clean up fugaku files and gromacs builds with gcc now with minor spac… #161
Conversation
…k modifications to fujitsu-mpi package
| @@ -128,9 +53,19 @@ compilers: | |||
| cxx: /vol0004/apps/oss/spack-v0.19/opt/spack/linux-rhel8-a64fx/gcc-8.5.0/gcc-12.2.0-sxcx7kmt3qiktffgzzvrj2wmup3g32bc/bin/g++ | |||
| f77: /vol0004/apps/oss/spack-v0.19/opt/spack/linux-rhel8-a64fx/gcc-8.5.0/gcc-12.2.0-sxcx7kmt3qiktffgzzvrj2wmup3g32bc/bin/gfortran | |||
| fc: /vol0004/apps/oss/spack-v0.19/opt/spack/linux-rhel8-a64fx/gcc-8.5.0/gcc-12.2.0-sxcx7kmt3qiktffgzzvrj2wmup3g32bc/bin/gfortran | |||
| flags: {} | |||
| flags: | |||
| ldflags: {"-lelf -ldl"} | |||
There was a problem hiding this comment.
(potential issue) Des this appear to concretize for you? I'm not familiar with this YAML syntax (it looks like a set of one element), typically one would write:
ldflags: "-lelf -ldl"
Also, would you be able to mention the error you get without this? I don't want to hold this PR up on it, but I may be able to make build suggestions that avoids the need for this.
There was a problem hiding this comment.
linking against fujitsu's blas lib (SSL2) with gcc results in missing symbols, which is why i need those flags
| providers: | ||
| mpi: [fujitsu-mpi, openmpi, mpich] | ||
| blas: [fujitsu-ssl2, openblas] | ||
| lapack: [fujitsu-ssl2, openblas] | ||
| scalapack: [fujitsu-ssl2, netlib-scalapack] | ||
| fftw-api: [fujitsu-fftw, fftw, rist-fftw] | ||
| fftw-api: [fujitsu-ssl2, fftw, rist-fftw] |
There was a problem hiding this comment.
This should not have an effect: fujitsu-ssl2 does not provide fftw-api. Spack can work around this because the expressions here are taken as preferences by spack (i.e. not hard requirements). In other words, this shouldn't cause anything to fail, but I also think it isn't necessary for anything to succeed.
There was a problem hiding this comment.
afaik, fujitsu-ssl2 provides fft functions, but for some reason the fftw.h header is missing. i'm also not yet 100% sure how to set it properly, when i compile (forgot) gromacs or hyper i see that it doesnt find the header and then builds fftw from source (which isn't the best approach, IMHO). i don't think 'fujitsu-fftw' exists, so the previous config is definitely wrong.
There was a problem hiding this comment.
What I mean is that because Spack's builtin fujitsu-ssl2 package.py does not have provides(fftw-api) spack will never assume it is useful as an fftw provider and will always try to use something else (the logs generated by ramble workspace setup should include the concretization, and I assume Spack will have chosen something besides fujitsu-ssl2 wherever an fftw implementation is needed).
There was a problem hiding this comment.
I see, thanks for the clarification
| compiler-fujitsu: | ||
| spack_spec: fj@{default_fj_version} | ||
| compiler-clang: | ||
| spack_spec: clang@{default_llvm_version} |
There was a problem hiding this comment.
(no change needed) This may be required until we invoke Ramble in a way that avoids having it look for additional compilers.
| OMPI_CC: /vol0004/apps/oss/spack-v0.19/opt/spack/linux-rhel8-a64fx/gcc-8.5.0/gcc-12.2.0-sxcx7kmt3qiktffgzzvrj2wmup3g32bc/bin/gcc | ||
| OMPI_CXX: /vol0004/apps/oss/spack-v0.19/opt/spack/linux-rhel8-a64fx/gcc-8.5.0/gcc-12.2.0-sxcx7kmt3qiktffgzzvrj2wmup3g32bc/bin/g++ | ||
| OMPI_FC: /vol0004/apps/oss/spack-v0.19/opt/spack/linux-rhel8-a64fx/gcc-8.5.0/gcc-12.2.0-sxcx7kmt3qiktffgzzvrj2wmup3g32bc/bin/gfortran | ||
| OMPI_F77: /vol0004/apps/oss/spack-v0.19/opt/spack/linux-rhel8-a64fx/gcc-8.5.0/gcc-12.2.0-sxcx7kmt3qiktffgzzvrj2wmup3g32bc/bin/gfortran |
There was a problem hiding this comment.
(potential issue) When we set OMPI_* variables, it is typically to point them at the underlying spack wrappers, it might be preferable to handle this as an update to fujitsu-mpi's package.py file in Spack.
There was a problem hiding this comment.
i understand. i can take a look how to fix fujitsu-mpi with an additional PR. for now its just to work around problems i run into with fugaku.
yes, i was able to build amg2023 and gromacs with gc with those changes. i'll check clang next, and verify that i didnt break anything for fujitsu compiler. |
…k modifications to fujitsu-mpi package