Abstract: A compiler system for deploying CNN models to FPGA-based high-performance accelerators is provided. The compiler system comprises a compiler front end and a compiler back end, as well as a runtime library and an accelerator; the compiler front end is used for the quantization of CNN models to be deployed based on the ONNX model and a training data set, and the IR Graph corresponding to the CNN model to be deployed is obtained; the compiler back end is used to determine instructions and configuration information; the instruction set architecture comprises CONY, MAXP, AVGP, ADD and FC, and the configuration information comprises pre-trained weights and bias, packet descriptors and key-value pairs representing control registers; the runtime library is used to store instructions and configuration files to DRAM, and FPGA is configured according to the configuration information of the model.