Research On Key Techniques Of Basic Image Processing Algorithms’ Optimization On GPU

Image processing mainly includes image compression,filtering,sampling,segmentation,and analysis.At present,it has important applications in many fields,including the widely used image recognition,and face recognition.With the expanding scale of images and increasing real-time requirements,improve the performance of image processing algorithms has become a research hotspot.Relative to CPU,GPU(Graphics Processing Units)has a great advantage in computing power and memory bandwidth.The development of GPGPU(General Purpose GPU)provides a solution to the increasing real-time requirements of image processing applications.Because of the cross-platform features of OpenCL,this thesis will use OpenCL to parallelize the image processing algorithm on GPU.Parallel optimization can significantly improve the performance of image processing algorithms,and thus accelerate the development of many fields.Meanwhile,the image processing algorithms are typically compute-intensive,memory-intensive,and have good parallel architecture.Therefore,it is suitable for image processing algorithms to implement on GPU.In this thesis,we will conduct research on image processing algorithms’ GPU implementation and the optimization methods on GPU.Due to the complex architecture and limitation of hardware resources on GPU,optimization of performance becomes the emphasis and difficulty of GPU programing.The essence of GPU program optimization is achieving the efficient mapping of algorithm features and hardware architecture features.The performance optimization in this thesis will be based on this theory.This thesis will focus on common image processing algorithms,including up-scale,down-scale,reduction,horizontion filter,vertical filter,convolution and overshoot control.These algorithms have different calculation and memory access characteristics.Therefore,this thesis will analyze the difference between different algorithms in calculation bottleneck and optimization methods,based on the GPU platform features.The optimizations include data transfer optimization,memory access optimization,NDRange optimization,instruction stream optimization,data sharing optimization and data related optimization.The major works of this thesis: 1)The implement and optimization of Sharpness,a synthesized image processing algorithm,on GPU.Analyze features and parallelism of the basic image processing algorithms in Sharpness.The adoptive optimizations in Sharpness are: Data Transfer Optimization,Kernel Fusion,Reduction Optimization,Vectorization,Data Locality Optimization,Border Optimization,etc.Meanwhile,this thesis also implements the SIMD optimization on CPU.Finally,the performance of basic CPU implement,SIMD optimized CPU implement,and GPU implement are compared.2)The implement and optimization of Laplacian,a synthesized image processing algorithm,on GPU.Analyze features and parallelism of the basic image processing algorithms in Laplacian.The adoptive optimizations in Laplacian are: Kernel Fusion,Reduce Global Synchronization,brief computation,Padding,Reduce conditional statement,and alignment of data.Implement the SIMD optimization on CPU.Finally,the performance of basic CPU implement,SIMD optimized CPU implement,and GPU implement are compared.Accelerating image processing algorithms by GPU can gain a significant performance improvement.Besides,the optimization methods,which should be based on the GPU platform features,have remarkable effects on performance.