Perceptron-based confidence estimation for value prediction

Data dependencies between instructions have traditionally limited the ability of processors to execute instructions in parallel. Data value predictors are used to overcome these dependencies by guessing the outcomes of instructions in a program. Because mispredictions can result in a significant performance decrease, most data value predictors include a confidence estimator that indicates whether a prediction should be used or not. This paper presents a global approach to confidence estimation in which the prediction accuracy of previous instructions is used to estimate the confidence of the current prediction. Perceptron-based neural networks are used to identify which past instructions affect the accuracy of a prediction and to decide based on their results whether the prediction is likely to be correct or not. Simulation studies compare this global confidence estimator to the more conventional local confidence estimator. These confidence estimators are tested on six SPEC2000 integer benchmark programs with three different prediction approaches: stride, last-value, and context. Results show that predictors using this global confidence estimator tend to predict significantly more instructions and incur fewer mispredictions than predictors using existing local confidence estimation approaches.