Predicting conditional branches with fusion-based hybrid predictors

Researchers have studied hybrid branch predictors that leverage the strengths of multiple standalone predictors. The common theme among the proposed techniques is a selection mechanism that chooses a prediction from among several component predictors. We make the observation that singling out one particular component predictor ignores the information of the nonselected components. We propose branch prediction fusion, originally inspired by work in the machine learning field, which combines or fuses the information from all of the components to arrive at a final prediction. Our 32 KB predictor achieves the same overall prediction accuracy as the 188 KB versions of the previous best performing predictors (the Multi-Hybrid and the global-local perceptron).