Convolutional operation device and method

A method for performing a single-layer convolutional neural network forward operation, comprising:

• pre-storing an input/output instruction at a starting address of an instruction storage unit;
• if operation is started, a control unit reading the input/output instruction from the starting address of the instruction storage unit, and a data access unit reading, according to a control signal decoded from the input/output instruction, all corresponding convolutional neural network computational instructions from an external address space, and caching all the instructions in the instruction storage unit;
• the control unit reading in a next input/output instruction from the instruction storage unit, and the data access unit reading all data required by a primary computation module from the external address space to a first storage unit of the primary computation module according to a control signal decoded from the next input/output instruction;
• the control unit reading in another input/output instruction from the instruction storage unit, and the data access unit reading convolutional kernel data required by secondary computation modules from the external address space according to a control signal decoded from another input/output instruction;
• the control unit reading in a next CONFIG instruction from the instruction storage unit, and the convolutional operation device configuring various constants required by computation of a present layer of a neural network according to a control signal decoded from the next CONFIG instruction;
• the control unit reading in a next COMPUTE instruction from the instruction storage unit, and the primary computation module sending input data in a convolutional window to the multiple secondary computation modules through an interconnection module according to a control signal decoded from the next COMPUTE instruction, storing the input data in second storage unit of the multiple secondary computation modules, and moving the convolutional window according to the instruction;
• computation units of the multiple secondary computation modules reading convolutional kernels from a third storage unit according to the control signal decoded from the COMPUTE instruction, reading the input data from the second storage units, completing convolutional operation of the input data and the convolutional kernels, and returning obtained output scalars through the interconnection module;
• splicing the output scalars returned by the multiple secondary computation modules into complete intermediate vectors step by step in the interconnection module;
• the primary computation module obtaining the intermediate vectors returned by the interconnection module, moving the convolutional window to traverse all the input data, splicing all the returned intermediate vectors into an intermediate result, reading offset data from the first storage unit according to the control signal decoded from the COMPUTE instruction, and adding the offset data and the intermediate result together to obtain an offset result through a vector addition unit;
• and then an activation unit activating the offset result and writing final output data back into the first storage unit;
• the control unit reading in yet another input/output instruction from the instruction storage unit, and the data access unit storing the output data of the first storage unit to a specified address of the external address space according to a control signal decoded from the next input/output instruction, then ending the operation.