矩阵向量乘在流处理器上的实现  

Implementation of Matrix-Vector Multiplication on Stream Processors

作　　者：杜静[1] 敖富江[1] 杨学军[1] 

机构地区：[1]国防科技大学计算机学院,湖南长沙410073

出　　处：《计算机工程与科学》2007年第11期131-133,144,共4页Computer Engineering & Science

基　　金：国家自然科学基金资助项目(60621003;60633050)

摘　　要：Imagine是一款致力于媒体处理的流体系结构芯片,而它是否适合科学计算应用是当前国际上的热点讨论问题。本文研究了典型科学计算核心——矩阵向量乘在Imagine处理器上的有效映射,主要开发了四种不同的流程序版本,并且提出了优化方法。实验结果表明,经过优化的矩阵向量乘程序能够开发Imagine处理器的潜能。Imagine is a stream processor aiming at media applications. However, whether scientific computing applications are suited for Imagine is a hot topic. In this paper, we propose an efficient stream program development technique on Imagine for matrix-vector multiplication, which is a representative scientific computing kernel. Especially, we propose four different program versions of this kernel, and explore efficient programming optimizations. The experimental results show that matrix-vector multiplication can efficiently exploit the tremendous potential of the Imagine stream processor through programming optimization.

关 键 词：矩阵向量乘Imagine三级存储层次 计算密集性 stripmining 

分 类 号：TP301[自动化与计算机技术—计算机系统结构]