Re: My vectorized code is slower than a loop

[Nicholas Jankowski]

On Mon, Oct 17, 2016 at 3:37 PM, Dave Cottingham 2
<address@hidden> wrote:
> In the process of writing a routine to compute the Qn statistic of Croux and
> Rousseeuw, I started from the Fortran code they published in their paper
> (C&R, "Time-efficient algorithms for two highly robust estimators of scale")
> and ran into unexpected trouble with one section. Here's a fairly literal
> translation of the original, where x is an array of length n, and left and
> right are also arrays of length n that contain integers:
>
> w = zeros(1, n);
> j = 1
> for i = 1:n
>   if(left(i) <= right(i))
>     for jj = left(i):right(i)
>       w(j) = x(i) - x(n+1-jj);
>       j += 1;
>     endfor
>   endif
> endfor
>
> Before we get to this section, left and right are such that at most n items
> will get copied into w.
>
> Not surprisingly, this code is pretty slow. I see no way to eliminate the
> outside loop on i; but getting rid of the inside loop on jj is
> straightforward:
>
> w = zeros(1, n);
> j = 1
> for i = 1:n
>   if(left(i) <= right(i))
>     w(j:j+right(i)-left(i)) = x(i) - x(n+1-[left(i):right(i)]);
>     j = j + right(i)-left(i)+1;
>   endif
> endfor
>
> Problem is, the second, vectorized, code takes twice as long as the first,
> non-vectorized, code.
>
> Anybody have an idea why? I'm stumped.
>
> I have attached a script that contains these two snippets, some setup, and
> tics and tocs, for anyone who wants to play with it.
>
> Thanks,
> Dave Cottingham


Don't have time to test at the moment, but have you tried using the
profiler? Octave's profiler isn't as thorough as Matlab's, but it
should be able to show you where your scripts are spending the most
time, and maybe why the latter is slower.

profile on;

*run script*

profile off

profshow

and/or

profexplore