CONVOLUTIONAL NEURAL NETWORK ACCELERATOR-BASED DATA REUSE METHOD

摘要

Disclosed is a convolutional neural network accelerator-based data reuse method, comprising: a tile scan module receives command information of a command module, the command information comprising a size of tile blocks for dividing a CNN operation; the tile scan module generates coordinates of tile blocks according to the size of the tile blocks, and sends the coordinates to an internal storage request module; the internal storage request module generates an internal storage read request, and sends the internal storage read request to an internal storage module; the internal storage module sequentially returns tile block data to an input activation weight cache unit, and the input activation weight cache unit saves the received tile block data so as to implement data reuse, and send the received tile block data to computing processing units (PEs).