A Compiler Design for a Programmable CNN Accelerator

Convolutional Neural Networks (CNNs) are widely used in many AI applications, such as image classification, target detection, and target tracking. Due to the increase of CNN computational complexity, hardware acceleration is necessary for inference. Programmable accelerators are promising because of their support for different CNN models. To program an existing programmable accelerator, dedicated instructions need to be generated. In this paper, a compiler is designed to generate the instructions. The compiler explores the best partition of CNN models, schedules the sequence of computing, and generates the instructions automatically. With the proposed compiler, the instruction-driven CNN accelerator achieves the throughput varied from 114 FPS (ResNet152) to 1130 FPS (AlexNet).