GPU程序访存行为分析和优化关键技术研究

Graphics Processing Unit (GPU) has attracted much interests in high performance computing area due to its powerful computing performance. With massive parallel computing cores, GPU's peak performance far ex ceeds that of contemporary CPU. However, real applications usually cannot efficiently exploit GPU's potential computing ability. The major factor lies in the "memory wall" problem. Therefore, analysis and optimizations oriented to on-chip memory hierarchies are of great importance to improve the performance of GPU programs. Based on deep analysis of GPU's multi-thread executing model, this program will perform researches on on-chip memory access analysis and optimizations for GPU programs. The key contents cover the following two aspects: proposing analysis model of the impact of on-chip memory accesses on the performance, including the software managed shared memory and the hardware managed cache; proposing data access optimizations for shared memory and cache respectively. With these results, we expect to improve the usage of GPU's on-chip memory hierarchies, reduce the off-chip memory accesses and efficiently exploit GPU's potential computing ability.

图形处理器(GPU)以其强大的峰值计算能力渐渐成为高性能计算领域研究的热点。由于集成了大量并行的计算核心，GPU 的峰值计算性能远远超过同时期的通用 CPU，然而真实程序往往很难将其计算潜力有效发挥出来，其中“存储墙”问题是制约GPU性能发挥的最主要因素，因此面向GPU片上存储层次的分析和优化对于提高GPU程序的性能具有重要意义。本项目基于对GPU执行特点的深入探索和分析，展开GPU片上存储访问的分析和优化研究，关键内容包括两方面：一是建立GPU片上存储层次访问对程序性能影响的分析模型，包括软件管理的shared memory和硬件管理的cache；二是提出若干面向shared memory和cache的数据访问优化技术。我们希望通过这些研究成果，有效提升GPU片上存储层次的利用率，减少片外访存，充分开发GPU的计算潜力。

由图形处理器（GPU）作为加速器构建异构并行系统已经成为高性能计算领域最常采用的技术路线之一。由于降低了核心的复杂度，GPU往往集成了大量的并行的计算核心，其峰值性能远超同时期的通用CPU，但真实程序的性能发挥严重依赖于访存性能的发挥。因此，面向GPU片上存储层次的分析和优化对于提高GPU程序的性能具有重要意义。本项目基于对GPU多线程执行模式的分析，建立了GPU片上存储层次访问对程序性能影响的分析模型，并基于此提出了若干面向shared memory和cache的数据访问优化技术，可以有效提升GPU片上存储层次的利用率，减少访存，从而提升GPU的实际计算性能。本项目在国内外会议和期刊上公开发表论文9篇。