Study of Register Value Branch Predictor Based on CNN

In modern processor architecture, a branch predictor is an important module whose design influences the performance of the processor. Nowadays, the factors affecting the performance of branch predictors have changed from common branches to hard-to-predict branches, which rarely show up but have low prediction accuracy. In particular, the data-related branches are a difficult problem to be handled by the current predictors. In this paper, we continue to study Zangeneh’s BranchNet CNN neural network branch predictor to explore the solution. Due to the observation that some data-related branches are associated with register values, we add BranchNet, a sub-network that handles register values in the offline training phase and modify the above sub-network according to the feedback from the preliminary experiments by adjusting the network parameters and redesigning the network structure. Through that process, it can better learn the relationship between register values and branch results. The experiments were carried out on the SPEC 2017 integer benchmarks. The network was trained and validated on the top five H2P branches of all benchmarks. The results show that the addition of register information reduces the MPKI rate of the top five H2P branches on average, compared to the BranchNet, by 17.32%. Analyzing the results, we find that the network with the addition of register value information has significantly improved the prediction accuracy of certain H2P branches, indicating that the addition of register values is effective in H2P branch prediction.