Low-level Code Auto-tuning for State-of-the-art Multicore Architectures

Abstract

The article presents an approach to automatic transformation of sequential code in parallel for a given multicore architecture. A generalized code transformation methodology based on its intermediate low-level representation and the mathematical apparatus of Petri nets with additional semantic relations is presented. The algorithm of automatic parallelization of the linear and cyclic structures based on equivalent transformations of the initial net of the sequential algorithm is presented. The paper shows the experiments for automated parallelization of the linear algorithm.