/src/libvpx/vp9/encoder/vp9_speed_features.c
Line | Count | Source (jump to first uncovered line) |
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 | | // Define 2 mesh density levels for FC_GRAPHICS_ANIMATION content type and non |
20 | | // FC_GRAPHICS_ANIMATION content type. |
21 | | static MESH_PATTERN best_quality_mesh_pattern[2][MAX_MESH_STEP] = { |
22 | | { { 64, 4 }, { 28, 2 }, { 15, 1 }, { 7, 1 } }, |
23 | | { { 64, 8 }, { 28, 4 }, { 15, 1 }, { 7, 1 } }, |
24 | | }; |
25 | | |
26 | | #if !CONFIG_REALTIME_ONLY |
27 | | // Define 3 mesh density levels to control the number of searches. |
28 | | #define MESH_DENSITY_LEVELS 3 |
29 | | static MESH_PATTERN |
30 | | good_quality_mesh_patterns[MESH_DENSITY_LEVELS][MAX_MESH_STEP] = { |
31 | | { { 64, 8 }, { 28, 4 }, { 15, 1 }, { 7, 1 } }, |
32 | | { { 64, 8 }, { 14, 2 }, { 7, 1 }, { 7, 1 } }, |
33 | | { { 64, 16 }, { 24, 8 }, { 12, 4 }, { 7, 1 } }, |
34 | | }; |
35 | | |
36 | | // Intra only frames, golden frames (except alt ref overlays) and |
37 | | // alt ref frames tend to be coded at a higher than ambient quality |
38 | 45.2k | static int frame_is_boosted(const VP9_COMP *cpi) { |
39 | 45.2k | return frame_is_kf_gf_arf(cpi); |
40 | 45.2k | } |
41 | | |
42 | | // Sets a partition size down to which the auto partition code will always |
43 | | // search (can go lower), based on the image dimensions. The logic here |
44 | | // is that the extent to which ringing artefacts are offensive, depends |
45 | | // partly on the screen area that over which they propagate. Propagation is |
46 | | // limited by transform block size but the screen area take up by a given block |
47 | | // size will be larger for a small image format stretched to full screen. |
48 | 0 | static BLOCK_SIZE set_partition_min_limit(VP9_COMMON *const cm) { |
49 | 0 | unsigned int screen_area = (cm->width * cm->height); |
50 | | |
51 | | // Select block size based on image format size. |
52 | 0 | if (screen_area < 1280 * 720) { |
53 | | // Formats smaller in area than 720P |
54 | 0 | return BLOCK_4X4; |
55 | 0 | } else if (screen_area < 1920 * 1080) { |
56 | | // Format >= 720P and < 1080P |
57 | 0 | return BLOCK_8X8; |
58 | 0 | } else { |
59 | | // Formats 1080P and up |
60 | 0 | return BLOCK_16X16; |
61 | 0 | } |
62 | 0 | } |
63 | | |
64 | | static void set_good_speed_feature_framesize_dependent(VP9_COMP *cpi, |
65 | | SPEED_FEATURES *sf, |
66 | 22.6k | int speed) { |
67 | 22.6k | VP9_COMMON *const cm = &cpi->common; |
68 | 22.6k | const int min_frame_size = VPXMIN(cm->width, cm->height); |
69 | 22.6k | const int is_480p_or_larger = min_frame_size >= 480; |
70 | 22.6k | const int is_720p_or_larger = min_frame_size >= 720; |
71 | 22.6k | const int is_1080p_or_larger = min_frame_size >= 1080; |
72 | 22.6k | const int is_2160p_or_larger = min_frame_size >= 2160; |
73 | 22.6k | const int boosted = frame_is_boosted(cpi); |
74 | | |
75 | | // speed 0 features |
76 | 22.6k | sf->partition_search_breakout_thr.dist = (1 << 20); |
77 | 22.6k | sf->partition_search_breakout_thr.rate = 80; |
78 | 22.6k | sf->use_square_only_thresh_high = BLOCK_SIZES; |
79 | 22.6k | sf->use_square_only_thresh_low = BLOCK_4X4; |
80 | | |
81 | 22.6k | if (is_480p_or_larger) { |
82 | | // Currently, the machine-learning based partition search early termination |
83 | | // is only used while VPXMIN(cm->width, cm->height) >= 480 and speed = 0. |
84 | 13 | sf->rd_ml_partition.search_early_termination = 1; |
85 | 13 | sf->recode_tolerance_high = 45; |
86 | 22.6k | } else { |
87 | 22.6k | sf->use_square_only_thresh_high = BLOCK_32X32; |
88 | 22.6k | } |
89 | 22.6k | if (is_720p_or_larger) { |
90 | 0 | sf->alt_ref_search_fp = 1; |
91 | 0 | } |
92 | | |
93 | 22.6k | if (!is_1080p_or_larger) { |
94 | 22.6k | sf->rd_ml_partition.search_breakout = 1; |
95 | 22.6k | if (is_720p_or_larger) { |
96 | 0 | sf->rd_ml_partition.search_breakout_thresh[0] = 0.0f; |
97 | 0 | sf->rd_ml_partition.search_breakout_thresh[1] = 0.0f; |
98 | 0 | sf->rd_ml_partition.search_breakout_thresh[2] = 0.0f; |
99 | 22.6k | } else { |
100 | 22.6k | sf->rd_ml_partition.search_breakout_thresh[0] = 2.5f; |
101 | 22.6k | sf->rd_ml_partition.search_breakout_thresh[1] = 1.5f; |
102 | 22.6k | sf->rd_ml_partition.search_breakout_thresh[2] = 1.5f; |
103 | 22.6k | } |
104 | 22.6k | } |
105 | | |
106 | 22.6k | if (!is_720p_or_larger) { |
107 | 22.6k | if (is_480p_or_larger) |
108 | 13 | sf->prune_single_mode_based_on_mv_diff_mode_rate = boosted ? 0 : 1; |
109 | 22.6k | else |
110 | 22.6k | sf->prune_single_mode_based_on_mv_diff_mode_rate = 1; |
111 | 22.6k | } |
112 | | |
113 | 22.6k | if (speed >= 1) { |
114 | 19.6k | sf->rd_ml_partition.search_early_termination = 0; |
115 | 19.6k | sf->rd_ml_partition.search_breakout = 1; |
116 | 19.6k | if (is_480p_or_larger) |
117 | 5 | sf->use_square_only_thresh_high = BLOCK_64X64; |
118 | 19.6k | else |
119 | 19.6k | sf->use_square_only_thresh_high = BLOCK_32X32; |
120 | 19.6k | sf->use_square_only_thresh_low = BLOCK_16X16; |
121 | 19.6k | if (is_720p_or_larger) { |
122 | 0 | sf->disable_split_mask = |
123 | 0 | cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT; |
124 | 0 | sf->partition_search_breakout_thr.dist = (1 << 22); |
125 | 0 | sf->rd_ml_partition.search_breakout_thresh[0] = -5.0f; |
126 | 0 | sf->rd_ml_partition.search_breakout_thresh[1] = -5.0f; |
127 | 0 | sf->rd_ml_partition.search_breakout_thresh[2] = -9.0f; |
128 | 19.6k | } else { |
129 | 19.6k | sf->disable_split_mask = DISABLE_COMPOUND_SPLIT; |
130 | 19.6k | sf->partition_search_breakout_thr.dist = (1 << 21); |
131 | 19.6k | sf->rd_ml_partition.search_breakout_thresh[0] = -1.0f; |
132 | 19.6k | sf->rd_ml_partition.search_breakout_thresh[1] = -1.0f; |
133 | 19.6k | sf->rd_ml_partition.search_breakout_thresh[2] = -1.0f; |
134 | 19.6k | } |
135 | 19.6k | #if CONFIG_VP9_HIGHBITDEPTH |
136 | 19.6k | if (cpi->Source->flags & YV12_FLAG_HIGHBITDEPTH) { |
137 | 0 | sf->rd_ml_partition.search_breakout_thresh[0] -= 1.0f; |
138 | 0 | sf->rd_ml_partition.search_breakout_thresh[1] -= 1.0f; |
139 | 0 | sf->rd_ml_partition.search_breakout_thresh[2] -= 1.0f; |
140 | 0 | } |
141 | 19.6k | #endif // CONFIG_VP9_HIGHBITDEPTH |
142 | 19.6k | } |
143 | | |
144 | 22.6k | if (speed >= 2) { |
145 | 0 | sf->use_square_only_thresh_high = BLOCK_4X4; |
146 | 0 | sf->use_square_only_thresh_low = BLOCK_SIZES; |
147 | 0 | if (is_720p_or_larger) { |
148 | 0 | sf->disable_split_mask = |
149 | 0 | cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT; |
