面向GPU的体系结构敏感型数值算法优化技术研究

GPU has been well recognized as the main platform for high performance computing along with the development of multi-core techniques. Its essential features in multilevel memory hierarchy and various thread management have distinguished GPU from any other previous computing platforms, such that the traditional optimization techniques cannot meet the high performance demand of arthitecture-sensitive numerical algorithms in real practices.To take full advantage of GPU's unique architecture, we must research the new optimization methods to make better the parallelism, locality and irregularity feature of numerical algorithms. In this work, we take quantitative measurement and analysis of various factors that affect the performance of arthitecture-sensitive numerical algorithms on GPU, and build the GPU performance model being described by these factors. Furthermore, targeting on the arthitecture-sensitive numerical algorithms, we breaking through traditional concept to study the autotuner technology to optimize the parallelism on multi-layers of the architecture in order to improve the locality of data access, thread load balancing and better data stream processing. We would like to study the unified optimization technique that serves both the regular and irregular computations, and guide the optimization of architecture-sensitive numerical algorithms with our GPU performance model. Our study can be applied to generate highly efficient codes for architecture-sensitive numerical algorithms, which are of significant importance both in scientific computing theory and real practices.

随着多核技术的不断发展，GPU已经成为高性能计算的主流平台，与以往相比，该系统的体系结构具有更多的存储层次和多样化的线程管理，传统的优化技术已经不能满足体系结构敏感型数值算法的性能要求。该类算法必须重构以实现深度挖掘自身的并行性、局部性和非规则计算特性,通过充分发挥GPU的体系结构优势,提高程序性能。为此，本课题面向GPU体系结构,通过定量的测试和分析影响体系结构敏感型数值算法执行效率的各种因素，形成GPU性能模型，刻画体系结构敏感性指标。在此基础上，研究体系结构敏感型数值算法的多层次优化方法及自动调优策略,改善访存局部性、线程间负载均衡、数据读写和流处理方式。研究规则计算和非规则计算统一的性能优化方法。并且将GPU性能模型用于指导体系结构敏感型数值算法的调优方法和策略。本项目研究成果可以很好地提高体系结构敏感型数值算法执行效率，具有重要的理论意义和应用价值。

随着多核技术的不断发展,GPU 已经成为高性能计算的主流平台,与以往相比 该系统的体系结构具有更多的存储层次和多样化的线程管理,传统的优化技术已经不能满足 体系结构敏感型数值算法的性能要求。该类算法必须重构以实现深度挖掘自身的并行性、局 部性和非规则计算特性,通过充分发挥 GPU 的体系结构优势,提高程序性能。为此,本课题面 向 GPU 体系结构,通过定量的测试和分析影响体系结构敏感型数值算法执行效率的各种因素 形成 GPU 性能模型,刻画体系结构敏感性指标。在此基础上,研究体系结构敏感型数值算法 的多层次优化方法及自动调优策略,改善访存局部性、线程间负载均衡、数据读写和流处理 方式。研究规则计算和非规则计算统一的性能优化方法。并且将 GPU 性能模型用于指导体系 结构敏感型数值算法的调优方法和策略。本项目研究成果可以很好地提高体系结构敏感型数 值算法执行效率,具有重要的理论意义和应用价值。