Better to use a context manager [1], with np.errstate(...), so a badly timed exception/ctrl-c doesn't leave the numpy error state differently.

[1] http://docs.scipy.org/doc/numpy/reference/generated/numpy.errstate.html

I'm inclined to think this shouldn't be an option at all, since np.seterr contexts can be set at any level in code execution and should probably be managed outside any specific functions (do we have any other funcks that do this?)

if warnings are common in this function, then I'd be ok with mentioning it in the docstring but I agree it shouldn't be an option.

The only other code I see that does this is in the new-ish all_world2pix implementation:

But I would make the same point there--it should probably be taken out or at least changed to use np.errstate.

Okay, I'll take it out and just put an example in the doc.

(minor): can you change this to something slightly more explicit, like, "The astropy.analytic_functions was added to contain analytic functions useful for astronomy"?

Ok

Best to also test the function with a single wavelength.

Actually, now that I look more closely, this particular test requires integration under the curve, so single wavelength cannot be tested here, but it is tested elsewhere.

in_x not being a Quantity does not guarantee that it does not have a unit already. E.g., with this a column with Angstrom units will become a Quantity with units of A**2. Instead, I would test on the object not having a unit and then letting the Quantity initializer take care, using again the equivalency framework (this makes things slightly slower for, say, a Column with frequency units, but I think that's OK since this is not the common path):

if not isinstance(in_x, u.Quantity):
    with u.add_enabled_equivalencies(u.spectral()):
        in_x = u.Quantity(in_x, u.AA)

I don't think it's necessary to do the full conversion here, because it is done anyway in blackbody_nu() later. But I can change isinstance(in_x, u.Quantity) to hasattr(in_x, 'unit').

Yes, that test should work too (assuming it passes u.spectral_energy(in_x)), but still rewrite the test so that Column with unit unset goes through and do use u.Quantity(in_x, u.AA), as it is substantially faster:

    if getattr(in_x, 'unit', None) is None:
        in_x = u.Quantity(in_x, u.AA)

In [6]: a = np.arange(100.)

In [7]: %timeit a * u.AA
10000 loops, best of 3: 42 µs per loop

In [8]: %timeit u.Quantity(a, u.AA)
100000 loops, best of 3: 7.34 µs per loop

What exactly do you mean by "Column with unit unset"? I am not familiar with Column.

In [15]: c = Column([1.,2.,3.], name='a')

In [16]: c.unit

In [17]: c.unit is None
Out[17]: True

So, if I had a table with a column with its unit unset, the code should assume the unit is Angstrom. But if it is, say, nm, it should pass the thing through.

I did a quick test, passing in Column gives wrong answer because it is not being converted to Hz properly. I think Column support for unit conversion is out of scope for this PR.

But yes, I can implement getattr and u.Quantity(in_x, ...) anyway. Just be clear, that does not mean Column gives the correct answer although nothing crashes.

I'd do the getattr anyway, and indeed not worry about Column otherwise. There has been endless discussion about how its unit should be treated, and I don't want to hold this up for that!