Optimizing: Getting rid of for loop

Hi, I've been scratching my head on this problem and keep seem to figureout a good way to do it.

I have a big matrix : size(A)=1024x1024x400 where the 3rd dimension is that of time and the two first of space. Here is an example of what I would want to do with it:

for x=1:1022
    for y=1:1022
B(y,x,1) = mean2(A(y:y+2,x:x+2,1));
    end
end

Doing it like that works but Matlab will do the calculations on one element at the time instead of all at the same time like in the following case:

B(:,:,1) = A(:,:,1).*A(:,:,2);

Is there a way to rewrite my initial code or use a function that would make it way faster? Thank you for any help.

Answer 1

[Guillaume]

Not sure what the mean2d function do as it's not a standard matlab function. Assuming it's calculating a mean, then probably

kernel = ones(3, 3, size(A, 3));
kernel = kernel ./ numel(kernel);  %convolution kernel for mean
B = convn(A, kernel, 'valid');

[Raphaël]

Yes it works thanks. One question, do you know the difference between conv2 and convn for this? Thanks

[Matt J]

Because of the separable nature of the kernel, it may be more efficient to do this in separable passes,

B=convn(A,[1,1,1]/9, 'valid');
B=convn(B,[1;1;1],'valid');