RNG Scope?

[coatless]

Any thoughts on ensuring the correct RNG functionality with the Rmath library?

For instance, if we create a cpp11 function that interfaces with Rf_norm():

cpp11::cpp_source(
    code = '
#include "cpp11/doubles.hpp"
#include "R.h"
#include "Rmath.h"

[[cpp11::register]]
double rng_cpp() {
  double x = Rf_rnorm(0, 1);
  return x;
}

[[cpp11::register]]
double rng_explicit_state_cpp() {
  GetRNGstate();
  double x = Rf_rnorm(0, 1);
  PutRNGstate();
  return x;
}', quiet = FALSE)

The subsequent running of the versions between C++ and R produce the same result without a seed being re-set.

set.seed(3161611)
rng_cpp()
#> [1] -0.9325398
rnorm(1)
#> [1] -0.9325398

Under the second function with an explicit Get/Put state call the desired results are achieved:

set.seed(3161611)
rng_explicit_state_cpp()
#> [1] -0.9325398
rnorm(1)
#> [1] 0.8282708

From looking at exported code it seems like there is no wrapping of C++ calls with GetRNGstate()/PutRNGstate(). Perhaps this should be added to the converting documents or addressed during code generation?

[jimhester]

Generally I think if you want this you should put the requisite calls to GetRNGstate() in your code.

Part of the motivation for cpp11 is to do less than Rcpp, and automatically including these calls in Rcpp has lead to issues like (RcppCore/Rcpp#823, RcppCore/Rcpp#1013)

It is something we should definitely document in the converting vignette however.