Value Prediction With Perceptrons

This paper presents a new technique for value prediction. It uses perceptrons, one of the simplest neural networks, for prediction of instruction output values. Perceptrons have been shown to be highly effective for conditional branch prediction. Current value predictors use two-bit saturating counters for value prediction and involve an exponential increase in hardware resources. This limits the length of value history that can be explored, resulting in reduced prediction accuracy. Perceptrons scale linearly with value history length and can explore longer history, so they can achieve higher prediction accuracy and higher instructions per cycle (IPC). We present a first paper that explores how to use perceptrons for value prediction. We consider a number of design space tradeoffs in this paper. First, we discuss the design of both finite context method (fcm) and stride predictors using perceptrons, and combine these mechanisms to realize a hybrid perceptron predictor. Second, we report on the accuracy of our hybrid perceptron predictor while running the SPEC2000 integer benchmarks. We find that we can improve prediction accuracy over an aggressive baseline design, while obtaining equivalent or better IPC. Finally, we compare results for predictors with equal hardware budgets and conclude that a perceptron predictor can still provide better value prediction accuracy that a more traditional design.