On Sat, Jun 1, 2013 at 11:47 PM, ET <sketchfoot.gmail.com> wrote:
> Hi,
>
> Apologies if this question has been asked before, but I could not find an
> answer when I searched.
>
> I wanted to know which compiler gives the best code performance for AMBER
> 12. I have compiled with both intel C++ & Fortran and GNU.
>
> There are less variations in the diff file using Gnu. I assume because the
>
> benchmark was run on an installation compiled with GNU?
>
Yes and no. I always use the Intel compilers, since the Fortran compiler
was known to provide faster executables (this is most apparent for pmemd on
CPUs, where the code was written to take advantage of vectorization
instructions that ifort tends to catch pretty well). I'm not sure if
anybody has done a test on recent gfortran vs. recent ifort: my instinct is
that the difference is not very pronounced.
There are a couple failures we 'know' about with the Intel compiler (those
should be going away for the next version of Amber, but we've just kind of
accepted them as innocuous). But both the Intel and GNU compilers get
tested thoroughly with released code.
>
> I am specifically interested in CUDA optimisation. Is there any benefit to
>
> using compilers from the Portland group?
>
No. In fact, I don't know that the Amber build supports CUDA + PGI. All
of the cuda performance comes from nvcc, so the choice of compiler is
irrelevant here. CPU compilers don't build code that runs on the GPU, and
with pmemd.cuda everything is run on the GPU. You could just as easily
rewrite the CPU part in Python or bash with minimal loss of performance ;).
HTH,
Jason
--
Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Candidate
352-392-4032
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Sat Jun 01 2013 - 21:30:02 PDT
