Workshops Archive

GPU Code Generation for OpenACC Kernels construct in a Modern MLIR-Based Compiler


Workshop: LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC

Authors: Ivan R. Ivanov (NVIDIA)

Abstract: The OpenACC kernels construct asks the compiler to discover and exploit parallelism inside a region that may mix parallel and sequential loops, intervening sequential code, reductions, and control flow that can guard those loops. Mapping such regions to GPUs is harder than compiling the more prescriptive parallel construct: the compiler must decide where synchronization and launch boundaries fall, not only how to schedule a single nest. We describe an MLIR-based Fortran OpenACC pipeline with three strategies for kernels: (1) splitting kernels into sequenced host-side device launches; (2) a single cooperative grid with grid-scoped barriers; and (3) CUDA Dynamic Parallelism (CDP), launching child grids from a device driver kernel. The strategies share most of the parallelism-lowering stack and diverge on how nests are ordered, synchronized, and bridged. We detail their CUDA mechanisms and mappings, and evaluate them on proxy applications and a synthetic suite that isolates the axes driving strategy choice.


Back to LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC Archive Listing

Back to Full Workshop Archive Listing