add custom easyblock for NVHPC (aka PGI)

[AndiH]

PGI beyond 20.4 is not called PGI any more but NVIDIA HPC SDK (NVHPC), with 20.7 being the first version; https://developer.nvidia.com/hpc-sdk.

This pull request is part of a set of three pull requests to introduce NVHPC to EasyBuild (1, 2, 3). The stack has been successfully deployed at JSC and is under use by our HPC users.
The EasyBuild files are created after discussions with the NVHPC product owners, taking into account HPC practices and EasyBuild conventions.

This request is for pulling the NVHPC EasyBlock. It is based on the PGI EasyBlock but cleaned-up and extended for NVHPC.

Additional options are included; for example default_cuda_version, compute_capability and several module_* options. The former two are passed to the NVHPC's install script which configures NVHPC for usage with a certain CUDA or device Compute Capability.
The module_* options refer to options being passed to the LMod module where additional directories beyond the default directories are included into the environment. As an example, nvhpc_nvshmem_basedir adds the directory of NVHPC in which NVSHMEM is included into CPATH and LD_LIBRARY_PATH with their corresponding sub-directories. The options exist to enable HPC centres to rather provide their own installations of the packaged libraries.

[boegel]

Please also take care of the code style suggestions.

Don't hesitate to reach out if you need any help!

[AndiH]

(Snapshot at end of work day today. Will continue after additional input. Note: Not tested on system.)

[AndiH]

Some more work done. @boegel can you have a look at the open questions?

[boegel]

Assignment is missing here, this can't work? .replace returns a new string value...

Also, I hate to be pedantic, but you're now just assuming you have a string value in default_compute_capability, which may not be the case (and if it's not, the crash will be pretty ugly).

So I'd add a small check:

if isinstance(default_compute_capability, str):
    default_compute_capability = default_compute_capability.replace('.', '')
else:
    raise EasyBuildError("Unexpected non-string value encountered for compute capability: %s",
                         default_compute_capability)

[AndiH]

Implemented in 2c91709

[boegel]

I would check all commands, and just use the command name (not the path):

compiler_cmds = ['nvc', 'nvc++', 'nvfortran']
custom_paths = {
    'files': [os.path.join(prefix, 'compilers', 'bin', x) for x in compiler_cmds + ['siterc']],
...
custom_commands = ["%s -v" % x for x in compiler_cmds]

[AndiH]

Implemented in c54b719.

[boegel]

@AndiH Please also revisit easybuilders/easybuild-easyconfigs#11391, even though the tests can't pass there as long as this easyblock isn't merged, since that's the test case for this easyblock (we'll merge both in quick succession).

We're getting there, I'm confident we can get this merged really soon now...

[AndiH]

@boegel Absolutely! I wanted to get the EasyBlock done first before I implement the easyconfig. I am working on that next!

[AndiH]

I think I've tackled all the remaining open issues!
I leave resolution of the open comments to you and will now start with the edits for the easyconfig.

[boegel]

@AndiH Tested this using a tweaked version of easybuilders/easybuild-easyconfigs#11391 (only changed the CUDA dep), ran into this crash:

== installing...
ERROR: Traceback (most recent call last):
  File "/user/boegel/easybuild-framework/easybuild/main.py", line 115, in build_and_install_software
    (ec_res['success'], app_log, err) = build_and_install_one(ec, init_env)
  File "/user/boegel/easybuild-framework/easybuild/framework/easyblock.py", line 3301, in build_and_install_one
    result = app.run_all_steps(run_test_cases=run_test_cases)
  File "/user/boegel/easybuild-framework/easybuild/framework/easyblock.py", line 3204, in run_all_steps
    self.run_step(step_name, step_methods)
  File "/user/boegel/easybuild/framework/easyblock.py", line 3059, in run_step
    step_method(self)()
  File "/tmp/eb-30tldro0/included-easyblocks-h6qfu1hb/easybuild/easyblocks/nvhpc.py", line 127, in install_step
    default_compute_capability = _before_default_compute_capability[0]
IndexError: list index out of range

[AndiH]

@boegel Did you specify a Compute Capability (via CLI or via easyconfig)?

(I test some combinations meanwhile!)

[AndiH]

Ok, this is a bit weird, but fixed in e4c6c34.

If cuda_compute_capabilities is not set in the easyconfig file and not via CLI, the value is not None, but rather []; for whatever reason. The added if ec_default_compute_capability checks for empty lists.

[boegel]

@AndiH Getting [] as default value for cuda_compute_capabilities makes sense, see:

$ eb -a | grep cuda_compute_capabilities
cuda_compute_capabilities   List of CUDA compute capabilities to build with (if supported) [default: []]

Thanks for fixing that issue!

I'm running some final tests now, but I think this PR and easybuilders/easybuild-easyconfigs#11391 are good to go (I'll deal with the failing tests in the latter once this PR is merged).

[AndiH]

Great! 🕺

I try to work soon on NVHPC 20.9, which has been released meanwhile!

[boegel]

lgtm, tested with easybuilders/easybuild-easyconfigs#11391

Thanks a lot for the effort and patience @AndiH !