FlexCNN: Design of a Universal and Easy-to-Deploy CNN Acceleration System

The widespread application of Convolutional Neural Networks (CNNs) has driven the development of dedicated accelerators. However, existing CNN accelerators still face two major challenges in practical applications: limited hardware versatility to support diverse convolution and activation operators, and high barriers to software-hardware collaborative deployment. To address these issues, this paper proposes FlexCNN, a CNN acceleration system that is both highly universal and convenient for deploying models. The system features a hardware architecture incorporating an image to column based general matrix multiplication (im2col-GEMM) unit compatible with multiple convolution types, a vector ALU supporting various activation functions, and a reconfigurable pooling unit. In addition, a customized toolchain based on the NCNN inference engine is developed to facilitate end-to-end deployment. Implemented on a Xilinx XCZU15EG FPGA, FlexCNN achieves an energy efficiency of 67.07 GOPS/W when executing YOLOv3 and demonstrates competitive performance across multiple network architectures.