Adaptive Information Processing: An Effective Way to Improve Perceptron Predictors

Perceptron branch predictors achieve high prediction accuracy by capturing correlation from very long histories. The required hardware, however, limits the effective history length to be explored, which in turn undermines the potential performance. In this paper, we propose an adaptive approach to dynamically reconfigure the input vector to a perceptron predictor to facilitate correlation exploitation. In this way, a much larger information set can be explored without increasing the size of a perceptron predictor. Along with carefully designed predictor parameters, the proposed scheme achieves significant improvements on prediction accuracy.