test_operators_and_elementwise_functions.py

"""
Test element-wise functions/operators against reference implementations.
"""
import cmath
import math
import operator
import builtins
from copy import copy
from enum import Enum, auto
from typing import Callable, List, NamedTuple, Optional, Sequence, TypeVar, Union

import pytest
from hypothesis import assume, given
from hypothesis import strategies as st

from . import _array_module as xp, api_version
from . import array_helpers as ah
from . import dtype_helpers as dh
from . import hypothesis_helpers as hh
from . import pytest_helpers as ph
from . import shape_helpers as sh
from . import xps
from .typing import Array, DataType, Param, Scalar, ScalarType, Shape


pytestmark = pytest.mark.unvectorized


EPS32 = xp.finfo(xp.float32).eps


def mock_int_dtype(n: int, dtype: DataType) -> int:
    """Returns equivalent of `n` that mocks `dtype` behaviour."""
    nbits = dh.dtype_nbits[dtype]
    mask = (1 << nbits) - 1
    n &= mask
    if dh.dtype_signed[dtype]:
        highest_bit = 1 << (nbits - 1)
        if n & highest_bit:
            n = -((~n & mask) + 1)
    return n


def isclose(
    a: float,
    b: float,
    maximum: float,
    *,
    rel_tol: float = 0.25,
    abs_tol: float = 1,
) -> bool:
    """Wraps math.isclose with very generous defaults.

    This is useful for many floating-point operations where the spec does not
    make accuracy requirements.
    """
    if math.isnan(a) or math.isnan(b):
        raise ValueError(f"{a=} and {b=}, but input must be non-NaN")
    if math.isinf(a):
        return math.isinf(b) or abs(b) > math.log(maximum)
    elif math.isinf(b):
        return math.isinf(a) or abs(a) > math.log(maximum)
    return math.isclose(a, b, rel_tol=rel_tol, abs_tol=abs_tol)


def isclose_complex(
    a: complex,
    b: complex,
    maximum: float,
    *,
    rel_tol: float = 0.25,
    abs_tol: float = 1,
) -> bool:
    """Like isclose() but specifically for complex values."""
    if cmath.isnan(a) or cmath.isnan(b):
        raise ValueError(f"{a=} and {b=}, but input must be non-NaN")
    if cmath.isinf(a):
        return cmath.isinf(b) or abs(b) > 2**(math.log2(maximum)//2)
    elif cmath.isinf(b):
        return cmath.isinf(a) or abs(a) > 2**(math.log2(maximum)//2)
    return cmath.isclose(a, b, rel_tol=rel_tol, abs_tol=abs_tol)


def default_filter(s: Scalar) -> bool:
    """Returns False when s is a non-finite or a signed zero.

    Used by default as these values are typically special-cased.
    """
    if isinstance(s, int):  # note bools are ints
        return True
    else:
        return math.isfinite(s) and s != 0


T = TypeVar("T")


def unary_assert_against_refimpl(
    func_name: str,
    in_: Array,
    res: Array,
    refimpl: Callable[[T], T],
    *,
    res_stype: Optional[ScalarType] = None,
    filter_: Callable[[Scalar], bool] = default_filter,
    strict_check: Optional[bool] = None,
    expr_template: Optional[str] = None,
):
    """
    Assert unary element-wise results are as expected.

    We iterate through every element in the input and resulting arrays, casting
    the respective elements (0-D arrays) to Python scalars, and assert against
    the expected output specified by the passed reference implementation, e.g.

        >>> x = xp.asarray([[0, 1], [2, 4]])
        >>> out = xp.square(x)
        >>> unary_assert_against_refimpl('square', x, out, lambda s: s ** 2)

        is equivalent to

        >>> for idx in np.ndindex(x.shape):
        ...     expected = int(x[idx]) ** 2
        ...     assert int(out[idx]) == expected

    Casting
    -------

    The input scalar type is inferred from the input array's dtype like so

        Array dtypes      | Python builtin type
        ----------------- | ---------------------
        xp.bool           | bool
        xp.int*, xp.uint* | int
        xp.float*         | float
        xp.complex*       | complex

    If res_stype=None (the default), the result scalar type is the same as the
    input scalar type. We can also specify the result scalar type ourselves, e.g.

        >>> x = xp.asarray([42., xp.inf])
        >>> out = xp.isinf(x)  # should be [False, True]
        >>> unary_assert_against_refimpl('isinf', x, out, math.isinf, res_stype=bool)

    Filtering special-cased values
    ------------------------------

    Values which are special-cased can be present in the input array, but get
    filtered before they can be asserted against refimpl.

    If filter_=default_filter (the default), all non-finite and floating zero
    values are filtered, e.g.

        >>> unary_assert_against_refimpl('sin', x, out, math.sin)

        is equivalent to

        >>> for idx in np.ndindex(x.shape):
        ...     at_x = float(x[idx])
        ...     if math.isfinite(at_x) or at_x != 0:
        ...         expected = math.sin(at_x)
        ...         assert math.isclose(float(out[idx]), expected)

    We can also specify the filter function ourselves, e.g.

        >>> def sqrt_filter(s: float) -> bool:
        ...    return math.isfinite(s) and s >= 0
        >>> unary_assert_against_refimpl('sqrt', x, out, math.sqrt, filter_=sqrt_filter)

        is equivalent to

        >>> for idx in np.ndindex(x.shape):
        ...     at_x = float(x[idx])
        ...     if math.isfinite(s) and s >=0:
        ...         expected = math.sin(at_x)
        ...         assert math.isclose(float(out[idx]), expected)

    Note we leave special-cased values in the input arrays, so as to ensure
    their presence doesn't affect the outputs respective to non-special-cased
    elements. We specifically test special case bevaiour in test_special_cases.py.

    Assertion strictness
    --------------------

    If strict_check=None (the default), integer elements are strictly asserted
    against, and floating elements are loosely asserted against, e.g.

        >>> unary_assert_against_refimpl('square', x, out, lambda s: s ** 2)

        is equivalent to

        >>> for idx in np.ndindex(x.shape):
        ...     expected = in_stype(x[idx]) ** 2
        ...     if in_stype == int:
        ...         assert int(out[idx]) == expected
        ...     else:  # in_stype == float
        ...         assert math.isclose(float(out[idx]), expected)

    Specifying strict_check as True or False will assert strictly/loosely
    respectively, regardless of dtype. This is useful for testing functions that
    have definitive outputs for floating inputs, i.e. rounding functions.

