Zero to consciously incompetent with {Rcpp}.

Rcpp

Exploring the world of {Rcpp} through the eyes of a complete novice.

Published

April 4, 2021

I was skimming through Hadley’s advanced R over Easter (as one does) and ended up spending quite a bit of time reading through the last chapter (anyone else read books back to front?), “Rewriting R code in c++”. My understanding of c++ was (is?) pretty much non-existent, but whenever I try to learn something new I find the process of writing about it helps things stick. So this post is exactly that, me fumbling my way through some very basic c++ using {Rcpp}.

This post is pretty much a rewrite of the early sections in chapter 25 of Advanced R, so don’t expect anything too exciting!

What & why

For a general introduction / overview to c++, I found https://www.learncpp.com/ to be a great place, and assumes no previous knowledge at all. Also this presentation by Dirk Eddelbuettel gives a great introduction from an R perspective. Rcpp for everyone by Masaki E. Tsuda is another fantastic resource and does a great job at showing all the different possibilities with Rcpp.

So, why would one use R with c++?

  1. When things gotta go fast (and R is slow)
  2. Solve a problem in a new way
  3. Make R do something new / extend R

I was particularly interested in point number 1, as who doesn’t enjoy trying to make things go fast?

How

The Rcpp package makes the process of R and c++ working together “easier, yet also more robust”1so we’ll be using this throughout the following examples. When writing c++ for a package, you’d use the .cpp file extension, however we can also include c++ chunks in rmarkdown (pretty cool) so here are a few examples…

You can use the Rcpp chunk setting, or set the chunk option engine = “Rcpp” to get things working in Rmarkdown

Say we want a simple function that doubles whatever number we put in, we could writing something like

// [[Rcpp::export]]
int doubleUP(int x) {
  int y = x * 2;
  return y;
}

Now there are a few things going on here, most notably we need to define the data type. In this case we use int for integer. Other options include doubles, strings, and logicals. We also need to use a semi-colon at the end of each statement. The // [[Rcpp::export]] bit makes the c++ function available to use in R, so lets do a quick test…

doubleUP(10)
[1] 20

Great! Now a single value is fine, but say we have a vector of values that we want to double. Would this work?

doubleUP(c(10, 20, 30))
Error in doubleUP(c(10, 20, 30)): Expecting a single value: [extent=3].

Oh dear. Looks like we need to make some adjustments to our function so it will accept more than one number.

Fortunately, we can use one of the Vector classes from Rcpp, NumericVector, IntegerVector, CharacterVector, and LogicalVector to make our function accept a vector. Here is the adjusted function…

#include <Rcpp.h>
using namespace Rcpp;

// [[Rcpp::export]]
NumericVector doubleUPPP (NumericVector x) {
  int n = x.size();
  NumericVector result(n);
  
  for(int i = 0; i < n; ++i) {
    result[i] = x[i] * 2;
  }
  return result;
}

Few extra things have happened and I’ll attempt to explain… First we create result which is a numeric vector the size of the vector that we pass in to the function. Then we use a for loop with the syntax for(init; check; increment). We initiate the loop with a new variable i = 0 and keep looping while i is less than n (the size of our vector). Finally the ++i tells the loop to increment the value of i each time the loop completes.

doubleUPPP has three P’s to indicate it accepts a vector. Maybe this naming convention will catch on?

We’ve also added the following lines to the chunk:

#include <Rcpp.h>
using namespace Rcpp;

which I think is a bit like using library(package) in R so we’re making the functions(?) from Rcpp available in our c++ code. Alternatively, we could have written Rcpp::NumericVector each time.

So, does it work?

doubleUPPP(c(10, 20, 30))
[1] 20 40 60

:). We could write a similar function in R and compare performance.

doubleR <- function(x) {
  
  total <- vector(mode = mode(x), length = length(x))
  
  for (i in seq_along(x)) {
    total[i] <- x[i] * 2
  }
  total
}

And if we compare:

big_vector <- runif(1e6)

bench::mark(
  doubleR(big_vector),
  doubleUPPP(big_vector),

)[1:6]
# A tibble: 2 × 6
  expression                  min   median `itr/sec` mem_alloc `gc/sec`
  <bch:expr>             <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
1 doubleR(big_vector)     156.3ms 156.85ms      6.38    7.67MB     2.13
2 doubleUPPP(big_vector)   8.34ms   8.55ms    116.      7.64MB    58.2

Quite the improvement! Now lets just ignore for a moment that we could simply use big_vector + big_vector to get the same result (and is faster than both doubleR and doubleUPPP…).

Rcpp also provides some familiar R classes such as List and DataFrame. Here is a useless function that accepts a DataFrame, multiplies two columns together and returns a DataFrame with the two columns and the result.

#include <Rcpp.h>
using namespace Rcpp;

// [[Rcpp::export]]
DataFrame carFacts(DataFrame df, std::string var1, std::string var2) {

  NumericVector var_1 = as<NumericVector>(df[var1]);
  NumericVector var_2 = as<NumericVector>(df[var2]);
  
  DataFrame data = DataFrame::create( Named(var1) = var_1,         
                                      Named(var2) = var_2,
                                      Named("Product") = var_1 * var_2); 
  return data;
}

Here it is in action with our old friend mtcars:

carFacts(mtcars, var1 = "hp", var2 = "mpg")[1:10,]
    hp  mpg Product
1  110 21.0  2310.0
2  110 21.0  2310.0
3   93 22.8  2120.4
4  110 21.4  2354.0
5  175 18.7  3272.5
6  105 18.1  1900.5
7  245 14.3  3503.5
8   62 24.4  1512.8
9   95 22.8  2166.0
10 123 19.2  2361.6

Nice! (Although my function is pretty useless in practice…)

Another cool feature is the ability to use R functions in c++ code. Here is another completely useless example using the emo::ji function with a string input.

#include <Rcpp.h>
using namespace Rcpp;

// [[Rcpp::export]]
RObject cppEMOJI(Function f, std::string x) {
  return f(x);
}

Here we’ve used the RObject class as a catchall for the output type (as we may not know it in advance). We’ve also used std::string as our input type. std refers to the standard library, from which we are using the string identifier. We could also use using namespace std; like we did for Rcpp.

If we run our function in R we get:

library(emo)
cppEMOJI(ji, "smile")
😄

Incredible.

Wrapping up

Before embarrassing myself any further with these useless functions, I think I’ll leave it there. The resources for learning about Rcpp are absolutely fantastic and if you’re interested I’d highly recommend starting with Hadley’s Advanced R chapter and the great Rcpp vignettes. There are also a whole load of showcase examples that you can use for inspiration.

Thanks for reading!

Acknowledgments

Preview image from https://github.com/RcppCore/Rcpp/issues/827, made by https://github.com/coatless

Footnotes

  1. https://cran.r-project.org/web/packages/Rcpp/vignettes/Rcpp-introduction.pdf↩︎