Example of multidimensional transform?

[shssoichiro]

Hi, I have a case where I need to implement a 3D FFT. This crate doesn't seem to have an existing utility function for it. I figured I could potentially create my own by combining multiple 1D FFTs, but it seems to have resulted in garbage output. Could an example be added to the repository of how to do a multidimensional transform correctly with this crate? (Even better would be if there were wrapper functions provided by the crate.)

My current attempt looks like this:

// Applies a real-to-complex 3-dimensional FFT to `real`
fn real_to_complex_3d(
  &self, real: &[f32; BLOCK_VOLUME], output: &mut [Complex<f32>; CCNT],
) {
  let scratch_len = self
    .ft
    .0
    .get_inplace_scratch_len()
    .max(self.ft.1.get_inplace_scratch_len())
    .max(self.ft.2.get_inplace_scratch_len());
  let mut scratch = vec![Complex::default(); scratch_len];

  let mut complex =
    real.iter().map(|val| Complex::from(*val)).collect::<Vec<_>>();
  self.ft.0.process_with_scratch(&mut complex, &mut scratch);
  self.ft.1.process_with_scratch(&mut complex, &mut scratch);
  self.ft.2.process_with_scratch(&mut complex, &mut scratch);

  output.iter_mut().zip(complex.iter()).for_each(|(out, inn)| {
    *out = *inn;
  });
}

[HEnquist]

Your solution is not far off, you "just" need to transpose the data between the FFT steps so that each FFT runs along the correct axis.

But have you looked at ndrustfft? I haven't tried it but looks like it should be able to do what you need.
https://crates.io/crates/ndrustfft

[shssoichiro]

I see, I'll "just" do that if I can. 😆 Thanks. 🙂

I did look at ndrustfft, but I wanted to avoid the dependency on ndarray. Last time I tried ndarray, it was quite slow. Although that was years ago, it looks like they've been iterating, so I guess that could also be worth trying once more.

[HEnquist]

Yeah there is quite a lot hidden in that "just" 😆
BTW roughly how large is your data? If it fits in the cpu cache, then you can get away with using simple transpose algorithms. But those get painfully slow for larger data. From a quick look I think ndrustfft is using something simple, could explain why you thought it was slow. Speeding up transposing is quite some work. This is for example what is used in RustFFT, and that's only for 2d! https://github.com/ejmahler/transpose/blob/master/src/out_of_place.rs

[shssoichiro]

I can't remember anymore, this was quite some time ago when I was messing around with some ML stuff in Rust. ndrustfft itself doesn't quite have just a "xd_transform" function, but they did have examples showing how to do it, so I ended up using that. (Although I haven't benchmarked it yet because there's a bug somewhere else in my code 🙂 ) The data I'm working with can be several MB but the transforms are in 16x16x3 chunks, so not terribly large and easily doable in cache.

[ejmahler]

I think RustFFT would benefit from something built in that does this. It seems like a common request, and easy enough to get wrong that people would benefit from a central utility.

I unfortunately don't have a lot of motivation for side projects right now, but if someone else makes a PR for a n-d FFT, tests it, etc I'd be happy to accept.

[preiter93]

Hi,

author of ndrustfft here; I crossed this issue by chance. The topic seems to be quite important because of its relevance to image processing and numerical computation, so I decided to do some tests.

Assuming we have a two-dimensional array of size n x n, the two approaches are:

"transpose"

• Transpose data, if necessary
• Compute 1 fft of size n x n
• Transpose back, if necessary

The transpose is done out of place, hopefully effciently with transposes very similar to
https://github.com/ejmahler/transpose
I think multidimensional transposes can be done in a similar way by adjusting
the row and col sizes appropriately, however, I haven't thought too deeply about that.

"ndrustfft"

• Iterate over 1d lanes -> Array1View
• Copy data into slice, if 1d view is not contiguous,
• Compute n ffts of size n

The code can be found here:
https://github.com/preiter93/ndfft_test
(This is only a proof of concept)

Here is a performance comparison (time in ms):

FFT along outer axis (transpose is not necessary)

N	ndrustfft	transpose
128	0.09	0.04
256	0.4	0.24
512	2.22	1.64
1024	10.3	4.77

FFT along inner axis (transpose is necessary)

N	ndrustfft	transpose
128	0.24	0.1
256	1.12	0.56
512	8.41	5.08
1024	42.92	19.49

So I think the transpose-based approach is promising. The advantages are
(1) Better Performance (~50%)
(2) No additional dependencies
However, the disadvantages are
(1) More memory is needed if an out of place transpose is used. In-place transposes are generally less efficient, see ejmahler/transpose#4.
(2) it is harder to parallelize, whereas ndrustfft is very easy to parallelize and scales quite well.

I don't know how much time I have to pursue this topic further, but I hope this has given a little insight.

[skewballfox]

hey I wasn't sure if I should post here or on #90 . Are you open to PRs for this? What are the constraints? I was thinking about doing something similar to how burn handles multiple tensor implementations: providing a trait that could be implemented on matrix/tensor types

[ejmahler]

I've opened pull request #181 to add multidimensional FFTs as a first-class feature of the library. I'd love some feedback from people here about the API and whether it covers your needs.