Need Gigaflops from GCC on P4

[Wonder]

Anybody know how to configure GCC?

I've got SlackWare 8, which for stability reasons has been set up entirely using the '386 instruction set. I think they also have the GCC compiler set up to only generate 386 instructions.

Using this, I can get 0.7 billion double precision floating point operations per second, computing the inner product of two long vectors (on a 1.8GHz Pentium4). That's not bad, but I think I should be able to get more like 8.0 Gflops (10X), if I can invoke the SIMD/SSE2 capabilities of my Pentium 4. The GCC compiler is supposed to support this. I just want to do a little sub-atomic physics at home in my spare time.

I think I may need the -b MACHINE option, which in turn requires (I think?) a library configured for P4 instead of i386. I've tried -mcpu=pentium4 and -march=pentium4 and -msse2 compile options, with no effect. I haven't tried -mfpmath=sse.

Do I need to recompile some libraries for pentium4? Do I need to set up GCC as if it's doing a cross-compile (for a different platform)? Am I even asking the right questions?

Thanks!

[ernestus]

Waw, what a level!
I think gnu.org is a good place to start (I think SSE & 3DNOW! are assembly supported, but may be I am wrong).
Luck.
E

[ernestus]

Ok, I was wrong, what I said is true in my gcc version. here go two useful links:

here is a discussion about the subject:
http://compilers.iecc.com/comparch/index/2002-05

look for the MMX/3Dnow!/SSE/SSE2 compilers thread. As you can see, SSE2 instructions may take longer than FPU instructions.

And here there is a depiction of what can you expect:
http://www.andrew.cmu.edu/~komarek/work/RobustCholeskyPerf/RobustCholeskyPerf.html

I realize x86 architecture is not appropiated for number crunching (look at the great results of the humble G3 300MHz Apple iBook)

Bye
E.