基于相似性的粗粒度可重构指令压缩  

Similarity aware context reduction for coarse grained reconfigurable architectures

作　　者：李锦超 赵仲元[1] 绳伟光[1] LI Jin-chao;ZHAO Zhong-yuan;SHENG Wei-guang(School of Electronic Information and Electrical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)

机构地区：[1]上海交通大学电子信息与电气工程学院,上海200240

出　　处：《微电子学与计算机》2020年第8期21-26,共6页Microelectronics & Computer

基　　金：国家自然科学基金项目(61176037,61201059)。

摘　　要：粗粒度可重构架构在能效比方面具有明显优势,然而其指令存储与传输过程的功耗代价过高.实验发现指令间具有明显的相似性,由此本文提出一种基于指令相似性的压缩技术,通过对指令的压缩、传输与解压,可以在不降低性能的前提下,优化架构的功耗和面积.针对同构和异构平台分别提出了指令分发模型和指令寄存器模型的解决方案,结合编译策略优化,最终与两种传统结构相比,面积效率比分别提升36%和181%,功耗效率比分别提升33%和118%.Coarse-Grained Reconfigurable Architectures(CGRAs)are promising computing platform with better efficiency.However,too much cost is wasted on the path of context storage and transmitting.With statistics of real case,there is obvious similarity between context.We propose one kind of context reduction method based on similarity of context.By making context compress,transmit,and decompress,this method can reduce power and area of CGRA obviously,with little overhead on performance.We propose context dispensing and register model for homogeneous heterogeneous platform.With the optimize of compiler,we get 33%and 118%promotion on power efficiency to tradition architecture,36%and 181%on area efficiency.

关 键 词：粗粒度可重构架构 指令压缩 能效比 异构计算 

分 类 号：TP303[自动化与计算机技术—计算机系统结构]