Optimizing Deep Learning Inference via Global Analysis and Tensor Expression

Nowadays, a large number of accelerators are proposed to increase the performance of AI applications, making it a big challenge to enhance existing AI programming frameworks to support these new accelerators. In this paper, we select TensorFlow to demonstrate how to port the AI programming framework to new hardwares, i.e., FPGA and Sunway TaihuLight here. FPGA and Sunway TaihuLight represent two distinct and significant hardware architectures for considering the retargeting process. We introduce our retargeting processes and experiences for these two platforms, from the source codes to the compilation processes. We compare the two retargeting approaches and demonstrate some preliminary experimental results.

In Network and Parallel Computing