Circshift matrix by different amount without for loop

Hello everyone. I have this 3 dimensional matrix called spectrums and I want to shift the 3rd dimension by different amounts given by vector shifts. Is there a way to do this without using a for loop? something like

shifted_spectrums=arrayfun(@(k) circshift(spectrums(:,:,k),shifts(2:end),2),spectrums);

but a solution that actualy works :P

PS: 3rd dimension of spectrums has 22 elements and shifts has 23 elements, that's why the 2:end.

[Jan]

What is the class and contents of shifts(2:end)? What is not "actually working"?

We cannot guess these details.

A for loop is faster than arrayfun in general. So please mention, why you want to avoid it.

Is this the loop version?

shifted_spectrums = zeros(size(spectrums));
for k = 1:size(spectrums, 3)
    shifted_spectrums(:, :, k) = circshift( ...
                        spectrum(:, :, k), shifts(k+1), 2);
end

Answer 1

[Javier Agustin Romero]

I've just got it working. A reshape is needed so the output has wanted size (sz1=1024, sz2=8192).

shifted_spectrums=reshape(cell2mat(arrayfun(@(k) circshift(spectrum(:,:,k)...
        ,shifts(k,1),2),1:Nfiles,'uni',0)),sz1/2,sz2,Nfiles);

This version is faster than the for loop

for i=2:Nfiles
        shifted_spectrums(:,:,i)=circshift(spectrums(:,:,i),shifts(i-1,1),2);
 end

Although not by much (16%, spectrums is a 512x8192x22 matrix)

[Jan]

Did you pre-allocate the output? This is essential for the performance:

shifted_spectrums = zeros(size(spectrums));

Answer 2

[Jan]

Try:

n         = 8192;
spectrums = rand(512, n, 22);
shifts    = randi(n, 1, 22);
tic
s1 = cell2mat(arrayfun(@(k) ...
      circshift(spectrums(:,:,k), shifts(k), 2),1:22,'uni',0));
toc

tic;
s2 = zeros(size(spectrums));
for k = 1:22
   s2(:,:,k)=circshift(spectrums(:,:,k), shifts(k), 2);
end
toc

For n=2000 I get:

arrayfun: 0.22sec
loop:     0.17sec

[Javier Agustin Romero]

Weird, with n=8192 I get:

arrayfun:   1.47sec
loop:       1.91sec

might be the Matlab version? I'm using 2016a. Anyhow, thank you very much for all your suggestions.

[Jan]

i7-3770 3.4GHz, 16 GB RAM, n = 8192:

R2009b:

Elapsed time is 2.660240 seconds.
Elapsed time is 2.062384 seconds.

R2016b:

Elapsed time is 2.613608 seconds.
Elapsed time is 2.292270 seconds.

R2018b:

Elapsed time is 2.682782 seconds.
Elapsed time is 1.207014 seconds.   ! Nice !

[Javier Agustin Romero]

Just ran it again and got

Elapsed time is 1.512578 seconds.
Elapsed time is 1.911443 seconds.

I'm working on a i7-7700HQ, 2.8GHz, 16GB RAM and got Matlab installed in a SSD (good to know that was a good investment). Anyway, what I find weird is that which method works best changes with PC's specs. I mean, shouldn't one method always work best and both times move up or down according to the PC specs? I am talking way over my head here though, have no idea.