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.
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"
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 }
23 #define MAX_MESH_SPEED 5 // Max speed setting for mesh motion method
25 good_quality_mesh_patterns[MAX_MESH_SPEED + 1][MAX_MESH_STEP] = {
26 { { 64, 8 }, { 28, 4 }, { 15, 1 }, { 7, 1 } },
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 { { 64, 16 }, { 24, 8 }, { 12, 4 }, { 7, 1 } },
31 { { 64, 16 }, { 24, 8 }, { 12, 4 }, { 7, 1 } },
33 static unsigned char good_quality_max_mesh_pct[MAX_MESH_SPEED + 1] = {
37 // Intra only frames, golden frames (except alt ref overlays) and
38 // alt ref frames tend to be coded at a higher than ambient quality
39 static int frame_is_boosted(const VP9_COMP *cpi) {
40 return frame_is_kf_gf_arf(cpi) || vp9_is_upper_layer_key_frame(cpi);
43 // Sets a partition size down to which the auto partition code will always
44 // search (can go lower), based on the image dimensions. The logic here
45 // is that the extent to which ringing artefacts are offensive, depends
46 // partly on the screen area that over which they propogate. Propogation is
47 // limited by transform block size but the screen area take up by a given block
48 // size will be larger for a small image format stretched to full screen.
49 static BLOCK_SIZE set_partition_min_limit(VP9_COMMON *const cm) {
50 unsigned int screen_area = (cm->width * cm->height);
52 // Select block size based on image format size.
53 if (screen_area < 1280 * 720) {
54 // Formats smaller in area than 720P
56 } else if (screen_area < 1920 * 1080) {
57 // Format >= 720P and < 1080P
60 // Formats 1080P and up
65 static void set_good_speed_feature_framesize_dependent(VP9_COMP *cpi,
68 VP9_COMMON *const cm = &cpi->common;
71 sf->partition_search_breakout_thr.dist = (1 << 20);
72 sf->partition_search_breakout_thr.rate = 80;
74 // Currently, the machine-learning based partition search early termination
75 // is only used while VPXMIN(cm->width, cm->height) >= 480 and speed = 0.
76 // TODO(yunqingwang): Re-enable when test failures are fixed.
77 // if (VPXMIN(cm->width, cm->height) >= 480) {
78 // sf->ml_partition_search_early_termination = 1;
82 sf->ml_partition_search_early_termination = 0;
84 if (VPXMIN(cm->width, cm->height) >= 720) {
85 sf->disable_split_mask =
86 cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT;
87 sf->partition_search_breakout_thr.dist = (1 << 23);
89 sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
90 sf->partition_search_breakout_thr.dist = (1 << 21);
95 if (VPXMIN(cm->width, cm->height) >= 720) {
96 sf->disable_split_mask =
97 cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT;
98 sf->adaptive_pred_interp_filter = 0;
99 sf->partition_search_breakout_thr.dist = (1 << 24);
100 sf->partition_search_breakout_thr.rate = 120;
102 sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY;
103 sf->partition_search_breakout_thr.dist = (1 << 22);
104 sf->partition_search_breakout_thr.rate = 100;
106 sf->rd_auto_partition_min_limit = set_partition_min_limit(cm);
108 // Use a set of speed features for 4k videos.
109 if (VPXMIN(cm->width, cm->height) >= 2160) {
110 sf->use_square_partition_only = 1;
111 sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
112 sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC;
113 sf->alt_ref_search_fp = 1;
114 sf->cb_pred_filter_search = 1;
115 sf->adaptive_interp_filter_search = 1;
116 sf->disable_split_mask = DISABLE_ALL_SPLIT;
121 if (VPXMIN(cm->width, cm->height) >= 720) {
122 sf->disable_split_mask = DISABLE_ALL_SPLIT;
123 sf->schedule_mode_search = cm->base_qindex < 220 ? 1 : 0;
124 sf->partition_search_breakout_thr.dist = (1 << 25);
125 sf->partition_search_breakout_thr.rate = 200;
127 sf->max_intra_bsize = BLOCK_32X32;
128 sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT;
129 sf->schedule_mode_search = cm->base_qindex < 175 ? 1 : 0;
130 sf->partition_search_breakout_thr.dist = (1 << 23);
131 sf->partition_search_breakout_thr.rate = 120;
135 // If this is a two pass clip that fits the criteria for animated or
136 // graphics content then reset disable_split_mask for speeds 1-4.