150 | 0 | sf->adaptive_pred_interp_filter = 0; |
151 | 0 | sf->partition_search_breakout_thr.dist = (1 << 24); |
152 | 0 | sf->partition_search_breakout_thr.rate = 120; |
153 | 0 | sf->rd_ml_partition.search_breakout = 0; |
154 | 0 | } else { |
155 | 0 | sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY; |
156 | 0 | sf->partition_search_breakout_thr.dist = (1 << 22); |
157 | 0 | sf->partition_search_breakout_thr.rate = 100; |
158 | 0 | sf->rd_ml_partition.search_breakout_thresh[0] = 0.0f; |
159 | 0 | sf->rd_ml_partition.search_breakout_thresh[1] = -1.0f; |
160 | 0 | sf->rd_ml_partition.search_breakout_thresh[2] = -4.0f; |
161 | 0 | } |
162 | 0 | sf->rd_auto_partition_min_limit = set_partition_min_limit(cm); |
163 | | |
164 | | // Use a set of speed features for 4k videos. |
165 | 0 | if (is_2160p_or_larger) { |
166 | 0 | sf->use_square_partition_only = 1; |
167 | 0 | sf->intra_y_mode_mask[TX_32X32] = INTRA_DC; |
168 | 0 | sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC; |
169 | 0 | sf->alt_ref_search_fp = 1; |
170 | 0 | sf->cb_pred_filter_search = 2; |
171 | 0 | sf->adaptive_interp_filter_search = 1; |
172 | 0 | sf->disable_split_mask = DISABLE_ALL_SPLIT; |
173 | 0 | } |
174 | 0 | } |
175 | | |
176 | 22.6k | if (speed >= 3) { |
177 | 0 | sf->rd_ml_partition.search_breakout = 0; |
178 | 0 | if (is_720p_or_larger) { |
179 | 0 | sf->disable_split_mask = DISABLE_ALL_SPLIT; |
180 | 0 | sf->schedule_mode_search = cm->base_qindex < 220 ? 1 : 0; |
181 | 0 | sf->partition_search_breakout_thr.dist = (1 << 25); |
182 | 0 | sf->partition_search_breakout_thr.rate = 200; |
183 | 0 | } else { |
184 | 0 | sf->max_intra_bsize = BLOCK_32X32; |
185 | 0 | sf->disable_split_mask = DISABLE_ALL_INTER_SPLIT; |
186 | 0 | sf->schedule_mode_search = cm->base_qindex < 175 ? 1 : 0; |
187 | 0 | sf->partition_search_breakout_thr.dist = (1 << 23); |
188 | 0 | sf->partition_search_breakout_thr.rate = 120; |
189 | 0 | } |
190 | 0 | } |
191 | | |
192 | | // If this is a two pass clip that fits the criteria for animated or |
193 | | // graphics content then reset disable_split_mask for speeds 1-4. |
194 | | // Also if the image edge is internal to the coded area. |
195 | 22.6k | if ((speed >= 1) && (cpi->oxcf.pass == 2) && |
196 | 22.6k | ((cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) || |
197 | 0 | (vp9_internal_image_edge(cpi)))) { |
198 | 0 | sf->disable_split_mask = DISABLE_COMPOUND_SPLIT; |
199 | 0 | } |
200 | | |
201 | 22.6k | if (speed >= 4) { |
202 | 0 | sf->partition_search_breakout_thr.rate = 300; |
203 | 0 | if (is_720p_or_larger) { |
204 | 0 | sf->partition_search_breakout_thr.dist = (1 << 26); |
205 | 0 | } else { |
206 | 0 | sf->partition_search_breakout_thr.dist = (1 << 24); |
207 | 0 | } |
208 | 0 | sf->disable_split_mask = DISABLE_ALL_SPLIT; |
209 | 0 | } |
210 | | |
211 | 22.6k | if (speed >= 5) { |
212 | 0 | sf->partition_search_breakout_thr.rate = 500; |
213 | 0 | } |
214 | 22.6k | } |
215 | | |
216 | | static double tx_dom_thresholds[6] = { 99.0, 14.0, 12.0, 8.0, 4.0, 0.0 }; |
217 | | static double qopt_thresholds[6] = { 99.0, 12.0, 10.0, 4.0, 2.0, 0.0 }; |
218 | | |
219 | | static void set_good_speed_feature_framesize_independent(VP9_COMP *cpi, |
220 | | VP9_COMMON *cm, |
221 | | SPEED_FEATURES *sf, |
222 | 22.6k | int speed) { |
223 | 22.6k | const VP9EncoderConfig *const oxcf = &cpi->oxcf; |
224 | 22.6k | const int boosted = frame_is_boosted(cpi); |
225 | 22.6k | int i; |
226 | | |
227 | 22.6k | sf->adaptive_interp_filter_search = 1; |
228 | 22.6k | sf->adaptive_pred_interp_filter = 1; |
229 | 22.6k | sf->adaptive_rd_thresh = 1; |
230 | 22.6k | sf->adaptive_rd_thresh_row_mt = 0; |
231 | 22.6k | sf->allow_skip_recode = 1; |
232 | 22.6k | sf->less_rectangular_check = 1; |
233 | 22.6k | sf->mv.auto_mv_step_size = 1; |
234 | 22.6k | sf->mv.use_downsampled_sad = 1; |
235 | 22.6k | sf->prune_ref_frame_for_rect_partitions = 1; |
236 | 22.6k | sf->temporal_filter_search_method = NSTEP; |
237 | 22.6k | sf->tx_size_search_breakout = 1; |
238 | 22.6k | sf->use_square_partition_only = !boosted; |
239 | 22.6k | sf->early_term_interp_search_plane_rd = 1; |
240 | 22.6k | sf->cb_pred_filter_search = 1; |
241 | 22.6k | sf->trellis_opt_tx_rd.method = sf->optimize_coefficients |
242 | 22.6k | ? ENABLE_TRELLIS_OPT_TX_RD_RESIDUAL_MSE |
243 | 22.6k | : DISABLE_TRELLIS_OPT; |
244 | 22.6k | sf->trellis_opt_tx_rd.thresh = boosted ? 4.0 : 3.0; |
245 | | |
246 | 22.6k | sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V; |
247 | 22.6k | sf->comp_inter_joint_search_iter_level = 1; |
248 | | |
249 | | // Reference masking is not supported in dynamic scaling mode. |
250 | 22.6k | sf->reference_masking = oxcf->resize_mode != RESIZE_DYNAMIC; |
251 | | |
252 | 22.6k | sf->rd_ml_partition.var_pruning = 1; |
253 | 22.6k | sf->rd_ml_partition.prune_rect_thresh[0] = -1; |
254 | 22.6k | sf->rd_ml_partition.prune_rect_thresh[1] = 350; |
255 | 22.6k | sf->rd_ml_partition.prune_rect_thresh[2] = 325; |
256 | 22.6k | sf->rd_ml_partition.prune_rect_thresh[3] = 250; |
257 | | |
258 | 22.6k | if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) { |
259 | 0 | sf->exhaustive_searches_thresh = (1 << 22); |
260 | 22.6k | } else { |
261 | 22.6k | sf->exhaustive_searches_thresh = INT_MAX; |
262 | 22.6k | } |
263 | | |
264 | 113k | for (i = 0; i < MAX_MESH_STEP; ++i) { |
265 | 90.5k | const int mesh_density_level = 0; |
266 | 90.5k | sf->mesh_patterns[i].range = |
267 | 90.5k | good_quality_mesh_patterns[mesh_density_level][i].range; |
268 | 90.5k | sf->mesh_patterns[i].interval = |
269 | 90.5k | good_quality_mesh_patterns[mesh_density_level][i].interval; |
270 | 90.5k | } |
271 | | |
272 | 22.6k | if (speed >= 1) { |
273 | 19.6k | sf->rd_ml_partition.var_pruning = !boosted; |
274 | 19.6k | sf->rd_ml_partition.prune_rect_thresh[1] = 225; |
275 | 19.6k | sf->rd_ml_partition.prune_rect_thresh[2] = 225; |
276 | 19.6k | sf->rd_ml_partition.prune_rect_thresh[3] = 225; |
277 | | |
278 | 19.6k | if (oxcf->pass == 2) { |
279 | 0 | TWO_PASS *const twopass = &cpi->twopass; |
280 | 0 | if ((twopass->fr_content_type == FC_GRAPHICS_ANIMATION) || |
281 | 0 | vp9_internal_image_edge(cpi)) { |
282 | 0 | sf->use_square_partition_only = !boosted; |
283 | 0 | } else { |
284 | 0 | sf->use_square_partition_only = !frame_is_intra_only(cm); |
285 | 0 | } |
286 | 19.6k | } else { |
287 | 19.6k | sf->use_square_partition_only = !frame_is_intra_only(cm); |
288 | 19.6k | } |
