METHOD FOR PERFORMING OPERATION OF CONVOLUTIONAL LAYERS IN CONVOLUTIONAL NEURAL NETWORK AND DEVICE THEREOF

摘要

The present invention relates to a method for performing operation of convolutional layers in a convolutional neural network (CNN) and a device thereof. According to the present invention, the method comprises the steps of: reading unfolded feature data and a first convolutional kernel from a dynamic random access memory (DRAM); padding the unfolded feature data; folding the unfolded feature data padded from folded feature data in at least one dimension; storing the folded feature data in a static random access memory (SRAM); folding the first convolutional kernel in at least one dimension to generate one or more folded convolutional kernels; storing the one or more folded convolutional kernels in the SRAM; and allowing a calculation unit performing convolutional operation for the folded feature data with the one or more folded convolutional kernels to read the folded feature data and the one or more folded convolutional kernels from the SRAM. Accordingly, channel use can be increased, a buffer footprint can be decreased and operational efficiency can be increased.;COPYRIGHT KIPO 2019