Biography
Enrollment Date: 2010
Graduation Date：2012
Degree：M.S.
Defense Date：2012.05.28
Advisors：Yangdong Deng
Department：Institute of Microelectronics，Tsinghua University
Title of Dissertation/Thesis：Polyhedron-based GPU Code Translation And Optimization
Abstract:
In this thesis, the polyhedral model in the field of compilation and optimization is applied to the problem of GPGPU code optimization. Based on a highly applicable performance model for GPGPU, an entire workflow for automatic GPGPU code transformation and optimization has been proposed. First, the thesis provides a brief survey on general-purpose parallel computing, GPGPU, and polyhedral-model-based loop optimizations, and reviews several typical existing tools for the automatic code generation and optimization for GPGPU. Then based on the existing works, in accordance to the architectural traits of GPGPU, we propose a GPGPU performance model based on hardware utilization. The model predicts runtime efficiency of kernels using thread-level parallelism and memory overhead with good accuracy, and serves as solid foundation for the automatic code transformation and generation techniques. Lastly, using the polyhedral representation as theory foundation, a complete framework for the automatic optimization of GPGPU code is proposed, along with a set of guidelines and strategies for optimization. The framework has been proven effective by tests on a set of typical CUDA programs, where the result programs measured performance similar to or higher than manual-optimized versions.