Data multiplexing method based on operator fusion on heterogeneous many-core architecture

The invention discloses a data multiplexing method based on operator fusion on a heterogeneous many-core architecture, and the method comprises the steps: carrying out the function fusion of at least two operators A and B which are sequentially called in a DNN operator library, obtaining a fusion operator C, and executing the following operations: S1, reading data from a main memory to a local memory by the fusion operator C, and taking the read data as the input of the operator A; s2, the operator A takes the obtained data as input, corresponding operation is carried out, function calculation of the operator A is completed, and at the moment, the operator A keeps a result in a local memory and does not write the result back to a main memory; s3, the operator A transmits the calculation result in the local storage to an operator B to serve as the input of the operator B; s4, the operator B takes the data from the operator A as input; s5, after the operator B completes the operation, writing a final calculatio