cupy.random.randint is slow

Hi all. It seems, numpy.random.randint is notably faster than cupy.random.randint. Is this expected?

I’ve also attained a dramatic speedup, sampling uniform floats with cupy and rounding them to int.

import cupy as cp
import numpy as np
import time

def runWithTime(f, iters):
    start = time.time()
    for i in range(iters):
        f()
    print(time.time() - start)

size = 100000
high = 1000000
iters = 50000

runWithTime(lambda: cp.random.uniform(low=0, high=high, size=size).astype('int', copy=False), iters)
runWithTime(lambda: np.random.randint(0, high, size=size), iters)
runWithTime(lambda: cp.random.randint(0, high, size=size), iters) 

prints

2.910490036010742
23.180258750915527
36.04059100151062

Is there a reason why we cannot rewrite cupy.random.randint, relying cupy.random.uniform?

• 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

[toslunar]

The current algorithm aligns with numpy. However the algorithm, rejection sampling, is not very efficient with GPU. numpy: https://github.com/numpy/numpy/blob/a72b89c7c2e30f5df5cf27f68b6afd45361934fd/numpy/random/src/distributions/distributions.c#L1072-L1081 cupy: https://github.com/cupy/cupy/blob/9a8cdf80b212f31278d7168851b1dac0c6d75f06/cupy/random/_generator.py#L689-L710

[akopich]

@anaruse Could you please clarify a little further? Is your argument based on the fact, that the uniform distribution is modeled by float64-values of equal probability for intervals of equal exponent field, like [0.5, 1)? randint seems to be implemented only for int32 and it is possible to distribute 2^52 equiprobable objects into 2^32 bins almost uniformly. And in general (not on [0.5,1)), we have even more (float64-valued) objects, which are not equiprobable.

Also, I’ve noticed, cupy.random.uniform might return high.

import cupy as cp

high = int(2**32) - 1
low = int(2**32) - 3
size = 100000000
print(cp.max(cp.random.uniform(low=low, high=high, size=size)) == high)

prints True. Is this expected behavior?

Speaking of “almost”. How do you test the random generators cupy provides?