137 // Also if the image edge is internal to the coded area.
138 if ((speed >= 1) && (cpi->oxcf.pass == 2) &&
139 ((cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ||
140 (vp9_internal_image_edge(cpi)))) {
141 sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
145 sf->partition_search_breakout_thr.rate = 300;
146 if (VPXMIN(cm->width, cm->height) >= 720) {
147 sf->partition_search_breakout_thr.dist = (1 << 26);
149 sf->partition_search_breakout_thr.dist = (1 << 24);
151 sf->disable_split_mask = DISABLE_ALL_SPLIT;
155 sf->partition_search_breakout_thr.rate = 500;
159 static double tx_dom_thresholds[6] = { 99.0, 14.0, 12.0, 8.0, 4.0, 0.0 };
160 static double qopt_thresholds[6] = { 99.0, 12.0, 10.0, 4.0, 2.0, 0.0 };
162 static void set_good_speed_feature_framesize_independent(VP9_COMP *cpi,
166 const int boosted = frame_is_boosted(cpi);
168 sf->tx_size_search_breakout = 1;
169 sf->adaptive_rd_thresh = 1;
170 sf->allow_skip_recode = 1;
171 sf->less_rectangular_check = 1;
172 sf->use_square_partition_only = !frame_is_boosted(cpi);
173 sf->use_square_only_threshold = BLOCK_16X16;
176 if (cpi->oxcf.pass == 2) {
177 TWO_PASS *const twopass = &cpi->twopass;
178 if ((twopass->fr_content_type == FC_GRAPHICS_ANIMATION) ||
179 vp9_internal_image_edge(cpi)) {
180 sf->use_square_partition_only = !frame_is_boosted(cpi);
182 sf->use_square_partition_only = !frame_is_intra_only(cm);
185 sf->use_square_partition_only = !frame_is_intra_only(cm);
188 sf->allow_txfm_domain_distortion = 1;
189 sf->tx_domain_thresh = tx_dom_thresholds[(speed < 6) ? speed : 5];
190 sf->allow_quant_coeff_opt = sf->optimize_coefficients;
191 sf->quant_opt_thresh = qopt_thresholds[(speed < 6) ? speed : 5];
193 sf->use_square_only_threshold = BLOCK_4X4;
194 sf->less_rectangular_check = 1;
196 sf->use_rd_breakout = 1;
197 sf->adaptive_motion_search = 1;
198 sf->mv.auto_mv_step_size = 1;
199 sf->adaptive_rd_thresh = 2;
200 sf->mv.subpel_iters_per_step = 1;
201 sf->mode_skip_start = 10;
202 sf->adaptive_pred_interp_filter = 1;
205 sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
206 sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
207 sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
208 sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
210 sf->recode_tolerance_low = 15;
211 sf->recode_tolerance_high = 30;
215 sf->recode_loop = ALLOW_RECODE_KFARFGF;
216 sf->tx_size_search_method =
217 frame_is_boosted(cpi) ? USE_FULL_RD : USE_LARGESTALL;
219 // Reference masking is not supported in dynamic scaling mode.
