1
0
mirror of https://github.com/FFmpeg/FFmpeg.git synced 2024-12-02 03:06:28 +02:00
FFmpeg/fftools/sync_queue.c

677 lines
21 KiB
C
Raw Normal View History

fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
/*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
#include <stdint.h>
#include <string.h>
#include "libavutil/avassert.h"
#include "libavutil/channel_layout.h"
#include "libavutil/cpu.h"
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
#include "libavutil/error.h"
#include "libavutil/fifo.h"
#include "libavutil/mathematics.h"
#include "libavutil/mem.h"
#include "libavutil/samplefmt.h"
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
#include "objpool.h"
#include "sync_queue.h"
/*
* How this works:
* --------------
* time: 0 1 2 3 4 5 6 7 8 9 10 11 12 13
* -------------------------------------------------------------------
* | | | | | | | | | | | | | |
* |
* stream 0| d=1 d=2 d=1 d=3
* |
*
* stream 1d=1 d=5
*
* |
* stream 2| d=1d=1d=1d=1 <- stream 2 is the head stream of the queue
* |
* ^ ^
* [stream 2 tail] [stream 2 head]
*
* We have N streams (N=3 in the diagram), each stream is a FIFO. The *tail* of
* each FIFO is the frame with smallest end time, the *head* is the frame with
* the largest end time. Frames submitted to the queue with sq_send() are placed
* after the head, frames returned to the caller with sq_receive() are taken
* from the tail.
*
* The head stream of the whole queue (SyncQueue.head_stream) is the limiting
* stream with the *smallest* head timestamp, i.e. the stream whose source lags
* furthest behind all other streams. It determines which frames can be output
* from the queue.
*
* In the diagram, the head stream is 2, because it head time is t=5, while
* streams 0 and 1 end at t=8 and t=9 respectively. All frames that _end_ at
* or before t=5 can be output, i.e. the first 3 frames from stream 0, first
* frame from stream 1, and all 4 frames from stream 2.
*/
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
typedef struct SyncQueueStream {
AVFifo *fifo;
AVRational tb;
/* number of audio samples in fifo */
uint64_t samples_queued;
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
/* stream head: largest timestamp seen */
int64_t head_ts;
int limiting;
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
/* no more frames will be sent for this stream */
int finished;
uint64_t frames_sent;
uint64_t samples_sent;
uint64_t frames_max;
int frame_samples;
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
} SyncQueueStream;
struct SyncQueue {
enum SyncQueueType type;
/* no more frames will be sent for any stream */
int finished;
/* sync head: the stream with the _smallest_ head timestamp
* this stream determines which frames can be output */
int head_stream;
/* the finished stream with the smallest finish timestamp or -1 */
int head_finished_stream;
// maximum buffering duration in microseconds
int64_t buf_size_us;
SyncQueueStream *streams;
unsigned int nb_streams;
// pool of preallocated frames to avoid constant allocations
ObjPool *pool;
int have_limiting;
uintptr_t align_mask;
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
};
static void frame_move(const SyncQueue *sq, SyncQueueFrame dst,
SyncQueueFrame src)
{
if (sq->type == SYNC_QUEUE_PACKETS)
av_packet_move_ref(dst.p, src.p);
else
av_frame_move_ref(dst.f, src.f);
}
/**
* Compute the end timestamp of a frame. If nb_samples is provided, consider
* the frame to have this number of audio samples, otherwise use frame duration.
*/
static int64_t frame_end(const SyncQueue *sq, SyncQueueFrame frame, int nb_samples)
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
{
if (nb_samples) {
int64_t d = av_rescale_q(nb_samples, (AVRational){ 1, frame.f->sample_rate},
frame.f->time_base);
return frame.f->pts + d;
}
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
return (sq->type == SYNC_QUEUE_PACKETS) ?
frame.p->pts + frame.p->duration :
frame.f->pts + frame.f->duration;
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
}
static int frame_samples(const SyncQueue *sq, SyncQueueFrame frame)
{
return (sq->type == SYNC_QUEUE_PACKETS) ? 0 : frame.f->nb_samples;
}
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
static int frame_null(const SyncQueue *sq, SyncQueueFrame frame)
{
return (sq->type == SYNC_QUEUE_PACKETS) ? (frame.p == NULL) : (frame.f == NULL);
}
static void tb_update(const SyncQueue *sq, SyncQueueStream *st,
const SyncQueueFrame frame)
{
AVRational tb = (sq->type == SYNC_QUEUE_PACKETS) ?
