We have been working to bring multi-architecture support using SYCL to the LLVM Clang project. Our original approach was to implement a "Plugin Interface" to add support for a PTX back-end and subsequently we have also added support for GCN enabling NVIDIA and AMD GPUs. This short presentation will outline our approach to designing this multi-architecture back-end and recent work to formalise the interface in the SYCL specification. This work is enabling researchers using the pre-exascale Perlmutter and Polaris supercomputers and exascale Summit supercomputer to write code using open standard SYCL and deploy on these machines.
Speakers: Hugh Delaney