Multi-row interleaved storage system and method based on mixed bit width neural network accelerator

The invention discloses a multi-row interleaving storage system and method based on a mixed bit width neural network accelerator, the storage system designs a plurality of interleaving modes according to layer fusion characteristics of adjacent convolutional layers, pooling layers, up-sampling layers, data recombination and the like, and realizes cross continuous storage of multi-row data of a feature map through an interleaving module. Extraction of interleaved feature map data and filling data is achieved through the de-interleaving module, the data are distributed to a multi-channel multi-line storage space of an input cache, addresses are continuously increased in the data reading and writing process, the burst transmission characteristics of a bus and DDR are fully utilized, high reading and writing efficiency is ensured, and the data reading and writing efficiency is improved. And the time overhead caused by frequent address switching is avoided. The idle waiting time of a calculation unit in the data l