frame.p->time_base : frame.f->time_base;
av_assert0(tb.num > 0 && tb.den > 0);
if (tb.num == st->tb.num && tb.den == st->tb.den)
return;
// timebase should not change after the first frame
av_assert0(!av_fifo_can_read(st->fifo));
if (st->head_ts != AV_NOPTS_VALUE)
st->head_ts = av_rescale_q(st->head_ts, st->tb, tb);
st->tb = tb;
}
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
static void finish_stream(SyncQueue *sq, unsigned int stream_idx)
{
SyncQueueStream *st = &sq->streams[stream_idx];
st->finished = 1;
if (st->limiting && st->head_ts != AV_NOPTS_VALUE) {
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
/* check if this stream is the new finished head */
if (sq->head_finished_stream < 0 ||
av_compare_ts(st->head_ts, st->tb,
sq->streams[sq->head_finished_stream].head_ts,
sq->streams[sq->head_finished_stream].tb) < 0) {
sq->head_finished_stream = stream_idx;
}
/* mark as finished all streams that should no longer receive new frames,
* due to them being ahead of some finished stream */
st = &sq->streams[sq->head_finished_stream];
for (unsigned int i = 0; i < sq->nb_streams; i++) {
SyncQueueStream *st1 = &sq->streams[i];
if (st != st1 && st1->head_ts != AV_NOPTS_VALUE &&
av_compare_ts(st->head_ts, st->tb, st1->head_ts, st1->tb) <= 0)
st1->finished = 1;
}
}
/* mark the whole queue as finished if all streams are finished */
for (unsigned int i = 0; i < sq->nb_streams; i++) {
if (!sq->streams[i].finished)
return;
}
sq->finished = 1;
}
static void queue_head_update(SyncQueue *sq)
{
if (sq->head_stream < 0) {
/* wait for one timestamp in each stream before determining
* the queue head */
for (unsigned int i = 0; i < sq->nb_streams; i++) {
SyncQueueStream *st = &sq->streams[i];
if (st->limiting && st->head_ts == AV_NOPTS_VALUE)
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
return;
}
// placeholder value, correct one will be found below
sq->head_stream = 0;
}
for (unsigned int i = 0; i < sq->nb_streams; i++) {
SyncQueueStream *st_head = &sq->streams[sq->head_stream];
SyncQueueStream *st_other = &sq->streams[i];
if (st_other->limiting && st_other->head_ts != AV_NOPTS_VALUE &&
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
av_compare_ts(st_other->head_ts, st_other->tb,
st_head->head_ts, st_head->tb) < 0)
sq->head_stream = i;
}
}
/* update this stream's head timestamp */
static void stream_update_ts(SyncQueue *sq, unsigned int stream_idx, int64_t ts)
{
SyncQueueStream *st = &sq->streams[stream_idx];
if (ts == AV_NOPTS_VALUE ||
(st->head_ts != AV_NOPTS_VALUE && st->head_ts >= ts))
return;
st->head_ts = ts;
/* if this stream is now ahead of some finished stream, then
* this stream is also finished */
if (sq->head_finished_stream >= 0 &&
av_compare_ts(sq->streams[sq->head_finished_stream].head_ts,
sq->streams[sq->head_finished_stream].tb,
ts, st->tb) <= 0)
finish_stream(sq, stream_idx);
/* update the overall head timestamp if it could have changed */
if (st->limiting &&
(sq->head_stream < 0 || sq->head_stream == stream_idx))
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
queue_head_update(sq);
}
/* If the queue for the given stream (or all streams when stream_idx=-1)
* is overflowing, trigger a fake heartbeat on lagging streams.
