On Tue, 20 Sep 2011 13:35:56 +0200, Daniel Vetter <dan...@ffwll.ch> wrote:
> On Sat, Sep 17, 2011 at 09:57:05PM +0100, Chris Wilson wrote:
> > And whilst you are here, can you incorporate this patch?
> > else if (obj->gtt_space &&
> > + obj->map_and_fenceable &&
> > obj->base.write_domain != I915_GEM_DOMAIN_CPU) {
> > ret = i915_gem_object_pin(obj, 0, true);
>
> Not sure this is good. On !llc machines, gtt_pwrite is much faster, so we
> want to move the objects into the mappable part of the gtt to benefit from
> that. Without this, they'll just stay wherever they are.
This turns out to be a big win for machines where it avoids the pipeline
stall due to the eviction of an active page and allowing us to utilize
the full GTT for vertex data.
10% on pnv, 60% on snb for x11perf -aa10text. (Though on SNB this is
eclipsed by using LLC and a test for obj->cache_level). And it is likely
to be an improvement with geometry bound game benchmarks like openarena.
-Chris
--
Chris Wilson, Intel Open Source Technology Centre
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/intel-gfx
