Fix numpy dtypes bug in cuda.to_gpu and cuda.copy

[tkerola]

I found a bug in cuda.to_gpu and cuda.copy that would not change the device when the device identifier is a numpy dtype. Users might use numpy to generate the device ids, and would therefore be subject to this subtle bug.

MWE

Illustrating the problem for to_gpu, but the same thing happens for copy.

from chainer import cuda
import numpy as np

gpus = [0, 1, 2, 3]
batch_size = 2
eval_gpus = np.repeat(gpus, batch_size)

xs = []
for gpu in eval_gpus:
    y = np.ones(1)
    x = cuda.to_gpu(y, device=gpu)
    xs.append(x)

for x in xs:
    print(x.device)

Output

Without this PR:

<CUDA Device 0>
<CUDA Device 0>
<CUDA Device 0>
<CUDA Device 0>
<CUDA Device 0>
<CUDA Device 0>
<CUDA Device 0>
<CUDA Device 0>

With this PR (expected):

<CUDA Device 0>
<CUDA Device 0>
<CUDA Device 1>
<CUDA Device 1>
<CUDA Device 2>
<CUDA Device 2>
<CUDA Device 3>
<CUDA Device 3>

[niboshi]

How about factoring out common part?:

_integer_types = six.integer_types + (numpy.integer,)
if six.PY2:
     _integer_types +=  (_newint,)

[tkerola]

That's a good idea, thanks.

[niboshi]

In this way bool is evaluated as True.
How about excluding bool explicitly?

[tkerola]

Good catch, done.

[niboshi]

Ah, I meant bool, not numpy.bool_.
(isinstance(numpy.bool_, numpy.integer) and isinstance(numpy.bool_, six.integer_types)) are both False, but isinstance(bool, six.integer_types) is True).

So the correct condition would be:
if type(arg) is not bool and isinstance(arg, _integer_types):

[tkerola]

Isn't my condition also correct? It works for both numpy.bool, numpy.bool_ and bool.

[niboshi]

I see. So the outcome would be the same, but it does duplicate check and slower (numpy.bool_ never satisfies the first condition). Also note that numpy.bool and bool are identical.

[tkerola]

Aha, bool and np.bool are identical. Then your solution is cleaner. Thanks! I will update it.

[niboshi]

Thank you for the fix.
jenkins, test this please

[niboshi]

LGTM!