*
* @return 1 if heartbeat triggered, 0 otherwise
*/
static int overflow_heartbeat(SyncQueue *sq, int stream_idx)
{
SyncQueueStream *st;
SyncQueueFrame frame;
int64_t tail_ts = AV_NOPTS_VALUE;
/* if no stream specified, pick the one that is most ahead */
if (stream_idx < 0) {
int64_t ts = AV_NOPTS_VALUE;
for (int i = 0; i < sq->nb_streams; i++) {
st = &sq->streams[i];
if (st->head_ts != AV_NOPTS_VALUE &&
(ts == AV_NOPTS_VALUE ||
av_compare_ts(ts, sq->streams[stream_idx].tb,
st->head_ts, st->tb) < 0)) {
ts = st->head_ts;
stream_idx = i;
}
}
/* no stream has a timestamp yet -> nothing to do */
if (stream_idx < 0)
return 0;
}
st = &sq->streams[stream_idx];
/* get the chosen stream's tail timestamp */
for (size_t i = 0; tail_ts == AV_NOPTS_VALUE &&
av_fifo_peek(st->fifo, &frame, 1, i) >= 0; i++)
tail_ts = frame_end(sq, frame, 0);
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
/* overflow triggers when the tail is over specified duration behind the head */
if (tail_ts == AV_NOPTS_VALUE || tail_ts >= st->head_ts ||
av_rescale_q(st->head_ts - tail_ts, st->tb, AV_TIME_BASE_Q) < sq->buf_size_us)
return 0;
/* signal a fake timestamp for all streams that prevent tail_ts from being output */
tail_ts++;
for (unsigned int i = 0; i < sq->nb_streams; i++) {
SyncQueueStream *st1 = &sq->streams[i];
int64_t ts;
if (st == st1 || st1->finished ||
(st1->head_ts != AV_NOPTS_VALUE &&
av_compare_ts(tail_ts, st->tb, st1->head_ts, st1->tb) <= 0))
continue;
ts = av_rescale_q(tail_ts, st->tb, st1->tb);
if (st1->head_ts != AV_NOPTS_VALUE)
ts = FFMAX(st1->head_ts + 1, ts);
stream_update_ts(sq, i, ts);
}
return 1;
}
int sq_send(SyncQueue *sq, unsigned int stream_idx, SyncQueueFrame frame)
{
SyncQueueStream *st;
SyncQueueFrame dst;
int64_t ts;
int ret, nb_samples;
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
av_assert0(stream_idx < sq->nb_streams);
st = &sq->streams[stream_idx];
if (frame_null(sq, frame)) {
finish_stream(sq, stream_idx);
return 0;
}
if (st->finished)
return AVERROR_EOF;
tb_update(sq, st, frame);
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
ret = objpool_get(sq->pool, (void**)&dst);
if (ret < 0)
return ret;
frame_move(sq, dst, frame);
nb_samples = frame_samples(sq, dst);
// make sure frame duration is consistent with sample count
if (nb_samples) {
av_assert0(dst.f->sample_rate > 0);
dst.f->duration = av_rescale_q(nb_samples, (AVRational){ 1, dst.f->sample_rate },
dst.f->time_base);
}
ts = frame_end(sq, dst, 0);
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
ret = av_fifo_write(st->fifo, &dst, 1);
if (ret < 0) {
frame_move(sq, frame, dst);
objpool_release(sq->pool, (void**)&dst);
return ret;
}
stream_update_ts(sq, stream_idx, ts);
st->samples_queued += nb_samples;
st->samples_sent += nb_samples;
if (st->frame_samples)
st->frames_sent = st->samples_sent / st->frame_samples;
else
st->frames_sent++;
if (st->frames_sent >= st->frames_max)
finish_stream(sq, stream_idx);
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
return 0;
}
static void offset_audio(AVFrame *f, int nb_samples)
{
const int planar = av_sample_fmt_is_planar(f->format);
const int planes = planar ? f->ch_layout.nb_channels : 1;
const int bps = av_get_bytes_per_sample(f->format);
const int offset = nb_samples * bps * (planar ? 1 : f->ch_layout.nb_channels);
av_assert0(bps > 0);
av_assert0(nb_samples < f->nb_samples);
for (int i = 0; i < planes; i++) {
f->extended_data[i] += offset;
if (i < FF_ARRAY_ELEMS(f->data))
f->data[i] = f->extended_data[i];
}
f->linesize[0] -= offset;
f->nb_samples -= nb_samples;
f->duration = av_rescale_q(f->nb_samples, (AVRational){ 1, f->sample_rate },
f->time_base);
f->pts += av_rescale_q(nb_samples, (AVRational){ 1, f->sample_rate },
f->time_base);
}
static int frame_is_aligned(const SyncQueue *sq, const AVFrame *frame)
{
// only checks linesize[0], so only works for audio
