replacing rand()

replacing rand()

[Julius Schmidt]

An infuriating discussion with martian67 this morning has reminded me
to run statistical tests on our rand().  As expected, it doesn't perform
too well (fails 9/15 SmallCrush, BigCrush still running, fails after 256
MB of PractRand).  I'm wondering whether we should replace with rand()
with something better as RNGs exist nowadays that are both simpler,
faster and produce better random numbers.

I would propose xoroshiro128+:

u64int state0, state1;

u64int
llrand(void)
{
 	u64int s0, s1, r;

 	s0 = state0;
 	s1 = state1;
 	r = s0 + s1;
 	s1 ^= s0;
 	state0 = (s0 << 55 | s0 >> 9) ^ s1 ^ (s1 << 14);
 	state1 = (s1 << 36 | s0 >> 28);
 	return r;
}

http://xoroshiro.di.unimi.it/xoroshiro128plus.c

Re: [9front] replacing rand()

[Julius Schmidt]

Corrections:

I forgot that lrand() returns bit 31 as zero.  Taking this into account
makes lrand() look pretty good actually.  It passes SmallCrush.  The lower
16 bit fail PractRand (but so does the lowest bit of xoroshiro128+).
The highest 16 bit (30:15) pass PractRand for at least 2 GB (still
running).

Also my code was slightly wrong.
state1 should be s1 >> 28, not s0 >> 28.

On Sat, 10 Sep 2016, Julius Schmidt wrote:

> An infuriating discussion with martian67 this morning has reminded me
> to run statistical tests on our rand().  As expected, it doesn't perform
> too well (fails 9/15 SmallCrush, BigCrush still running, fails after 256
> MB of PractRand).  I'm wondering whether we should replace with rand()
> with something better as RNGs exist nowadays that are both simpler,
> faster and produce better random numbers.
>
> I would propose xoroshiro128+:
>
> u64int state0, state1;
>
> u64int
> llrand(void)
> {
> 	 u64int s0, s1, r;
>
> 	 s0 = state0;
> 	 s1 = state1;
> 	 r = s0 + s1;
> 	 s1 ^= s0;
> 	 state0 = (s0 << 55 | s0 >> 9) ^ s1 ^ (s1 << 14);
> 	 state1 = (s1 << 36 | s0 >> 28);
> 	 return r;
> }
>
> http://xoroshiro.di.unimi.it/xoroshiro128plus.c
>
>