Re: ATLAS on the IA64 ?

[R Clint Whaley <rwhaley@cs.utk.edu>]

Daniel,

>Has anyone had any success in building ATLAS on an IA64 platform.
>
>So far I have succeeded using gcc but I am disappointed by the performance
>(only around 1 Gflops/s).

I got access to a 666Mhz IA64 through Compaq's Testdrive program, and
I also got a max performance of roughly 1Gflop, a disappointing less
then 1/2 of peak.  However, that machine only had gcc, and I was unable
to find any way to throw any IA64-specific flags; since compilers are so
important for VLIW, I was not unduly alarmed by the depressed performance.
I noticed that ATLAS's best case used only 62 registers, instead of the
hardware 128.  My guess is gcc is not giving them to me; this is important,
since the IA64's FPU does not use the L1 cache . . .

Another aspect to consider on performance is memory bandwidth.  The IA64
may do 4 flops/cycle like an IBM, but does anyone know if it has the memory
bandwidth to feed it?  If it has PC-style memory, it would not surprise me
to see memory holding us to around 60% of peak, even if we got everything
else rolling . . .

I didn't have access to any other compilers; this is doubly unfortunate, as
gcc only achieved 1/2 peak even using the all-register code (which runs at
peak on PIII, for example) . . .

Has anyone heard any performance numbers for the IA64 in general?  It would
be informative to hear if anyone else is getting better performance than
this on real computations.  Unfortunately, there were no vendor blas to
compare against on the machine I had access to . . .

Cheers,
Clint