better support for mpicc, mpicxx, and mpifort

[leofang]

Currently the MPI wrapper compilers mpicc, mpicxx, and mpifort are not guaranteed to work out of box unless the corresponding host compilers (gcc_linux-64, gxx_linux-64, and gfortran_linux-64 in Linux) are also installed in the Conda environment, see conda-forge/openmpi-feedstock#34 (comment):

The openmpi package doesn't depend on the compilers because it's most often installed as a runtime dependency, not as a compiler. To use the mpi compiler wrappers and the compilers they wrap, the conda compilers must be installed. # 35 adds openmpi-mpicc etc. as new outputs so that users have a simple package name to install to pull in both dependencies needed to compiler with openmpi (openmpi and the relevant compiler)

A solution just merged to conda-forge feedstocks yesterday (conda-forge/mpich-feedstock/pull/32
and ongoing conda-forge/openmpi-feedstock/pull/35) was to add additional outputs in the MPI recipes, including mpich-mpicc, mpich-mpicxx, mpich-mpifort, and their openmpi counterparts.

Let us wait for a while and see if it is a sustainable solution before adapting it. In the meanwhile one can always conda install the host compilers to make the MPI wrapper compilers work.

[leofang]

more context: conda-forge/mpich-feedstock/issues/25

[mrakitin]

Good to know. Thanks for the links. Agree to wait a bit and see how good this solution is.

[minrk]

I'm not sure which packages you are using these in, but FWIW if you have a package that requires MPI as a compiler, you can depend on both mpi and the given compiler as runtime requirements, e.g.:

requirements:
  run:
    - mpich
    - {{ compiler('c') }}

This is the ultimately the same as the new mpich-mpicc package, but more explicit if it's an actual requirement. The mpich-mpicc package and friends are meant really only for users, where the {{ compiler('c') }} template variable isn't available. I would use this pattern in recipes, regardless of the presence of the new convenience packages.

[leofang]

Thanks @minrk for the comment. This is indeed better for packages requiring an MPI compiler. This issue was opened mainly to document my observation that using MPI as a compiler, as you put it, requires a bit of change. However, I suspect most of our users don't need any MPI compiler.