    Expressions in errors
    ---------------------

    Assertion error messages include an expression, by default using func_name
    like so

        >>> x = xp.asarray([42., xp.inf])
        >>> out = xp.isinf(x)
        >>> out
        [False, False]
        >>> unary_assert_against_refimpl('isinf', x, out, math.isinf, res_stype=bool)
        AssertionError: out[1]=False, but should be isinf(x[1])=True ...

    We can specify the expression template ourselves, e.g.

        >>> x = xp.asarray(True)
        >>> out = xp.logical_not(x)
        >>> out
        True
        >>> unary_assert_against_refimpl(
        ...    'logical_not', x, out, expr_template='(not {})={}'
        ... )
        AssertionError: out=True, but should be (not True)=False ...

    """
    if in_.shape != res.shape:
        raise ValueError(f"{res.shape=}, but should be {in_.shape=}")
    if expr_template is None:
        expr_template = func_name + "({})={}"
    in_stype = dh.get_scalar_type(in_.dtype)
    if res_stype is None:
        res_stype = dh.get_scalar_type(res.dtype)
    if res.dtype == xp.bool:
        m, M = (None, None)
    elif res.dtype in dh.complex_dtypes:
        m, M = dh.dtype_ranges[dh.dtype_components[res.dtype]]
    else:
        m, M = dh.dtype_ranges[res.dtype]
    if in_.dtype in dh.complex_dtypes:
        component_filter = copy(filter_)
        filter_ = lambda s: component_filter(s.real) and component_filter(s.imag)
    for idx in sh.ndindex(in_.shape):
        scalar_i = in_stype(in_[idx])
        if not filter_(scalar_i):
            continue
        try:
            expected = refimpl(scalar_i)
        except OverflowError:
            continue
        if res.dtype != xp.bool:
            if res.dtype in dh.complex_dtypes:
                if expected.real <= m or expected.real >= M:
                    continue
                if expected.imag <= m or expected.imag >= M:
                    continue
            else:
                if expected <= m or expected >= M:
                    continue
        scalar_o = res_stype(res[idx])
        f_i = sh.fmt_idx("x", idx)
        f_o = sh.fmt_idx("out", idx)
        expr = expr_template.format(f_i, expected)
        # TODO: strict check floating results too
        if strict_check == False or res.dtype in dh.all_float_dtypes:
            msg = (
                f"{f_o}={scalar_o}, but should be roughly {expr} [{func_name}()]\n"
                f"{f_i}={scalar_i}"
            )
            if res.dtype in dh.complex_dtypes:
                assert isclose_complex(scalar_o, expected, M), msg
            else:
                assert isclose(scalar_o, expected, M), msg
        else:
            assert scalar_o == expected, (
                f"{f_o}={scalar_o}, but should be {expr} [{func_name}()]\n"
                f"{f_i}={scalar_i}"
            )


def binary_assert_against_refimpl(
    func_name: str,
    left: Array,
    right: Array,
    res: Array,
    refimpl: Callable[[T, T], T],
    *,
    res_stype: Optional[ScalarType] = None,
    filter_: Callable[[Scalar], bool] = default_filter,
    strict_check: Optional[bool] = None,
    left_sym: str = "x1",
    right_sym: str = "x2",
    res_name: str = "out",
    expr_template: Optional[str] = None,
):
    """
    Assert binary element-wise results are as expected.

    See unary_assert_against_refimpl for more information.
    """
    if expr_template is None:
        expr_template = func_name + "({}, {})={}"
    in_stype = dh.get_scalar_type(left.dtype)
    if res_stype is None:
        res_stype = dh.get_scalar_type(left.dtype)
    if res_stype is None:
        res_stype = in_stype
    if res.dtype == xp.bool:
        m, M = (None, None)
    elif res.dtype in dh.complex_dtypes:
        m, M = dh.dtype_ranges[dh.dtype_components[res.dtype]]
    else:
        m, M = dh.dtype_ranges[res.dtype]
    if left.dtype in dh.complex_dtypes:
        component_filter = copy(filter_)
        filter_ = lambda s: component_filter(s.real) and component_filter(s.imag)
    for l_idx, r_idx, o_idx in sh.iter_indices(left.shape, right.shape, res.shape):
        scalar_l = in_stype(left[l_idx])
        scalar_r = in_stype(right[r_idx])
        if not (filter_(scalar_l) and filter_(scalar_r)):
            continue
        try:
            expected = refimpl(scalar_l, scalar_r)
        except OverflowError:
            continue
        if res.dtype != xp.bool:
            if res.dtype in dh.complex_dtypes:
                if expected.real <= m or expected.real >= M:
                    continue
                if expected.imag <= m or expected.imag >= M:
                    continue
            else:
                if expected <= m or expected >= M:
                    continue
        scalar_o = res_stype(res[o_idx])
        f_l = sh.fmt_idx(left_sym, l_idx)
        f_r = sh.fmt_idx(right_sym, r_idx)
        f_o = sh.fmt_idx(res_name, o_idx)
        expr = expr_template.format(f_l, f_r, expected)
        if strict_check == False or res.dtype in dh.all_float_dtypes:
            msg = (
                f"{f_o}={scalar_o}, but should be roughly {expr} [{func_name}()]\n"
                f"{f_l}={scalar_l}, {f_r}={scalar_r}"
            )
            if res.dtype in dh.complex_dtypes:
                assert isclose_complex(scalar_o, expected, M), msg
            else:
                assert isclose(scalar_o, expected, M), msg
        else:
            assert scalar_o == expected, (
                f"{f_o}={scalar_o}, but should be {expr} [{func_name}()]\n"
                f"{f_l}={scalar_l}, {f_r}={scalar_r}"
            )


def right_scalar_assert_against_refimpl(
    func_name: str,
    left: Array,
    right: Scalar,
    res: Array,
    refimpl: Callable[[T, T], T],
    *,
    res_stype: Optional[ScalarType] = None,
    filter_: Callable[[Scalar], bool] = default_filter,
    strict_check: Optional[bool] = None,
    left_sym: str = "x1",
    res_name: str = "out",
    expr_template: str = None,
):
    """
    Assert binary element-wise results from scalar operands are as expected.