220 sf->reference_masking = cpi->oxcf.resize_mode != RESIZE_DYNAMIC ? 1 : 0;
222 sf->mode_search_skip_flags =
223 (cm->frame_type == KEY_FRAME)
225 : FLAG_SKIP_INTRA_DIRMISMATCH | FLAG_SKIP_INTRA_BESTINTER |
226 FLAG_SKIP_COMP_BESTINTRA | FLAG_SKIP_INTRA_LOWVAR;
227 sf->disable_filter_search_var_thresh = 100;
228 sf->comp_inter_joint_search_thresh = BLOCK_SIZES;
229 sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX;
230 sf->allow_partition_search_skip = 1;
231 sf->recode_tolerance_low = 15;
232 sf->recode_tolerance_high = 45;
236 sf->use_square_partition_only = !frame_is_intra_only(cm);
237 sf->tx_size_search_method =
238 frame_is_intra_only(cm) ? USE_FULL_RD : USE_LARGESTALL;
239 sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED;
240 sf->adaptive_pred_interp_filter = 0;
241 sf->adaptive_mode_search = 1;
242 sf->cb_partition_search = !boosted;
243 sf->cb_pred_filter_search = 1;
244 sf->alt_ref_search_fp = 1;
245 sf->recode_loop = ALLOW_RECODE_KFMAXBW;
246 sf->adaptive_rd_thresh = 3;
247 sf->mode_skip_start = 6;
248 sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
249 sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC;
250 sf->adaptive_interp_filter_search = 1;
254 sf->use_square_partition_only = 1;
255 sf->tx_size_search_method = USE_LARGESTALL;
256 sf->mv.search_method = BIGDIA;
257 sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED_MORE;
258 sf->adaptive_rd_thresh = 4;
259 if (cm->frame_type != KEY_FRAME)
260 sf->mode_search_skip_flags |= FLAG_EARLY_TERMINATE;
261 sf->disable_filter_search_var_thresh = 200;
262 sf->use_lp32x32fdct = 1;
263 sf->use_fast_coef_updates = ONE_LOOP_REDUCED;
264 sf->use_fast_coef_costing = 1;
265 sf->motion_field_mode_search = !boosted;
270 sf->optimize_coefficients = 0;
271 sf->mv.search_method = HEX;
272 sf->disable_filter_search_var_thresh = 500;
273 for (i = 0; i < TX_SIZES; ++i) {
274 sf->intra_y_mode_mask[i] = INTRA_DC;
275 sf->intra_uv_mode_mask[i] = INTRA_DC;
277 sf->mv.reduce_first_step_size = 1;
278 sf->simple_model_rd_from_var = 1;
282 static void set_rt_speed_feature_framesize_dependent(VP9_COMP *cpi,
285 VP9_COMMON *const cm = &cpi->common;
288 if (VPXMIN(cm->width, cm->height) >= 720) {
289 sf->disable_split_mask =
290 cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT;
292 sf->disable_split_mask = DISABLE_COMPOUND_SPLIT;
297 if (VPXMIN(cm->width, cm->height) >= 720) {
298 sf->disable_split_mask =
299 cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT;
301 sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY;
306 sf->partition_search_breakout_thr.rate = 200;
307 if (VPXMIN(cm->width, cm->height) >= 720) {
308 sf->partition_search_breakout_thr.dist = (1 << 25);
310 sf->partition_search_breakout_thr.dist = (1 << 23);
315 sf->encode_breakout_thresh =
316 (VPXMIN(cm->width, cm->height) >= 720) ? 800 : 300;
320 static void set_rt_speed_feature_framesize_independent(
321 VP9_COMP *cpi, SPEED_FEATURES *sf, int speed, vp9e_tune_content content) {
322 VP9_COMMON *const cm = &cpi->common;
323 const int is_keyframe = cm->frame_type == KEY_FRAME;
324 const int frames_since_key = is_keyframe ? 0 : cpi->rc.frames_since_key;
325 sf->static_segmentation = 0;
326 sf->adaptive_rd_thresh = 1;
327 sf->use_fast_coef_costing = 1;
328 sf->allow_exhaustive_searches = 0;
329 sf->exhaustive_searches_thresh = INT_MAX;
331 sf->copy_partition_flag = 0;
332 sf->use_source_sad = 0;
333 sf->use_simple_block_yrd = 0;
336 sf->allow_txfm_domain_distortion = 1;
337 sf->tx_domain_thresh = 0.0;
338 sf->allow_quant_coeff_opt = 0;
339 sf->quant_opt_thresh = 0.0;
340 sf->use_square_partition_only = !frame_is_intra_only(cm);
341 sf->less_rectangular_check = 1;
342 sf->tx_size_search_method =
343 frame_is_intra_only(cm) ? USE_FULL_RD : USE_LARGESTALL;
345 sf->use_rd_breakout = 1;
347 sf->adaptive_motion_search = 1;
348 sf->adaptive_pred_interp_filter = 1;
349 sf->mv.auto_mv_step_size = 1;
350 sf->adaptive_rd_thresh = 2;
351 sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V;
352 sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V;
353 sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V;
357 sf->mode_search_skip_flags =
358 (cm->frame_type == KEY_FRAME)
360 : FLAG_SKIP_INTRA_DIRMISMATCH | FLAG_SKIP_INTRA_BESTINTER |
361 FLAG_SKIP_COMP_BESTINTRA | FLAG_SKIP_INTRA_LOWVAR;
362 sf->adaptive_pred_interp_filter = 2;
364 // Reference masking only enabled for 1 spatial layer, and if none of the
365 // references have been scaled. The latter condition needs to be checked
366 // for external or internal dynamic resize.
