Convolutional neural network hardware accelerator based on Winograd algorithm and calculation method

Abstract

The invention discloses a Winograd algorithm-based convolutional neural network hardware accelerator and a calculation method. The accelerator comprises an input buffer module, a PE array, a post-processing module, a data distributor, a control module, a configuration module, an output buffer module and a storage layer; the configuration module decodes the instruction into configuration information; the control module is used for controlling the PE array, the data distributor and other modules to complete different calculations; the data distributor reads the data from the storage layer by adopting different address mapping modes and data distribution modes, and distributes the data to the input buffer; the input buffer module sends the buffered calculation data to the PE array for calculation; the PE array performs calculation path reconstruction according to the configuration information and calculates the data; and the post-processing module performs operations such as multi-channel accumulation and activation function processing on a calculation result. According to the method, the convolution calculation speed can be increased, the data migration loss is reduced, and the performance of the whole accelerator is improved.

Description

technical field [0001] The invention relates to the field of deep convolutional neural network calculations, in particular to a method and device for parallel accelerated calculation of convolutional pipelines based on the Winograd algorithm. Background technique [0002] The increasing complexity of the current neural network structure has dramatically increased the amount of calculation and data transfer of the neural network, while the serial calculation characteristics of the CPU and the small number of cores make its calculation efficiency low, while the GPU energy consumption is relatively low, so It is necessary to design efficient hardware neural network accelerators to improve the computational efficiency of neural networks. [0003] Convolution calculation is the basic calculation of neural network. Traditionally, it is generally performed by sliding window convolution. This calculation method brings a huge amount of calculation and data transfer to the entire neur...

AI Technical Summary

Problems solved by technology

[0004] The Winograd algorithm contains continuous matrix multiplication, which will increase the access time of intermediate results and reduce the calculation speed. At the same time, the Winograd algorithm needs a long calculation process to obtain the calculation results, and will occupy a lot of hardware resources.

Images