[cairo] [PATCH] SSE2 patch for pixman
andrelrt at gmail.com
Sat Apr 5 07:04:46 PDT 2008
Do you know? cairo-perf with 10000 iterations took about 13 hours in
P4!!! Well, I had so many free time that's finally reinstall the Linux
in my Core2 Quad machine.
In attach has the results of cairo-perf, with 10000 iterations, for
MMX and SSE2 fast-paths. There are two versions, one with Mobile
Pentium 4 3.2GHz and another with Core2 Quad 2.4GHz processor.
There are a small slowdown with test paint_similar_rgba_source-256,
but only in P4 machine, I'll investigate it. In another side, the top
speedup with Core2 machine was *only* 2.5x (compared with 8x in P4
one). I'll investigate it too.
But for me this patch is the final one.
On Tue, Apr 1, 2008 at 11:13 PM, André Tupinambá <andrelrt at gmail.com> wrote:
> Hi everyone,
> We finnish the code for SSE2 fast-paths in pixman. In attach we have
> the patch and two perfs, with only 50 iterations. I'll run overnight
> the perf with 10000 iterations with MMX and SSE2 fast-paths and will
> send tomorrow morning. I saw a slowdown in paint_image_rgba_source,
> I'll investigate this now with VTune analyser. Well, I only translate
> the MMX code to SSE2 one, I'll start the optimization process now ;)
> I ran my tests always with Mobile Pentium 4 3.2Ghz, so I don't know
> how the code performs with Core2 or AMD processors. Could Someone post
> the perf results with this processors?
> BTW, Rodrigo Kumpera is starting another proof of concept for dynamic
> code generation for use with pixman compose functions. Pedro Lamarão
> is running new OpenMP experiments too. Probably we will translate this
> code again in a near future, maybe with threads :)
> Best Regards,
> André Tupinambá
-------------- next part --------------
A non-text attachment was scrubbed...
Size: 65898 bytes
Desc: not available
Url : http://lists.cairographics.org/archives/cairo/attachments/20080405/6b29384f/attachment-0001.bin
More information about the cairo