A Tool for Automatically Suggesting Source-Code Optimizations for Complex GPU Kernels

Abstract- Future computing systems, from handhelds to su-percomputers, will undoubtedly be more parallel and heter-ogeneous than today’s systems to provide more performance and energy efficiency. Thus, GPUs are increasingly being used to accelerate general-purpose applications, including applications with data-dependent, irregular control flow and memory access patterns. However, the growing com-plexity, exposed memory hierarchy, incoherence, heteroge-neity, and parallelism will make accelerator-based systems progressively more difficult to program. In the foreseeable future, the vast majority of programmers will no longer be able to extract additional performance or energy-savings from next-generation systems because the programming will be too difficult. Automatic performance analysis and optimi-zation recommendation tools have the potential to avert this situation. They embody expert knowledge and make it avail-able to software developers when needed. In this paper, we describe and evaluate such a tool. It quantifies performance characteristics of GPU code through profiling, employs ma-chine learning models to estimate the suitability and benefit of several known source-code optimizations, ranks the opti-mizations, and suggests the most promising ones to the user if the expected speedup is sufficiently high. 1