2 * This file is part of Libav.
4 * Libav 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.
9 * Libav 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.
14 * You should have received a copy of the GNU Lesser General Public
15 * License along with Libav; if not, write to the Free Software
16 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 * Frame multithreading support functions
22 * @see doc/multithreading.txt
32 #include "compat/w32pthreads.h"
37 #include "pthread_internal.h"
40 #include "libavutil/avassert.h"
41 #include "libavutil/buffer.h"
42 #include "libavutil/common.h"
43 #include "libavutil/cpu.h"
44 #include "libavutil/frame.h"
45 #include "libavutil/log.h"
46 #include "libavutil/mem.h"
49 * Context used by codec threads and stored in their AVCodecInternal thread_ctx.
51 typedef struct PerThreadContext {
52 struct FrameThreadContext *parent;
56 pthread_cond_t input_cond; ///< Used to wait for a new packet from the main thread.
57 pthread_cond_t progress_cond; ///< Used by child threads to wait for progress to change.
58 pthread_cond_t output_cond; ///< Used by the main thread to wait for frames to finish.
60 pthread_mutex_t mutex; ///< Mutex used to protect the contents of the PerThreadContext.
61 pthread_mutex_t progress_mutex; ///< Mutex used to protect frame progress values and progress_cond.
63 AVCodecContext *avctx; ///< Context used to decode packets passed to this thread.
65 AVPacket avpkt; ///< Input packet (for decoding) or output (for encoding).
66 uint8_t *buf; ///< backup storage for packet data when the input packet is not refcounted
67 int allocated_buf_size; ///< Size allocated for buf
69 AVFrame *frame; ///< Output frame (for decoding) or input (for encoding).
70 int got_frame; ///< The output of got_picture_ptr from the last avcodec_decode_video() call.
71 int result; ///< The result of the last codec decode/encode() call.
74 STATE_INPUT_READY, ///< Set when the thread is awaiting a packet.
75 STATE_SETTING_UP, ///< Set before the codec has called ff_thread_finish_setup().
76 STATE_GET_BUFFER, /**<
77 * Set when the codec calls get_buffer().
78 * State is returned to STATE_SETTING_UP afterwards.
80 STATE_SETUP_FINISHED ///< Set after the codec has called ff_thread_finish_setup().
84 * Array of frames passed to ff_thread_release_buffer().
85 * Frames are released after all threads referencing them are finished.
87 AVFrame *released_buffers;
88 int num_released_buffers;
89 int released_buffers_allocated;
91 AVFrame *requested_frame; ///< AVFrame the codec passed to get_buffer()
92 int requested_flags; ///< flags passed to get_buffer() for requested_frame
96 * Context stored in the client AVCodecInternal thread_ctx.
98 typedef struct FrameThreadContext {
99 PerThreadContext *threads; ///< The contexts for each thread.
100 PerThreadContext *prev_thread; ///< The last thread submit_packet() was called on.
102 pthread_mutex_t buffer_mutex; ///< Mutex used to protect get/release_buffer().
104 int next_decoding; ///< The next context to submit a packet to.
105 int next_finished; ///< The next context to return output from.
108 * Set for the first N packets, where N is the number of threads.
109 * While it is set, ff_thread_en/decode_frame won't return any results.
112 int die; ///< Set when threads should exit.
113 } FrameThreadContext;
116 * Codec worker thread.
