[Mesa-dev] [PATCH] intel: Improve teximage perf for Google Chrome paint rects

[Daniel Vetter]

On Thu, Sep 13, 2012 at 04:22:20PM +0300, Chad Versace wrote:
> On 09/11/2012 10:40 PM, Daniel Vetter wrote:
> > Only quick read-through but I'd have expected a has_llc check in there
> > - if vlv is anything like the previous platforms wc gtt will be much
> > faster there.
> 
> I'm not too familiar with why one would want to check has_llc here. I see that
> has_llc is checked in only one place in i965, in a function that also writes to
> a drm_intel_bo_map'd bo.
> 
> Are you suggesting that I replace the check for `intel->gen >= 6` with
> `intel->has_llc`?

Yep, since vlv is  gen7, but doesn't have llc caching. Hence the write
through the cpu map needs to be manually clflushed, which will be dead
slow. Also, if some unlucky soul runs this on an older kernel on snb
without llc enabled, he'll similarly suffer.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch