Weight data storage method and neural network processor based on the method

A neural network processor, comprising:

• a weight and index storage unit configured to store effective weights and an index associated with the effective weights, wherein the effective weights are non-zero weights associated with a weight convolution kernel matrix and wherein the index associated with the effective weights is used to mark respective positions of the effective weights in the weight convolution kernel matrix;
• a neuron index unit configured to acquire, based on the index associated with the effective weights, convolution domain neurons corresponding to the effective weights;
• and
• a computation array unit configured to perform convolution operations associated with the effective weights and the convolution domain neurons, wherein the index associated with the effective weights is acquired by:
• expanding the weight convolution kernel matrix to a K by K square matrix;
• equally dividing the square matrix into P2 sub-matrixes, wherein P is an integer greater than or equal to 2;
• recursively and equally dividing one or more sub-matrixes of the P2 sub-matrixes that comprise an effective weight until leaf sub-matrixes resulting from the dividing comprise only one weight, wherein a multi-level sub-matrix is obtained from the dividing;
• and
• for each of the leaf sub-matrixes that is associated with an effective weight, marking an effective path branch associated with the leaf sub-matrix with a bit value of 1.

A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to at least:

• search for effective weights in a weight convolution kernel matrix and acquire an index associated with the effective weights, wherein the effective weights are non-zero weights and the index associated with the effective weights is used to mark respective positions of the effective weights in the weight convolution kernel matrix;
• expand the weight convolution kernel matrix to a K by K square matrix;
• equally divide the square matrix into P2 sub-matrixes, wherein P is an integer greater than or equal to 2;
• recursively and equally divide one or more sub-matrixes of the P2 sub-matrixes that comprise an effective weight until leaf sub-matrixes resulting from the dividing comprise only one weight, wherein a multi-level sub-matrix is obtained from the dividing;
• for each of the leaf sub-matrixes that is associated with an effective weight, mark an effective path branch associated with the leaf sub-matrix with a bit value of 1 to obtain the index associated with effective weights;
• and
• store the effective weights and the index associated with the effective weights.