Merge commit '86bfcfcf2364bc837b7bb582c66a8a15a332414f'
[ffmpeg.git] / libavcodec / frame_thread_encoder.c
1 /*
2  * Copyright (c) 2012 Michael Niedermayer <michaelni@gmx.at>
3  *
4  * This file is part of FFmpeg.
5  *
6  * FFmpeg is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * FFmpeg is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with FFmpeg; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20
21 #include "frame_thread_encoder.h"
22
23 #include "libavutil/fifo.h"
24 #include "libavutil/avassert.h"
25 #include "libavutil/imgutils.h"
26 #include "avcodec.h"
27 #include "internal.h"
28 #include "thread.h"
29
30 #if HAVE_PTHREADS
31 #include <pthread.h>
32 #elif HAVE_W32THREADS
33 #include "w32pthreads.h"
34 #elif HAVE_OS2THREADS
35 #include "os2threads.h"
36 #endif
37
38 #define MAX_THREADS 64
39 #define BUFFER_SIZE (2*MAX_THREADS)
40
41 typedef struct{
42     void *indata;
43     void *outdata;
44     int64_t return_code;
45     unsigned index;
46 } Task;
47
48 typedef struct{
49     AVCodecContext *parent_avctx;
50     pthread_mutex_t buffer_mutex;
51
52     AVFifoBuffer *task_fifo;
53     pthread_mutex_t task_fifo_mutex;
54     pthread_cond_t task_fifo_cond;
55
56     Task finished_tasks[BUFFER_SIZE];
57     pthread_mutex_t finished_task_mutex;
58     pthread_cond_t finished_task_cond;
59
60     unsigned task_index;
61     unsigned finished_task_index;
62
63     pthread_t worker[MAX_THREADS];
64     int exit;
65 } ThreadContext;
66
67 static void * attribute_align_arg worker(void *v){
68     AVCodecContext *avctx = v;
69     ThreadContext *c = avctx->internal->frame_thread_encoder;
70     AVPacket *pkt = NULL;
71
72     while(!c->exit){
73         int got_packet, ret;
74         AVFrame *frame;
75         Task task;
76
77         if(!pkt) pkt= av_mallocz(sizeof(*pkt));
78         if(!pkt) continue;
79         av_init_packet(pkt);
80
81         pthread_mutex_lock(&c->task_fifo_mutex);
82         while (av_fifo_size(c->task_fifo) <= 0 || c->exit) {
83             if(c->exit){
84                 pthread_mutex_unlock(&c->task_fifo_mutex);
85                 goto end;
86             }
87             pthread_cond_wait(&c->task_fifo_cond, &c->task_fifo_mutex);
88         }
89         av_fifo_generic_read(c->task_fifo, &task, sizeof(task), NULL);
90         pthread_mutex_unlock(&c->task_fifo_mutex);
91         frame = task.indata;
92
93         ret = avcodec_encode_video2(avctx, pkt, frame, &got_packet);
94         pthread_mutex_lock(&c->buffer_mutex);
95         c->parent_avctx->release_buffer(c->parent_avctx, frame);
96         pthread_mutex_unlock(&c->buffer_mutex);
97         av_freep(&frame);
98         if(got_packet) {
99             av_dup_packet(pkt);
100         } else {
101             pkt->data = NULL;
102             pkt->size = 0;
103         }
104         pthread_mutex_lock(&c->finished_task_mutex);
105         c->finished_tasks[task.index].outdata = pkt; pkt = NULL;
106         c->finished_tasks[task.index].return_code = ret;
107         pthread_cond_signal(&c->finished_task_cond);
108         pthread_mutex_unlock(&c->finished_task_mutex);
109     }
110 end:
111     av_free(pkt);
112     pthread_mutex_lock(&c->buffer_mutex);
113     avcodec_close(avctx);
114     pthread_mutex_unlock(&c->buffer_mutex);
115     av_freep(&avctx);
116     return NULL;
117 }
118
119 int ff_frame_thread_encoder_init(AVCodecContext *avctx, AVDictionary *options){
120     int i=0;
121     ThreadContext *c;
122
123
124     if(   !(avctx->thread_type & FF_THREAD_FRAME)
125        || !(avctx->codec->capabilities & CODEC_CAP_INTRA_ONLY))
126         return 0;
127
128     if(!avctx->thread_count) {
129         avctx->thread_count = ff_get_logical_cpus(avctx);
130         avctx->thread_count = FFMIN(avctx->thread_count, MAX_THREADS);
131     }
132
133     if(avctx->thread_count <= 1)
134         return 0;
135
136     if(avctx->thread_count > MAX_THREADS)
137         return AVERROR(EINVAL);
138
139     av_assert0(!avctx->internal->frame_thread_encoder);
140     c = avctx->internal->frame_thread_encoder = av_mallocz(sizeof(ThreadContext));
141     if(!c)
142         return AVERROR(ENOMEM);
143
144     c->parent_avctx = avctx;
145
146     c->task_fifo = av_fifo_alloc(sizeof(Task) * BUFFER_SIZE);
147     if(!c->task_fifo)
148         goto fail;
149
150     pthread_mutex_init(&c->task_fifo_mutex, NULL);
151     pthread_mutex_init(&c->finished_task_mutex, NULL);
152     pthread_mutex_init(&c->buffer_mutex, NULL);
153     pthread_cond_init(&c->task_fifo_cond, NULL);
154     pthread_cond_init(&c->finished_task_cond, NULL);
