Exploring a multi-resolution GPU programming model for Chapel

There is a growing need to support accelerators, especially GPU accelerators, since they are a common source of performance improvement in HPC clusters. As for GPU programming with Chapel, typically programmers first start with writing forall loops and run these loops on CPUs as a proof-of-concept. If the resulting CPU performance is not sufficient for their needs, their next step could be to try the automatic compiler-based GPU code generation techniques [1], [2]. For portions that remain as performance bottlenecks, even after automatic compilation approaches, the next step is to consider writing GPU kernels using CUDA/HIP/OpenCL and invoking these kernels from the Chapel program using the GPUIterator [3], [4] and Chapel's C interoperability feature.