Using unsupported modules in R packages with RcppEigen

[serkor1]

Hi @eddelbuettel,

I am trying to use the unsupported modules in Eigen. This is my something.cpp-file:

//[[Rcpp::depends(RcppEigen)]]
#include <RcppEigen.h>
#include <unsupported/Eigen/CXX11/Tensor>

// [[Rcpp::export]]
Rcpp::XPtr<Eigen::Tensor<float, 2>> createFloatTensor(const Eigen::Map<Eigen::MatrixXd>& mat)
{
   
    Eigen::Index rows = mat.rows();
    Eigen::Index cols = mat.cols();

   
    auto* t = new Eigen::Tensor<float, 2>(rows, cols);

    
    for(Eigen::Index i = 0; i < rows; i++)
    {
        for(Eigen::Index j = 0; j < cols; j++)
        {
            (*t)(i, j) = static_cast<float>(mat(i, j));
        }
    }


    return Rcpp::XPtr<Eigen::Tensor<float, 2>>(t, true);
}

If I do Rcpp::sourceCpp("src/somethting.cpp") I have no issues with compiling and running the functions in R. However, if I build my R package with with R CMD build . + R CMD install I get the following error:

 RcppExports.cpp:15:19: error: ‘Tensor’ is not a member of ‘Eigen’
      15 | Rcpp::XPtr<Eigen::Tensor<float, 2>> createFloatTensor(const Eigen::Map<Eigen::MatrixXd>& mat);
         |                   ^~~~~~

If I compile the R package without // [[Rcpp::export]] there is no issues. I have mingled with the Makevars a bit, too:

PKG_CXX_STD = CXX23
PKG_CPPFLAGS = -I../inst/include

But nothing really works. Am I doing something wrong, or is this a bug?

[eddelbuettel]

I think I have an idea. It's a subtle one, and described (but easy to miss) in the Rcpp Attributes vignette.

When you use sourceCpp() your file as an #include for these headers. And [[Rcpp::depends(RcppEigen)]] and, at the package level, LinkingTo: get the directory added. Now, by putting these types into the interface they are now also in RcppExports.cpp. But look at that (generated) file: it has no #include. So you have to 'force' that and that can be done by adding a special file in inst/include, typically $NAMEOFTHEPACKAGE_types.h containing the includes you need so that they end up in the compilation of RcppExports too.

[serkor1]

Oh! OK, yes - that makes sense, I think. Is that why my RcppExports.cpp looks like this:

// Generated by using Rcpp::compileAttributes() -> do not edit by hand
// Generator token: 10BE3573-1514-4C36-9D1C-5A225CD40393

#include <RcppEigen.h>
#include <Rcpp.h>

And not like this (as I believe it should do):

// Generated by using Rcpp::compileAttributes() -> do not edit by hand
// Generator token: 10BE3573-1514-4C36-9D1C-5A225CD40393

#include <RcppEigen.h>
#include <Rcpp.h>
#include <unsupported/Eigen/CXX11/Tensor>

I will give the vignette a thorough read - thank you so much 👍

[eddelbuettel]

You can't edit RcppExports.cpp, it gets overwritten. So I think what you need is (at least the) one line in a file inst/include/MyPackage_types.h containing

#include <unsupported/Eigen/CXX11/Tensor>

[eddelbuettel]

Actually, you have a second problem. There are no as<>() or wrap() methods for these tensors so you have to do "something" to unpack them into type RcppEigen and friends understand and can convert.

[serkor1]

You can't edit RcppExports.cpp, it gets overwritten. So I think what you need is (at least the) one line in a file inst/include/MyPackage_types.h containing

#include <unsupported/Eigen/CXX11/Tensor>

You know - I spent, maybe, 4-5 hours on this. I knew the solution was a one-line solution, but the key question was "where to put it". And you provided me with the answer - it works like charm now. Thank you so much! 😆

I am not going to use the on the R side, I will only export it as pointer and do stuff with it in the C++ side. So I should be fine without it (I think!)!

Again, thank you so much.

[serkor1]

If it is not too much, I have a follow-up question for my curiosity. Why does putting #include <unsupported/Eigen/CXX11/Tensor> in inst/include/MyPackage_types.h work when building the package, and not if only included in any .cpp or .h file in src/?

[eddelbuettel]

Yes. You can of course use the Tensor types internally to compute but on the interface to R you need to stick to known types. or write converters.

If you can think of a clear place to document the need for the PACKAGENAME_types.h header, let me know. Believe it or not, I also spend time running into the same issues. But hey the third or fourth time one usually remembers ...

[serkor1]

If anything, Id opt for placing it in the Thirteen Simple Steps for Creating An R Package
with an External C++ Library-vignette, and in chapter 5 of your book too! This is where I have been frantically been looking for solutions to my problem!

[eddelbuettel]

@serkor1 (and @jjhvive who seems to follow the discussion) : How is the following?

[serkor1]

@eddelbuettel thank you for including me in the documentation process. I do not know if I am the right person to ask, as my experience with C++ is limited to the last 3-4 months - so I might be misunderstanding parts of your snippet.

In the snippet you provided you write:

(...) in which creating a PACKAGE_types.h (...) will help (...)

This phrasing, for me at least, suggests that including #include statements are optional in PACKAGE_types.h when using external libraries. But based on the issue posted here, it was a requirement for the R package to compile and run correctly.

[eddelbuettel]

How about a slightly stronger version:

One aspect worth noting is that if you include a type in your function interface it will
also be part of the generated \code{RcppExports.cpp}. In this case adding a file
\code{PACKAGE_types.h} (where \code{PACKAGE} is to be replaced with the name of your
package) containing the required \code{#include} statement for the type(s) will permit
compilation; see the 'Rcpp Attributes' vignette for details \citep{CRAN:Rcpp:Attributes}.

[serkor1]

Yes! This is clearer!

[eddelbuettel]

This is now PR #1355 over at the Rcpp repo.