[FFmpeg-devel] MXF D10 regression tests

[Reimar Döffinger]

On Wed, Mar 18, 2009 at 03:40:52PM +0100, Michael Niedermayer wrote:
> iam not asking for all of it to be taken into account but rather that
> there is some framework in which such things can be filled in where
> they are needed.
> (we also have avcodec_align_dimensions())

Well, you are the maintainer, I leave it to you to come up with a
solution. At least it should be clear enough what exactly is the issue
and can be fixed without access to a PPC machine.

> currently STRIDE_ALIGN is 8 for x86* and that works out exactly because 420
> and the linesize relations gurantee 16 byte alignment for luma.

I want to make completely clear: there is nothing in the code that
guarantees any "linesize relations", it is pure luck it works, and the
only reason is that the width is required to be a multiple of 16 and
STRIDE_ALIGN << chroma_x_shift divides (more precisely is equal to)
this on x86.

> if 444 is ever added it will need some special handling like svq1 does,
> also other archs than x86 set STRIDE_ALIGN to 16 while they possibly
> like x86 really just mean 16 for luma ...

Since (probably) none of that is documented that won't be fixed anytime
soon, at least not by someone who does not know the asm by heart already.