I have a project in R which uses C++ functions to speed up computations. In my src-folder I have a C++ file which calls functions I have defined in different header files and C++ files in the same folder. To provide a minimum working example I have created a new project with some very small files of which I describe the contents below.
src/test_func.cpp:
#include "RcppArmadillo.h"
#include "extra1.h"
// [[Rcpp::export]]
arma::vec OLS(const arma::mat & X, const arma::vec & y)
{
arma::vec beta = compute_b(X, y);
return beta;
}
// [[Rcpp::export]]
arma::vec sparse_solver()
{
arma::sp_mat X = arma::sprandu<arma::sp_mat>(1000, 1000, 0.1);
arma::vec y(1000, arma::fill::randu);
arma::vec beta = compute_sp_b(X, y);
return beta;
}
src/extra1.h:
#include "RcppArmadillo.h"
arma::vec compute_b(const arma::mat & X, const arma::vec & y);
arma::vec compute_sp_b(const arma::sp_mat & X, const arma::vec & y);
src/extra1.cpp:
#define ARMA_USE_SUPERLU 1
#include "RcppArmadillo.h"
#include "extra1.h"
arma::vec compute_b(const arma::mat & X, const arma::vec & y)
{
arma::vec b = arma::solve(X.t() * X, X.t() * y);
return b;
}
arma::vec compute_sp_b(const arma::sp_mat & X, const arma::vec & y)
{
arma::vec b = arma::spsolve(X, y, "superlu");
return b;
}
Using this code via
Sys.setenv("PKG_LIBS" = "-lsuperlu")
Rcpp::sourceCpp("src/test_func.cpp")
X = matrix(rnorm(20), nrow = 10)
b = matrix(rnorm(2))
y = X %*% b + matrix(rnorm(10, 0, 0.1))
OLS(X, y)
sparse_solver()
works exactly as expected, I am able to use my functions in R without any hiccups. However, my project has grown and has far more than a single header file and corresponding cpp file and using this to recompile everything if I have made small changes to a single file is starting to take quite a bit of time. Additionally, I read in some posts online that Rcpp::sourceCpp() is not meant for compiling multiple source files and that creating a package is recommended.
To achieve this I followed instructions online and created the file structure using RcppArmadillo::RcppArmadillo.package.skeleton("Test2", example_code = FALSE), copied my source code to the src folder and ran Rcpp::compileAttributes("Test2"). Then I pressed the option Install and Restart in the Build tab in Rstudio and got the following error:
Error: package or namespace load failed for ‘Test2’ in dyn.load(file, DLLpath = DLLpath, ...):
... some_path/Test2.so: undefined symbol: set_default_options
Error: loading failed
Execution halted
ERROR: loading failed
Which has probably to do with the use of SuperLU by my code. If I comment out the #define ARMA_USE_SUPERLU 1 and replace "superlu" by "lapack" in arma::spsolve() the package builds without issues (but this is not really a feasible solution, as it turns the sparse matrix into a dense one). I have no experience building R packages and I can't find an example package online which uses SuperLU to solve sparse systems of equations to figure out how to get this package to build. At the moment I don't really care about portability, this package is still purely for personal use, but lately some additional reasons to turn it into an installable package have come up (I would like to also be able to call these functions from Python using the rpy2 library, for which it appears to be necessary to be an installed package).
In short, my question is if someone knows how to build packages using SuperLU or knows of an R package that uses SuperLU so I can take a look at how they were able to get it to compile.
Edit: my operating system is Ubuntu 20.04
