How do I perform muliple FFTs on a stack of data in parallel on a GPU?

Hi, I have a stack of images, and I want to compute the 2d fourier-transform of each image in parallel. For example, if I was not using a GPU, I would be doing something like this:

A = rand(256,256,100);
B = zeros(256, 256, 100);
for j = 1:100
    B(:,:,j) = fftshift(fft2(A(:,:,j)));
end;

I would like to do each iteration of this for-loop in parallel. Is there a way to do that using gpuArray and arrayfun, or do I have to define my own cuda kernel?

Answer 1

[Edric Ellis]

FFT2 in MATLAB already applies to each 'page' of a 3-dimensional array, and this is true on the GPU too. Unfortunately, FFTSHIFT doesn't work that way, so you need something like this:

A = gpuArray.rand(256, 256, 100);
B = fftshift(fftshift(fft2(A), 1), 2);

[Adam]

Thanks for your help! This definitely makes things very simple! Out of curiosity, is there a standard way of independently processing multiple independent 'pages' simultaneously by some arbitrary function? For example, if I have a function:

b = myfun(a)

where b and a are both n-by-n, how would I use arrayfun to apply 'myfun' to each page of a n-by-n-by-m gpuArray, (thus returning another 3D gpuArray)?