如何在不重新发明轮子的情况下为自定义类型生成均匀分布的随机实数？

Question

I was going to use std::uniform_real_distribution with some non-builtin floating-point-like types, eg half_float::half or boost::multiprecision::float128 . But what I get is

/opt/gcc-5.2/include/c++/5.2.0/bits/random.h:1868:7: error: static assertion failed: template argument not a floating point type

from g++ 5.2. Here's the example program (compile with g++ -std=c++11 -fext-numeric-literals test.cpp -o test ):

#include <random>
#include <boost/multiprecision/float128.hpp>
#include <half.hpp>

template<typename Float>
void test()
{
    std::random_device rd;
    std::mt19937 mt(rd());
    std::uniform_real_distribution<Float> rnd(Float(1),Float(10));
}

int main()
{
    test<float>();
    test<double>();
    test<long double>();
    test<boost::multiprecision::float128>(); // doesn't compile
    test<half_float::half>(); // doesn't compile
}

So, how is one supposed to generate uniformly distributed random reals for such custom types? Is there a way to go without reinventing the wheel?

Answer 1

As you can see from here , uniform_real_distribution is defined as:

template< class RealType = double >
class uniform_real_distribution;

Where RealType is:

The result type generated by the generator. The effect is undefined if this is not one of float, double, or long double.

It seems that your compiler explicitly forbids to use a custom type as a solution to deal with (let me say) unaccepted types . Therefore I would say that you have no chance to get it working.

Answer 2

The standard random number facilities which work with RealType type as defined in the Standard, must have float , double or long double as type template argument, as in 26.5.1.1.d:

Throughout this subclause 26.5, the effect of instantiating a template:

<...>

d) that has a template type parameter named RealType is undefined unless the corresponding template argument is cv-unqualified and is one of float , double , or long double .

One solution is to replace std::uniform_real_distribution with boost::random::uniform_real_distribution . The latter does accept non-builtin float types. Here's an example with a reasonable type ^* :

#include <random>
#include <boost/random/uniform_real_distribution.hpp>
#include <boost/multiprecision/float128.hpp>

template<typename Float>
void test()
{
    std::random_device rd;
    std::mt19937 mt(rd());
    boost::random::uniform_real_distribution<Float> rnd(Float(1),Float(10));
}

int main()
{
    test<float>();
    test<double>();
    test<long double>();
    test<boost::multiprecision::float128>();
}

_{^* half_float::half doesn't work because its multiplication by a constant results in a float , while half_float::half(float) constructor is explicit, and Boost's implementation doesn't call it explicitly}