| abstract |
The invention discloses a separable image processing neural network accelerator and an acceleration method. Through a low-speed interface module, an accelerator instruction is loaded into an instruction storage module, a control module obtains an instruction from the instruction storage module, and a control signal is obtained through an instruction decoding module. , schedule each module and cooperate with each module and the storage multiplexing interface module; the high-speed interface module respectively loads the input feature map, weight, quantization factor, bias parameter and batch normalization parameter to the corresponding storage module according to the instruction ;Convolution calculation, arithmetic logic processing, pooling operation and upsampling operation are performed according to the structure of the neural network; the calculation of each module is repeated until the calculation of all layers of the neural network is completed; according to the jump instruction, jump to the first instruction Continue to process the next input feature map. Through the combination of high and low speed interfaces, the implementation method is simple, the accuracy is reliable, each module is relatively independent, and the scalability is strong. |