Biography
Enrollment Date: 2009
Graduation Date：2012
Degree：M.S.
Defense Date：2012.05.24
Advisors：Yangdong Deng
Department：Institute of Microelectronics，Tsinghua University
Title of Dissertation/Thesis：Polyhedral Modeling Based CUDA Source-to-Source Optimization Framework
Abstract:
In this paper, we propose a model-driven source-to-source code optimization framework for general purpose computing on graphics processing units (GPGPU). Our framework is based on a re-formulation of the polyhedral loop transformation theory under the context of GPGPU. A performance model is developed by simultaneously thread level parallelism, instruction level parallelism, and memory level parallelism. The model can be analytically derived from a GPU program’s polyhedral model. Then affine transformations can be applied to an initial parallelized solution under the guidance of the performance model. The experiment results demonstrate the effectiveness of our work. On average, the code generated by our work outperforms a leading GPGPU compiler and NVIDIA handcrafted CUBLAS 4.0 by 22% and 17%, respectively.
We summarize our contributions as follows: 1) A model-driven automatic source-to-source optimization framework for GPGPU is developed to optimize the performance of GPGPU programs. Under the guidance of a performance model, we formulate a few GPGPU optimization problems under the polyhedral framework and propose efficient solutions. 2) We propose a performance model by simultaneously considering thread level parallelism, instruction level parallelism, and memory level parallelism. The performance allows tradeoffs among different optimization approaches for improved overall performance. Experimental results prove the effectiveness of the proposed metric. 3) The proposed optimization framework is applied to a set of extensively used CUDA kernels and observe superior or comparable performance to handcrafted code and CUBLAS 4.0.