speedhq: fix decoding artifacts
[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 #include "avcodec.h"
28 #include "internal.h"
29 #include "pthread_internal.h"
30 #include "thread.h"
31
32 #include "libavutil/avassert.h"
33 #include "libavutil/common.h"
34 #include "libavutil/cpu.h"
35 #include "libavutil/mem.h"
36 #include "libavutil/thread.h"
37
38 typedef int (action_func)(AVCodecContext *c, void *arg);
39 typedef int (action_func2)(AVCodecContext *c, void *arg, int jobnr, int threadnr);
40
41 typedef struct SliceThreadContext {
42     pthread_t *workers;
43     action_func *func;
44     action_func2 *func2;
45     void *args;
46     int *rets;
47     int job_count;
48     int job_size;
49
50     pthread_cond_t last_job_cond;
51     pthread_cond_t current_job_cond;
52     pthread_mutex_t current_job_lock;
53     unsigned current_execute;
54     int current_job;
55     int done;
56
57     int *entries;
58     int entries_count;
59     int thread_count;
60     pthread_cond_t *progress_cond;
61     pthread_mutex_t *progress_mutex;
62 } SliceThreadContext;
63
64 static void* attribute_align_arg worker(void *v)
65 {
66     AVCodecContext *avctx = v;
67     SliceThreadContext *c = avctx->internal->thread_ctx;
68     unsigned last_execute = 0;
69     int our_job = c->job_count;
70     int thread_count = avctx->thread_count;
71     int self_id;
72
73     pthread_mutex_lock(&c->current_job_lock);
74     self_id = c->current_job++;
75     for (;;){
76         int ret;
77         while (our_job >= c->job_count) {
78             if (c->current_job == thread_count + c->job_count)
79                 pthread_cond_signal(&c->last_job_cond);
80
81             while (last_execute == c->current_execute && !c->done)
82                 pthread_cond_wait(&c->current_job_cond, &c->current_job_lock);
83             last_execute = c->current_execute;
84             our_job = self_id;
85
86             if (c->done) {
87                 pthread_mutex_unlock(&c->current_job_lock);
88                 return NULL;
89             }
90         }
91         pthread_mutex_unlock(&c->current_job_lock);
92
93         ret = c->func ? c->func(avctx, (char*)c->args + our_job*c->job_size):
94                                 c->func2(avctx, c->args, our_job, self_id);
95         if (c->rets)
96             c->rets[our_job%c->job_count] = ret;
97
98         pthread_mutex_lock(&c->current_job_lock);
99         our_job = c->current_job++;
100     }
101 }
102
103 void ff_slice_thread_free(AVCodecContext *avctx)
104 {
105     SliceThreadContext *c = avctx->internal->thread_ctx;
106     int i;
107
108     pthread_mutex_lock(&c->current_job_lock);
109     c->done = 1;
110     pthread_cond_broadcast(&c->current_job_cond);
111     for (i = 0; i < c->thread_count; i++)
112         pthread_cond_broadcast(&c->progress_cond[i]);
113     pthread_mutex_unlock(&c->current_job_lock);
114
115     for (i=0; i<avctx->thread_count; i++)
116          pthread_join(c->workers[i], NULL);
117
118     for (i = 0; i < c->thread_count; i++) {
119         pthread_mutex_destroy(&c->progress_mutex[i]);
120         pthread_cond_destroy(&c->progress_cond[i]);
121     }
122
123     pthread_mutex_destroy(&c->current_job_lock);
124     pthread_cond_destroy(&c->current_job_cond);
125     pthread_cond_destroy(&c->last_job_cond);
126
127     av_freep(&c->entries);
128     av_freep(&c->progress_mutex);
129     av_freep(&c->progress_cond);
130
131     av_freep(&c->workers);
132     av_freep(&avctx->internal->thread_ctx);
133 }
134
135 static av_always_inline void thread_park_workers(SliceThreadContext *c, int thread_count)
136 {
137     while (c->current_job != thread_count + c->job_count)
138         pthread_cond_wait(&c->last_job_cond, &c->current_job_lock);
139     pthread_mutex_unlock(&c->current_job_lock);
140 }
141
142 static int thread_execute(AVCodecContext *avctx, action_func* func, void *arg, int *ret, int job_count, int job_size)
143 {
144     SliceThreadContext *c = avctx->internal->thread_ctx;
145
146     if (!(avctx->active_thread_type&FF_THREAD_SLICE) || avctx->thread_count <= 1)
147         return avcodec_default_execute(avctx, func, arg, ret, job_count, job_size);
148
149     if (job_count <= 0)
150         return 0;
151
152     pthread_mutex_lock(&c->current_job_lock);
153
154     c->current_job = avctx->thread_count;
155     c->job_count = job_count;
156     c->job_size = job_size;
157     c->args = arg;
158     c->func = func;
159     if (ret) {
160         c->rets = ret;
161     } else {
162         c->rets = NULL;
163     }
164     c->current_execute++;
165     pthread_cond_broadcast(&c->current_job_cond);
166
167     thread_park_workers(c, avctx->thread_count);
168
169     return 0;
170 }
171
172 static int thread_execute2(AVCodecContext *avctx, action_func2* func2, void *arg, int *ret, int job_count)
