Signals Approaches to Computer Architecture Prediction Mechanisms

North Carolina State University, Raleigh NC

Investigators

Abstract

This research is focused on developing a new, symmetric framework for computer architecture prediction motivated by the signals approach to data analysis. This research uses styles of prediction in two dimensions: periodic vs. computational and local history vs. global history. This is built on top of the principles of discrete signal analysis. There has been a significant amount of work done on prediction mechanisms, especially for branch behavior. The initial focus of this new work will be on value prediction because it is the least well understood area. However, the framework and methodology are applicable to branch prediction, dependence prediction, etc. This research will investigate new prediction schemes using two techniques: analogy between local and global schemes, and data analysis techniques in the case of a computational predictor. Global predictors can lose accuracy due to pipeline delays. This research project will investigate methods for avoiding the delay penalty for global predictors. The detailed microarchitectural simulation involves a high number of tradeoffs. A detailed study of predictors in that context is a significant study that is included in this project. There is also considerable work to be performed on combining predictors in cost effective and practical ways.