Notes.txt

Visual studio compiler flags:
	/W4
	/WX
	/sdl			???
	/O2				in Release mode, same as /Og /Oi /Ot /Oy /Ob2 /GF /Gy )
	/GL				in Release mode, does it make any difference?
	/arch:AVX2		faster than /arch:AVX, mainly due to vectorized exp() in BoostTrainer
	/fp:fast		much faster than /fp:strict and /fp:precise
	/std:c++17
	/permissive-
	/openmp
    /Zc:__cplusplus         pdqsort uses __cplusplus to detect C++11

========================================================================================================================

some profiling numbers from Visual Studio 2019, providing justification for using 
splitmix, pdq_sort, VeryFastBernoulliDistribution and FastBernoulliDistribution

............................................

clock cycles per call

splitmix			2.8
pcg					3.5
xorshift			3.6
std::mt19937		7.5
std::minstd_rand	8.7
std::minstd_rand0	8.9
std::mt19937_64		8.9

..........................................

pdqsort_branchless is 2.2 times faster than std::sort

..........................................

the following code snippets were used to profile different implementations of the Bernoulli distribution
the rng is an instance of splitmix

while (n != 0) {
    bool b = VeryFastBernoulliDistribution(m, n)(rng);
    bool b = BD(m, n)(rng);
    m -= b;
    --n;
}
3.8 clock cycles per iteration

while (n != 0) {
    bool b = FastBernoulliDistribution(m, n)(rng);
    bool b = BD(m, n)(rng);
    m -= b;
    --n;
}
8.9 clock cycles per iteration

while (n != 0) {
    bool b = std::bernoulli_distribution(double(m) / n)(rng);
    m -= b;
    --n;
}
51.7 clock cycles per iteration!