GPU上基本图像处理算法性能优化关键技术研究

发布时间：2019-02-15 04:39

【摘要】：图像处理主要包括图像压缩、图像滤波、图像采样、图像分割和图像分析等,目前在众多实际应有领域都有重要应有,如目前应用广泛的图像识别,人脸识别等,其技术核心均是图像处理问题。随着这些领域图像处理规模的不断扩大以及对实时性能要求的不断提高,如何提高图像处理算法的性能已经成为当前的研究热点。GPU(Graphics Processing Units)在处理能力和存储带宽上相对CPU都有无可比拟的优势,其发展为图像处理应用的实时性要求提供了解决方案。鉴于OpenCL的跨平台特性,本文将使用OpenCL在GPU上实现并行图像处理的相关操作,可大幅提升图像处理算法性能,对于前面提到的相关问题无疑是一个很好的解决方法。同时,图像处理算法通常具有数据量大以及计算访存密集的特点,在GPU上并行化处理是切实可行的解决方案。因此,本文将针对图像处理算法在GPU上的实现以及优化方法进行研究。由于GPU架构的复杂性和硬件资源的限制,性能优化成为GPU编程的难点和重点,GPU优化的本质是实现算法特征向底层硬件架构特征的高效映射。本文将结合图像处理算法特性和底层硬件架构特征开展图像处理算法在GPU计算平台上的性能优化研究。本文针对常用的图像处理算法类型进行研究,包括上采样、下采样、归约、水平滤波、垂直滤波、卷积,过冲控制等。由于这些算法分别具有不同的计算访存特征,因此,本文将结合GPU硬件平台特性,从数据传输优化、访存优化、NDRange优化、指令流优化、数据共享优化和数据相关优化等角度总结具有不同特征的图像处理算法在GPU上的性能瓶颈和优化方法。本文的主要工作如下:1)Sharpness综合图像处理算法在GPU计算平台上的实现和优化。对Sharpness中所包含的基本图像处理算法进行了算法分析和并行性分析。对GPU上的Sharpness算法采取的优化方式包括:数据传输优化,kernel融合,归约优化,向量化及数据本地化优化,边界优化和其他基本优化方法。同时研究了Sharpness算法在SIMD上的优化。对比分析了Sharpness算法CPU版本的性能、SIMD优化后版本的性能以及GPU优化后版本的性能。2)同时也分析了Laplacian综合图像处理算法,对Laplacian中所包含的基本图像处理算法进行了算法分析和并行性分析。对GPU上的Laplacian算法采用的优化方法包括:kernel融合、减少全局同步并精简算法,添加padding、减少条件判断并解决数据对齐问题。同时介绍了Laplacian算法在SIMD上的优化。对比分析了Laplacian算法CPU版本的性能、SIMD优化后版本的性能以及GPU优化后版本的性能。实验结果可以看出,GPU加速图像处理算法具有无可比拟的优势,同时由于GPU硬件架构的特性,为移植到GPU上的算法进行针对硬件架构的优化对于性能有较大影响。
[Abstract]:Image processing mainly includes image compression, image filtering, image sampling, image segmentation and image analysis. The core of its technology is image processing. With the increasing scale of image processing in these fields and the increasing demand for real-time performance, How to improve the performance of image processing algorithm has become a research hotspot in the current research. GPU (Graphics Processing Units) has an unparalleled advantage over CPU in terms of processing capacity and storage bandwidth. It provides a solution for the real-time application of image processing. In view of the cross-platform characteristics of OpenCL, this paper will use OpenCL to implement parallel image processing operations on GPU, which can greatly improve the performance of image processing algorithm. It is undoubtedly a good solution to the related problems mentioned above. At the same time, image processing algorithms usually have the characteristics of large amount of data and dense computing access. Parallel processing is a feasible solution on GPU. Therefore, this paper will focus on the implementation of image processing algorithm on GPU and the optimization method. Because of the complexity of GPU architecture and the limitation of hardware resources, performance optimization has become the difficulty and focus of GPU programming. The essence of GPU optimization is to realize the efficient mapping between algorithm features and underlying hardware architecture features. In this paper, the performance optimization of image processing algorithm on GPU computing platform is studied based on the characteristics of image processing algorithm and underlying hardware architecture. In this paper, the types of image processing algorithms are studied, including up-sampling, down-sampling, reduction, horizontal filtering, vertical filtering, convolution, overshoot control and so on. Because these algorithms have different computational memory access characteristics, this paper will combine the characteristics of GPU hardware platform, from data transmission optimization, memory access optimization, NDRange optimization, instruction flow optimization, The performance bottlenecks and optimization methods of image processing algorithms with different characteristics on GPU are summarized from the point of view of data sharing optimization and data correlation optimization. The main work of this paper is as follows: 1) the realization and optimization of Sharpness synthetic image processing algorithm on GPU computing platform. The algorithm analysis and parallelism analysis of the basic image processing algorithm included in Sharpness are carried out. The optimization methods of Sharpness algorithm on GPU include: data transmission optimization, kernel fusion, reduction optimization, vectorization and data localization optimization, boundary optimization and other basic optimization methods. At the same time, the optimization of Sharpness algorithm on SIMD is studied. The performance of the CPU version of the Sharpness algorithm, the performance of the optimized version of SIMD and the performance of the optimized version of GPU are compared and analyzed. 2) at the same time, the comprehensive image processing algorithm of Laplacian is also analyzed. The algorithm analysis and parallelism analysis of the basic image processing algorithm included in Laplacian are carried out. The optimization methods for Laplacian algorithm on GPU include: kernel fusion, reducing global synchronization and reducing algorithm, adding padding, to reduce conditional judgment and solving the problem of data alignment. At the same time, the optimization of Laplacian algorithm on SIMD is introduced. The performance of CPU version of Laplacian algorithm, the performance of optimized version of SIMD and the performance of optimized version of GPU are compared and analyzed. The experimental results show that the GPU accelerated image processing algorithm has unparalleled advantages. At the same time because of the characteristics of the GPU hardware architecture the optimization of the hardware architecture for the algorithm transplanted to GPU has a great impact on the performance.
【学位授予单位】：中国科学院大学(中国科学院工程管理与信息技术学院)
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】