[Pixman] [PATCH] Adding AVX2 implementation of the OVER and REVERSE-OVER operator

[Matt Turner]

On Wed, Aug 29, 2018 at 12:09 PM Matt Turner <mattst88 at gmail.com> wrote:
>
> On Wed, Aug 22, 2018 at 10:03 AM raghuveer devulapalli
> <raghuveer.devulapalli at intel.com> wrote:
> >
> > The AVX2 implementation of OVER and REVERSE OVER operator was
> > found to be upto 2.2 times faster (depending on the array size) than
> > the corresponding SSE2 version. The AVX2 and SSE2 were benchmarked
> > on a Intel(R) Core(TM) i5-6260U CPU @ 1.80GHz
> >
> > Moving the helper functions in pixman-sse2.c to pixman-sse2.h. The AVX2
> > implementation uses the SSE2 version for manipulating pixels that are not
> > 32 byte aligned and hence, it made sense to separate the SSE2 helper
> > functions into a separate file to be included in the AVX2 file rather
> > than duplicate code.
>
> Let's please move the helpers into pixman-sse2.h in a separate commit
> from the one that adds AVX2 code paths.
>
> We typically have more substantial benchmarks in the commit message.

I ran all of the cairo traces in the benchmarks directory and couldn't
measure any difference. You'll have to describe your benchmarking.