lavc/pthread_slice: release entries
[ffmpeg.git] / libavcodec / pthread_slice.c
1 /*
2  * This file is part of FFmpeg.
3  *
4  * FFmpeg is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU Lesser General Public
6  * License as published by the Free Software Foundation; either
7  * version 2.1 of the License, or (at your option) any later version.
8  *
9  * FFmpeg is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
12  * Lesser General Public License for more details.
13  *
14  * You should have received a copy of the GNU Lesser General Public
15  * License along with FFmpeg; if not, write to the Free Software
16  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
17  */
18
19 /**
20  * @file
21  * Slice multithreading support functions
22  * @see doc/multithreading.txt
23  */
24
25 #include "config.h"
26
27 #if HAVE_PTHREADS
28 #include <pthread.h>
29 #elif HAVE_W32THREADS
30 #include "compat/w32pthreads.h"
31 #elif HAVE_OS2THREADS
32 #include "compat/os2threads.h"
33 #endif
34
35 #include "avcodec.h"
36 #include "internal.h"
37 #include "pthread_internal.h"
38 #include "thread.h"
39
40 #include "libavutil/common.h"
41 #include "libavutil/cpu.h"
42 #include "libavutil/mem.h"
43
44 typedef int (action_func)(AVCodecContext *c, void *arg);
45 typedef int (action_func2)(AVCodecContext *c, void *arg, int jobnr, int threadnr);
46
47 typedef struct SliceThreadContext {
48     pthread_t *workers;
49     action_func *func;
50     action_func2 *func2;
51     void *args;
52     int *rets;
53     int rets_count;
54     int job_count;
55     int job_size;
56
57     pthread_cond_t last_job_cond;
58     pthread_cond_t current_job_cond;
59     pthread_mutex_t current_job_lock;
60     unsigned current_execute;
61     int current_job;
62     int done;
63
64     int *entries;
65     int entries_count;
66     int thread_count;
67     pthread_cond_t *progress_cond;
68     pthread_mutex_t *progress_mutex;
69 } SliceThreadContext;
70
71 static void* attribute_align_arg worker(void *v)
72 {
73     AVCodecContext *avctx = v;
74     SliceThreadContext *c = avctx->internal->thread_ctx;
75     unsigned last_execute = 0;
76     int our_job = c->job_count;
77     int thread_count = avctx->thread_count;
78     int self_id;
79
80     pthread_mutex_lock(&c->current_job_lock);
81     self_id = c->current_job++;
82     for (;;){
83         while (our_job >= c->job_count) {
84             if (c->current_job == thread_count + c->job_count)
85                 pthread_cond_signal(&c->last_job_cond);
86
87             while (last_execute == c->current_execute && !c->done)
88                 pthread_cond_wait(&c->current_job_cond, &c->current_job_lock);
89             last_execute = c->current_execute;
90             our_job = self_id;
91
92             if (c->done) {
93                 pthread_mutex_unlock(&c->current_job_lock);
94                 return NULL;
95             }
96         }
97         pthread_mutex_unlock(&c->current_job_lock);
98
99         c->rets[our_job%c->rets_count] = c->func ? c->func(avctx, (char*)c->args + our_job*c->job_size):
100                                                    c->func2(avctx, c->args, our_job, self_id);
101
102         pthread_mutex_lock(&c->current_job_lock);
103         our_job = c->current_job++;
104     }
105 }
106
107 void ff_slice_thread_free(AVCodecContext *avctx)
108 {
109     SliceThreadContext *c = avctx->internal->thread_ctx;
110     int i;
111
112     pthread_mutex_lock(&c->current_job_lock);
113     c->done = 1;
114     pthread_cond_broadcast(&c->current_job_cond);
115     for (i = 0; i < c->thread_count; i++)
116         pthread_cond_broadcast(&c->progress_cond[i]);
117     pthread_mutex_unlock(&c->current_job_lock);
118
119     for (i=0; i<avctx->thread_count; i++)
120          pthread_join(c->workers[i], NULL);
121
122     for (i = 0; i < c->thread_count; i++) {
123         pthread_mutex_destroy(&c->progress_mutex[i]);
124         pthread_cond_destroy(&c->progress_cond[i]);
125     }
126
127     pthread_mutex_destroy(&c->current_job_lock);
128     pthread_cond_destroy(&c->current_job_cond);
129     pthread_cond_destroy(&c->last_job_cond);
130
131     av_freep(&c->entries);
132     av_freep(&c->progress_mutex);
133     av_freep(&c->progress_cond);
134
135     av_freep(&c->workers);
136     av_freep(&avctx->internal->thread_ctx);
137 }
138
139 static av_always_inline void thread_park_workers(SliceThreadContext *c, int thread_count)
140 {
141     while (c->current_job != thread_count + c->job_count)
142         pthread_cond_wait(&c->last_job_cond, &c->current_job_lock);
143     pthread_mutex_unlock(&c->current_job_lock);
144 }
145
146 static int thread_execute(AVCodecContext *avctx, action_func* func, void *arg, int *ret, int job_count, int job_size)
147 {
148     SliceThreadContext *c = avctx->internal->thread_ctx;
149     int dummy_ret;
150
151     if (!(avctx->active_thread_type&FF_THREAD_SLICE) || avctx->thread_count <= 1)
152         return avcodec_default_execute(avctx, func, arg, ret, job_count, job_size);