118 * Automatically calls ff_thread_finish_setup() if the codec does
119 * not provide an update_thread_context method, or if the codec returns
122 static attribute_align_arg void *frame_worker_thread(void *arg)
124 PerThreadContext *p = arg;
125 FrameThreadContext *fctx = p->parent;
126 AVCodecContext *avctx = p->avctx;
127 const AVCodec *codec = avctx->codec;
130 if (p->state == STATE_INPUT_READY && !fctx->die) {
131 pthread_mutex_lock(&p->mutex);
132 while (p->state == STATE_INPUT_READY && !fctx->die)
133 pthread_cond_wait(&p->input_cond, &p->mutex);
134 pthread_mutex_unlock(&p->mutex);
137 if (fctx->die) break;
139 if (!codec->update_thread_context && avctx->thread_safe_callbacks)
140 ff_thread_finish_setup(avctx);
142 pthread_mutex_lock(&p->mutex);
143 av_frame_unref(p->frame);
145 p->result = codec->decode(avctx, p->frame, &p->got_frame, &p->avpkt);
147 if ((p->result < 0 || !p->got_frame) && p->frame->buf[0]) {
148 if (avctx->internal->allocate_progress)
149 av_log(avctx, AV_LOG_ERROR, "A frame threaded decoder did not "
150 "free the frame on failure. This is a bug, please report it.\n");
151 av_frame_unref(p->frame);
154 if (p->state == STATE_SETTING_UP) ff_thread_finish_setup(avctx);
156 p->state = STATE_INPUT_READY;
158 pthread_mutex_lock(&p->progress_mutex);
159 pthread_cond_signal(&p->output_cond);
160 pthread_mutex_unlock(&p->progress_mutex);
162 pthread_mutex_unlock(&p->mutex);
169 * Update the next thread's AVCodecContext with values from the reference thread's context.
171 * @param dst The destination context.
172 * @param src The source context.
173 * @param for_user 0 if the destination is a codec thread, 1 if the destination is the user's thread
175 static int update_context_from_thread(AVCodecContext *dst, AVCodecContext *src, int for_user)
180 dst->time_base = src->time_base;
181 dst->width = src->width;
182 dst->height = src->height;
183 dst->pix_fmt = src->pix_fmt;
185 dst->coded_width = src->coded_width;
186 dst->coded_height = src->coded_height;
188 dst->has_b_frames = src->has_b_frames;
189 dst->idct_algo = src->idct_algo;
191 dst->bits_per_coded_sample = src->bits_per_coded_sample;
192 dst->sample_aspect_ratio = src->sample_aspect_ratio;
193 dst->dtg_active_format = src->dtg_active_format;
195 dst->profile = src->profile;
196 dst->level = src->level;
198 dst->bits_per_raw_sample = src->bits_per_raw_sample;
199 dst->ticks_per_frame = src->ticks_per_frame;
200 dst->color_primaries = src->color_primaries;
202 dst->color_trc = src->color_trc;
203 dst->colorspace = src->colorspace;
204 dst->color_range = src->color_range;
205 dst->chroma_sample_location = src->chroma_sample_location;
207 dst->hwaccel = src->hwaccel;
208 dst->hwaccel_context = src->hwaccel_context;
209 dst->internal->hwaccel_priv_data = src->internal->hwaccel_priv_data;
213 dst->coded_frame = src->coded_frame;
215 if (dst->codec->update_thread_context)
216 err = dst->codec->update_thread_context(dst, src);
223 * Update the next thread's AVCodecContext with values set by the user.
225 * @param dst The destination context.
226 * @param src The source context.
227 * @return 0 on success, negative error code on failure
229 static int update_context_from_user(AVCodecContext *dst, AVCodecContext *src)
231 #define copy_fields(s, e) memcpy(&dst->s, &src->s, (char*)&dst->e - (char*)&dst->s);
232 dst->flags = src->flags;
234 dst->draw_horiz_band= src->draw_horiz_band;
235 dst->get_buffer2 = src->get_buffer2;
236 #if FF_API_GET_BUFFER
237 FF_DISABLE_DEPRECATION_WARNINGS
238 dst->get_buffer = src->get_buffer;
239 dst->release_buffer = src->release_buffer;
240 FF_ENABLE_DEPRECATION_WARNINGS
243 dst->opaque = src->opaque;
244 dst->debug = src->debug;
246 dst->slice_flags = src->slice_flags;
247 dst->flags2 = src->flags2;
249 copy_fields(skip_loop_filter, subtitle_header);
251 dst->frame_number = src->frame_number;
252 dst->reordered_opaque = src->reordered_opaque;
254 if (src->slice_count && src->slice_offset) {
255 if (dst->slice_count < src->slice_count) {
256 int *tmp = av_realloc(dst->slice_offset, src->slice_count *
257 sizeof(*dst->slice_offset));
259 av_free(dst->slice_offset);
260 return AVERROR(ENOMEM);
262 dst->slice_offset = tmp;
264 memcpy(dst->slice_offset, src->slice_offset,
265 src->slice_count * sizeof(*dst->slice_offset));
267 dst->slice_count = src->slice_count;
272 /// Releases the buffers that this decoding thread was the last user of.
