-
Notifications
You must be signed in to change notification settings - Fork 88
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Building Ginkgo on Perlmutter #1367
Comments
Thanks for reporting this issue.
It should solve the problem. |
I'd like to get to the bottom of this though - we've had some workaround in place for nvhpc, but it seems like Perlmutter's setup deviates from the ways we've seen nvhpc installed so far. IMO this is something that needs to be fixed on the CMake level, we can't keep adding system-specific workarounds. |
This fixed my issue with CUDA! Thank you so much. Separately, when the compilation was nearly complete, it seems like ginkgo failed to detect MPI. I have cray-mpich/8.1.25 loaded, so I'm not sure why it's an issue. I tried adding this location to CPATH and LIBRARY_PATH and encountered the same issue. From the setup, it seems like MPI was detected though, but in a different version than what I have loaded.
|
@MrQuell From the log in your first post, Ginkgo has correctly detected MPI. The issue you see is only in one of our examples. If you don't use this example, and just use our library, then you can ignore the issue, while we work on patching this. |
It seems like something might be going wrong in CMake's MPI detection - the include path should be detected. Could you attach your CMakeCache.txt or look up the values of MPI_*_INCLUDE_DIRS in there? For a full deep-dive into what is going on, we could also use the output of |
Here are the instances of looking at MPI_*_INCLUDE_DIRS in CMakeCache.txt.
I am interested in running some examples, but I can complete the build by excluding the examples. I can run the debug cmake later and paste over its output. |
A minimal reproducer using |
@MrQuell can you try a fresh build (including the CUDA environment variable fixes) where you add |
For context, here the corresponding CMake issue: https://gitlab.kitware.com/cmake/cmake/-/issues/25093 |
It does seem like this cmake issue was resolved. However, I still get the error that mpi.h is not detected. I am up to date with github and I did all the suggested commands. Perhaps cmake is not up to date on perlmutter with the latest commit from the issue you raised. export CPATH="${CUDATOOLKIT_HOME}/../../math_libs/include:$CPATH" |
Just to make sure, did you use |
I appologize for the confusion. I did use -DGINKGO_BUILD_CUDA=ON. I must have copied an earlier command from my history that failed. When I run that command now, at the bottom, I get the following
|
Got it, thanks. Now #1368 should fix all issues in one go, can you give it a try? I also checked on Perlmutter with your modules enabled, the only thing I needed to load on top was |
I'm a new user of the Ginkgo Project and I'm attempting to build Ginkgo on Perlmutter, a NERSC cluster. Everything in cmake seems to pass and it's detecting CUDA, but it cannot recognize certain modules in cuda, such as cublas, cusparse, curand, and cufft. Upon further investigation, it seems like the cuda folder and math_libs folder have different functions in them.
The methods of cublas, cusparse, curand, and cufft are all in math_libs, not in cuda. How can I get Ginkgo to detect the methods in that folder?
Here are the modules I have loaded
Here is the print out of what happens when I try to build Ginkgo
The text was updated successfully, but these errors were encountered: