Workshop: 2026 International Workshop on Performance, Portability, and Productivity in HPC
Authors: Rintaro Togo, Keichi Takahashi, Tomonori Hayami, Kota Sakakura, and Takashi Soga (The University of Osaka); Arata Endo (Nara Institute of Science and Technology); and Susumu Date (The University of Osaka)
Abstract: The architecture of computing infrastructures is becoming increasingly diverse, making performance portability critical. In the Julia language, JACC (Julia for Accelerators) provides performance portability across CPUs and GPUs. However, JACC was not designed with long-vector architectures in mind, which effectively hide memory latency and deliver high sustained performance in memory-bound applications. In this paper, we extend the performance portability of Julia to long-vector architectures by developing an execution backend for the NEC Vector Engine within JACC. We compare JACC applications against native implementations on three GPU models and two VE models, with little or no architecture-specific code depending on the benchmark. Our results show that the relative performance exceeds 90% on all architectures in most cases, and demonstrate that our backend extends the performance portability of JACC from CPUs and GPUs to long-vector architectures.
Back to 2026 International Workshop on Performance, Portability, and Productivity in HPC Archive Listing
Back to Full Workshop Archive Listing