A CPU-FPGA Based Heterogeneous Accelerator for RepVGG

RepVGG is a Convolutional Neural Network (CNN) that has a VGG-like inference-time body with a good performance. Compared with other devices, the FPGA accelerator has the advantages of flexibility and efficiency. In this paper, we proposed a heterogeneous CPU-FPGA accelerator for RepVGG. Specific computing kernels are designed for convolution layers. Weight and bias parameters are quantized to 8 bits without a large loss of accuracy. The memory access performance and computing performance of each layer of the network are also optimized. The accelerator achieves 758 Gop/s. Compared with the CPU, the accelerator is more than three times faster.