一个面向存储优化GPGPU编译器  

GPGPU Compiler for Memory Optimization

作　　者：李小强[1] 安虹[1] 吴石磊[1] 

机构地区：[1]中国科学技术大学计算机科学与技术学院,合肥230027

出　　处：《小型微型计算机系统》2014年第4期872-877,共6页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(60970023)资助;国家"九七三"重点基础研究发展计划项目(2011CB302501)资助;国家"八六三"高技术研究发展计划项目(2012AA010902;2012AA010901)资助;国家科技重大专项项目(2009ZX01036-001-002;2011ZX01028-001-002-3)资助

摘　　要：近年来,图形处理器(GPU)在通用计算领域的应用越来越广泛,较高的性能功耗比以及合理的价格使其迅速普及.CUDA则是一种在当今图形处理器上进行通用编程的代表语言.但是由于CUDA过多的暴露了底层硬件的细节,使得程序员必须熟悉GPU的硬件细节才能写出高效运行的程序.而且由于GPU硬件在不断进化,使得为某一代GPU开发的程序很难高效的在新的硬件平台上执行.这大大增加了使用CUDA进行编程的难度和工作量.因此本文提出了一个面向存储优化的GPGPU编译器,将高层的制导语言翻译成优化的CUDA语言.硬件无关的高层制导语言提供了简单的编程模型和可移植性.应用了多种存储优化技术的优化编译器则负责生成高性能的CUDA程序.使用Rodinia测试程序集在Tesla C2050 GPU上的实验结果表明,本文的方案达到了较好的性能和可移植性的平衡.In recent years, GPU is becoming popular for its high performance per watt and reasonable price. CUDA is a representative language to program GPU for general purpose computing. Programmers must have a thorough understanding of the underlying hard- ware to write high performance programs. Optimized CUDA code is normally hardware dependent and code optimized for one plat- form is likely to be inefficient on a different platform because of different hardware parameters. All these problems make CUDA pro- gramming difficult and onerous. In this paper, a GPGPU compiler for memory optimization is proposed. The compiler compile high level directive based language to optimized CUDA code. Hardware independent high level directives provide a simple programming model and portability of source code. The optimizing compiler which implements lots of memory system optimization techniques gen- erates high performance CUDA programs. Experiment results on Tesla C2050 GPU using Rodinia benchmark shows that the compiler achieves good balance of performance and portability.

关 键 词：制导语言 源到源编译 存储优化 GPU 

分 类 号：TP314[自动化与计算机技术—计算机软件与理论]