Why are the 2 lines at the end so time consuming?

Hi, I just ran the profiler on a function and was quite surprised.

EDIT: New run

function outputVector = addAndReplace(inputVector,addVector,atElement,addValue)

0.25   10000   18 if not(exist('addValue','var')) 
               19     addValue = 0;
               20 end
               21 
0.01   10000   22 if isscalar(atElement) 
               23     atElement = [atElement,atElement];
               24 end
               25 
< 0.01   10000   26 len = length(inputVector)+length(addVector)-1-(atElement(2)-atElement(1)); 
0.03   10000   27 sizeVector = size(inputVector); 
< 0.01   10000   28 if sizeVector(1) > 1; 
               29     outputVector = zeros(len,1);
< 0.01   10000   30 else 
2.28   10000   31     outputVector = zeros(1,len); 
< 0.01   10000   32 end 
0.15   10000   33 outputVector(1:atElement(1)-1) = inputVector(1:atElement(1)-1); 
0.07   10000   34 outputVector(atElement(1):atElement(1)+length(addVector)-1) = addVector; 
12.55   10000   35 outputVector(atElement(1)+length(addVector):end) = inputVector(atElement(2)+1:end)+addValue;

The function was created to replace elements between adElement(1) and adElement(2) (start and end index) with the vector addVector. and if necessary, add a value addValue to the last elements. This to save myself some headache while creating and debugging the code.

Is there anyone who knows why the final lines takes much more time to execute than creating an empty matrix? I have heard something about matlab creates one pointer and copies it for every element when creating a zeros matrix, can this be the reason why line 30 is faster?

BR/ Patrik

[Patrik Ek]

Thankful for answer BR/Patrik

[Niklas Nylén]

Can you provide the input data that you called the function with, i.e. the values of InputVector, addVector, atElement and addValue.

[Patrik Ek]

this example should do

inputVector = 1:100000;
addVector = 1:100;
atElement = [500 650];
addValue = 4;
profile clear; profile on;
for k = 1:10000
  addAndReplace(inputVector,addVector,atElement,addValue);
end
profile viewer;

[Patrik Ek]

Do anyone know where to get these answers at least? That would be appreciated as well, since I understand that this requires a fundamental understanding of matlab.

[Sara]

Why is there this for loop:

for k = 1:10000
  addAndReplace(inputVector,addVector,atElement,addValue);
end

Since you don't use k, the for loop is useless.

[Patrik Ek]

Probably, but I need to run the function enough time to properly show the issue. One time is to short.

[Niklas Nylén]

I do not have an answer, but interestingly row 34 is very quick for me (0.15 s) while row 36 is very time consuming (7.98 s). Could it be a difference in the input data or will different Matlab versions/OS's handle the code differently?

[Patrik Ek]

@Niklas Nylén, sorry I take the blame for this one. The question was old and the original run was lost in time. I edit so that the example is updated.

Answer 1

[Matt J]

SUBSREF operations are costly. Because you are extracting data from inputVector, memory allocation is required. A whole new vector gets allocated/created to hold extracted data such as

 inputVector(atElement(2)+1:end)

and

 inputVector(1:atElement(1)-1)

[Patrik Ek]

Thanks for that! Now if you have time for a following question, is there a solution to that? I will still accept the answer right now by the way.

[Matt J]

Depends on what is really changing throughout the loop. In your example, nothing is.

Answer 2

[per isakson]

I run your example on R2013a 64bit, Win7, and and old vanilla desktop. Our line number differs. You stripped off 18 lines in the beginning of the function?

Comments

• obsolete it is remarkable that my line 18 (your 34) in my case use very little time.
• obsolete my line 16 use much more time than your corresponding line.
• have you tried to use Windows The Task Manager to see if copies of the vector are created?
• I'm not surprised that manipulating arrays takes time

The working of the Matlab acceletator is a secret of The Mathworks. However, see Array resizing performance and Internal Matlab memory optimizations at Undocumented Matlab. And there was a post on performance and "In-place Operations on Data" at Loren on the Art of MATLAB. (At the moment it says error 404)

 2.07   10000   16         outputVector = zeros(1,len); 
 0.01   10000   17     end 
 0.18   10000   18     outputVector(1:atElement(1)-1) = inputVector(1:atElement(1)-1); 
 0.03   10000   19     outputVector(atElement(1):atElement(1)+length(addVector)-1) = addVector; 
 13.37   10000   20     outputVector(atElement(1)+length(addVector):end) ... 
                 21     = inputVector(atElement(2)+1:end)+addValue;
  0.35   10000   22 end

[Patrik Ek]

Sorry about the times, see edit. And regarding the missing lines I stripped of the help part to keep the amount of text down in the question.

Answer 3

[Sara]

I don't know why it's slow on those line, but since you know the pieces of the final array, instead of calculating its dimension, you can do:

p1 = inputVector(1:atElement(1)-1);
p2 = addVector;
p3 = inputVector(atElement(2)+1:end)+addValue;
if(size(inputVector,1) == 1)
    outputVector = [p1,p2,p3];
else
    outputVector = [p1;p2;p3];
end

It's faster than your code and it does the same thing.

[Patrik Ek]

I see, I guess I will do some changes. Thanks a lot!