A new version of TRNG (Tina’s Random Number Generator Library) has been released. TRNG may be utilized in sequential as well as in parallel Monte Carlo simulations. It does not depend on a specific parallelization technique, e.g., POSIX threads, MPI or others. As an outstanding new feature of the latest TRNG release 4.11 it also supports CUDA. See TRNG documentation for details. If you need a modern C++ random number generator library for sequential or parallel Monte Carlo then check out TRNG 4.11.

The following piece of C++ code exemplifies how to use TRNG in a CUDA program. See TRNG documentation for details.

#include <cstdlib>
#include <iostream>
#include <trng/yarn5s.hpp>
#include <trng/uniform01_dist.hpp>

__global__
void parallel_pi(long samples, long *in, trng::yarn5s r) {
  long rank=threadIdx.x;
  long size=blockDim.x;
  r.jump(2*(rank*samples/size));      // jump ahead
  trng::uniform01_dist<float> u;      // random number distribution
  in[rank]=0;                         // local number of points in circle
  for (long i=rank*samples/size; i<(rank+1)*samples/size; ++i) {
    float x=u(r), y=u(r);             // choose random x- and y-coordinates
    if (x*x+y*y<=1)                   // is point in circle?
      ++in[rank];                     // increase thread-local counter
  }
}

int main(int argc, char *argv[]) {
  const long samples=1000000l;             // total number of points in square
  const int size=128;                      // number of threads
  long *in_device;
  cudaMalloc(&in_device, size*sizeof(*in_device));
  trng::yarn5s r;
  // start parallel Monte Carlo
  parallel_pi<<<1, size>>>(samples, in_device, r);
  // gather results
  long *in=new long[size];
  cudaMemcpy(in, in_device, size*sizeof(*in), cudaMemcpyDeviceToHost);
  long sum=0;
  for (int rank=0; rank<size; ++rank) 
    sum+=in[rank];
  // print result
  std::cout << "pi = " << 4.0*sum/samples << std::endl;
  return EXIT_SUCCESS;
}