Consider this code:

clear; close all;
n = 1e6;
m = 1e3;
s = 1e-3;
p = round(s^2 * n * m);
I = randperm(n, p);
J = randperm(m, p);
phi = sparse(I, J, 1, n, m);
Gx = randn(n, 1);
for k = 1:3
tic
L = Gx .* phi;
toc
end
Gx(1) = NaN;
for k = 1:3
tic
L = Gx .* phi;
toc
end

Now consider the output on my macbook pro M1 running Matlab 2024a (so through rosetta).

Elapsed time is 0.001591 seconds.
Elapsed time is 0.000291 seconds.
Elapsed time is 0.000282 seconds.
Elapsed time is 0.573762 seconds.
Elapsed time is 0.548177 seconds.
Elapsed time is 0.560880 seconds.

Notice the very large difference in computation times between the no-NaN version of Gx and the version that contains a single NaN value.

The solution is quite obvious, just remove the NaN values with your favorite method (isnan, isfinite, etc.). So this is not my question.

What intregues me is:
why the computation time changes so drastically?Consider this code:

clear; close all;
n = 1e6;
m = 1e3;
s = 1e-3;
p = round(s^2 * n * m);
I = randperm(n, p);
J = randperm(m, p);
phi = sparse(I, J, 1, n, m);
Gx = randn(n, 1);
for k = 1:3
tic
L = Gx .* phi;
toc
end
Gx(1) = NaN;
for k = 1:3
tic
L = Gx .* phi;
toc
end

Now consider the output on my macbook pro M1 running Matlab 2024a (so through rosetta).

Elapsed time is 0.001591 seconds.
Elapsed time is 0.000291 seconds.
Elapsed time is 0.000282 seconds.
Elapsed time is 0.573762 seconds.
Elapsed time is 0.548177 seconds.
Elapsed time is 0.560880 seconds.

Notice the very large difference in computation times between the no-NaN version of Gx and the version that contains a single NaN value.

The solution is quite obvious, just remove the NaN values with your favorite method (isnan, isfinite, etc.). So this is not my question.

What intregues me is:
why the computation time changes so drastically? Consider this code:

clear; close all;
n = 1e6;
m = 1e3;
s = 1e-3;
p = round(s^2 * n * m);
I = randperm(n, p);
J = randperm(m, p);
phi = sparse(I, J, 1, n, m);
Gx = randn(n, 1);
for k = 1:3
tic
L = Gx .* phi;
toc
end
Gx(1) = NaN;
for k = 1:3
tic
L = Gx .* phi;
toc
end

Now consider the output on my macbook pro M1 running Matlab 2024a (so through rosetta).

Elapsed time is 0.001591 seconds.
Elapsed time is 0.000291 seconds.
Elapsed time is 0.000282 seconds.
Elapsed time is 0.573762 seconds.
Elapsed time is 0.548177 seconds.
Elapsed time is 0.560880 seconds.

Notice the very large difference in computation times between the no-NaN version of Gx and the version that contains a single NaN value.

The solution is quite obvious, just remove the NaN values with your favorite method (isnan, isfinite, etc.). So this is not my question.

What intregues me is:
why the computation time changes so drastically? sparse, mtimes MATLAB Answers — New Questions

​