netcvode test tolerances may be overly generous to accomodate NVIDIA compilers

[olupton]

Context

The test introduced in #1752 did not pass when the NVIDIA HPC compilers nvc and nvc++ were used. #1822 was introduced to make the tests pass by tolerating larger differences.

Overview of the issue

Some tolerances, in particular this 25% margin:

nrn/test/cover/test_netcvode.py

Lines 569 to 571 in c127b54

	# Note the very large tolerance, this seems relatively unstable with the
	# NVIDIA compilers
	chk(key + " tvec size", tvec.size(), tol=0.25)

seem worthy of further investigation. Differences between compilers are to be expected at some level, but this seems high.

See the discussion on the follow-up PR here: #1822 (comment).

Expected result/behaviour

The results obtained with the NVIDIA compilers should be compatible with those obtained with GCC/Intel/Clang with a margin that is plausibly related to optimisations (e.g. FMA and enabled fast-math optimisations).

Assuming the difference is understood and fixed at source, the test tolerance should be reduced again.

NEURON setup

• Version: master
• Installation method: CMake build
• OS + Version: BB5
• Compiler + Version: NVHPC 22.3

Minimal working example - MWE

Modify the tolerance linked above to be smaller again, run the relevant test and investigate its failure.

[nrnhines]

I did not succeed in installing a working nvcc/nvc++ on my desktop (cuda refused to install). What is the name of the BB5 these days. Last time I logged in it was [email protected]. I guess I'm out of date with:

Hostname: bbpv1.epfl.ch
User: hines
ModuleCmd_Load.c(213):ERROR:105: Unable to locate a modulefile for 'hpe-mpi'
ModuleCmd_Load.c(213):ERROR:105: Unable to locate a modulefile for 'gcc/6.4.0'
ModuleCmd_Load.c(213):ERROR:105: Unable to locate a modulefile for 'python/3.6.5

Do I need a list of module loads? And what is a reasonable cmake to look into this issue? If I've entered a minefield, I'm happy to step back.