Data Value Prediction Methods and Performance CS/ECE752 Project Report

Elimination of data dependencies through value prediction can help to extract more parallelism. High accuracy value prediction is thus essential in achieving significant performance improvement by exposing more ILP. Various data value prediction methods have been studied [SS97],[WF97]. In this paper, the Shifting Locality Stride Predictor (SLSP) and Repeating Peak Predictor(RPP) are introduced, which will have lower miss rates than available computational predictors(Stride and last value) on certain common value patterns. Their implementation and integration with existing predictors through a counter based selector are discussed, followed by a comparison of the performance between the proposed predictors and Stride and Last Value predictor. The results strengthened our belief that these predictors can help to improve prediction accuracy by an average of 1.9% and 3.1% respectively. We also observed that prediction accuracy will increase with the prediction table size and level off after certain size, though it is not sensitive to the initial values of selection counter.