Hi Frédéric, I'm working with André on the sse2 optimizations. It would be great to cooperate with you to avoid overlapping. What about using a parallel git tree for our patches while they don't land mainline? Once the basic infrastructure I sent over is committed, working in parallel will be easier. About the deviation on results is mostly due to André's patch not checking for opaque colours and using streaming writes. From my tests streaming writes, for some reason, don't have stable performance across difference cpus. Aligned stores works very well, thou. Cheers, Rodrigo Kumpera <div class="gmail_quote">On Fri, Mar 14, 2008 at 2:42 PM, Frédéric Plourde <<a href="mailto:frederic.plourde@polymtl.ca">frederic.plourde@polymtl.ca</a>> wrote: <blockquote class="gmail_quote" style="border-left: 1px solid rgb(204, 204, 204); margin: 0pt 0pt 0pt 0.8ex; padding-left: 1ex;"> Hi André ! My name if Frederic Plourde (Fplourde on irc - "freenode" server, "cairo" channel, by the way... if you wanna join ;-)) I'm working with mozilla and I've been working on mmx for some time now.   Sorry I didn't answer sooner to you email, I was out of town for the last 3 days. Indeed, the mmx and sse projects are running and we intend to optimize pixman as much as we can until next version of firefox. When you came in 3 or 4 days ago, I was a little surprised to discover you had been working on the exact same things as I was working on. But I can't complain... your contributions will sure help pixman perform faster.  The only thing is : I really hope we'll manage to cooperate efficiently on mmx and sse2 optimizations as I intend to keep in touch with your team on a regular basis.  :-) Best, -fred- P.S. - I saw you preliminary results, attached to your email... And I was wondering about the very high standard deviations you got ... something like 30%... It's quite high.. Do you have any insight on what could explain it ? André Tupinambá a écrit : > Hi everyone, > > I just subscribe this mailing list and saw some archives about SSE(2) > optimization in pixman library. Are this (sub)project running? > > I'm studing the pixman some days and I think that I can help too. So I > wrote a proof of concept changing the fbCompositeSrc_8888x8888mmx to > use SSE2 instructions. > > In attachment there are the patch and some cairo-perf's results. I > don't know much about the autoconf tool so to compile the code I > needed to run the configure like this: > > CFLAGS=-msse2 ./configure > > And them change the config.h file to include this: > > #define USE_SSE 1 > > The firsts tests results are REALLY good, but I'm not sure that's > works fine and not introduce some bad pixels in the result images. The > perf was runned using: > > taskset 2 ./cairo-perf -i 500 paint_image_rgba_over > perf-sse2-run1 > > Best Regards > > André Tupinambá > > PS: If you saw anything weird in my English, please forgive me. I'm > from Brazil and English ain't my first language. Well, the comments > are LGPL too, fell free to fix them :) > > ------------------------------------------------------------------------ > > _______________________________________________ > cairo mailing list > <a href="mailto:cairo@cairographics.org">cairo@cairographics.org</a> > <a href="http://lists.cairographics.org/mailman/listinfo/cairo" target="_blank">http://lists.cairographics.org/mailman/listinfo/cairo</a> _______________________________________________ cairo mailing list <a href="mailto:cairo@cairographics.org">cairo@cairographics.org</a> <a href="http://lists.cairographics.org/mailman/listinfo/cairo" target="_blank">http://lists.cairographics.org/mailman/listinfo/cairo</a> </blockquote></div>