When does using RNGScope make a difference? - rcpp

Question

In Rcpp documentation, I often find the recommendation to place Rcpp::RNGScope scope; before using random draws within Rcpp. I wondered what exactly this does, because I've only ever seen it described as "ensures RNG state gets set/reset".
Then, I tested a bit, but I can't seem to come up with an example where doing this makes any difference. I used an example from here. My tests were:
#include <Rcpp.h>
using namespace Rcpp;
// [[Rcpp::export]]
NumericVector noscope() {
Rcpp::Function rt("rt");
return rt(5, 3);
}
// [[Rcpp::export]]
NumericVector withscope() {
RNGScope scope;
Rcpp::Function rt("rt");
return rt(5, 3);
}
and then
set.seed(45)
noscope() # [1] 0.6438 -0.6082 -1.9710 -0.1402 -0.6482
set.seed(45)
withscope() # [1] 0.6438 -0.6082 -1.9710 -0.1402 -0.6482
set.seed(45)
rt(5, 3) # [1] 0.6438 -0.6082 -1.9710 -0.1402 -0.6482
So, my question is twofold. First, when does RNGScope make a difference, and what exactly does it do different from not using it? Second, does anyone have a code example which shows different results with and without it?
If RNGScope was deprecated in a newer release, then I'm sorry for asking.

Answer 1

When using Rcpp attributes, the automagically generated interface to your code will automatically insert the appropriate construction of the RNGScope object -- so it's already being done for you behind the scenes in this case. For example, if you write sourceCpp(..., verbose = TRUE), you'll see output like this:
Generated extern "C" functions
--------------------------------------------------------
#include <Rcpp.h>
RcppExport SEXP sourceCpp_38808_timesTwo(SEXP xSEXP) {
BEGIN_RCPP
Rcpp::RObject __result;
Rcpp::RNGScope __rngScope;
Rcpp::traits::input_parameter< NumericVector >::type x(xSEXP);
__result = Rcpp::wrap(timesTwo(x));
return __result;
END_RCPP
}
Note the automatic construction of the RNGScope object.
You only need to construct that object manually if you are operating outside of the realm of Rcpp attributes.

Answer 2

This all becomes a little clearer once you read the original documentation in the Writing R Extensions manual, Section 6.3, "Random Numbers".
All that RNGScope scope does is the automagic calls to "get" and "put" in order to keep the state of the RNG sane.
The problem with your test code, as explained by Kevin, is that it already happens for you. So you can only test by going through .Call() by hand, in which case you will for sure leave the RNG in a mess if you use it and do not get/put properly.