Chiebot-Mirror
/
FFmpeg
mirror of https://github.com/FFmpeg/FFmpeg.git
8
0
Fork 0
Code Issues Projects Releases Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
110328 Commits
37 Branches
400 Tags
465 MiB
Tag: Branch: Tree: b088d59347
Branches Tags
${ item.name }
Create tag ${ searchTerm }
Create branch ${ searchTerm }
from 'b088d59347'
${ noResults }
FFmpeg/fftools/sync_queue.c

673 lines
21 KiB
Raw Normal View History Unescape

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.
2 years ago
/*
* 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"
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
#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.
2 years ago
#include "libavutil/error.h"
#include "libavutil/fifo.h"
#include "libavutil/mathematics.h"
#include "libavutil/mem.h"
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
#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.
2 years ago
#include "objpool.h"
#include "sync_queue.h"
fftools/sync_queue: document overall design
3 years ago
/*
* 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 1│d=1│ │ d=5 │
* └───┘ └───────────────────────┘
* | ┌───┐┌───┐┌───┐┌───┐
* stream 2| │d=1││d=1││d=1││d=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.
2 years ago
typedef struct SyncQueueStream {
AVFifo *fifo;
AVRational tb;
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
/* 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.
2 years ago
/* stream head: largest timestamp seen */
int64_t head_ts;
fftools/ffmpeg: use the sync queues to handle -frames Same issues apply to it as to -shortest. Changes the results of the following tests: - matroska-flac-extradata-update The test reencodes two input FLAC streams into three output FLAC streams. The last output stream is limited to 8 frames. The current code results in the first two output streams having 12 frames, after this commit all three streams have 8 frames and are the same length. This new result is better, since it is predictable. - mkv-1242 The test streamcopies one video and one audio stream, video is limited to 11 frames. The new result shortens the audio stream so that it is not longer than the video.
3 years ago
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.
2 years ago
/* no more frames will be sent for this stream */
int finished;
fftools/ffmpeg: use the sync queues to handle -frames Same issues apply to it as to -shortest. Changes the results of the following tests: - matroska-flac-extradata-update The test reencodes two input FLAC streams into three output FLAC streams. The last output stream is limited to 8 frames. The current code results in the first two output streams having 12 frames, after this commit all three streams have 8 frames and are the same length. This new result is better, since it is predictable. - mkv-1242 The test streamcopies one video and one audio stream, video is limited to 11 frames. The new result shortens the audio stream so that it is not longer than the video.
3 years ago
uint64_t frames_sent;
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
uint64_t samples_sent;
fftools/ffmpeg: use the sync queues to handle -frames Same issues apply to it as to -shortest. Changes the results of the following tests: - matroska-flac-extradata-update The test reencodes two input FLAC streams into three output FLAC streams. The last output stream is limited to 8 frames. The current code results in the first two output streams having 12 frames, after this commit all three streams have 8 frames and are the same length. This new result is better, since it is predictable. - mkv-1242 The test streamcopies one video and one audio stream, video is limited to 11 frames. The new result shortens the audio stream so that it is not longer than the video.
3 years ago
uint64_t frames_max;
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
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.
2 years ago
} 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;
fftools/sync_queue: support operation with no limiting streams ffmpeg CLI will not create such queues currently, but this will become useful in following commits.
2 years ago
int have_limiting;
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
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.
2 years ago
};
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);
}
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
/**
* 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.
2 years ago
{
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
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.
2 years ago
return (sq->type == SYNC_QUEUE_PACKETS) ?
frame.p->pts + frame.p->duration :
fftools/ffmpeg: replace AVFrame.pkt_duration with duration Mistakenly reintroduced in 4740fea7ddf.
2 years ago
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.
2 years ago
}
fftools/sync_queue: make sure audio duration matches sample count For audio AVFrames, nb_samples is typically more trustworthy than duration. Since sync queues look at durations, make sure they match the sample count. The last audio frame in the fate-shortest test is now gone. This is more correct, since it outlasts the last video frame.
2 years ago
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.
2 years ago
static int frame_null(const SyncQueue *sq, SyncQueueFrame frame)
{
return (sq->type == SYNC_QUEUE_PACKETS) ? (frame.p == NULL) : (frame.f == NULL);
}
fftools/sync_queue: use timebase from input frames/packets They are always properly set now. Avoid a separate timebase-setting call, which duplicates the knowledge of the timebase being used.
2 years ago
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.
2 years ago
static void finish_stream(SyncQueue *sq, unsigned int stream_idx)
{
SyncQueueStream *st = &sq->streams[stream_idx];
st->finished = 1;
fftools/ffmpeg: use the sync queues to handle -frames Same issues apply to it as to -shortest. Changes the results of the following tests: - matroska-flac-extradata-update The test reencodes two input FLAC streams into three output FLAC streams. The last output stream is limited to 8 frames. The current code results in the first two output streams having 12 frames, after this commit all three streams have 8 frames and are the same length. This new result is better, since it is predictable. - mkv-1242 The test streamcopies one video and one audio stream, video is limited to 11 frames. The new result shortens the audio stream so that it is not longer than the video.
3 years ago
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.
2 years ago
/* 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];
fftools/ffmpeg: use the sync queues to handle -frames Same issues apply to it as to -shortest. Changes the results of the following tests: - matroska-flac-extradata-update The test reencodes two input FLAC streams into three output FLAC streams. The last output stream is limited to 8 frames. The current code results in the first two output streams having 12 frames, after this commit all three streams have 8 frames and are the same length. This new result is better, since it is predictable. - mkv-1242 The test streamcopies one video and one audio stream, video is limited to 11 frames. The new result shortens the audio stream so that it is not longer than the video.
