[gmx-developers] Assembler loops on Itanium2 Montecito (again)

Rafael R. Pappalardo rafapa at us.es
Tue Feb 26 18:26:15 CET 2008


On Tuesday 26 February 2008 16:21:40 Carsten Kutzner wrote:
> Rafael R. Pappalardo wrote:
> > I don't have enough expertize to fix the problem with assembly loops on
> > Montecito. Following Carsten Kutzner suggestion I have compiled GROMACS
> > enabling fortran loops and obtained a considerable speed up with respect
> > to the C version.
> >
> > My suggestion is to include an specific test for Montecito CPUs and
> > disable the assembly loops similarly to what it's done when the
> > NOASSEMBLYLOOPS environment variable is set. I don't know if after the
> > check for cpuid it's still possible to dissable the use of assembly
> > loops.
> >
> > In this way Gromacs could run in a system with a mixture of Itanium2
> > processors. Running in Montecito processors will use fortran loops and on
> > non-Montecito ones the assembly loops without the user needing to change
> > the environment which cannot be feasible if there is a pool of cpus and
> > you don't know in advance which type of processor will run the job.
>
> Hi Rafael,
>
> why don't you disable the assembly code at install time then? This way
> Gromacs would run on all Itanium CPUs, regardless if Montecito or not.
>
> ./configure --disable-ia64-asm --enable-fortran --enable-mpi
>
> Hope that helps,
>   Carsten
>
> > Thanks for the attention.
> >
> > Best regards,
> >
> > Rafael

I was assuming that the assembly code is faster on non Montecito cpus. I read 
somewhere that it achieve something like 80% of peak performance.

Thanks for the suggestion. What I am doing know if setting or not 
NOASSEMBLYLOOPS when I am sure where it will run.

Rafael


-- 
Dr. Rafael R. Pappalardo
Dept. Physical Chemistry, Univ. de Sevilla (Spain)
e-mail: rafapa at us.es



More information about the gromacs.org_gmx-developers mailing list