I have been building a convolution library for multichannel audio and am using FFTW as the main FFT transform library. There are 4 channels of audio in play at any one time and they are interleaved, so one "frame" contains 4 separate samples - one from each channel.

I have been using the Real to Complex 2D functions up until now, but I am starting to be concerned that multichannel audio is perhaps not the right application for this - even though it fits with the documented data structure.

Should I be using 4 separate 1D transforms instead of a single 2D FFT pass?

  • $\begingroup$ Use whichever is more efficient. Test it and judge yourself. I would prefer 1D for simplicity. $\endgroup$
    – Fat32
    Dec 19 '16 at 11:30
  • $\begingroup$ I'm really trying to determine whether there is a technical reason why I shouldn't do this. The code I have at the moment is humming along nicely, but I am running into a few challenges implementing deconvolution in a similar manner. Does a 2D FFTW treat each "channel" as entirely separate, as if it is a separate unique FFT process across each channel, or is there interaction between the channels spectrally? $\endgroup$
    – Mark
    Dec 19 '16 at 11:32

You should consider what you're doing.

Let me draw your audio data in the manner I understood it's in your memory. I'll call your channels A,B,C,D, and number their respective samples increasingly:

A0 B0 C0 D0 A1 B1 C1 D1 A2 B2 C2 D2 …

Now, if you want to apply a 2D FFT, your data needs to have the structure of 2D data – i.e. a matrix. You didn't quite specify, but I assume you've got a matrix

$$ \mathbf S = \begin{pmatrix} A_0 & B_0 & C_0 & D_0\\ A_1 & B_1 & C_1 & D_1\\ A_2 & B_2 & C_2 & D_2\\ A_3 & B_3 & C_3 & D_3\\ \vdots & \vdots & \vdots & \vdots\\ A_{N-2} & B_{N-2} & C_{N-2} & D_{N-2}\\ A_{N-1} & B_{N-1} & C_{N-1} & D_{N-1}\\ \end{pmatrix} $$ And you're applying the $4\times N$ DFT (FFT):

$$\begin{align} \hat{\mathbf{S}} &= \text{DFT}_{4\times N}\left\{\mathbf S\right\}\\ &= \left(\frac1{\sqrt{4N}}\sum_{n=0}^N\sum_{m=0}^4 S_{n,m}\, e^{j2\pi\left(\frac{nk}N + \frac{ml}4\right)} \right)_{k,l}\tag1\\ &=\left(\frac1{2\sqrt N}\sum_{n=0}^N A_n e^{j2\pi\left(\frac{nk}N + 0\right)} + B_n e^{j2\pi\left(\frac{nk}N + \frac l4\right)} + C_n e^{j2\pi\left(\frac{nk}N + \frac {2l}4\right)} + D_n e^{j2\pi\left(\frac{nk}N + \frac {3l}4\right)} \right)_{k,l}\\ &=\left(\frac1{2\sqrt N}\sum_{n=0}^N A_n e^{j2\pi\frac{nk}N} + B_n e^{j2\pi\frac{nk}N} e^{j\frac12\pi l} + C_n e^{j2\pi\frac{nk}N} e^{j\pi l} + D_n e^{j2\pi\frac{nk}N} e^{j\pi \frac32 l} \right)_{k,l} \end{align} $$

Looking at the last line: you've got to ask yourself whether you really want to have the DFT over each of your 4-channel frames – that sounds intuitively wrong.

Look at eq. $(1)$: you can swap the summations; hence, it doesn't matter if you first take the FFT of every channel column vector in "time direction", and then the row-wise FFT of the resulting matrix in "channels direction", or first do the row-wise FFT over all four channels, and then over all times.

I think you really want 4 1D-FFTs of length $N$, and not one $N\times 4$ 2D-FFT.

Luckily, the advanced 1D-FFT real-to-complex API of FFTw makes it really easy to do four identical FFTs on interleaved data. I annotated the API with the values that seem to be apporpriate for your use case:

fftw_plan fftw_plan_many_dft_r2c(int rank, // Dimensionality == 1 
        const int *n,       // FFT input length == N
        int howmany,        // Amount of FFTs == 4
        double *in,     // pointer to A0
        const int *inembed, // pad/truncate input ? No --> NULL
        int istride,        // Distance between input elements that belong to the *same* FFT == 4
        int idist,      // Distance between the first elements of different FFTs == 1
        fftw_complex *out,  // pointer to output space
        const int *onembed, // see above inembed
        int ostride,        // same as above
        int odist,      // same as above
        unsigned flags);    // allow destruction of input buffer? Measure or guess the best algorithm?
  • $\begingroup$ Thankyou Marcus. 4 x 1D FFTs it is. Back to the compiler... $\endgroup$
    – Mark
    Dec 19 '16 at 13:06
  • $\begingroup$ @Mark hope my edit helps $\endgroup$ Dec 19 '16 at 13:19
  • $\begingroup$ I'm a novice at all of this - surprised I have managed to get as far as I have - so ploughing on until I hit the next brick wall. I'll be hitting you up for some advice on deconvolution soon! $\endgroup$
    – Mark
    Dec 19 '16 at 13:20

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.