Workshops Archive

Code Generation with MLIR for the OPS Structured-Mesh DSL


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

Authors: Archie Dascombe, Prakanth Thilakaraj, and Gihan R. Mudalige (University of Warwick)

Abstract: Structured-mesh applications underpin high-performance scientific simulations, yet performance portability across CPUs and GPUs demands substantial compiler and runtime support. We present an MLIR-based compiler framework for OPS, a structured-mesh DSL, that captures OPS abstractions at runtime and JIT-lowers them to OpenMP and CUDA backends. A key contribution is adapting the predominantly value-semantic xDSL stencil compilation stack to OPS’s buffer-semantic model, where pre-allocated datasets have memory identity and update semantics. This demonstrates that shared MLIR stencil infrastructure can accommodate lower-level, production-oriented DSLs without redesigning their programming model. The framework introduces an OPS dialect and staged lowerings for parallel loops, data accesses, stencils, and execution policies. Evaluation of representative structured-mesh applications, including the OpenSBLI research code on multicore CPUs and an NVIDIA H100 GPU, shows performance competitive with OPS’s mature source-to-source backends, substantial memory-bandwidth utilization, closely matched GPU kernel times, and amortized JIT overheads, providing a practical path for modernizing DSL compilers.


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

Back to Full Workshop Archive Listing