An LSTM Acceleration Method Based on Embedded Neural Network Accelerator

With the maturity of neural network technology, chips to accelerate neural network inference are emerging endlessly. Faced with the emerging complex neural network operators (such as LSTM) that are constantly evolving in neural network algorithms, it is unrealistic to modify the hardware design of the neural network inference chip to support the evolving new operators. Therefore, it has important research significance and practical value to make existing hardware support new operators through software. We propose an LSTM acceleration method based on an embedded neural network accelerator. Split the LSTM operator into multiple basic operators supported by the neural network accelerator by software, and optimize it. Finally, the embedded neural network accelerator supports LSTM operators quickly and efficiently. Experimental results show that the execution efficiency of LRCN model deployed on a low-power accelerator is x1.6 and X1.3 higher than that on CPU and GPU, respectively.