[FFmpeg-devel] MXF D10 regression tests

Måns Rullgård mans
Wed Mar 18 22:46:47 CET 2009


Michael Niedermayer <michaelni at gmx.at> writes:

> On Wed, Mar 18, 2009 at 09:19:26PM +0000, M?ns Rullg?rd wrote:
>> Michael Niedermayer <michaelni at gmx.at> writes:
>> 
>> > On Wed, Mar 18, 2009 at 08:33:48PM +0000, M?ns Rullg?rd wrote:
>> >> Michael Niedermayer <michaelni at gmx.at> writes:
>> >> 
>> >> > what should be done is
>> >> > * identify and document which code exactly needs this relation
>> >> >   (your patch is a start for this)
>> >> > * find what speed effect it has to fix that code
>> >> > * depending on the possibility to fix the code without speed
>> >> >   regressions document which buffers require the relation and
>> >> >   fix the code used to allocate them
>> >> > * fix STRIDE_ALIGN for each arch (really a job for the
>> >> >   maintainers of that archs)
>> >> 
>> >> Do we have to do all of that before we can fix the encoding on PPC and
>> >> ARM?
>> >
>> > well, we cant fix the code requireing the relation without
>> > knowing where it is and without knowing the speed effect.
>> > That leaves
>> > * changing STRIDE_ALIGN to 8
>> 
>> It has to be 16, or it will crash.
>
> in svq1 or somewhere else?

A lot of functions need 16-byte alignment on ARM.  I don't remember
which ones exactly.

>> > * changing default_get_buffer() to maintain the stride relation with a big
>> >   FIXME.
>> 
>> How big a fixme?
>
> big enough so its not ignored until godot

Will it make the code any worse than it is now?

-- 
M?ns Rullg?rd
mans at mansr.com




More information about the ffmpeg-devel mailing list