[FFmpeg-devel] [PATCH v4 4/8] avformat/mux: add proper support for full N:M bitstream filtering

[Andreas Rheinhardt]

Marton Balint:
> 
> 
> On Tue, 5 May 2020, Andreas Rheinhardt wrote:
> 
>> Marton Balint:
>>>
>>>
>>> On Tue, 5 May 2020, Andreas Rheinhardt wrote:
>>>
>>>> Marton Balint:
>>>>> Previously only 1:1 bitstream filters were supported, the end of the
>>>>> stream was
>>>>> not signalled to the bitstream filters and time base changes were
>>>>> ignored.
>>>>>
>>>>> This change also allows muxers to set up bitstream filters regardless
>>>>> of the
>>>>> autobsf flag during write_header instead of during check_bitstream
>>>>> and those
>>>>> bitstream filters will always be executed.
>>>>
>>>> Ignoring the autobsf flag is not ok IMO. The muxer should not add a bsf
>>>> when the user explicitly didn't want this.
>>>>
>>>
>>> The user should not be allowed to create broken files, and the only way
>>> to achieve that is to force the BSF.
>>
>> No, one could also check in the muxer that the packets are ok (these
>> checks could be disabled if the bsf is enabled, of course) and error out
>> if it is not.
> 
> And that would duplicate some functionaltiy for negigable performance
> gain of not always doing a BSF.
> 
>>
>>> I don't see a use case for
>>> disabling BSF either for MXFenc of GXFenc. (in fact, from the top of my
>>> head I can't see a use case for disabling automatically inserted BSF-s
>>> in other muxers).
>>>
>> A user who already knows that his packets are ok can do so to avoid the
>> (admittedly small) overhead of these bsfs. (And even if you think that
>> the option should be removed/should be made non-public (for internal
>> testing only), then this does not change that this option exists and
>> removing it would require the usual deprecation period (and consensus
>> from other devs, of course; I personally don't have formed an opinion
>> about this yet).)
> 
> With my patch the autobsf option is still honored for all existing
> bitstream filters which are inserted mid-stream (in check_bitstream).
> Only the ones which are inserted at write_header(), so in mxfenc and
> gxfenc are forced.
> 
>> One use-case is if a user runs the pcm_rechunk bsf filter manually and
>> then sends the output to both a framecrc as well as the mxf muxers in
>> order to be able to test more thoroughly whether any write errors
>> happened. Or where you want to write the same audio to multiple mxf
>> muxers, all with the same audio settings.
> 
> And for them the additional BSF which is auto inserted will only pass on
> their packets, it will only *check* for correct packets, exactly what
> you suggested. I cannot accept the performance gain as argument.
> 
>>
>>> So if you feel strongly about this and we still want to avoid creating
>>> corrupt files then the BSFs should be processed in mxf_interleave and
>>> gxf_interleave separately from the auto inserted bitstream filters. Do
>>> you think that is better? I don't think so.
>>>
>> That would just be a worse form of automatically inserted bsf (and it
>> would not work for av_write_frame()).
> 
> Please suggest something reasonable to follow up on this.
> 
How about: You error out in write_header if there is an audio track and
the user does not want automatically inserted bitstream filters?

- Andreas