273 static void release_delayed_buffers(PerThreadContext *p)
275 FrameThreadContext *fctx = p->parent;
277 while (p->num_released_buffers > 0) {
280 pthread_mutex_lock(&fctx->buffer_mutex);
282 // fix extended data in case the caller screwed it up
283 av_assert0(p->avctx->codec_type == AVMEDIA_TYPE_VIDEO);
284 f = &p->released_buffers[--p->num_released_buffers];
285 f->extended_data = f->data;
288 pthread_mutex_unlock(&fctx->buffer_mutex);
292 static int submit_packet(PerThreadContext *p, AVPacket *avpkt)
294 FrameThreadContext *fctx = p->parent;
295 PerThreadContext *prev_thread = fctx->prev_thread;
296 const AVCodec *codec = p->avctx->codec;
298 if (!avpkt->size && !(codec->capabilities & CODEC_CAP_DELAY)) return 0;
300 pthread_mutex_lock(&p->mutex);
302 release_delayed_buffers(p);
306 if (prev_thread->state == STATE_SETTING_UP) {
307 pthread_mutex_lock(&prev_thread->progress_mutex);
308 while (prev_thread->state == STATE_SETTING_UP)
309 pthread_cond_wait(&prev_thread->progress_cond, &prev_thread->progress_mutex);
310 pthread_mutex_unlock(&prev_thread->progress_mutex);
313 err = update_context_from_thread(p->avctx, prev_thread->avctx, 0);
315 pthread_mutex_unlock(&p->mutex);
320 av_buffer_unref(&p->avpkt.buf);
323 p->avpkt.buf = av_buffer_ref(avpkt->buf);
325 av_fast_malloc(&p->buf, &p->allocated_buf_size, avpkt->size + FF_INPUT_BUFFER_PADDING_SIZE);
326 p->avpkt.data = p->buf;
327 memcpy(p->buf, avpkt->data, avpkt->size);
328 memset(p->buf + avpkt->size, 0, FF_INPUT_BUFFER_PADDING_SIZE);
331 p->state = STATE_SETTING_UP;
332 pthread_cond_signal(&p->input_cond);
333 pthread_mutex_unlock(&p->mutex);
336 * If the client doesn't have a thread-safe get_buffer(),
337 * then decoding threads call back to the main thread,
338 * and it calls back to the client here.
341 FF_DISABLE_DEPRECATION_WARNINGS
342 if (!p->avctx->thread_safe_callbacks && (
343 #if FF_API_GET_BUFFER
344 p->avctx->get_buffer ||
346 p->avctx->get_buffer2 != avcodec_default_get_buffer2)) {
347 FF_ENABLE_DEPRECATION_WARNINGS
348 while (p->state != STATE_SETUP_FINISHED && p->state != STATE_INPUT_READY) {
349 pthread_mutex_lock(&p->progress_mutex);
350 while (p->state == STATE_SETTING_UP)
351 pthread_cond_wait(&p->progress_cond, &p->progress_mutex);
353 if (p->state == STATE_GET_BUFFER) {
354 p->result = ff_get_buffer(p->avctx, p->requested_frame, p->requested_flags);
355 p->state = STATE_SETTING_UP;
356 pthread_cond_signal(&p->progress_cond);
358 pthread_mutex_unlock(&p->progress_mutex);
362 fctx->prev_thread = p;
363 fctx->next_decoding++;
368 int ff_thread_decode_frame(AVCodecContext *avctx,
369 AVFrame *picture, int *got_picture_ptr,
372 FrameThreadContext *fctx = avctx->internal->thread_ctx;
373 int finished = fctx->next_finished;
378 * Submit a packet to the next decoding thread.
381 p = &fctx->threads[fctx->next_decoding];
382 err = update_context_from_user(p->avctx, avctx);
384 err = submit_packet(p, avpkt);
388 * If we're still receiving the initial packets, don't return a frame.
