A HIP port of the TeaLeaf mini-app starting from the current CUDA version on UK-MAC. This port used hipify-perl (https://github.com/ROCm-Developer-Tools/HIPIFY) to convert the CUDA kernels. Extra work was required for CUDA specific code (references to specific warpsizes relevant to Nvidia hardware etc.).
In most cases one needs to load the appropriate modules (including those for the accelerator in question).
Then for AMD GPUs (with the Cray compiler)
make clean; make COMPILER=CRAY
Turn on HIP kernel use insert use_cuda_kernels in tea.in.
Observed performance uplift compared to directive based versions (OpenMP, OpenACC) is anywhere from 30-40% depending on node count for a sufficiently large enough benchmark to make using a GPU sensible. This is consistent with performance observed for the CUDA version of CloverLeaf.