[Rcpp-devel] Problem passing Armadillo objects back to R

Davor Cubranic cubranic at stat.ubc.ca
Wed Apr 6 17:12:14 CEST 2011


To make it harder for users to shoot themselves in the foot with 'List::create(Named("x") = wrap(X), ...', is it possible to make it simply illegal to assign an explicit "wrap" to a "Named"? This would force everyone to use implicit wraps, which as you said are safe.

Davor

On 2011-04-06, at 4:22 AM, Romain Francois wrote:

> Hmmmmmmmm. I'm afraid I can't fix this.
> 
> Here is the story: both wrap(X) and wrap(X) create a new SEXP and allocate memory. When wrap(Y) is called, since the created object is large, R calls the GC to collect unprotected SEXP to free some space. So my guess is that at that point the SEXP that is generated by wrap(X) is collected since it is not protected.
> 
> I can not deal with this inside ::create, because it is too late.
> 
> 
> A Workaround is to use this :
> 
> return Rcpp::List::create(
> Rcpp::Named("X") = X,
> Rcpp::Named("Y") = Y
> );
> 
> i.e. use the implicit wrap. here wrap will be called much later and things are fine.
> 
> 
> The issue is that when calling wrap it creates an unprotected SEXP, and Rcpp is not responsible for these. If the programmer passes an unprotected SEXP to ::create, it is his responsability to protect it. So another alternative would be:
> 
> SEXP xx = PROTECT( wrap( X ) ) ;
> SEXP yy = PROTECT( wrap( Y ) ) ;
> Rcpp::List result =  Rcpp::List::create(
> Rcpp::Named("X") = xx,
> Rcpp::Named("Y") = yy
> );
> UNPROTECT(2) ;
> return res ;
> 
> yes, this is ugly.
> 
> Romain
> 
> Le 06/04/11 11:42, Romain Francois a écrit :
>> 
>> Good news. I can reproduce it.
>> Bas news. This is likely a PROTECT problem with the Rcpp::List constructor.
>> 
>> Wearing my bugbuster costume, and getting ready for a fight.
>> 
>> Romain
>> 
>> Le 07/03/11 23:03, "Günter J. Hitsch" a écrit :
>>> 
>>> I'm new to Rcpp and RcppArmadillo---so far I like it a lot! Thanks to
>>> the developers for their good work.
>>> 
>>> I run into a peculiar kind of problem when I try to pass a "large"
>>> Armadillo object back to R. Here's some code to replicate the problem in
>>> stylized form:
>>> 
>>> 
>>> extern "C" SEXP testFun(SEXP L_)
>>> {
>>> const long L = Rcpp::as<long>(L_);
>>> arma::mat X(L,1);
>>> arma::mat Y(L,1);
>>> X.fill(1);
>>> Y.fill(2);
>>> return Rcpp::List::create(
>>> Rcpp::Named("X") = Rcpp::wrap(X),
>>> Rcpp::Named("Y") = Rcpp::wrap(Y)
>>> );
>>> }
>>> 
>>> 
>>> I compile (both using g++ and the Intel Compiler --- choice of compiler
>>> makes no difference) and then call from R:
>>> 
>>> 
>>> L = 1000000
>>> ret = .Call("testFun", as.integer(L))
>>> print(ret$X[1:5,])
>>> print(ret$Y[1:5,])
>>> 
>>> 
>>> Here's what I often get as output:
>>> 
>>> [1] 1 1 1 1 1
>>> [1] 1 1 1 1 1
>>> 
>>> However, the second rows should be all 2's!
>>> 
>>> When I try to pass smaller matrices, for example by setting L=100000,
>>> the problem goes away:
>>> 
>>> [1] 1 1 1 1 1
>>> [1] 2 2 2 2 2
>>> 
>>> 
>>> Also, the problem does not arise when I create and pass pack objects of
>>> type Rcpp::NumericMatrix; so far I've see the problem only with
>>> Armadillo objects. I've encountered this on two Macs running OS X
>>> 10.6.6, R 2.12.2, and I'm using the latest versions of Rcpp and
>>> RcppArmadillo.
>>> 
>>> Help on this matter is appreciated!
>>> 
>>> Günter
>>> 
>>> 
>>> 
>>> _______________________________________________
>>> Rcpp-devel mailing list
>>> Rcpp-devel at lists.r-forge.r-project.org
>>> https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/rcpp-devel
>> 
>> 
> 
> 
> -- 
> Romain Francois
> Professional R Enthusiast
> +33(0) 6 28 91 30 30
> http://romainfrancois.blog.free.fr
> http://romain-francois.com
> |- http://bit.ly/fhqbRC : Rcpp workshop in Chicago on April 28th
> |- http://bit.ly/dFyZGB : Hydraulique au Montpellier Comedie Club
> `- http://bit.ly/eVXit9 : Eponyme : 40 minutes stand up
> 
> 
> _______________________________________________
> Rcpp-devel mailing list
> Rcpp-devel at lists.r-forge.r-project.org
> https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/rcpp-devel



More information about the Rcpp-devel mailing list