[Radiance-general] Compile error on AMD running Fedora 4

[atelier iebele abel]

Hi Ian,

I used your recommended gcc options an got an rendering speed 
improvement of 1.035. That is 3.5 % (!).
I used gcc 4.0(etc, I dont know exactly) but did not find any option 
like "auto vectorization". We'll have to wait for that coming up in the 
future i think.
Thanks a lot for figuring out these settings so far !

Iebele




Ian Tester wrote:

> iebele wrote:
>
>> Current rmake command is:
>> #!/bin/sh
>> exec make "SPECIAL=" \
>>         "OPT=-m64 -O3 -mfpmath=sse -DSPEED=200" \
>>         "MACH=-Dlinux -D_FILE_OFFSET_BITS=64 -L/usr/X11R6/lib
>> -I/usr/include/X11 -DNOSTEREO" \
>>         ARCH=IBMPC "COMPAT=bmalloc.o erf.o getpagesize.o" \
>>         INSTDIR=/usr/local/bin \
>>         LIBDIR=/usr/local/lib/ray \
>>         ESUFFIX= \
>>         CC=gcc "$@" -f Rmakefile
>
>
>
> What version of gcc is installed on your system?
> Check using gcc -v
>
> Looking at the man page for gcc-3.4 on my Debian (Athlon XP) system, 
> I'd suggest the following settings:
>
> -march=k8    to enable the AMD64 instructions, implies the same
>         argument to -mtune.
>
> -msse2        to enable SSE2 instructions. Without this, the -mfpmath=sse
>         option is useless.
>
> -mfpmath=sse,387 this will give you a boost if the SSE unit is separate
>         from the old x87 unit but I don't know enough about the K8
>         architecture to know if it does. It will try to use both
>         execution units, but the man page says it still
>         experimental.
>
> -ffast-math    the man page says this is dangerous but I've never 
> noticed          any problems myself. Try it if you want.
>
> Apart from that, wait for upcoming GCC-4.x releases to flesh out its 
> optimisation features, in particular auto vectorization. The SSE2 
> instruction set can handle double precision FP math, better than the 
> single precision in SSE or Altivec. You should theoretically be able 
> to get some pretty awesome performance out of that processor. Reality 
> is another matter though... :(
>
> My other suggestion would be to check out some Gentoo web sites. Those 
> guys are nuts about compiler optimisations :P
>
> bye
>
> _______________________________________________
> Radiance-general mailing list
> Radiance-general at radiance-online.org
> http://www.radiance-online.org/mailman/listinfo/radiance-general
>