Automatic Program Analysis For Data Parallel Kernels

Date: 2011-06-30

Time: 11:00

Room: BBL 165

Speaker: Calin Juravle

Title: Automatic Program Analysis for Data Parallel Kernels

Abstract

It is widely known that GPUs have more computational power and expose a far greater level of parallelism than conventional CPUs. Despite the high potential, GPUs are not yet a popular choice in practice, mainly because of their high programming complexity. The aim of this thesis is to explore automatic program analyses that will enable automatic or guided transformation of programs from sequential versions to data parallel GPU kernels. Our main contributions consist of identification and implementation of key automatic program analyses that enable such transformations. The work serves as a foundation for an automatic GPU parallelization system.

We note that no knowledge about GPU architectures or GPU programming models is required to follow this presentation. But, because of the time constraints, some basic knowledge about static program analysis (i.e. abstract interpretation and monotone frameworks) is assumed. Although not essential to get the overall picture, this knowledge will be helpful to understand the underlying analysis framework.