面向多任务的定制指令模式提取

Custom instruction pattern identification for multi-task environment

为提高嵌入式处理器对计算密集型应用的处理能力,并增强定制指令的适应性,提出一种面向多任务的定制指令模式自动化提取方法。该方法以目标应用的热点代码数据流图集合作为分析对象,通过流图频率加权的方式实现目标任务的优先级调整,并挖掘隐藏于任务程序热点中的频繁计算模式作为定制指令模式。安全加密和媒体处理优化实例结果表明,该方法可提高定制指令的适应性和利用率,其优化效果优于传统独立设计方式。

To improve the performance of computation intensive applications for embedded processors,and enhance the adaptability of custom instructions,an automatic custom instruction pattern identification method for multiple applications is proposed.Dataset combined by data-flow graphs of hot-spots from object applications is taken as analysis target,priorities adjustment is fulfilled though frequency weighting process for data-flow graphs,and common frequent computation patterns hiding in hot-spots of application programs are searched and taken as custom instruction patterns.Two case studies,encryption and media processing,are presented to demonstrate that adaptability and utility of custom instructions are improved by adopting the proposed method,and performance improvement acquired is better than the existed task-independent method.