Re: [AMBER-Developers] CUDA Conformance Test not deterministic

From: Scott Le Grand <varelse2005.gmail.com>
Date: Sat, 14 Nov 2020 15:34:31 -0800

Cool I'm setting it to 2 e - 6 instead of 1e - 6

On Sat, Nov 14, 2020, 15:33 Jason Swails <jason.swails.gmail.com> wrote:

> Increasing the tolerance. These are tiny differences compared to most in
> our regression suite.
>
> Thanks,
> Jason
>
> --
> Jason M. Swails
>
> > On Nov 14, 2020, at 5:28 PM, Scott Le Grand <varelse2005.gmail.com>
> wrote:
> >
> > This test appears to be ever so slightly nondeterministic. About 1 in 4
> > runs on TOT is showing this tiny difference. I suspect is because DPFP
> > energy accumulation is not quite deterministic. Which is preferable here?
> > Increasing the tolerance by a factor 2 or ignoring the test?
> >
> > possible FAILURE: check md_SC_NVT_SC_-1.o.dif
> > /media/work/slegrand/amber/test/cuda/gti/SC_Correction/complex
> > 3158c3158
> > < Etot = 3.9278 EKtot = 146.6202 EPtot =
> > 145.6113
> >> Etot = 3.9277 EKtot = 146.6202 EPtot =
> > 145.6113
> > ### Maximum absolute error in matching lines = 1.00e-04 at line 3158
> field 3
> > ### Maximum relative error in matching lines = 2.55e-05 at line 3158
> field 3
> > _______________________________________________
> > AMBER-Developers mailing list
> > AMBER-Developers.ambermd.org
> > http://lists.ambermd.org/mailman/listinfo/amber-developers
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Sat Nov 14 2020 - 16:00:02 PST
Custom Search