Implementation of Convolution Function Through Fast Fourier Transform in Convolution Neural Networks Computation

This paper investigates the possibility of performing the convolution function in the frequency domain using the fast Fourier transform. The paper presents the proposed structure of the convolution function implementation, comparing the computational complexity of convolution when performed by the classical method and using the FFT. The pros and cons of performing the convolution function using FFT are described. The implementation of the classical convolution function and convolution by FFT in VHDL is done and the results are given: running time, required external memory bandwidth, FPGA resources occupied. Several possible solutions to the increased load on the external memory are proposed.