An Infrastructure for Tackling Input-Sensitivity of GPU Program Optimizations |
| |
Authors: | Xipeng Shen Yixun Liu Eddy Z. Zhang Poornima Bhamidipati |
| |
Affiliation: | 1. Computer Science Department, College of William and Mary, Williamsburg, VA, USA 2. Radiology and Imaging Sciences, Clinical Center, National Institutes of Health, 10 Center Drive, Building 10, Rm 1C515, Bethesda, MD, 20892-1182, USA 3. Department of Computer Science, Rutgers, The State University of New Jersey, New Brunswick, NJ, 08901, USA 4. Capital One, Williamsburg, VA, 23185, USA
|
| |
Abstract: | Graphic processing units (GPU) have become increasingly adopted for the enhancement of computing throughput. However, the development of a high-quality GPU application is challenging, due to the large optimization space and complex unpredictable effects of optimizations on GPU program performance. Many recent efforts have been employing empirical search-based auto-tuners to tackle the problem, but few of them have concentrated on the influence of program inputs on the optimizations. In this paper, based on a set of CUDA and OpenCL kernels, we report some evidences on the importance for auto-tuners to adapt to program input changes, and present a framework, G-ADAPT+, to address the influence by constructing cross-input predictive models for automatically predicting the (near-)optimal configurations for an arbitrary input to a GPU program. G-ADAPT+ is based on source-to-source compilers, specifically, Cetus and ROSE. It supports the optimizations of both CUDA and OpenCL programs. |
| |
Keywords: | |
本文献已被 SpringerLink 等数据库收录! |
|