Add Numba to rolling.apply

[mroeschke]

• closes Use Numba to improve the performance difference between rolling.apply #25
• tests added / passed
• passes black pandas
• passes git diff upstream/master -u -- "*.py" | flake8 --diff

[mroeschke]

Here's the performance comparison so far. Unfortunately cython is still faster for the ndarray case.

# This branch
In [1]: s = pd.Series(range(10000))

In [2]: f = lambda x: np.sum(x) + 5

# raw is unused; row is always an ndarray
In [4]: %timeit s.rolling(10).apply(f, raw=False)
219 ms ± 29.4 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

In [5]: %timeit s.rolling(10).apply(f, raw=True)
226 ms ± 26.6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

# Master
In [1]: s = pd.Series(range(10000))

In [2]: f = lambda x: np.sum(x) + 5

# row passed as a Series
In [4]: %timeit s.rolling(10).apply(f, raw=False)
1.19 s ± 5.81 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

# row passed as a ndarray
In [5]: %timeit s.rolling(10).apply(f, raw=True)
32.3 ms ± 366 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

[mroeschke]

Additionally, apply currently supports passing *args and **kwargs into the function per row. When njitting the passed function from the user, passing kwargs is currently unsupported. numba/numba#2916. And binding these kwargs with functools.partial beforehand is unsupported, numba/numba#4587

[mroeschke]

I was able to solve the performance problem in aa9644c. The biggest issue was compiling the njit rolling_apply function every time. If we dynamically create the rolling apply function with the passed argument, cache the function, and call it again, performance beats cython.

In [1]: s = pd.Series(range(10000))

# r, a Rolling object, will cache the apply functions
In [2]: r = s.rolling(10)

In [3]: f = lambda x: np.sum(x) + 5

In [4]: %timeit r.apply(f, raw=False)
2.16 ms ± 204 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)

[mroeschke]

Here are the ASV benchmarks:

       before           after         ratio
     [68a663be]       [3c49d034]
     <master>         <feature/rolling_apply_numba>
+         289±2ms        350±0.9ms     1.21  rolling.Apply.time_rolling('Series', 1000, 'float', <function sum at 0x105754620>, True)
+         290±7ms          350±1ms     1.20  rolling.Apply.time_rolling('Series', 1000, 'int', <function sum at 0x105754620>, True)
+         297±8ms          353±2ms     1.19  rolling.Apply.time_rolling('DataFrame', 1000, 'int', <function sum at 0x105754620>, True)
-         250±3ms        215±0.6ms     0.86  rolling.Apply.time_rolling('Series', 10, 'float', <function sum at 0x105754620>, True)
-         255±1ms        215±0.8ms     0.84  rolling.Apply.time_rolling('Series', 10, 'int', <function sum at 0x105754620>, True)
-         308±5ms        246±0.9ms     0.80  rolling.Apply.time_rolling('Series', 10, 'float', <function Apply.<lambda> at 0x11b8c6730>, True)
-        310±20ms          241±4ms     0.78  rolling.Apply.time_rolling('DataFrame', 10, 'float', <function Apply.<lambda> at 0x11b8c6730>, True)
-        340±40ms        245±0.6ms     0.72  rolling.Apply.time_rolling('Series', 10, 'int', <function Apply.<lambda> at 0x11b8c6730>, True)
-      13.9±0.06s          380±1ms     0.03  rolling.Apply.time_rolling('Series', 1000, 'int', <function Apply.<lambda> at 0x11b8c6730>, False)
-       14.0±0.1s        380±0.8ms     0.03  rolling.Apply.time_rolling('Series', 1000, 'float', <function Apply.<lambda> at 0x11b8c6730>, False)
-         14.1±0s          380±1ms     0.03  rolling.Apply.time_rolling('DataFrame', 1000, 'float', <function Apply.<lambda> at 0x11b8c6730>, False)
-      13.7±0.01s        349±0.8ms     0.03  rolling.Apply.time_rolling('Series', 1000, 'int', <function sum at 0x105754620>, False)
-      13.9±0.06s        350±0.6ms     0.03  rolling.Apply.time_rolling('Series', 1000, 'float', <function sum at 0x105754620>, False)
-      13.7±0.07s        245±0.8ms     0.02  rolling.Apply.time_rolling('Series', 10, 'float', <function Apply.<lambda> at 0x11b8c6730>, False)
-       14.0±0.1s        246±0.5ms     0.02  rolling.Apply.time_rolling('Series', 10, 'int', <function Apply.<lambda> at 0x11b8c6730>, False)
-      13.5±0.01s        215±0.3ms     0.02  rolling.Apply.time_rolling('Series', 10, 'float', <function sum at 0x105754620>, False)

The raw=False benchmarks are partially misleading since it means cython is handling pandas objects while numba is always handling numpy arrays (numba cannot operate in nopython mode with pandas objects).

[mroeschke]

Additional notes:

• I had to xfail a couple of tests. To operate rolling.apply in nopython mode for max performance, apply cannot accept arbitrary functions.
• the passed function to apply cannot accept *args and **kwargs because it's unsupported in numba.