asv_bench/benchmarks/arithmetic.py - main - githubmirror/pandas

asv_bench/benchmarks/arithmetic.py
479 строк12 KB
jbrockmendel
PERF: datetimelike addition (#56373)
09 дек 2023, 21:36
09 дек 2023, 21:36a6c0ae4
import operator
import warnings

import numpy as np

import pandas as pd
from pandas import (
    DataFrame,
    Index,
    Series,
    Timestamp,
    date_range,
    to_timedelta,
)

from .pandas_vb_common import numeric_dtypes

try:
    import pandas.core.computation.expressions as expr
except ImportError:
    import pandas.computation.expressions as expr
try:
    import pandas.tseries.holiday
except ImportError:
    pass


class IntFrameWithScalar:
    params = [
        [np.float64, np.int64],
        [2, 3.0, np.int32(4), np.float64(5)],
        [
            operator.add,
            operator.sub,
            operator.mul,
            operator.truediv,
            operator.floordiv,
            operator.pow,
            operator.mod,
            operator.eq,
            operator.ne,
            operator.gt,
            operator.ge,
            operator.lt,
            operator.le,
        ],
    ]
    param_names = ["dtype", "scalar", "op"]

    def setup(self, dtype, scalar, op):
        arr = np.random.randn(20000, 100)
        self.df = DataFrame(arr.astype(dtype))

    def time_frame_op_with_scalar(self, dtype, scalar, op):
        op(self.df, scalar)


class OpWithFillValue:
    def setup(self):
        # GH#31300
        arr = np.arange(10**6)
        df = DataFrame({"A": arr})
        ser = df["A"]

        self.df = df
        self.ser = ser

    def time_frame_op_with_fill_value_no_nas(self):
        self.df.add(self.df, fill_value=4)

    def time_series_op_with_fill_value_no_nas(self):
        self.ser.add(self.ser, fill_value=4)


class MixedFrameWithSeriesAxis:
    params = [
        [
            "eq",
            "ne",
            "lt",
            "le",
            "ge",
            "gt",
            "add",
            "sub",
            "truediv",
            "floordiv",
            "mul",
            "pow",
        ]
    ]
    param_names = ["opname"]

    def setup(self, opname):
        arr = np.arange(10**6).reshape(1000, -1)
        df = DataFrame(arr)
        df["C"] = 1.0
        self.df = df
        self.ser = df[0]
        self.row = df.iloc[0]

    def time_frame_op_with_series_axis0(self, opname):
        getattr(self.df, opname)(self.ser, axis=0)

    def time_frame_op_with_series_axis1(self, opname):
        getattr(operator, opname)(self.df, self.ser)

    # exclude comparisons from the params for time_frame_op_with_series_axis1
    #  since they do not do alignment so raise
    time_frame_op_with_series_axis1.params = [params[0][6:]]


class FrameWithFrameWide:
    # Many-columns, mixed dtypes

    params = [
        [
            # GH#32779 has discussion of which operators are included here
            operator.add,
            operator.floordiv,
            operator.gt,
        ],
        [
            # (n_rows, n_columns)
            (1_000_000, 10),
            (100_000, 100),
            (10_000, 1000),
            (1000, 10_000),
        ],
    ]
    param_names = ["op", "shape"]

    def setup(self, op, shape):
        # we choose dtypes so as to make the blocks
        #  a) not perfectly match between right and left
        #  b) appreciably bigger than single columns
        n_rows, n_cols = shape

        if op is operator.floordiv:
            # floordiv is much slower than the other operations -> use less data
            n_rows = n_rows // 10

        # construct dataframe with 2 blocks
        arr1 = np.random.randn(n_rows, n_cols // 2).astype("f8")
        arr2 = np.random.randn(n_rows, n_cols // 2).astype("f4")
        df = pd.concat([DataFrame(arr1), DataFrame(arr2)], axis=1, ignore_index=True)
        # should already be the case, but just to be sure
        df._consolidate_inplace()

