基于数据流图的专用指令生成方法  

Application specific instruction generation based on data flow graph

作　　者：李德贤[1] 严晓浪[1] 彭剑英[1] 

机构地区：[1]浙江大学超大规模集成电路研究所,浙江杭州310027

出　　处：《浙江大学学报（工学版）》2009年第1期103-107,共5页Journal of Zhejiang University：Engineering Science

基　　金：国家"863"高技术研究发展计划资助项目(2005AA1Z1271)

摘　　要：提出了一种基于数据流图(DFG)的专用指令生成方法.以目标应用的高级语言参考代码为起点,通过编译器前端转化为三地址中间格式,得到有向数据流图.使用提出的基于处理器体系架构约束的静态搜索算法对得到的数据流图进行搜索,得到满足约束的操作组合.结合典型测试序列的动态运行数据对搜索结果进行进一步的筛选,确定对于目标应用性能起关键作用的操作组合,以此作为专用加速指令.该方法实现了数据流图提取的自动化,并结合了静态数据流图搜索与动态结果筛选.通过在视频压缩专用处理器设计中的应用,证明此方法可以快速高效地进行专用指令的自动生成.An automatic application specific instruction generation approach based on the data flow graph （DFG） of target application was proposed. The high level reference code of target application was parsed by the compiler frontend to get the three-address intermediate representation, and the directed data flow graph was generated. A processor-architecture-constraint based static DFG search algorithm was designed to traverse the graph and extract all micro-operation combinations meeting the constraints. Searching results were further selected using the dynamic run-time statistics of application code to find operation combinations significantly influencing the target application performance. Then the operation combinations were chosen as accelerating instruction sets. The approach can automatically extract the data flow graph and combines the static graph traversing with dynamic results selection. The case study of application-specific processor for video compression shows that the approach can efficiently generate the application specific instruction set.

关 键 词：专用指令集处理器 数据流图 指令生成 

分 类 号：TN47[电子电信—微电子学与固体电子学]