Consider moving R to GCC level, splitting MPI extensions into a separate bundle

[jfgrimm]

Unlike Python, Perl, etc., R sits at foss level due to requiring BLAS/LAPACK and MPI for various extensions. This pulls any easyconfigs depending on R all the way to foss, even if (and this often seems to be the case) they are not MPI-aware.
I would propose moving R itself to GCC (for flexiblas), and splitting any extensions that need to be at foss level into a separate bundle, similar to Python/SciPy-bundle

[branfosj]

MPI is definitely needed for Rmpi, but I am not sure much else beyond that and there are not a lot of R packages that depend on Rmpi (as far as I know).

No package in R-4.1.2-foss-2021b.eb or R-bundle-Bioconductor-3.14-foss-2021b-R-4.1.2.eb requires Rmpi.

[jfgrimm]

had a look at packages depending on MPI (directly or indirectly) in the current R 4.1.2 easyconfig:

R package	dependency	library linked
Rssa	FFTW	FFTW/3.3.10-gompi-2021b/lib/libfftw3.so.3
hdf5r	HDF5	HDF5/1.12.1-gompi-2021b/lib/libhdf5_hl.so.200
Rmpi	OpenMPI	OpenMPI/4.1.1-GCC-11.2.0/lib/libmpi.so.40
sf	netCDF	netCDF/4.8.1-gompi-2021b/lib/libnetcdf.so.19
terra	netCDF	netCDF/4.8.1-gompi-2021b/lib/libnetcdf.so.19
fftw	FFTW	FFTW/3.3.10-gompi-2021b/lib/libfftw3.so.3
fftwtools	FFTW	FFTW/3.3.10-gompi-2021b/lib/libfftw3.so.3
imager	FFTW	FFTW/3.3.10-gompi-2021b/lib/libfftw3.so.3
rgdal	netCDF	netCDF/4.8.1-gompi-2021b/lib/libnetcdf.so.19
ncdf4	netCDF	netCDF/4.8.1-gompi-2021b/lib/libnetcdf.so.19

[bartoldeman]

At Compute Canada we compile R with the gcccoreflexiblas toolchain:
https://github.com/ComputeCanada/easybuild-computecanada-config/blob/main/toolchains/gcccoreflexiblas.py
but this is a bit tricky as BLIS and OpenBLAS are at the GCC level (they have to be since Fortran interfaces differ between GCC and Intel), and having BLIS at both GCCcore and Intel can be confusing with an HMNS.

We also have
https://github.com/ComputeCanada/easybuild-computecanada-config/blob/main/toolchains/gccflexiblas.py

But it may be ok to simply use flexiblas as a dependency instead, and FFTW-serial.
Would need to look up if R can use HDF5/netCDF's MPI interfaces or not, if not it could use the -serial versions of those too.

[Micket]

R now sits at gfbf level, which is as low as it can get in our toolchains, becuase R itself depends on BLAS/LAPACK (it cant ever be moved to compiler level).