OLD | NEW |
| (Empty) |
1 /* | |
2 * Copyright (c) 2014 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 "./vpx_config.h" | |
12 | |
13 #include "vpx_mem/vpx_mem.h" | |
14 | |
15 #include "vp9/common/vp9_reconinter.h" | |
16 | |
17 #include "vp9/decoder/vp9_dthread.h" | |
18 #include "vp9/decoder/vp9_decoder.h" | |
19 | |
20 #if CONFIG_MULTITHREAD | |
21 static INLINE void mutex_lock(pthread_mutex_t *const mutex) { | |
22 const int kMaxTryLocks = 4000; | |
23 int locked = 0; | |
24 int i; | |
25 | |
26 for (i = 0; i < kMaxTryLocks; ++i) { | |
27 if (!pthread_mutex_trylock(mutex)) { | |
28 locked = 1; | |
29 break; | |
30 } | |
31 } | |
32 | |
33 if (!locked) | |
34 pthread_mutex_lock(mutex); | |
35 } | |
36 #endif // CONFIG_MULTITHREAD | |
37 | |
38 static INLINE void sync_read(VP9LfSync *const lf_sync, int r, int c) { | |
39 #if CONFIG_MULTITHREAD | |
40 const int nsync = lf_sync->sync_range; | |
41 | |
42 if (r && !(c & (nsync - 1))) { | |
43 pthread_mutex_t *const mutex = &lf_sync->mutex_[r - 1]; | |
44 mutex_lock(mutex); | |
45 | |
46 while (c > lf_sync->cur_sb_col[r - 1] - nsync) { | |
47 pthread_cond_wait(&lf_sync->cond_[r - 1], mutex); | |
48 } | |
49 pthread_mutex_unlock(mutex); | |
50 } | |
51 #else | |
52 (void)lf_sync; | |
53 (void)r; | |
54 (void)c; | |
55 #endif // CONFIG_MULTITHREAD | |
56 } | |
57 | |
58 static INLINE void sync_write(VP9LfSync *const lf_sync, int r, int c, | |
59 const int sb_cols) { | |
60 #if CONFIG_MULTITHREAD | |
61 const int nsync = lf_sync->sync_range; | |
62 int cur; | |
63 // Only signal when there are enough filtered SB for next row to run. | |
64 int sig = 1; | |
65 | |
66 if (c < sb_cols - 1) { | |
67 cur = c; | |
68 if (c % nsync) | |
69 sig = 0; | |
70 } else { | |
71 cur = sb_cols + nsync; | |
72 } | |
73 | |
74 if (sig) { | |
75 mutex_lock(&lf_sync->mutex_[r]); | |
76 | |
77 lf_sync->cur_sb_col[r] = cur; | |
78 | |
79 pthread_cond_signal(&lf_sync->cond_[r]); | |
80 pthread_mutex_unlock(&lf_sync->mutex_[r]); | |
81 } | |
82 #else | |
83 (void)lf_sync; | |
84 (void)r; | |
85 (void)c; | |
86 (void)sb_cols; | |
87 #endif // CONFIG_MULTITHREAD | |
88 } | |
89 | |
90 // Implement row loopfiltering for each thread. | |
91 static void loop_filter_rows_mt(const YV12_BUFFER_CONFIG *const frame_buffer, | |
92 VP9_COMMON *const cm, | |
93 struct macroblockd_plane planes[MAX_MB_PLANE], | |
94 int start, int stop, int y_only, | |
95 VP9LfSync *const lf_sync) { | |
96 const int num_planes = y_only ? 1 : MAX_MB_PLANE; | |
97 int r, c; // SB row and col | |
98 const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2; | |
99 | |
100 for (r = start; r < stop; r += lf_sync->num_workers) { | |
101 const int mi_row = r << MI_BLOCK_SIZE_LOG2; | |
102 MODE_INFO *const mi = cm->mi + mi_row * cm->mi_stride; | |
103 | |
104 for (c = 0; c < sb_cols; ++c) { | |
105 const int mi_col = c << MI_BLOCK_SIZE_LOG2; | |
106 LOOP_FILTER_MASK lfm; | |
107 int plane; | |
108 | |
109 sync_read(lf_sync, r, c); | |
110 | |
111 vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col); | |
112 vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride, &lfm); | |
113 | |
114 for (plane = 0; plane < num_planes; ++plane) { | |
115 vp9_filter_block_plane(cm, &planes[plane], mi_row, &lfm); | |
116 } | |
117 | |
118 sync_write(lf_sync, r, c, sb_cols); | |
119 } | |
120 } | |
121 } | |
122 | |
123 // Row-based multi-threaded loopfilter hook | |
124 static int loop_filter_row_worker(VP9LfSync *const lf_sync, | |
125 LFWorkerData *const lf_data) { | |
126 loop_filter_rows_mt(lf_data->frame_buffer, lf_data->cm, lf_data->planes, | |
127 lf_data->start, lf_data->stop, lf_data->y_only, lf_sync); | |
128 return 1; | |
129 } | |
130 | |
131 // VP9 decoder: Implement multi-threaded loopfilter that uses the tile | |
132 // threads. | |
133 void vp9_loop_filter_frame_mt(VP9LfSync *lf_sync, | |
134 YV12_BUFFER_CONFIG *frame, | |
135 struct macroblockd_plane planes[MAX_MB_PLANE], | |
136 VP9_COMMON *cm, | |
137 VP9Worker *workers, int nworkers, | |
138 int frame_filter_level, | |
139 int y_only) { | |
140 const VP9WorkerInterface *const winterface = vp9_get_worker_interface(); | |
141 // Number of superblock rows and cols | |
142 const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2; | |
143 const int tile_cols = 1 << cm->log2_tile_cols; | |
144 const int num_workers = MIN(nworkers, tile_cols); | |
145 int i; | |
146 | |
147 if (!frame_filter_level) return; | |