        # TODO: GH#33198 the setting here shouldn't need two steps
        arr1 = np.random.randn(n_rows, max(n_cols // 4, 3)).astype("f8")
        arr2 = np.random.randn(n_rows, n_cols // 2).astype("i8")
        arr3 = np.random.randn(n_rows, n_cols // 4).astype("f8")
        df2 = pd.concat(
            [DataFrame(arr1), DataFrame(arr2), DataFrame(arr3)],
            axis=1,
            ignore_index=True,
        )
        # should already be the case, but just to be sure
        df2._consolidate_inplace()

        self.left = df
        self.right = df2

    def time_op_different_blocks(self, op, shape):
        # blocks (and dtypes) are not aligned
        op(self.left, self.right)

    def time_op_same_blocks(self, op, shape):
        # blocks (and dtypes) are aligned
        op(self.left, self.left)


class Ops:
    params = [[True, False], ["default", 1]]
    param_names = ["use_numexpr", "threads"]

    def setup(self, use_numexpr, threads):
        self.df = DataFrame(np.random.randn(20000, 100))
        self.df2 = DataFrame(np.random.randn(20000, 100))

        if threads != "default":
            expr.set_numexpr_threads(threads)
        if not use_numexpr:
            expr.set_use_numexpr(False)

    def time_frame_add(self, use_numexpr, threads):
        self.df + self.df2

    def time_frame_mult(self, use_numexpr, threads):
        self.df * self.df2

    def time_frame_multi_and(self, use_numexpr, threads):
        self.df[(self.df > 0) & (self.df2 > 0)]

    def time_frame_comparison(self, use_numexpr, threads):
        self.df > self.df2

    def teardown(self, use_numexpr, threads):
        expr.set_use_numexpr(True)
        expr.set_numexpr_threads()


class Ops2:
    def setup(self):
        N = 10**3
        self.df = DataFrame(np.random.randn(N, N))
        self.df2 = DataFrame(np.random.randn(N, N))

        self.df_int = DataFrame(
            np.random.randint(
                np.iinfo(np.int16).min, np.iinfo(np.int16).max, size=(N, N)
            )
        )
        self.df2_int = DataFrame(
            np.random.randint(
                np.iinfo(np.int16).min, np.iinfo(np.int16).max, size=(N, N)
            )
        )

        self.s = Series(np.random.randn(N))

    # Division

    def time_frame_float_div(self):
        self.df // self.df2

    def time_frame_float_div_by_zero(self):
        self.df / 0

    def time_frame_float_floor_by_zero(self):
        self.df // 0

    def time_frame_int_div_by_zero(self):
        self.df_int / 0

    # Modulo

    def time_frame_int_mod(self):
        self.df_int % self.df2_int

    def time_frame_float_mod(self):
        self.df % self.df2

    # Dot product

    def time_frame_dot(self):
        self.df.dot(self.df2)

    def time_series_dot(self):
        self.s.dot(self.s)

    def time_frame_series_dot(self):
        self.df.dot(self.s)


class Timeseries:
    params = [None, "US/Eastern"]
    param_names = ["tz"]

    def setup(self, tz):
        N = 10**6
        halfway = (N // 2) - 1
        self.s = Series(date_range("20010101", periods=N, freq="min", tz=tz))
        self.ts = self.s[halfway]

        self.s2 = Series(date_range("20010101", periods=N, freq="s", tz=tz))
        self.ts_different_reso = Timestamp("2001-01-02", tz=tz)

    def time_series_timestamp_compare(self, tz):
        self.s <= self.ts

    def time_series_timestamp_different_reso_compare(self, tz):
        self.s <= self.ts_different_reso

    def time_timestamp_series_compare(self, tz):
        self.ts >= self.s

    def time_timestamp_ops_diff(self, tz):
        self.s2.diff()

    def time_timestamp_ops_diff_with_shift(self, tz):
        self.s - self.s.shift()


class IrregularOps:
    def setup(self):
        N = 10**5
        idx = date_range(start="1/1/2000", periods=N, freq="s")
        s = Series(np.random.randn(N), index=idx)
        self.left = s.sample(frac=1)
        self.right = s.sample(frac=1)

    def time_add(self):
        self.left + self.right


class TimedeltaOps:
    def setup(self):
        self.td = to_timedelta(np.arange(1000000))
        self.ts = Timestamp("2000")

    def time_add_td_ts(self):
        self.td + self.ts


class CategoricalComparisons:
    params = ["__lt__", "__le__", "__eq__", "__ne__", "__ge__", "__gt__"]
    param_names = ["op"]

    def setup(self, op):
        N = 10**5
        self.cat = pd.Categorical(list("aabbcd") * N, ordered=True)

    def time_categorical_op(self, op):
        getattr(self.cat, op)("b")


class IndexArithmetic:
    params = ["float", "int"]
    param_names = ["dtype"]

    def setup(self, dtype):
        N = 10**6
        if dtype == "float":
            self.index = Index(np.arange(N), dtype=np.float64)
        elif dtype == "int":
            self.index = Index(np.arange(N), dtype=np.int64)

