[Mesa-dev] Testing out Kristian's fast mtx patch

[Kristian Høgsberg]

On Mon, Oct 9, 2017 at 7:45 PM, Timothy Arceri <tarceri at itsqueeze.com> wrote:
> After a recent discussion about this code from 2015 I was curious
> to give it a try. The outstanding review item was that we shouldn't
> be replacing the C11 mtx type/functions with our own, so I've renamed
> the fast path to simple_mtx* and added a couple of patches to make use
> of it.
>
> The idea is this fast mtx can be used in place of the full mtx
> implementation when its of type mtx_plain.
>
> I though if anywhere we might see a change in the drawoverhead piglit
> test but I didn't see any real change.
>
> Anyway since I've made the updates I thought I'd send it out. Maybe
> someone else might find some better results. Kristian reported a 10%
> increase in some internal Intel benchmarks, I wonder if thats still the
> case.

Hi Tim, thanks for reviving this. The one case where I saw that big
win was one of Intels internal micro-benchmarks (SynMark). It was
bottlenecking on bind/unbind of various objects and much of the
overhead was in the locking there. Any real world usecase without
excessive bind/unbind probably won't see any improvement.

Kristian