    See unary_assert_against_refimpl for more information.
    """
    if expr_template is None:
        expr_template = func_name + "({}, {})={}"
    if left.dtype in dh.complex_dtypes:
        component_filter = copy(filter_)
        filter_ = lambda s: component_filter(s.real) and component_filter(s.imag)
    if filter_(right):
        return  # short-circuit here as there will be nothing to test
    in_stype = dh.get_scalar_type(left.dtype)
    if res_stype is None:
        res_stype = dh.get_scalar_type(left.dtype)
    if res_stype is None:
        res_stype = in_stype
    if res.dtype == xp.bool:
        m, M = (None, None)
    elif left.dtype in dh.complex_dtypes:
        m, M = dh.dtype_ranges[dh.dtype_components[left.dtype]]
    else:
        m, M = dh.dtype_ranges[left.dtype]
    for idx in sh.ndindex(res.shape):
        scalar_l = in_stype(left[idx])
        if not (filter_(scalar_l) and filter_(right)):
            continue
        try:
            expected = refimpl(scalar_l, right)
        except OverflowError:
            continue
        if left.dtype != xp.bool:
            if res.dtype in dh.complex_dtypes:
                if expected.real <= m or expected.real >= M:
                    continue
                if expected.imag <= m or expected.imag >= M:
                    continue
            else:
                if expected <= m or expected >= M:
                    continue
        scalar_o = res_stype(res[idx])
        f_l = sh.fmt_idx(left_sym, idx)
        f_o = sh.fmt_idx(res_name, idx)
        expr = expr_template.format(f_l, right, expected)
        if strict_check == False or res.dtype in dh.all_float_dtypes:
            msg = (
                f"{f_o}={scalar_o}, but should be roughly {expr} [{func_name}()]\n"
                f"{f_l}={scalar_l}"
            )
            if res.dtype in dh.complex_dtypes:
                assert isclose_complex(scalar_o, expected, M), msg
            else:
                assert isclose(scalar_o, expected, M), msg
        else:
            assert scalar_o == expected, (
                f"{f_o}={scalar_o}, but should be {expr} [{func_name}()]\n"
                f"{f_l}={scalar_l}"
            )


# When appropriate, this module tests operators alongside their respective
# elementwise methods. We do this by parametrizing a generalised test method
# with every relevant method and operator.
#
# Notable arguments in the parameter's context object:
# - The function object, which for operator test cases is a wrapper that allows
#   test logic to be generalised.
# - The argument strategies, which can be used to draw arguments for the test
#   case. They may require additional filtering for certain test cases.
# - right_is_scalar (binary parameters only), which denotes if the right
#   argument is a scalar in a test case. This can be used to appropriately
#   adjust draw filtering and test logic.


func_to_op = {v: k for k, v in dh.op_to_func.items()}
all_op_to_symbol = {**dh.binary_op_to_symbol, **dh.inplace_op_to_symbol}
finite_kw = {"allow_nan": False, "allow_infinity": False}


class UnaryParamContext(NamedTuple):
    func_name: str
    func: Callable[[Array], Array]
    strat: st.SearchStrategy[Array]

    @property
    def id(self) -> str:
        return self.func_name

    def __repr__(self):
        return f"UnaryParamContext(<{self.id}>)"


def make_unary_params(
    elwise_func_name: str,
    dtypes: Sequence[DataType],
    *,
    min_version: str = "2021.12",
) -> List[Param[UnaryParamContext]]:
    dtypes = [d for d in dtypes if not isinstance(d, xp._UndefinedStub)]
    assert len(dtypes) > 0  # sanity check
    if api_version < "2022.12":
        dtypes = [d for d in dtypes if d not in dh.complex_dtypes]
    dtypes_strat = st.sampled_from(dtypes)
    strat = hh.arrays(dtype=dtypes_strat, shape=hh.shapes())
    func_ctx = UnaryParamContext(
        func_name=elwise_func_name, func=getattr(xp, elwise_func_name), strat=strat
    )
    op_name = func_to_op[elwise_func_name]
    op_ctx = UnaryParamContext(
        func_name=op_name, func=lambda x: getattr(x, op_name)(), strat=strat
    )
    if api_version < min_version:
        marks = pytest.mark.skip(
            reason=f"requires ARRAY_API_TESTS_VERSION >= {min_version}"
        )
    else:
        marks = ()
    return [
        pytest.param(func_ctx, id=func_ctx.id, marks=marks),
        pytest.param(op_ctx, id=op_ctx.id, marks=marks),
    ]


class FuncType(Enum):
    FUNC = auto()
    OP = auto()
    IOP = auto()


shapes_kw = {"min_side": 1}


class BinaryParamContext(NamedTuple):
    func_name: str
    func: Callable[[Array, Union[Scalar, Array]], Array]
    left_sym: str
    left_strat: st.SearchStrategy[Array]
    right_sym: str
    right_strat: st.SearchStrategy[Union[Scalar, Array]]
    right_is_scalar: bool
    res_name: str

    @property
    def id(self) -> str:
        return f"{self.func_name}({self.left_sym}, {self.right_sym})"

    def __repr__(self):
        return f"BinaryParamContext(<{self.id}>)"


def make_binary_params(
    elwise_func_name: str, dtypes: Sequence[DataType]
) -> List[Param[BinaryParamContext]]:
    dtypes = [d for d in dtypes if not isinstance(d, xp._UndefinedStub)]
    assert len(dtypes) > 0  # sanity check
    shared_oneway_dtypes = st.shared(hh.oneway_promotable_dtypes(dtypes))
    left_dtypes = shared_oneway_dtypes.map(lambda D: D.result_dtype)
    right_dtypes = shared_oneway_dtypes.map(lambda D: D.input_dtype)

    def make_param(
        func_name: str, func_type: FuncType, right_is_scalar: bool
    ) -> Param[BinaryParamContext]:
        if right_is_scalar:
            left_sym = "x"
            right_sym = "s"
        else:
            left_sym = "x1"
            right_sym = "x2"