289 | | |
290 | 19.6k | sf->allow_txfm_domain_distortion = 1; |
291 | 19.6k | sf->tx_domain_thresh = tx_dom_thresholds[(speed < 6) ? speed : 5]; |
292 | 19.6k | sf->trellis_opt_tx_rd.method = sf->optimize_coefficients |
293 | 19.6k | ? ENABLE_TRELLIS_OPT_TX_RD_SRC_VAR |
294 | 19.6k | : DISABLE_TRELLIS_OPT; |
295 | 19.6k | sf->trellis_opt_tx_rd.thresh = qopt_thresholds[(speed < 6) ? speed : 5]; |
296 | 19.6k | sf->less_rectangular_check = 1; |
297 | 19.6k | sf->use_rd_breakout = 1; |
298 | 19.6k | sf->adaptive_motion_search = 1; |
299 | 19.6k | sf->adaptive_rd_thresh = 2; |
300 | 19.6k | sf->mv.subpel_search_level = 1; |
301 | 19.6k | if (cpi->oxcf.content != VP9E_CONTENT_FILM) sf->mode_skip_start = 10; |
302 | 19.6k | sf->allow_acl = 0; |
303 | | |
304 | 19.6k | sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V; |
305 | 19.6k | if (cpi->oxcf.content != VP9E_CONTENT_FILM) { |
306 | 19.6k | sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V; |
307 | 19.6k | sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V; |
308 | 19.6k | } |
309 | | |
310 | 19.6k | sf->recode_tolerance_low = 15; |
311 | 19.6k | sf->recode_tolerance_high = 30; |
312 | | |
313 | 19.6k | sf->exhaustive_searches_thresh = |
314 | 19.6k | (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ? (1 << 23) |
315 | 19.6k | : INT_MAX; |
316 | 19.6k | sf->use_accurate_subpel_search = USE_4_TAPS; |
317 | 19.6k | } |
318 | | |
319 | 22.6k | if (speed >= 2) { |
320 | 0 | sf->rd_ml_partition.var_pruning = 0; |
321 | 0 | if (oxcf->vbr_corpus_complexity) |
322 | 0 | sf->recode_loop = ALLOW_RECODE_FIRST; |
323 | 0 | else |
324 | 0 | sf->recode_loop = ALLOW_RECODE_KFARFGF; |
325 | |
|
326 | 0 | sf->tx_size_search_method = |
327 | 0 | frame_is_boosted(cpi) ? USE_FULL_RD : USE_LARGESTALL; |
328 | |
|
329 | 0 | sf->mode_search_skip_flags = |
330 | 0 | (cm->frame_type == KEY_FRAME) |
331 | 0 | ? 0 |
332 | 0 | : FLAG_SKIP_INTRA_DIRMISMATCH | FLAG_SKIP_INTRA_BESTINTER | |
333 | 0 | FLAG_SKIP_COMP_BESTINTRA | FLAG_SKIP_INTRA_LOWVAR; |
334 | 0 | sf->disable_filter_search_var_thresh = 100; |
335 | 0 | sf->comp_inter_joint_search_iter_level = 2; |
336 | 0 | sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX; |
337 | 0 | sf->recode_tolerance_high = 45; |
338 | 0 | sf->enhanced_full_pixel_motion_search = 0; |
339 | 0 | sf->prune_ref_frame_for_rect_partitions = 0; |
340 | 0 | sf->rd_ml_partition.prune_rect_thresh[1] = -1; |
341 | 0 | sf->rd_ml_partition.prune_rect_thresh[2] = -1; |
342 | 0 | sf->rd_ml_partition.prune_rect_thresh[3] = -1; |
343 | 0 | sf->mv.subpel_search_level = 0; |
344 | |
|
345 | 0 | if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) { |
346 | 0 | for (i = 0; i < MAX_MESH_STEP; ++i) { |
347 | 0 | int mesh_density_level = 1; |
348 | 0 | sf->mesh_patterns[i].range = |
349 | 0 | good_quality_mesh_patterns[mesh_density_level][i].range; |
350 | 0 | sf->mesh_patterns[i].interval = |
351 | 0 | good_quality_mesh_patterns[mesh_density_level][i].interval; |
352 | 0 | } |
353 | 0 | } |
354 | |
|
355 | 0 | sf->use_accurate_subpel_search = USE_2_TAPS; |
356 | 0 | } |
357 | | |
358 | 22.6k | if (speed >= 3) { |
359 | 0 | sf->use_square_partition_only = !frame_is_intra_only(cm); |
360 | 0 | sf->tx_size_search_method = |
361 | 0 | frame_is_intra_only(cm) ? USE_FULL_RD : USE_LARGESTALL; |
362 | 0 | sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED; |
363 | 0 | sf->adaptive_pred_interp_filter = 0; |
364 | 0 | sf->adaptive_mode_search = 1; |
365 | 0 | sf->cb_partition_search = !boosted; |
366 | 0 | sf->cb_pred_filter_search = 2; |
367 | 0 | sf->alt_ref_search_fp = 1; |
368 | 0 | sf->recode_loop = ALLOW_RECODE_KFMAXBW; |
369 | 0 | sf->adaptive_rd_thresh = 3; |
370 | 0 | sf->mode_skip_start = 6; |
371 | 0 | sf->intra_y_mode_mask[TX_32X32] = INTRA_DC; |
372 | 0 | sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC; |
373 | |
|
374 | 0 | if (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) { |
375 | 0 | for (i = 0; i < MAX_MESH_STEP; ++i) { |
376 | 0 | int mesh_density_level = 2; |
377 | 0 | sf->mesh_patterns[i].range = |
378 | 0 | good_quality_mesh_patterns[mesh_density_level][i].range; |
379 | 0 | sf->mesh_patterns[i].interval = |
380 | 0 | good_quality_mesh_patterns[mesh_density_level][i].interval; |
381 | 0 | } |
382 | 0 | } |
383 | 0 | } |
384 | | |
385 | 22.6k | if (speed >= 4) { |
386 | 0 | sf->use_square_partition_only = 1; |
387 | 0 | sf->tx_size_search_method = USE_LARGESTALL; |
388 | 0 | sf->mv.search_method = BIGDIA; |
389 | 0 | sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED_MORE; |
390 | 0 | sf->adaptive_rd_thresh = 4; |
391 | 0 | if (cm->frame_type != KEY_FRAME) |
392 | 0 | sf->mode_search_skip_flags |= FLAG_EARLY_TERMINATE; |
393 | 0 | sf->disable_filter_search_var_thresh = 200; |
394 | 0 | sf->use_lp32x32fdct = 1; |
395 | 0 | sf->use_fast_coef_updates = ONE_LOOP_REDUCED; |
396 | 0 | sf->use_fast_coef_costing = 1; |
397 | 0 | sf->motion_field_mode_search = !boosted; |
398 | 0 | } |
399 | | |
400 | 22.6k | if (speed >= 5) { |
401 | 0 | sf->optimize_coefficients = 0; |
402 | 0 | sf->mv.search_method = HEX; |
403 | 0 | sf->disable_filter_search_var_thresh = 500; |
404 | 0 | for (i = 0; i < TX_SIZES; ++i) { |
405 | 0 | sf->intra_y_mode_mask[i] = INTRA_DC; |
406 | 0 | sf->intra_uv_mode_mask[i] = INTRA_DC; |
407 | 0 | } |
408 | 0 | sf->mv.reduce_first_step_size = 1; |
409 | 0 | sf->simple_model_rd_from_var = 1; |
410 | 0 | } |
411 | 22.6k | } |
412 | | #endif // !CONFIG_REALTIME_ONLY |
413 | | |
414 | | static void set_rt_speed_feature_framesize_dependent(VP9_COMP *cpi, |
415 | | SPEED_FEATURES *sf, |
416 | 22.6k | int speed) { |
417 | 22.6k | VP9_COMMON *const cm = &cpi->common; |
418 | | |
419 | 22.6k | if (speed >= 1) { |
420 | 22.6k | if (VPXMIN(cm->width, cm->height) >= 720) { |
421 | 0 | sf->disable_split_mask = |
422 | 0 | cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT; |
423 | 22.6k | } else { |
424 | 22.6k | sf->disable_split_mask = DISABLE_COMPOUND_SPLIT; |
425 | 22.6k | } |
426 | 22.6k | } |
427 | | |
428 | 22.6k | if (speed >= 2) { |
429 | 0 | if (VPXMIN(cm->width, cm->height) >= 720) { |
430 | 0 | sf->disable_split_mask = |
431 | 0 | cm->show_frame ? DISABLE_ALL_SPLIT : DISABLE_ALL_INTER_SPLIT; |
432 | 0 | } else { |
433 | 0 | sf->disable_split_mask = LAST_AND_INTRA_SPLIT_ONLY; |
434 | 0 | } |
435 | 0 | } |
436 | | |
437 | 22.6k | if (speed >= 5) { |
438 | 0 | sf->partition_search_breakout_thr.rate = 200; |
