quasi-cumsum by matrix indexes

Here's an example of what I'm trying to do:

indexes = randi(4,1000,10);
K       = rand(4,6);
Q       = zeros(1000,10+6);

for k = 1:10
Q(:,k:k+5) = Q(:,k:k+5) + K(indexes(:,k),:);
end

Any idea if this is vectorizable? THANKS!

Answer 1

[Sean de Wolski]

Why bother?

That loop takes less than seven 10,000ths of a second on my system. You could run this loop billions of times in the time it might take you to possibly gain a 10000th of a second by vectorizing.

More

Even more reason to not vectorize since memory requirements will be huge and slow it down more than a for-loop.

Your numbers didn't scale up but here's the loop with comparable numbers run the same amount of times:

indexes = randi(100,3000,150);
K       = rand(3000,47);
Q       = zeros(3000,1000+46);
tic
for jj = 1:10
  for k = 1:100
      Q(:,k:k+46) = Q(:,k:k+46) + K(indexes(:,k),:);
  end
end
toc

Elapsed time is 1.330083 seconds.

I would guess any vectorization, if possible, would be significantly slower due to the memory requirements.

Answer 2

[darius]

Hi Sean, thanks for your answer. Note the numbers are just for illustration. Replace 4,1000,10, and 6 with 100, 3000, 150, and 47, then loop the loop 1000 times, and you'll see why vectorization might be helpful.