藉由多面體模型 (Polyhedral Model) 改善同質多核心系統之資料局部性

Abstract

The parallel programming model on homogeneous multiprocessors includes OpenMP, MPI, and others. OpenMP is the most widely used in these common parallel programming models. On the other hand, the performance bottleneck on homogeneous multiprocessors is usually the memory access within a nested loop. Therefore, programmers would adopt some loop optimizations to enhance data locality and reduce external memory access times for the parallel execution on homogeneous multiprocessors. The common techniques of loop optimizations are loop interchange and loop tiling. The source code of the nested loops would usually translate to the intermediate representation (IR) for applying the loop optimizations. After a series of optimizations, IR would translate back to the source code or the executable program. The common IR includes abstract syntax tree, polyhedral model, and others. Using polyhedral IR could avoid causing side-effects between transformation orders, such as increasing the code size and enhancing the transformation complexity. In this thesis, we present the implementation of a polyhedral source-to-source transformation framework that could optimize the nested loops for the data locality improvement. The framework could find the DOALL loops automatically and adopt loop tiling to enhance the data locality. According to the simulation results, the programs could increase 14.1% performance in average by performing loop tiling. Moreover, programmers could develop loop optimizations quickly based on our polyhedral framework.