Tiling for GPU Environments

Abstract

Graphics processing units (GPUs) are a good platform for solving large numerical problems, as they support simultaneous execution of thousands of threads. Porting applications on to them is an active area of research. In this research, we port a matrix multiplication application to a GPU with emphasis on good tile size. The research indicates that tile size should be judiciously chosen in keeping with the problem size. We also propose a new iterative algorithm to solve NP-complete subset sum problem and port it to a GPU architecture as well. The algorithm is very promising and it gives, for instance, an average performance improvement of 0.8% over a standard CPU implementation of existing recursive version.