Hardware-software co-design for accelerating deep learning inference

An artificial intelligence chip, comprising:

• an interface;
• a memory on which one or more programs are stored;
• and
• one or more processors operatively connected to the interface and the memory, configured for:
• dividing a convolution task into at least one convolution subtask and at least one corresponding pooling subtask, and determining a parallel computing unit corresponding to each convolution subtask, wherein the convolution task is configured for performing convolution, batch normalization, non-linear computation, and pooling operations on to-be-processed data of the convolution task, the convolution subtask is configured for performing convolution, batch normalization, and non-linear computation operations in the convolution task on to-be-processed data of the convolution subtask, and the pooling subtask is configured for performing a pooling operation in the convolution task on an execution result of the corresponding convolution subtask;
• transmitting each divided convolution subtask to the parallel computing unit corresponding to the convolution subtask, and controlling the parallel computing unit to execute the received convolution subtask;
• transmitting a pooling parameter of a target pooling subtask to the pooling computation unit in response to receiving indication information, which is transmitted by the parallel computing unit, indicating that the convolution subtask has been completely executed, and controlling the pooling computation unit to execute the target pooling subtask, wherein the target pooling subtask refers to a pooling subtask corresponding to the convolution subtask which corresponds to the parallel computing unit transmitting the received indication information;
• merging the execution results of the respective pooling subtasks received from the pooling computation unit to obtain an execution result of the convolution task;
• one or more parallel computing units operatively connected to the one or more processors and the memory, configured for:
• executing the received convolution subtasks, and transmitting execution results as to-be-pooled data to the pooling computation unit;
• and
• a pooling computation unit operatively connected to the one or more processors, the memory and the one or more parallel computing units, configured for:
• executing the respective pooling subtasks and transmitting execution results to the one or more processors.

A computer readable storage medium, on which a computer program is stored, wherein when the computer program is executed by one or more processors, the one or more processors are caused to:

• divide a convolution task into at least one convolution subtask and at least one corresponding pooling subtask, and determine a parallel computing unit corresponding to each convolution subtask, wherein the convolution task is configured for performing convolution, batch normalization, non-linear computation, and pooling operations on to-be-processed data of the convolution task, the convolution subtask is configured for performing convolution, batch normalization, and non-linear computation operations in the convolution task on to-be-processed data of the convolution subtask, and the pooling subtask is configured for performing a pooling operation in the convolution task on an execution result of the corresponding convolution subtask;
• transmit each divided convolution subtask to the parallel computing unit corresponding to the convolution subtask, and control the parallel computing unit to execute the received convolution subtask;
• transmit a pooling parameter of a target pooling subtask to the pooling computation unit in response to receiving indication information, which is transmitted by the parallel computing unit, indicating that the convolution subtask has been completely executed, and control the pooling computation unit to execute the target pooling subtask, wherein the target pooling subtask refers to a pooling subtask corresponding to the convolution subtask which corresponds to the parallel computing unit transmitting the received indication information;
• merge the execution results of the respective pooling subtasks received from the pooling computation unit to obtain an execution result of the convolution task.