slow re-assignment of values to an array in a structure

Here is my problem: Why does line 6 below take two orders of magnitude more time to compute than other instances of the same command?

1.>> x = ones(100,100,250);

2.>> tic; x(:,:,23) = rand(100,100) ; toc

Elapsed time is 0.000433 seconds.

3.>> tic; x(:,:,23) = rand(100,100) ; toc

Elapsed time is 0.000420 seconds.

4.>> tic; x(:,:,23) = rand(100,100) ; toc

Elapsed time is 0.000417 seconds.

5.>> s.x = x;

6.>> tic; x(:,:,23) = rand(100,100) ; toc

Elapsed time is 0.013906 seconds.

7.>> tic; x(:,:,23) = rand(100,100) ; toc

Elapsed time is 0.000435 seconds.

8.>> tic; x(:,:,23) = rand(100,100) ; toc

Elapsed time is 0.000537 seconds.

*******

In my application I send a large structure of arrays into a subroutine and I observe it running super slowly and I think this is the issue.

Any explanation?

Answer 1

[Matt Fig]

Sorry, at first I didn't read the whole question.

My guess is that line 5 doesn't actually copy the data in x. When you change x the copy happens. Before the copy happens, s.x only "points" to the values in x. This is a MATLAB "efficiency" known as copy-on-write.

See Loren's response number 6.

.

.

EDIT

Here is how really to see this:

format debug
x = [2,2] % Look at pr,  this is the pointer in memory
s.x = x;  % Make assignment...
s.x  % Same pr here as above for x.
% Now change x and look at the pointers...
x = 5;  % Make change.
x
s.x % Different pr now

[Jiro Doke]

Exactly.

If you put tic/toc around steps 1 and 5, you'll notice that step 5 takes very little time because it's not really creating a copy of "x". and the extra amount of time taken in step 6 is comparable to the amount of time it takes to create an array of ones (step 1).

[Matt Fig]

Nice approach, Jiro! Here is even more evidence of what happens, as if we needed any at this point:

>> clear all, pack

>> tic,x = rand(10000);toc

Elapsed time is 5.906108 seconds.

>> tic,y = x;toc

Elapsed time is 0.000041 seconds.

>> tic,x(1) = 3;toc

??? Out of memory. Type HELP MEMORY for your options.

Now you just gotta ask, how could this have run MATLAB out of memory?

[Jiro Doke]

Nice.

Answer 2

[Jonathan]

Thanks guys! Much appreciated. So it is the accursed "copy on write" thing. If only I could pass array pointers to Matlab subroutines.....