GPURoofline: A Model for guiding performance optimizations on GPUs

Abstract

Performance optimization on GPUs requires deep technical knowledge of the underlying hardware. Modern GPU architectures are becoming more and more diversified, which further exacerbates the already difficult problem. This paper presents GPURoofline, an empirical model for guiding optimizations on GPUs. The goal is to help non-expert programmers with limited knowledge of GPU architectures implement high performance GPU kernels. The model addresses this problem by exploring potential performance bottlenecks and evaluating whether specific optimization techniques bring any performance improvement. To demonstrate the usage of the model, we optimize four representative kernels with different computation densities, namely matrix transpose, Laplace transform, integral and face-dection, on both NVIDIA and AMD GPUs. Experimental results show that under the guidance of GPURoofline, performance of those kernels achieves 3.74~14.8 times speedup compared to their naïve implementations on both NVIDIA and AMD GPU platforms. © 2012 Springer-Verlag.