391 if (fctx->delaying) {
392 if (fctx->next_decoding >= (avctx->thread_count-1)) fctx->delaying = 0;
400 * Return the next available frame from the oldest thread.
401 * If we're at the end of the stream, then we have to skip threads that
402 * didn't output a frame, because we don't want to accidentally signal
403 * EOF (avpkt->size == 0 && *got_picture_ptr == 0).
407 p = &fctx->threads[finished++];
409 if (p->state != STATE_INPUT_READY) {
410 pthread_mutex_lock(&p->progress_mutex);
411 while (p->state != STATE_INPUT_READY)
412 pthread_cond_wait(&p->output_cond, &p->progress_mutex);
413 pthread_mutex_unlock(&p->progress_mutex);
416 av_frame_move_ref(picture, p->frame);
417 *got_picture_ptr = p->got_frame;
418 picture->pkt_dts = p->avpkt.dts;
421 * A later call with avkpt->size == 0 may loop over all threads,
422 * including this one, searching for a frame to return before being
423 * stopped by the "finished != fctx->next_finished" condition.
424 * Make sure we don't mistakenly return the same frame again.
428 if (finished >= avctx->thread_count) finished = 0;
429 } while (!avpkt->size && !*got_picture_ptr && finished != fctx->next_finished);
431 update_context_from_thread(avctx, p->avctx, 1);
433 if (fctx->next_decoding >= avctx->thread_count) fctx->next_decoding = 0;
435 fctx->next_finished = finished;
437 /* return the size of the consumed packet if no error occurred */
438 return (p->result >= 0) ? avpkt->size : p->result;
441 void ff_thread_report_progress(ThreadFrame *f, int n, int field)
444 int *progress = f->progress ? (int*)f->progress->data : NULL;
446 if (!progress || progress[field] >= n) return;
448 p = f->owner->internal->thread_ctx;
450 if (f->owner->debug&FF_DEBUG_THREADS)
451 av_log(f->owner, AV_LOG_DEBUG, "%p finished %d field %d\n", progress, n, field);
453 pthread_mutex_lock(&p->progress_mutex);
455 pthread_cond_broadcast(&p->progress_cond);
456 pthread_mutex_unlock(&p->progress_mutex);
459 void ff_thread_await_progress(ThreadFrame *f, int n, int field)
462 int *progress = f->progress ? (int*)f->progress->data : NULL;
464 if (!progress || progress[field] >= n) return;
466 p = f->owner->internal->thread_ctx;
468 if (f->owner->debug&FF_DEBUG_THREADS)
469 av_log(f->owner, AV_LOG_DEBUG, "thread awaiting %d field %d from %p\n", n, field, progress);
471 pthread_mutex_lock(&p->progress_mutex);
472 while (progress[field] < n)
473 pthread_cond_wait(&p->progress_cond, &p->progress_mutex);
474 pthread_mutex_unlock(&p->progress_mutex);
477 void ff_thread_finish_setup(AVCodecContext *avctx) {
478 PerThreadContext *p = avctx->internal->thread_ctx;
480 if (!(avctx->active_thread_type&FF_THREAD_FRAME)) return;
482 pthread_mutex_lock(&p->progress_mutex);
483 p->state = STATE_SETUP_FINISHED;
484 pthread_cond_broadcast(&p->progress_cond);
485 pthread_mutex_unlock(&p->progress_mutex);
488 /// Waits for all threads to finish.
