drm/radeon/kms: improve performance of blit-copy
ihadzic at research.bell-labs.com
Thu Oct 13 12:18:17 PDT 2011
On Thu, 13 Oct 2011, Roland Scheidegger wrote:
> I guess it isn't possible to temporarily disable some RBEs or otherwise
> reconfigure the chip that you could get the same performance for the
> high-end chips?
According to the conversation I had with Alex, this *is* possible but
requires the pipeline and cache flush. So it is unclear what the overall
gain will be given the flush penalty.
Also, this phenomena occurs only when GTT is involved in the copy.
VRAM-to-VRAM copy in which there is no host memory involved (for which I
added a benchmark, but didn't report in my note yesterday), high-end
devices are beating low-end ones big time .... they better be ;-)
So if we can get RBE-reduction to work, it should be turned on only when
one of the BOs is in GTT domain. I looked at what it would take to do
this, and it's doable, but requires hacks at many places.
More information about the dri-devel