基于程序特征分析的流处理器VLIW压缩技术与解压实现  

Program characteristics analysis-based VLIW compression and decompression on stream processor

作　　者：管茂林[1] 何义 杨乾明[1] 张春元[1] 

机构地区：[1]国防科技大学计算机学院,湖南长沙410073 [2]北京油料研究所,北京102300

出　　处：《国防科技大学学报》2012年第1期138-143,共6页Journal of National University of Defense Technology

基　　金：国家自然科学基金资助项目(61033008;60903041);教育部博士点基金资助项目(20104307110002)

摘　　要：代码体积和代码稀疏是VLIW处理器一直存在的问题。通过对一系列典型应用在流处理器上的程序特征进行分析,提出了一种新的VLIW分域压缩技术,剔除各个子域中的空操作,并设计了分布式指令存储器对压缩后的代码进行解压缩执行。实验证明,该技术能够减少MASA流处理器中近39%的片外指令访存,降低约65%的片上指令存储器空间需求;同时使得指令存储器面积和系统面积分别减少了约37%和8.9%。Huge code cubage and poor code density have always been a serious problem in VLIW processors. Through analyzing the code characteristics of a series of typical applications on stream processor, a novel domain-divided VLIW code compression scheme to eliminate the nop operations in each domain, and a distributed instruction memory to decompress and execute of the codes are proposed. The experiments show that this scheme （：an reduce nearly 39% of the oil-chip instruction memory accessing and about 65% of the on-chip instruction memory space demand. Meanwhile, it can also depress the area of on-chip instruction memory and MASA stream processor by 37% and 8.9% respectively.

关 键 词：特征分析 流处理器 代码压缩 分布式指令存储器 

分 类 号：TP302[自动化与计算机技术—计算机系统结构]