Enable Datetime/Timedelta dtypes in Masked UDFs

[brandon-b-miller]

Closes #9432.

Enables UDFs that do this:

import cudf

df = cudf.DataFrame({'a':['2011-01-01'], 'b':[1]})
df['a'] = df['a'].astype('datetime64[ns]')
df['b'] = df['b'].astype('timedelta64[ns]')

def f(row):
    return row['a'] + row['b']

res = df.apply(f, axis=1)

0   2011-01-01 00:00:00.000000001
dtype: datetime64[ns]

[brandon-b-miller]

cc @gmarkall

[codecov]

Codecov Report

Merging #9451 (d12a474) into branch-21.12 (ab4bfaa) will increase coverage by 0.03%.
The diff coverage is n/a.

❗ Current head d12a474 differs from pull request most recent head 773e711. Consider uploading reports for the commit 773e711 to get more accurate results

@@               Coverage Diff                @@
##           branch-21.12    #9451      +/-   ##
================================================
+ Coverage         10.79%   10.82%   +0.03%     
================================================
  Files               116      117       +1     
  Lines             18869    19451     +582     
================================================
+ Hits               2036     2106      +70     
- Misses            16833    17345     +512     

Impacted Files	Coverage Δ
python/cudf/cudf/__init__.py	0.00% <ø> (ø)
python/cudf/cudf/_lib/__init__.py	0.00% <ø> (ø)
python/cudf/cudf/io/csv.py	0.00% <0.00%> (ø)
python/cudf/cudf/io/hdf.py	0.00% <0.00%> (ø)
python/cudf/cudf/io/orc.py	0.00% <0.00%> (ø)
python/cudf/cudf/_version.py	0.00% <0.00%> (ø)
python/cudf/cudf/core/abc.py	0.00% <0.00%> (ø)
python/cudf/cudf/api/types.py	0.00% <0.00%> (ø)
python/cudf/cudf/io/dlpack.py	0.00% <0.00%> (ø)
... and 64 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 7fa2738...773e711. Read the comment docs.

[isVoid]

Just one question: numpy rejects operators between a datetime64 with a datetime64 type. Will numba do the same type check here?

[galipremsagar]

Just one question: numpy rejects operators between a datetime64 with a datetime64 type. Will numba do the same type check here?

Does numpy reject subtract operation on datetime64 vs datetime64 too? I'm aware that this is an operation supported in pandas and cudf.

[brandon-b-miller]

Just one question: numpy rejects operators between a datetime64 with a datetime64 type. Will numba do the same type check here?

Does numpy reject subtract operation on datetime64 vs datetime64 too? I'm aware that this is an operation supported in pandas and cudf.

This brings up an interesting point in general. The space of allowed operations is a complex function allowed(lhs, rhs, op) -> bool. Technically, we should test all of them, right? There's some functions that might be able to be adapted for this generally here but the result might be thousands of tests. Since all these tests require different jit compilation, it would be a huge inflation of the pytest runtime. I have been debating what to do about this without a great answer yet.

[isVoid]

Does numpy reject subtract operation on datetime64 vs datetime64 too?

Numpy does support subtraction, only rejects addition

There's some functions that might be able to be adapted for this generally here but the result might be thousands of tests.

I'm in general favor of testing the delegations when we can't have a full coverage (or full coverage cost is high), and add tests when corner cases pops up.

[brandon-b-miller]

Just one question: numpy rejects operators between a datetime64 with a datetime64 type. Will numba do the same type check here?

Does numpy reject subtract operation on datetime64 vs datetime64 too? I'm aware that this is an operation supported in pandas and cudf.

This seems to be handled with this diff:

import cudf
x = cudf.DataFrame({
    'a': ['2011-01-01'],
    'b': ['2011-02-02']
})
x['a'] = x['a'].astype('datetime64[ns]')
x['b'] = x['b'].astype('datetime64[ns]')

def f(row):
    return row['a'] - row['b']

res = x.apply(f, axis=1)

0    -32 days 00:00:00
dtype: timedelta64[ns]

added a few tests for this.

[gmarkall]

This looks good to me at the moment. I had one question on the diff, which probably stems from something I don't understand about datetime types.

[brandon-b-miller]

@gpucibot merge