ENH: add mm->q floordiv

[tylerjereddy]

Motivated by #12120 (comment)

Add support for floor division between timedelta64 operands with type signature: mm->q

I tried mm->q for closer mimic with built-in timedelta, but that makes it tricky to handle NaT-> NaN.

Maybe there's a cleaner / more efficient way to deal with negative values.

[eric-wieser]

Hmm, seems weird to be to have floordiv return float - I'm not sure if that's better or worse than losing NaTs.

[tylerjereddy]

@shoyer Should I switch to returning an int and do something else with NaT?

[tylerjereddy]

Based on in-person discussion with @shoyer & @eric-wieser I should likely revise this to return integer values.

For the case of NaT in the numerator, denominator, or both, use error flag to emit warning & return 0.

[shoyer]

Note that this behavior will be consistent with other cases where ufuncs have undefined behavior, e.g., dividing by 0.

…

On Sat, Dec 1, 2018 at 10:52 AM Tyler Reddy ***@***.***> wrote: Based on in-person discussion with @shoyer <https://github.com/shoyer> & @eric-wieser <https://github.com/eric-wieser> I should likely revise this to return integer values. For the case of NaT in the numerator, denominator, or both, use error flag to emit warning & return 0. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#12308 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ABKS1rmE-7xKAN9eO-7XIP5s-eMhlP_Mks5u0s_sgaJpZM4YLE9d> .

[tylerjereddy]

Refactored to return integer type -- tests pass locally anyway. I wonder what reviewers will think about the way I intercepted floor division for type resolution between m8 types -- is the strcmp too fragile? It is readable at least.

[eric-wieser]

Can you construct a test that would fail if we lost precision through a float?

[tylerjereddy]

Can you construct a test that would fail if we lost precision through a float?

Yes, I've added precision unit tests for both code paths based on integers that a double can't represent exactly.

Since NPY_MIN_LONGLONG is cast to NaT, its special case handling from the longlong inner loop was removed as well.

[mattip]

Thanks Tyler