Provide access to the global RNG

[rstub]

For some applications it makes sense to use a RNG directly. It is of course always possible to generate a new RNG, but that has its own independent state. It would be more convenient, if one could get access to (a shared pointer holding) the global RNG.

[hsloot]

I revisited the issue and noticed that my previous suggestion was not helpful. I came up with an alternative, simple approach (see https://github.com/hsloot/dqrng) using an Rcpp::XPtr to expose the underlying raw pointer to the RNG. This seems to work (also in parallel clusters). Benchmarks show that this approach comes with little to no overhead (see https://github.com/hsloot/dqrng/blob/master/performance/global_rng.md).

If you like the approach, I could provide a PR.

Notes on the implementation

The following method can be used to expose the raw pointer to the global RNG in an Rcpp::XPtr.

// src/dqrng.cpp
...
// after rng is initialized
// [[Rcpp::export(rng = false)]]
SEXP get_sxprng() {
  return Rcpp::XPtr<dqrng::rng64_t::element_type>(rng.get(), false);
}

This creates bindings via the Rcpp interface in dqrng_RcppExports.h. Using this method directly would be a little difficult, so I created a wrapper method intended to be used by others in a separate header dqrng_get_rng.h.

// inst/include/dqrng_get_rng.h
#ifndef DQRNG_GET_RNG_H
#define DQRNG_GET_RNG_H 1

#include <Rcpp.h>
#include "dqrng.h"
#include "dqrng_generator.h"

namespace dqrng {

dqrng::rng64_t::element_type* get_rng() {
	return Rcpp::XPtr<dqrng::rng64_t::element_type>(dqrng::get_sxprng()).get();
}

}

#endif // DQRNG_GET_RNG_H

Alternatives

I am not sure if it is also possible to directly expose the shared_ptr or the raw pointer without going through the Rcpp::XPtr; however, I could not make this to work.

[rstub]

Hi @hsloot.
Sorry for the late reply. I had missed your comments, but came to a similar solution when I started playing around with it, c.f. https://github.com/daqana/dqrng/tree/access-rng. Would you still be interested in providing a PR?

[hsloot]

I tried merging both approaches and cleaning it up a bit; see #58.