05 - Sparse matrices
In armadillo4r: An 'Armadillo' Interface

Supported sparse matrix classes

Since v0.9.0, armadillo4r links to the Matrix package and uses its public C API to convert between R sparse matrices and Armadillo SpMat<double>. The as_SpMat(SEXP) function accepts any sparse matrix class that the Matrix package can express:

CsparseMatrix subclasses (dgCMatrix, dsCMatrix, dtCMatrix, lgCMatrix, ...) -- the fast path, read directly via the CHOLMOD bridge.
RsparseMatrix and TsparseMatrix subclasses -- transparently coerced to CsparseMatrix first.
Symmetric (dsCMatrix) matrices are expanded to general (both triangles populated) before constructing the SpMat.
Unit-triangular (dtCMatrix with diag = "U") matrices have their implicit unit diagonal materialized.

as_dgCMatrix(SpMat<double>) always returns a general dgCMatrix.

The strategy is to use Matrix's M_sexp_as_cholmod_sparse to wrap the input without copying the slot data, then build the Armadillo SpMat from the CSC arrays. The output path uses M_cholmod_sparse_as_sexp to construct the returned dgCMatrix.

Note that cpp4r does not provide sparse matrices as it does for the dense data types doubles_matrix<> or integers_matrix<>. armadillo4r uses SEXP to provide a method to convert between R sparse classes and SpMat objects.

Here is an example of how to convert a dgCMatrix object to a SpMat object and export the resulting operation to R:

[[cpp4r::register]] SEXP sum_matrices_(SEXP x) {
  // Convert from any sparse Matrix object to SpMat
  SpMat<double> A = as_SpMat(x);

  // Create a matrix B with a diagonal of random numbers
  SpMat<double> B(A.n_rows, A.n_cols);
  for (uword i = 0; i < A.n_rows; ++i) {
    B(i, i) = randu<double>();
  }

  A += B; // Add the two matrices

  // Convert back to dgCMatrix and return
  return as_dgCMatrix(A);
}