Interactive Refactoring for GPU Parallelization of Affine Loops
Kostadin Damevski, Madhan Muralimanohar · 2013
Abstract—Considerable recent attention has been given to the problem of porting existing code to heterogeneous comput-ing architectures, such as GPUs. In this paper, we describe a novel, interactive refactoring tool that allows for quick and easy transformation of affine loops to execute on GPUs. Compared to previous approaches, our refactoring approach interactively combines the user’s knowledge with that of an automated parallelizer to produce parallel CUDA GPU code. The generated code retains the structure of the original loop in order to remain maintainable. The refactoring tool also computes and displays profitability metrics, intended to advise the user of the performance potential of the generated code. Keywords-I.