A Cycle-Accurate, Execution-Driven GPU Simulation Framework

• Main Authors: Lin-Chieh Shangkuan, 上官林傑
• Other Authors: Chia-Lin Yang
• Format: Others
• Language: en_US
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/52983893521463807432

碩士 === 國立臺灣大學 === 資訊工程學研究所 === 94 === Graphics processing unit (GPU) is designed for accelerating the graphics rendering manipulations. Their highly-parallel structure makes them more effective than CPUs for a range of graphics rendering algorithms. Modern GPUs become increasingly hard to evaluate because it needs to support more complex funcionts and the architecture details are not released by the GPU vendors. To study the GPU design, this thesis proposes a cycle-accurate, execution-driven GPU simulation framework. In this framework, the GPU simulator core is modeled as a pipelined processor and there is also a detailed timing-model of memory system within it for more accurate simulation. The GPU simulator executes rendering commands that are converted from the stream of OpenGL function calls and simulates the behaviours in a cycle-accurate fashion. The OpenGL trace is captured from real 3D games (e.g., Quake 3). To demonstrate the applicability of the framework, this thesis also introduces a study on graphics memory system. I analyze the performance effect by applying different memory access scheduling policies. The experimental results shows that an adaptive policy is the most effective.