Parallel Image Processing on the Sunway Many-Core Processor

Abstract

Many image processing algorithms have been parallelized successfully on many-core processors, such as GPU and Intel Xeon Phi. In this paper, we choose the Sunway many-core processor SW26010, which is a new processor designed and made in China that constitutes the current NO.1 supercomputer Sunway TaihuLight. This paper firstly introduces the architecture of Sunway SW26010 processor and two representative image processing algorithms: local binary pattern (LBP) and histogram of oriented gradient (HOG). Furthermore we propose a method of parallel implementation, and the experimental results of this method show that the speedup can be up to 170 for LBP and 33 for HOG. Then two optimized methods are brought forward based on this parallel implementation, including the optimization of program and parallel design. We optimize the program by using the method that combined step transmission and software prefetching. From the experiment results of the first optimization we can know that the maximum speedup can reach 310 for LBP with processing high-resolution images and 83 for HOG. Then we optimize the parallel design by using a coarse grained parallel method, and the experimental results show that the speedup can be up to 370 for LBP and 95 for HOG when processing low-resolution images. Finally, we investigate the scalability of our parallelism on the Sunway TaihuLight with different number of processor nodes, and the experiment results prove that the two algorithms' parallel design and implementation have better expansibility.