CASIA OpenIR  > 毕业生  > 硕士学位论文
GPU中流多处理器设计技术
其他题名Design Techniques of Streaming Multiprocessors in GPU
王骏也
学位类型工程硕士
导师李立健
2015-05-25
学位授予单位中国科学院大学
学位授予地点中国科学院自动化研究所
学位专业计算机技术
关键词Gpu 流多处理器 流水线 指令系统 仿真验证 Gpu Streaming Multiprocessors Pipelining Instruction Set Simulation
摘要GPU(Graphic Processor Unit)是一种拥有几何编程能力和像素处理能力的图像芯片。流多处理器是GPU中最基本的运算单元,是GPU的核心部分,流多处理器的设计关系到 GPU 的整体计算性能。在航空航天应用中,遥感图像处理具有计算量大、数据量大的问题,在FPGA中设计高速低功耗的流多处理器组成高性能的 GPU 软核来加速遥感图像的处理具有重要意义和潜在的应用价值。本文针对GPU中流多处理器的设计技术开展研究,主要内容如下: (1)对 GPU 体系结构的发展历史和国内外研究现状进行了综述,介绍了GPU 体系结构特点、开发平台及主要应用领域。结合遥感图像并行化处理的迫切需求,论述了研究 GPU 中流多处理器体系结构的重要意义。深入分析了NVIDIA G80/GT200 和 AMD R600 的GPU体系结构,并对 NVIDIA PTX 指令集和AMD R600 指令集进行了总结和分析。 (2)在NVIDIA G80 系列GPU 中流多处理器架构的基础上,设计了一种流多处理器的基本结构,包括流水线的设计和存储系统的设计。对流多处理器执行指令时遇到的warp分歧问题和线程切换问题进行了深入的研究及实现。 (3)设计了一种流多处理器的基本指令系统。 给出了指令系统的设计文档,介绍了指令系统中的指令编码格式、寻址方式、寄存器的定义、程序状态字及条 件码的定义。对各类指令的格式、编码和功能进行了详细地解释。 (4)对流多处理器进行了仿真验证,并对仿真波形进行了详细地解释与分析。仿真波形结果表明,流多处理器运行正常,设计方法正确可行,取得了预期的结果,成功实现了各种指令的功能,解决了 warp分歧和线程切换等关键问题。
其他摘要GPU is an image processing chip with geometric programming and pixel processing abilities. A Streaming Multiprocessors is the basic arithmetic unit in GPU. The design of Streaming Multiprocessors is crucial to the overall performance of GPU. In the application of aerospace, processing the remote sensing image is very time-consuming due to its nature as being computationally intensive. Consequently, it is very important and potential to design a GPU soft core consisting of Streaming Multiprocessors that has high performance and low power consumption in FPGA. This thesis carry out research on the design techniques of Streaming Multiprocessors in GPU. The main contents of this thesis are listed as follows: Firstly, this thesis give a comprehensive introduction of the history and recent advancement on the research of GPU architecture. Specifically, the characteristics of the architecture, the development platform for GPU programming and the main application areas of GPU are introduced. According to the urgent demands for parallelly processing of remote sensing image, this thesis address the significance of the research on the architecture design for Streaming Multiprocessors in GPU. Moreover, in-depth analysis of the architecture for NVIDIA G80/GT200 and AMD R600 are provided. Subsequently, the instruction set of NVIDIA PTX and AMD R600 are summarized and analyzed. Secondly, a basic architecture design of Streaming Multiprocessors in GPU based on the GPU of NVIDIA G80 is given, including the design of pipelining and storage system. In-depth research and thorough realization are conducted to solve the problem of warp divergence and thread switch when instructions are execuated in the Streaming Multiprocessors. Thirdly, a basic instruction set for the Streaming Multiprocessors is designed. The design document of the instruction set is also provided. Detailed explanation are given on the encoding format, the addressing mode, the definitions of registers, the definitions of the program status word and the condition code of the instruction set. Also, the format, the encoding and the function of all instructions are explained. Finally, to verify the function of the designed Streaming Multiprocessors, simulation of the designed Streaming Multiprocessors is conducted. The simulation waveforms are detailed explains and analysis. The simulation results show that the Streaming Multiprocessors run smoothly and the design approach i...
其他标识符2012E8014661096
语种中文
文献类型学位论文
条目标识符http://ir.ia.ac.cn/handle/173211/7746
专题毕业生_硕士学位论文
推荐引用方式
GB/T 7714
王骏也. GPU中流多处理器设计技术[D]. 中国科学院自动化研究所. 中国科学院大学,2015.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
CASIA_2012E801466109(2762KB) 暂不开放CC BY-NC-SA请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[王骏也]的文章
百度学术
百度学术中相似的文章
[王骏也]的文章
必应学术
必应学术中相似的文章
[王骏也]的文章
相关权益政策
暂无数据
收藏/分享
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。