Method for realizing high performance of radix-2 one-dimensional FFT (Fast Fourier Transform) based on domestic SW 26010 processor

Abstract

The invention provides a method for realizing high performance of a radix-2 one-dimensional FFT (Fast Fourier Transform) based on a domestic SW 26010 processor. Based on the domestic processor SW26010 platform, various optimization technologies such as a row or column register communication mechanism in an auxiliary core, an access memory-calculation overlapping double-buffer mechanism and a vector operation of a 256-bit single-instruction stream multiple-data stream are designed, meanwhile a two-layer decomposition-based Stockham FFT calculation framework is provided and a decomposition rule is a Cooley-Turkey algorithm, a four-layer structure framework of an interface layer, a main core layer, an auxiliary core layer and a kernel layer is designed for calculation of the radix-2 one-dimensional FFT, and thus the problem of limitation of an access memory bandwidth for the FFT calculation is effectively solved and the calculation performance of the radix-2 one-dimensional FFT is effectively improved. Compared with an open source FFTW (Fastest Fourier Transform in the West) library, the calculation performance of the radix-2 one-dimensional FFT based on the platform is rapidly improved; and a floating-point operation per second of the FFT calculation is taken as an example, an average speed-up ratio of the FFT calculation is 34.4 and a maximum speed-up ratio reaches 50.3.

Description

technical field [0001] The invention belongs to the field of Fourier transform, and in particular relates to a high-performance realization method of radix-2 one-dimensional FFT based on the domestic Shenwei 26010 processor. Background technique [0002] Fast Fourier Transform (FFT) is a fast calculation method of discrete Fourier transform. Discrete Fourier Transform (DFT) means that the Fourier transform is expressed as a discrete state in both the time domain and the frequency domain, and the time-domain sampling of the signal is transformed into the frequency-domain sampling of the discrete-time Fourier transform. DFT converts continuous and complex problems in natural science and engineering technology into discrete and simple operations. For a one-dimensional input sequence with a data size of N, the DFT calculation formula is as follows: [0003] [0004] Among them, ω N is the twiddle factor sequence, ω N = e -i2π / N ,e ix =cos x+i sin x, It can be seen fro...

AI Technical Summary

Problems solved by technology

[0008] The technology of the present invention solves the problem: it overcomes the problem that the open-source FFTW function library based on the prior art is directly applied to this platform with low performance, and provides a high-performance radix-2 one-dimensional fast Fourier transform based on the domestic Shenwei 26010 processor. The implementation method, designing a variety of high-performance optimization methods, and proposing a two-layer decomposition FFT algorithm structure, which can be effectively applied to radix-2 one-dimensional FFT calculations and fully improve the performance of the FFT function library

Images