Floating point exceptions when using -fpe0

[nichannah]

When the model is built with -fpe0 (intel) or ffpe-trap=invalid,zero,overflow (gnu) a couple of FPEs are triggered. This issue will fix those.

In addition, should we consider making these options standard? I can't find any mention of a down-side on intel/amd procs.

[jwdGFDL]

In general, setting FPE exception handling will slow performance. We
haven't tested it recently. But historically the performance loss is
quite signficant.

On 07/09/2015 10:23 AM, Nic Hannah wrote:

When the model is built with -fpe0 (intel) or
ffpe-trap=invalid,zero,overflow (gnu) a couple of FPEs are triggered.
This issue will fix those.

In addition, should we consider making these options standard? I can't
find any mention of a down-side on intel/amd procs.

—
Reply to this email directly or view it on GitHub
NOAA-GFDL#194.

Jeff Durachta
Engineering Lead for Modeling Services
NOAA Geophysical Fluid Dynamics Lab
Forrestal Campus, Princeton University
201 Forrestal Road
Princeton, NJ 08540
Office: +1-609-987-5054

[adcroft]

I think we could (should?) use these copmiler options in the DEBUG=1 mode. Right now, the FFLAGS_DEBUG in ncrc-gnu.mk just turns off optimization and turns on bounds checking. I agree with @mom6jwd that we would not worry about catching these exceptions for production executables.

[nichannah]

I have created an mkmf issue NOAA-GFDL/mkmf#6 to add -fpe0 for debug builds.