ROMJIST Volume 22, No. 3-4, 2019, pp. 215-227
Arpad GELLERT, Maria VINȚAN, Lucian VINȚAN Perceptron-Based Selective Load Value Prediction in a Multicore Architecture
ABSTRACT: In this work we have extended the load value prediction concept, previously implemented by us in a multicore architecture, with perceptron-based classification of load instructions into the predictable and unpredictable categories. The prediction scheme has been integrated into the Sniper multicore simulator. The goal of the load value predictor is to anticipate the values of critical load instructions and to unlock in a speculative manner the subsequent dependent instructions’ processing. Since high prediction accuracies are necessary, we use the predicted value only if the corresponding load instruction is classified as being predictable. The evaluations performed on the Splash-2 parallel benchmarks have shown an average relative speedup of 4.21% over the baseline multicore architecture, with a maximum of about 17%.KEYWORDS: Multicore architecture; issue bottleneck; selective load value prediction; perceptron; speculative execution; Splash-2 parallel benchmarks; Sniper multicore simulatorRead full text (pdf)