Static and Dynamic Branch Prediction Using Neural Networks

In this short paper we investigated a new static branch prediction technique. The main idea of this technique is to use a large body of different programs (benchmarks) to identify and infer common C program behaviour. Then, this knowledge is used to predict new “unseen” branches belonging to new programs. The common behaviour is represented as a set of static features of branches that are mapped using a neural network to the probability that the branch will be taken. In this way the predictor does not predict a program behaviour based on previous execution of the same program or based on some program profiles but uses the knowledge gathered from other programs (knowledge experience). Also we combined static and a dynamic neural branch predictor in order to investigate how much influences the static predictor the dynamic one.