GPU矩阵乘法的性能定量分析模型

发布时间：2018-03-11 14:02

本文选题：GPU　切入点：GPGPU-Sim　出处：《计算机科学》2015年12期 　论文类型：期刊论文

【摘要】：性能评价和优化是设计高效率并行程序必不可少的重要工作,存储系统的性能高低直接影响到处理器的整体性能。利用GPGPU-Sim对GPU的存储层次结构进行了模拟,找出了SM数量与存储控制器数量之间最佳配置关系。矩阵乘法是科学计算领域中的基本组成部分,是一种具有计算和访存密集特点的典型应用,其性能是GPU高性能计算的一个重要指标。性能模型作为并行系统性能评价的新的技术解决方案,具有许多其它性能评价方法无法比拟的优势。建立了一个性能模型,模型通过对指令流水线、共享存储器访存、全局存储器访存进行定量分析,找到了程序运行瓶颈,提高了执行速度。实验证明,该模型具有实用性,并有效地实现了矩阵乘法的优化。
[Abstract]:Performance evaluation and optimization is an essential work in designing efficient parallel programs. The performance of storage system directly affects the overall performance of the processor. The memory hierarchy of GPU is simulated by GPGPU-Sim. The optimal configuration relationship between SM number and memory controller number is found. Matrix multiplication is a basic part of scientific computing field, and it is a typical application with the characteristics of computing and memory access. As a new technical solution for parallel system performance evaluation, the performance model has many advantages that cannot be compared with other performance evaluation methods. Through quantitative analysis of instruction pipeline, shared memory access and global memory access, the model finds the bottleneck of program running and improves the execution speed. Experiments show that the model is practical. The optimization of matrix multiplication is realized effectively.
【作者单位】：武汉大学计算机学院;湖北工程学院计算机与信息科学学院;
【基金】：国家自然科学基金(61370092) 湖北省自然科学基金(2013CFC005) 湖北省中青年创新团队(T201410)资助
【分类号】：TP332.22

【参考文献】