        if right_is_scalar:
            left_strat = hh.arrays(dtype=left_dtypes, shape=hh.shapes(**shapes_kw))
            right_strat = right_dtypes.flatmap(lambda d: hh.from_dtype(d, **finite_kw))
        else:
            if func_type is FuncType.IOP:
                shared_oneway_shapes = st.shared(hh.oneway_broadcastable_shapes())
                left_strat = hh.arrays(
                    dtype=left_dtypes,
                    shape=shared_oneway_shapes.map(lambda S: S.result_shape),
                )
                right_strat = hh.arrays(
                    dtype=right_dtypes,
                    shape=shared_oneway_shapes.map(lambda S: S.input_shape),
                )
            else:
                mutual_shapes = st.shared(
                    hh.mutually_broadcastable_shapes(2, **shapes_kw)
                )
                left_strat = hh.arrays(
                    dtype=left_dtypes, shape=mutual_shapes.map(lambda pair: pair[0])
                )
                right_strat = hh.arrays(
                    dtype=right_dtypes, shape=mutual_shapes.map(lambda pair: pair[1])
                )

        if func_type is FuncType.FUNC:
            func = getattr(xp, func_name)
        else:
            op_sym = all_op_to_symbol[func_name]
            expr = f"{left_sym} {op_sym} {right_sym}"
            if func_type is FuncType.OP:

                def func(l: Array, r: Union[Scalar, Array]) -> Array:
                    locals_ = {}
                    locals_[left_sym] = l
                    locals_[right_sym] = r
                    return eval(expr, locals_)

            else:

                def func(l: Array, r: Union[Scalar, Array]) -> Array:
                    locals_ = {}
                    locals_[left_sym] = xp.asarray(l, copy=True)  # prevents mutating l
                    locals_[right_sym] = r
                    exec(expr, locals_)
                    return locals_[left_sym]

            func.__name__ = func_name  # for repr

        if func_type is FuncType.IOP:
            res_name = left_sym
        else:
            res_name = "out"

        ctx = BinaryParamContext(
            func_name,
            func,
            left_sym,
            left_strat,
            right_sym,
            right_strat,
            right_is_scalar,
            res_name,
        )
        return pytest.param(ctx, id=ctx.id)

    op_name = func_to_op[elwise_func_name]
    params = [
        make_param(elwise_func_name, FuncType.FUNC, False),
        make_param(op_name, FuncType.OP, False),
        make_param(op_name, FuncType.OP, True),
    ]
    iop_name = f"__i{op_name[2:]}"
    if iop_name in dh.inplace_op_to_symbol.keys():
        params.append(make_param(iop_name, FuncType.IOP, False))
        params.append(make_param(iop_name, FuncType.IOP, True))

    return params


def binary_param_assert_dtype(
    ctx: BinaryParamContext,
    left: Array,
    right: Union[Array, Scalar],
    res: Array,
    expected: Optional[DataType] = None,
):
    if ctx.right_is_scalar:
        in_dtypes = left.dtype
    else:
        in_dtypes = [left.dtype, right.dtype]  # type: ignore
    ph.assert_dtype(
        ctx.func_name, in_dtype=in_dtypes, out_dtype=res.dtype, expected=expected, repr_name=f"{ctx.res_name}.dtype"
    )


def binary_param_assert_shape(
    ctx: BinaryParamContext,
    left: Array,
    right: Union[Array, Scalar],
    res: Array,
    expected: Optional[Shape] = None,
):
    if ctx.right_is_scalar:
        in_shapes = [left.shape]
    else:
        in_shapes = [left.shape, right.shape]  # type: ignore
    ph.assert_result_shape(
        ctx.func_name, in_shapes=in_shapes, out_shape=res.shape, expected=expected, repr_name=f"{ctx.res_name}.shape"
    )


def binary_param_assert_against_refimpl(
    ctx: BinaryParamContext,
    left: Array,
    right: Union[Array, Scalar],
    res: Array,
    op_sym: str,
    refimpl: Callable[[T, T], T],
    *,
    res_stype: Optional[ScalarType] = None,
    filter_: Callable[[Scalar], bool] = default_filter,
    strict_check: Optional[bool] = None,
):
    expr_template = "({} " + op_sym + " {})={}"
    if ctx.right_is_scalar:
        right_scalar_assert_against_refimpl(
            func_name=ctx.func_name,
            left_sym=ctx.left_sym,
            left=left,
            right=right,
            res_stype=res_stype,
            res_name=ctx.res_name,
            res=res,
            refimpl=refimpl,
            expr_template=expr_template,
            filter_=filter_,
            strict_check=strict_check,
        )
    else:
        binary_assert_against_refimpl(
            func_name=ctx.func_name,
            left_sym=ctx.left_sym,
            left=left,
            right_sym=ctx.right_sym,
            right=right,
            res_stype=res_stype,
            res_name=ctx.res_name,
            res=res,
            refimpl=refimpl,
            expr_template=expr_template,
            filter_=filter_,
            strict_check=strict_check,
        )


def _convert_scalars_helper(x1, x2):
    """Convert python scalar to arrays, record the shapes/dtypes of arrays.

    For inputs being scalars or arrays, return the dtypes and shapes of array arguments,
    and all arguments converted to arrays.

    dtypes are separate to help distinguishing between 
    `py_scalar + f32_array -> f32_array` and `f64_array + f32_array -> f64_array`
    """
    if dh.is_scalar(x1):
        in_dtypes = [x2.dtype]
        in_shapes = [x2.shape]
        x1a, x2a = xp.asarray(x1), x2
    elif dh.is_scalar(x2):
        in_dtypes = [x1.dtype]
        in_shapes = [x1.shape]
        x1a, x2a = x1, xp.asarray(x2)
    else:
        in_dtypes = [x1.dtype, x2.dtype]
        in_shapes = [x1.shape, x2.shape]
        x1a, x2a = x1, x2

    return in_dtypes, in_shapes, (x1a, x2a)


def _assert_correctness_binary(
    name, func, in_dtypes, in_shapes, in_arrs, out, expected_dtype=None, **kwargs
):
    x1a, x2a = in_arrs
    ph.assert_dtype(name, in_dtype=in_dtypes, out_dtype=out.dtype, expected=expected_dtype)
    ph.assert_result_shape(name, in_shapes=in_shapes, out_shape=out.shape)
    binary_assert_against_refimpl(name, x1a, x2a, out, func, **kwargs)


