一种适用于向量处理器的新型归约网络  

Novel Reduce Network for Vector Processor

作　　者：庄巍[1] 杨惠[1] 刘衡竹[1] 

机构地区：[1]国防科技大学计算机学院微电子与微处理器研究所,长沙410073

出　　处：《小型微型计算机系统》2012年第11期2498-2502,共5页Journal of Chinese Computer Systems

基　　金：核高基重大专项(2009ZX01034-001-006)资助

摘　　要：现代通信和图像处理等应用领域中向量运算的宽度灵活多变,且程序可循环执行.为了支持宽度的灵活多变,研究并提出一种新颖的分层计算模型,在基于有16个向量运算单元的向量处理器上用该模型构建了一种可配置成多组多宽度的归约树结构,向量宽度为2至16.设计了控制模块,在无需修改程序的情况下可通过隐式自增的方式指定归约目标,使单程序可以多次循环执行.对所实现的归约网络部件在TSMC65nm工艺库下用Synopsys的Design Complier工具进行综合,结果显示其工作频率可达700MHz,面积为92363.04um2,功耗为17.30 mW.在相同工艺库和工作频率下,该设计的性能开销比是一种精简归约网络的2.36倍.The properties of vector computation in communication of modern time and image processing are variety of vector width and looply executing of program.In order to support the variety of vector width,a novel model of hierarchical computing is designed.Basing on it and aiming at a vector processor which consists of 16 vector processing elements,a reduce tree which supports the modes of grouping and various width from 2 to 16 is constructed.A controlling module which can implicitly decide the destination of reduction is presented so that the program can be executed looply without being modified.The novel reduce network unit was synthesized in 65nm CMOS technology by Design Complier of Synopsys.Its frequency can reach 700MHz,its area is 92363.04um2 and its power is 17.30 mW.In the same process library and frequency,its ratio of performance to cost is 2.36 times as great as a reduced reduce network.

关 键 词：向量运算 归约网络 分组多宽度 程序循环执行 

分 类 号：TP303[自动化与计算机技术—计算机系统结构]