speed up the code

hi everyone.

can you help me speed up this code?

    tic
    InputOverlap = magic(64)
    SDR_Overlap = InputOverlap;
    SDR = (zeros (64,64)) ;
    Radius = 2;
    InputOverlap = [InputOverlap(:,1:Radius) InputOverlap InputOverlap(:,end+1-Radius:end)];
    InputOverlap = [InputOverlap(1:Radius,:) ; InputOverlap ; InputOverlap(end+1-Radius:end,:) ];

        for r=1:64

        for c=1:64

            Neighbour= InputOverlap(r:r+2*Radius,c:c+2*Radius);
            Kmax = max(Neighbour(:)) ;
            if (SDR_Overlap(r,c)>0)&(SDR_Overlap(r,c)>=  Kmax) 
              SDR(r,c) = 1;
            else
              SDR(r,c) = 0;
            end

        end
        end
        toc

Thanks.

[Jan]

Start with omitting:

 else
   SDR(r,c) = 0;

SDR is initialized to zero already.

The editor should show a hint that && is more efficient than &. Consider these MLint messages.

The main part of your code happens before the loop. Most of all displaying the magic matrix is slow. I guess, you want to measure the time inside the loop only, don't you?

[Tiki Tiki]

Yes. My problem is time in the loop. I remove SDR(r,c) = 0 by setting it is zeros before loop.

But time consumes still high. How can I remove loop in this case?

Please help me. Thank.

Answer 1

[Jan]

This is slightly faster:

tic
for r = 1:64
  for c = 1:64
      Neighbour = InputOverlap(r:r+2*Radius, c:c+2*Radius);
      Kmax      = max(Neighbour(:));
      SDR(r,c)  = (SDR_Overlap(r,c) >= Kmax);
  end
end
toc

Is the SDR_Overlap(r,c)>0 test useful?

[Tiki Tiki]

Yes. it is a little faster. i also remove SDR_Overlap(r,c)>0. but this code still consume much time.

so i need optimze more. Can you help me how to remove loop in this case?

I have gpu. but dont undertand to use it.

[Jan]

Use movmax to replace the loops.

Does the padding of the input matrix belong to the problem? With movmax and 'EndPoints' set to 'shrink' you can omit the padding.

can you post some real input data? Especially the dimensions matter.