AIRV: Enabling Deep Learning Inference on RISC-V.

Recently the emerging RISC-V instruction set architecture (ISA) has been widely adopted by both academia and industry. Meanwhile, various artificial intelligence (AI) applications have been extensively deployed in cloud, edge, mobile and IoT devices due to latest breakthroughs in deep learning algorithms and techniques. Therefore, there is an increasing need for enabling deep learning inference on RISC-V. However, at present mainstream machine learning frameworks have not been ported to RISC-V, which poses challenges to deep learning application developers. In this paper, we explore approaches to enabling deep learning inference on RISC-V. Experimental results show that in our work, there is a great gap between the performance of deep learning inference on RISC-V and that on x86; thus compared with direct compilation on RISC-V, cross-compilation on x86 is a better option to significantly improve development efficiency.