Speaker: Gavin Davidson from the University of Glasgow

Title: Automatic Detection of Parallelism in Scientific Fortran using Algorithmic Skeletons and OpenCL

Abstract: General purpose graphics hardware represents a powerful and affordable tool for climate scientists who look to model the vastly complex systems that make up the Earth's atmosphere. However, levering these highly parallel devices in this field requires an understanding of tools like OpenCL and effective strategies for parallelisation. We present a source to source compiler that automatically detects parallelism in scientific Fortran code and produces programs parallelised using OpenCL. Our approach requires no directives or extra information from the user and output code is composed with use of algorithmic skeletons. We evaluate our work using a version of the large eddy simulator along with synthetic test cases and show that this approach yields performance increases with minimal effort.