numpy.dot has undocumented dtype behavior, resulting in different values than (a * b).sum()

[itamarst]

I would expect (a * b).sum() and np.dot(a, b) to be the same for 1D arrays. However, the former upscales(?) the dtype, and the latter doesn't, so they give different results. This may be expected behavior, but if so the np.dot docs should mention the dtype behavior.

Reproducing code example:

In [36]: a = np.array([254, 2, 200], dtype=np.uint8)

In [37]: b = np.array([True, False, True], dtype=np.bool)

In [38]: (a * b).sum()
Out[38]: 454

In [39]: np.dot(a, b)
Out[39]: 198

In [40]: np.dot(a, b).dtype
Out[40]: dtype('uint8')

Numpy/Python version information:

Numpy 1.15.0, Python 3.6.5

[sturlamolden]

Integer overflow: 198 + 256 = 454.

Should np.dot check for overflow before outputting an integer?

[mattip]

The current design is to choose the output dtype with no consideration overflow. Overflow issues seem to frequently trip up users, perhaps an overall vision of how to deal with them is needed

[itamarst]

As mentioned above, a minimal approach would be to document the expected behavior better.

[eric-wieser]

If anything, I'd consider the behavior of sum to be more surprising here, and in need of more documentation