I am trying to use a boost's random number generator with std::random_shuffle, but it seems that I'm not doing too well, because there are mismatches between expected function signatures.

Code:
#include <iostream>
#include <algorithm>
#include <boost/random.hpp>

int main()
{
    typedef boost::minstd_rand Generator;
    typedef boost::uniform_smallint<std::size_t> Distribution;
    Generator gen;
    Distribution distr(0, 256);
    boost::variate_generator<Generator, Distribution> gen2(gen, distr);

    int array[] = {1, 2, 3, 4, 5, 6, 7, 8, 9};
    const int array_size = sizeof(array)/sizeof(array[0]);
    std::random_shuffle(array, array+array_size, gen);
    std::random_shuffle(array, array+array_size, gen2);
    for (std::size_t i = 0; i != array_size; ++i) {
        std::cout << array[i] << ' ';
    }
}
The first attempt, with Generator, fails because its operator() apparently doesn't take any arguments that random_shuffle requires.

The second attempt, with variate_generator, fails because variate_generator tries to call Distribution::operator()(Generator&, value) whereas the distribution expects operator()(Generator&).

Now, I can work around it by using a proxy object whose operator() does what random_shuffle expects:
Code:
struct RandomWrapper
{
     VariateGenerator rng;
    std::size_t operator() (std::size_t n) {return rng() % n;}
};
However, this still doesn't seem right. For example, by using mod I'm throwing away any advantages that the distribution might provide? And I could use generator alone? Or should I create a new distribution object for each call?

The random distribution doesn't need to be perfect (no Monte Carlo). Random numbers are simply used to shuffle a bunch of small arrays and to pick random indices from other, slightly larger arrays elsewhere (in a multithreading program, so rand() is not suitable) to produce random puzzles. Efficiency is important though.

Does anyone have any advice how I should proceed?