The Performance Optimization Of Parallel Image Processing Based On CUDA

With the rapid development of computer technology, digital image processing technology can exert a powerful influence on astronomy, biology, nuclear medicine, law implementation, technology for national defense and engineering manufacture. In real world, new demands about image processing technology are proposed, for example, processing and real-time processing on large-scale image data. In order to meet the demands, researches integrate the image processing with parallel computing technique and propose parallel image processing technology.There are various platforms for parallel computing, for example, compute unified device architecture (CUDA) designed by NVIDIA is widely due to the strong computing power of GPU (graphic processing unit) for realizing general parallel computing. CUDA is becoming a hot research topic in image processing.This paper focuses on the performance optimization strategies based on CUDA, implements parallel image processing using those methods, and then achieves the best optimal performance under certain conditions.Firstly the existing technologies of parallel computing and parallel image processing are summarized, CUDA programming model and the architecture of hardware and software are introduced. Secondly a number of optimization strategies are discussed to improve the algorithms’performance based on CUDA architecture, followed by theory analysis. After that a new programming framework named GPIP (GPGPU Image Processing) is introduced, the key technology in parallelizing image processing is also given. Take erosion, dilation, open and close operation (which are belong to morphology) for example, the parallelization on CUDA and optimization performance of the above algorithms are achieved according to the optimization methods proposed in the paper.Based on above work, we did lots of experiments, the results shown that the optimization methods proposed in this paper are helpful in parallelizing image processing algorithms based on CUDA architecture, and achieving best optimal performance under certain conditions.