367 sf->reference_masking = (cpi->svc.number_spatial_layers == 1);
368 if (sf->reference_masking == 1 &&
369 (cpi->external_resize == 1 ||
370 cpi->oxcf.resize_mode == RESIZE_DYNAMIC)) {
371 MV_REFERENCE_FRAME ref_frame;
372 static const int flag_list[4] = { 0, VP9_LAST_FLAG, VP9_GOLD_FLAG,
374 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
375 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
376 if (yv12 != NULL && (cpi->ref_frame_flags & flag_list[ref_frame])) {
377 const struct scale_factors *const scale_fac =
378 &cm->frame_refs[ref_frame - 1].sf;
379 if (vp9_is_scaled(scale_fac)) sf->reference_masking = 0;
384 sf->disable_filter_search_var_thresh = 50;
385 sf->comp_inter_joint_search_thresh = BLOCK_SIZES;
386 sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX;
387 sf->lf_motion_threshold = LOW_MOTION_THRESHOLD;
388 sf->adjust_partitioning_from_last_frame = 1;
389 sf->last_partitioning_redo_frequency = 3;
390 sf->use_lp32x32fdct = 1;
391 sf->mode_skip_start = 11;
392 sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V;
396 sf->use_square_partition_only = 1;
397 sf->disable_filter_search_var_thresh = 100;
398 sf->use_uv_intra_rd_estimate = 1;
399 sf->skip_encode_sb = 1;
400 sf->mv.subpel_iters_per_step = 1;
401 sf->adaptive_rd_thresh = 4;
402 sf->mode_skip_start = 6;
403 sf->allow_skip_recode = 0;
404 sf->optimize_coefficients = 0;
405 sf->disable_split_mask = DISABLE_ALL_SPLIT;
406 sf->lpf_pick = LPF_PICK_FROM_Q;
411 sf->last_partitioning_redo_frequency = 4;
412 sf->adaptive_rd_thresh = 5;
413 sf->use_fast_coef_costing = 0;
414 sf->auto_min_max_partition_size = STRICT_NEIGHBORING_MIN_MAX;
415 sf->adjust_partitioning_from_last_frame =
416 cm->last_frame_type != cm->frame_type ||
417 (0 == (frames_since_key + 1) % sf->last_partitioning_redo_frequency);
418 sf->mv.subpel_force_stop = 1;
419 for (i = 0; i < TX_SIZES; i++) {
420 sf->intra_y_mode_mask[i] = INTRA_DC_H_V;
421 sf->intra_uv_mode_mask[i] = INTRA_DC;
423 sf->intra_y_mode_mask[TX_32X32] = INTRA_DC;
424 sf->frame_parameter_update = 0;
425 sf->mv.search_method = FAST_HEX;
427 sf->inter_mode_mask[BLOCK_32X32] = INTER_NEAREST_NEAR_NEW;
428 sf->inter_mode_mask[BLOCK_32X64] = INTER_NEAREST;
429 sf->inter_mode_mask[BLOCK_64X32] = INTER_NEAREST;
430 sf->inter_mode_mask[BLOCK_64X64] = INTER_NEAREST;
431 sf->max_intra_bsize = BLOCK_32X32;
432 sf->allow_skip_recode = 1;
436 sf->use_quant_fp = !is_keyframe;
437 sf->auto_min_max_partition_size =
438 is_keyframe ? RELAXED_NEIGHBORING_MIN_MAX : STRICT_NEIGHBORING_MIN_MAX;
439 sf->default_max_partition_size = BLOCK_32X32;
440 sf->default_min_partition_size = BLOCK_8X8;
441 sf->force_frame_boost =
443 (frames_since_key % (sf->last_partitioning_redo_frequency << 1) == 1);
444 sf->max_delta_qindex = is_keyframe ? 20 : 15;
445 sf->partition_search_type = REFERENCE_PARTITION;
446 if (cpi->oxcf.rc_mode == VPX_VBR && cpi->oxcf.lag_in_frames > 0 &&
447 cpi->rc.is_src_frame_alt_ref) {
448 sf->partition_search_type = VAR_BASED_PARTITION;
450 sf->use_nonrd_pick_mode = 1;
451 sf->allow_skip_recode = 0;
452 sf->inter_mode_mask[BLOCK_32X32] = INTER_NEAREST_NEW_ZERO;
453 sf->inter_mode_mask[BLOCK_32X64] = INTER_NEAREST_NEW_ZERO;
454 sf->inter_mode_mask[BLOCK_64X32] = INTER_NEAREST_NEW_ZERO;
455 sf->inter_mode_mask[BLOCK_64X64] = INTER_NEAREST_NEW_ZERO;
456 sf->adaptive_rd_thresh = 2;
457 // This feature is only enabled when partition search is disabled.
