I'm running the following code
a = normrnd(0,1,5000, 300);
b = normrnd(0, 1, 5000, 50);
tic
xp = repmat(a,1,1,50) .* permute(repmat(b, 1,1, 300), [1 3 2]);
toc
which outputs
Elapsed time is 0.425773 seconds.
Apparently the vast majority of execution time is spent on the repmat and permute part. I thought some alternative implementation with bsxfun might help to speed it up quite a bit, but I don't see exactly how. Is there any better way of dealing with this? Thank you!
You can reduce running time if instead of repmat you use singleton expansion, either implicit (requires Matlab R2016b or newer):
xp = a .* permute(b, [1 3 2]);
or explicit with bsxfun:
xp = bsxfun(@times, a , permute(b, [1 3 2]));
Results on my machine, using timeit (which is more accurate than tic, toc):
>> a = normrnd(0,1,5000, 300);
>> b = normrnd(0, 1, 5000, 50);
>> timeit(@() repmat(a,1,1,50) .* permute(repmat(b, 1, 1, 300), [1 3 2]))
ans =
0.706580436900000
>> timeit(@() a .* permute(b, [1 3 2]))
ans =
0.167270436900000
>> timeit(@() bsxfun(@times, a , permute(b, [1 3 2])))
ans =
0.161594036900000
The likely explanation of these reduced times is that in the repmat approach the extended arrays are explicitly constructed, which requires memory allocation.