add -ffpe-summary=none to stand-alone MPAS Makefile for gnu

[mark-petersen]

Currently, gnu-compiled executables produce a summary table of exceptions at the end, which is some version of:

Note: The following floating-point exceptions are signalling: IEEE_DENORMAL
Note: The following floating-point exceptions are signalling: IEEE_INVALID_FLAG IEEE_UNDERFLOW_FLAG IEEE_DENORMAL
Note: The following floating-point exceptions are signalling: IEEE_INVALID_FLAG IEEE_DENORMAL

and prints one line per processor on finalize for both debug and optimized runs. When running on hundreds of cores, this is a headache.

Looking at on-line material 1 2 3 other modeling groups see this with gnu 5+ as well, and say to ignore it or add -ffpe-summary=none. See gnu compile flags.

I have had numerous students and post-docs tell me they had errors when they see this. The error messages are confusing and not useful, because there is no back-trace. Note that actual IEEE errors are caught during debug with -ffpe-trap=invalid,zero,overflow and include a back-trace. The -ffpe-trap is the normal way we catch the errors.

This affects MPAS-seaice and MALI as well. This Makefile is not used for E3SM compiling.

[BFB]

[mark-petersen]

Tested pr regression suite on grizzly with make gfortran and cori with make gnu-nersc. Passes pr suite for both debug and optimized on both. Ran some QU240 and baroclinic channel forward runs with before/after executables, and could see that summary table of IEEE exceptions summary was removed.

[ndkeen]

Ah. We might want to use this flag in other places as well. Thanks Mark.

[mark-petersen]

In this process, I found that gnu-nersc is not using -ffpe-trap=invalid,zero,overflow like the other gnu compilers in the Makefile. I have to think that was an oversight. I added an intentional divide by zero, and indeed the current settings don't produce an error, but with this last commit, we get a floating-point exception and trace-back, as expected.

[philipwjones]

Can you also please add the flag to the gfortran-clang target?

I built/ran a version and confirm that it ran successfully and did not generate the usual underflow messages. They were kinda annoying so thanks @mark-petersen

[mark-petersen]

Added -fbounds-check -fbacktrace to gnu-nersc debug flags, so it matches gfortran. I tested on cori, including with and without an intentional divide by zero, and it all works as expected. Also added -ffpe-summary=none to gfortran-clang per @philipwjones's request.

@matthewhoffman and @akturner if you could run a standard test on MALI and MPAS-seaice with this Makefile, I think that is enough to test this PR.

[mark-petersen]

@xylar I think additional testing from you is not required. You can read my testing above. You've been involved with Makefile changes in the past, so wanted you to see this PR.

[xylar]

This looks good to me and passed the nightly regression suite on my laptop with gnu compilers. I see:

Note: The following floating-point exceptions are signalling: IEEE_DENORMAL
Note: The following floating-point exceptions are signalling: IEEE_UNDERFLOW_FLAG IEEE_DENORMAL

without these changes but not with them. Thanks for this work!

[matthewhoffman]

I tested on Cori with gnu using MALI, and ran into no problems. It will be nice to eliminate those confusing output messages!

[akturner]

Passed MPAS-Seaice testing

[jonbob]

passes sanity testing, merged to next

[jonbob]

merged to master