Selecting the Best Compiler Optimizations: A Bayesian Network Approach

Abstract

After presenting our DSE approach for finding good compiler optimizations, we present our autotuning framework to tackle the problem of selecting the best compiler passes. It leverages machine learning and an application characterization to find the most promising optimization passes given an application. This chapter proposes COBAYN: Compiler autotuning framework using Bayesian Networks . An autotuning methodology based on machine learning to speed up application performance and to reduce the cost of the compiler optimization phases. The proposed framework is based on the application characterization done dynamically by using independent micro-architecture features and Bayesian networks. The chapter also presents an evaluation based on static analysis and hybrid feature collection approaches. Besides, we compare our approach against several state-of-the-art machine-learning models. Experiments are carried out on an ARM embedded platform and GCC compiler by considering two benchmark suites with 39 applications. The set of compiler configurations selected by the model (less than 7% of the search space), demonstrated an application performance speedup of up to 4.6$\times$ on Polybench (1.85$\times$ on average) and 3.1$\times$ on Cbench (1.54$\times$ on average) with respect to standard optimization levels. Moreover, the comparison of the proposed technique with (i) random iterative compilation, (ii) machine learning-based iterative compilation and (iii) non-iterative predictive modeling techniques, shows on average, 1.2$\times$, 1.37$\times$ and 1.48$\times$ speedup, respectively. Finally, the proposed method demonstrates 4$\times$ and 3$\times$ speedup, respectively on cBench and Polybench, in terms of exploration efficiency given the same quality of the solutions generated by the random iterative compilation model.