I found http://eel.is/c++draft/rand.eng.mers which, if it has been adopted by the Standard, seems in Points 6, 8 actually to fully specify the seeding, from initial values as well as from sequents (std::seed_seq), and thus, hopefully, actually seeding in both forms seems well-defined by the standard. (Otherwise it would be a strange oversight.) Oliver P.S. And it seems that the GCC implementation does what is described in that document. > Consider the following program MT.cpp: > > #include <iostream> > #include <random> > #include <vector> > #include <cstdint> > > int main() { > typedef std::mt19937_64::result_type ui; > std::seed_seq s {ui(10),ui(10),ui(0)}; > std::vector<std::uint32_t> seeds(3); > s.generate(seeds.begin(), seeds.end()); > for (const auto x : seeds) std::cout << x << " "; > std::cout << "\n"; > std::mt19937_64 g(s); > std::cout << g() << " " << g() << "\n"; > } > > Compiled with > > g++ --std=c++11 -Wall MT.cpp > it yields with gcc 4.7.4 and gcc 5.4 > > > ./a.out > 927367489 2598207009 681269367 > 14538134740155241067 15440504459514664889 > > As far as I understand it, only the first line is defined by the standard, > but not the second: how the seed-sequence is used by the generator g is up > to the compiler. > > Thus the need to do the seeding ourselves. > > Oliver >
