Introducing Rcpp::algorithm

Daniel C. Dillon — written Jun 25, 2016 — source

Introduction

A while back, I saw a post on StackOverflow where the user was trying to use Rcpp::sugar::sum() on an RcppParallel::RVector. Obviously, this does not work (as Rcpp Sugar pertains to Rcpp types, but not RcppParallel which cannot rely on SEXP-based representation to allow multi-threaded execution). It raised the question “Why doesn’t something more generic exist to provide functions with R semantics that can be used on arbitrary data structures?” As a result, I set out to create a set of such functions in Rcpp::algorithm which follow the pattern of std::algorithm.

Rcpp::algorithm

Currently Rcpp::algorithm contains only a few simple functions. If these are found to be useful, more will be added. Examples of using the currently implemented iterator-based functions are below.

sum, sum_nona, prod, and prod_nona

#include <Rcpp.h>

using namespace Rcpp;

// [[Rcpp::export]]
double sum_of_matrix_row(NumericMatrix m, int row) {
    NumericMatrix::Row r = m.row(row);

    return algorithm::sum(r.begin(), r.end());
}

min, max, and mean

#include <Rcpp.h>

using namespace Rcpp;

// [[Rcpp::export]]
double mean_of_matrix_row(NumericMatrix m, int row) {
    NumericMatrix::Row r = m.row(row);

    return algorithm::mean(r.begin(), r.end());
}

log, exp, and sqrt

#include <Rcpp.h>

using namespace Rcpp;

// [[Rcpp::export]]
NumericVector log_of_matrix_row(NumericMatrix m, int row) {
    NumericMatrix::Row r = m.row(row);

    NumericVector retval(m.cols());
    algorithm::log(r.begin(), r.end(), retval.begin());

    return retval;
}

Additional Benefits

Through the coding of these simple “algorithms”, a few needs arose.

First, the ability to deduce the appropriate C numeric type given an Rcpp iterator was necessary. This gave birth to the Rcpp::algorithm::helpers::decays_to_ctype and Rcpp::algorithm::helpers::ctype type traits. Given a type, these allow you to determine whether it can be cast to a C numeric type and which type that would be.

Second, the need arose for more information about R types. This gave birth to the Rcpp::algorithm::helpers::rtype traits. These are defined as follows:

template< typename T >
struct rtype_helper {};

template<>
struct rtype_helper< double > {
    typedef double type;
    static RCPP_CONSTEXPR int RTYPE = REALSXP;
    static inline double NA() { return NA_REAL; }
    static inline RCPP_CONSTEXPR double ZERO() { return 0.0; }
    static inline RCPP_CONSTEXPR double ONE() { return 1.0; }
};

template<>
struct rtype_helper< int > {
    typedef int type;
    static RCPP_CONSTEXPR int RTYPE = INTSXP;
    static inline int NA() { return NA_INTEGER; }
    static inline RCPP_CONSTEXPR int ZERO() { return 0; }
    static inline RCPP_CONSTEXPR int ONE() { return 1; }
};

template< typename T >
struct rtype {
    typedef typename rtype_helper< typename ctype< T >::type >::type type;
    typedef rtype_helper< typename ctype< T >::type > helper_type;
    static RCPP_CONSTEXPR int RTYPE = helper_type::RTYPE;
    static inline T NA() { return helper_type::NA(); }
    static inline RCPP_CONSTEXPR T ZERO() { return helper_type::ZERO(); }
    static inline RCPP_CONSTEXPR T ONE() { return helper_type::ONE(); }
};

These additional benefits may actually prove more useful than the algorithms themselves. Only time will tell.

Wrapping Up

There are now some simple iterator-based algorithms that can be used with any structure that supports iterators. They apply the same semantics as the analogous Rcpp::sugar functions, but give us more flexibility in their usage. If you find these to be useful, feel free to request more.

tags: sugar

Extending R with C++ and Fortran — Dirk Eddelbuettel and JBrandon Duck-Mayr
Benchmarking Rcpp code with RcppClock — Zach DeBruine
Simulation Smoother using RcppArmadillo — Tomasz Woźniak