How can I speed up a nested for loop without parfor or spmd?

I'm trying to run a nested for loop (shared below) on a data set that is 128x128x5263 and is complex. It's taking too long, over 25 minutes.

These are the things I tried:

1. Running it in a function
2. Using parfor: but it doesn't work because the loop outputs are not independent of each other.
3. Using spmd, but that did not help with the speed.

Is there another way I can do this?

Code below:

NOTE: k_data_shift is a 128x128x5263 complex data-set.

k_d1=zeros(size(k_data_shift,1),size(k_data_shift,2),size(k_data_shift,3));
k_d1(:,ss(1),1)=k_data_shift(:,ss(1),1);
%
for j=2:num_frames;
    k_d1(:,:,j)=k_d1(:,:,j)+k_d1(:,:,j-1);
    for i=1:num_frames;
        k_d1(:,ss(i),i)=k_data_shift(:,ss(i),i);
    end
end

Thanks!

[per isakson]

And what's ss ?

Answer 1

[Prateekshya]

Hello Radhika,

Apart from the things you have already tried, you can use methods like pre-allocation, vectorization and profiling to optimize your code further. Here is a sample code for the same:

% Assuming num_frames is defined and ss is an index array
num_frames = size(k_data_shift, 3);
k_d1 = zeros(size(k_data_shift));

% Initialize the first frame
k_d1(:, ss(1), 1) = k_data_shift(:, ss(1), 1);

% Iterate over frames
for j = 2:num_frames
    % Accumulate the sum from the previous frame
    k_d1(:, :, j) = k_d1(:, :, j-1);
    
    % Update specific slices
    % Vectorize the inner loop if possible
    k_d1(:, ss(1:num_frames), j) = k_data_shift(:, ss(1:num_frames), j);
end

You can find more details on the above-mentioned techniques in the below links:

• Pre-allocation: https://www.mathworks.com/help/matlab/matlab_prog/preallocating-arrays.html
• Vectorization: https://www.mathworks.com/help/matlab/matlab_prog/vectorization.html
• Profiling: https://www.mathworks.com/help/matlab/matlab_prog/profiling-for-improving-performance.html

I hope this helps!