458 sf->reuse_inter_pred_sby = 1;
459 sf->coeff_prob_appx_step = 4;
460 sf->use_fast_coef_updates = is_keyframe ? TWO_LOOP : ONE_LOOP_REDUCED;
461 sf->mode_search_skip_flags = FLAG_SKIP_INTRA_DIRMISMATCH;
462 sf->tx_size_search_method = is_keyframe ? USE_LARGESTALL : USE_TX_8X8;
463 sf->simple_model_rd_from_var = 1;
464 if (cpi->oxcf.rc_mode == VPX_VBR) sf->mv.search_method = NSTEP;
468 if (content == VP9E_CONTENT_SCREEN) {
469 for (i = 0; i < BLOCK_SIZES; ++i)
470 sf->intra_y_mode_bsize_mask[i] = INTRA_DC_TM_H_V;
472 for (i = 0; i < BLOCK_SIZES; ++i)
474 sf->intra_y_mode_bsize_mask[i] = INTRA_DC;
476 // Use H and V intra mode for block sizes <= 16X16.
477 sf->intra_y_mode_bsize_mask[i] = INTRA_DC_H_V;
480 if (content == VP9E_CONTENT_SCREEN) {
481 sf->short_circuit_flat_blocks = 1;
483 if (cpi->oxcf.rc_mode == VPX_CBR &&
484 cpi->oxcf.content != VP9E_CONTENT_SCREEN) {
485 sf->limit_newmv_early_exit = 1;
486 if (!cpi->use_svc) sf->bias_golden = 1;
491 sf->partition_search_type = VAR_BASED_PARTITION;
492 // Turn on this to use non-RD key frame coding mode.
493 sf->use_nonrd_pick_mode = 1;
494 sf->mv.search_method = NSTEP;
495 sf->mv.reduce_first_step_size = 1;
496 sf->skip_encode_sb = 0;
497 if (cpi->oxcf.rc_mode == VPX_CBR && content != VP9E_CONTENT_SCREEN) {
498 // Enable short circuit for low temporal variance.
499 sf->short_circuit_low_temp_var = 1;
501 if (cpi->use_svc) sf->base_mv_aggressive = 1;
505 sf->adaptive_rd_thresh = 3;
506 sf->mv.search_method = FAST_DIAMOND;
507 sf->mv.fullpel_search_step_param = 10;
508 if (cpi->svc.number_temporal_layers > 2 &&
509 cpi->svc.temporal_layer_id == 0) {
510 sf->mv.search_method = NSTEP;
511 sf->mv.fullpel_search_step_param = 6;
513 if (!cpi->external_resize) sf->use_source_sad = 1;
514 if (sf->use_source_sad) {
515 // For SVC allocate for top layer.