@pytest.mark.parametrize("ctx", make_unary_params("abs", dh.numeric_dtypes))
@given(data=st.data())
def test_abs(ctx, data):
    x = data.draw(ctx.strat, label="x")
    # abs of the smallest negative integer is out-of-scope
    if x.dtype in dh.int_dtypes:
        assume(xp.all(x > dh.dtype_ranges[x.dtype].min))

    out = ctx.func(x)

    if x.dtype in dh.complex_dtypes:
        assert out.dtype == dh.dtype_components[x.dtype]
    else:
        ph.assert_dtype(ctx.func_name, in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape(ctx.func_name, out_shape=out.shape, expected=x.shape)
    unary_assert_against_refimpl(
        ctx.func_name,
        x,
        out,
        abs,  # type: ignore
        res_stype=float if x.dtype in dh.complex_dtypes else None,
        expr_template="abs({})={}",
        # filter_=lambda s: (
        #     s == float("infinity") or (cmath.isfinite(s) and not ph.is_neg_zero(s))
        # ),
    )


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_acos(x):
    out = xp.acos(x)
    ph.assert_dtype("acos", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("acos", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.acos if x.dtype in dh.complex_dtypes else math.acos
    filter_ = default_filter if x.dtype in dh.complex_dtypes else lambda s: default_filter(s) and -1 <= s <= 1
    unary_assert_against_refimpl(
        "acos", x, out, refimpl, filter_=filter_
    )


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_acosh(x):
    out = xp.acosh(x)
    ph.assert_dtype("acosh", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("acosh", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.acosh if x.dtype in dh.complex_dtypes else math.acosh
    filter_ = default_filter if x.dtype in dh.complex_dtypes else lambda s: default_filter(s) and s >= 1
    unary_assert_against_refimpl(
        "acosh", x, out, refimpl, filter_=filter_
    )


@pytest.mark.parametrize("ctx,", make_binary_params("add", dh.numeric_dtypes))
@given(data=st.data())
def test_add(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)

    with hh.reject_overflow():
        res = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, res)
    binary_param_assert_shape(ctx, left, right, res)
    binary_param_assert_against_refimpl(ctx, left, right, res, "+", operator.add)


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_asin(x):
    out = xp.asin(x)
    ph.assert_dtype("asin", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("asin", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.asin if x.dtype in dh.complex_dtypes else math.asin
    filter_ = default_filter if x.dtype in dh.complex_dtypes else lambda s: default_filter(s) and -1 <= s <= 1
    unary_assert_against_refimpl(
        "asin", x, out, refimpl, filter_=filter_
    )


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_asinh(x):
    out = xp.asinh(x)
    ph.assert_dtype("asinh", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("asinh", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.asinh if x.dtype in dh.complex_dtypes else math.asinh
    unary_assert_against_refimpl("asinh", x, out, refimpl)


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_atan(x):
    out = xp.atan(x)
    ph.assert_dtype("atan", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("atan", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.atan if x.dtype in dh.complex_dtypes else math.atan
    unary_assert_against_refimpl("atan", x, out, refimpl)


@given(*hh.two_mutual_arrays(dh.real_float_dtypes))
def test_atan2(x1, x2):
    out = xp.atan2(x1, x2)
    _assert_correctness_binary(
        "atan",
        cmath.atan2 if x1.dtype in dh.complex_dtypes else math.atan2,
        in_dtypes=[x1.dtype, x2.dtype],
        in_shapes=[x1.shape, x2.shape],
        in_arrs=[x1, x2],
        out=out,
    )


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_atanh(x):
    out = xp.atanh(x)
    ph.assert_dtype("atanh", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("atanh", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.atanh if x.dtype in dh.complex_dtypes else math.atanh
    filter_ = default_filter if x.dtype in dh.complex_dtypes else lambda s: default_filter(s) and -1 < s < 1
    unary_assert_against_refimpl(
        "atanh",
        x,
        out,
        refimpl,
        filter_=filter_,
    )


@pytest.mark.parametrize(
    "ctx", make_binary_params("bitwise_and", dh.bool_and_all_int_dtypes)
)
@given(data=st.data())
def test_bitwise_and(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)

    res = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, res)
    binary_param_assert_shape(ctx, left, right, res)
    if left.dtype == xp.bool:
        refimpl = operator.and_
    else:
        refimpl = lambda l, r: mock_int_dtype(l & r, res.dtype)
    binary_param_assert_against_refimpl(ctx, left, right, res, "&", refimpl)


@pytest.mark.parametrize(
    "ctx", make_binary_params("bitwise_left_shift", dh.all_int_dtypes)
)
@given(data=st.data())
def test_bitwise_left_shift(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)
    if ctx.right_is_scalar:
        assume(right >= 0)
    else:
        assume(not xp.any(ah.isnegative(right)))

    res = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, res)
    binary_param_assert_shape(ctx, left, right, res)
    nbits = dh.dtype_nbits[res.dtype]
    binary_param_assert_against_refimpl(
        ctx, left, right, res, "<<", lambda l, r: l << r if r < nbits else 0
    )


@pytest.mark.parametrize(
    "ctx", make_unary_params("bitwise_invert", dh.bool_and_all_int_dtypes)
)
@given(data=st.data())
def test_bitwise_invert(ctx, data):
    x = data.draw(ctx.strat, label="x")

    out = ctx.func(x)

    ph.assert_dtype(ctx.func_name, in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape(ctx.func_name, out_shape=out.shape, expected=x.shape)
    if x.dtype == xp.bool:
        refimpl = operator.not_
    else:
        refimpl = lambda s: mock_int_dtype(~s, x.dtype)
    unary_assert_against_refimpl(ctx.func_name, x, out, refimpl, expr_template="~{}={}")


@pytest.mark.parametrize(
    "ctx", make_binary_params("bitwise_or", dh.bool_and_all_int_dtypes)
)
@given(data=st.data())
def test_bitwise_or(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)

    res = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, res)
    binary_param_assert_shape(ctx, left, right, res)
    if left.dtype == xp.bool:
        refimpl = operator.or_
    else:
        refimpl = lambda l, r: mock_int_dtype(l | r, res.dtype)
    binary_param_assert_against_refimpl(ctx, left, right, res, "|", refimpl)


@pytest.mark.parametrize(
    "ctx", make_binary_params("bitwise_right_shift", dh.all_int_dtypes)
)
@given(data=st.data())
def test_bitwise_right_shift(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)
    if ctx.right_is_scalar:
        assume(right >= 0)
    else:
        assume(not xp.any(ah.isnegative(right)))

    res = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, res)
    binary_param_assert_shape(ctx, left, right, res)
    binary_param_assert_against_refimpl(
        ctx, left, right, res, ">>", lambda l, r: mock_int_dtype(l >> r, res.dtype)
    )


