Methods and apparatus to tile walk a tensor for convolution operations

A system for running a convolutional layer of a neural network, comprising:

• a memory configured to store an input tensor of the one or more convolutions, the input tensor comprising a plurality of tiles;
• and
• an accelerator associated with the memory, the accelerator comprising:
• memory units to store the plurality of tiles,
• convolution engines configured to operate in parallel to run the convolutional layer, a convolution engine configured to perform a plurality of multiply-accumulate (MAC) operations on a tile in the input tensor of the convolutional layer, different convolution engines configured to perform MAC operations on different tiles in the input tensor of the convolutional layer in parallel, and
• a controller configured to control, based on a first iterator and a second iterator, iterations of the convolutional layer over the plurality of tiles, an iteration of the convolutional layer comprising the plurality of MAC operations performed by the convolution engine on the tile in the input tensor of the convolutional layer, different iterations conducted in parallel by different ones of the convolution engines over different ones of the plurality of tiles, the first iterator configured to count tiles along a first dimension of the input tensor that have been processed by the convolution engines, the second iterator configured to count tiles along a second dimension of the input tensor that have been processed by the convolution engines.

A method for running a convolutional layer of a neural network, comprising:

• receiving a plurality of tiles in an input tensor of the convolutional layer, wherein each tile is a portion of the input tensor;
• storing the plurality of tiles in one or more memory units, each memory unit storing one or more of the plurality of tiles;
• providing the plurality of tiles to convolution engines to run the convolutional layer, a convolution engine configured to perform a plurality of multiply-accumulate (MAC) operations on a tile in the input tensor of the convolutional layer, different convolution engines configured to perform MAC operations on different tiles in the input tensor of the convolutional layer in parallel;
• and
• controlling, based on a first iterator and a second iterator, iterations of the convolutional layer over the plurality of tiles, an iteration of the convolutional layer comprising the plurality of MAC operations performed by the convolution engine on the tile in the input tensor of the convolutional layer, different iterations conducted in parallel by different ones of the convolution engines over different ones of the plurality of tiles, the first iterator configured to count tiles along a first dimension of the input tensor that have been processed by the convolution engines, the second iterator configured to count tiles along a second dimension of the input tensor that have been processed by the convolution engines.