US 11,537,860 B2 · Neural net work processing

Abstract

A neural network processor is disclosed that includes a combined convolution and pooling circuit that can perform both convolution and pooling operations. The circuit can perform a convolution operation by a multiply circuit determining products of corresponding input feature map and convolution kernel weight values, and an add circuit accumulating the products determined by the multiply circuit in storage. The circuit can perform an average pooling operation by the add circuit accumulating input feature map data values in the storage, a divisor circuit determining a divisor value, and a division circuit dividing the data value accumulated in the storage by the determined divisor value. The circuit can perform a maximum pooling operation by a maximum circuit determining a maximum value of input feature map data values, and storing the determined maximum value in the storage.

Claims16

Dependent claims

A neural network processor operable to execute a neural network, the neural network processor comprising:

a combined convolution and pooling circuit configured to perform convolution and pooling operations for the neural network being executed by the neural network processor;
the combined convolution and pooling circuit comprising:
a multiply circuit configured to determine a product of data values;
storage configured to store data values;
and
an add circuit configured to add a data value to a data value stored in the storage so as to accumulate data values in the storage;
the combined convolution and pooling circuit further comprising at least one of:
a divisor circuit configured to determine a divisor value to use to determine a mean of data values accumulated in the storage;
and
a maximum circuit configured to determine a maximum value of data values, and to store the determined maximum value in the storage;
wherein the combined convolution and pooling circuit further comprises:
a padding circuit configured to pad an array of input data values, wherein the padding circuit is selectively configurable to:
pad an array of input data values with zeros;
and to:
pad an array of input data values with a minimum value for the array.

12 more elements

A method of performing convolution and/or pooling operations when executing a neural network on a neural network processor, the neural network processor comprising:

a combined convolution and pooling circuit configured to perform convolution and pooling operations for the neural network being executed by the neural network processor;
the combined convolution and pooling circuit comprising:
a multiply circuit configured to determine a product of data values;
storage configured to store data values;
a divisor circuit configured to determine a divisor value to use to determine a mean of data values accumulated in the storage;
and
a combined add and maximum circuit comprising:
an add circuit configured to add a data value to a data value stored in the storage so as to accumulate data values in the storage;
and
a maximum circuit configured to determine a maximum value of data values, and to store the determined maximum value in the storage;
wherein the combined add and maximum circuit is selectively configurable to:
add a data value to a data value stored in the storage so as to accumulate data values in the storage;
and to:
determine a maximum value of data values, and store the determined maximum value in the storage;
the method comprising:
the neural network processor executing the neural network;
and
the combined convolution and pooling circuit performing at least one of a convolution operation, an average pooling operation and a maximum pooling operation for the neural network executing on the neural network processor;
wherein the combined convolution and pooling circuit performing the convolution operation for the neural network executing on the neural network processor comprises:
the multiply circuit determining products of corresponding input feature map and convolution kernel weight values;
and
the combined add and maximum circuit accumulating the products determined by the multiply circuit in the storage;
and
the combined convolution and pooling circuit performing the average pooling operation for the neural network executing on the neural network processor comprises:
the combined add and maximum circuit accumulating input feature map data values in the storage;
and
the divisor circuit determining a divisor value to use to determine a mean of the input feature map data values accumulated in the storage by the add circuit;
and
the combined convolution and pooling circuit performing the maximum pooling operation for the neural network executing on the neural network processor comprises:
the combined add and maximum circuit determining a maximum value of input feature map data values, and storing the determined maximum value in the storage.

27 more elements

A method of performing convolution and/or pooling operations when executing a neural network on a neural network processor, the neural network processor comprising:

a combined convolution and pooling circuit configured to perform convolution and pooling operations for the neural network being executed by the neural network processor;
the combined convolution and pooling circuit comprising:
a multiply circuit configured to determine a product of data values;
storage configured to store data values;
an add circuit configured to add a data value to a data value stored in the storage so as to accumulate data values in the storage;
a divisor circuit configured to determine a divisor value to use to determine a mean of data values accumulated in the storage;
and
a maximum circuit configured to determine a maximum value of data values, and to store the determined maximum value in the storage;
wherein the multiply circuit comprises a bypass circuit, and the multiply circuit is selectively configurable to:
determine a product of data values;
and to:
pass data values onwards unchanged via the bypass circuit;
the method comprising:
the neural network processor executing the neural network;
and
the combined convolution and pooling circuit performing at least one of a convolution operation, an average pooling operation and a maximum pooling operation for the neural network executing on the neural network processor;
wherein the combined convolution and pooling circuit performing the convolution operation for the neural network executing on the neural network processor comprises:
the multiply circuit determining products of corresponding input feature map and convolution kernel weight values;
and
the add circuit accumulating the products determined by the multiply circuit in the storage;
and
wherein the combined convolution and pooling circuit performing the average pooling operation for the neural network executing on the neural network processor comprises:
the multiply circuit passing input feature map data values onwards unchanged to the add circuit via the bypass circuit;
the add circuit accumulating the input feature map data values in the storage;
and
the divisor circuit determining a divisor value to use to determine a mean of the input feature map data values accumulated in the storage;
and
wherein the combined convolution and pooling circuit performing the maximum pooling operation for the neural network executing on the neural network processor comprises:
the multiply circuit passing input feature map data values onwards unchanged to the maximum circuit via the bypass circuit;
and
the maximum circuit determining a maximum value of the input feature map data values, and storing the determined maximum value in the storage.

28 more elements

Dependencies are inferred from the claim text.

Neural net work processing

Life of this patent

Neural net work processing

Life of this patent

Abstract

Claims16