The purpose of this library is to provide a fast FFT (Fast Fourier Transform) implementation, entirely in pure C#. This library is adapted from General Purpose FFT Package by Ooura, modified to be compatible with the .NET Standard complex number type.

This library was created as an FFT implementation for NumFlat, a general purpose numerical computing library written in pure C#. If you're interested in numerical computing with C#, please check out NumFlat.

• Fast: More than four times as fast as the managed FFT implementation in Math.NET Numerics.
• Lightweight: Small code size, with no dependencies other than .NET Standard 2.1.

The NuGet package is available:

Install-Package FftFlat

If you don't want to add a DLL, copy the .cs files to your project.

First, add a using statement for the FftFlat namespace.

using FftFlat;

To perform FFT or IFFT, create an instance of FastFourierTransform and call the appropriate method.

var samples = new Complex[1024];
samples[0] = 1;

var fft = new FastFourierTransform(1024);
fft.Forward(samples);

Ooura's original FFT implementation is based on a different definition from that used in Math.NET Numerics. FFtFlat adjusts this difference, ensuring its results match those of Math.NET Numerics' FFT.

Normalization is only done during the IFFT. This is similar to using FourierOptions.AsymmetricScaling for FFT in Math.NET Numerics.

Note that the FastFourierTransform object is not thread-safe. If performing FFT across multiple threads, ensure a separate instance is provided for each thread.

In this demo video, the spectrum is visualized in real-time as sound is played using AudioStream of RayLib-CsLo.

https://www.youtube.com/watch?v=KTpG_z_ejZ0

The following is a benchmark comparing this with other pure C# FFT implementations. In this benchmark, the time taken to perform FFT and IFFT on a random signal was measured. The FFT lengths used were powers of two, ranging from 256 to 8192.

• ✅ FFT for power-of-two length samples
• ✅ Real FFT
• ⬜ Other transformations (such as cosine transform)
• ⬜ Support for 32-bit floating-point numbers
• ⬜ FFT for arbitrary length samples

FftFlat is available under the MIT license.