Workshops Archive

JForce: Fortran on AI Hardware via OpenMP Offloading and XLA JIT Compilation


Workshop: Thirteenth Workshop on Accelerator Programming and Directives (WACCPD 2026)

Authors: Muyao Xiao (Institute of Science Tokyo), Jens Domke (RIKEN CCS), William S. Moses (University of Illinois at Urbana-Champaign), and Ivan R. Ivanov and Toshio Endo (Institute of Science Tokyo)

Abstract: Modern accelerator increasingly rely on tensor-oriented execution units and compiler stacks originally developed for AI workloads, yet Fortran HPC applications cannot readily exploit them through existing toolchains. The OpenMP \verb|workdistribute| construct defines a restricted class of Fortran array computations whose semantics closely resemble tensor computation graphs. However, executing such regions on domain-specific accelerators still requires bridging the semantic gap between Fortran/OpenMP and tensor compiler infrastructures. We present JForce, a Just-In-Time compilation and execution framework that intercepts the high-level MLIR representation of an OpenMP region, translates it into StableHLO, and compiles it using XLA. JForce delegates tensor computation to XLA while retaining the standard OpenMP runtime for host--device data mapping and integration with the remaining OpenMP code. We evaluate JForce on CPUs, NVIDIA and AMD GPUs and Google TPUs, demonstrating portable execution across heterogeneous backends and competitive performance for sufficiently large or repeatedly executed workloads.


Back to Thirteenth Workshop on Accelerator Programming and Directives (WACCPD 2026) Archive Listing

Back to Full Workshop Archive Listing