Workshop: LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC
Authors: Benjamin Glick, Subhadeep Bhattacharya, Michael YH Wang, Xiakun Lu, and Pouya Kousha (NVIDIA)
Abstract: GPU applications combine kernels produced by multiple compiler frontends with production device libraries implemented in CUDA C++. Reimplementing libraries per frontend is costly, especially communication libraries whose code depends on host-created transport state. We show linkable device IR can serve as an interoperability boundary for stateful GPU libraries without sacrificing baseline performance. We package \nvshmem{} and \nccl{} implementations as LLVM bitcode and CUDA LTOIR and consume them from CUDA C++, CUTLASS Python, and numba-cuda-mlir. Across RMA benchmarks, packaged IR preserves latency and bandwidth relative to conventional CUDA C++ linkage, reaching 98~GB/s. Across both libraries, CUTLASS Python stays within 3.7\% of CUDA C++ latency and delivers at least 90.7\% of its bandwidth, reaching \nccl{} bandwidth parity. Fused GEMM+AllGather experiments show packaged communication IR composes with tensor-core computation, while LTOIR remains within 0.57\% of baseline across every tested \nccl{} shape. Results establish linkable device IR as a performant multi-frontend library boundary.
Back to LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC Archive Listing
Back to Full Workshop Archive Listing