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

[Rafael R. Pappalardo]

On Tuesday 26 February 2008 19:52:00 Erik Lindahl wrote:
> Hi,
>
> On Feb 26, 2008, at 6:26 PM, Rafael R. Pappalardo wrote:
> > 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.
>
> It should be; provided the instruction latency timings haven't changed
> we keep the pipeline full.
> On the other hand, we first wrote the loops for McKinley, so there
> might have been some instruction timing differences, and ia64 is
> incredibly sensitive to those.
>
> Considering everybody seems to be going x86 nowadays I'm not sure if
> it's worth looking into re-scheduling the ia64 assembly, in particular
> if you're reasonably happy with the fortran performance :-)
>
> Cheers,
>
> Erik
>

I agree with not devoting more efforts to ia64 assembly from you, albeit I'll 
try to contact some people at SGI and see if they can find a solution. What I 
would like to see implemented if a test so if the CPU is not Itanium or 
Itanium 2 McKinley the use of assembly loops is disabled. The advantage is 
that the same executable could be run in a cluster composed of a mixture of  
McKinley  and Montecito CPUs. Do you believe it's possible to include such 
test?

Best regards,

Rafael


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