MATLAB Answers

Compare inv(X) Vs mex inv (X)

9 views (last 30 days)
zohar
zohar on 8 Jul 2012
Hi all
I wrote a simple program which calculates inverse matrix size (100*100).
function y = test_inv(x) %#codegen
% test inverse matrix
y = inv(x);
end
I created mex function test_inv_mex and tested running time.
x = randn(100);
tic;
for n = 1: 100
y = test_inv(x);
end
toc;
tic;
for n = 1: 100
y = test_inv_mex(x);
end
toc;
This is my resulte :
Elapsed time is 0.054418 seconds.
Elapsed time is 0.613990 seconds.
My question is :
Why the mex function run much slower than regular one ?
B.R
Zohar

  4 Comments

Show 1 older comment
Jan
Jan on 8 Jul 2012
@zohar: Please explain, how you have "created the mex function". It matters if you have a hand-coded function, call the optimized ATLAS or BLAS functions, or if you let the code generator do this, which is implied by the comment "%#codegen" appearing in your code. Good answers require no guessing of important details.
@Yash: Without doubt you a right. But zohar's question concerned another topic.
zohar
zohar on 8 Jul 2012
Hi Yash,
I understand the disadvantage using inv(X).
Why the mex function run much slower than the regular one ?
zohar
zohar on 8 Jul 2012
Hi Jan,
I am new with codegen, but I saw that if I unchecked everything under project setting general the mex function run faster but still slower than regular one! (but warnnings appears ).
Any help ?

Sign in to comment.

Accepted Answer

Mike Hosea
Mike Hosea on 9 Jul 2012
Edited: Mike Hosea on 10 Jul 2012
I think it's rather likely with a difference of that magnitude that you have left memory integrity checks on. Try
>> cfg = coder.config('mex');
>> cfg.IntegrityChecks = false;
>> cfg.ResponsivenessChecks = false;
>> codegen test_inv -args {zeros(100)} -config cfg
>> x = randn(100);
tic;
for n = 1: 100
y = test_inv(x);
end
toc;
tic;
for n = 1: 100
y = test_inv_mex(x);
end
toc;
Elapsed time is 0.082537 seconds.
Elapsed time is 0.109494 seconds.
The remaining difference is not surprising since there is some overhead and we are essentially just comparing the compiled and optimized MATLAB implementation to the compiled MATLAB Coder implementation. MATLAB Coder is not intended to speed up MATLAB linear algebra. They are mostly compiled and highly optimized routines. If it could do much of that, we'd just speed up the MATLAB implementation. If you want to see it speed something up, you usually need to have a substantial amount of your code in there. Also, in this particular case there are some differences in the algorithms. MATLAB Coder is generating unblocked code, and here the matrix is large enough that I would rather use a blocked algorithm.

  1 Comment

zohar
zohar on 11 Jul 2012
Hi Mike, Thanks. +1

Sign in to comment.

More Answers (0)

Sign in to answer this question.

Tags