MN implementation method and device based on DMA architecture, equipment and storage medium

The embodiment of the invention provides an MN implementation method and device based on a DMA architecture, equipment and a storage medium, and relates to the technical field of neural networks. Effective data points of the input data and the mask data in the input MN unit are the same, so that the input data and the mask data are stored in a queue form, the effective data points in the two kinds of data can be in one-to-one correspondence according to queue digits, and standby data pairs are formed; the mask data can have a zero result and a non-zero result, so that the effective data of the input data can be screened according to the mask data, and the input data corresponding to the non-zero mask is screened as the effective data to be output. Therefore, the difficulty of sorting and screening complex data can be reduced, and the accuracy and universality of data screening and carrying can be improved. 本发明实施例提供一种基于DMA架构的MN实现方法、装置、设备及存储介质，涉及神经网络技术领域。输入MN单元中的输入数据和掩码数据的有效数据点是相同的，所以以队列形式存储输入数据和掩码数据，可以使得两种数据中的