Speed up: Parfor loop vs Vectorization

I'm trying to speed up my code, which is bottlenecking here:

tic
FD = zeros(28, Rotations);
parfor i = 1:Rotations
    CodePR = rotate (CodeP, i*(Rotations - 1)*angle, [0 0]);
    for j = 1:28
        FD(j, i) = (area(intersect(intersect(CodePR, MaskP), WholeVD(j))))/(AreaFD(j));
    end
end
A = FD;
toc

I came up with that solution:

tic
FD = zeros(28, Rotations);
WholeVD = transpose(WholeVD);
AreaFD = transpose(AreaFD);
parfor i = 1:Rotations
    CodePR = rotate (CodeP, i*(Rotations - 1)*angle, [0 0]);
    FD(:,i) = area(intersect(intersect(CodePR, MaskP), WholeVD))./AreaFD;
end
B = FD;
toc

which is like 20 times (for my rig at least). The bad thing is that with actual dataset it consumes too much RAM (and eventually crashes matlab) so I can't evaluate that. Getting rid or PARFOR helps with memory problem but effectivly slower (by amount of CPU cores). Is there any walk arounds to reduce memory usage so I could utilize more threads?

The whole code is in attachment (fold everything, described part is not foldable). Thanks.

Answer 1

[Mohammad Sami]

R2020a introduced Threads based parallel pool. This may reduce the memory issues.

You can create a threads based parpool before using parfor.

pool = parpool("threads");

https://www.mathworks.com/help/releases/R2020a/parallel-computing/choose-between-thread-based-and-process-based-environments.html

[max fire]

Thanks, seems a legit way to improve memory usage. I have to update to 2020a to check this, tho.

[Raymond Norris]

Running your code (which I'm asusming isn't the entire number of iteration) ran under "threads", so I think that's a good approach to potentially solving your parfor memory issues. Both "local" and "threads" ran in the same amount of time.

[max fire]

Oh, that's nice of you, Raymond, thanks.