OLD | NEW |
1 /* | 1 /* |
2 * Copyright (c) 2014 The WebM project authors. All Rights Reserved. | 2 * Copyright (c) 2014 The WebM project authors. All Rights Reserved. |
3 * | 3 * |
4 * Use of this source code is governed by a BSD-style license | 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 | 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 | 6 * tree. An additional intellectual property rights grant can be found |
7 * in the file PATENTS. All contributing project authors may | 7 * in the file PATENTS. All contributing project authors may |
8 * be found in the AUTHORS file in the root of the source tree. | 8 * be found in the AUTHORS file in the root of the source tree. |
9 */ | 9 */ |
10 | 10 |
11 #include "./vpx_config.h" | 11 #include "./vpx_config.h" |
12 | |
13 #include "vpx_mem/vpx_mem.h" | 12 #include "vpx_mem/vpx_mem.h" |
14 | 13 #include "vp9/common/vp9_loopfilter_thread.h" |
15 #include "vp9/common/vp9_reconinter.h" | 14 #include "vp9/common/vp9_reconinter.h" |
16 | 15 |
17 #include "vp9/decoder/vp9_dthread.h" | |
18 #include "vp9/decoder/vp9_decoder.h" | |
19 | |
20 #if CONFIG_MULTITHREAD | 16 #if CONFIG_MULTITHREAD |
21 static INLINE void mutex_lock(pthread_mutex_t *const mutex) { | 17 static INLINE void mutex_lock(pthread_mutex_t *const mutex) { |
22 const int kMaxTryLocks = 4000; | 18 const int kMaxTryLocks = 4000; |
23 int locked = 0; | 19 int locked = 0; |
24 int i; | 20 int i; |
25 | 21 |
26 for (i = 0; i < kMaxTryLocks; ++i) { | 22 for (i = 0; i < kMaxTryLocks; ++i) { |
27 if (!pthread_mutex_trylock(mutex)) { | 23 if (!pthread_mutex_trylock(mutex)) { |
28 locked = 1; | 24 locked = 1; |
29 break; | 25 break; |
(...skipping 51 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
81 } | 77 } |
82 #else | 78 #else |
83 (void)lf_sync; | 79 (void)lf_sync; |
84 (void)r; | 80 (void)r; |
85 (void)c; | 81 (void)c; |
86 (void)sb_cols; | 82 (void)sb_cols; |
87 #endif // CONFIG_MULTITHREAD | 83 #endif // CONFIG_MULTITHREAD |
88 } | 84 } |
89 | 85 |
90 // Implement row loopfiltering for each thread. | 86 // Implement row loopfiltering for each thread. |
91 static void loop_filter_rows_mt(const YV12_BUFFER_CONFIG *const frame_buffer, | 87 static INLINE |
92 VP9_COMMON *const cm, | 88 void thread_loop_filter_rows(const YV12_BUFFER_CONFIG *const frame_buffer, |
93 struct macroblockd_plane planes[MAX_MB_PLANE], | 89 VP9_COMMON *const cm, |
94 int start, int stop, int y_only, | 90 struct macroblockd_plane planes[MAX_MB_PLANE], |
95 VP9LfSync *const lf_sync) { | 91 int start, int stop, int y_only, |
| 92 VP9LfSync *const lf_sync) { |
96 const int num_planes = y_only ? 1 : MAX_MB_PLANE; | 93 const int num_planes = y_only ? 1 : MAX_MB_PLANE; |
97 int r, c; // SB row and col | 94 const int use_420 = y_only || (planes[1].subsampling_y == 1 && |
| 95 planes[1].subsampling_x == 1); |
98 const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2; | 96 const int sb_cols = mi_cols_aligned_to_sb(cm->mi_cols) >> MI_BLOCK_SIZE_LOG2; |
| 97 int mi_row, mi_col; |
99 | 98 |
100 for (r = start; r < stop; r += lf_sync->num_workers) { | 99 for (mi_row = start; mi_row < stop; |
101 const int mi_row = r << MI_BLOCK_SIZE_LOG2; | 100 mi_row += lf_sync->num_workers * MI_BLOCK_SIZE) { |
102 MODE_INFO *const mi = cm->mi + mi_row * cm->mi_stride; | 101 MODE_INFO *const mi = cm->mi + mi_row * cm->mi_stride; |
103 | 102 |
104 for (c = 0; c < sb_cols; ++c) { | 103 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) { |
105 const int mi_col = c << MI_BLOCK_SIZE_LOG2; | 104 const int r = mi_row >> MI_BLOCK_SIZE_LOG2; |
| 105 const int c = mi_col >> MI_BLOCK_SIZE_LOG2; |
106 LOOP_FILTER_MASK lfm; | 106 LOOP_FILTER_MASK lfm; |
107 int plane; | 107 int plane; |
108 | 108 |
109 sync_read(lf_sync, r, c); | 109 sync_read(lf_sync, r, c); |
110 | 110 |
111 vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col); | 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); | 112 |
| 113 // TODO(JBB): Make setup_mask work for non 420. |
| 114 if (use_420) |
| 115 vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride, |