439 | 0 | if (VPXMIN(cm->width, cm->height) >= 720) { |
440 | 0 | sf->partition_search_breakout_thr.dist = (1 << 25); |
441 | 0 | } else { |
442 | 0 | sf->partition_search_breakout_thr.dist = (1 << 23); |
443 | 0 | } |
444 | 0 | } |
445 | | |
446 | 22.6k | if (speed >= 7) { |
447 | 0 | sf->encode_breakout_thresh = |
448 | 0 | (VPXMIN(cm->width, cm->height) >= 720) ? 800 : 300; |
449 | 0 | } |
450 | 22.6k | } |
451 | | |
452 | | static void set_rt_speed_feature_framesize_independent( |
453 | 22.6k | VP9_COMP *cpi, SPEED_FEATURES *sf, int speed, vp9e_tune_content content) { |
454 | 22.6k | VP9_COMMON *const cm = &cpi->common; |
455 | 22.6k | SVC *const svc = &cpi->svc; |
456 | 22.6k | const int is_keyframe = cm->frame_type == KEY_FRAME; |
457 | 22.6k | const int frames_since_key = is_keyframe ? 0 : cpi->rc.frames_since_key; |
458 | 22.6k | sf->static_segmentation = 0; |
459 | 22.6k | sf->adaptive_rd_thresh = 1; |
460 | 22.6k | sf->adaptive_rd_thresh_row_mt = 0; |
461 | 22.6k | sf->use_fast_coef_costing = 1; |
462 | 22.6k | sf->exhaustive_searches_thresh = INT_MAX; |
463 | 22.6k | sf->allow_acl = 0; |
464 | 22.6k | sf->copy_partition_flag = 0; |
465 | 22.6k | sf->use_source_sad = 0; |
466 | 22.6k | sf->use_simple_block_yrd = 0; |
467 | 22.6k | sf->adapt_partition_source_sad = 0; |
468 | 22.6k | sf->use_altref_onepass = 0; |
469 | 22.6k | sf->use_compound_nonrd_pickmode = 0; |
470 | 22.6k | sf->nonrd_keyframe = 0; |
471 | 22.6k | sf->svc_use_lowres_part = 0; |
472 | 22.6k | sf->overshoot_detection_cbr_rt = NO_DETECTION; |
473 | 22.6k | sf->disable_16x16part_nonkey = 0; |
474 | 22.6k | sf->disable_golden_ref = 0; |
475 | 22.6k | sf->enable_tpl_model = 0; |
476 | 22.6k | sf->enhanced_full_pixel_motion_search = 0; |
477 | 22.6k | sf->use_accurate_subpel_search = USE_2_TAPS; |
478 | 22.6k | sf->nonrd_use_ml_partition = 0; |
479 | 22.6k | sf->variance_part_thresh_mult = 1; |
480 | 22.6k | sf->cb_pred_filter_search = 0; |
481 | 22.6k | sf->force_smooth_interpol = 0; |
482 | 22.6k | sf->rt_intra_dc_only_low_content = 0; |
483 | 22.6k | sf->mv.enable_adaptive_subpel_force_stop = 0; |
484 | | |
485 | 22.6k | if (speed >= 1) { |
486 | 22.6k | sf->allow_txfm_domain_distortion = 1; |
487 | 22.6k | sf->tx_domain_thresh = 0.0; |
488 | 22.6k | sf->trellis_opt_tx_rd.method = DISABLE_TRELLIS_OPT; |
489 | 22.6k | sf->trellis_opt_tx_rd.thresh = 0.0; |
490 | 22.6k | sf->use_square_partition_only = !frame_is_intra_only(cm); |
491 | 22.6k | sf->less_rectangular_check = 1; |
492 | 22.6k | sf->tx_size_search_method = |
493 | 22.6k | frame_is_intra_only(cm) ? USE_FULL_RD : USE_LARGESTALL; |
494 | | |
495 | 22.6k | sf->use_rd_breakout = 1; |
496 | | |
497 | 22.6k | sf->adaptive_motion_search = 1; |
498 | 22.6k | sf->adaptive_pred_interp_filter = 1; |
499 | 22.6k | sf->mv.auto_mv_step_size = 1; |
500 | 22.6k | sf->adaptive_rd_thresh = 2; |
501 | 22.6k | sf->intra_y_mode_mask[TX_32X32] = INTRA_DC_H_V; |
502 | 22.6k | sf->intra_uv_mode_mask[TX_32X32] = INTRA_DC_H_V; |
503 | 22.6k | sf->intra_uv_mode_mask[TX_16X16] = INTRA_DC_H_V; |
504 | 22.6k | } |
505 | | |
506 | 22.6k | if (speed >= 2) { |
507 | 0 | sf->mode_search_skip_flags = |
508 | 0 | (cm->frame_type == KEY_FRAME) |
509 | 0 | ? 0 |
510 | 0 | : FLAG_SKIP_INTRA_DIRMISMATCH | FLAG_SKIP_INTRA_BESTINTER | |
511 | 0 | FLAG_SKIP_COMP_BESTINTRA | FLAG_SKIP_INTRA_LOWVAR; |
512 | 0 | sf->adaptive_pred_interp_filter = 2; |
513 | | |
514 | | // Reference masking only enabled for 1 spatial layer, and if none of the |
515 | | // references have been scaled. The latter condition needs to be checked |
516 | | // for external or internal dynamic resize. |
517 | 0 | sf->reference_masking = (svc->number_spatial_layers == 1); |
518 | 0 | if (sf->reference_masking == 1 && |
519 | 0 | (cpi->external_resize == 1 || |
520 | 0 | cpi->oxcf.resize_mode == RESIZE_DYNAMIC)) { |
521 | 0 | MV_REFERENCE_FRAME ref_frame; |
522 | 0 | for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) { |
523 | 0 | const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame); |
524 | 0 | if (yv12 != NULL && |
525 | 0 | (cpi->ref_frame_flags & ref_frame_to_flag(ref_frame))) { |
526 | 0 | const struct scale_factors *const scale_fac = |
527 | 0 | &cm->frame_refs[ref_frame - 1].sf; |
528 | 0 | if (vp9_is_scaled(scale_fac)) sf->reference_masking = 0; |
529 | 0 | } |
530 | 0 | } |
531 | 0 | } |
532 | |
|
533 | 0 | sf->disable_filter_search_var_thresh = 50; |
534 | 0 | sf->comp_inter_joint_search_iter_level = 2; |
535 | 0 | sf->auto_min_max_partition_size = RELAXED_NEIGHBORING_MIN_MAX; |
536 | 0 | sf->lf_motion_threshold = LOW_MOTION_THRESHOLD; |
537 | 0 | sf->adjust_partitioning_from_last_frame = 1; |
538 | 0 | sf->last_partitioning_redo_frequency = 3; |
539 | 0 | sf->use_lp32x32fdct = 1; |
540 | 0 | sf->mode_skip_start = 11; |
541 | 0 | sf->intra_y_mode_mask[TX_16X16] = INTRA_DC_H_V; |
542 | 0 | } |
543 | | |
544 | 22.6k | if (speed >= 3) { |
545 | 0 | sf->use_square_partition_only = 1; |
546 | 0 | sf->disable_filter_search_var_thresh = 100; |
547 | 0 | sf->use_uv_intra_rd_estimate = 1; |
548 | 0 | sf->skip_encode_sb = 1; |
549 | 0 | sf->mv.subpel_search_level = 0; |
550 | 0 | sf->adaptive_rd_thresh = 4; |
551 | 0 | sf->mode_skip_start = 6; |
552 | 0 | sf->allow_skip_recode = 0; |
553 | 0 | sf->optimize_coefficients = 0; |
554 | 0 | sf->disable_split_mask = DISABLE_ALL_SPLIT; |
555 | 0 | sf->lpf_pick = LPF_PICK_FROM_Q; |
556 | 0 | } |
557 | | |
558 | 22.6k | if (speed >= 4) { |
559 | 0 | int i; |
560 | 0 | if (cpi->oxcf.rc_mode == VPX_VBR && cpi->oxcf.lag_in_frames > 0) |
561 | 0 | sf->use_altref_onepass = 1; |
562 | 0 | sf->mv.subpel_force_stop = QUARTER_PEL; |
563 | 0 | for (i = 0; i < TX_SIZES; i++) { |
564 | 0 | sf->intra_y_mode_mask[i] = INTRA_DC_H_V; |
565 | 0 | sf->intra_uv_mode_mask[i] = INTRA_DC; |
566 | 0 | } |
567 | 0 | sf->intra_y_mode_mask[TX_32X32] = INTRA_DC; |
568 | 0 | sf->frame_parameter_update = 0; |
569 | 0 | sf->mv.search_method = FAST_HEX; |
570 | 0 | sf->allow_skip_recode = 0; |
571 | 0 | sf->max_intra_bsize = BLOCK_32X32; |
572 | 0 | sf->use_fast_coef_costing = 0; |
573 | 0 | sf->use_quant_fp = !is_keyframe; |
574 | 0 | sf->inter_mode_mask[BLOCK_32X32] = INTER_NEAREST_NEW_ZERO; |
575 | 0 | sf->inter_mode_mask[BLOCK_32X64] = INTER_NEAREST_NEW_ZERO; |
576 | 0 | sf->inter_mode_mask[BLOCK_64X32] = INTER_NEAREST_NEW_ZERO; |
577 | 0 | sf->inter_mode_mask[BLOCK_64X64] = INTER_NEAREST_NEW_ZERO; |
578 | 0 | sf->adaptive_rd_thresh = 2; |
