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
mutex_lock(pthread_mutex_t * const mutex)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
sync_read(VP9LfSync * const lf_sync,int r,int c)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 mutex_lock(&lf_sync->mutex_[r - 1]);
44
45 while (c > lf_sync->cur_sb_col[r - 1] - nsync) {
46 pthread_cond_wait(&lf_sync->cond_[r - 1],
47 &lf_sync->mutex_[r - 1]);
48 }
49 pthread_mutex_unlock(&lf_sync->mutex_[r - 1]);
50 }
51 #else
52 (void)lf_sync;
53 (void)r;
54 (void)c;
55 #endif // CONFIG_MULTITHREAD
56 }
57
sync_write(VP9LfSync * const lf_sync,int r,int c,const int sb_cols)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.
loop_filter_rows_mt(const YV12_BUFFER_CONFIG * const frame_buffer,VP9_COMMON * const cm,MACROBLOCKD * const xd,int start,int stop,int y_only,VP9LfSync * const lf_sync,int num_lf_workers)91 static void loop_filter_rows_mt(const YV12_BUFFER_CONFIG *const frame_buffer,
92 VP9_COMMON *const cm, MACROBLOCKD *const xd,
93 int start, int stop, int y_only,
94 VP9LfSync *const lf_sync, int num_lf_workers) {
95 const int num_planes = y_only ? 1 : MAX_MB_PLANE;
96 int r, c; // SB row and col
97 LOOP_FILTER_MASK lfm;
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 += num_lf_workers) {
101 const int mi_row = r << MI_BLOCK_SIZE_LOG2;
102 MODE_INFO **mi_8x8 = cm->mi_grid_visible + 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 int plane;
107
108 sync_read(lf_sync, r, c);
109
110 vp9_setup_dst_planes(xd, frame_buffer, mi_row, mi_col);
111 vp9_setup_mask(cm, mi_row, mi_col, mi_8x8 + mi_col, cm->mi_stride, &lfm);
112
113 for (plane = 0; plane < num_planes; ++plane) {
114 vp9_filter_block_plane(cm, &xd->plane[plane], mi_row, &lfm);
115 }
116
117 sync_write(lf_sync, r, c, sb_cols);
118 }
119 }
120 }
121
122 // Row-based multi-threaded loopfilter hook
loop_filter_row_worker(void * arg1,void * arg2)123 static int loop_filter_row_worker(void *arg1, void *arg2) {
124 (void)arg2;
125 TileWorkerData *const tile_data = (TileWorkerData*)arg1;
126 LFWorkerData *const lf_data = &tile_data->lfdata;
127
128 loop_filter_rows_mt(lf_data->frame_buffer, lf_data->cm, &lf_data->xd,
129 lf_data->start, lf_data->stop, lf_data->y_only,
130 lf_data->lf_sync, lf_data->num_lf_workers);
131 return 1;
132 }
133
134 // VP9 decoder: Implement multi-threaded loopfilter that uses the tile
135 // threads.
vp9_loop_filter_frame_mt(VP9D_COMP * pbi,VP9_COMMON * cm,MACROBLOCKD * xd,int frame_filter_level,int y_only,int partial_frame)136 void vp9_loop_filter_frame_mt(VP9D_COMP *pbi,
137 VP9_COMMON *cm,
138 MACROBLOCKD *xd,
139 int frame_filter_level,
140 int y_only, int partial_frame) {
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 int i;
144 (void)xd;
145 (void)partial_frame;
146
147 // Allocate memory used in thread synchronization.
148 // This always needs to be done even if frame_filter_level is 0.
149 if (!cm->current_video_frame || cm->last_height != cm->height) {
150 VP9LfSync *const lf_sync = &pbi->lf_row_sync;
151
152 if (cm->last_height != cm->height) {
153 const int aligned_last_height =
154 ALIGN_POWER_OF_TWO(cm->last_height, MI_SIZE_LOG2);
155 const int last_sb_rows =
156 mi_cols_aligned_to_sb(aligned_last_height >> MI_SIZE_LOG2) >>
157 MI_BLOCK_SIZE_LOG2;
158
159 vp9_loop_filter_dealloc(lf_sync, last_sb_rows);
160 }
161
162 vp9_loop_filter_alloc(cm, lf_sync, sb_rows, cm->width);
163 }
164
165 if (!frame_filter_level) return;
166
167 vp9_loop_filter_frame_init(cm, frame_filter_level);
168
169 // Initialize cur_sb_col to -1 for all SB rows.
