Overall, automatic per-kernel block/workgroup-shape selection improves OPS performance across GPUs, provided shape sensitivity and timing stability are interpreted together.
Abstract
Large-scale structured-mesh stencil applications can run for hours or days on modern HPC GPU platforms, making launch-configuration choices important. OPS enables high-level stencil codes to target multiple GPU architectures, but its GPU backends rely on a default or user-selected global block/workgroup shape, even though different kernels, problem geometries, and architectures may favor different configurations. This paper presents a transparent per-kernel block/workgroup-shape autotuning extension for OPS. The extension explores valid block/workgroup shapes during execution, records the fastest configuration for each kernel, and enables precomputed replay of the selected configurations in subsequent runs. We evaluate the approach on structured-mesh proxy and academic research applications, covering 459 unique kernels across 3 NVIDIA and 2 AMD GPUs. The evaluation reports speedup distributions, representative wall-time summaries, timing variability, and input problem shape effects. Replaying the best per-kernel launch shapes improves 923 of 995 measured application/input/GPU cases over the default OPS shape. Depending on the GPU, the median speedup ranges from 1.02x to 1.18x, with individual configurations of LAPL2D, TTI3D, and OpenSBLI-based TGV solvers exceeding 1.5x. The results also show that performance sensitivity is not limited to launch shape. At fixed cell count, changing the input problem shape changes precomputed-replay runtime by 0.8-34.6% in representative cases, while kernel-level timing variability depends on the interaction between runtime contribution, input problem shape, and GPU/backend behavior. Overall, automatic per-kernel block/workgroup-shape selection improves OPS performance across GPUs, provided shape sensitivity and timing stability are interpreted together.
GPU architecture research requires representative benchmarks that can execute on new designs. However, many vendor-tailored workloads and library kernels cannot run directly on research architectures because their implementations are closed-source or rely on programming models unsupported by the target architecture. Mi...
Huan-Zhi Pu, Chulhyung Park, Chihyo Ahn et al.· 0 citations
CUDA is the dominant GPU programming model in HPC and industrial accelerator software, and a large body of production code is written directly in it. Deploying that code on non-NVIDIA accelerators has traditionally required source translation, backend-specific rewrites, or a full rewrite in a new programming model. Thi...
Beau Johnston, Chris Kitching, Matthew Ireland et al.· Workshop Proceedings of the...· 0 citations
Achieving performance portability is challenging for irregular scientific applications built as repeated multi-kernel GPU pipelines, because throughput depends on out-of-kernel decisions as much as on kernel efficiency. Using molecular docking as a representative case study, we propose an out-of-kernel tuning methodolo...
G. Accordi, Leonardo Beltrame, Theresa Pollinger et al.· 0 citations
SlideDP is presented, a synchronous data-parallel runtime for shared-host multi-GPU systems that maintains one authoritative host state, decouples communication routes from state layout, and pipelines parameter delivery, gradient aggregation, and CPU updates across ranks and chunks.
Rui-Jia Yang, Shi-Yuan Lin, Yu-Long Ao et al.· arXiv.org· 1 citation
MorphX is a GPU multitasking system that embraces the runtime adjustability of kernels as the first principle, and transparently transforms traditional kernels into morphable ones at compile time, and leverages an in-place profiling mechanism to orchestrate fine-grained scheduling based on real-time resource contention...
Ting-Xu Ren, Ru-Wen Fan, Hao Guo et al.· Proceedings of the ACM SIGOP...· 1 citation
Behind the expanding use of large language models (LLMs) across application domains lies a growing dependence on specialized GPU kernel libraries. These libraries underpin the correctness and efficiency of training and inference across diverse GPU platforms, implementing attention, GEMM, quantization, and low-precision...
Yan-Jie Gao, Jia-Yi Bi, Hao-Xiang Lin et al.· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.