Random number generation with C++11 and VSL

I have to generate a large amount of Gaussian random numbers and I was curious how good stock RNGs are, as opposed to rolling our own Box-Muller transform over rand(). Options were the new RNG in C++11 with GCC, and Intel's VSL in MKL with ICC.

The C++11 version is simple. It uses a base RNG, probably a linear congruential engine, to generate uniform random numbers. Then it applies a transformation, which happens to be the Box-Muller transform. The code is:

#include <iostream>
#include <ctime>
#include <random>

using namespace std;

int main(int argc, char **argv) {

    const int nrolls=500000000;

    default_random_engine generator;
    normal_distribution<double> distribution(0.0,1.0);

    time_t start_time = time(NULL);

    double number = 0;
    for (int i=0; i<nrolls; ++i) {
      number = distribution(generator);
    }

    cout << difftime(time(NULL),start_time) << " " << number << endl;

    return 0;
}

The VSL version uses the same split between base RNG and transformation. Apart from the Box-Muller transform, it is also able to use the inverse cumulative distribution function (ICDF) method to generate a Gaussian random number. A crucial difference is that it can vectorize the generation, which yields a massive speedup in theory. The code is here, switching around the numbers in nrolls and N flips between vectorized and non-vectorized variants:

#include <iostream>
#include <ctime>
#include "mkl_vsl.h"
#define SEED 0
#define BRNG VSL_BRNG_MCG31
#define METHOD VSL_METHOD_DGAUSSIAN_ICDF

using namespace std;

main() {
    const int nrolls= 1;//500000000;
    const int N = 500000000;//500000000;
    double* r=new double[N];
    double a = 0, sigma = 1.0;
    VSLStreamStatePtr stream;
    int errcode;
    errcode = vslNewStream( &stream, BRNG, SEED );
    time_t start_time = time(NULL);
    for (int i=0; i<nrolls; ++i) {
      errcode = vdRngGaussian( METHOD, stream, N, r, a, sigma );
    }
    cout << difftime(time(NULL),start_time) << endl;
    errcode = vslDeleteStream( &stream );
    delete r;
}

The table below summarizes the results with linear congruential engines as the base RNG. I did not see much difference between the execution time by changing the base RNG. The hardware was AMD, so it is possible that the Intel compiler and VSL give poorer results. The results are in seconds.

Without vectorization, VSL performs poorly. The kind of transformation also makes a big difference. With vectorized generation and the ICDF method, we are looking at a speedup around 8x.

I would also be curious to see how the results compare to cuRAND. Nvidia claim a 75x speedup over MKL with Keppler hardware, which sounds exciting.