| Computer vision commonly used image processing technology for image preprocessing and feature extraction,its purpose is to eliminate irrelevant information in the image,has been widely used in industry,medicine,military and other fields.Image edge detection algorithm is based on the edge of the image to identify the edge part of the image.this thesis selects the two most widely used image edge detection algorithms for parallel research,namely Sobel edge detection and Canny edge detection algorithm.Embedded system and CUDA architecture have been adapted to such image edge detection algorithm,but the deployment of such image processing algorithm by domestic DCU accelerators is still blank.To realize this kind of image edge detection algorithm efficiently on domestic DCU platform and expand the ecosystem of domestic DCU accelerators,Sobel edge detection algorithm and Canny edge detection algorithm are taken as examples to transplant and optimize these algorithms for DCU platform.The main work of this thesis is as follows:(1)The hardware structure,operation support environment and bottom tool chain of DCU accelerator are studied,and the characteristics and advantages of the target hardware platform are mastered.The transplantation of Sobel edge detection algorithm and Canny edge detection algorithm on DCU platform is realized and the correctness of the results is verified.(2)A parallel image edge detection algorithm for DCU platform is proposed.Taking Sobel edge detection algorithm and Canny edge detection algorithm as examples,the parallelism of their serial algorithm is analyzed and the parallelization program segment is found.Coarse-grained parallelism is realized by one-to-one correspondence between thread blocks and sub-image blocks,and fine-grained parallelism is realized by one-to-one correspondence between thread blocks and pixel points in image.Sobel edge detection and Canny edge detection algorithm are parallel rewritten based on DCU platform.(3)Combined with the hardware characteristics of DCU target platform,a parallel optimization method of image edge detection algorithm for DCU platform is proposed.The acceleration ratio is improved by using compiler optimization and Open MP parallel optimization.To improve the data locality and pipeline execution efficiency,loop unrolling technology is used to unroll inner loop of core computing part.The delay between processor and memory is hidden by data prefetch technology to improve the time and space locality of program.Cache hit ratio is improved by hiding global memory latency and using texture memory.Experimental results show that the proposed method effectively improves the parallel execution efficiency of Sobel edge detection and Canny edge detection.For DCU target platform,3*3,5*5,and 7*7 convolution kernels are used as examples to input images of different sizes for experiments.The results show that the overall execution efficiency of Sobel edge detection algorithm on DCU platform is improved by 1.398～2.712 times.The overall execution efficiency of Canny edge detection algorithm is improved by 1.342～2.561 times. |
