inconsistent behavior of keepdims for CUB-based reductions

It appears that keepdims does not keep all of the reduction dimensions when CUB is enabled and axis is a tuple containing more than one axis.

• Conditions (you can just paste the output of python -c 'import cupy; cupy.show_config()')

CuPy Version          :  7.0.0rc1 (build from master branch)
CUDA Root             : /usr/local/cuda
CUDA Build Version    : 10010
CUDA Driver Version   : 10010
CUDA Runtime Version  : 10010

• Code to reproduce

import cupy
x = cupy.ones((64, 64))

cupy.cuda.cub_enabled=False
out = x.sum(axis=(0, 1), keepdims=True)
print("CUB DISABLED shape={}".format(out.shape))

cupy.cuda.cub_enabled=True
out = x.sum(axis=(0, 1), keepdims=True)
print("CUB ENABLED shape={}".format(out.shape))

gives

CUB DISABLED shape=(1, 1)
CUB ENABLED shape=(1,)

cc @leofang

[grlee77]

Apparently I didn’t think it was a problem

I think the origin of the problem was an oversight on my part in #2517. I enabled reduction over all axes there, but did not ensure that keepdims preserved all of the axes.

[grlee77]

I am working on a fix for this currently.