基于CUDA的简化并行编程方案设计

发布时间：2018-02-20 03:38

本文关键词： CUDA GPGPU 并行编程内存管理多地址空间集群　出处：《大连理工大学》2013年硕士论文　论文类型：学位论文

【摘要】：NVIDIA公司设计的基于GPGPU的通用并行计算架构CUDA已经在各个领域得到广泛的应用。本文针对GPGPU设备外部的并行设计了基于CUDA的简化并行编程方案。首先,本文针对大批量数据场景提出了GPGPU设备间并行编程简化方案,以辅助本文提出的多地址空间封装技术能够独立地运行于集群环境中。此外,本文提出的GPGPU设备间并行编程解决方案引入了数据复用技术,对视频处理任务场景提供了优化。其次,本文针对CUDA并行编程中内存管理与多地址空间架构具有较高耦合度这一问题,提出了多地址空间封装技术来简化CPU与GPGPU设备间的并行编程。基于多地址空间封装技术,本文提出了源码转换技术,以进一步简化CUDA编程中的内存管理,允许编程者对于内存相关变量采取只声明不维护的编码方式。在实验中,本文将多地址空间封装技术与CUDA的统一虚拟寻址技术进行了对比,实验结果表明：当计算所需数据位于其它设备中时,在执行访存密集型任务时,多地址空间封装技术占有性能优势；在执行计算密集型任务时,统一虚拟寻址技术占有性能优势。最后,本文还通过从NPP库中选取的三个具有代表性的算法再次验证了上述结论,并在实验中测量了多地址空间封装技术的额外开销,实验结果表明其额外开销仅为0.04%至0.07%。
[Abstract]:CUDA, a general parallel computing architecture based on GPGPU, which is designed by NVIDIA Company, has been widely used in various fields. In this paper, a simplified parallel programming scheme based on CUDA is designed for the concurrent design of GPGPU devices. In this paper, we propose a simplified scheme of parallel programming between GPGPU devices for mass data scenarios, which can assist the multi-address space encapsulation technology proposed in this paper to run independently in the cluster environment. The parallel programming solution between GPGPU devices proposed in this paper introduces the technology of data reuse, which provides the optimization of the video processing task scene. In order to solve the problem of high coupling between memory management and multi-address space architecture in CUDA parallel programming, a multi-address space encapsulation technique is proposed to simplify the parallel programming between CPU and GPGPU devices. Based on the multi-address space encapsulation technology, this paper proposes the source code conversion technology to further simplify the memory management in CUDA programming, allowing the programmer to adopt the coding method of declaring and not maintaining the memory-related variables. In this paper, the multi-address space encapsulation technology is compared with the unified virtual addressing technology of CUDA. The experimental results show that when the computing data is located in other devices, the memory access intensive task is performed. The multi-address space encapsulation technology has the performance advantage, and the unified virtual addressing technology has the performance advantage when performing computation-intensive tasks. Finally, This paper also verifies the above conclusions again through three representative algorithms selected from the NPP library, and measures the extra cost of the multi-address space encapsulation technology in the experiment. The experimental results show that the extra cost is only 0.04% to 0.07.
【学位授予单位】：大连理工大学
【学位级别】：硕士
【学位授予年份】：2013
【分类号】：TP338.6

【相似文献】