Why is diag faster than zeros?

I get some surprising timings when i allocate a zero matrix using zeros and diag. For instance when i run the code:

N = 10000;
tic
A = diag(zeros(1,N));
toc
tic
B = zeros(N);
toc

I get the output:

Elapsed time is 0.007424 seconds.
Elapsed time is 0.181349 seconds.

Suggesting that that it should be 24 times faster to allocate using a zero diagonal diag comapred to just using zeros. I find this very weird so I was wondering if someone could explain/educate me on what is happening? :)

Thanks,

Peter

[Walter Roberson]

Did you cross-check with timeit() from the FEX ? With two different functions? It is possible that time is being mis-accounted .

[Peter]

I do not know about timeit(), but using cputime i get a similar result only know the factor is only 3-4.

Answer 1

[James Tursa]

More discussion related to this can be found on this thread:

http://www.mathworks.com/matlabcentral/answers/58055-faster-way-to-initilize-arrays-via-empty-matrix-multiplication

There is no documentation on how the various functions and operations allocate variables, but the guess is that some functions and operations draw from a pool of previously zero'ed out memory while others don't, hence the timing differences.

Answer 2

[Jan]

The JIT acceleration can eliminate dead code. So please try:

N = 10000;
tic
for k = 1:100
  A = diag(zeros(1,N));
  A(1) = 1;
  clear('A');
end
toc

tic
for k = 1:100
  B = zeros(N);
  B(1) = 1;
  clear('B');
end
toc

Now copy this inside a function M-file and compare the times again.

Dummy loops do not reflect the behavior in a real program necessarily.