489 static void park_frame_worker_threads(FrameThreadContext *fctx, int thread_count)
493 for (i = 0; i < thread_count; i++) {
494 PerThreadContext *p = &fctx->threads[i];
496 if (p->state != STATE_INPUT_READY) {
497 pthread_mutex_lock(&p->progress_mutex);
498 while (p->state != STATE_INPUT_READY)
499 pthread_cond_wait(&p->output_cond, &p->progress_mutex);
500 pthread_mutex_unlock(&p->progress_mutex);
505 void ff_frame_thread_free(AVCodecContext *avctx, int thread_count)
507 FrameThreadContext *fctx = avctx->internal->thread_ctx;
508 const AVCodec *codec = avctx->codec;
511 park_frame_worker_threads(fctx, thread_count);
513 if (fctx->prev_thread && fctx->prev_thread != fctx->threads)
514 update_context_from_thread(fctx->threads->avctx, fctx->prev_thread->avctx, 0);
518 for (i = 0; i < thread_count; i++) {
519 PerThreadContext *p = &fctx->threads[i];
521 pthread_mutex_lock(&p->mutex);
522 pthread_cond_signal(&p->input_cond);
523 pthread_mutex_unlock(&p->mutex);
526 pthread_join(p->thread, NULL);
529 codec->close(p->avctx);
533 release_delayed_buffers(p);
534 av_frame_free(&p->frame);
537 for (i = 0; i < thread_count; i++) {
538 PerThreadContext *p = &fctx->threads[i];
540 pthread_mutex_destroy(&p->mutex);
541 pthread_mutex_destroy(&p->progress_mutex);
542 pthread_cond_destroy(&p->input_cond);
543 pthread_cond_destroy(&p->progress_cond);
544 pthread_cond_destroy(&p->output_cond);
545 av_buffer_unref(&p->avpkt.buf);
547 av_freep(&p->released_buffers);
550 av_freep(&p->avctx->priv_data);
551 av_freep(&p->avctx->slice_offset);
554 av_freep(&p->avctx->internal);
558 av_freep(&fctx->threads);
559 pthread_mutex_destroy(&fctx->buffer_mutex);
560 av_freep(&avctx->internal->thread_ctx);
563 int ff_frame_thread_init(AVCodecContext *avctx)
565 int thread_count = avctx->thread_count;
566 const AVCodec *codec = avctx->codec;
567 AVCodecContext *src = avctx;
568 FrameThreadContext *fctx;
576 int nb_cpus = av_cpu_count();
577 av_log(avctx, AV_LOG_DEBUG, "detected %d logical cores\n", nb_cpus);
578 // use number of cores + 1 as thread count if there is more than one
580 thread_count = avctx->thread_count = FFMIN(nb_cpus + 1, MAX_AUTO_THREADS);
582 thread_count = avctx->thread_count = 1;
585 if (thread_count <= 1) {
586 avctx->active_thread_type = 0;
590 avctx->internal->thread_ctx = fctx = av_mallocz(sizeof(FrameThreadContext));
592 fctx->threads = av_mallocz(sizeof(PerThreadContext) * thread_count);
593 pthread_mutex_init(&fctx->buffer_mutex, NULL);
596 for (i = 0; i < thread_count; i++) {
597 AVCodecContext *copy = av_malloc(sizeof(AVCodecContext));
598 PerThreadContext *p = &fctx->threads[i];
600 pthread_mutex_init(&p->mutex, NULL);
601 pthread_mutex_init(&p->progress_mutex, NULL);
602 pthread_cond_init(&p->input_cond, NULL);
603 pthread_cond_init(&p->progress_cond, NULL);
604 pthread_cond_init(&p->output_cond, NULL);
606 p->frame = av_frame_alloc();
608 err = AVERROR(ENOMEM);
616 err = AVERROR(ENOMEM);
622 copy->internal = av_malloc(sizeof(AVCodecInternal));
623 if (!copy->internal) {
624 err = AVERROR(ENOMEM);
627 *copy->internal = *src->internal;
628 copy->internal->thread_ctx = p;
629 copy->internal->pkt = &p->avpkt;
635 err = codec->init(copy);
637 update_context_from_thread(avctx, copy, 1);
639 copy->priv_data = av_malloc(codec->priv_data_size);
640 if (!copy->priv_data) {
641 err = AVERROR(ENOMEM);
644 memcpy(copy->priv_data, src->priv_data, codec->priv_data_size);
645 copy->internal->is_copy = 1;
647 if (codec->init_thread_copy)
