[Rcpp-devel] Is there a way to pass references to R

Dirk Eddelbuettel edd at debian.org
Sat Dec 22 01:40:49 CET 2012


On 21 December 2012 at 16:16, Nick Matzke wrote:
| I'm working on a problem that involves calculating a very 
| large, sparse transition probability matrix, P.  It takes 
| significant time to run through the whole matrix, check all 
| the transitions, and put in the appropriate probabilities, 
| based on multiplying parameter values j, d, e, etc.
| Each time I update j, or d, or e, I would rather not go 
| re-process which cells of P get j, j*d, d*e, etc.  It would 
| be better to just do this once, with j, d, and e each just 
| referring to a memory address.  I could then update the 
| values at that memory address and re-use the matrix P 
| without re-building it again and again.
| Currently, my script builds P via an Rcpp function, then 
| passes it to R, where it gets used (with the values 
| calculated) a few hundred times (via another Rcpp function); 
| then j, d, and e are updated and I have to re-calculate P.
| So, my question is, is there a way to make P, currently a 
| vector of float vectors, into a vector of float pointers, 
| which can be passed to R and then back to Rcpp?  Or is this 
| just ridiculous?

There will be lots of different ways to skin this cat. Sparse matrices may be
one, I have no experience there. They now exist in Armadillo, though without
[at current] glue code for as<> and wrap in RcppArmadillo; this does exist in
Eigen and RcppEigen. Others may help you on sparse stuff.

You can also define you own structure to hold those vectors, and then simply
pass one single pointer around using the Rcpp::XPtr wrapper class for R's
external pointer. If data volume is an issue, external pointers are your
friends. (Some of) the database class package uses this, and as does the
bigmemory familiy of packages.  

But you should do some measuring and profiling before you go off and
rearchitect your application based on a hunch.  Great slide deck seen today
and retweeted:  

  The only good intuition: 'I should time time this' 
  -- Andrei Alexandrescu

  Via http://isocpp.org/blog/2012/12/three-optimization-tips-alexandrescu

and I think he has that exactly right.  And he has about a gazillion times as
much street cred on C++ as I do ....


PS You need to work on your signature. Not sure it's long and detailed enough.

Dirk Eddelbuettel | edd at debian.org | http://dirk.eddelbuettel.com  

More information about the Rcpp-devel mailing list