A method for neural network processing, wherein the method is applied to a device for neural network processing, the device comprises a first computing array, a second computing array, and a control module, the control module includes a first control unit and a second control unit, the second control unit is configured to be in communication with the first control unit, a module or region where the first computing array and a first storage module are located is denoted as a convolution accelerate unit, and a second storage module is located outside the convolution accelerate unit;
- the method comprises:
- executing, by the first computing array, a first type of neural network operation, wherein the first computing array is a three-dimensional multiply accumulate (3D MAC) array, and the first type of neural network operation is a convolution operation;
- executing, by the second computing array, a second type of neural network operation, the second type of neural network operation being different from the first type of neural network operation;
- reading, from the first storage module, a weight required for the convolution operation into the first computing array by the first control unit;
- and reading, from the second storage module, feature data required for the convolution operation into the first computing array by the second control unit, the second storage module being disposed separately from the first storage module;
- reading, from the second storage module, data required for the second type of neural network operation into the second computing array by the second control unit;
- wherein the reading, from the second storage module, the feature data required for the convolution operation into the first computing array by the second control unit, comprises:
- reading, from a cache, the feature data required for the convolution operation into the first computing array by the second control unit when the first computing array is to execute the convolution operation, the cache being configured to cache the feature data required for the convolution operation that is read from the second storage module;
- wherein the reading, from the cache, the feature data required for the convolution operation into the first computing array by the second control unit when the first computing array is to execute the convolution operation, comprises:
- determining, by the second control unit, whether the cache has the feature data required for the convolution operation executed by the first computing array when the first computing array is to execute the convolution operation;
- and
- if a determination result is yes, reading, from the cache, the feature data required for the convolution operation into the first computing array by the second control unit;
- or
- if a determination result is no, reading, from the second storage module, required feature data into the cache, and reading, from the cache, the feature data required for the convolution operation into the first computing array by the second control unit;
- wherein the operation of reading the weight required for the convolution operation by the first control unit and the operation of reading the feature data required for the convolution operation by the second control unit are executed simultaneously, and after the convolution operation of the first computing array is completed, the second control unit stores a computation result obtained by the first computing array into the second storage module.