PERF: optimize Block.getitem_block

[jbrockmendel]

Performance comparison is based on the the asv groupby.Apply.time_scalar_function_single_col, which is the one in which disabling the libreduction path has the biggest impact.

import pandas as pd
import numpy as np

N = 10 ** 4
labels = np.random.randint(0, 2000, size=N)
labels2 = np.random.randint(0, 3, size=N)
df = pd.DataFrame(
    {
        "key": labels,
        "key2": labels2,
        "value1": np.random.randn(N),
        "value2": ["foo", "bar", "baz", "qux"] * (N // 4),
    }
)

In [4]: %prun -s cumtime df.groupby("key").apply(lambda x: 1) 

master-but-with-fast_apply-disabled:

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.000    0.000    0.081    0.081 groupby.py:822(apply)
        1    0.000    0.000    0.081    0.081 groupby.py:871(_python_apply_general)
        1    0.006    0.006    0.080    0.080 ops.py:157(apply)
     1989    0.002    0.000    0.069    0.000 ops.py:933(__iter__)
     1988    0.003    0.000    0.066    0.000 ops.py:966(_chop)
     1988    0.004    0.000    0.059    0.000 managers.py:724(get_slice)
     1988    0.002    0.000    0.035    0.000 managers.py:730(<listcomp>)
     5964    0.009    0.000    0.033    0.000 blocks.py:283(getitem_block)
     5971    0.005    0.000    0.021    0.000 blocks.py:247(make_block_same_class)
     5974    0.007    0.000    0.013    0.000 blocks.py:115(__init__)
     1988    0.003    0.000    0.011    0.000 base.py:4064(__getitem__)
     1990    0.003    0.000    0.008    0.000 managers.py:120(__init__)
     1988    0.002    0.000    0.007    0.000 numeric.py:105(_shallow_copy)
     1992    0.002    0.000    0.007    0.000 blocks.py:2379(__init__)
     1988    0.002    0.000    0.005    0.000 base.py:485(_shallow_copy)
     1990    0.002    0.000    0.004    0.000 frame.py:432(__init__)
     1992    0.002    0.000    0.003    0.000 base.py:450(_simple_new)

PR-but-with-fast_apply-disabled

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.000    0.000    0.061    0.061 groupby.py:822(apply)
        1    0.000    0.000    0.061    0.061 groupby.py:871(_python_apply_general)
        1    0.005    0.005    0.058    0.058 ops.py:157(apply)
     1991    0.002    0.000    0.048    0.000 ops.py:933(__iter__)
     1990    0.003    0.000    0.046    0.000 ops.py:966(_chop)
     1990    0.004    0.000    0.039    0.000 managers.py:724(get_slice)
     1990    0.002    0.000    0.017    0.000 managers.py:730(<listcomp>)
     5970    0.009    0.000    0.015    0.000 blocks.py:297(getitem_block)
     1990    0.002    0.000    0.011    0.000 base.py:4064(__getitem__)
     1992    0.003    0.000    0.007    0.000 managers.py:120(__init__)
     1990    0.002    0.000    0.007    0.000 numeric.py:105(_shallow_copy)
     1990    0.002    0.000    0.005    0.000 base.py:485(_shallow_copy)
     1992    0.002    0.000    0.004    0.000 frame.py:432(__init__)
     5970    0.002    0.000    0.003    0.000 blocks.py:116(_simple_new)
     1994    0.002    0.000    0.003    0.000 base.py:450(_simple_new)
     1992    0.001    0.000    0.002    0.000 managers.py:126(<listcomp>)
    22438    0.002    0.000    0.002    0.000 {built-in method builtins.isinstance}

master

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
        1    0.000    0.000    0.009    0.009 groupby.py:822(apply)
        1    0.000    0.000    0.009    0.009 groupby.py:871(_python_apply_general)
        1    0.000    0.000    0.008    0.008 ops.py:157(apply)
        1    0.000    0.000    0.005    0.005 ops.py:961(fast_apply)
        1    0.003    0.003    0.005    0.005 {pandas._libs.reduction.apply_frame_axis0}
     1994    0.001    0.000    0.002    0.000 base.py:4064(__getitem__)
        1    0.000    0.000    0.001    0.001 ops.py:135(_get_splitter)
        1    0.000    0.000    0.001    0.001 ops.py:268(group_info)
        1    0.000    0.000    0.001    0.001 generic.py:1206(_wrap_applied_output)
        1    0.000    0.000    0.001    0.001 ops.py:285(_get_compressed_codes)

[jreback]

do these actually make a diff?

[jbrockmendel]

i havent measured this independently, so not sure how much it matters in this case (tried it for some Timestamp methods and got a decent boost). it allows cython to do some inlining, at the cost of disallowing subclassing.