Neural-network pooling

A circuit, comprising:

• a multiply-add array configured to perform an operation for a first layer in a neural network, wherein the first layer is a linear layer;
• a pre-pooler configured to perform a pre-pooling operation for a second layer in the neural network;
• a post-pooler configured to perform a post-pooling operation for a third layer in the neural network;
• and
• a tensor buffer configured to store a tensor generated by a fourth layer in the neural network, wherein inputs of the pre-pooler and the multiply-add array are coupled to the tensor buffer and an output of the pre-pooler is coupled to the tensor buffer, and wherein an input of the post-pooler is coupled to an output of the multiply-add array,
• wherein the multiply-add array is the only multiply-add array coupled to the tensor buffer and the pre-pooler and the post-pooler are separate from the multiply-add array.

A method comprising:

• storing a tensor outputted from a previous layer in a neural network in a tensor buffer;
• processing the tensor using a first lower layer in the neural network by reading the tensor from a first page of the tensor buffer;
• storing the processed tensor in a scratch pad page of the tensor buffer;
• performing a first operation on the processed tensor using a multiply-add array,
• wherein the first operation corresponds to an upper layer in the neural network that is downstream from the first lower layer;
• while performing the first operation, performing a first pooling operation using a pre-pooler corresponding to a second lower layer in the neural network by reading the tensor from the first page of the tensor buffer;
• and
• while performing the first operation and the first pooling operation, performing a second pooling operation in a post-pooler using data generated by the upper layer;
• wherein the multiply-add array is the only multiply-add array coupled to the tensor buffer and the pre-pooler and the post-pooler are separate from the multiply-add array.