Device and method for allocating intermediate data from an artificial neural network

A method for allocating a global memory size, the method comprising:

• defining several possible placements of intermediate data blocks, generated by layers of an artificial neural network, in a global memory area;
• developing an initial sequence of placements of the intermediate data blocks of each layer from a first layer to a last layer of the neural network, each placement of the initial sequence being selected from the several possible placements of the intermediate data block to be placed, the initial sequence allowing definition of a planned size of the global memory area, the initial sequence being set as a candidate sequence;
• memorizing unselected placements for the intermediate data blocks of each layer;
• developing at least one sequence of placements that is modified compared to the initial sequence from a given placement of the initial sequence, the developing comprising modifying the given placement by a memorized unselected placement;
• in accordance with the planned size of the global memory area obtained by the modified sequence being less than the planned size of the global memory area of the candidate sequence, setting the modified sequence to be the candidate sequence;
• using the sequence of placements, set as the candidate sequence once each modified sequence has been developed, as the placements of the intermediate data blocks for the memory allocations;
• selecting the global memory size to be equal to the planned size obtained by the candidate sequence;
• and
• allocating the memory according to successive determined placements.

An integrated circuit configured to receive a multilayer artificial neural network, the integrated circuit comprising:

• a memory configured to memorize intermediate data blocks generated by each of a plurality of layers of the neural network in a global memory area;
• and
• a processing unit coupled to the memory, the processing unit configured to:
• define several possible placements of the intermediate data blocks of each layer in the global memory area;
• develop an initial sequence of placements of the intermediate data blocks of each layer from a first layer to a last layer of the neural network, each placement of the initial sequence being selected from the several possible placements of the intermediate data block to be placed, the initial sequence allowing definition of a planned size of the global memory area, the initial sequence being set as a candidate sequence;
• memorize unselected placements for the intermediate data blocks of each layer;
• develop at least one sequence of placements that is modified compared to the initial sequence from a given placement of the initial sequence, the developing comprising modifying the given placement by a memorized unselected placement;
• in accordance with the planned size of the global memory area obtained by the modified sequence being less than the planned size of the global memory area of the candidate sequence, set the modified sequence to be the candidate sequence;
• use the sequence of placements, set as the candidate sequence once each modified sequence has been developed, as the placements of the intermediate data blocks for in the global memory area;
• and
• during an execution of the neural network:
• select a global memory size to be equal to the planned size obtained by the candidate sequence;
• and
• allocate the memory according to successive determined placements.