| 116 &lfm); |
113 | 117 |
114 for (plane = 0; plane < num_planes; ++plane) { | 118 for (plane = 0; plane < num_planes; ++plane) { |
115 vp9_filter_block_plane(cm, &planes[plane], mi_row, &lfm); | 119 if (use_420) |
| 120 vp9_filter_block_plane(cm, &planes[plane], mi_row, &lfm); |
| 121 else |
| 122 vp9_filter_block_plane_non420(cm, &planes[plane], mi + mi_col, |
| 123 mi_row, mi_col); |
116 } | 124 } |
117 | 125 |
118 sync_write(lf_sync, r, c, sb_cols); | 126 sync_write(lf_sync, r, c, sb_cols); |
119 } | 127 } |
120 } | 128 } |
121 } | 129 } |
122 | 130 |
123 // Row-based multi-threaded loopfilter hook | 131 // Row-based multi-threaded loopfilter hook |
124 static int loop_filter_row_worker(VP9LfSync *const lf_sync, | 132 static int loop_filter_row_worker(VP9LfSync *const lf_sync, |
125 LFWorkerData *const lf_data) { | 133 LFWorkerData *const lf_data) { |
126 loop_filter_rows_mt(lf_data->frame_buffer, lf_data->cm, lf_data->planes, | 134 thread_loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes, |
127 lf_data->start, lf_data->stop, lf_data->y_only, lf_sync); | 135 lf_data->start, lf_data->stop, lf_data->y_only, |
| 136 lf_sync); |
128 return 1; | 137 return 1; |
129 } | 138 } |
130 | 139 |
131 // VP9 decoder: Implement multi-threaded loopfilter that uses the tile | 140 static void loop_filter_rows_mt(YV12_BUFFER_CONFIG *frame, |
132 // threads. | 141 VP9_COMMON *cm, |
133 void vp9_loop_filter_frame_mt(VP9LfSync *lf_sync, | 142 struct macroblockd_plane planes[MAX_MB_PLANE], |
134 YV12_BUFFER_CONFIG *frame, | 143 int start, int stop, int y_only, |
135 struct macroblockd_plane planes[MAX_MB_PLANE], | 144 VP9Worker *workers, int nworkers, |
136 VP9_COMMON *cm, | 145 VP9LfSync *lf_sync) { |
137 VP9Worker *workers, int nworkers, | |
138 int frame_filter_level, | |
139 int y_only) { | |
140 const VP9WorkerInterface *const winterface = vp9_get_worker_interface(); | 146 const VP9WorkerInterface *const winterface = vp9_get_worker_interface(); |
141 // Number of superblock rows and cols | 147 // Number of superblock rows and cols |
142 const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2; | 148 const int sb_rows = mi_cols_aligned_to_sb(cm->mi_rows) >> MI_BLOCK_SIZE_LOG2; |
| 149 // Decoder may allocate more threads than number of tiles based on user's |
| 150 // input. |
143 const int tile_cols = 1 << cm->log2_tile_cols; | 151 const int tile_cols = 1 << cm->log2_tile_cols; |
144 const int num_workers = MIN(nworkers, tile_cols); | 152 const int num_workers = MIN(nworkers, tile_cols); |
145 int i; | 153 int i; |
146 | 154 |
147 if (!frame_filter_level) return; | |
148 | |
149 if (!lf_sync->sync_range || cm->last_height != cm->height || | 155 if (!lf_sync->sync_range || cm->last_height != cm->height || |
150 num_workers > lf_sync->num_workers) { | 156 num_workers > lf_sync->num_workers) { |
151 vp9_loop_filter_dealloc(lf_sync); | 157 vp9_loop_filter_dealloc(lf_sync); |
152 vp9_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers); | 158 vp9_loop_filter_alloc(lf_sync, cm, sb_rows, cm->width, num_workers); |
153 } | 159 } |
154 | 160 |
155 vp9_loop_filter_frame_init(cm, frame_filter_level); | |
156 | |
157 // Initialize cur_sb_col to -1 for all SB rows. | 161 // 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); | 162 vpx_memset(lf_sync->cur_sb_col, -1, sizeof(*lf_sync->cur_sb_col) * sb_rows); |
159 | 163 |
160 // Set up loopfilter thread data. | 164 // Set up loopfilter thread data. |
161 // The decoder is capping num_workers because it has been observed that using | 165 // 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 | 166 // 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 | 167 // 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 | 168 // 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 | 169 // tries to use more threads for the loopfilter, it will hurt performance |
166 // because of contention. If the multithreading code changes in the future | 170 // because of contention. If the multithreading code changes in the future |