    def time_add(self, dtype):
        self.index + 2

    def time_subtract(self, dtype):
        self.index - 2

    def time_multiply(self, dtype):
        self.index * 2

    def time_divide(self, dtype):
        self.index / 2

    def time_modulo(self, dtype):
        self.index % 2


class NumericInferOps:
    # from GH 7332
    params = numeric_dtypes
    param_names = ["dtype"]

    def setup(self, dtype):
        N = 5 * 10**5
        self.df = DataFrame(
            {"A": np.arange(N).astype(dtype), "B": np.arange(N).astype(dtype)}
        )

    def time_add(self, dtype):
        self.df["A"] + self.df["B"]

    def time_subtract(self, dtype):
        self.df["A"] - self.df["B"]

    def time_multiply(self, dtype):
        self.df["A"] * self.df["B"]

    def time_divide(self, dtype):
        self.df["A"] / self.df["B"]

    def time_modulo(self, dtype):
        self.df["A"] % self.df["B"]


class DateInferOps:
    # from GH 7332
    def setup_cache(self):
        N = 5 * 10**5
        df = DataFrame({"datetime64": np.arange(N).astype("datetime64[ms]")})
        df["timedelta"] = df["datetime64"] - df["datetime64"]
        return df

    def time_subtract_datetimes(self, df):
        df["datetime64"] - df["datetime64"]

    def time_timedelta_plus_datetime(self, df):
        df["timedelta"] + df["datetime64"]

    def time_add_timedeltas(self, df):
        df["timedelta"] + df["timedelta"]


hcal = pd.tseries.holiday.USFederalHolidayCalendar()
# These offsets currently raise a NotImplementedError with .apply_index()
non_apply = [
    pd.offsets.Day(),
    pd.offsets.BYearEnd(),
    pd.offsets.BYearBegin(),
    pd.offsets.BQuarterEnd(),
    pd.offsets.BQuarterBegin(),
    pd.offsets.BMonthEnd(),
    pd.offsets.BMonthBegin(),
    pd.offsets.CustomBusinessDay(),
    pd.offsets.CustomBusinessDay(calendar=hcal),
    pd.offsets.CustomBusinessMonthBegin(calendar=hcal),
    pd.offsets.CustomBusinessMonthEnd(calendar=hcal),
    pd.offsets.CustomBusinessMonthEnd(calendar=hcal),
]
other_offsets = [
    pd.offsets.YearEnd(),
    pd.offsets.YearBegin(),
    pd.offsets.QuarterEnd(),
    pd.offsets.QuarterBegin(),
    pd.offsets.MonthEnd(),
    pd.offsets.MonthBegin(),
    pd.offsets.DateOffset(months=2, days=2),
    pd.offsets.BusinessDay(),
    pd.offsets.SemiMonthEnd(),
    pd.offsets.SemiMonthBegin(),
]
offsets = non_apply + other_offsets


class OffsetArrayArithmetic:
    params = offsets
    param_names = ["offset"]

    def setup(self, offset):
        N = 10000
        rng = date_range(start="1/1/2000", periods=N, freq="min")
        self.rng = rng
        self.ser = Series(rng)

    def time_add_series_offset(self, offset):
        with warnings.catch_warnings(record=True):
            self.ser + offset

    def time_add_dti_offset(self, offset):
        with warnings.catch_warnings(record=True):
            self.rng + offset


class ApplyIndex:
    params = other_offsets
    param_names = ["offset"]

    def setup(self, offset):
        N = 10000
        rng = date_range(start="1/1/2000", periods=N, freq="min")
        self.rng = rng

    def time_apply_index(self, offset):
        self.rng + offset


class BinaryOpsMultiIndex:
    params = ["sub", "add", "mul", "div"]
    param_names = ["func"]

    def setup(self, func):
        array = date_range("20200101 00:00", "20200102 0:00", freq="s")
        level_0_names = [str(i) for i in range(30)]

        index = pd.MultiIndex.from_product([level_0_names, array])
        column_names = ["col_1", "col_2"]

        self.df = DataFrame(
            np.random.rand(len(index), 2), index=index, columns=column_names
        )

        self.arg_df = DataFrame(
            np.random.randint(1, 10, (len(level_0_names), 2)),
            index=level_0_names,
            columns=column_names,
        )

    def time_binary_op_multiindex(self, func):
        getattr(self.df, func)(self.arg_df, level=0)


from .pandas_vb_common import setup  # noqa: F401 isort:skip