降低函数调用分支别名的神经网络预测器

Neural branch prediction with alias reducing between function calls

通过对程序调用过程中分支预测空间特性的分析,发现传统神经网络算法在不同函数调用相同子函数时容易出现别名效应,进而提出了一种基于子函数权重索引离散的神经网络分支预测器。该预测器通过调用信息堆栈记录函数调用中的父函数的路径信息,并用该信息离散子函数权重索引,有效降低了由于不同父函数调用相同子函数造成的别名效应。实验结果显示,基于该方法的神经网络分支预测器的预测错误率降低1%～10%。

With the spatial property analysis of the branch prediction, found the traditional neural branch prediction method had alias effect between function calls. Proposed a new receptor accessing mechanism with index dispersing in sub-function. It saved the path information in function calling stack for functions and dispersed the receptor index with the path information in branch prediction. It would differentiate the branch of sub-function in different function calling and could eliminate the prediction alias effect between function calls. Experiment shows that the neural prediction with alias reducing can decrease the misprediction rate by 1%～10%.