Application-specific instruction-set processors (ASIPs) with extended instructions (Els) can effectively improve performance and meet time-to-market requirements.However,there are three problems to be solved when customizing Els for embedded systems.First,design space increases exponentially as applications become more complex.Second,limited on-chip resources restrict the number and complexity of Els. Third, current instruction-set extension (ISE) algorithms can hardly run on embedded systems due to their high complexity. We propose a fast automatic method called GISEES to address these problems. GISEES can effectively prune design space by enumerating Els around typical operations of applications,and reduce resources overhead and the number of inserted multiplexers through sharing resources based on finding the maximal common equivalent substring (MCES). Experiment results show that,GISEES features linear complexity and can generate Els with higher efficiency, which means it is more suitable for customizing power efficient instruction-set extension.%面向应用的指令集处理器通过增加扩展指令可有效提升处理器的性能,满足上市时间要求.然而为嵌入式系统定制扩展指令需解决以下3个问题:设计空间随应用复杂度的增加指数增加,有限的片上资源限制了扩展指令的数量和复杂度,现有指令集扩展算法复杂度高难以在嵌入式系统上运行.本文提出了一种快速的指令集扩展方法GISEES.该方法以应用的典型操作为中心产生扩展指令以裁剪了设计空间,并采用基于最大公共等价子串的资源共享策略减少资源开销和插入的多路选择器的数量.实验结果表明,该方法具有线性复杂度,可产生效率更高的扩展指令,更适合为嵌入式系统定制高效的扩展指令.
展开▼