@pytest.mark.parametrize(
    "ctx", make_binary_params("bitwise_xor", dh.bool_and_all_int_dtypes)
)
@given(data=st.data())
def test_bitwise_xor(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)

    res = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, res)
    binary_param_assert_shape(ctx, left, right, res)
    if left.dtype == xp.bool:
        refimpl = operator.xor
    else:
        refimpl = lambda l, r: mock_int_dtype(l ^ r, res.dtype)
    binary_param_assert_against_refimpl(ctx, left, right, res, "^", refimpl)


@given(hh.arrays(dtype=hh.real_dtypes, shape=hh.shapes()))
def test_ceil(x):
    out = xp.ceil(x)
    ph.assert_dtype("ceil", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("ceil", out_shape=out.shape, expected=x.shape)
    unary_assert_against_refimpl("ceil", x, out, math.ceil, strict_check=True)


@pytest.mark.min_version("2023.12")
@given(x=hh.arrays(dtype=hh.real_dtypes, shape=hh.shapes()), data=st.data())
def test_clip(x, data):
    # Ensure that if both min and max are arrays that all three of x, min, max
    # are broadcast compatible.
    shape1, shape2 = data.draw(hh.mutually_broadcastable_shapes(2,
                                                                base_shape=x.shape),
                                label="min.shape, max.shape")

    dtypes = hh.real_floating_dtypes if dh.is_float_dtype(x.dtype) else hh.int_dtypes

    min = data.draw(st.one_of(
        st.none(),
        hh.scalars(dtypes=st.just(x.dtype)),
        hh.arrays(dtype=dtypes, shape=shape1),
    ), label="min")
    max = data.draw(st.one_of(
        st.none(),
        hh.scalars(dtypes=st.just(x.dtype)),
        hh.arrays(dtype=dtypes, shape=shape2),
    ), label="max")

    # min > max is undefined (but allow nans)
    assume(min is None or max is None or not xp.any(ah.less(xp.asarray(max), xp.asarray(min))))

    kw = data.draw(
        hh.specified_kwargs(
            ("min", min, None),
            ("max", max, None)),
        label="kwargs")

    out = xp.clip(x, **kw)

    # min and max do not participate in type promotion
    ph.assert_dtype("clip", in_dtype=x.dtype, out_dtype=out.dtype)

    shapes = [x.shape]
    if min is not None and not dh.is_scalar(min):
        shapes.append(min.shape)
    if max is not None and not dh.is_scalar(max):
        shapes.append(max.shape)
    expected_shape = sh.broadcast_shapes(*shapes)
    ph.assert_shape("clip", out_shape=out.shape, expected=expected_shape)

    # This is based on right_scalar_assert_against_refimpl and
    # binary_assert_against_refimpl. clip() is currently the only ternary
    # elementwise function and the only function that supports arrays and
    # scalars. However, where() (in test_searching_functions) is similar
    # and if scalar support is added to it, we may want to factor out and
    # reuse this logic.

    def refimpl(_x, _min, _max):
        # Skip cases where _min and _max are integers whose values do not
        # fit in the dtype of _x, since this behavior is unspecified.
        if dh.is_int_dtype(x.dtype):
            if _min is not None and _min not in dh.dtype_ranges[x.dtype]:
                return None
            if _max is not None and _max not in dh.dtype_ranges[x.dtype]:
                return None

        # If min or max are float64 and x is float32, they will need to be
        # downcast to float32. This could result in a round in the wrong
        # direction meaning the resulting clipped value might not actually be
        # between min and max. This behavior is unspecified, so skip any cases
        # where x is within the rounding error of downcasting min or max.
        if x.dtype == xp.float32:
            if min is not None and not dh.is_scalar(min) and min.dtype == xp.float64 and math.isfinite(_min):
                _min_float32 = float(xp.asarray(_min, dtype=xp.float32))
                if math.isinf(_min_float32):
                    return None
                tol = abs(_min - _min_float32)
                if math.isclose(_min, _min_float32, abs_tol=tol):
                    return None
            if max is not None and not dh.is_scalar(max) and max.dtype == xp.float64 and math.isfinite(_max):
                _max_float32 = float(xp.asarray(_max, dtype=xp.float32))
                if math.isinf(_max_float32):
                    return None
                tol = abs(_max - _max_float32)
                if math.isclose(_max, _max_float32, abs_tol=tol):
                    return None

        if (math.isnan(_x)
            or (_min is not None and math.isnan(_min))
            or (_max is not None and math.isnan(_max))):
            return math.nan
        if _min is _max is None:
            return _x
        if _max is None:
            return builtins.max(_x, _min)
        if _min is None:
            return builtins.min(_x, _max)
        return builtins.min(builtins.max(_x, _min), _max)

    stype = dh.get_scalar_type(x.dtype)
    min_shape = () if min is None or dh.is_scalar(min) else min.shape
    max_shape = () if max is None or dh.is_scalar(max) else max.shape

    for x_idx, min_idx, max_idx, o_idx in sh.iter_indices(
            x.shape, min_shape, max_shape, out.shape):
        x_val = stype(x[x_idx])
        if min is None or dh.is_scalar(min):
            min_val = min
        else:
            min_val = stype(min[min_idx])
        if max is None or dh.is_scalar(max):
            max_val = max
        else:
            max_val = stype(max[max_idx])
        expected = refimpl(x_val, min_val, max_val)
        if expected is None:
            continue
        out_val = stype(out[o_idx])
        if math.isnan(expected):
            assert math.isnan(out_val), (
                f"out[{o_idx}]={out[o_idx]} but should be nan [clip()]\n"
                f"x[{x_idx}]={x_val}, min[{min_idx}]={min_val}, max[{max_idx}]={max_val}"
            )
        else:
            if out.dtype == xp.float32:
                # conversion to builtin float is prone to roundoff errors
                close_enough = math.isclose(out_val, expected, rel_tol=EPS32)
            else:
                close_enough = out_val == expected

            assert close_enough, (
                f"out[{o_idx}]={out[o_idx]} but should be {expected} [clip()]\n"
                f"x[{x_idx}]={x_val}, min[{min_idx}]={min_val}, max[{max_idx}]={max_val}"
            )


