Synthesising graphics card programs from DSLs

Abstract

Over the last five years, graphics cards have become a tempting target for scientific computing, thanks to unrivaled peak performance, often producing a runtime speed-up of x10 to x25 over comparable CPU solutions. However, this increase can be difficult to achieve, and doing so often requires a fundamental rethink. This is especially problematic in scientific computing, where experts do not want to learn yet another architecture. In this paper we develop a method for automatically parallelising recursive functions of the sort found in scientific papers. Using a static analysis of the function dependencies we identify sets - partitions - of independent elements, which we use to synthesise an efficient GPU implementation using polyhedral code generation techniques. We then augment our language with DSL extensions to support a wider variety of applications, and demonstrate the effectiveness of this with three case studies, showing significant performance improvement over equivalent CPU methods, and similar efficiency to hand-tuned GPU implementations.