MATLAB Answers

0

Different speed of execution of the same code in different versions Matlab. 2014b and 2017a

Nickolay Ternovoy さんによって質問されました 2018 年 6 月 3 日
最新アクティビティ Walter Roberson
さんによって コメントされました 2018 年 6 月 4 日
Hello. The question is this. Why are the same code executed at different speeds in different versions of matlab? Versions 2014b and 2017a.
Below are screenshots with the difference in code execution in different versions.
I noticed one strange thing. If I write code without declaring a function, then the difference in execution between versions is huge.
If the code is described as a function, then speed is greatly increased
code without declaring a function:
clear
b = rand(1000000,10);
i=1;
tic
while i <= 1000
res = b(:,1).*b(:,10);
i = i+1;
end
toc
result Matlab 2014b :
Elapsed time is 1.962136 seconds.
result Matlab 2017a :
Elapsed time is 8.741258 seconds.
code with function:
function f=Untitled
clear
b = rand(1000000,10);
i=1;
tic
while i <= 1000
res = b(:,1).*b(:,10);
i = i+1;
end
toc
end
result Matlab 2014b :
Elapsed time is 1.757147 seconds.
result Matlab 2017a :
Elapsed time is 1.885382 seconds.
p.s. Untitled is the name of the script / function

  6 件のコメント

Thanks for the advice. Untitled is the name of my function / script.
Code and results of execution will update in title p.s. updated screenshots
"A script has access to the base workspace"
Only for scripts run from the command line (and a few other cases). The documentation states that "When you call a script from a function, the script uses the function workspace", which is not the base workspace.
@Stephen, I was already corrected by Walter. I misspoke because I only use scripts in the base workspace for debugging, so for all my use cases it is in fact the base workspace. But, yes, you are correct, it's not actually the base workspace necessarily, but the workspace of the calling function.

サインイン to comment.

製品


リリース

R2017a

2 件の回答

Answer by Philip Borghesani on 4 Jun 2018
 Accepted Answer

This feels like a bug to me and I filed a bug report. Walter is correct that coding this as a function is expected to produce faster code but there is no reason for this big a performance difference.
If you can't easily convert your code to a function I suggest posting more information about the problem you are trying to solve with actual code, or creating a support call for the best solutions to this performance issue.

  3 件のコメント

For me it remains unclear why the code execution as a script and as a function in the Matlab2014b faster than in version Matlab2017a
I'll try to file a bug report. Thanks for the advice.
p.s. as I understood the bug report is carried out through technical support. but for me technical support is not possible, because I am a student license holder
When I add timeit() calls on the multiplication, I can see that they take essentially the same time for script or function.
I saw some hints that it might have to do with the management of the variable the result is stored in, but the connection was not strong and not nearly enough to account for the timing differences.
Philip is staff, and has filed a bug report about it already.

サインイン to comment.


Answer by Walter Roberson
on 3 Jun 2018

The Just In Time engine historically compiled functions more than it compiled scripts. This had to do with the fact that scripts were more free to "poof" variables into existence, so decisions involving variables that might be set in a script call had to be made at run time, whereas static analysis for functions could be more thorough.
In R2015b a new Execution Engine started making more assumptions about what was happening in scripts, and started declaring that some potential changes in scripts would no longer be paid attention to or would now be errors. As a result, performance of code that included scripts improved.

  2 件のコメント

performance of code that included scripts improved
But not in the case mentioned by the OP:
2014b: 1.962136 seconds.
2017a: 8.741258 seconds.
Ah, yes, I am able to replicate the script slowdown for that code between R2014a and R2018a.

サインイン to comment.



Translated by