Clear Filters
Clear Filters

For loop faster than vectors?

29 views (last 30 days)
Jérôme
Jérôme on 17 Sep 2021
Commented: Image Analyst on 18 Sep 2021
I was thinking that for loops were usually slower than vector or matrix operations, but a simple example showed me the opposite.
Here is a comparison of different implementations of a counter. The results for five runs of the code below are the following:
0.0730 0.0608 0.0691
0.0705 0.0543 0.0669
0.0745 0.0546 0.0709
0.0756 0.0571 0.0798
0.0755 0.0573 0.0729
I am surprised to see that the function equation1 (multiplication of a vector by a scalar) is slower than counter (for loop).
I am also surprised to see that the function equation1 is slower than equation2, just because of the vector is stored in a variable.
Does anyone have an explanation?
N = 10^7;
step = 0.123456789;
f1 = @() equation1(step, N);
t1 = timeit(f1);
f2 = @() equation2(step, N);
t2 = timeit(f2);
f3 = @() counter(step, N);
t3 = timeit(f3);
disp([t1 t2 t3])
% Functions under test
function out = equation1(step, N)
n = 1:N;
out = n * step;
end
function out = equation2(step, N)
out = (1:N) * step;
end
function out = counter(step, N)
out = zeros(1,N);
out(1) = step;
for n = 2:N
out(n) = out(n-1) + step;
end
end
  1 Comment
Jérôme
Jérôme on 17 Sep 2021
I just noticed that the results are quite different when using Run an Time. With the code below, the results are:
  • test>counter: 1.055 s
  • test>equation1: 0.097 s
  • test>equation2: 0.077 s
I guess in such situations, tic toc and timeit are more accurate measurements than the profiler.
N = 10^7;
step = 0.123456789;
out1 = equation1(step, N);
out2 = equation2(step, N);
out3 = counter(step, N);
% Functions under test
same as in the original post

Sign in to comment.

Accepted Answer

Image Analyst
Image Analyst on 17 Sep 2021
I'm not surprised. See my attached demo:
Vectorized won 403 times out of 1000 = 40.3%.
For loop won 597 times out of 1000 = 59.70%.
The average time for the vectorized method = 7.3006e-06.
The average time for the for loop method = 1.40819e-05.
It often depends on how your vectorization was done and how many elements there were. It might be one way for a handful of elements but the opposite for tens of millions of elements.
  2 Comments
Jérôme
Jérôme on 17 Sep 2021
Thanks for your feedback.
I guess this common rule of avoiding for loops and trying to vectorize operations as much as possible should not be considered a general truth anymore, due to the evolutions and improvements of Matlab over time.
Image Analyst
Image Analyst on 18 Sep 2021
@Jérôme, yes it's a myth that just won't die. That said, quite often or usually vectorized is faster, and makes for smaller, cleaner code that is often, though not always, more intuitive and easier to read and understand.

Sign in to comment.

More Answers (0)

Products


Release

R2021a

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!