More GPU CI - Githubissues

QMCPACK / qmcpack

Main repository for QMCPACK, an open-source production level many-body ab initio Quantum Monte Carlo code for computing the electronic structure of atoms, molecules, and solids with full performance portable GPU support

http://www.qmcpack.org

Other

299 stars 139 forks source link

More GPU CI #3402

Closed ye-luo closed 2 years ago

ye-luo commented 3 years ago

We have self-hosted legacy CUDA CI runs which are quite fast. So consider to add more.

We need

[x] ENABLE_CUDA=ON. It should be low risk as legacy CUDA. Can be GCC + CUDA. cmake -DCMAKE_C_COMPILER=mpicc -DCMAKE_CXX_COMPILER=mpicxx -DBUILD_AFQMC=ON -DENABLE_CUDA=ON -DCUDA_ARCH=sm_xx ..
[x] ENABLE_CUDA=ON ENABLE_OFFLOAD=ON. This requires Clang 13 release at least rc2 cmake -DCMAKE_C_COMPILER=mpicc -DCMAKE_CXX_COMPILER=mpicxx -DBUILD_AFQMC=ON -DENABLE_CUDA=ON -DCUDA_ARCH=sm_xx -DCUDA_HOST_COMPILER=which gcc-DENABLE_OFFLOAD=ON -DOFFLOAD_ARCH=sm_XX -DUSE_OBJECT_TARGET=ON ..
[x] CUDA2HIP cases. Maybe start with build only without running. Add -DQMC_CUDA2HIP=ON to the above and drop CUDA_ARCH

prckent commented 3 years ago

Agreed. We have bandwidth for these.

Please can you put the exact cmake build lines and define "latest Clang". The most recent release? Develop?

prckent commented 3 years ago

For point 2 I think we should just wait to the LLVM13 release, currently alleged for 21 September. @williamfgc Is there an easy way to get and use the RC versions? Preferences?

ye-luo commented 3 years ago

@williamfgc can we start to add the above CI variants. the first one only needs gcc. The second one probably we can just use llvm 12.0.1 for the moment.

williamfgc commented 3 years ago

@ye-luo @prckent let me look into LLVM13 and installing things on sulfur since we are not using containers for GPU.

ye-luo commented 3 years ago

LLVM13 has been released.

williamfgc commented 3 years ago

I think this issue has been addressed, thanks @ye-luo and @prckent for bringing this up. There is only one lingering thing that requires digging deeper as there are numerical instabilities on potential Clang14Dev-MPI-CUDA-AFQMC-Offload-Complex-Mixed Clang14Dev-MPI-CUDA-AFQMC-Offload-Complex (not present) configurations.