[FFmpeg-devel] [PATCH] lavf/mxfenc: Bump EDIT_UNITS_PER_BODY

[Marton Balint]

On Mon, 13 Feb 2023, Tomas Härdin wrote:

> mån 2023-02-13 klockan 20:33 +0100 skrev Marton Balint:
>> 
>> 
>> On Mon, 13 Feb 2023, Tomas Härdin wrote:
>> 
>> > Passes FATE because we don't have any tests that mux files with a
>> > whopping 250 frames. Tested with a jpeg2000 Tears of Steel sample.
>> 
>> Have you checked what BMX does? Because as far as I remember for 
>> RDD9 (XDCAM HD) this should not exceed 10 seconds.
>
> I have not checked BMX. Also such constraints should be explicitly
> documented and there should be tests for them. An option like -xdcam
> might be appropriate. Currently there is no justification at all in the
> code for the choice of 250.

Intent of RDD9 compliance.

>> 
>> I suggest adding an option for this with -1 default, and unless the
>> user 
>> overrides it make it 238 in case of MPEG2 essence, and infinite (0) 
>> otherwise.
>
> I wouldn't be opposed to having it use a dynamic array. As for MPEG2,
> that does stink a bit but perhaps it avoids causing grief for most
> users. Here we again come to the problem that we don't know the
> workflows in which mxf* are used, and I don't like that we're cargo
> culting defaults.

I think we have a pretty good idea that MPEG2 in MXF usually means some 
broadcast realted use, therefore intent of RDD9 compliance by default is 
not insane at all for MPEG2 essence. Can we please keep it as default for 
MPEG2?

>
>> 
>> There is also code which allocates the index entries in 
>> EDIT_UNITS_PER_BODY increments, that probably should be replaced with
>> av_dynarray2_add...
>
> That (re)allocation happens at most twice for assuming GOP size <
> EDIT_UNITS_PER_BODY
>
> The reason I bring this all up is because opening MXF files muxed by
> lavf over HTTP is slow as hell. This will be true for any other storage
> where seeking incurs a non-trivial cost. HDDs come to mind. Perhaps
> that is really an mxfdec problem, but then parsing becomes even hairier
> than it already is. We'd have to binary search the file when seeking,
> hoping to find the necessary index table segments on-the-fly..

A sane muxer duplicates all index table segments in the footer, so a smart 
demuxer can read the whole index from there. Yes, mxfdec problem. I have a 
WIP patch somewhere fixing that.

Regards,
Marton