Optimizing stencil application on multi-thread GPU architecture using stream programming model

  • Authors:

  • Fang Xudong;Tang Yuhua;Wang Guibin;Tang Tao;Zhang Ying

  • Affiliations:

  • National Laboratory for Parallel and Distributed Processing, School of Computer, National University of Defense Technology, Changsha, Hunan, China;National Laboratory for Parallel and Distributed Processing, School of Computer, National University of Defense Technology, Changsha, Hunan, China;National Laboratory for Parallel and Distributed Processing, School of Computer, National University of Defense Technology, Changsha, Hunan, China;National Laboratory for Parallel and Distributed Processing, School of Computer, National University of Defense Technology, Changsha, Hunan, China;National Laboratory for Parallel and Distributed Processing, School of Computer, National University of Defense Technology, Changsha, Hunan, China

  • Venue:
  • ARCS'10 Proceedings of the 23rd international conference on Architecture of Computing Systems
  • Year:
  • 2010

Quantified Score

Hi-index 0.00

Visualization

Abstract

With fast development of GPU hardware and software, using GPUs to accelerate non-graphics CPU applications is becoming inevitable trend. GPUs are good at performing ALU-intensive computation and feature high peak performance; however, how to harness GPUs' powerful computing capacity to accelerate the applications in the field of scientific computing still remains a big challenge. In this paper, we implement the whole application Mgrid taken from Spec2000 benchmarks on an AMD GPU and propose several optimization strategies for stencil computations in the naive GPU code. We first improve thread utilization through using vector types and multiple output streams mechanism provided by the Brook+ programming language. By tuning thread granularity, we try to hit the right balance between locality within each thread and parallelism among threads. Then, we reorganize the stream layout by transforming the 3D data stream into the 2D stream in the block manner. Through stream reorganization, more data locality in the cache is exploited. Further, we propose branch elimination to convert control dependence to data dependence, catering to GPUs' powerful ALU-intensive processing capability. Finally, we redistribute computations between CPU and GPU to make more advisable computing resources usage considering different problem sizes. We demonstrate the effectiveness of our proposed optimization strategies on an AMD Radeon HD4870 GPU using the Brook+ programming language. Using a double-precision floating-point implementation, the experimental results show that the optimized GPU version of Mgrid gains 2.38x speedup compared to the naive GPU code and obtains as high as 15.06x speedup versus the CPU implementation run on an Intel Xeon E5405 CPU.