516 if (cpi->content_state_sb == NULL &&
518 cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1)) {
519 cpi->content_state_sb = (uint8_t *)vpx_calloc(
520 (cm->mi_stride >> 3) * ((cm->mi_rows >> 3) + 1), sizeof(uint8_t));
526 sf->adaptive_rd_thresh = 4;
527 // Enable partition copy
528 if (!cpi->use_svc && !cpi->resize_pending && cpi->resize_state == ORIG &&
529 !cpi->external_resize && cpi->oxcf.resize_mode == RESIZE_NONE)
530 sf->copy_partition_flag = 1;
532 if (sf->copy_partition_flag) {
533 cpi->max_copied_frame = 4;
534 if (cpi->prev_partition == NULL) {
535 cpi->prev_partition = (BLOCK_SIZE *)vpx_calloc(
536 cm->mi_stride * cm->mi_rows, sizeof(BLOCK_SIZE));
538 if (cpi->prev_segment_id == NULL) {
539 cpi->prev_segment_id = (int8_t *)vpx_calloc(
540 (cm->mi_stride >> 3) * ((cm->mi_rows >> 3) + 1), sizeof(int8_t));
542 if (cpi->prev_variance_low == NULL) {
543 cpi->prev_variance_low = (uint8_t *)vpx_calloc(
544 (cm->mi_stride >> 3) * ((cm->mi_rows >> 3) + 1) * 25,
547 if (cpi->copied_frame_cnt == NULL) {
548 cpi->copied_frame_cnt = (uint8_t *)vpx_calloc(
549 (cm->mi_stride >> 3) * ((cm->mi_rows >> 3) + 1), sizeof(uint8_t));
553 sf->mv.subpel_force_stop = (content == VP9E_CONTENT_SCREEN) ? 3 : 2;
554 if (content == VP9E_CONTENT_SCREEN) sf->lpf_pick = LPF_PICK_MINIMAL_LPF;
555 // Only keep INTRA_DC mode for speed 8.
558 for (i = 0; i < BLOCK_SIZES; ++i)
559 sf->intra_y_mode_bsize_mask[i] = INTRA_DC;
561 if (!cpi->use_svc && cpi->oxcf.rc_mode == VPX_CBR &&
562 content != VP9E_CONTENT_SCREEN) {
563 // More aggressive short circuit for speed 8.
564 sf->short_circuit_low_temp_var = 3;
565 // Use level 2 for noisey cases as there is a regression in some
566 // noisy clips with level 3.
567 if (cpi->noise_estimate.enabled && cm->width >= 1280 &&
569 NOISE_LEVEL noise_level =
570 vp9_noise_estimate_extract_level(&cpi->noise_estimate);
571 if (noise_level >= kMedium) sf->short_circuit_low_temp_var = 2;
573 // Since the short_circuit_low_temp_var is used, reduce the
574 // adaptive_rd_thresh level.
575 sf->adaptive_rd_thresh = 2;
577 sf->limit_newmv_early_exit = 0;
578 if (cm->width > 640 && cm->height > 480) sf->use_simple_block_yrd = 1;
580 // Turn off adaptive_rd_thresh if row_mt is on for all the non-rd paths. This
581 // causes too many locks in realtime mode in certain platforms (Android ARM,
583 if (speed >= 5 && cpi->row_mt && cpi->num_workers > 1) {
584 sf->adaptive_rd_thresh = 0;
588 void vp9_set_speed_features_framesize_dependent(VP9_COMP *cpi) {
589 SPEED_FEATURES *const sf = &cpi->sf;
590 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
591 RD_OPT *const rd = &cpi->rd;
594 // best quality defaults
595 // Some speed-up features even for best quality as minimal impact on quality.
596 sf->partition_search_breakout_thr.dist = (1 << 19);
597 sf->partition_search_breakout_thr.rate = 80;
598 sf->ml_partition_search_early_termination = 0;
600 if (oxcf->mode == REALTIME) {
601 set_rt_speed_feature_framesize_dependent(cpi, sf, oxcf->speed);
602 } else if (oxcf->mode == GOOD) {
603 set_good_speed_feature_framesize_dependent(cpi, sf, oxcf->speed);
606 if (sf->disable_split_mask == DISABLE_ALL_SPLIT) {
607 sf->adaptive_pred_interp_filter = 0;
610 if (cpi->encode_breakout && oxcf->mode == REALTIME &&
611 sf->encode_breakout_thresh > cpi->encode_breakout) {
612 cpi->encode_breakout = sf->encode_breakout_thresh;
615 // Check for masked out split cases.
