Adaptive Auto-Tuning Framework for Global Exploration of Stencil Optimization on GPUs

Qingxiao Sun,Yi Liu,Hailong Yang,Zhonghui Jiang,Zhongzhi Luan,Depei Qian

DOI: https://doi.org/10.1109/tpds.2023.3325630

IF: 5.3

2023-12-01

IEEE Transactions on Parallel and Distributed Systems

Abstract:Stencil computations are widely used in high performance computing (HPC) applications. Many HPC platforms utilize the high computation capability of GPUs to accelerate stencil computations. In recent years, stencils have become more diverse in terms of stencil order, memory accesses and computation patterns. To adapt diverse stencils to GPUs, a variety of optimization techniques have been proposed. Due to the diversity of stencil patterns and GPU architectures, no single optimization technique fits all stencils. Therefore, stencil auto-tuning mechanisms have been proposed to conduct parameter search for a given combination of optimization techniques. However, parameter search for an inappropriate optimization combination (OC) misses the globally optimal solution. To address the above problems, we propose GSTuner, an adaptive auto-tuning framework that efficiently determines the optimal parameter setting of the global optimization space for stencils on GPUs. Specifically, GSTuner represents stencil patterns as neighboring features and unifies feature vectors of OCs through data pre-processing. In addition, GSTuner samples parameter settings from superior OCs via the quota-based reward policy and regression mechanisms. After that, GSTuner employs the genetic algorithm that considers sub-population similarity to reduce the cost of evolutionary search. The experiment results show that GSTuner can identify better performing settings with higher auto-tuning speed compared to the state-of-the-art works.

computer science, theory & methods,engineering, electrical & electronic