面向弹性波模拟方法中复杂Stencil的GPU性能优化.pdfVIP

面向弹性波模拟方法中复杂Stencil 的GPU 性能优化 1 2+ 1 12 吕子鉷 , 付昊桓 , 赵文来 , 杨广文 1(清华大学 计算机科学与技术系,北京 100084) 2(清华大学 地球系统科学研究中心,北京 100084) Optimizing Complex Stencils in Elastic Wave Simulation Methods on GPUs 1 2+ 1 12 LÜ Zi-Hong , FU Hao-Huan , ZHAO Wen-Lai , YANG Guang-Wen 1(Department of Computer Science and Technology, Tsinghua University, Beijing 100084, China ) 2(Ministry of Education Key Laboratory for Earth System Modeling, and Center for Earth System Science, Tsinghua University, Beijing 100084, China) + Corresponding author : Phn: +86- 10 Fax: +86- 10 E-mail : haohuan@ Abstract : The paper focuses on optimizing complex stencils which generally involves hundreds of variables and thousands of operations. We choose the Weighted Nearly Analytic Discrete (WNAD) stencil, which is one of the most complex stencils used in elastic wave modeling, as our target to explore the optimization techniques on GPUs. Besides optimizations on data reuse in fast buffers, tuning on the mapping of parallelism and etc., we further include graph-oriented code transformation and instruction re-scheduling into a unified work flow to balance the utilization of various system resources and achieve the best possible performance. We explore the balanced combination on register usage and computation amount by extracting and transforming the expression graph of the original program. Automated instruction re-scheduling is included to improve instruction-level parallelism. By applying these optimization techniques to the WNAD stencil kernel on NVIDIA Kepler GPU, we manage to improve the performance from 231 GFlops to 972 GFlops. Key words: stencil; GPU; complex stencil; optimization; tuning 摘 要: 本文针对包含数