3 years ago
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.
2 years ago
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];
fftools/ffmpeg: use the sync queues to handle -frames Same issues apply to it as to -shortest. Changes the results of the following tests: - matroska-flac-extradata-update The test reencodes two input FLAC streams into three output FLAC streams. The last output stream is limited to 8 frames. The current code results in the first two output streams having 12 frames, after this commit all three streams have 8 frames and are the same length. This new result is better, since it is predictable. - mkv-1242 The test streamcopies one video and one audio stream, video is limited to 11 frames. The new result shortens the audio stream so that it is not longer than the video.
3 years ago
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.
2 years ago
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 */
fftools/ffmpeg: use the sync queues to handle -frames Same issues apply to it as to -shortest. Changes the results of the following tests: - matroska-flac-extradata-update The test reencodes two input FLAC streams into three output FLAC streams. The last output stream is limited to 8 frames. The current code results in the first two output streams having 12 frames, after this commit all three streams have 8 frames and are the same length. This new result is better, since it is predictable. - mkv-1242 The test streamcopies one video and one audio stream, video is limited to 11 frames. The new result shortens the audio stream so that it is not longer than the video.
3 years ago
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.
2 years ago
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++)
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
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.
2 years ago
/* 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;
fftools/sync_queue: make sure audio duration matches sample count For audio AVFrames, nb_samples is typically more trustworthy than duration. Since sync queues look at durations, make sure they match the sample count. The last audio frame in the fate-shortest test is now gone. This is more correct, since it outlasts the last video frame.
2 years ago
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.
2 years ago
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;
fftools/sync_queue: use timebase from input frames/packets They are always properly set now. Avoid a separate timebase-setting call, which duplicates the knowledge of the timebase being used.
2 years ago
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.
2 years ago
ret = objpool_get(sq->pool, (void**)&dst);
if (ret < 0)
return ret;
frame_move(sq, dst, frame);
fftools/sync_queue: make sure audio duration matches sample count For audio AVFrames, nb_samples is typically more trustworthy than duration. Since sync queues look at durations, make sure they match the sample count. The last audio frame in the fate-shortest test is now gone. This is more correct, since it outlasts the last video frame.
2 years ago
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);
}
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
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.
2 years ago
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);
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
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++;
fftools/ffmpeg: use the sync queues to handle -frames Same issues apply to it as to -shortest. Changes the results of the following tests: - matroska-flac-extradata-update The test reencodes two input FLAC streams into three output FLAC streams. The last output stream is limited to 8 frames. The current code results in the first two output streams having 12 frames, after this commit all three streams have 8 frames and are the same length. This new result is better, since it is predictable. - mkv-1242 The test streamcopies one video and one audio stream, video is limited to 11 frames. The new result shortens the audio stream so that it is not longer than the video.
3 years ago
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.
2 years ago
return 0;
}
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
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;
return 0;
}
// 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;
}
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.
2 years ago
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];
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
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.
2 years ago
SyncQueueFrame peek;
int64_t ts;
int cmp = 1;
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
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.
2 years ago
av_fifo_peek(st->fifo, &peek, 1, 0);
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
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.
2 years ago
/* 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.
fftools/sync_queue: support operation with no limiting streams ffmpeg CLI will not create such queues currently, but this will become useful in following commits.
2 years ago
* 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.
2 years ago
*/
fftools/sync_queue: support operation with no limiting streams ffmpeg CLI will not create such queues currently, but this will become useful in following commits.
2 years ago
if (cmp <= 0 || ts == AV_NOPTS_VALUE || !sq->have_limiting) {
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
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.
2 years ago
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;
}
fftools/ffmpeg: use the sync queues to handle -frames Same issues apply to it as to -shortest. Changes the results of the following tests: - matroska-flac-extradata-update The test reencodes two input FLAC streams into three output FLAC streams. The last output stream is limited to 8 frames. The current code results in the first two output streams having 12 frames, after this commit all three streams have 8 frames and are the same length. This new result is better, since it is predictable. - mkv-1242 The test streamcopies one video and one audio stream, video is limited to 11 frames. The new result shortens the audio stream so that it is not longer than the video.
3 years ago
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.
2 years ago
{
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;
fftools/ffmpeg: use the sync queues to handle -frames Same issues apply to it as to -shortest. Changes the results of the following tests: - matroska-flac-extradata-update The test reencodes two input FLAC streams into three output FLAC streams. The last output stream is limited to 8 frames. The current code results in the first two output streams having 12 frames, after this commit all three streams have 8 frames and are the same length. This new result is better, since it is predictable. - mkv-1242 The test streamcopies one video and one audio stream, video is limited to 11 frames. The new result shortens the audio stream so that it is not longer than the video.
3 years ago
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.
2 years ago
fftools/sync_queue: support operation with no limiting streams ffmpeg CLI will not create such queues currently, but this will become useful in following commits.
2 years ago
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.
2 years ago
return sq->nb_streams++;
}
fftools/ffmpeg: use the sync queues to handle -frames Same issues apply to it as to -shortest. Changes the results of the following tests: - matroska-flac-extradata-update The test reencodes two input FLAC streams into three output FLAC streams. The last output stream is limited to 8 frames. The current code results in the first two output streams having 12 frames, after this commit all three streams have 8 frames and are the same length. This new result is better, since it is predictable. - mkv-1242 The test streamcopies one video and one audio stream, video is limited to 11 frames. The new result shortens the audio stream so that it is not longer than the video.
3 years ago
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);
}
fftools/sync_queue: allow requesting a specific number of audio samples This will be made useful in following commits.
2 years ago
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.
2 years ago
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);
}