148 | |
149 if (!lf_sync->sync_range || cm->last_height != cm->height || | |
150 num_workers > lf_sync->num_workers) { | |
151 vp9_loop_filter_dealloc(lf_sync); | |
152 vp9_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers); | |
153 } | |
154 | |
155 vp9_loop_filter_frame_init(cm, frame_filter_level); | |
156 | |
157 // Initialize cur_sb_col to -1 for all SB rows. | |
158 vpx_memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows); | |
159 | |
160 // Set up loopfilter thread data. | |
161 // The decoder is capping num_workers because it has been observed that using | |
162 // more threads on the loopfilter than there are cores will hurt performance | |
163 // on Android. This is because the system will only schedule the tile decode | |
164 // workers on cores equal to the number of tile columns. Then if the decoder | |
165 // tries to use more threads for the loopfilter, it will hurt performance | |
166 // because of contention. If the multithreading code changes in the future | |
167 // then the number of workers used by the loopfilter should be revisited. | |
168 for (i = 0; i < num_workers; ++i) { | |
169 VP9Worker *const worker = &workers[i]; | |
170 LFWorkerData *const lf_data = &lf_sync->lfdata[i]; | |
171 | |
172 worker->hook = (VP9WorkerHook)loop_filter_row_worker; | |
173 worker->data1 = lf_sync; | |
174 worker->data2 = lf_data; | |
175 | |
176 // Loopfilter data | |
177 vp9_loop_filter_data_reset(lf_data, frame, cm, planes); | |
178 lf_data->start = i; | |
179 lf_data->stop = sb_rows; | |
180 lf_data->y_only = y_only; | |
181 | |
182 // Start loopfiltering | |
183 if (i == num_workers - 1) { | |
184 winterface->execute(worker); | |
185 } else { | |
186 winterface->launch(worker); | |
187 } | |
188 } | |
189 | |
190 // Wait till all rows are finished | |
191 for (i = 0; i < num_workers; ++i) { | |
192 winterface->sync(&workers[i]); | |
193 } | |
194 } | |
195 | |
196 // Set up nsync by width. | |
197 static int get_sync_range(int width) { | |
198 // nsync numbers are picked by testing. For example, for 4k | |
199 // video, using 4 gives best performance. | |
200 if (width < 640) | |
201 return 1; | |
202 else if (width <= 1280) | |
203 return 2; | |
204 else if (width <= 4096) | |
205 return 4; | |
206 else | |
207 return 8; | |
208 } | |
209 | |
210 // Allocate memory for lf row synchronization | |
211 void vp9_loop_filter_alloc(VP9LfSync *lf_sync, VP9_COMMON *cm, int rows, | |
212 int width, int num_workers) { | |
213 lf_sync->rows = rows; | |
214 #if CONFIG_MULTITHREAD | |
215 { | |
216 int i; | |
217 | |
218 CHECK_MEM_ERROR(cm, lf_sync->mutex_, | |
219 vpx_malloc(sizeof(*lf_sync->mutex_) * rows)); | |
220 if (lf_sync->mutex_) { | |
221 for (i = 0; i < rows; ++i) { | |
222 pthread_mutex_init(&lf_sync->mutex_[i], NULL); | |
223 } | |
224 } | |
225 | |
226 CHECK_MEM_ERROR(cm, lf_sync->cond_, | |
227 vpx_malloc(sizeof(*lf_sync->cond_) * rows)); | |
228 if (lf_sync->cond_) { | |
229 for (i = 0; i < rows; ++i) { | |
230 pthread_cond_init(&lf_sync->cond_[i], NULL); | |
231 } | |
232 } | |
233 } | |
234 #endif // CONFIG_MULTITHREAD | |
235 | |
236 CHECK_MEM_ERROR(cm, lf_sync->lfdata, | |
237 vpx_malloc(num_workers * sizeof(*lf_sync->lfdata))); | |
238 lf_sync->num_workers = num_workers; | |
239 | |
240 CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col, | |
241 vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows)); | |
242 | |
243 // Set up nsync. | |
244 lf_sync->sync_range = get_sync_range(width); | |
245 } | |
246 | |
247 // Deallocate lf synchronization related mutex and data | |
248 void vp9_loop_filter_dealloc(VP9LfSync *lf_sync) { | |
249 if (lf_sync != NULL) { | |
250 #if CONFIG_MULTITHREAD | |
251 int i; | |
252 | |
253 if (lf_sync->mutex_ != NULL) { | |
254 for (i = 0; i < lf_sync->rows; ++i) { | |
255 pthread_mutex_destroy(&lf_sync->mutex_[i]); | |
256 } | |
257 vpx_free(lf_sync->mutex_); | |
258 } | |
259 if (lf_sync->cond_ != NULL) { | |
260 for (i = 0; i < lf_sync->rows; ++i) { | |
261 pthread_cond_destroy(&lf_sync->cond_[i]); | |
262 } | |
263 vpx_free(lf_sync->cond_); | |
264 } | |
265 #endif // CONFIG_MULTITHREAD | |
266 vpx_free(lf_sync->lfdata); | |
267 vpx_free(lf_sync->cur_sb_col); | |
268 // clear the structure as the source of this call may be a resize in which | |
269 // case this call will be followed by an _alloc() which may fail. | |
270 vp9_zero(*lf_sync); | |
271 } | |
272 } | |
OLD | NEW |