xoshiro256++ implementation in task.c

[stefan-zobel]

julia/src/task.c

Line 670 in 272eaf9

uint64_t t = s0 << 17;

Shouldn't this be s1 as in

uint64_t t = s1 << 17; ?

Note: s1, not s0

See Fig. 5 ("C code for a xoshiro256+/xoshiro256++/xoshiro256** generator") on page 6 in the Blackman / Vigna paper : https://arxiv.org/pdf/1805.01407.pdf

[stevengj]

cc @chethega and @rfourquet, who wrote this code.

[GregPlowman]

Also, implementation in task.c seems inconsistent with stdlib version:

julia/src/task.c

Line 672 in 272eaf9

uint64_t res = ((tmp << 23) | (tmp >> 41)) + s0;

julia/stdlib/Random/src/Xoshiro.jl

Line 57 in 3a4198e

res = tmp << 23 | tmp >> 41

Shouldn't this be:

res = (tmp << 23 | tmp >> 41) + s0

See ref: https://prng.di.unimi.it/xoshiro256plusplus.c

[PallHaraldsson]

It seems obvious that both the implementations are wrong (while fixes trivial), so this needs to be added to 1.7 milestone.

[bjarthur]

possibly related?? https://discourse.julialang.org/t/same-random-sequence-on-gpu-and-cpu/67881/6?u=bjarthur

[vtjnash]

This is not undefined behavior, so unrelated. There is perhaps a fast-fill array method that is vectorized for longer arrays? I am not sure if it was supposed to show a difference for simple sequences like that.

[JeffBezanson]

fixed by #42201