Using Armadillo with SuperLU
This article is originally published at https://gallery.rcpp.org
Armadillo is a very versatile C++ library for linear algebra, brought to R via the RcppArmadillo package. It has proven to be very useful and popular, and is (as of February 2017) used by well over 300 CRAN packages as indicated by the reverse depends / linking-to on its CRAN page. Well over a dozen earlier posts on this Rcpp Gallery site also demonstrate the popularity of the package as well as different usage patterns.
Armadillo has a core focus on dense matrices, but continues to expand its capabilities for sparse matrices. Basic operation are supported directly via the templated header files, along with calls back into the default (dense) LAPACK and BLAS libraries we can access easily as R uses them.
Armadillo also supports dedicated sparse solvers via the
SuperLU package. However, this requires access to
the SuperLU library. Many Linux distributions ship
it, see e.g. the Debian package page and the
Ubuntu package page; there is also a
homebrew recipe for OS X /
macOS (or other systems using
brew). As of this writing, the version in the current Ubuntu
release is behind the version Debian. But it is the 5.2.* version that is in Debian that is also
required with current Armadillo versions 7.700.* so we prepared ‘backports’ via
Dirk’s PPA repo.)
In a nutshell, two things need to happen:
One needs to define the required variable
ARMA_USE_SUPERLUwhich has to be done before the Armadillo headers are included. One possibility (shown below) is a
#definestatement right in the code file.
One needs to tell the linker to use the SuperLU library. This step is of course not perfectly portable, and does require that the library be installed. (A genuinely portable solution would either need to test for presence of SuperLU, or include its sources. Both aspects are beyond the scope of this post._
Now that R knows about this library, we can present the code requiring it:
This code snippet defines a function
superLuDemo() which we can call from R:
As the data generated here is random, we did not bother printing the (dense) result vector of length 1000.
Please visit source website for post related comments.