[Mesa-dev] [PATCH] r600g: Various patches

[Marek Olšák]

2011/6/12 Mathias Fröhlich <Mathias.Froehlich at gmx.net>:
> On Sunday, June 12, 2011 16:53:08 Marek Olšák wrote:
>> 2011/6/12 Mathias Fröhlich <Mathias.Froehlich at gmx.net>:
>> > 0005-r600g-Make-use-of-PIPE_TRANSFER_DISCARD_WHOLE_RESOUR.patch:
>> > This makes use of the DISCARD_WHOLE_RESOURCE flag for memory mappings.
>>
>> That won't work. The old buffer may still be set as a vertex
>> buffer/texture. If you replace the whole buffer like that, all the
>> states in all existing pipe contexts where the old buffer is set in
>> must be re-emitted as well (e.g. by marking them as dirty), which is unfun.
> Hmm, I see. This 'find all context' stuff is really bad ...
> So, forget about this last one.
>
> Ok, then I wonder if it would work when this is just implemented at the
> state_stracker level in st_cb_bufferobjects.c:
>
> There is the implementation of glBufferData which detaches itself from the
> underlying pipe_resource by unreferencing the old one and creating a new one.
> So, given this works well in all cases (does it?),

No, it doesn't work either. Thanks for point out a serious bug in
st/mesa. I have added a new piglit test for this: vbo-bufferdata

We shouldn't just grab a new buffer without updating all the places
it's bound in no matter if it's a driver or mesa core.

Not sure what the correct fix would be. We must still reallocate the
buffer if the old and new buffer sizes don't match.

Marek