[Rcpp-devel] Most efficient way to apply function over rows of a matrix with Rcpp?

[Kevin Ushey]

On Thu, Jan 17, 2013 at 7:25 PM, Dirk Eddelbuettel <edd at debian.org> wrote:

>
> Hi Kevin,
>
> First off, congrats regarding RcppRool. Nice to see another package on
> CRAN. I had some follow-up questions as to whether you planned to push this
> further: the mean at n+1 is really just adding 1/n of the difference
> between
> the newest and the oldest... but getting these things right is hard.
> Anyway,
> I didn't really have time poke around so consider these questions not
> asked :)
>

Thanks! I do plan on extending the package further, but I need to iron out
some of the bugs and makes the pseudo-API available within more accessible
first. I'm definitely open to new ideas / additions though.

>
> On 17 January 2013 at 18:09, Kevin Ushey wrote:
> | Hi guys,
> |
> | I'm trying to use Rcpp to implement an 'apply'-type function for
> matrices,
> | whereby I apply a function to each row or column of a matrix. I'm
> testing it
> | here with the 'mean' just to see how well we can do vs. the highly
> optimized
> | rowMeans, colMeans functions, but of course any other function taking a
> vector
> | and returning a scalar might fit here.
> |
> | I'm wondering if I can do better. Please see the gist in the link here;
> you can
> | sourceCpp it in R.
> |
> | https://gist.github.com/4561281
> |
> | I try to limit the amount of copying as much as possible with
> | NumericMatrix::Column and NumericMatrix::Row.
> |
> | Note that the Rcpp solution over columns is just as fast as colMeans,
> but over
> | rows it's a fair bit slower relative to rowMeans. Is there any way I
> could
> | improve this?
>
> I never measured. I always stopped at the simple sapply/lapply
> implementations I showed in the talks/workshops.
>
> One problem is column-wise storage of vectors. For kicks, can you see what
> happens when for rowMeans you do a transpose and then compute colMeans ?
>

It's still faster to just compute the rowMeans compared to column means on
the transpose, for different sizes of matrices... Interestingly though, the
Rcpp rowMeans actually runs faster (!) than base-R rowMeans for a matrix w/
1E4 rows, 1E2 columns, so the relative speed does depend on the size /
shape of the matrix. I'm always a little surprised to learn these things.

Regardless, an Rcpp 'apply' solution is almost always faster than the
corresponding R 'apply' solution.


>
> | I plan to submit an expanded version of this to the Rcpp gallery so any
> advice
> | is appreciated!
>
> That;s the spirit -- let's grow http://gallery.Rcpp.org which is meant to
> be
> open for contributions.
>

I will have something nifty to submit tomorrow. :)


>
> Cheers,  Dirk
>
> --
> Dirk Eddelbuettel | edd at debian.org | http://dirk.eddelbuettel.com
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.r-forge.r-project.org/pipermail/rcpp-devel/attachments/20130117/eb491f5a/attachment-0001.html>