Consider supporting HIP-over-CUDA

Some folks are interested in running HIP codes on NVIDIA GPU platforms like summit, which now has a system-provided rocm-hip module that provides a HIP-over-CUDA shim. This is not as crazy as it sounds; due to the deliberate similarity between the libraries, the shim is extremely lightweight (in some cases header-only), and delivers comparable performance. The benefit is the ability to integrate libraries written specifically for HIP and still run them on an NVIDIA platform.

UPC++ has a currently undocumented/unsupported configure -with-hip-platform=nvidia mode that allows it to run using the HIP-over-CUDA shim. This was deployed mostly to assist in correctness testing by maintainers, but thanks to GASNet is robust enough to provide ROCmRDMA-over-GDR on supported networks, so is not necessarily a terrible way to run.

We should consider whether and how to document support for such configurations.

Comments (4)