by hardcoding the rounding instead of relying
on NumPy functions, so that this op should work
on CPU and on GPU.

Tests (CPU and GPU) passed on my computer, let's
see if jenkins/travis are ok with that.