153
154     if (job_count <= 0)
155         return 0;
156
157     pthread_mutex_lock(&c->current_job_lock);
158
159     c->current_job = avctx->thread_count;
160     c->job_count = job_count;
161     c->job_size = job_size;
162     c->args = arg;
163     c->func = func;
164     if (ret) {
165         c->rets = ret;
166         c->rets_count = job_count;
167     } else {
168         c->rets = &dummy_ret;
169         c->rets_count = 1;
170     }
171     c->current_execute++;
172     pthread_cond_broadcast(&c->current_job_cond);
173
174     thread_park_workers(c, avctx->thread_count);
175
176     return 0;
177 }
178
179 static int thread_execute2(AVCodecContext *avctx, action_func2* func2, void *arg, int *ret, int job_count)
180 {
181     SliceThreadContext *c = avctx->internal->thread_ctx;
182     c->func2 = func2;
183     return thread_execute(avctx, NULL, arg, ret, job_count, 0);
184 }
185
186 int ff_slice_thread_init(AVCodecContext *avctx)
187 {
188     int i;
189     SliceThreadContext *c;
190     int thread_count = avctx->thread_count;
191
192 #if HAVE_W32THREADS
193     w32thread_init();
194 #endif
195
196     if (!thread_count) {
197         int nb_cpus = av_cpu_count();
198         if  (avctx->height)
199             nb_cpus = FFMIN(nb_cpus, (avctx->height+15)/16);
200         // use number of cores + 1 as thread count if there is more than one
201         if (nb_cpus > 1)
202             thread_count = avctx->thread_count = FFMIN(nb_cpus + 1, MAX_AUTO_THREADS);
203         else
204             thread_count = avctx->thread_count = 1;
205     }
206
207     if (thread_count <= 1) {
208         avctx->active_thread_type = 0;
209         return 0;
210     }
211
212     c = av_mallocz(sizeof(SliceThreadContext));
213     if (!c)
214         return -1;
215
216     c->workers = av_mallocz_array(thread_count, sizeof(pthread_t));
217     if (!c->workers) {
218         av_free(c);
219         return -1;
220     }
221
222     avctx->internal->thread_ctx = c;
223     c->current_job = 0;
224     c->job_count = 0;
225     c->job_size = 0;
226     c->done = 0;
227     pthread_cond_init(&c->current_job_cond, NULL);
228     pthread_cond_init(&c->last_job_cond, NULL);
229     pthread_mutex_init(&c->current_job_lock, NULL);
230     pthread_mutex_lock(&c->current_job_lock);
231     for (i=0; i<thread_count; i++) {
232         if(pthread_create(&c->workers[i], NULL, worker, avctx)) {
233            avctx->thread_count = i;
234            pthread_mutex_unlock(&c->current_job_lock);
235            ff_thread_free(avctx);
236            return -1;
237         }
238     }
239
240     thread_park_workers(c, thread_count);
241
242     avctx->execute = thread_execute;
243     avctx->execute2 = thread_execute2;
244     return 0;
245 }
246
247 void ff_thread_report_progress2(AVCodecContext *avctx, int field, int thread, int n)
248 {
249     SliceThreadContext *p = avctx->internal->thread_ctx;
250     int *entries = p->entries;
251
252     pthread_mutex_lock(&p->progress_mutex[thread]);
253     entries[field] +=n;
254     pthread_cond_signal(&p->progress_cond[thread]);
255     pthread_mutex_unlock(&p->progress_mutex[thread]);
256 }
257
258 void ff_thread_await_progress2(AVCodecContext *avctx, int field, int thread, int shift)
259 {
260     SliceThreadContext *p  = avctx->internal->thread_ctx;
261     int *entries      = p->entries;
262
263     if (!entries || !field) return;
264
265     thread = thread ? thread - 1 : p->thread_count - 1;
266
267     pthread_mutex_lock(&p->progress_mutex[thread]);
268     while ((entries[field - 1] - entries[field]) < shift){
269         pthread_cond_wait(&p->progress_cond[thread], &p->progress_mutex[thread]);
270     }
271     pthread_mutex_unlock(&p->progress_mutex[thread]);
272 }
273
274 int ff_alloc_entries(AVCodecContext *avctx, int count)
275 {
276     int i;
277
278     if (avctx->active_thread_type & FF_THREAD_SLICE)  {
279         SliceThreadContext *p = avctx->internal->thread_ctx;
280         p->thread_count  = avctx->thread_count;
281         p->entries       = av_mallocz_array(count, sizeof(int));
282
283         p->progress_mutex = av_malloc_array(p->thread_count, sizeof(pthread_mutex_t));
284         p->progress_cond  = av_malloc_array(p->thread_count, sizeof(pthread_cond_t));
285
286         if (!p->entries || !p->progress_mutex || !p->progress_cond) {
287             av_freep(&p->entries);
288             av_freep(&p->progress_mutex);
289             av_freep(&p->progress_cond);
290             return AVERROR(ENOMEM);
291         }
292         p->entries_count  = count;
293
294         for (i = 0; i < p->thread_count; i++) {
295             pthread_mutex_init(&p->progress_mutex[i], NULL);
296             pthread_cond_init(&p->progress_cond[i], NULL);
297         }
298     }
299
300     return 0;
301 }
302
303 void ff_reset_entries(AVCodecContext *avctx)
304 {
305     SliceThreadContext *p = avctx->internal->thread_ctx;
306     memset(p->entries, 0, p->entries_count * sizeof(int));
307 }