Vectorize for-loop that changes overlapping parts of an array

Hi everyone,

I have problem speeding up my code. I have a very large (20-40m datapoints) logical array and at every position where it reads "1", the next N datapoints should be set to "1" a well.

I have the following code that is quite slow since it has a for loop:

N=5000-1;
logarray; %logical array of size 40000000x1 
inx=find(logarray);  %inx can be quite large as well, 1-5m points
for ii=1:length(inx)  
    logarray(inx(ii):(inx(ii)+N))=1;
end

This code works, but is very slow (more than 10 secs). I tried something like

 logarray(find(inx):(find(inx)+100))=1;

but this did not work. I was wondering whether vectorisation would be a solution here? Mind that overlapping parts of the logical array are changed by the loop.

Does anyone have a good suggestion?

Thanks, Maximilian

Answer 1

[Jos (10584)]

Easy when using convolution:

logarray = logical([0 1 1 0 0 0 1 0 0 0 1 0 0 0 0 1 0 1 0 0 0 0 1])
N = 2 ;
out = conv2(double(logarray(:)), ones(N+1,1)) ;
out = out(1:numel(logarray))>0 ;
out = reshape(out,size(logarray)) ; 
disp([logarray(:) ; out(:)]) ;

[Jos (10584)]

taking about 2 seconds on my old mac for a 3e7 elements and N = 5000

[Maximilian Rüppell]

Works like a charm - thanks!

Answer 2

[Roger Stafford]

It isn't perhaps that you have a for-loop that is slowing you down so much as that your loop is repeating the writing of ones so much in overlapping intervals. You might try the following which first combines overlapping intervals before doing the writing of ones.

   N=5000-1;

   f1 = find(logarray);
   f2 = min(f1+N,length(logarray));
   f3 = zeros(size(f1));
   f4 = zeros(size(f2));
   ie = 1;
   f3(ie) = f1(ie);
   f4(ie) = f2(ie);
   for ib = 1:size(f1,1)-1
     if f1(ib+1) > f4(ie)+1
       ie = ie+1;
       f3(ie) = f1(ib+1); 
     end
     f4(ie) = f2(ib+1);
   end
   f3 = f3(1:ie); % The intervals defined by f3 and f4 don't overlap
   f4 = f4(1:ie);
   for ib = 1:ie
     logarray(f3(ib):f4(ib)) = 1;
   end

[Roger Stafford]

I thought of another way you could try which uses 'accumarray'. Assume 'logarray' is a column vector.

   N = 1000;
   f = find(logarray);
   t = ones(size(f));
   logarray = min(cumsum(accumarray([f;f+N+1],[t;-t],...
              [length(logarray)+N+1,1])),1);
   logarray(end) = [];