@pytest.mark.min_version("2022.12")
@pytest.mark.skipif(hh.complex_dtypes.is_empty, reason="no complex data types to draw from")
@given(hh.arrays(dtype=hh.complex_dtypes, shape=hh.shapes()))
def test_conj(x):
    out = xp.conj(x)
    ph.assert_dtype("conj", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("conj", out_shape=out.shape, expected=x.shape)
    unary_assert_against_refimpl("conj", x, out, operator.methodcaller("conjugate"))


@pytest.mark.min_version("2023.12")
@given(*hh.two_mutual_arrays(dh.real_float_dtypes))
def test_copysign(x1, x2):
    out = xp.copysign(x1, x2)
    ph.assert_dtype("copysign", in_dtype=[x1.dtype, x2.dtype], out_dtype=out.dtype)
    ph.assert_result_shape("copysign", in_shapes=[x1.shape, x2.shape], out_shape=out.shape)
    binary_assert_against_refimpl("copysign", x1, x2, out, math.copysign)


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_cos(x):
    out = xp.cos(x)
    ph.assert_dtype("cos", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("cos", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.cos if x.dtype in dh.complex_dtypes else math.cos
    unary_assert_against_refimpl("cos", x, out, refimpl)


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_cosh(x):
    out = xp.cosh(x)
    ph.assert_dtype("cosh", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("cosh", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.cosh if x.dtype in dh.complex_dtypes else math.cosh
    unary_assert_against_refimpl("cosh", x, out, refimpl)


@pytest.mark.parametrize("ctx", make_binary_params("divide", dh.all_float_dtypes))
@given(data=st.data())
def test_divide(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)
    if ctx.right_is_scalar:
        assume  # TODO: assume what?

    res = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, res)
    binary_param_assert_shape(ctx, left, right, res)
    binary_param_assert_against_refimpl(
        ctx,
        left,
        right,
        res,
        "/",
        operator.truediv,
        filter_=lambda s: cmath.isfinite(s) and s != 0,
    )


@pytest.mark.parametrize("ctx", make_binary_params("equal", dh.all_dtypes))
@given(data=st.data())
def test_equal(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)

    out = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, out, xp.bool)
    binary_param_assert_shape(ctx, left, right, out)
    if not ctx.right_is_scalar:
        # We manually promote the dtypes as incorrect internal type promotion
        # could lead to false positives. For example
        #
        #     >>> xp.equal(
        #     ...     xp.asarray(1.0, dtype=xp.float32),
        #     ...     xp.asarray(1.00000001, dtype=xp.float64),
        #     ... )
        #
        # would erroneously be True if float64 downcasted to float32.
        promoted_dtype = dh.promotion_table[left.dtype, right.dtype]
        left = xp.astype(left, promoted_dtype)
        right = xp.astype(right, promoted_dtype)
    binary_param_assert_against_refimpl(
        ctx, left, right, out, "==", operator.eq, res_stype=bool
    )


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_exp(x):
    out = xp.exp(x)
    ph.assert_dtype("exp", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("exp", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.exp if x.dtype in dh.complex_dtypes else math.exp
    unary_assert_against_refimpl("exp", x, out, refimpl)


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_expm1(x):
    out = xp.expm1(x)
    ph.assert_dtype("expm1", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("expm1", out_shape=out.shape, expected=x.shape)
    if x.dtype in dh.complex_dtypes:
        def refimpl(z):
            # There's no cmath.expm1. Use
            #
            # exp(x+yi) - 1
            # = exp(x)exp(yi) - 1
            # = exp(x)(cos(y) + sin(y)i) - 1
            # = (exp(x) - 1)cos(y) + (cos(y) - 1) + exp(x)sin(y)i
            # = expm1(x)cos(y) - 2sin(y/2)^2 + exp(x)sin(y)i
            #
            # where 1 - cos(y) = 2sin(y/2)^2 is used to avoid loss of
            # significance near y = 0.
            re, im = z.real, z.imag
            return math.expm1(re)*math.cos(im) - 2*math.sin(im/2)**2 + 1j*math.exp(re)*math.sin(im)
    else:
        refimpl = math.expm1
    unary_assert_against_refimpl("expm1", x, out, refimpl)


@given(hh.arrays(dtype=hh.real_dtypes, shape=hh.shapes()))
def test_floor(x):
    out = xp.floor(x)
    ph.assert_dtype("floor", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("floor", out_shape=out.shape, expected=x.shape)
    if x.dtype in dh.complex_dtypes:
        def refimpl(z):
            return complex(math.floor(z.real), math.floor(z.imag))
    else:
        refimpl = math.floor
    unary_assert_against_refimpl("floor", x, out, refimpl, strict_check=True)


@pytest.mark.parametrize("ctx", make_binary_params("floor_divide", dh.real_dtypes))
@given(data=st.data())
def test_floor_divide(ctx, data):
    left = data.draw(
        ctx.left_strat.filter(lambda x: not xp.any(x == 0)), label=ctx.left_sym
    )
    right = data.draw(ctx.right_strat, label=ctx.right_sym)
    if ctx.right_is_scalar:
        assume(right != 0)
    else:
        assume(not xp.any(right == 0))

    res = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, res)
    binary_param_assert_shape(ctx, left, right, res)
    binary_param_assert_against_refimpl(ctx, left, right, res, "//", operator.floordiv)


@pytest.mark.parametrize("ctx", make_binary_params("greater", dh.real_dtypes))
@given(data=st.data())
def test_greater(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)

    out = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, out, xp.bool)
    binary_param_assert_shape(ctx, left, right, out)
    if not ctx.right_is_scalar:
        # See test_equal note
        promoted_dtype = dh.promotion_table[left.dtype, right.dtype]
        left = xp.astype(left, promoted_dtype)
        right = xp.astype(right, promoted_dtype)
    binary_param_assert_against_refimpl(
        ctx, left, right, out, ">", operator.gt, res_stype=bool
    )


@pytest.mark.parametrize("ctx", make_binary_params("greater_equal", dh.real_dtypes))
@given(data=st.data())
def test_greater_equal(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)

    out = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, out, xp.bool)
    binary_param_assert_shape(ctx, left, right, out)
    if not ctx.right_is_scalar:
        # See test_equal note
        promoted_dtype = dh.promotion_table[left.dtype, right.dtype]
        left = xp.astype(left, promoted_dtype)
        right = xp.astype(right, promoted_dtype)
    binary_param_assert_against_refimpl(
        ctx, left, right, out, ">=", operator.ge, res_stype=bool
    )


