| The mainstream type of high-performance computing architectures at present is CPU/GPU and CPU/MIC heterogeneous cooperative computing system, which provides a powerful computing capability, versatility and effectiveness. However, because of the complexity of hardware architecture and the specialty of programming model, user programming and performance optimization has become a key in overall system performance.As the growing popular of high-resolution images, the speed of edge detection has become more and more important in the follow-up study. In this paper, we focus on the problem that how to maximize the speed of edge detection and the efficiency of the code in particular heterogeneous architecture system platform.In this paper, we studied the combination of edge detection algorithm and GPU/MIC, at the same time, we had an in-depth research on the edge detection algorithm and the CPU/GPU/MIC heterogeneous architectures. First of all, this paper reviewed the development of GPU hardware, CUDA programming model and other related knowledge, analyzed the parallelism of edge detection algorithm and implemented it on CPU and CPU/GPU. We made the processing speed faster through a series of optimization methods like loop unrolling, instruction optimization and the shared memory access optimization. By comparing with CPU/MIC heterogeneous architecture horizontally, we analyzed the calculation of CPU/GPU's characteristic. Based on hardware configuration, this paper presented a parameter selecting method to make the program intelligently select operating parameters. After the analysis of the Gaussian filter and Roberts edge detection of CUDA program, the integration of the Kernel function had being used to accelerate the speed of the program. Finally, by the study of Kirsch algorithm, we proposed an improvement plan, and the improvement plan been realized and optimized on GPU. The effectiveness of the improvement plan is verified by experimental results. |
PDF Full Text Request