Try just removing the "<compile=optimized>" from the top of that particular
source file.
What is the performance hit when you do that? (And more importantly, does
it solve the problem?)
On Wed, Jul 25, 2012 at 4:04 AM, <steinbrt.rci.rutgers.edu> wrote:
> Hi,
>
> >> xmolnu 40.3614616595491
> >> -xmol -40.3635833689616
> [...]
> > First: are you sure that x(1,i), xmolnu, and xmol are exactly the same on
> > V0 and V1?
>
> yes, at least to the degree that a write (6,*) prints them. The numbers
> above have 15 digits, I believe that is the expected amount for our _REAL_
> numbers.
>
> > Second: try adding -fltconsistency and/or -fp-model precise to the flags
> > for
>
> Compiling with additional switches solves the problem indeed, at a cost:
>
> -fltconsistency: 30% runtime increase
> -O0 instead of -O3 -xHost 100% runtime increase
> -fp-model precise 10% runtime increase
>
> (runtimes measured on my simple 10 md-step test case.)
>
> So it looks like '-fp-model precise' works best and should be added to the
> compiler directives. Is there a way to tell 'make' that only softcore.F90
> should need this flag? In config.h, it is set for the whole compile.
>
> Is there a similar compiler directive for the Portland group compilers
> which appear to show the same behaviour (I couldnt verify that, since I
> dont have those)
>
> Thanks Dave,
>
> Thomas
>
> Dr. Thomas Steinbrecher
> formerly at the
> BioMaps Institute
> Rutgers University
> 610 Taylor Rd.
> Piscataway, NJ 08854
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
--
Jason M. Swails
Quantum Theory Project,
University of Florida
Ph.D. Candidate
352-392-4032
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Wed Jul 25 2012 - 05:30:03 PDT
