An End-to-End Quantization Framework for Fixed Point Fast Fourier Transform Hardware Implementation via Deep Neural Network
Wenqian Cui, Shunqing Zhang, Zhiyong Chen, Shan Cecilia Cao, Shugong Xu · 2020
Fast Fourier Transform (FFT) plays an important role in signal processing nowadays. The quantization problem of sensed/sampled information has been revealed recently. In this paper, we present an end-to-end quantization framework via deep neural networks (DNNs). We have jointly optimized the system of signal quantization and de-quantization, and applied this system before fixed point FFT operation. The system can extract the features of the signal and reduce the hardware consumption of signal processing. Through numerous experiment, we find that the FFT calculation error of our quantization system is not smaller than results of uniform quantization and we analyzed underlying reasons. The designs of system brings a lot of new ideas to the future work. Our quantization system based on DNN can be used not only in FFT calculation, but also in other linear/non-linear symmetric operations.