AMBER Archive (2007)

Subject: Re: AMBER: AMBER8 compiler with intel 9.1 compiler

From: Scott Brozell (sbrozell_at_scripps.edu)
Date: Tue May 22 2007 - 21:20:26 CDT

Hi,

On Fri, 18 May 2007, Peter Z. Qin wrote:

> On 17 May 2007, dac wrote:
>
> > 2. If so, try just compiling ew_fft.f (and/or ew_recip.f) at low
> > optimization, everything else at high optimzation. My guess is that
> > ew_fft.f
> > is the culprit...there seem to be things in there that the intel
> > optimizers don't like.
>
> I modified the ew_fft.f by removing "! <compile=optimized>" and also applied
> budfix64. Then I recompile with optimization on. However, failures occur at
> the same cases (cytosine, dhfr, etc.)

! <compile=optimized>
is irrelevant to non-humans.
Try really compiling ew_fft.f with low opts:

cd src/sander
make sander
rm ew_fft.o
make -e FOPTFLAGS='$(FFLAGS)' ew_fft.o
make sander
then re-test

This recipe is designed to be bullet proof to amber-builder-non-experts.

> > If you want, the next thing to do is to try to figure out which routine
> > doesn't
> > work with optimization turned on. Here is a suggested path, (for you or
> > someone else):
> >
> > 1. Do dhfr one step with verbose=2 in the &ewald namelist. Look at the
> > differences between optimized and unoptimzed code. If you are lucky, the
> > differences will be in the reciprocal contribution to the electrostatic
> > energy.
>
>
> With verbose=2, I have ran dhfr/Run.dhfr. Part of the difference between
> mdout.dhfr is listed below. The version with MPI flag is compiled without
> optimization, and the version with no Flags is compiled as indicated above.
>
> The big difference seems to be Eevir vs. Ecoulomb.
>
> 188,189c191,192
> < Iso virial = -187936.116173447459
> < Eevir vs. Ecoulomb = 0.000535006138
> ---
> > Iso virial = -187933.775401700172
> > Eevir vs. Ecoulomb = 0.000037834768
>
> What next?

Try the above first before diving into the verbose output.

Scott

-----------------------------------------------------------------------
The AMBER Mail Reflector
To post, send mail to amber_at_scripps.edu
To unsubscribe, send "unsubscribe amber" to majordomo_at_scripps.edu