579 | 0 | sf->use_fast_coef_updates = is_keyframe ? TWO_LOOP : ONE_LOOP_REDUCED; |
580 | 0 | sf->mode_search_skip_flags = FLAG_SKIP_INTRA_DIRMISMATCH; |
581 | 0 | sf->tx_size_search_method = is_keyframe ? USE_LARGESTALL : USE_TX_8X8; |
582 | 0 | sf->partition_search_type = VAR_BASED_PARTITION; |
583 | 0 | } |
584 | | |
585 | 22.6k | if (speed >= 5) { |
586 | 0 | sf->use_altref_onepass = 0; |
587 | 0 | sf->use_quant_fp = !is_keyframe; |
588 | 0 | sf->auto_min_max_partition_size = |
589 | 0 | is_keyframe ? RELAXED_NEIGHBORING_MIN_MAX : STRICT_NEIGHBORING_MIN_MAX; |
590 | 0 | sf->default_max_partition_size = BLOCK_32X32; |
591 | 0 | sf->default_min_partition_size = BLOCK_8X8; |
592 | 0 | sf->force_frame_boost = |
593 | 0 | is_keyframe || |
594 | 0 | (frames_since_key % (sf->last_partitioning_redo_frequency << 1) == 1); |
595 | 0 | sf->max_delta_qindex = is_keyframe ? 20 : 15; |
596 | 0 | sf->partition_search_type = REFERENCE_PARTITION; |
597 | 0 | if (cpi->oxcf.rc_mode == VPX_VBR && cpi->oxcf.lag_in_frames > 0 && |
598 | 0 | cpi->rc.is_src_frame_alt_ref) { |
599 | 0 | sf->partition_search_type = VAR_BASED_PARTITION; |
600 | 0 | } |
601 | 0 | sf->use_nonrd_pick_mode = 1; |
602 | 0 | sf->allow_skip_recode = 0; |
603 | 0 | sf->inter_mode_mask[BLOCK_32X32] = INTER_NEAREST_NEW_ZERO; |
604 | 0 | sf->inter_mode_mask[BLOCK_32X64] = INTER_NEAREST_NEW_ZERO; |
605 | 0 | sf->inter_mode_mask[BLOCK_64X32] = INTER_NEAREST_NEW_ZERO; |
606 | 0 | sf->inter_mode_mask[BLOCK_64X64] = INTER_NEAREST_NEW_ZERO; |
607 | 0 | sf->adaptive_rd_thresh = 2; |
608 | | // This feature is only enabled when partition search is disabled. |
609 | 0 | sf->reuse_inter_pred_sby = 1; |
610 | 0 | sf->coeff_prob_appx_step = 4; |
611 | 0 | sf->use_fast_coef_updates = is_keyframe ? TWO_LOOP : ONE_LOOP_REDUCED; |
612 | 0 | sf->mode_search_skip_flags = FLAG_SKIP_INTRA_DIRMISMATCH; |
613 | 0 | sf->tx_size_search_method = is_keyframe ? USE_LARGESTALL : USE_TX_8X8; |
614 | 0 | sf->simple_model_rd_from_var = 1; |
615 | 0 | if (cpi->oxcf.rc_mode == VPX_VBR) sf->mv.search_method = NSTEP; |
616 | |
|
617 | 0 | if (!is_keyframe) { |
618 | 0 | int i; |
619 | 0 | if (content == VP9E_CONTENT_SCREEN) { |
620 | 0 | for (i = 0; i < BLOCK_SIZES; ++i) |
621 | 0 | if (i >= BLOCK_32X32) |
622 | 0 | sf->intra_y_mode_bsize_mask[i] = INTRA_DC_H_V; |
623 | 0 | else |
624 | 0 | sf->intra_y_mode_bsize_mask[i] = INTRA_DC_TM_H_V; |
625 | 0 | } else { |
626 | 0 | for (i = 0; i < BLOCK_SIZES; ++i) |
627 | 0 | if (i > BLOCK_16X16) |
628 | 0 | sf->intra_y_mode_bsize_mask[i] = INTRA_DC; |
629 | 0 | else |
630 | | // Use H and V intra mode for block sizes <= 16X16. |
631 | 0 | sf->intra_y_mode_bsize_mask[i] = INTRA_DC_H_V; |
632 | 0 | } |
633 | 0 | } |
634 | 0 | if (content == VP9E_CONTENT_SCREEN) { |
635 | 0 | sf->short_circuit_flat_blocks = 1; |
636 | 0 | } |
637 | 0 | if (cpi->oxcf.rc_mode == VPX_CBR && |
638 | 0 | cpi->oxcf.content != VP9E_CONTENT_SCREEN) { |
639 | 0 | sf->limit_newmv_early_exit = 1; |
640 | 0 | if (!cpi->use_svc) sf->bias_golden = 1; |
641 | 0 | } |
642 | | // Keep nonrd_keyframe = 1 for non-base spatial layers to prevent |
643 | | // increase in encoding time. |
644 | 0 | if (cpi->use_svc && svc->spatial_layer_id > 0) sf->nonrd_keyframe = 1; |
645 | 0 | if (cm->frame_type != KEY_FRAME && cpi->resize_state == ORIG && |
646 | 0 | cpi->oxcf.rc_mode == VPX_CBR && !cpi->rc.disable_overshoot_maxq_cbr) { |
647 | 0 | if (cm->width * cm->height <= 352 * 288 && !cpi->use_svc && |
648 | 0 | cpi->oxcf.content != VP9E_CONTENT_SCREEN) |
649 | 0 | sf->overshoot_detection_cbr_rt = RE_ENCODE_MAXQ; |
650 | 0 | else |
651 | 0 | sf->overshoot_detection_cbr_rt = FAST_DETECTION_MAXQ; |
652 | 0 | } |
653 | 0 | if (cpi->oxcf.rc_mode == VPX_VBR && cpi->oxcf.lag_in_frames > 0 && |
654 | 0 | cm->width <= 1280 && cm->height <= 720) { |
655 | 0 | sf->use_altref_onepass = 1; |
656 | 0 | sf->use_compound_nonrd_pickmode = 1; |
657 | 0 | } |
658 | 0 | if (cm->width * cm->height > 1280 * 720) sf->cb_pred_filter_search = 2; |
659 | 0 | if (!cpi->external_resize) sf->use_source_sad = 1; |
660 | 0 | } |
661 | | |
662 | 22.6k | if (speed >= 6) { |
663 | 0 | if (cpi->oxcf.rc_mode == VPX_VBR && cpi->oxcf.lag_in_frames > 0) { |
664 | 0 | sf->use_altref_onepass = 1; |
665 | 0 | sf->use_compound_nonrd_pickmode = 1; |
666 | 0 | } |
667 | 0 | sf->partition_search_type = VAR_BASED_PARTITION; |
668 | 0 | sf->mv.search_method = NSTEP; |
669 | 0 | sf->mv.reduce_first_step_size = 1; |
670 | 0 | sf->skip_encode_sb = 0; |
671 | |
|
672 | 0 | if (sf->use_source_sad) { |
673 | 0 | sf->adapt_partition_source_sad = 1; |
674 | 0 | sf->adapt_partition_thresh = |
675 | 0 | (cm->width * cm->height <= 640 * 360) ? 40000 : 60000; |
676 | 0 | if (cpi->content_state_sb_fd == NULL && |
677 | 0 | (!cpi->use_svc || |
678 | 0 | svc->spatial_layer_id == svc->number_spatial_layers - 1)) { |
679 | 0 | CHECK_MEM_ERROR(&cm->error, cpi->content_state_sb_fd, |
680 | 0 | (uint8_t *)vpx_calloc( |
681 | 0 | (cm->mi_stride >> 3) * ((cm->mi_rows >> 3) + 1), |
682 | 0 | sizeof(uint8_t))); |
683 | 0 | } |
684 | 0 | } |
685 | 0 | if (cpi->oxcf.rc_mode == VPX_CBR && content != VP9E_CONTENT_SCREEN) { |
686 | | // Enable short circuit for low temporal variance. |
687 | 0 | sf->short_circuit_low_temp_var = 1; |
688 | 0 | } |
689 | 0 | if (svc->temporal_layer_id > 0) { |
690 | 0 | sf->adaptive_rd_thresh = 4; |
691 | 0 | sf->limit_newmv_early_exit = 0; |
692 | 0 | sf->base_mv_aggressive = 1; |
693 | 0 | } |
694 | 0 | if (cm->frame_type != KEY_FRAME && cpi->resize_state == ORIG && |
695 | 0 | cpi->oxcf.rc_mode == VPX_CBR && !cpi->rc.disable_overshoot_maxq_cbr) |
696 | 0 | sf->overshoot_detection_cbr_rt = FAST_DETECTION_MAXQ; |
697 | 0 | } |
698 | | |
699 | 22.6k | if (speed >= 7) { |
700 | 0 | sf->adapt_partition_source_sad = 0; |
701 | 0 | sf->adaptive_rd_thresh = 3; |
702 | 0 | sf->mv.search_method = FAST_DIAMOND; |
703 | 0 | sf->mv.fullpel_search_step_param = 10; |
704 | | // For SVC: use better mv search on base temporal layer, and only |
705 | | // on base spatial layer if highest resolution is above 640x360. |
706 | 0 | if (svc->number_temporal_layers > 2 && svc->temporal_layer_id == 0 && |
707 | 0 | (svc->spatial_layer_id == 0 || |
708 | 0 | cpi->oxcf.width * cpi->oxcf.height <= 640 * 360)) { |
709 | 0 | sf->mv.search_method = NSTEP; |
710 | 0 | sf->mv.fullpel_search_step_param = 6; |
711 | 0 | } |
712 | 0 | if (svc->temporal_layer_id > 0 || svc->spatial_layer_id > 1) { |
713 | 0 | sf->use_simple_block_yrd = 1; |
