Thread Subject:

Slow repmat

Can someone explain why these performance numbers are an order of magnitude different?

x = [1 2 3]';

>> tic; for i = 1:100000; repmat(x,1,10).^2;end;toc
Elapsed time is 4.612648 seconds.

>> tic; for i = 1:100000; x(:,ones(1,10)).^2;end;toc
Elapsed time is 0.354172 seconds.

"Abel" wrote in message <k5puo6$eq4$1@newscl01ah.mathworks.com>...
> Can someone explain why these performance numbers are an order of magnitude different?
>
> x = [1 2 3]';
>
> >> tic; for i = 1:100000; repmat(x,1,10).^2;end;toc
> Elapsed time is 4.612648 seconds.
>
> >> tic; for i = 1:100000; x(:,ones(1,10)).^2;end;toc
> Elapsed time is 0.354172 seconds.


The first is implemented as an mfile. The second one runs builtin code.

"Matt J" wrote in message <k5pvgi$h86$1@newscl01ah.mathworks.com>...
> The first is implemented as an mfile. The second one runs builtin code.

that's obvious, but I'm not quite sure that explains an order of magnitude performance difference.

"Abel" wrote in message <k5rcm1$fee$1@newscl01ah.mathworks.com>...

> that's obvious, but I'm not quite sure that explains an order of magnitude performance difference.

Of course that obviously explains. The overhead of repmat takes T to perform, the memory duplication take T/10 with the small case you test

Try similar test to replicate (1000 x 1) to (1000 x 1000) matrix and see what you get.

Bruno

so is it .m-file overhead vs. built-in or memory duplication or both?

> so is it .m-file overhead vs. built-in or memory duplication or both?

You can edit repmat.m and take a look the code, it calls a(one(...),...) at the end. The rest of the time is error checking, parse input, meaning overhead.

Bruno

Thanks Bruno. I really appreciate your comments.

I'm just finding it hard to digest that a wrapper function (repmat) has such an impact on performance.

On 10/19/2012 11:55 AM, Abel wrote:
> Thanks Bruno. I really appreciate your comments.
> I'm just finding it hard to digest that a wrapper function (repmat) has
> such an impact on performance.

Might try to run profile on repmat() and see where the time is spent
inside it...

--

On 10/19/2012 11:55 AM, Abel wrote:
> Thanks Bruno. I really appreciate your comments.
> I'm just finding it hard to digest that a wrapper function (repmat) has
> such an impact on performance.

Oh, besides my other comment, this isn't new--I get almost the same
ratio of 10:1 in R14 here...

--

"Abel" wrote in message <k5s0lc$46n$1@newscl01ah.mathworks.com>...
> I'm just finding it hard to digest that a wrapper function (repmat) has such an impact on performance.

The problem is that calling m functions inside loops is costly

If you want speed (and future) go to Julia

http://julialang.org/