Research on Full-Chip Programming for Sunway Heterogeneous Many-core Processor
Wei Wu, Hong Qian, Qi Zhu, Jue Wang, XingJian Fan · 2021
Programming on many-core processors is a challenging task. It's a difficult topic to program and compile on heterogeneous many-core architectures in high-performance computing area. The bottom-level programming support on Sunway many-core processors is insufficient and can hardly satisfy the growing need in applications. This paper conducts an thread accelerated programming model and a multi-mode accelerating thread library on Sunway architectures. A fast thread group mode is proposed, which can reduce the consumption of thread spawn and join according to hardware's features. Compared to GPU, it achieves a speedup of 24.88. We also designs a globally sharing executing mode, which supports many-core acceleration programming from the view of full-chip. Evaluation results on the Parboil benchmark shows that, the average performance is 1.16 times as NVIDIA V100 GPU.