avcodec/huffyuvenc: frame multi-threading support
[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 "compat/w32pthreads.h"
34 #elif HAVE_OS2THREADS
35 #include "compat/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         av_frame_unref(frame);
96         pthread_mutex_unlock(&c->buffer_mutex);
97         av_frame_free(&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 = av_cpu_count();
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->codec_id == AV_CODEC_ID_HUFFYUV ||
137         avctx->codec_id == AV_CODEC_ID_FFVHUFF) {
138         // huffyuv doesnt support these with multiple frame threads currently
139         if (avctx->context_model > 0 || (avctx->flags & CODEC_FLAG_PASS1))
140             return 0;
141     }
142
143     if(avctx->thread_count > MAX_THREADS)
144         return AVERROR(EINVAL);
145
146     av_assert0(!avctx->internal->frame_thread_encoder);
147     c = avctx->internal->frame_thread_encoder = av_mallocz(sizeof(ThreadContext));
148     if(!c)
149         return AVERROR(ENOMEM);
150
151     c->parent_avctx = avctx;
152
153     c->task_fifo = av_fifo_alloc(sizeof(Task) * BUFFER_SIZE);
154     if(!c->task_fifo)
155         goto fail;
156
157     pthread_mutex_init(&c->task_fifo_mutex, NULL);
158     pthread_mutex_init(&c->finished_task_mutex, NULL);
159     pthread_mutex_init(&c->buffer_mutex, NULL);
160     pthread_cond_init(&c->task_fifo_cond, NULL);
161     pthread_cond_init(&c->finished_task_cond, NULL);
162
163     for(i=0; i<avctx->thread_count ; i++){
164         AVDictionary *tmp = NULL;
165         void *tmpv;
166         AVCodecContext *thread_avctx = avcodec_alloc_context3(avctx->codec);
167         if(!thread_avctx)
168             goto fail;
169         tmpv = thread_avctx->priv_data;
170         *thread_avctx = *avctx;
171         thread_avctx->priv_data = tmpv;
172         thread_avctx->internal = NULL;
173         memcpy(thread_avctx->priv_data, avctx->priv_data, avctx->codec->priv_data_size);
174         thread_avctx->thread_count = 1;
175         thread_avctx->active_thread_type &= ~FF_THREAD_FRAME;
176
177         av_dict_copy(&tmp, options, 0);
178         av_dict_set(&tmp, "threads", "1", 0);
179         if(avcodec_open2(thread_avctx, avctx->codec, &tmp) < 0) {
180             av_dict_free(&tmp);
181             goto fail;
182         }
183         av_dict_free(&tmp);
184         av_assert0(!thread_avctx->internal->frame_thread_encoder);
185         thread_avctx->internal->frame_thread_encoder = c;
186         if(pthread_create(&c->worker[i], NULL, worker, thread_avctx)) {
187             goto fail;
188         }
189     }
190
191     avctx->active_thread_type = FF_THREAD_FRAME;
192
193     return 0;
194 fail:
195     avctx->thread_count = i;
196     av_log(avctx, AV_LOG_ERROR, "ff_frame_thread_encoder_init failed\n");
197     ff_frame_thread_encoder_free(avctx);
198     return -1;
199 }
200
201 void ff_frame_thread_encoder_free(AVCodecContext *avctx){
202     int i;
203     ThreadContext *c= avctx->internal->frame_thread_encoder;
204
205     pthread_mutex_lock(&c->task_fifo_mutex);
206     c->exit = 1;
207     pthread_cond_broadcast(&c->task_fifo_cond);
208     pthread_mutex_unlock(&c->task_fifo_mutex);
209
210     for (i=0; i<avctx->thread_count; i++) {
211          pthread_join(c->worker[i], NULL);
212     }
213
214     pthread_mutex_destroy(&c->task_fifo_mutex);
215     pthread_mutex_destroy(&c->finished_task_mutex);
216     pthread_mutex_destroy(&c->buffer_mutex);
217     pthread_cond_destroy(&c->task_fifo_cond);
218     pthread_cond_destroy(&c->finished_task_cond);
219     av_fifo_free(c->task_fifo); c->task_fifo = NULL;
220     av_freep(&avctx->internal->frame_thread_encoder);
221 }
222
223 int ff_thread_video_encode_frame(AVCodecContext *avctx, AVPacket *pkt, const AVFrame *frame, int *got_packet_ptr){
224     ThreadContext *c = avctx->internal->frame_thread_encoder;
225     Task task;
226     int ret;
227
228     av_assert1(!*got_packet_ptr);
229
230     if(frame){
231         if(!(avctx->flags & CODEC_FLAG_INPUT_PRESERVED)){
232             AVFrame *new = av_frame_alloc();
233             if(!new)
234                 return AVERROR(ENOMEM);
235             pthread_mutex_lock(&c->buffer_mutex);
236             ret = ff_get_buffer(c->parent_avctx, new, 0);
237             pthread_mutex_unlock(&c->buffer_mutex);
238             if(ret<0)
239                 return ret;
240             new->pts = frame->pts;
241             new->quality = frame->quality;
242             new->pict_type = frame->pict_type;
243             av_image_copy(new->data, new->linesize, (const uint8_t **)frame->data, frame->linesize,
244                           avctx->pix_fmt, avctx->width, avctx->height);
245             frame = new;
246         }
247
248         task.index = c->task_index;
249         task.indata = (void*)frame;
250         pthread_mutex_lock(&c->task_fifo_mutex);
251         av_fifo_generic_write(c->task_fifo, &task, sizeof(task), NULL);
252         pthread_cond_signal(&c->task_fifo_cond);
253         pthread_mutex_unlock(&c->task_fifo_mutex);
254
255         c->task_index = (c->task_index+1) % BUFFER_SIZE;
256
257         if(!c->finished_tasks[c->finished_task_index].outdata && (c->task_index - c->finished_task_index) % BUFFER_SIZE <= avctx->thread_count)
258             return 0;
259     }
260
261     if(c->task_index == c->finished_task_index)
262         return 0;
263
264     pthread_mutex_lock(&c->finished_task_mutex);
265     while (!c->finished_tasks[c->finished_task_index].outdata) {
266         pthread_cond_wait(&c->finished_task_cond, &c->finished_task_mutex);
267     }
268     task = c->finished_tasks[c->finished_task_index];
269     *pkt = *(AVPacket*)(task.outdata);
270     if(pkt->data)
271         *got_packet_ptr = 1;
272     av_freep(&c->finished_tasks[c->finished_task_index].outdata);
273     c->finished_task_index = (c->finished_task_index+1) % BUFFER_SIZE;
274     pthread_mutex_unlock(&c->finished_task_mutex);
275
276     return task.return_code;
277 }