Applying Perceptrons to Speculation in Computer Architecture

Modern microprocessors make use of speculation, or predictions about future program behavior, to optimize the execution of programs. Perceptrons are simple neural networks that can be highly useful in speculation for their ability to examine larger quantities of available data than more commonly used approaches, and identify which data lead to accurate results. This work first studies how perceptrons can be made to predict accurately when they directly replace the traditional pattern table predictor. Different training methods, perceptron topologies, and interference reduction strategies are evaluated. Perceptrons are then applied to two speculative applications: data value prediction and dataflow critical path prediction. Several novel perceptron-based prediction strategies are proposed for each application that can take advantage of a wider scope of past data in making predictions than previous predictors could. These predictors are evaluated against local tablebased approaches on a custom cycle-accurate processor simulator, and are shown on average to have both superior accuracy and higher instruction-percycle performance. This work is addressed to computer architects and computer engineering researchers.