fix definition of fosscuda/2019b to make sure it works with hierarchical MNS

[akesandgren]

(created using eb --new-pr)

edit (@boegel): just to clarify: CUDA should not be installed in Core/ (with system compiler), but using compiler-only toolchain like GCC/iccfort, to ensure that the CUDA module correctly extends $MODULEPATH (not using GCCcore since that would result in always extending $MODULEPATH with path to Compiler/GCC-CUDA/ modules)

[bartoldeman]

@akesandgren one question that puzzles me: how do you distinguish the GCCcore CUDA from CUDA in the intelcuda toolchain in a HMNS. They both have the name CUDA and are both visible via module avail when iccifort is loaded right?

[akesandgren]

The intelcuda (or rather iccifortcuda) would also use CUDA at GCCcore level if i get this right. So it would be the same CUDA. At least that's the plan.

[bartoldeman]

I need to test this properly with the default HMNS but just wondering, how does the GCCcore CUDA extend MODULEPATH?

[akesandgren]

By adding Compiler/GCC-CUDA/version to modulepath. And works as expected now that GCC is just 8.3.0, i.e. the same as GCCcore.

[bartoldeman]

I am still puzzled where in intelcuda Compiler/intel-CUDA/version is added to MODULEPATH then, if they share the same CUDA.

At Compute Canada we have an alternative solution where CUDA is split in CUDAcore (system toolchain) and CUDA at compiler level (GCC, iccifort, PGI). Then you can design a gcccorecuda toolchain to combine GCCcore with CUDAcore, or even a CUDAcore toolchain with just CUDAcore and the system compiler, see https://github.com/ComputeCanada/easybuild-computecanada-config/tree/master/toolchains

[akesandgren]

@bartoldeman You're quite right, CUDA att GCCcore level for use in fosscuda/intelcuda does not work.
If one tries that there will be only one module and it will expand MODULEPATH to GCC-CUDA....

@boegel What do you think about doing it the CC way? with CUDAcore at system and dummies at Compiler (GCC, iccifort, PGI) level

[boegel]

@boegel What do you think about doing it the CC way? with CUDAcore at system and dummies at Compiler (GCC, iccifort, PGI) level

As discussed in Slack, I'm not really in favor of introducing CUDAcore just avoid duplicate CUDA installations in different places in the hierarchy.

Those installations are fairly small, and there's no real problem in having them duplicated (it's just a bit "dirty").

Introducing CUDAcore could cause confusion among end users (even if we hide it from plain sight using hidden = True).

[akesandgren]

Test report by @akesandgren
SUCCESS
Build succeeded for 4 out of 4 (4 easyconfigs in this PR)
b-an03.hpc2n.umu.se - Linux ubuntu 16.04, Intel(R) Xeon(R) CPU E5-2690 v4 @ 2.60GHz, Python 2.7.12
See https://gist.github.com/e599f9a28fd72391508979df65ff59d6 for a full test report.

[boegel]

Test report by @boegel
SUCCESS
Build succeeded for 4 out of 4 (4 easyconfigs in this PR)
node3151.skitty.os - Linux centos linux 7.7.1908, Intel(R) Xeon(R) Gold 6140 CPU @ 2.30GHz, Python 3.6.8
See https://gist.github.com/2819eab2611ba6e6672d2f1153713de3 for a full test report.

[boegel]

Test report by @boegel
SUCCESS
Build succeeded for 4 out of 4 (4 easyconfigs in this PR)
generoso - Linux centos linux 7.6.1810, Intel(R) Xeon(R) CPU E5-2660 v4 @ 2.00GHz, Python 3.6.8
See https://gist.github.com/80f2576a95a3472cc1d07c4703992b3f for a full test report.

remark addressed

[boegel]

Going in, thanks @akesandgren!