On Wednesday, 22 January 2014 at 11:02:10 UTC, monarch_dodra
wrote:
On Wednesday, 22 January 2014 at 09:33:39 UTC, terchestor wrote:
My problem is the cast to integral type (line 80), downgrading
performances of circa 50%.
Is there any possibility to avoid it?
I'd suggest you get rid of "high/low" completly. There are prng
*adaptors* that can create a uniform distribution *from* a
natural prng (std.random.uniform). If you do this, you code
becomes (starting line 73):
auto uni = (s1 ^ s2 ^ s3);
static if (isFloatingPoint!result_t)
{
return uni / Tconst_tau0;
}
else
{
return uni;
}
Heck, you could get rid of floating point support entirely, and
generate only integrals. Then, you just let "uniform" do the
tough work:
auto rnd = tausworthe!(ulong, uint)(randomSeed);
auto rndF = rnd.uniform!"[)"(0.0, 1.0); //0 - 1 double range
auto rnd1024 = rnd.uniform!"[)"(0, 1024); //[0 .. 1024)
integral range.
Actually, for the purpose I've in mind, a DSP library, which is
time critical (and that's why I need the fastest possible PRNG),
I'd better decoupling the range processing and create a class
implementing the range interface to store the numbers generated
by uniform to serve them on demand: a composite pattern is better
than inheritance.
Nitpick:
/// get seed from current time
auto se = Clock.currTime().stdTime;
This is known to be a bad seed. If you need a "good enough"
seed, use "std.random.randomSeed". Although arguably (IMO), it
is better place the burden of seeding on the caller.
Fine remark worth adopting.
