Update Frontier machine/compilers following system update

[grnydawn]

• Retain current software modules instead of updating to the latest versions to prioritize reliability.
• Add linker options to use GCC 12.2, addressing linker errors.
• Utilize Fortran linker to resolve additional linker errors.
• Replace hipcc with mpicxx for MPICXX macro in the GPU compiler definitions.
• Adjust compiler priority to prioritize reliability over performance.
• Temporarily comment out ADIOS2 configurations

[github-actions]

PR Preview Action v1.4.7
🚀 Deployed preview to https://E3SM-Project.github.io/E3SM/pr-preview/pr-6579/
on branch gh-pages at 2024-09-05 18:59 UTC

[rljacob]

How was this tested?

[grnydawn]

The following table summarizes the test results. Test results are from running the e3sm_developer test suite without debug cases. It took an excessive amount of time to build debug cases with the crayclang compiler.

Test Result	crayclang (Current)	crayclang (Latest)	amdclang (Current)	amdclang (Latest)	gnu (Current)	gnu (Latest)
PASS	63	11	58	44	65	71
FAIL	9	60	14	38	7	12
Total	72	71	72	82	72	83
Avg. Build Time (secs)	1137	1619	289	293	133	123

[grnydawn]

The crayclang compiler (both the current and latest versions) has internal compiler issues(segfault from optcg compiler internal module) and excessive compile times : OLCF tickets for the latest version: OLCFHELP-19210, OLCFHELP-19356, and OLCFHELP-19435.

The amdclang compiler has a segfault issue during the compilation of some test cases. Essentially, AMD has asked us to wait until they release a new Fortran compiler based on LLVM.

[rljacob]

Try testing with this suite: e3sm_gpucxx

[grnydawn]

@rljacob , I got following test result with e3sm_gpucxx test suite.

	amdclang	amdclanggpu	crayclang	crayclanggpu	gnu	gnugpu
SCREAM	FAIL MODEL_BUILD	FAIL SHAREDLIB_BUILD	FAIL MODEL_BUILD	FAIL MODEL_BUILD	FAIL MODEL_BUILD	FAIL MODEL_BUILD
MMF	PASS	PASS	PEND MODEL_BUILD	PEND MODEL_BUILD	PASS	PASS

All scream failures except amdclanggpu compiler case are caused by the following error:

-- Configuring incomplete, errors occurred!
...
CMake Error at cmake/build_eamxx.cmake:34 (include):
include could not find requested file:
No macro file found: .../cmake_macros/frontier.cmake in build_eamxx.cmake

[grnydawn]

The following cmake statements in "$E3SM/components/cmake/build_eamxx.cmake" caused the error:

set(SCREAM_MACH_FILE_ROOT ${CMAKE_SOURCE_DIR}/eamxx/cmake/machine-files)
if (EXISTS ${SCREAM_MACH_FILE_ROOT}/${MACH}-${COMPILER}.cmake)
    include(${SCREAM_MACH_FILE_ROOT}/${MACH}-${COMPILER}.cmake)
else()
    include(${SCREAM_MACH_FILE_ROOT}/${MACH}.cmake)

The only cmake file that Scream currently supports for Frontier is "frontier-scream-gpu.cmake". So, none of the above machine/compiler names worked.

[jgfouca]

@grnydawn , eamxx won't work if there's no eamxx machine file. Eamxx clears all E3SM settings and manages its own flags, so you will end up with no compiler flags if there's no eamxx machine file.

[jgfouca]

Eamxx is currently using it's own machine files for frontier, so you probably don't need to worry about this for now.

[rljacob]

What stops EAMxx from using the regular E3SM build system components?

[jgfouca]

@rljacob , a long time ago, we thought eamxx would need to manage it's own flags, even when part of a CIME build. I think we have since decided that we should just use the CIME system but I haven't gotten around to changing things yet.

[grnydawn]

If there are no objections, I will merge this PR to the next branch, and, if no new issues arise, into the master branch.