add toolchain definition for Fujitsu ARM toolchain

[migueldiascosta]

being tested at Fugaku

[migueldiascosta]

@boegel see easybuilders/easybuild-easyconfigs#12863 (comment) about the hierarchy and the naming of ffmpi, do you have a better suggestion? (e.g. FujitsuCore? FujitsuCoreMPI?)

basically, this will be the toolchain used for most of what's in both GCCcore and gompi...

[boegel]

I replied in that issue, but to be clear: the Ffmpi class name is just an internal thing, and sort of separate with the actual toolchain name (but it's better to keep those in sync, of course).

[migueldiascosta]

apparently -Kfp_precision doesn't work in clang mode, closest we can get seems to be -Knoeval,nofast_matmul,nofp_contract,nofp_relaxed,noilfunc (and at least the tests that require precise in GMP do pass with these flags)

[migueldiascosta]

-SSL2/-SSL2BLAMP and -SCALAPACK are only meant to be used when linking and generate a warning on compiling that is harmless but pollutes the logs

easybuild prepends -L to variables appended to LDFLAGS, so some refactoring seems to be necessary (?) in order to do this properly (is it worth it?)

[boegel]

@migueldiascosta If you think it's important enough, please open an issue to follow up on this.

[migueldiascosta]

turns out having -Nclang only in these flags messes up with configure scripts that probe for the compiler without them, and so detect the traditional mode instead (hitting this with the libffi included in _ctypes in Python 2.7.18)

one solution that seems to work is to move them to COMPILER_CC and COMPILER_CCX above (but I think then it's crucial to always enclose $CC etc. in double quotes, since they will include a space?)

[migueldiascosta]

same as above, these don't seem to work in clang mode... SVE is set by default, so the alternative seems to be -Knosimd to disable...

[migueldiascosta]

I replied in that issue, but to be clear: the Ffmpi class name is just an internal thing, and sort of separate with the actual toolchain name (but it's better to keep those in sync, of course).

@boegel well, besides the name there is also the hierarchy, the reason I used only two levels is that the module provides compiler, mpi and linalg;

but even though they are there, easybuild doesn't have to enable all at once, e.g. we can have an additional non-mpi level "below" Ffmpi and/or move FujitsuSSL to the full toolchain level, would that be clearer?

[migueldiascosta]

-Kfast is actually -O3 plus a bunch of other stuff, so planning to play it safe and change to:

    DEFAULT_OPT_LEVEL: 'O2',
    'opt': 'Kfast',

In addition to the possibility of using 'opt': True in easyconfigs, -Kfast will be explicitly enabled in some easyblocks, e.g. fftw, berkeleygw (, fortranpythonpackage?)

[migueldiascosta]

but even though they are there, easybuild doesn't have to enable all at once, e.g. we can have an additional non-mpi level "below" Ffmpi and/or move FujitsuSSL to the full toolchain level, would that be clearer?

this does seem clearer (and cleaner), done in migueldiascosta@c095190

(at the time of writing, the commit doesn't show up in this PR, did I mess up or is GitHub acting up? will check back later)

[migueldiascosta]

(weird, commit c095190 only showed up here after I pushed 558487e)

[boegel]

@migueldiascosta I'll assume this is fully functional in the current state (can't check myself, because I don't have access to Fugaku), so I'll go ahead and merge this, mostly because I want to ensure that the changes to easybuild/toolchains/fft/fftw.py don't have any negative side effects as we're ramping up to the extensive regression test for EasyBuild v4.4.0.

Follow-up PRs to fix specific issues are welcome, of course! :)