DCR置信度评估方案  

DCR Confidence Estimation Scheme

作　　者：朱霞[1] 高德远[1] 樊晓桠[1] 

机构地区：[1]西北工业大学计算机科学与工程系,西安710072

出　　处：《计算机学报》2004年第8期1121-1128,共8页Chinese Journal of Computers

基　　金：国家自然科学基金项目"可重构计算中控制配置策略研究"( 60 2 73 0 88)资助

摘　　要：指令间的控制相关对处理器性能的影响比数据相关带来的影响更为严重 .虽然转移预测能够提高处理器的性能 ,但不同转移指令的可预测性是不同的 .置信度评估方案可被用来评测转移预测方案的可信度 ,该文利用误预测集群性 ,设计了DCR(DecreaseConstantorReset)置信度评估方案 .仿真结果表明 ,DCR方案相对文献中的JRS方案和MDC方案 ,否定测试的预测值PVN分别提高 2 9 5 %和 4 2 19% ;特异性值SPEC分别提高 0 5 3%和15 1 8% ,从而更适合在多线程处理器的双路径执行时使用 .最后 ,该文讨论参数变化对DCR方案性能的影响 .Control dependence is more critical to the performance of microprocessor than data dependence. Many branch prediction schemes are studied to improve performance, but the predictability of different branch instructions is different. It is known that Confidence estimation can be used to evaluate whether a branch prediction will be correct or not. In this paper, Decrease Constant or Reset (DCR) confidence estimation mechanism, which can take advantage of the clustering of mispredicted branches, is proposed. The implementation of the DCR mechanism is discussed in details and the performance change of DCR mechanism with the parameter change is analyzed. Compared with the JRS scheme, the DCR scheme can improve the SPEC and PVN metrics of confidence estimation by 0.53% and 29.5%, respectively. Meanwhile, compared with Misprediction Distance Counter (MDC) scheme, the DCR scheme can improve the SPEC and PVN metrics of confidence estimation by 151.8% and 42.19%, respectively. Thus the DCR scheme is suitable to be used in dual path execution in multithreaded microprocessors.

关 键 词：转移预测 置信度评估 双路径执行 

分 类 号：TP303[自动化与计算机技术—计算机系统结构]