170 vpx_memset(pbi->lf_row_sync.cur_sb_col, -1,
171 sizeof(*pbi->lf_row_sync.cur_sb_col) * sb_rows);
172
173 // Set up loopfilter thread data.
174 for (i = 0; i < pbi->num_tile_workers; ++i) {
175 VP9Worker *const worker = &pbi->tile_workers[i];
176 TileWorkerData *const tile_data = (TileWorkerData*)worker->data1;
177 LFWorkerData *const lf_data = &tile_data->lfdata;
178
179 worker->hook = (VP9WorkerHook)loop_filter_row_worker;
180
181 // Loopfilter data
182 lf_data->frame_buffer = get_frame_new_buffer(cm);
183 lf_data->cm = cm;
184 lf_data->xd = pbi->mb;
185 lf_data->start = i;
186 lf_data->stop = sb_rows;
187 lf_data->y_only = y_only; // always do all planes in decoder
188
189 lf_data->lf_sync = &pbi->lf_row_sync;
190 lf_data->num_lf_workers = pbi->num_tile_workers;
191
192 // Start loopfiltering
193 if (i == pbi->num_tile_workers - 1) {
194 vp9_worker_execute(worker);
195 } else {
196 vp9_worker_launch(worker);
197 }
198 }
199
200 // Wait till all rows are finished
201 for (i = 0; i < pbi->num_tile_workers; ++i) {
202 vp9_worker_sync(&pbi->tile_workers[i]);
203 }
204 }
205
206 // Set up nsync by width.
get_sync_range(int width)207 static int get_sync_range(int width) {
208 // nsync numbers are picked by testing. For example, for 4k
209 // video, using 4 gives best performance.
210 if (width < 640)
211 return 1;
212 else if (width <= 1280)
213 return 2;
214 else if (width <= 4096)
215 return 4;
216 else
217 return 8;
218 }
219
220 // Allocate memory for lf row synchronization
vp9_loop_filter_alloc(VP9_COMMON * cm,VP9LfSync * lf_sync,int rows,int width)221 void vp9_loop_filter_alloc(VP9_COMMON *cm, VP9LfSync *lf_sync, int rows,
222 int width) {
223 #if CONFIG_MULTITHREAD
224 int i;
225
226 CHECK_MEM_ERROR(cm, lf_sync->mutex_,
227 vpx_malloc(sizeof(*lf_sync->mutex_) * rows));
228 for (i = 0; i < rows; ++i) {
229 pthread_mutex_init(&lf_sync->mutex_[i], NULL);
230 }
231
232 CHECK_MEM_ERROR(cm, lf_sync->cond_,
233 vpx_malloc(sizeof(*lf_sync->cond_) * rows));
234 for (i = 0; i < rows; ++i) {
235 pthread_cond_init(&lf_sync->cond_[i], NULL);
236 }
237 #endif // CONFIG_MULTITHREAD
238
239 CHECK_MEM_ERROR(cm, lf_sync->cur_sb_col,
240 vpx_malloc(sizeof(*lf_sync->cur_sb_col) * rows));
241
242 // Set up nsync.
243 lf_sync->sync_range = get_sync_range(width);
244 }
245
246 // Deallocate lf synchronization related mutex and data
vp9_loop_filter_dealloc(VP9LfSync * lf_sync,int rows)247 void vp9_loop_filter_dealloc(VP9LfSync *lf_sync, int rows) {
248 #if CONFIG_MULTITHREAD
249 if (lf_sync != NULL) {
250 int i;
251
252 if (lf_sync->mutex_ != NULL) {
253 for (i = 0; i < rows; ++i) {
254 pthread_mutex_destroy(&lf_sync->mutex_[i]);
255 }
256 vpx_free(lf_sync->mutex_);
257 }
258 if (lf_sync->cond_ != NULL) {
259 for (i = 0; i < rows; ++i) {
260 pthread_cond_destroy(&lf_sync->cond_[i]);
261 }
262 vpx_free(lf_sync->cond_);
263 }
264
265 vpx_free(lf_sync->cur_sb_col);
266 // clear the structure as the source of this call may be a resize in which
267 // case this call will be followed by an _alloc() which may fail.
268 vpx_memset(lf_sync, 0, sizeof(*lf_sync));
269 }
270 #else
271 (void)rows;
272 if (lf_sync != NULL) {
273 vpx_free(lf_sync->cur_sb_col);
274 vpx_memset(lf_sync, 0, sizeof(*lf_sync));
275 }
276 #endif // CONFIG_MULTITHREAD
277 }
278