Subset of a Rcpp Matrix that matches a logical statement

Question

In R, if we have a data matrix, say a 100 by 10 matrix X, and a 100-elements vector t with possible values (0, 1, 2, 3), we can easily find a submatrix y of X using a simple syntax:

y = X[t == 1, ]

But, the problem is, how can I do that with Rcpp's NumericMatrix ?
(Or, more generally, how can I do that in C++'s any containers ?)

Thanks to Dirk's hint, it seems that

NumericMatrix X(dataX);
IntegerVector T(dataT);
mat Xmat(X.begin(), X.nrow(), X.ncol(), false);
vec tIdx(T.begin(), T.size(), false); 
mat y = X.rows(find(tIdx == 1));

Can do what I want, but that seems too lengthy.

Answer 1

I would love to see this as sugar. Unfortunately, I am not qualified to implement it though. Here are still a number of different solutions I played with.

First, I had to make some modifications to Gong-Yi Liao code to get this to work (colvec instead of vec for tIdx and Xmat.rows(... instead of X.rows(...:

mat Xmat(X.begin(), X.nrow(), X.ncol(), false);
colvec tIdx(T.begin(), T.size(), false); 
mat y = Xmat.rows(find(tIdx == 1));

Second, here are three function with benchmarks that all subset matrices based on a logical statement. The functions take arma or rcpp arguments and return values Two are based on Gong-Yi Liao's solution and one is a simple loop-based solution.

n(rows)=100, p(T==1)=0.3

                expr   min     lq median     uq    max
1  submat_arma(X, T) 5.009 5.3955 5.8250 6.2250 28.320
2 submat_arma2(X, T) 4.859 5.2995 5.6895 6.1685 45.122
3  submat_rcpp(X, T) 5.831 6.3690 6.7465 7.3825 20.876
4        X[T == 1, ] 3.411 3.9380 4.1475 4.5345 27.981

n(rows)=10000, p(T==1)=0.3

                expr     min       lq   median       uq      max
1  submat_arma(X, T) 107.070 113.4000 125.5455 141.3700 1468.539
2 submat_arma2(X, T)  76.179  80.4295  88.2890 100.7525 1153.810
3  submat_rcpp(X, T) 244.242 247.3120 276.6385 309.2710 1934.126
4        X[T == 1, ] 229.884 236.1445 263.5240 289.2370 1876.980

submat.cpp

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

using namespace Rcpp;
using namespace arma;

// arma in; arma out
// [[Rcpp::export]]
mat submat_arma(arma::mat X, arma::colvec T) {
    mat y = X.rows(find(T == 1));
    return y;
}

// rcpp in; arma out
// [[Rcpp::export]]
mat submat_arma2(NumericMatrix X, NumericVector T) {
    mat Xmat(X.begin(), X.nrow(), X.ncol(), false);
    colvec tIdx(T.begin(), T.size(), false); 
    mat y = Xmat.rows(find(tIdx == 1));
    return y;
}

// rcpp in; rcpp out
// [[Rcpp::export]]
NumericMatrix submat_rcpp(NumericMatrix X, LogicalVector condition) { 
    int n=X.nrow(), k=X.ncol();
    NumericMatrix out(sum(condition),k);
    for (int i = 0, j = 0; i < n; i++) {
        if(condition[i]) {
            out(j,_) = X(i,_);
            j = j+1;
        }
    }
    return(out);
}


/*** R
library("microbenchmark")

# simulate data
n=100
p=0.3
T=rbinom(n,1,p)
X=as.matrix(cbind(rnorm(n),rnorm(n)))

# compare output
identical(X[T==1,],submat_arma(X,T))
identical(X[T==1,],submat_arma2(X,T))
identical(X[T==1,],submat_rcpp(X,T))

# benchmark
microbenchmark(X[T==1,],submat_arma(X,T),submat_arma2(X,T),submat_rcpp(X,T),times=500)

# increase n
n=10000
p=0.3
T=rbinom(n,1,p)
X=as.matrix(cbind(rnorm(n),rnorm(n)))
# benchmark
microbenchmark(X[T==1,],submat_arma(X,T),submat_arma2(X,T),submat_rcpp(X,T),times=500)

*/

Answer 2

The closest I know of is the combination of the find() function combined with the submat() function in Armadillo accessible via RcppArmadillo.

Edit: This is of course something we could add via a patch. If anybody is sufficiently motivated to try this, please come to the rcpp-devel mailing list.