[Mesa-dev] i965: Improved shader_time accuracy, using send-from-GRF

[Kenneth Graunke]

On 03/19/2013 05:05 PM, Eric Anholt wrote:
> I've been experimenting with the simd16 replicated data message
> (non-gen7-specific fast clears), and was bothered that shader_time wasn't
> noting any time spent in our clear FS currently, so I wasn't totally sure
> which apps I should be looking at.
>
> Many apps won't spend any time in the clear FS since they don't clear
> color, in which case the FS isn't dispatched.  But for those that do
> (like minecraft), it would be nice to see how much is spent clearing.  I
> also knew that we were significantly under-counting VS time by not
> counting URB payload setup, which I'd worked on a patchset for
> previously.
>
> This series fixes those two problems.  Now we know that 7% of non-EOT
> instruction time in minecraft is in the clear color setup (wow), and that
> VSes are a bit of a bigger deal than we previously knew (36% of minecraft,
> 23% of ETQW, 7.5% of GLB2.7)
>
> Branch is at shadertime-from-grfs of my tree.

This series is great!  Thanks so much for fixing this; it definitely 
gives more accurate times now.

Other than my proposed tidying of patch 2 and trivial comments on patch 
5, this series is:

Reviewed-by: Kenneth Graunke <kenneth at whitecape.org>