Evaluating OpenCL Applications with Heterogeneous Virtual Platforms

• Main Authors: Kuan-Wen Su, 蘇冠文
• Other Authors: Shih-Hao Hung
• Format: Others
• Language: en_US
• Published: 2013
• Online Access: http://ndltd.ncl.edu.tw/handle/42410742052340427565

碩士 === 國立臺灣大學 === 資訊網路與多媒體研究所 === 101 === With a heterogeneous architecture, a modern computer systems combines multi-core CPUs and multiple accelerator cores to perform applications efficiently. As heterogeneous system architectures rapidly become the mainstream, various types of processors such as GPUs, FPGAs and DSPs, can be chosen to optimize a system for specific applications. OpenCL provides a standard framework for leveraging different types of processors as a computing device. To aid the design of heterogeneous systems and the evaluation of the OpenCL applications, we develop a heterogeneous virtual platform. Since operating system is responsible for task scheduling, file systems, I/O operations, and networking, we believe a full-system emulation is essential for evaluating data-intensive OpenCL applications. In this thesis, we describe a heterogeneous virtual platform with multi-core CPUs and GPUs for executing the OpenCL applications. Furthermore, we developed an event-driven performance analysis toolkit, PSET, for performance profiling of OpenCL applications on the virtual platform. We demonstrate how our heterogeneous virtual platform works by analyzing the OpenCL applications with the help from graphical visualization and show that our framework are useful tools for system-wide optimization for developers and architects.