hipSYCL Gets New Compilation Toolchain For Taking SYCL Directly To CUDA & ROCm
Intel is making very interesting moves and among them are working on upstreaming SYCL support in LLVM. It appears this SYCL play of theirs will be a key part of their "oneAPI" effort that is supposed to be out by year's end.
Separately, others in the ecosystem are making strides around SYCL. The hipSYCL open-source project that has been working on implementing SYCL over NVIDIA's CUDA and AMD HIP/ROCm scored another major achievement.
The hipSYCL project has introduced a new compilation toolchain for compiling SYCL code directly for CUDA and ROCm by making use of the LLVM Clang CUDA / HIP front-end. Previously hipSYCL has relied upon source-to-source transformations.
Big news: We now have a new compilation toolchain in #hipSYCL based on a clang plugin. This allows us to compile #SYCL code *directly* for #CUDA and #ROCm with #clang's CUDA/HIP frontend, without source-to-source transformation! Faster compilation and more robust! pic.twitter.com/TcCGApPiY4— Aksel Alpay (@illuhad) May 12, 2019
With this Clang-based plug-in they are now able to go directly into Clang / LLVM IR and bypassing the source transformations and thus should allow for more optimizations and other improvements moving forward.
Those wanting to learn more about hipSYCL itself can visit GitHub.