Second success: Uniform integer distribution from 1,...,n:
typedef std::mt19937_64::result_type uint_t;
// Returns element of 1,...,n with uniform probability:
class Uniform {
std::mt19937_64& g;
const uint_t n;
const uint_t scaling;
const uint_t past;
public :
Uniform(std::mt19937_64& g, const uint_t n) : g(g), n(n), scaling(std::numeric_limits<uint_t>::max() / n), past(n * scaling) {}
gen_uint_t operator ()() {
gen_uint_t result;
do result = g(); while (result >= past);
return result/scaling + 1;
}
};
Now only missing seeding for std::mt19937_64.
This seems most complicated.
Apparently it is in
https://gcc.gnu.org/viewcvs/gcc/trunk/libstdc%2B%2B-v3/include/bits/random.tcc?revision=246542&view=markup
Lines 346 - 389.
I would guess that this functionality is most specialised, and most likely to be different
for different compilers.
Really a pitty that this was not standardised.
Oliver
On Tue, Apr 11, 2017 at 02:02:42PM +0100, Oliver Kullmann wrote:
> thanks for the info.
>
> My first, little, basically trivially success, is replacing the
> Bernoulli-distribution for the default case:
>
> inline bool bernoulli(std::mt19937_64& g) noexcept {
> typedef std::mt19937_64::result_type uint_t;
> constexpr uint_t half = std::numeric_limits<uint_t>::max() / 2;
> return g() < half;
> }
>
> Checked that it reproduces the behaviour of gcc 4.7.4 and 5.4.
>
> Replacing std::uniform_int_distribution (only for std::mt19937_64) seems harder,
> and yet I have problems understanding the code.
>
> One could just generalise the above, and since only 32-bits intervals
> are used, that likely would suffice for a decent distribution.
>
> However g++ seems to do something more advanced.
>
> On Sun, Apr 09, 2017 at 05:31:54PM -0600, Martin Sebor wrote:
> > It's worth keeping in mind that copying libstdc++ code will have
> > licensing implications. A different caveat is that conforming C++
> > programs can define explicit specializations of standard library
> > templates only for user-defined types. I.e., it's not valid to
> > define one's own std::uniform_int_ditribution class template or
> > specialization of it on a fundamental type such as int.
> >
>
> I just want to extract the mathematical content, and then write simple
> code for it, like above.
>
> If somebody knows some underlying papers describing the algorithms for
> uniform_int_distribution and seeding std::mt19937_64, that would be great.
>
> > Another option is to browse the sources online:
> >
> > https://gcc.gnu.org/viewcvs/gcc/trunk/libstdc%2B%2B-v3/include/bits/random.h?view=markup
> >
>
> Thanks, I found that most useful.
>
> Oliver
>
