Realization and optimization of Sobel edge detection algorithm for Domestic DCU accelerators

Konferenz: MEMAT 2022 - 2nd International Conference on Mechanical Engineering, Intelligent Manufacturing and Automation Technology
07.01.2022 - 09.01.2022 in Guilin, China

Tagungsband: MEMAT 2022

Seiten: 7Sprache: EnglischTyp: PDF

Autoren:
Zhang, Hao; Ma, Ke (School of Information Engineering, Zhengzhou University, China)
Han, Lin; Xie, Jingming (National Supercomputing Center in Zhengzhou, Zhengzhou, China)

Inhalt:
Sobel edge detection algorithm is one of the commonly used methods in image segmentation algorithm based on boundary, but the deployment of this kind of image processing algorithm is still blank in domestic DCU accelerators. To efficiently implement Sobel edge detection algorithm on DCU multi-core platform, a parallel Sobel edge detection algorithm for DCU platform is designed. Based on the domestic DCU accelerators, the Sobel algorithm was parallelized, its granularity was divided, and its parallel implementation was heterogeneous. The parallel execution efficiency of Sobel algorithm was improved by hiding the global memory delay and using texture memory to improve the cache hit ratio. The results show that the parallel algorithm on DCU platform achieves the acceleration ratio of 1.32~2.22 compared with the serial algorithm and achieves the parallelization of Sobel edge detection algorithm efficiently on DCU platform.