167 // then the number of workers used by the loopfilter should be revisited. | 171 // then the number of workers used by the loopfilter should be revisited. |
168 for (i = 0; i < num_workers; ++i) { | 172 for (i = 0; i < num_workers; ++i) { |
169 VP9Worker *const worker = &workers[i]; | 173 VP9Worker *const worker = &workers[i]; |
170 LFWorkerData *const lf_data = &lf_sync->lfdata[i]; | 174 LFWorkerData *const lf_data = &lf_sync->lfdata[i]; |
171 | 175 |
172 worker->hook = (VP9WorkerHook)loop_filter_row_worker; | 176 worker->hook = (VP9WorkerHook)loop_filter_row_worker; |
173 worker->data1 = lf_sync; | 177 worker->data1 = lf_sync; |
174 worker->data2 = lf_data; | 178 worker->data2 = lf_data; |
175 | 179 |
176 // Loopfilter data | 180 // Loopfilter data |
177 vp9_loop_filter_data_reset(lf_data, frame, cm, planes); | 181 vp9_loop_filter_data_reset(lf_data, frame, cm, planes); |
178 lf_data->start = i; | 182 lf_data->start = start + i * MI_BLOCK_SIZE; |
179 lf_data->stop = sb_rows; | 183 lf_data->stop = stop; |
180 lf_data->y_only = y_only; | 184 lf_data->y_only = y_only; |
181 | 185 |
182 // Start loopfiltering | 186 // Start loopfiltering |
183 if (i == num_workers - 1) { | 187 if (i == num_workers - 1) { |
184 winterface->execute(worker); | 188 winterface->execute(worker); |
185 } else { | 189 } else { |
186 winterface->launch(worker); | 190 winterface->launch(worker); |
187 } | 191 } |
188 } | 192 } |
189 | 193 |
190 // Wait till all rows are finished | 194 // Wait till all rows are finished |
191 for (i = 0; i < num_workers; ++i) { | 195 for (i = 0; i < num_workers; ++i) { |
192 winterface->sync(&workers[i]); | 196 winterface->sync(&workers[i]); |
193 } | 197 } |
194 } | 198 } |
195 | 199 |
| 200 void vp9_loop_filter_frame_mt(YV12_BUFFER_CONFIG *frame, |
| 201 VP9_COMMON *cm, |
| 202 struct macroblockd_plane planes[MAX_MB_PLANE], |
| 203 int frame_filter_level, |
| 204 int y_only, int partial_frame, |
| 205 VP9Worker *workers, int num_workers, |
| 206 VP9LfSync *lf_sync) { |
| 207 int start_mi_row, end_mi_row, mi_rows_to_filter; |
| 208 |
| 209 if (!frame_filter_level) return; |
| 210 |
| 211 start_mi_row = 0; |
| 212 mi_rows_to_filter = cm->mi_rows; |
| 213 if (partial_frame && cm->mi_rows > 8) { |
| 214 start_mi_row = cm->mi_rows >> 1; |
| 215 start_mi_row &= 0xfffffff8; |
| 216 mi_rows_to_filter = MAX(cm->mi_rows / 8, 8); |
| 217 } |
| 218 end_mi_row = start_mi_row + mi_rows_to_filter; |
| 219 vp9_loop_filter_frame_init(cm, frame_filter_level); |
| 220 |
| 221 loop_filter_rows_mt(frame, cm, planes, start_mi_row, end_mi_row, |
| 222 y_only, workers, num_workers, lf_sync); |
| 223 } |
| 224 |
196 // Set up nsync by width. | 225 // Set up nsync by width. |
197 static int get_sync_range(int width) { | 226 static INLINE int get_sync_range(int width) { |
198 // nsync numbers are picked by testing. For example, for 4k | 227 // nsync numbers are picked by testing. For example, for 4k |
199 // video, using 4 gives best performance. | 228 // video, using 4 gives best performance. |
200 if (width < 640) | 229 if (width < 640) |
201 return 1; | 230 return 1; |
202 else if (width <= 1280) | 231 else if (width <= 1280) |
203 return 2; | 232 return 2; |
204 else if (width <= 4096) | 233 else if (width <= 4096) |
205 return 4; | 234 return 4; |
206 else | 235 else |
207 return 8; | 236 return 8; |
(...skipping 55 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
263 vpx_free(lf_sync->cond_); | 292 vpx_free(lf_sync->cond_); |
264 } | 293 } |
265 #endif // CONFIG_MULTITHREAD | 294 #endif // CONFIG_MULTITHREAD |
266 vpx_free(lf_sync->lfdata); | 295 vpx_free(lf_sync->lfdata); |
267 vpx_free(lf_sync->cur_sb_col); | 296 vpx_free(lf_sync->cur_sb_col); |
268 // clear the structure as the source of this call may be a resize in which | 297 // 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. | 298 // case this call will be followed by an _alloc() which may fail. |
270 vp9_zero(*lf_sync); | 299 vp9_zero(*lf_sync); |
271 } | 300 } |
272 } | 301 } |
OLD | NEW |