LBPM icon indicating copy to clipboard operation
LBPM copied to clipboard

GPU support for Rocm

Open hpcresearch-ht opened this issue 5 years ago • 2 comments

Hi I would like to run LBPM on AMD-GPU but there is no GPU Aware MPI present for ROCm yet As you have mentioned in your wiki "For GPU support, it is necessary to have CUDA along with a GPU-aware MPI implementation. Otherwise, the LBPM routines should behave identically irrespective of the underlying hardware. " Is it possible for me to run it on AMD-GPU after hipification and changing the configurations to hipcc

Thanks in advance

hpcresearch-ht avatar Feb 10 '20 10:02 hpcresearch-ht

Hi,

We have explored this a bit and the cuda -> hip translation process is fairly straightforward. The only sticking point is that LBPM uses some cuda 10 features to perform parallel reductions on the GPU, and I do not believe these are supported by HIP right now. Otherwise it is a fairly straight shot.

It is reasonable to anticipate that LBPM will be compatible with both AMD and NVIDIA GPU in the longer term. This is something that we are actively developing as a part of the Frontier CAAR program.

JamesEMcClure avatar Feb 12 '20 15:02 JamesEMcClure

Thanks for the response @JamesEMcClure Yeah there are two fucntions that are not defined by HIP yet sycl_shift_down and this_thread_group() I have tried and ported and in 33 testcases 3 testcases were not running and i am getting segmentation fault TestForceD3Q19 TestBubbleDFH TestFluxBC I beilieve they are failing because of GPU Aware MPI which is causing segmentation fault

Thanks

hpcresearch-ht avatar Feb 21 '20 05:02 hpcresearch-ht