Automatic Loop Transformation Selection with the Aid of Kohonen's Self-Organizing Maps for Parallelizing Compilers.
Mohamed Ahamed Mead, Hesham E. ElDeeb, Salwa M. Nassar · Parallel and Distributed Processing Techniques and Applications · 2008
Determine the appropriate loop transformations is an essential process in the automatic parallelization field. The sequence of loop transformation to be applied also must be considered. Selection of loop transformations faces many challenges, it is needed an experienced. In this paper, An Intelligent Loop Transformation Selector (ILTS); as a part of parallelizing tool project, was developed to overcome on these challenges and imitate an experienced. A Kohonen’s Self-Organizing Map (SOM) neural network is used to select the appropriate loop transformation or sequence of them. Neural Networks offer intelligent transformations selection to reduce or eliminate the dependencies and maximize the parallelization in the sequential code. The experimental results show that ILTS chooses loop transformations successfully in most cases. This tool can be integrated with any parallelizing compiler to enhance loop transformation selection process.