Revisiting the Perceptron Predictor Again

D. Tarjan and K. Skadron
In Univ. of Virginia Dept. of Computer Science Tech. Report CS-2004-28, Sept 2004

Abstract
We introduce a new kind of branch predictor, the hashed perceptron predictor, which merges the concepts behind the gshare and perceptron branch predictors. This is done by fetching the perceptron weights using the exclusive-or of branch addresses and branch history. This predictor can achieve superior accuracy to a path-based and a global perceptron predictor, previously the most accurate fully dynamic branch predictors known in the literature, at the same storage budgets. Additionally, it reduces the number of adders by a factor of four compared to a path-based perceptron. We also show how such a predictor can be ahead pipelined to yield one cycle effective latency, making it the first standalone perceptron predictor. On the SPEC integer set of benchmarks, the hashed ahead-pipelined path-based perceptron predictor (hashed perceptron for short) improves accuracy by 20% over a path-based perceptron and improves IPC by 5.8%. We believe these improvements make the perceptron predictor a promising choice as a branch predictor for a future high-performance microprocessor.

Available in pdf

The previous perceptron papers should probably be read in the following order:
Vintan: Towards a High Performance Neural Branch Predictor(need subscription to IEEExplore)
Jiménez:Neural Methods for Dynamic Branch Prediction
Seznec:Redundant History Skewed Perceptron Predictors: pushing limits on global history branch predictors
Ipek:On Accurate and Efficient Perceptron-Based Branch Prediction
Jiménez:Fast Path-Based Neural Branch Prediction
Seznec:Revisiting the perceptron predictor

The 1st Championship Branch Prediction produced a bunch of interesting papers using perceptrons, building upon the older work.


Anybody have any additional suggestions? Comments?