Adaptive mapping and parameter selection scheme to improve automatic code generation for GPUs.

Abstract

Graphics Processing Units (GPUs) are today’s most powerful coprocessors for accelerating massive data-parallel algorithms. However, programmers are forced to adopt new programming paradigms to take full advantage of their computing capabilities; this requires significant programming and maintenance effort. As a result, there is an increasing interest in the development of tools for automatic mapping of sequential code to GPUs. Current automatic tools require both a deep knowledge on the GPU architecture and the algorithm being mapped, which makes the mapping process a labor-intensive task. This paper proposes a technique that improves the code mapping of one of these tools, PPCG, removing the need for any user interaction. It relies on data reuse estimations to explore the mapping space and compute appropriate values for the number of threads per threadblock and tile sizes. Our results show speedups of 3x on average compared to the default code generated by PPCG.