PI-1554: Redundant History Skewed Perceptron Predictors: pushing
limits on global history branch predictors
http://www.irisa.fr/bibli/publi/pi/2003/1554/1554.html
André Seznec
19 pages - septembre 2003

Abstract
In this paper, we build upon the original perceptron predictor to
propose the redundant history skewed perceptron predictor. We show that
the potential accuracy of a global history branch predictor based on
perceptron was significantly underestimated in the former work by
Jimenez and Lin. We establish that, when considering a hypothetical
conflict-free perceptron predictor, the potential accuracy can be
significantly higher when a redundant representation of the global
history or path is used. In particular, we show that using redundant
history associated with very long history lengths (up to 128 bits) may
lead to significant improvement of the potential prediction accuracy
when using the perceptron predictor. Despite using a larger number of
counters in each perceptron, at equal finite-size storage budgets, the
accuracy of a redundant history perceptron predictor is significantly
higher than the one of a perceptron predictor using standard branch
history. We also show that splitting the perceptron predictor table in
several physical tables allows to further improve the accuracy through
the use of different hashing functions mixing branch history and
address. At equivalent storage budgets, the redundant history skewed
perceptron predictor outperforms (and sometimes very significantly) both
the original perceptron predictor and more conventional global history
predictors on every of our benchmarks.

Résumé: Dans ce rapport, nous définissons un nouveau prédicteur de
branchement, le skewed perceptron à historique redondant. Nous montrons
que la précision du schéma de prédiction du prédicteur à perceptron a
été largement sous-estimé par les premiewrs travaux de Jimènez et Lin.