A qusetion about the FiNUFFT “guru” interface of the function "finufft1d1"

[zengletian1491]

Dear developers:
For the function “finufft1d1(x, c,+1,1e-12,1024)”, suppose the input parameter “c” be a double-precision matrix with 6144 rows and 1200 columns. Suppose the input parameter “x” be a column vector with 6144 rows and 1 column that represents the source non-uniform points but varies for each column of “c”. Does the FiNUFFT “guru” interface “finufft_plan” work for this situation?

I will provide a MCE (minimally complete example code) as follows written by MATLAB code:

kr = ([-3072 : 3071].'/6144 * 1.2 + 33.33) *（4.0 * pi）; // 6144×1 column vector
kx = (33.33 + [-600: 599]./1200 * 4.0) *（2.0 * pi）; // 1×1200 row vector
c = zeros(6144,1200);
for k = 1 : 1200
c(:,k) = finufft1d1((sqrt(kr.^2 - kx(k).^2) - 362.27)/(15.98)*pi, data_in(:,k),-1,1e-12,6144);
end

Note that data_in is a 6144×1200 matrix of double-precision type (such as SAR data). In the above code, the term “sqrt(kr.^2 - kx(k).^2) - 362.27)/(15.98)*pi” varies with “k”. That is, the source non-uniform points are different each transfer, so the FiNUFFT “vectorized” interface does not work.

We refrence to the finufft-readthedocs-io-en-latest on the website https://finufft.readthedocs.io/en/latest/matlab.html , the screenshot of which are as follows:

As demonstrated in the screenshot, the FiNUFFT "guru" interface is capable of handling the situations previously mentioned. However, the validity of this assertion remains uncertain.

Thanks very much!

[ahbarnett]

Dear Letian,

I'm not sure what you mean by " this assertion remains uncertain" :)

I remember sending you the above code on 4/29/24.

If you're looking for how to do this with guru, and what is the associated speedup, here you go.
Note how I started to clean up your code (it was unusual that you had number of modes equal to number of input points, so I used different variables, N vs M, to clarify that):

% MCE for zengletian, on repeated 1d NUFFTs with new NU pts.
% Apparently a SAR application.
% Note you needed data_in for it to run. I tidied it up.
% Barnett 10/31/25
clear
maxNumCompThreads(1);    % better since transforms so small

% note it's neater to have a single variable control sizes...
N = 6144;  % num modes
M = N;    % num 1D k-space points, weirdly same as modes (ie, square xform)
nx = 1200;  % number of 1D transforms
tol = 1e-12;
kr = ([-M/2:M/2-1].'/M * 1.2 + 33.33) *(4.0 * pi);     % 6144x1 column vect
kx = (33.33 + [-nx/2:nx/2-1]./nx * 4.0) *(2.0 * pi);   % 1x1200 row vector
data_in = rand(M,nx) + 1i*rand(M,nx);     % stack of complex data at k pts

c = zeros(N,nx,'like',data_in);     % allocate for complex output

% 1) naive repeated FINUFFT calls... slow
tic
for k = 1:nx
  x = (sqrt(kr.^2 - kx(k)^2) - 362.27)/(15.98)*pi;  % N 1D k-space pts
  c(:,k) = finufft1d1(x, data_in(:,k),-1, tol, N);
end
toc     % 0.60 sec

% 2) Guru interface...
tic
ntrans = 1;       % since you only do one transform for each set of x pts
opts = struct();  % opts.nthreads=1; opts.debug=2;
p = finufft_plan(1,N,-1,ntrans,tol,opts);
for k = 1:nx
  x = (sqrt(kr.^2 - kx(k)^2) - 362.27)/(15.98)*pi;  % N 1D k-space pts
  p.setpts(x);
  c(:,k) = p.execute(data_in(:,k));
end
p.delete;
toc     % 0.28 sec: 2x speedup

I hope you can learn from this example.
Note single-threaded is better for such small problems, at least on my Ryzen laptop, here compiled using GCC 11.4 with FFT=DUCC (not FFTW).

Then you could also parallelize over k. For example, see the C++ demos:
https://github.com/flatironinstitute/finufft/blob/master/examples/threadsafe1d1.cpp
https://github.com/flatironinstitute/finufft/blob/master/examples/threadsafe2d2f.cpp

It would be pretty easy to try the first one out in matlab.

Let me know how it goes, Alex