616 for (i = 0; i < MAX_REFS; ++i) {
617 if (sf->disable_split_mask & (1 << i)) {
618 rd->thresh_mult_sub8x8[i] = INT_MAX;
622 // With row based multi-threading, the following speed features
623 // have to be disabled to guarantee that bitstreams encoded with single thread
624 // and multiple threads match
625 if (cpi->oxcf.row_mt_bit_exact) {
626 sf->adaptive_rd_thresh = 0;
627 sf->allow_exhaustive_searches = 0;
628 sf->adaptive_pred_interp_filter = 0;
632 void vp9_set_speed_features_framesize_independent(VP9_COMP *cpi) {
633 SPEED_FEATURES *const sf = &cpi->sf;
634 VP9_COMMON *const cm = &cpi->common;
635 MACROBLOCK *const x = &cpi->td.mb;
636 const VP9EncoderConfig *const oxcf = &cpi->oxcf;
639 // best quality defaults
640 sf->frame_parameter_update = 1;
641 sf->mv.search_method = NSTEP;
642 sf->recode_loop = ALLOW_RECODE_FIRST;
643 sf->mv.subpel_search_method = SUBPEL_TREE;
644 sf->mv.subpel_iters_per_step = 2;
645 sf->mv.subpel_force_stop = 0;
646 sf->optimize_coefficients = !is_lossless_requested(&cpi->oxcf);
647 sf->mv.reduce_first_step_size = 0;
648 sf->coeff_prob_appx_step = 1;
649 sf->mv.auto_mv_step_size = 0;
650 sf->mv.fullpel_search_step_param = 6;
651 sf->comp_inter_joint_search_thresh = BLOCK_4X4;
652 sf->tx_size_search_method = USE_FULL_RD;
653 sf->use_lp32x32fdct = 0;
654 sf->adaptive_motion_search = 0;
655 sf->adaptive_pred_interp_filter = 0;
656 sf->adaptive_mode_search = 0;
657 sf->cb_pred_filter_search = 0;
658 sf->cb_partition_search = 0;
659 sf->motion_field_mode_search = 0;
660 sf->alt_ref_search_fp = 0;
661 sf->use_quant_fp = 0;
662 sf->reference_masking = 0;
663 sf->partition_search_type = SEARCH_PARTITION;
664 sf->less_rectangular_check = 0;
665 sf->use_square_partition_only = 0;
666 sf->use_square_only_threshold = BLOCK_SIZES;
667 sf->auto_min_max_partition_size = NOT_IN_USE;
668 sf->rd_auto_partition_min_limit = BLOCK_4X4;
669 sf->default_max_partition_size = BLOCK_64X64;
670 sf->default_min_partition_size = BLOCK_4X4;
671 sf->adjust_partitioning_from_last_frame = 0;
672 sf->last_partitioning_redo_frequency = 4;
673 sf->disable_split_mask = 0;
674 sf->mode_search_skip_flags = 0;
675 sf->force_frame_boost = 0;
676 sf->max_delta_qindex = 0;
677 sf->disable_filter_search_var_thresh = 0;
678 sf->adaptive_interp_filter_search = 0;
679 sf->allow_partition_search_skip = 0;
680 sf->allow_txfm_domain_distortion = 0;
681 sf->tx_domain_thresh = 99.0;
682 sf->allow_quant_coeff_opt = sf->optimize_coefficients;
683 sf->quant_opt_thresh = 99.0;
686 for (i = 0; i < TX_SIZES; i++) {
687 sf->intra_y_mode_mask[i] = INTRA_ALL;
688 sf->intra_uv_mode_mask[i] = INTRA_ALL;
690 sf->use_rd_breakout = 0;
691 sf->skip_encode_sb = 0;
692 sf->use_uv_intra_rd_estimate = 0;
693 sf->allow_skip_recode = 0;
694 sf->lpf_pick = LPF_PICK_FROM_FULL_IMAGE;
695 sf->use_fast_coef_updates = TWO_LOOP;
696 sf->use_fast_coef_costing = 0;
697 sf->mode_skip_start = MAX_MODES; // Mode index at which mode skip mask set
698 sf->schedule_mode_search = 0;
699 sf->use_nonrd_pick_mode = 0;
700 for (i = 0; i < BLOCK_SIZES; ++i) sf->inter_mode_mask[i] = INTER_ALL;
701 sf->max_intra_bsize = BLOCK_64X64;
702 sf->reuse_inter_pred_sby = 0;
703 // This setting only takes effect when partition_search_type is set
704 // to FIXED_PARTITION.
