[Mesa-dev] r600g on rv635 and broken mipmaps

[Alex Deucher]

2011/5/5 Mathias Fröhlich <Mathias.Froehlich at gmx.net>:
>
> Hi all,
>
> On Thursday, May 05, 2011 04:32:03 you wrote:
>> Okay my guess at the problem is that:
>>
>> the CP tracks coherency but the SURFACE_BASE_UPDATE stuff might rely
>> on the base in the CB being the same as the texture BASE which it won't
>> be in the case where we are rendering to mip sublevels. Though I've no idea
>> how to workaround this without explicit flushes.
>
> Hmm, may be.
> I also thought that the surface sync packet has some special case optimzations
> for some of the probably often used flags that lead to that kind of behaviour.
>
> May be that 'flush all' in case of a new framebuffer target for these kind of
> chips is again a good idea instead of the finegrained flush dest caches.
>
> May be Alex finds some undocumented ideas somewhere in his bag :).

Apparently the CB/DB surface sync stuff has a number of issues on
r6xx, so we should just use event_write flushes for CB/DB.  A single
event write flush takes care of all dst caches.  Something like this
untested patch perhaps:
http://people.freedesktop.org/~agd5f/r600g_event_flush.diff

Alex