714 | 0 | if (svc->non_reference_frame) |
715 | 0 | sf->mv.subpel_search_method = SUBPEL_TREE_PRUNED_EVENMORE; |
716 | 0 | } |
717 | 0 | if (cpi->use_svc && cpi->row_mt && cpi->oxcf.max_threads > 1) |
718 | 0 | sf->adaptive_rd_thresh_row_mt = 1; |
719 | | // Enable partition copy. For SVC only enabled for top spatial resolution |
720 | | // layer. |
721 | 0 | cpi->max_copied_frame = 0; |
722 | 0 | if (!cpi->last_frame_dropped && cpi->resize_state == ORIG && |
723 | 0 | !cpi->external_resize && |
724 | 0 | (!cpi->use_svc || |
725 | 0 | (svc->spatial_layer_id == svc->number_spatial_layers - 1 && |
726 | 0 | !svc->last_layer_dropped[svc->number_spatial_layers - 1]))) { |
727 | 0 | sf->copy_partition_flag = 1; |
728 | 0 | cpi->max_copied_frame = 2; |
729 | | // The top temporal enhancement layer (for number of temporal layers > 1) |
730 | | // are non-reference frames, so use large/max value for max_copied_frame. |
731 | 0 | if (svc->number_temporal_layers > 1 && |
732 | 0 | svc->temporal_layer_id == svc->number_temporal_layers - 1) |
733 | 0 | cpi->max_copied_frame = 255; |
734 | 0 | } |
735 | | // For SVC: enable use of lower resolution partition for higher resolution, |
736 | | // only for 3 spatial layers and when config/top resolution is above VGA. |
737 | | // Enable only for non-base temporal layer frames. |
738 | 0 | if (cpi->use_svc && svc->use_partition_reuse && |
739 | 0 | svc->number_spatial_layers == 3 && svc->temporal_layer_id > 0 && |
740 | 0 | cpi->oxcf.width * cpi->oxcf.height > 640 * 480) |
741 | 0 | sf->svc_use_lowres_part = 1; |
742 | | // For SVC when golden is used as second temporal reference: to avoid |
743 | | // encode time increase only use this feature on base temporal layer. |
744 | | // (i.e remove golden flag from frame_flags for temporal_layer_id > 0). |
745 | 0 | if (cpi->use_svc && svc->use_gf_temporal_ref_current_layer && |
746 | 0 | svc->temporal_layer_id > 0) |
747 | 0 | cpi->ref_frame_flags &= (~VP9_GOLD_FLAG); |
748 | 0 | if (cm->width * cm->height > 640 * 480) sf->cb_pred_filter_search = 2; |
749 | 0 | } |
750 | | |
751 | 22.6k | if (speed >= 8) { |
752 | 0 | sf->adaptive_rd_thresh = 4; |
753 | 0 | sf->skip_encode_sb = 1; |
754 | 0 | if (cpi->svc.number_spatial_layers > 1 && !cpi->svc.simulcast_mode) |
755 | 0 | sf->nonrd_keyframe = 0; |
756 | 0 | else |
757 | 0 | sf->nonrd_keyframe = 1; |
758 | 0 | if (!cpi->use_svc) cpi->max_copied_frame = 4; |
759 | 0 | if (cpi->row_mt && cpi->oxcf.max_threads > 1) |
760 | 0 | sf->adaptive_rd_thresh_row_mt = 1; |
761 | | // Enable ML based partition for low res. |
762 | 0 | if (!frame_is_intra_only(cm) && cm->width * cm->height <= 352 * 288) { |
763 | 0 | sf->nonrd_use_ml_partition = 1; |
764 | 0 | } |
765 | 0 | #if CONFIG_VP9_HIGHBITDEPTH |
766 | 0 | if (cpi->Source->flags & YV12_FLAG_HIGHBITDEPTH) |
767 | 0 | sf->nonrd_use_ml_partition = 0; |
768 | 0 | #endif |
769 | 0 | if (content == VP9E_CONTENT_SCREEN) sf->mv.subpel_force_stop = HALF_PEL; |
770 | 0 | sf->rt_intra_dc_only_low_content = 1; |
771 | 0 | if (!cpi->use_svc && cpi->oxcf.rc_mode == VPX_CBR && |
772 | 0 | content != VP9E_CONTENT_SCREEN) { |
773 | | // More aggressive short circuit for speed 8. |
774 | 0 | sf->short_circuit_low_temp_var = 3; |
775 | | // Use level 2 for noisey cases as there is a regression in some |
776 | | // noisy clips with level 3. |
777 | 0 | if (cpi->noise_estimate.enabled && cm->width >= 1280 && |
778 | 0 | cm->height >= 720) { |
779 | 0 | NOISE_LEVEL noise_level = |
780 | 0 | vp9_noise_estimate_extract_level(&cpi->noise_estimate); |
781 | 0 | if (noise_level >= kMedium) sf->short_circuit_low_temp_var = 2; |
782 | 0 | } |
783 | | // Since the short_circuit_low_temp_var is used, reduce the |
784 | | // adaptive_rd_thresh level. |
785 | 0 | if (cm->width * cm->height > 352 * 288) |
786 | 0 | sf->adaptive_rd_thresh = 1; |
787 | 0 | else |
788 | 0 | sf->adaptive_rd_thresh = 2; |
789 | 0 | } |
790 | 0 | sf->limit_newmv_early_exit = 0; |
791 | 0 | sf->use_simple_block_yrd = 1; |
792 | 0 | if (cm->width * cm->height > 352 * 288) sf->cb_pred_filter_search = 2; |
793 | 0 | } |
794 | | |
795 | 22.6k | if (speed >= 9) { |
796 | | // Only keep INTRA_DC mode for speed 9. |
797 | 0 | if (!is_keyframe) { |
798 | 0 | int i = 0; |
799 | 0 | for (i = 0; i < BLOCK_SIZES; ++i) |
800 | 0 | sf->intra_y_mode_bsize_mask[i] = INTRA_DC; |
801 | 0 | } |
802 | 0 | sf->cb_pred_filter_search = 2; |
803 | 0 | sf->mv.enable_adaptive_subpel_force_stop = 1; |
804 | 0 | sf->mv.adapt_subpel_force_stop.mv_thresh = 1; |
805 | 0 | sf->mv.adapt_subpel_force_stop.force_stop_below = QUARTER_PEL; |
806 | 0 | sf->mv.adapt_subpel_force_stop.force_stop_above = HALF_PEL; |
807 | | // Disable partition blocks below 16x16, except for low-resolutions. |
808 | 0 | if (cm->frame_type != KEY_FRAME && cm->width >= 320 && cm->height >= 240) |
809 | 0 | sf->disable_16x16part_nonkey = 1; |
810 | | // Allow for disabling GOLDEN reference, for CBR mode. |
811 | 0 | if (cpi->oxcf.rc_mode == VPX_CBR) sf->disable_golden_ref = 1; |
812 | 0 | if (cpi->rc.avg_frame_low_motion < 70) sf->default_interp_filter = BILINEAR; |
813 | 0 | if (cm->width * cm->height >= 640 * 360) sf->variance_part_thresh_mult = 2; |
814 | 0 | } |
815 | | |
816 | | // Disable split to 8x8 for low-resolution at very high Q. |
817 | | // For variance partition (speed >= 6). Ignore the first few frames |
818 | | // as avg_frame_qindex starts at max_q (worst_quality). |
819 | 22.6k | if (cm->frame_type != KEY_FRAME && cm->width * cm->height <= 320 * 240 && |
820 | 22.6k | sf->partition_search_type == VAR_BASED_PARTITION && |
821 | 22.6k | cpi->rc.avg_frame_qindex[INTER_FRAME] > 208 && |
822 | 22.6k | cpi->common.current_video_frame > 8) |
823 | 0 | sf->disable_16x16part_nonkey = 1; |
824 | | |
825 | 22.6k | if (sf->nonrd_use_ml_partition) |
826 | 0 | sf->partition_search_type = ML_BASED_PARTITION; |
827 | | |
828 | 22.6k | if (sf->use_altref_onepass) { |
829 | 0 | if (cpi->rc.is_src_frame_alt_ref && cm->frame_type != KEY_FRAME) { |
830 | 0 | sf->partition_search_type = FIXED_PARTITION; |
831 | 0 | sf->always_this_block_size = BLOCK_64X64; |
832 | 0 | } |
833 | 0 | if (cpi->count_arf_frame_usage == NULL) { |
834 | 0 | CHECK_MEM_ERROR( |
835 | 0 | &cm->error, cpi->count_arf_frame_usage, |
836 | 0 | (uint8_t *)vpx_calloc((cm->mi_stride >> 3) * ((cm->mi_rows >> 3) + 1), |
837 | 0 | sizeof(*cpi->count_arf_frame_usage))); |
838 | 0 | } |
839 | 0 | if (cpi->count_lastgolden_frame_usage == NULL) |
