Combining local and global history hashing in perceptron branch prediction

Abstract

As instruction issue rates and depths of pipelining continue to increase, branch prediction becomes a performance hurdle for modern processors. Extremely high branch prediction accuracy is essential to achieve these processors’ potential performance. Many perceptron branch predictors have been investigated to improve the dynamic branch prediction in recent years. This thesis introduces combining local history hashing and global history hashing in perceptron branch prediction. The proposed perceptron predictor utilizes self-history as well as global history in indexing different weights of a perceptron. The simulation results show that our proposed perceptron predictor is more accurate than the one using either global history hashing or local history hashing alone. Moreover, our proposed perceptron predictor improves the misprediction rate by up to 26.9% over path-based neural predictor, and 17.2% over hashed perceptron predictor at 8,192 perceptrons configuration. And it is 12.2% more accurate than the global hashed perceptron predictor when the number of perceptrons is 256. In addition, a reducing aliasing approach is employed to improve the accuracy of our proposed perceptron predictor by diminishing the impact of destructive interference. The simulation results provide evidence that the reducing aliasing approach has relatively significant improvement on our proposed perceptron predictor configured with smaller number of perceptrons. Keywords: branch prediction, perceptrons, hashing, neural networks.