173 {
174     SliceThreadContext *c = avctx->internal->thread_ctx;
175     c->func2 = func2;
176     return thread_execute(avctx, NULL, arg, ret, job_count, 0);
177 }
178
179 int ff_slice_thread_init(AVCodecContext *avctx)
180 {
181     int i;
182     SliceThreadContext *c;
183     int thread_count = avctx->thread_count;
184
185 #if HAVE_W32THREADS
186     w32thread_init();
187 #endif
188
189     // We cannot do this in the encoder init as the threads are created before
190     if (av_codec_is_encoder(avctx->codec) &&
191         avctx->codec_id == AV_CODEC_ID_MPEG1VIDEO &&
192         avctx->height > 2800)
193         thread_count = avctx->thread_count = 1;
194
195     if (!thread_count) {
196         int nb_cpus = av_cpu_count();
197         if  (avctx->height)
198             nb_cpus = FFMIN(nb_cpus, (avctx->height+15)/16);
199         // use number of cores + 1 as thread count if there is more than one
200         if (nb_cpus > 1)
201             thread_count = avctx->thread_count = FFMIN(nb_cpus + 1, MAX_AUTO_THREADS);
202         else
203             thread_count = avctx->thread_count = 1;
204     }
205
206     if (thread_count <= 1) {
207         avctx->active_thread_type = 0;
208         return 0;
209     }
210
211     c = av_mallocz(sizeof(SliceThreadContext));
212     if (!c)
213         return -1;
214
215     c->workers = av_mallocz_array(thread_count, sizeof(pthread_t));
216     if (!c->workers) {
217         av_free(c);
218         return -1;
219     }
220
221     avctx->internal->thread_ctx = c;
222     c->current_job = 0;
223     c->job_count = 0;
224     c->job_size = 0;
225     c->done = 0;
226     pthread_cond_init(&c->current_job_cond, NULL);
227     pthread_cond_init(&c->last_job_cond, NULL);
228     pthread_mutex_init(&c->current_job_lock, NULL);
229     pthread_mutex_lock(&c->current_job_lock);
230     for (i=0; i<thread_count; i++) {
231         if(pthread_create(&c->workers[i], NULL, worker, avctx)) {
232            avctx->thread_count = i;
233            pthread_mutex_unlock(&c->current_job_lock);
234            ff_thread_free(avctx);
235            return -1;
236         }
237     }
238
239     thread_park_workers(c, thread_count);
240
241     avctx->execute = thread_execute;
242     avctx->execute2 = thread_execute2;
243     return 0;
244 }
245
246 void ff_thread_report_progress2(AVCodecContext *avctx, int field, int thread, int n)
247 {
248     SliceThreadContext *p = avctx->internal->thread_ctx;
249     int *entries = p->entries;
250
251     pthread_mutex_lock(&p->progress_mutex[thread]);
252     entries[field] +=n;
253     pthread_cond_signal(&p->progress_cond[thread]);
254     pthread_mutex_unlock(&p->progress_mutex[thread]);
255 }
256
257 void ff_thread_await_progress2(AVCodecContext *avctx, int field, int thread, int shift)
258 {
259     SliceThreadContext *p  = avctx->internal->thread_ctx;
260     int *entries      = p->entries;
261
262     if (!entries || !field) return;
263
264     thread = thread ? thread - 1 : p->thread_count - 1;
265
266     pthread_mutex_lock(&p->progress_mutex[thread]);
267     while ((entries[field - 1] - entries[field]) < shift){
268         pthread_cond_wait(&p->progress_cond[thread], &p->progress_mutex[thread]);
269     }
270     pthread_mutex_unlock(&p->progress_mutex[thread]);
271 }
272
273 int ff_alloc_entries(AVCodecContext *avctx, int count)
274 {
275     int i;
276
277     if (avctx->active_thread_type & FF_THREAD_SLICE)  {
278         SliceThreadContext *p = avctx->internal->thread_ctx;
279
280         if (p->entries) {
281             av_assert0(p->thread_count == avctx->thread_count);
282             av_freep(&p->entries);
283         }
284
285         p->thread_count  = avctx->thread_count;
286         p->entries       = av_mallocz_array(count, sizeof(int));
287
288         if (!p->progress_mutex) {
289             p->progress_mutex = av_malloc_array(p->thread_count, sizeof(pthread_mutex_t));
290             p->progress_cond  = av_malloc_array(p->thread_count, sizeof(pthread_cond_t));
291         }
292
293         if (!p->entries || !p->progress_mutex || !p->progress_cond) {
294             av_freep(&p->entries);
295             av_freep(&p->progress_mutex);
296             av_freep(&p->progress_cond);
297             return AVERROR(ENOMEM);
298         }
299         p->entries_count  = count;
300
301         for (i = 0; i < p->thread_count; i++) {
302             pthread_mutex_init(&p->progress_mutex[i], NULL);
303             pthread_cond_init(&p->progress_cond[i], NULL);
304         }
305     }
306
307     return 0;
308 }
309
310 void ff_reset_entries(AVCodecContext *avctx)
311 {
312     SliceThreadContext *p = avctx->internal->thread_ctx;
313     memset(p->entries, 0, p->entries_count * sizeof(int));
314 }