av_assert0(frame->nb_samples > 0);
av_assert0(sq->align_mask);
// only check data[0], because we always offset all data pointers
// by the same offset, so if one is aligned, all are
if (!((uintptr_t)frame->data[0] & sq->align_mask) &&
!(frame->linesize[0] & sq->align_mask) &&
frame->linesize[0] > sq->align_mask)
return 1;
return 0;
}
static int receive_samples(SyncQueue *sq, SyncQueueStream *st,
AVFrame *dst, int nb_samples)
{
SyncQueueFrame src;
int ret;
av_assert0(st->samples_queued >= nb_samples);
ret = av_fifo_peek(st->fifo, &src, 1, 0);
av_assert0(ret >= 0);
// peeked frame has enough samples and its data is aligned
// -> we can just make a reference and limit its sample count
if (src.f->nb_samples > nb_samples && frame_is_aligned(sq, src.f)) {
ret = av_frame_ref(dst, src.f);
if (ret < 0)
return ret;
dst->nb_samples = nb_samples;
offset_audio(src.f, nb_samples);
st->samples_queued -= nb_samples;
goto finish;
}
// otherwise allocate a new frame and copy the data
ret = av_channel_layout_copy(&dst->ch_layout, &src.f->ch_layout);
if (ret < 0)
return ret;
dst->format = src.f->format;
dst->nb_samples = nb_samples;
ret = av_frame_get_buffer(dst, 0);
if (ret < 0)
goto fail;
ret = av_frame_copy_props(dst, src.f);
if (ret < 0)
goto fail;
dst->nb_samples = 0;
while (dst->nb_samples < nb_samples) {
int to_copy;
ret = av_fifo_peek(st->fifo, &src, 1, 0);
av_assert0(ret >= 0);
to_copy = FFMIN(nb_samples - dst->nb_samples, src.f->nb_samples);
av_samples_copy(dst->extended_data, src.f->extended_data, dst->nb_samples,
0, to_copy, dst->ch_layout.nb_channels, dst->format);
if (to_copy < src.f->nb_samples)
offset_audio(src.f, to_copy);
else {
av_frame_unref(src.f);
objpool_release(sq->pool, (void**)&src);
av_fifo_drain2(st->fifo, 1);
}
st->samples_queued -= to_copy;
dst->nb_samples += to_copy;
}
finish:
dst->duration = av_rescale_q(nb_samples, (AVRational){ 1, dst->sample_rate },
dst->time_base);
return 0;
fail:
av_frame_unref(dst);
return ret;
}
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
static int receive_for_stream(SyncQueue *sq, unsigned int stream_idx,
SyncQueueFrame frame)
{
SyncQueueStream *st_head = sq->head_stream >= 0 ?
&sq->streams[sq->head_stream] : NULL;
SyncQueueStream *st;
av_assert0(stream_idx < sq->nb_streams);
st = &sq->streams[stream_idx];
if (av_fifo_can_read(st->fifo) &&
(st->frame_samples <= st->samples_queued || st->finished)) {
int nb_samples = st->frame_samples;
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
SyncQueueFrame peek;
int64_t ts;
int cmp = 1;
if (st->finished)
nb_samples = FFMIN(nb_samples, st->samples_queued);
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
av_fifo_peek(st->fifo, &peek, 1, 0);
ts = frame_end(sq, peek, nb_samples);
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
/* check if this stream's tail timestamp does not overtake
* the overall queue head */
if (ts != AV_NOPTS_VALUE && st_head)
cmp = av_compare_ts(ts, st->tb, st_head->head_ts, st_head->tb);
/* We can release frames that do not end after the queue head.
* Frames with no timestamps are just passed through with no conditions.
* Frames are also passed through when there are no limiting streams.
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
*/
if (cmp <= 0 || ts == AV_NOPTS_VALUE || !sq->have_limiting) {
if (nb_samples &&
(nb_samples != peek.f->nb_samples || !frame_is_aligned(sq, peek.f))) {
int ret = receive_samples(sq, st, frame.f, nb_samples);
if (ret < 0)
return ret;
} else {
frame_move(sq, frame, peek);
objpool_release(sq->pool, (void**)&peek);
av_fifo_drain2(st->fifo, 1);
av_assert0(st->samples_queued >= frame_samples(sq, frame));
st->samples_queued -= frame_samples(sq, frame);
}
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
return 0;
}
}
return (sq->finished || (st->finished && !av_fifo_can_read(st->fifo))) ?
