Closed
Description
In [2]: np.arange(20, dtype='uint') << 64
Out[2]: array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], dtype=uint64)
In [3]: np.arange(10, dtype='uint') << 64
Out[3]: array([0, 0, 0, 0, 0, 0, 0, 0, 8, 9], dtype=uint64)
In [4]: np.arange(5, dtype='uint') << 64
Out[4]: array([0, 1, 2, 3, 4], dtype=uint64)My understanding is that while numpy.left_shift should execute a rotation (despite the docs not mentioning this word), the << operator should execute a shift (that is, drop overflowing bits). At least, things seem to work this way with values below 64:
In [5]: (np.arange(20, dtype='uint') << 63 == (np.arange(20, dtype='uint') % 2 * 2) ** 63).all()
Out[5]: True
In [6]: (np.arange(10, dtype='uint') << 63 == (np.arange(10, dtype='uint') % 2 * 2) ** 63).all()
Out[6]: True
In [7]: (np.arange(5, dtype='uint') << 63 == (np.arange(5, dtype='uint') % 2 * 2) ** 63).all()
Out[7]: TrueInstead with 64 some cases (In [2]:) shift, some (In [4]:) rotate, and some (In [3]:) mix the two behaviors!
I'm using uint for simplicity, but the behavior is not limited to unsigned ints:
In [8]: np.arange(20) << 64
Out[8]: array([0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0])
In [9]: np.arange(5) << 64
Out[9]: array([0, 1, 2, 3, 4])Metadata
Metadata
Assignees
Labels
No labels