基于CNN的间接分支预测器分析  

Analysis of Indirect Branch Predictor Based on CNN

作　　者：裴清辰 周晶 王永文[1] PEI Qingchen;ZHOU Jing;WANG Yongwen(National University of Defense Technology,Hinan 410003,China)

机构地区：[1]国防科技大学,湖南410003

出　　处：《集成电路应用》2022年第9期9-13,共5页Application of IC

摘　　要：阐述在CPU分支预测器对间接跳转进行预测时,大量的预测失效往往是由少数指令造成的。传统的分支预测器难以对这些指令进行有效预测。探讨BranchNet预测器,提出了基于CNN的间接跳转指令分支预测器。该预测器使用特定的分支路径数据进行训练,从而能针对单一的低准确率指令的进行优化,降低预测失效率。实验结果显示,基于CNN的预测器使预测失效率平均降低了10.2%,预测器准确度有明显提升。When the CPU branch predictor predicts indirect jumps, a large number of failures are often caused by a small number of instructions. Traditional branch predictors have difficulty predicting these instructions efficiently. In this paper, a branch predictor for indirect jump instruction based on CNN is proposed based on the BranchNet predictor. The predictor is trained by specific branch data, so that it can be optimized for a single low-accuracy instruction and reduce the prediction failure rate. The experimental results show that the CNN-based predictor reduces the prediction failure rate by 10.2%on average, and the accuracy of the predictor is significantly improved.

关 键 词：分支预测器 间接分支指令 卷积神经网络 静态预测 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TP368.1[自动化与计算机技术—控制科学与工程]