Probably Dance

Some libraries are so small they’re almost not worth sharing. But the C++ standard has a giant hole in that it doesn’t provide an easy way to quickly generate truly random numbers: It has std::mt19937_64 which quickly generates pseudo-random numbers, and it has std::random_device, which slowly generates truly random numbers.

The easiest way to quickly generate truly random numbers is to use a std::random_device to seed a std::mt19937_64. That way we pay a one-time cost of using random device to generate a seed, and then have quick random numbers after that. Except that the standard doesn’t provide a way to do that. In fact it’s more dangerous than that: It provides an easy wrong way to do it (use a std::random_device to generate a single int and use that single int as the seed) and it provides a slow, slightly wrong way to do it. (use a std::random_device to fill a std::seed_seq and use that as the seed) There’s a proposal to fix this, (that link also contains reasons for why the existing methods are wrong) but I’ve actually been using a tiny class for this:

struct random_seed_seq
{
    template<typename It>
    void generate(It begin, It end)
    {
        for (; begin != end; ++begin)
        {
            *begin = device();
        }
    }

    static random_seed_seq & get_instance()
    {
        static thread_local random_seed_seq result;
        return result;
    }

private:
    std::random_device device;
};

(the license for the code in this blog post is the Unlicense)

This class has the same generate() function that std::seed_seq has and can be used to initialize a std::mt19937_64. The static get_instance() function is a small convenience to make initialization easier so that you can write this:

std::mt19937_64 random_source{random_seed_seq::get_instance()};

Without the get_instance() function this would have to be a two-liner.

Finally a lot of code doesn’t care where their random numbers come from. Sometimes you just want a random float in the range from zero to one and you don’t want to have to set up a random engine and a random distribution. In that case you can write something like this:

float random_float_0_1()
{
    static thread_local std::mt19937_64 randomness(random_seed_seq::get_instance());
    static thread_local std::uniform_real_distribution<float> distribution;
    return distribution(randomness);
}

And just like that we have easy, fast, high quality floating point numbers. Well we do if your compiler is GCC. On my machine this last function is slightly faster than the old-school “rand() * (1.0f / RAND_MAX)”. This function takes 11ns, and the old-school method takes 14 nanoseconds. (measured with Google Benchmark) I attribute most of that to the Mersenne Twister being a very fast random number generator.

When I compiled it with Clang however this new function takes 80ns. Stepping through the assembly generated by both compilers reveals that the problem is that Clang doesn’t inline aggressively enough. There are some calls to compute the logarithm of the upper bound and lower bound in the uniform_real_distribution. GCC inlines those expensive calls away, Clang does not.

Not sure what to do about that last problem: The problem is with how std::uniform_real_distribution is defined: It takes the upper bound and lower bound as runtime arguments. In my code listing above they are the default arguments of 0 and 1, but since Clang doesn’t inline the call, it doesn’t know that they are constants. The only way I see around that is to re-implement std::uniform_real_distribution with constants. But that’s beyond the scope of this blog post.

This blog post was only supposed to be about the random_seed_seq. The other code are just examples showing how you could use it. So let’s not worry about the details of std::uniform_real_distribution, and end this by saying that you should probably use random_seed_seq to seed your random number generators.

It’s a tiny class that I find myself needing all the time. Hopefully it will also be useful for you.