In some specific case, preallocation is slower than doing nothing.

Hi,

Please see the below screenshot. I think in this specific case, preallocation does not help.

for k=1:5
    clearvars
    sig_smpl = 10001;
    num_mode = 3;

    tic
    for i=1:100
        eta = -0.5 + rand(sig_smpl,1);  % -0.5 ~ 0.5 random generation from uniform distribution
        h = -0.5 + rand(sig_smpl,1);
        
        for j=1:num_mode
            G_U(:, j) = conv(eta, h);
            G_L(:, j) = conv(h, eta);
        end
    end
    toc
end

%%

for k=1:5
    clearvars
    sig_smpl = 10001;
    num_mode = 3;
    G_U = zeros(sig_smpl * 2 - 1, num_mode);
    G_L = zeros(sig_smpl * 2 - 1, num_mode);
    
    tic
    for i=1:100
        eta = -0.5 + rand(sig_smpl,1);  % -0.5 ~ 0.5 random generation from uniform distribution
        h = -0.5 + rand(sig_smpl,1);
        
        for j=1:num_mode
            G_U(:, j) = conv(eta, h);
            G_L(:, j) = conv(h, eta);
        end
    end
    toc
end

What did I wrong?

p.s. Is there any way copy and past of my live script to here with results ('Output')?

[Walter Roberson]

Don't use clearvars in a loop: it interferes with optimization.

[Sangmin Lee]

@Walter Roberson Thanks for the advice!

Answer 1

[Stephen23]

"What did I wrong?"

Your timing comparison is 99% meaningless.

Your are comparing the times of 100 loop iterations (the i loop), but only on the first loop iteration does the array get expanded by the j loop, after that the G_U and G_L matrices already exist and their content will simply get updated. So for the other 99 loop iterations that you are timing (of the i loop) there is practically no difference between your two versions.

You will not get a meaningful comparison of that first loop (the only one where preallocation makes any difference), when its timing is merged in with 99 other loop iterations (where both versions are effectively preallocated, thus no meaningful difference in your comparison). In the end you are basically measuring OS noise of all 100 iterations, because for your small arrays that swamps everything else.

The i loop seems to achieve nothing anyway. Did you add it as an attempt to compare the timing?

For that matter, the j loop does not seem to serve any purpose either.

[Sangmin Lee]

Thanks!

Now I changed the code as below, and see the difference.

Probably preallocation only have evident advantage when the matrix is relatively large...

If I decrease num_mode to 100, the difference is negligible.

Thanks for the advice!