840 | 0 | CHECK_MEM_ERROR( |
841 | 0 | &cm->error, cpi->count_lastgolden_frame_usage, |
842 | 0 | (uint8_t *)vpx_calloc((cm->mi_stride >> 3) * ((cm->mi_rows >> 3) + 1), |
843 | 0 | sizeof(*cpi->count_lastgolden_frame_usage))); |
844 | 0 | } |
845 | 22.6k | if (svc->previous_frame_is_intra_only) { |
846 | 0 | sf->partition_search_type = FIXED_PARTITION; |
847 | 0 | sf->always_this_block_size = BLOCK_64X64; |
848 | 0 | } |
849 | | // Special case for screen content: increase motion search on base spatial |
850 | | // layer when high motion is detected or previous SL0 frame was dropped. |
851 | 22.6k | if (cpi->oxcf.content == VP9E_CONTENT_SCREEN && cpi->oxcf.speed >= 5 && |
852 | 22.6k | (svc->high_num_blocks_with_motion || svc->last_layer_dropped[0])) { |
853 | 0 | sf->mv.search_method = NSTEP; |
854 | | // TODO(marpan/jianj): Tune this setting for screensharing. For now use |
855 | | // small step_param for all spatial layers. |
856 | 0 | sf->mv.fullpel_search_step_param = 2; |
857 | 0 | } |
858 | | // TODO(marpan): There is regression for aq-mode=3 speed <= 4, force it |
859 | | // off for now. |
860 | 22.6k | if (speed <= 3 && cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) |
861 | 0 | cpi->oxcf.aq_mode = 0; |
862 | | // For all speeds for rt mode: if the deadline mode changed (was good/best |
863 | | // quality on previous frame and now is realtime) set nonrd_keyframe to 1 to |
864 | | // avoid entering rd pickmode. This causes issues, such as: b/310663186. |
865 | 22.6k | if (cpi->oxcf.mode != cpi->deadline_mode_previous_frame) |
866 | 0 | sf->nonrd_keyframe = 1; |
867 | 22.6k | } |
868 | | |
869 | 45.2k | void vp9_set_speed_features_framesize_dependent(VP9_COMP *cpi, int speed) { |
870 | 45.2k | SPEED_FEATURES *const sf = &cpi->sf; |
871 | 45.2k | const VP9EncoderConfig *const oxcf = &cpi->oxcf; |
872 | 45.2k | RD_OPT *const rd = &cpi->rd; |
873 | 45.2k | int i; |
874 | | |
875 | | // best quality defaults |
876 | | // Some speed-up features even for best quality as minimal impact on quality. |
877 | 45.2k | sf->partition_search_breakout_thr.dist = (1 << 19); |
878 | 45.2k | sf->partition_search_breakout_thr.rate = 80; |
879 | 45.2k | sf->rd_ml_partition.search_early_termination = 0; |
880 | 45.2k | sf->rd_ml_partition.search_breakout = 0; |
881 | | |
882 | 45.2k | if (oxcf->mode == REALTIME) |
883 | 22.6k | set_rt_speed_feature_framesize_dependent(cpi, sf, speed); |
884 | 22.6k | #if !CONFIG_REALTIME_ONLY |
885 | 22.6k | else if (oxcf->mode == GOOD) |
886 | 22.6k | set_good_speed_feature_framesize_dependent(cpi, sf, speed); |
887 | 45.2k | #endif |
888 | | |
889 | 45.2k | if (sf->disable_split_mask == DISABLE_ALL_SPLIT) { |
890 | 0 | sf->adaptive_pred_interp_filter = 0; |
891 | 0 | } |
892 | | |
893 | 45.2k | if (cpi->encode_breakout && oxcf->mode == REALTIME && |
894 | 45.2k | sf->encode_breakout_thresh > cpi->encode_breakout) { |
895 | 0 | cpi->encode_breakout = sf->encode_breakout_thresh; |
896 | 0 | } |
897 | | |
898 | | // Check for masked out split cases. |
899 | 316k | for (i = 0; i < MAX_REFS; ++i) { |
900 | 271k | if (sf->disable_split_mask & (1 << i)) { |
901 | 84.5k | rd->thresh_mult_sub8x8[i] = INT_MAX; |
902 | 84.5k | } |
903 | 271k | } |
904 | | |
905 | | // With row based multi-threading, the following speed features |
906 | | // have to be disabled to guarantee that bitstreams encoded with single thread |
907 | | // and multiple threads match. |
908 | | // It can be used in realtime when adaptive_rd_thresh_row_mt is enabled since |
909 | | // adaptive_rd_thresh is defined per-row for non-rd pickmode. |
910 | 45.2k | if (!sf->adaptive_rd_thresh_row_mt && cpi->row_mt_bit_exact && |
911 | 45.2k | oxcf->max_threads > 1) |
912 | 0 | sf->adaptive_rd_thresh = 0; |
913 | 45.2k | } |
914 | | |
915 | 45.2k | void vp9_set_speed_features_framesize_independent(VP9_COMP *cpi, int speed) { |
916 | 45.2k | SPEED_FEATURES *const sf = &cpi->sf; |
917 | 45.2k | #if !CONFIG_REALTIME_ONLY |
918 | 45.2k | VP9_COMMON *const cm = &cpi->common; |
919 | 45.2k | #endif |
920 | 45.2k | MACROBLOCK *const x = &cpi->td.mb; |
921 | 45.2k | const VP9EncoderConfig *const oxcf = &cpi->oxcf; |
922 | 45.2k | int i; |
923 | | |
924 | | // best quality defaults |
925 | 45.2k | sf->frame_parameter_update = 1; |
926 | 45.2k | sf->mv.search_method = NSTEP; |
927 | 45.2k | sf->recode_loop = ALLOW_RECODE_FIRST; |
928 | 45.2k | sf->mv.subpel_search_method = SUBPEL_TREE; |
929 | 45.2k | sf->mv.subpel_search_level = 2; |
930 | 45.2k | sf->mv.subpel_force_stop = EIGHTH_PEL; |
931 | 45.2k | sf->optimize_coefficients = !is_lossless_requested(&cpi->oxcf); |
932 | 45.2k | sf->mv.reduce_first_step_size = 0; |
933 | 45.2k | sf->coeff_prob_appx_step = 1; |
934 | 45.2k | sf->mv.auto_mv_step_size = 0; |
935 | 45.2k | sf->mv.fullpel_search_step_param = 6; |
936 | 45.2k | sf->mv.use_downsampled_sad = 0; |
937 | 45.2k | sf->comp_inter_joint_search_iter_level = 0; |
938 | 45.2k | sf->tx_size_search_method = USE_FULL_RD; |
939 | 45.2k | sf->use_lp32x32fdct = 0; |
940 | 45.2k | sf->adaptive_motion_search = 0; |
941 | 45.2k | sf->enhanced_full_pixel_motion_search = 1; |
942 | 45.2k | sf->adaptive_pred_interp_filter = 0; |
943 | 45.2k | sf->adaptive_mode_search = 0; |
944 | 45.2k | sf->prune_single_mode_based_on_mv_diff_mode_rate = 0; |
945 | 45.2k | sf->cb_pred_filter_search = 0; |
946 | 45.2k | sf->early_term_interp_search_plane_rd = 0; |
947 | 45.2k | sf->cb_partition_search = 0; |
948 | 45.2k | sf->motion_field_mode_search = 0; |
949 | 45.2k | sf->alt_ref_search_fp = 0; |
950 | 45.2k | sf->use_quant_fp = 0; |
951 | 45.2k | sf->reference_masking = 0; |
952 | 45.2k | sf->partition_search_type = SEARCH_PARTITION; |
953 | 45.2k | sf->less_rectangular_check = 0; |
954 | 45.2k | sf->use_square_partition_only = 0; |
955 | 45.2k | sf->use_square_only_thresh_high = BLOCK_SIZES; |
956 | 45.2k | sf->use_square_only_thresh_low = BLOCK_4X4; |
957 | 45.2k | sf->auto_min_max_partition_size = NOT_IN_USE; |
958 | 45.2k | sf->rd_auto_partition_min_limit = BLOCK_4X4; |
959 | 45.2k | sf->default_max_partition_size = BLOCK_64X64; |
960 | 45.2k | sf->default_min_partition_size = BLOCK_4X4; |
961 | 45.2k | sf->adjust_partitioning_from_last_frame = 0; |
962 | 45.2k | sf->last_partitioning_redo_frequency = 4; |
963 | 45.2k | sf->disable_split_mask = 0; |
964 | 45.2k | sf->mode_search_skip_flags = 0; |
965 | 45.2k | sf->force_frame_boost = 0; |
966 | 45.2k | sf->max_delta_qindex = 0; |
