[Mesa-dev] Performance glxSwapBuffers 32 bit vs. 64 bit

[Jose Fonseca]

----- Original Message -----
> 
> Am Freitag, 11. November 2011 14:33 CET, Michel Dänzer
> <michel at daenzer.net> schrieb:
>  
> > On Fre, 2011-11-11 at 14:15 +0100, Theiss, Ingo wrote:
> > > Am Freitag, 11. November 2011 12:09 CET, Michel Dänzer
> > > <michel at daenzer.net> schrieb:
> > > 
> > > > So It makes sense to find a glReadPixels in VirtualGL's
> > > > glxSwapBuffers.
> > > > 
> > > > Ah. I thought the time measurements in Ingo's original post
> > > > were for the
> > > > Mesa glXSwapBuffers, not the VirtualGL one. If it's the latter,
> > > > then
> > > 
> > > > this makes sense.
> > > > 
> > > > Ingo, I noticed that your 64-bit and 32-bit drivers were built
> > > > from
> > > > slightly different Git snapshots. Is the problem still the same
> > > > if you
> > > > build both from the same, current snapshot?
> > > > 
> > > > If yes, have you compared the compiler flags that end up being
> > > > used in
> > > > both cases? E.g., in 64-bit mode SSE is always available, so
> > > > there might
> > > > be some auto-vectorization going on in that case.
> > > 
> > > I´ve rebuild my 64-bit and 32-bit drivers from a fresh Git
> > > snapshot
> > > and turned on all processor optimizations in both builds.
> > > But nevertheless the readback performance measured inside
> > > VirtualGL is
> > > only half of the 64-bit readback performance and of course the
> > > rendered window sceene is noticeable slower to :-(
> > > 
> > > Here are the compiler flags used.
> > > 
> > > 32-bit:
> > > 
> > > CFLAGS: -O2 -Wall -g -m32 -march=amdfam10 -mtune=amdfam10
> > > -fno-omit-frame-pointer -Wall -Wmissing-prototypes -std=c99
> > > -ffast-math -fno-strict-aliasing -fno-builtin-memcmp -m32 -O2
> > > -Wall -g -m32 -march=amdfam10 -mtune=amdfam10 -fno-omit
> > > -frame-pointer -fPIC -m32
> > 
> > Have you tried adding -mfpmath=sse to the 32-bit CFLAGS? According
> > to my
> > gcc documentation, that option is enabled by default in 64-bit mode
> > but
> > disabled in 32-bit mode.
> > 
> > Anyway, I guess there's room for optimization in glReadPixels...
> 
> Ok I have added -mfpmath=sse to the 32-bit CFLAGS and the readback
> performance increased from 30.44 Mpixels/sec to 48.92 Mpixel/sec. We
> are getting closer to the 64-bit performance.

hmm. you should try -msse2 too. It's implied on 64bits, and I'm not sure if -march/-mfpmath=sse by itself will enable the intrinsics.

Jose