pixman with and without SSE2 benchmarks?

[Matthias Hopf]

On Sep 28, 08 16:53:05 +0200, Roland Scheidegger wrote:
> On 28.09.2008 11:52, Clemens Eisserer wrote:
> >> I'd bet against that :-). Core 2 has magnificent SSE performance indeed,
> >> but that's true for MMX just as well.
> > Well, Core2 (and AMD K10) got support for 128bit operations per clock,
> > whereas previous processors only supported 64bit at once, and took 2
> > cycles for 128 bit operations.
> > MMX is just 64-bit, so it should't matter much there.
> Ah yes you're right. Somehow I thought it would execute mmx twice as
> fast as well - clearly that's not true. Should indeed be much closer on
> older cpus.

OTOH pixman w/ SSE does have some severe rendering issues - read
bugzilla.novell.org, #426370.

I haven't created an upstream bug yet, I can do that if that sounds
reasonable.

Bottom line is that fbCompositeSolid_nx8888sse2 is broken. It renders
stripes, 4 pixels with the correct alpha, 4 pixels with 1-alpha,
repeating. We have seen this on 16bpp as well, but I assume that the
source always was 32bpp. So the repeating pattern is 32 bytes or 128
bits... looking familiar?

The other rendering paths haven't been extensively tested from our side.
Is there a pixman validation tool? I don't know of any, maybe rendertest
could be used, but so far it apparently didn't hit this case. Maybe it
doesn't do solid rendering, haven't looked at the code.


I have close to no experience with SSE2, so I cannot really fix this
issue myself. Andre, I CC'ed you because you committed the SSE2
implementation.

Thanks

Matthias

-- 
Matthias Hopf <mhopf at suse.de>      __        __   __
Maxfeldstr. 5 / 90409 Nuernberg   (_   | |  (_   |__          mat at mshopf.de
Phone +49-911-74053-715           __)  |_|  __)  |__  R & D   www.mshopf.de