Workshop: Thirteenth Workshop on Accelerator Programming and Directives (WACCPD 2026)
Authors: Koki Morita and Tetsuya Hoshino (Nagoya University), Ryosuke Ando (The University of Tokyo), Akihiro Ida (Japan Agency for Marine-Earth Science and Technology), and Daichi Mukunoki and Takahiro Katagiri (Nagoya University)
Abstract: The Fast Domain Partitioning Method (FDPM) reduces the time and memory complexity of boundary-integral-equation simulations of earthquake rupture. Its matrix-generation phase exposes independent element pairs, but each pair's three slip-direction series produce hundreds of coefficients reused by closed-form evaluations throughout a time window. A one-thread-per-pair mapping exceeds register capacity, whereas distributing only the evaluation across threads routes coefficients through device memory. GPU-kernel and thread boundaries therefore cannot be chosen independently. We propose dependency-preserving work decomposition that jointly selects both boundaries from producer–consumer dependencies and GPU resource and execution constraints. For FDPM, one fused kernel handles each series and one warp lane owns each term-specific coefficient–evaluation chain, keeping coefficients on chip and exchanging only low-volume scalars. On an NVIDIA H100, matrix generation runs about 12× faster than a direct OpenACC port on the same GPU. With additional time-evolution optimizations, the whole simulation runs 45× faster than the production A64FX baseline.
Back to Thirteenth Workshop on Accelerator Programming and Directives (WACCPD 2026) Archive Listing
Back to Full Workshop Archive Listing