Re: vectorization help

[Jordi Gutiérrez Hermoso]

On 3 October 2011 18:26, Liam Groener <address@hidden> wrote:
> On Oct 3, 2011, at 12:10 PM, William Krekeler wrote:
>
> Thanks for the useful and very informative response. That definitely works.
> I'll have to read more on broadcasting to see if I can apply it elsewhere.
>
> My code has to be cross-compatible with ML. As an FYI the bsxfun version
> will work in matlab 7.8.0.347:
>    Cube = bsxfun(@rdivide, permute(RC, [1, 3, 2]), Cube);
>
> I'm curious; does this actually save any significant time over your for loop
> version?

Actually, it doesn't. It's in fact slower! But I don't want you to
believe this means "broadcasting is slower". In fact, we have a
problem with the implementation of bsxfun right now, which I want to
fix. It turns out the order of the dimensions affects the speed:

    rc = rand( 1, 1, 448 );
    cube = rand( 160, 512, 448 );

    p = perms([1,2,3]);

    for i = 1:rows(p)
      tic
      x = bsxfun(@rdivide, permute(rc,p(i,:)), permute(cube, p(i,:)));
      toc
    endfor

I have investigated the problem, and I'm convinced that the problem is
the number of function calls to divide changes depending on the
dimensions of the inputs. I will work on fixing this over the next few
days. Looks like a fun problem.

But this doesn't mean "don't use bsxfun". This slowness is a minor
problem that we can fix in the implementation, which is another reason
to vectorise code: when the code is vectorised, it's much easier to
optimise in Octave's C++ implementation, so you can gain future
optimisation if you vectorise today.

- Jordi G. H.