm, sf, oxcf->speed);
770
771 cpi->full_search_sad = vp9_full_search_sad;
772 cpi->diamond_search_sad = vp9_diamond_search_sad;
773
774 // Slow quant, dct and trellis not worthwhile for first pass
775 // so make sure they are always turned off.
776 if (oxcf->pass == 1) sf->optimize_coefficients = 0;
777
778 // No recode for 1 pass.
779 if (oxcf->pass == 0) {
780 sf->recode_loop = DISALLOW_RECODE;
781 sf->optimize_coefficients = 0;
782 }
783
784 if (sf->mv.subpel_force_stop == 3) {
785 // Whole pel only
786 cpi->find_fractional_mv_step = vp9_skip_sub_pixel_tree;
787 } else if (sf->mv.subpel_search_method == SUBPEL_TREE) {
788 cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree;
789 } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED) {
790 cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned;
791 } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED_MORE) {
792 cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned_more;
793 } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED_EVENMORE) {
794 cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned_evenmore;
795 }
796
797 x->optimize = sf->optimize_coefficients == 1 && oxcf->pass != 1;
798
799 x->min_partition_size = sf->default_min_partition_size;
800 x->max_partition_size = sf->default_max_partition_size;
801
802 if (!cpi->oxcf.frame_periodic_boost) {
803 sf->max_delta_qindex = 0;
804 }
805
806 // With row based multi-threading, the following speed features
807 // have to be disabled to guarantee that bitstreams encoded with single thread
808 // and multiple threads match.
809 // It can be used in realtime when adaptive_rd_thresh_row_mt is enabled since
810 // adaptive_rd_thresh is defined per-row for non-rd pickmode.
811 if (!sf->adaptive_rd_thresh_row_mt && cpi->row_mt_bit_exact)
812 sf->adaptive_rd_thresh = 0;
813
814 // This is only used in motion vector unit test.
815 if (cpi->oxcf.motion_vector_unit_test == 1)
816 cpi->find_fractional_mv_step = vp9_return_max_sub_pixel_mv;
817 else if (cpi->oxcf.motion_vector_unit_test == 2)
818 cpi->find_fractional_mv_step = vp9_return_min_sub_pixel_mv;
819 }
820