Speed up indexing / Repmat operation

I have to following code I'm running:

M=zeros(r,c);
M (:,I)=repmat(Q,1,sum(I));
end

I is a 1-by-c logical vector and Q is r-by-1

So, M starts out as all zeros, and I need certain columns of M (the ones where the value in I is equal to 1) to take on the values in vector Q.

The above code runs fine, but it is surprisingly slow (I'm doing this on a 2400-by-1000 matrix and looping 45 times.) About half of the time is spent doing the repmat() operation and half the time is spent on M (:,I)= part. I've tried replacing the "repmat" with a bsxfun() operation; it gives the same, correct result, but isn't really any faster.

Is there any way I can speed this up?

Answer 1

[Star Strider]

The bsxfun funciton is always faster than repmat in such situations. I’m not certain that I understand what you want ‘M’ to be.

See if this does what you want:

r = 7;                                          % Create Data
c = 5;                                          % Create Data
I = logical(randi([0 1], 1, c));                % Create Data
Q = randi(9, r, 1);                             % Create Data
M=zeros(r,c);
M = bsxfun(@times, Q, I);

[Jeff Ack]

This is perfect. Took the calculation time from 0.48 down to 0.2. Thanks!

[Star Strider]

My pleasure!

[James Tursa]

You don't really need the M=zeros(r,c) pre-allocation anymore, since the next line sets M directly.

Answer 2

[Thorsten]

 nnz(I)

may be faster than sum for logical matrices.

Another way could be to pre-allocate cells of ones for all possible values of sum(I), like

C{1} = 1;
C{2} = ones(1, 2);
C{3} = ones(1, 3);

and so on

and then use

M (:,I) = Q*C{sum(I)};