705 sf->always_this_block_size = BLOCK_16X16;
706 sf->search_type_check_frequency = 50;
707 sf->encode_breakout_thresh = 0;
708 // Recode loop tolerance %.
709 sf->recode_tolerance_low = 12;
710 sf->recode_tolerance_high = 25;
711 sf->default_interp_filter = SWITCHABLE;
712 sf->simple_model_rd_from_var = 0;
713 sf->short_circuit_flat_blocks = 0;
714 sf->short_circuit_low_temp_var = 0;
715 sf->limit_newmv_early_exit = 0;
717 sf->base_mv_aggressive = 0;
719 // Some speed-up features even for best quality as minimal impact on quality.
720 sf->adaptive_rd_thresh = 1;
721 sf->tx_size_search_breakout = 1;
723 if (oxcf->mode == REALTIME)
724 set_rt_speed_feature_framesize_independent(cpi, sf, oxcf->speed,
726 else if (oxcf->mode == GOOD)
727 set_good_speed_feature_framesize_independent(cpi, cm, sf, oxcf->speed);
729 cpi->full_search_sad = vp9_full_search_sad;
730 cpi->diamond_search_sad = vp9_diamond_search_sad;
732 sf->allow_exhaustive_searches = 1;
733 if (oxcf->mode == BEST) {
734 if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION)
735 sf->exhaustive_searches_thresh = (1 << 20);
737 sf->exhaustive_searches_thresh = (1 << 21);
738 sf->max_exaustive_pct = 100;
739 for (i = 0; i < MAX_MESH_STEP; ++i) {
740 sf->mesh_patterns[i].range = best_quality_mesh_pattern[i].range;
741 sf->mesh_patterns[i].interval = best_quality_mesh_pattern[i].interval;
744 int speed = (oxcf->speed > MAX_MESH_SPEED) ? MAX_MESH_SPEED : oxcf->speed;
745 if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION)
746 sf->exhaustive_searches_thresh = (1 << 22);
748 sf->exhaustive_searches_thresh = (1 << 23);
749 sf->max_exaustive_pct = good_quality_max_mesh_pct[speed];
751 sf->exhaustive_searches_thresh = sf->exhaustive_searches_thresh << 1;
753 for (i = 0; i < MAX_MESH_STEP; ++i) {
754 sf->mesh_patterns[i].range = good_quality_mesh_patterns[speed][i].range;
755 sf->mesh_patterns[i].interval =
756 good_quality_mesh_patterns[speed][i].interval;
760 // Slow quant, dct and trellis not worthwhile for first pass
761 // so make sure they are always turned off.
762 if (oxcf->pass == 1) sf->optimize_coefficients = 0;
764 // No recode for 1 pass.
765 if (oxcf->pass == 0) {
766 sf->recode_loop = DISALLOW_RECODE;
767 sf->optimize_coefficients = 0;
770 if (sf->mv.subpel_force_stop == 3) {
772 cpi->find_fractional_mv_step = vp9_skip_sub_pixel_tree;
773 } else if (sf->mv.subpel_search_method == SUBPEL_TREE) {
774 cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree;
775 } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED) {
776 cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned;
777 } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED_MORE) {
778 cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned_more;
779 } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED_EVENMORE) {
780 cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned_evenmore;
783 x->optimize = sf->optimize_coefficients == 1 && oxcf->pass != 1;
785 x->min_partition_size = sf->default_min_partition_size;
786 x->max_partition_size = sf->default_max_partition_size;
788 if (!cpi->oxcf.frame_periodic_boost) {
789 sf->max_delta_qindex = 0;
792 // With row based multi-threading, the following speed features
793 // have to be disabled to guarantee that bitstreams encoded with single thread
794 // and multiple threads match
795 if (cpi->oxcf.row_mt_bit_exact) {
796 sf->adaptive_rd_thresh = 0;
797 sf->allow_exhaustive_searches = 0;
798 sf->adaptive_pred_interp_filter = 0;