cupy.random.uniform(low=low, high=high) returns high

import cupy as cp

cp.random.seed(seed=13)
high = 2**32 - 1
low = 2**32 - 3
size = 100000000
(cp.random.uniform(low=low, high=high, size=size) == high).sum()

returns array(11), while array(0) is expected, as “Samples are drawn from a uniform distribution over the half-open interval [low, high)”.

Interestingly, the documentation of standard random.uniform says “The end-point value b may or may not be included in the range depending on floating-point rounding in the equation a + (b-a) * random()”. RandomState._scale_kernel seems to do the same thing, so it might be a documentation problem.

BTW, the behavior looks like it is consistent with numpy.

• Conditions

CuPy Version          : 6.0.0
CUDA Root             : /usr/local/cuda
CUDA Build Version    : 10000
CUDA Driver Version   : 10010
CUDA Runtime Version  : 10000
cuDNN Build Version   : 7301
cuDNN Version         : 7605
NCCL Build Version    : 1000
NCCL Runtime Version  : (unknown)
Linux 4.4.180-102-default 
Intel(R) Xeon(R) Gold 5118 CPU @ 2.30GHz
Tesla P100

[akopich]

@toslunar We’re dealing with the expression round(low+round(r*round(high-low))). Let D denote the largest distance between the two consecutive floats in the interval [low, high]. The change of the rounding policy will shift the value of round(high-low) by at most D. As long as |r|<=1, the value r*round(high-low) will be also shifted by at most D. Hence, round(r*round(high-low)) shall not be shifted by more than 2D and round(low+round(r*round(high-low))) – by 3D. Clearly, the mean will not be shifted by more than 3D either. And it doesn’t strike me as a significant change.

The result with x[x == high] = low (maybe x[x >= high] = low) looks similar to the round-toward-negative…

Why, is it low on RHS? Why not smth like std::nextafter(high, 0.0)?

Thank you for referencing the numpy issue. Looks like they found it possible to just change the rounding mode, but lest they break the others’ code, they just edited the docs.

[anaruse]

Ah, I think I finally figured out what you’re saying is the issue. In numpy.random.uniform, "The high limit may be included in the returned array of floats due to floating-point rounding in the equation low + (high-low) * random_sample(). ". You’re saying that the problem is that there is no such an explanation in the cupy.random.uniform document?