[FFmpeg-devel] [PATCH v3 0/1] avformat/demux: fix accurate probing of durations in mpegts/ps

[Nicolas Gaullier]

>Envoyé : lundi 29 avril 2024 19:35
>À : FFmpeg development discussions and patches <ffmpeg-devel at ffmpeg.org>
>Objet : Re: [FFmpeg-devel] [PATCH v3 0/1] avformat/demux: fix accurate probing of durations in mpegts/ps
>
>>Envoyé : lundi 22 avril 2024 14:32
>>À : ffmpeg-devel at ffmpeg.org
>>Objet : Re: [FFmpeg-devel] [PATCH v3 0/1] avformat/demux: fix accurate 
>>probing of durations in mpegts/ps
>>
>>>De : Nicolas Gaullier <nicolas.gaullier at cji.paris> Envoyé : mardi 2 
>>>avril 2024 23:26 Objet : [PATCH v3 0/1] avformat/demux: fix accurate 
>>>probing of durations in mpegts/ps
>>>
>>>v3: rebased after ed9363052f4b8b8 applied tonight (add 
>>>duration_probesize AVOption)
>>>
>>>Note: I have no other plan for demux/probing; with these two patches, I can cover my use cases, especially mpegts-concats.
>>>
>>>For remembering, previous cover-letters:
>>>
>>>v1--------------------
>>>ff_read_packet() is more lightweight, but it leads to important issues when looking for accurate durations.
>>>As a side effect, the code looks also simpler with regular av_read_frame() calls.
>>>1)Updates in the fate tests do exhibit most of the results.
>>>
>>>2)See also more directly the case of an audio PES containing many frames:
>>>>ffprobe tests/data/lavf/lavf.ts -select_streams a -show_entries 
>>>>stream=duration -of flat
>>>Before patch:
>>>  streams.stream.0.duration="0.757556"
>>>After patch:
>>>  streams.stream.0.duration="1.018778"
>>>
>>>3)Here is an additional (commonplace) sample to demonstrate the benefit for twofields-encoded video:
>>>>https://0x0.st/HFbm.ts (say h264-50i_mp2.ts)
>>>
>>>>ffprobe h264-50i_mp2.ts -show_entries stream=duration -of flat
>>>Before patch:
>>>  streams.stream.0.duration="2.060000"
>>>  streams.stream.1.duration="1.176000"
>>>After patch:
>>>  streams.stream.0.duration="2.080000"
>>>  streams.stream.1.duration="1.200000"
>>>
>>>
>>>v2--------------------
>>>v1: There was an issue with teletext where resolution is set just once at decoder init (teletext resolution is fixed/hard coded), so it is somewhat fragile: when a demuxer context update occurs, it is lost/overriden by >>>avcodec_parameters_to_context(sti->avctx, st->codecpar) in read_frame_internal.
>>>They could have been other scenario besides teletext, I don't know.
>>>v2: So now at estimate_timings_from_pts, with one or more seeking involved (seeking is detected by the mpegts demuxer and set last_vn=-1, so pmt is forced/updated and results in demuxer context update), it is required to >>>preserve the info in codecpar at first.
>>>Thanks to Michael for reporting the issue.
>>>
>>>
>>>Nicolas Gaullier (1):
>>>  avformat/demux: Fix accurate probing of durations in mpegts/ps
>>>
>>> libavformat/demux.c                           |  36 ++--
>>> tests/ref/fate/concat-demuxer-simple2-lavf-ts | 170 +++++++++---------
>>> tests/ref/fate/ts-opus-demux                  |   4 +-
>>> 3 files changed, 100 insertions(+), 110 deletions(-)
>>
>>Ping for review ?
>>The patch (https://patchwork.ffmpeg.org/project/ffmpeg/list/?series=11386) still applies on current master.
>>Thanks.

Another ping...
Thanks