967 | 45.2k | sf->disable_filter_search_var_thresh = 0; |
968 | 45.2k | sf->adaptive_interp_filter_search = 0; |
969 | 45.2k | sf->allow_txfm_domain_distortion = 0; |
970 | 45.2k | sf->tx_domain_thresh = 99.0; |
971 | 45.2k | sf->trellis_opt_tx_rd.method = |
972 | 45.2k | sf->optimize_coefficients ? ENABLE_TRELLIS_OPT : DISABLE_TRELLIS_OPT; |
973 | 45.2k | sf->trellis_opt_tx_rd.thresh = 99.0; |
974 | 45.2k | sf->allow_acl = 1; |
975 | 45.2k | sf->enable_tpl_model = oxcf->enable_tpl_model; |
976 | 45.2k | sf->prune_ref_frame_for_rect_partitions = 0; |
977 | 45.2k | sf->temporal_filter_search_method = MESH; |
978 | 45.2k | sf->allow_skip_txfm_ac_dc = 0; |
979 | | |
980 | 226k | for (i = 0; i < TX_SIZES; i++) { |
981 | 181k | sf->intra_y_mode_mask[i] = INTRA_ALL; |
982 | 181k | sf->intra_uv_mode_mask[i] = INTRA_ALL; |
983 | 181k | } |
984 | 45.2k | sf->use_rd_breakout = 0; |
985 | 45.2k | sf->skip_encode_sb = 0; |
986 | 45.2k | sf->use_uv_intra_rd_estimate = 0; |
987 | 45.2k | sf->allow_skip_recode = 0; |
988 | 45.2k | sf->lpf_pick = LPF_PICK_FROM_FULL_IMAGE; |
989 | 45.2k | sf->use_fast_coef_updates = TWO_LOOP; |
990 | 45.2k | sf->use_fast_coef_costing = 0; |
991 | 45.2k | sf->mode_skip_start = MAX_MODES; // Mode index at which mode skip mask set |
992 | 45.2k | sf->schedule_mode_search = 0; |
993 | 45.2k | sf->use_nonrd_pick_mode = 0; |
994 | 633k | for (i = 0; i < BLOCK_SIZES; ++i) sf->inter_mode_mask[i] = INTER_ALL; |
995 | 45.2k | sf->max_intra_bsize = BLOCK_64X64; |
996 | 45.2k | sf->reuse_inter_pred_sby = 0; |
997 | | // This setting only takes effect when partition_search_type is set |
998 | | // to FIXED_PARTITION. |
999 | 45.2k | sf->always_this_block_size = BLOCK_16X16; |
1000 | 45.2k | sf->search_type_check_frequency = 50; |
1001 | 45.2k | sf->encode_breakout_thresh = 0; |
1002 | | // Recode loop tolerance %. |
1003 | 45.2k | sf->recode_tolerance_low = 12; |
1004 | 45.2k | sf->recode_tolerance_high = 25; |
1005 | 45.2k | sf->default_interp_filter = SWITCHABLE; |
1006 | 45.2k | sf->simple_model_rd_from_var = 0; |
1007 | 45.2k | sf->short_circuit_flat_blocks = 0; |
1008 | 45.2k | sf->short_circuit_low_temp_var = 0; |
1009 | 45.2k | sf->limit_newmv_early_exit = 0; |
1010 | 45.2k | sf->bias_golden = 0; |
1011 | 45.2k | sf->base_mv_aggressive = 0; |
1012 | 45.2k | sf->rd_ml_partition.prune_rect_thresh[0] = -1; |
1013 | 45.2k | sf->rd_ml_partition.prune_rect_thresh[1] = -1; |
1014 | 45.2k | sf->rd_ml_partition.prune_rect_thresh[2] = -1; |
1015 | 45.2k | sf->rd_ml_partition.prune_rect_thresh[3] = -1; |
1016 | 45.2k | sf->rd_ml_partition.var_pruning = 0; |
1017 | 45.2k | sf->use_accurate_subpel_search = USE_8_TAPS; |
1018 | | |
1019 | | // Some speed-up features even for best quality as minimal impact on quality. |
1020 | 45.2k | sf->adaptive_rd_thresh = 1; |
1021 | 45.2k | sf->tx_size_search_breakout = 1; |
1022 | 45.2k | sf->tx_size_search_depth = 2; |
1023 | | |
1024 | 45.2k | sf->exhaustive_searches_thresh = |
1025 | 45.2k | (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ? (1 << 20) |
1026 | 45.2k | : INT_MAX; |
1027 | 45.2k | { |
1028 | 45.2k | const int mesh_density_level = |
1029 | 45.2k | (cpi->twopass.fr_content_type == FC_GRAPHICS_ANIMATION) ? 0 : 1; |
1030 | 226k | for (i = 0; i < MAX_MESH_STEP; ++i) { |
1031 | 181k | sf->mesh_patterns[i].range = |
1032 | 181k | best_quality_mesh_pattern[mesh_density_level][i].range; |
1033 | 181k | sf->mesh_patterns[i].interval = |
1034 | 181k | best_quality_mesh_pattern[mesh_density_level][i].interval; |
1035 | 181k | } |
1036 | 45.2k | } |
1037 | | |
1038 | 45.2k | if (oxcf->mode == REALTIME) |
1039 | 22.6k | set_rt_speed_feature_framesize_independent(cpi, sf, speed, oxcf->content); |
1040 | 22.6k | #if !CONFIG_REALTIME_ONLY |
1041 | 22.6k | else if (oxcf->mode == GOOD) |
1042 | 22.6k | set_good_speed_feature_framesize_independent(cpi, cm, sf, speed); |
1043 | 45.2k | #endif |
1044 | | |
1045 | 45.2k | cpi->diamond_search_sad = vp9_diamond_search_sad; |
1046 | | |
1047 | | // Slow quant, dct and trellis not worthwhile for first pass |
1048 | | // so make sure they are always turned off. |
1049 | 45.2k | if (oxcf->pass == 1) sf->optimize_coefficients = 0; |
1050 | | |
1051 | | // No recode for 1 pass. |
1052 | 45.2k | if (oxcf->pass == 0) { |
1053 | 45.2k | sf->recode_loop = DISALLOW_RECODE; |
1054 | 45.2k | sf->optimize_coefficients = 0; |
1055 | 45.2k | } |
1056 | | |
1057 | 45.2k | if (sf->mv.subpel_force_stop == FULL_PEL) { |
1058 | | // Whole pel only |
1059 | 0 | cpi->find_fractional_mv_step = vp9_skip_sub_pixel_tree; |
1060 | 45.2k | } else if (sf->mv.subpel_search_method == SUBPEL_TREE) { |
1061 | 45.2k | cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree; |
1062 | 45.2k | } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED) { |
1063 | 0 | cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned; |
1064 | 0 | } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED_MORE) { |
1065 | 0 | cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned_more; |
1066 | 0 | } else if (sf->mv.subpel_search_method == SUBPEL_TREE_PRUNED_EVENMORE) { |
1067 | 0 | cpi->find_fractional_mv_step = vp9_find_best_sub_pixel_tree_pruned_evenmore; |
1068 | 0 | } |
1069 | | |
1070 | | // This is only used in motion vector unit test. |
1071 | 45.2k | if (cpi->oxcf.motion_vector_unit_test == 1) |
1072 | 0 | cpi->find_fractional_mv_step = vp9_return_max_sub_pixel_mv; |
1073 | 45.2k | else if (cpi->oxcf.motion_vector_unit_test == 2) |
1074 | 0 | cpi->find_fractional_mv_step = vp9_return_min_sub_pixel_mv; |
1075 | | |
1076 | 45.2k | x->optimize = sf->optimize_coefficients == 1 && oxcf->pass != 1; |
1077 | | |
1078 | 45.2k | x->min_partition_size = sf->default_min_partition_size; |
1079 | 45.2k | x->max_partition_size = sf->default_max_partition_size; |
1080 | | |
1081 | 45.2k | if (!cpi->oxcf.frame_periodic_boost) { |
1082 | 45.2k | sf->max_delta_qindex = 0; |
1083 | 45.2k | } |
1084 | | |
1085 | | // With row based multi-threading, the following speed features |
1086 | | // have to be disabled to guarantee that bitstreams encoded with single thread |
1087 | | // and multiple threads match. |
1088 | | // It can be used in realtime when adaptive_rd_thresh_row_mt is enabled since |
1089 | | // adaptive_rd_thresh is defined per-row for non-rd pickmode. |
1090 | 45.2k | if (!sf->adaptive_rd_thresh_row_mt && cpi->row_mt_bit_exact && |
1091 | 45.2k | oxcf->max_threads > 1) |
1092 | 0 | sf->adaptive_rd_thresh = 0; |
1093 | 45.2k | } |