Workshop: LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC
Authors: Kevin Sala (Lawrence Livermore National Laboratory); Rahulkumar Gayatri (Lawrence Berkeley National Laboratory); Krishna Chaitanya Sankisa, Krzysztof Parzyszek, and Michael Klemm (Advanced Micro Devices (AMD)); and Johannes Doerfert (Lawrence Livermore National Laboratory)
Abstract: The OpenMP API has steadily evolved into a comprehensive, highly portable alternative to vendor-specific APIs for GPU acceleration. The upcoming OpenMP 6.1 specification introduces two key features for code acceleration: access to dynamic shared memory and multidimensional, kernel-like programming. In this paper, we provide an overview of these newly standardized features for C++ and Fortran codebases. We implement a functional prototype based on the official LLVM infrastructure, extending both the Clang and Flang compilers to support these additions. We evaluate our implementation across several C++ and Fortran benchmarks, confirming that these new OpenMP capabilities effectively help close the performance and expressiveness gap with vendor-specific APIs. During the evaluation, we also observe a few remaining opportunities where OpenMP could become even more competitive for specific computational patterns.
Back to LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC Archive Listing
Back to Full Workshop Archive Listing