648 err = codec->init_thread_copy(copy);
653 if (!pthread_create(&p->thread, NULL, frame_worker_thread, p))
660 ff_frame_thread_free(avctx, i+1);
665 void ff_thread_flush(AVCodecContext *avctx)
668 FrameThreadContext *fctx = avctx->internal->thread_ctx;
672 park_frame_worker_threads(fctx, avctx->thread_count);
673 if (fctx->prev_thread) {
674 if (fctx->prev_thread != &fctx->threads[0])
675 update_context_from_thread(fctx->threads[0].avctx, fctx->prev_thread->avctx, 0);
678 fctx->next_decoding = fctx->next_finished = 0;
680 fctx->prev_thread = NULL;
681 for (i = 0; i < avctx->thread_count; i++) {
682 PerThreadContext *p = &fctx->threads[i];
683 // Make sure decode flush calls with size=0 won't return old frames
685 av_frame_unref(p->frame);
687 release_delayed_buffers(p);
689 if (avctx->codec->flush)
690 avctx->codec->flush(p->avctx);
694 int ff_thread_get_buffer(AVCodecContext *avctx, ThreadFrame *f, int flags)
696 PerThreadContext *p = avctx->internal->thread_ctx;
701 if (!(avctx->active_thread_type & FF_THREAD_FRAME))
702 return ff_get_buffer(avctx, f->f, flags);
704 if (p->state != STATE_SETTING_UP &&
705 (avctx->codec->update_thread_context || !avctx->thread_safe_callbacks)) {
706 av_log(avctx, AV_LOG_ERROR, "get_buffer() cannot be called after ff_thread_finish_setup()\n");
710 if (avctx->internal->allocate_progress) {
712 f->progress = av_buffer_alloc(2 * sizeof(int));
714 return AVERROR(ENOMEM);
716 progress = (int*)f->progress->data;
718 progress[0] = progress[1] = -1;
721 pthread_mutex_lock(&p->parent->buffer_mutex);
722 FF_DISABLE_DEPRECATION_WARNINGS
723 if (avctx->thread_safe_callbacks || (
724 #if FF_API_GET_BUFFER
725 !avctx->get_buffer &&
727 avctx->get_buffer2 == avcodec_default_get_buffer2)) {
728 FF_ENABLE_DEPRECATION_WARNINGS
729 err = ff_get_buffer(avctx, f->f, flags);
731 p->requested_frame = f->f;
732 p->requested_flags = flags;
733 p->state = STATE_GET_BUFFER;
734 pthread_mutex_lock(&p->progress_mutex);
735 pthread_cond_signal(&p->progress_cond);
737 while (p->state != STATE_SETTING_UP)
738 pthread_cond_wait(&p->progress_cond, &p->progress_mutex);
742 pthread_mutex_unlock(&p->progress_mutex);
745 if (!avctx->thread_safe_callbacks && !avctx->codec->update_thread_context)
746 ff_thread_finish_setup(avctx);
749 av_buffer_unref(&f->progress);
751 pthread_mutex_unlock(&p->parent->buffer_mutex);
756 void ff_thread_release_buffer(AVCodecContext *avctx, ThreadFrame *f)
758 PerThreadContext *p = avctx->internal->thread_ctx;
759 FrameThreadContext *fctx;
761 FF_DISABLE_DEPRECATION_WARNINGS
762 int can_direct_free = !(avctx->active_thread_type & FF_THREAD_FRAME) ||
763 avctx->thread_safe_callbacks ||
765 #if FF_API_GET_BUFFER
766 !avctx->get_buffer &&
768 avctx->get_buffer2 == avcodec_default_get_buffer2);
769 FF_ENABLE_DEPRECATION_WARNINGS
771 if (!f->f || !f->f->buf[0])
774 if (avctx->debug & FF_DEBUG_BUFFERS)
775 av_log(avctx, AV_LOG_DEBUG, "thread_release_buffer called on pic %p\n", f);
777 av_buffer_unref(&f->progress);
780 if (can_direct_free) {
781 av_frame_unref(f->f);
786 pthread_mutex_lock(&fctx->buffer_mutex);
788 if (p->num_released_buffers + 1 >= INT_MAX / sizeof(*p->released_buffers))
790 tmp = av_fast_realloc(p->released_buffers, &p->released_buffers_allocated,
791 (p->num_released_buffers + 1) *
792 sizeof(*p->released_buffers));
795 p->released_buffers = tmp;
797 dst = &p->released_buffers[p->num_released_buffers];
798 av_frame_move_ref(dst, f->f);
800 p->num_released_buffers++;
803 pthread_mutex_unlock(&fctx->buffer_mutex);