Data transfers in neural processing

A non-transitory computer-readable storage medium comprising a set of computer-readable instructions stored thereon which, when executed by at least one processor, cause the at least one processor to:

• select a predetermined format of subdivision in terms of height, width and depth of a feature map having at least height, width and depth dimensions, from a plurality of predetermined formats of subdivision, corresponding to a first set of subportions of a first portion of feature map data of a neural network, the first portion of feature map data being representative of a first part of the feature map, wherein the predetermined format of subdivision is selected based on a property of a neural network layer associated with the feature map data to be transferred, wherein the property of the neural network layer comprises predetermined dimensions of a selection of output feature map data of the neural network layer or a position of the neural network layer within a sequence of neural network layers;
• and
• schedule the transfer of the feature map data between local storage and external storage using the selected predetermined format of subdivision;
• wherein the feature map data of the neural network is to be transferred between local storage and external storage by:
• receiving the first portion of feature map data from the local storage;
• selecting the first set of subportions of the first portion of feature map data;
• compressing the subportions of feature map data to produce a first plurality of sections of compressed feature map data, each section corresponding to a respective subportion;
• instructing the storage of the first plurality of sections of compressed feature map data into external storage;
• receiving a second plurality of sections of compressed feature map data from the external storage, the second plurality of sections corresponding to a second portion of feature map data, the second portion of feature map data being representative of a respective second part of the feature map;
• decompressing the second plurality of sections to produce a second set of subportions of the second portion of feature map data;
• and
• storing the second portion of feature map data in local storage.

A data processing system comprising local storage circuitry and external storage circuitry arranged to store feature map data of a neural network, and processing circuitry arranged to:

• receive a first portion of feature map data from the local storage, the first portion of feature map data being representative of a respective first part of a feature map, the feature map having at least height, width and depth dimensions;
• select a first set of subportions of the first portion of feature map data;
• compress the subportions of feature map data to produce a first plurality of sections of compressed feature map data, each section corresponding to a respective subportion;
• instruct the storage of the first plurality of sections of compressed feature map data into the external storage;
• receive a second plurality of sections of compressed feature map data from the external storage, the second plurality of sections corresponding to a second portion of feature map data, the second portion of feature map data being representative of a respective second part of the feature map;
• decompress the second plurality of sections to produce a second set of subportions of the second portion of feature map data;
• and
• store the second portion of feature map data in local storage,
• wherein the first and second sets of subportions each correspond to a predetermined format of subdivision, in terms of height, width and depth, of the feature map and selecting the predetermined format of subdivision from a plurality of predetermined formats of subdivision is based on a property of a neural network layer associated with the feature map data being transferred, wherein the property of the neural network layer comprises predetermined dimensions of a selection of output feature map data of the neural network layer or a position of the neural network layer within a sequence of neural network layers.