AVERROR_EOF : AVERROR(EAGAIN);
}
static int receive_internal(SyncQueue *sq, int stream_idx, SyncQueueFrame frame)
{
int nb_eof = 0;
int ret;
/* read a frame for a specific stream */
if (stream_idx >= 0) {
ret = receive_for_stream(sq, stream_idx, frame);
return (ret < 0) ? ret : stream_idx;
}
/* read a frame for any stream with available output */
for (unsigned int i = 0; i < sq->nb_streams; i++) {
ret = receive_for_stream(sq, i, frame);
if (ret == AVERROR_EOF || ret == AVERROR(EAGAIN)) {
nb_eof += (ret == AVERROR_EOF);
continue;
}
return (ret < 0) ? ret : i;
}
return (nb_eof == sq->nb_streams) ? AVERROR_EOF : AVERROR(EAGAIN);
}
int sq_receive(SyncQueue *sq, int stream_idx, SyncQueueFrame frame)
{
int ret = receive_internal(sq, stream_idx, frame);
/* try again if the queue overflowed and triggered a fake heartbeat
* for lagging streams */
if (ret == AVERROR(EAGAIN) && overflow_heartbeat(sq, stream_idx))
ret = receive_internal(sq, stream_idx, frame);
return ret;
}
int sq_add_stream(SyncQueue *sq, int limiting)
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
{
SyncQueueStream *tmp, *st;
tmp = av_realloc_array(sq->streams, sq->nb_streams + 1, sizeof(*sq->streams));
if (!tmp)
return AVERROR(ENOMEM);
sq->streams = tmp;
st = &sq->streams[sq->nb_streams];
memset(st, 0, sizeof(*st));
st->fifo = av_fifo_alloc2(1, sizeof(SyncQueueFrame), AV_FIFO_FLAG_AUTO_GROW);
if (!st->fifo)
return AVERROR(ENOMEM);
/* we set a valid default, so that a pathological stream that never
* receives even a real timebase (and no frames) won't stall all other
* streams forever; cf. overflow_heartbeat() */
st->tb = (AVRational){ 1, 1 };
st->head_ts = AV_NOPTS_VALUE;
st->frames_max = UINT64_MAX;
st->limiting = limiting;
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
sq->have_limiting |= limiting;
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
return sq->nb_streams++;
}
void sq_limit_frames(SyncQueue *sq, unsigned int stream_idx, uint64_t frames)
{
SyncQueueStream *st;
av_assert0(stream_idx < sq->nb_streams);
st = &sq->streams[stream_idx];
st->frames_max = frames;
if (st->frames_sent >= st->frames_max)
finish_stream(sq, stream_idx);
}
void sq_frame_samples(SyncQueue *sq, unsigned int stream_idx,
int frame_samples)
{
SyncQueueStream *st;
av_assert0(sq->type == SYNC_QUEUE_FRAMES);
av_assert0(stream_idx < sq->nb_streams);
st = &sq->streams[stream_idx];
st->frame_samples = frame_samples;
sq->align_mask = av_cpu_max_align() - 1;
}
fftools/ffmpeg: rework -shortest implementation The -shortest option (which finishes the output file at the time the shortest stream ends) is currently implemented by faking the -t option when an output stream ends. This approach is fragile, since it depends on the frames/packets being processed in a specific order. E.g. there are currently some situations in which the output file length will depend unpredictably on unrelated factors like encoder delay. More importantly, the present work aiming at splitting various ffmpeg components into different threads will make this approach completely unworkable, since the frames/packets will arrive in effectively random order. This commit introduces a "sync queue", which is essentially a collection of FIFOs, one per stream. Frames/packets are submitted to these FIFOs and are then released for further processing (encoding or muxing) when it is ensured that the frame in question will not cause its stream to get ahead of the other streams (the logic is similar to libavformat's interleaving queue). These sync queues are then used for encoding and/or muxing when the -shortest option is specified. A new option – -shortest_buf_duration – controls the maximum number of queued packets, to avoid runaway memory usage. This commit changes the results of the following tests: - copy-shortest[12]: the last audio frame is now gone. This is correct, since it actually outlasts the last video frame. - shortest-sub: the video packets following the last subtitle packet are now gone. This is also correct.
2022-06-10 14:21:42 +02:00
SyncQueue *sq_alloc(enum SyncQueueType type, int64_t buf_size_us)
{
SyncQueue *sq = av_mallocz(sizeof(*sq));
if (!sq)
return NULL;
sq->type = type;
sq->buf_size_us = buf_size_us;
sq->head_stream = -1;
sq->head_finished_stream = -1;
sq->pool = (type == SYNC_QUEUE_PACKETS) ? objpool_alloc_packets() :
objpool_alloc_frames();
if (!sq->pool) {
av_freep(&sq);
return NULL;
}
return sq;
}
void sq_free(SyncQueue **psq)
{
SyncQueue *sq = *psq;
if (!sq)
return;
for (unsigned int i = 0; i < sq->nb_streams; i++) {
SyncQueueFrame frame;
while (av_fifo_read(sq->streams[i].fifo, &frame, 1) >= 0)
objpool_release(sq->pool, (void**)&frame);
av_fifo_freep2(&sq->streams[i].fifo);
}
av_freep(&sq->streams);
objpool_free(&sq->pool);
av_freep(psq);
}