155
156     for(i=0; i<avctx->thread_count ; i++){
157         AVDictionary *tmp = NULL;
158         void *tmpv;
159         AVCodecContext *thread_avctx = avcodec_alloc_context3(avctx->codec);
160         if(!thread_avctx)
161             goto fail;
162         tmpv = thread_avctx->priv_data;
163         *thread_avctx = *avctx;
164         thread_avctx->priv_data = tmpv;
165         thread_avctx->internal = NULL;
166         memcpy(thread_avctx->priv_data, avctx->priv_data, avctx->codec->priv_data_size);
167         thread_avctx->thread_count = 1;
168         thread_avctx->active_thread_type &= ~FF_THREAD_FRAME;
169
170         av_dict_copy(&tmp, options, 0);
171         av_dict_set(&tmp, "threads", "1", 0);
172         if(avcodec_open2(thread_avctx, avctx->codec, &tmp) < 0) {
173             av_dict_free(&tmp);
174             goto fail;
175         }
176         av_dict_free(&tmp);
177         av_assert0(!thread_avctx->internal->frame_thread_encoder);
178         thread_avctx->internal->frame_thread_encoder = c;
179         if(pthread_create(&c->worker[i], NULL, worker, thread_avctx)) {
180             goto fail;
181         }
182     }
183
184     avctx->active_thread_type = FF_THREAD_FRAME;
185
186     return 0;
187 fail:
188     avctx->thread_count = i;
189     av_log(avctx, AV_LOG_ERROR, "ff_frame_thread_encoder_init failed\n");
190     ff_frame_thread_encoder_free(avctx);
191     return -1;
192 }
193
194 void ff_frame_thread_encoder_free(AVCodecContext *avctx){
195     int i;
196     ThreadContext *c= avctx->internal->frame_thread_encoder;
197
198     pthread_mutex_lock(&c->task_fifo_mutex);
199     c->exit = 1;
200     pthread_cond_broadcast(&c->task_fifo_cond);
201     pthread_mutex_unlock(&c->task_fifo_mutex);
202
203     for (i=0; i<avctx->thread_count; i++) {
204          pthread_join(c->worker[i], NULL);
205     }
206
207     pthread_mutex_destroy(&c->task_fifo_mutex);
208     pthread_mutex_destroy(&c->finished_task_mutex);
209     pthread_mutex_destroy(&c->buffer_mutex);
210     pthread_cond_destroy(&c->task_fifo_cond);
211     pthread_cond_destroy(&c->finished_task_cond);
212     av_fifo_free(c->task_fifo); c->task_fifo = NULL;
213     av_freep(&avctx->internal->frame_thread_encoder);
214 }
215
216 int ff_thread_video_encode_frame(AVCodecContext *avctx, AVPacket *pkt, const AVFrame *frame, int *got_packet_ptr){
217     ThreadContext *c = avctx->internal->frame_thread_encoder;
218     Task task;
219     int ret;
220
221     av_assert1(!*got_packet_ptr);
222
223     if(frame){
224         if(!(avctx->flags & CODEC_FLAG_INPUT_PRESERVED)){
225             AVFrame *new = avcodec_alloc_frame();
226             if(!new)
227                 return AVERROR(ENOMEM);
228             pthread_mutex_lock(&c->buffer_mutex);
229             ret = ff_get_buffer(c->parent_avctx, new);
230             pthread_mutex_unlock(&c->buffer_mutex);
231             if(ret<0)
232                 return ret;
233             new->pts = frame->pts;
234             new->quality = frame->quality;
235             new->pict_type = frame->pict_type;
236             av_image_copy(new->data, new->linesize, (const uint8_t **)frame->data, frame->linesize,
237                           avctx->pix_fmt, avctx->width, avctx->height);
238             frame = new;
239         }
240
241         task.index = c->task_index;
242         task.indata = (void*)frame;
243         pthread_mutex_lock(&c->task_fifo_mutex);
244         av_fifo_generic_write(c->task_fifo, &task, sizeof(task), NULL);
245         pthread_cond_signal(&c->task_fifo_cond);
246         pthread_mutex_unlock(&c->task_fifo_mutex);
247
248         c->task_index = (c->task_index+1) % BUFFER_SIZE;
249
250         if(!c->finished_tasks[c->finished_task_index].outdata && (c->task_index - c->finished_task_index) % BUFFER_SIZE <= avctx->thread_count)
251             return 0;
252     }
253
254     if(c->task_index == c->finished_task_index)
255         return 0;
256
257     pthread_mutex_lock(&c->finished_task_mutex);
258     while (!c->finished_tasks[c->finished_task_index].outdata) {
259         pthread_cond_wait(&c->finished_task_cond, &c->finished_task_mutex);
260     }
261     task = c->finished_tasks[c->finished_task_index];
262     *pkt = *(AVPacket*)(task.outdata);
263     if(pkt->data)
264         *got_packet_ptr = 1;
265     av_freep(&c->finished_tasks[c->finished_task_index].outdata);
266     c->finished_task_index = (c->finished_task_index+1) % BUFFER_SIZE;
267     pthread_mutex_unlock(&c->finished_task_mutex);
268
269     return task.return_code;
270 }