Fixes for cray-mpich+cuda on daint #51
Conversation
gtl is needed even at runtime
cuda and rocm are needed at runtime
libstdc++.so is missing on daint but present on beverin
|
How does making CUDA a runtime dependency instead of link time help with detection of MPI? Is there an example of why this matters (I imagine that there are some applications that want to find |
|
The problem is not MPI per say but the fact that the mpi detection fails because the cuda runtime library is not found (related to the gtl library). I might be inclined to close my own PR because it is not really a good solution. |
|
I see: the edit: Thinking about this some more, the error might be elsewhere, because we link CUDA against GTL using RPATH, so it should be found. Can you provide the compilation command and it's output, for me to look into further. |
In the migration to spack 1.0 with #30 we inadvertently switch from using spack compiler wrappers (with all their support for correctly setting build flags) to using the "plain" regular compiler. This ended up with build problems (see spack/spack-packages#1193) because not using the spack compiler wrapper means not using `SPACK_STORE_*` variables that contains information about include/libs of all dependencies. This was more damaging for some build systems than others, but it might be that this triggered also subtle problems (possibly #51?).
|
@mtaillefumier : could you please check the latest version of #57 introduced a key fix for |
2 participants
The detection of
mpican fail whencray-mpichis installed withcudaenabled.