Dynamic dispatch for sparse matrices

Fabian Scheipl — written Mar 20, 2014 — source

We want to do matrix multiplication for 3 cases:

  • dense times dense
  • sparse times dense for sparse matrices of class dgCMatrix
  • sparse times dense for sparse matrices of class indMatrix,

using R’s Matrix package for sparse matrices in R and RcppArmadillo for C++ linear algebra:

// [[Rcpp::depends(RcppArmadillo)]]
#include <RcppArmadillo.h>
using namespace Rcpp ;

arma::mat matmult_sp(const arma::sp_mat X, const arma::mat Y){
    arma::mat ret = X * Y;
    return ret;
};
arma::mat matmult_dense(const arma::mat X, const arma::mat Y){
    arma::mat ret = X * Y;
    return ret;
};
arma::mat matmult_ind(const SEXP Xr, const arma::mat Y){
    // pre-multiplication with index matrix is a permutation of Y's rows: 
    arma::uvec perm =  as<S4>(Xr).slot("perm");
    arma::mat ret = Y.rows(perm - 1);
    return ret;
};

//[[Rcpp::export]]
arma::mat matmult_cpp(SEXP Xr, const arma::mat Y) {
    if (Rf_isS4(Xr)) {
        if(Rf_inherits(Xr, "dgCMatrix")) {
            return matmult_sp(as<arma::sp_mat>(Xr), Y) ;
        } ;
        if(Rf_inherits(Xr, "indMatrix")) {
            return matmult_ind(Xr, Y) ; 
        } ;
        stop("unknown class of Xr") ;
    } else {
        return matmult_dense(as<arma::mat>(Xr), Y) ;
    } 
}

Set up test cases:

library(Matrix)
Loading required package: methods
library(rbenchmark)
set.seed(12211212)
n <- 1000
d <- 50
p <- 30  

X <- matrix(rnorm(n*d), n, d)
X_sp <- as(diag(n)[,1:d], "dgCMatrix")
X_ind <- as(sample(1:d, n, rep=TRUE), "indMatrix")
Y <- matrix(1:(d*p), d, p)

Check exception handling:

tryCatch(matmult_cpp(as(X_ind, "ngTMatrix"), Y),
         error = print)
<Rcpp::exception: unknown class of Xr>

Dense times dense:

all.equal(X%*%Y, matmult_cpp(X, Y))
[1] TRUE
benchmark(X%*%Y, 
          matmult_cpp(X, Y),
          replications=100)[,1:4]
               test replications elapsed relative
2 matmult_cpp(X, Y)          100   0.086     1.00
1           X %*% Y          100   0.098     1.14

dgCMatrix times dense:

all.equal(as(X_sp%*%Y, "matrix"), matmult_cpp(X_sp, Y),
          check.attributes = FALSE)
[1] TRUE
benchmark(X_sp%*%Y, 
          matmult_cpp(X_sp, Y),
          replications=100)[,1:4]
                  test replications elapsed relative
2 matmult_cpp(X_sp, Y)          100   0.006    1.000
1           X_sp %*% Y          100   0.013    2.167

indMatrix times dense:

all.equal(X_ind%*%Y, matmult_cpp(X_ind, Y))
[1] TRUE
benchmark(X_ind%*%Y, 
          matmult_cpp(X_ind, Y),
          replications=100)[,1:4]
                   test replications elapsed relative
2 matmult_cpp(X_ind, Y)          100   0.008      1.0
1           X_ind %*% Y          100   0.012      1.5

Based on this Q&A on StackOverflow, thanks again to Kevin Ushey for his helpful comment.

tags: sparse 

Related Articles