Investigate HIP performance difference with spatial indices

LLNL / axom

CS infrastructure components for HPC applications

BSD 3-Clause "New" or "Revised" License

153 stars 27 forks source link

Investigate HIP performance difference with spatial indices #1282

Open bmhan12 opened 7 months ago

bmhan12 commented 7 months ago

PR #1278 records an unexpected performance difference between HIP and CUDA (~4x slower) for the BVH and Implicit Grid spatial indices.

Some potential sources of improvement:

Verify Axom is using HIP intrinsics: https://github.com/LLNL/axom/blob/0a47cb26b8dd5d358cdee6fa0653f326ff98f834/src/axom/core/utilities/BitUtilities.hpp#L21-L36
Try newer rocm versions (>= 5.7.0)
Consider using RAJA::unsafeAtomicAdd (link and link to suggestions)

kennyweiss commented 1 week ago

Presumably, given the updated query results in https://github.com/LLNL/axom/pull/1419, this is still an issue for our hip-based vs. cuda-based ImplicitGrid