@pytest.mark.min_version("2023.12")
@given(*hh.two_mutual_arrays(dh.real_float_dtypes))
def test_hypot(x1, x2):
    out = xp.hypot(x1, x2)
    _assert_correctness_binary(
        "hypot",
        math.hypot,
        in_dtypes=[x1.dtype, x2.dtype],
        in_shapes=[x1.shape, x2.shape],
        in_arrs=[x1, x2],
        out=out
    )


@pytest.mark.min_version("2022.12")
@pytest.mark.skipif(hh.complex_dtypes.is_empty, reason="no complex data types to draw from")
@given(hh.arrays(dtype=hh.complex_dtypes, shape=hh.shapes()))
def test_imag(x):
    out = xp.imag(x)
    ph.assert_dtype("imag", in_dtype=x.dtype, out_dtype=out.dtype, expected=dh.dtype_components[x.dtype])
    ph.assert_shape("imag", out_shape=out.shape, expected=x.shape)
    unary_assert_against_refimpl("imag", x, out, operator.attrgetter("imag"))


@given(hh.arrays(dtype=hh.numeric_dtypes, shape=hh.shapes()))
def test_isfinite(x):
    out = xp.isfinite(x)
    ph.assert_dtype("isfinite", in_dtype=x.dtype, out_dtype=out.dtype, expected=xp.bool)
    ph.assert_shape("isfinite", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.isfinite if x.dtype in dh.complex_dtypes else math.isfinite
    unary_assert_against_refimpl("isfinite", x, out, refimpl, res_stype=bool)


@given(hh.arrays(dtype=hh.numeric_dtypes, shape=hh.shapes()))
def test_isinf(x):
    out = xp.isinf(x)
    ph.assert_dtype("isfinite", in_dtype=x.dtype, out_dtype=out.dtype, expected=xp.bool)
    ph.assert_shape("isinf", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.isinf if x.dtype in dh.complex_dtypes else math.isinf
    unary_assert_against_refimpl("isinf", x, out, refimpl, res_stype=bool)


@given(hh.arrays(dtype=hh.numeric_dtypes, shape=hh.shapes()))
def test_isnan(x):
    out = xp.isnan(x)
    ph.assert_dtype("isnan", in_dtype=x.dtype, out_dtype=out.dtype, expected=xp.bool)
    ph.assert_shape("isnan", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.isnan if x.dtype in dh.complex_dtypes else math.isnan
    unary_assert_against_refimpl("isnan", x, out, refimpl, res_stype=bool)


@pytest.mark.parametrize("ctx", make_binary_params("less", dh.real_dtypes))
@given(data=st.data())
def test_less(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)

    out = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, out, xp.bool)
    binary_param_assert_shape(ctx, left, right, out)
    if not ctx.right_is_scalar:
        # See test_equal note
        promoted_dtype = dh.promotion_table[left.dtype, right.dtype]
        left = xp.astype(left, promoted_dtype)
        right = xp.astype(right, promoted_dtype)
    binary_param_assert_against_refimpl(
        ctx, left, right, out, "<", operator.lt, res_stype=bool
    )


@pytest.mark.parametrize("ctx", make_binary_params("less_equal", dh.real_dtypes))
@given(data=st.data())
def test_less_equal(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)

    out = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, out, xp.bool)
    binary_param_assert_shape(ctx, left, right, out)
    if not ctx.right_is_scalar:
        # See test_equal note
        promoted_dtype = dh.promotion_table[left.dtype, right.dtype]
        left = xp.astype(left, promoted_dtype)
        right = xp.astype(right, promoted_dtype)
    binary_param_assert_against_refimpl(
        ctx, left, right, out, "<=", operator.le, res_stype=bool
    )


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_log(x):
    out = xp.log(x)
    ph.assert_dtype("log", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("log", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.log if x.dtype in dh.complex_dtypes else math.log
    filter_ = default_filter if x.dtype in dh.complex_dtypes else lambda s: default_filter(s) and s > 0
    unary_assert_against_refimpl(
        "log", x, out, refimpl, filter_=filter_
    )


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_log1p(x):
    out = xp.log1p(x)
    ph.assert_dtype("log1p", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("log1p", out_shape=out.shape, expected=x.shape)
    # There isn't a cmath.log1p, and implementing one isn't straightforward
    # (see
    # https://stackoverflow.com/questions/78318212/unexpected-behaviour-of-log1p-numpy).
    # For now, just use log(1+p) for complex inputs, which should hopefully be
    # fine given the very loose numerical tolerances we use. If it isn't, we
    # can try using something like a series expansion for small p.
    if x.dtype in dh.complex_dtypes:
        refimpl = lambda z: cmath.log(1+z)
    else:
        refimpl = math.log1p
    filter_ = default_filter if x.dtype in dh.complex_dtypes else lambda s: default_filter(s) and s > -1
    unary_assert_against_refimpl(
        "log1p", x, out, refimpl, filter_=filter_
    )


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_log2(x):
    out = xp.log2(x)
    ph.assert_dtype("log2", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("log2", out_shape=out.shape, expected=x.shape)
    if x.dtype in dh.complex_dtypes:
        refimpl = lambda z: cmath.log(z)/math.log(2)
    else:
        refimpl = math.log2
    filter_ = default_filter if x.dtype in dh.complex_dtypes else lambda s: default_filter(s) and s > 0
    unary_assert_against_refimpl(
        "log2", x, out, refimpl, filter_=filter_
    )


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_log10(x):
    out = xp.log10(x)
    ph.assert_dtype("log10", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("log10", out_shape=out.shape, expected=x.shape)
    if x.dtype in dh.complex_dtypes:
        refimpl = lambda z: cmath.log(z)/math.log(10)
    else:
        refimpl = math.log10
    filter_ = default_filter if x.dtype in dh.complex_dtypes else lambda s: default_filter(s) and s > 0
    unary_assert_against_refimpl(
        "log10", x, out, refimpl, filter_=filter_
    )


def logaddexp_refimpl(l: float, r: float) -> float:
    try:
        return math.log(math.exp(l) + math.exp(r))
    except ValueError: # raised for log(0.)
        raise OverflowError


@pytest.mark.min_version("2023.12")
@given(*hh.two_mutual_arrays(dh.real_float_dtypes))
def test_logaddexp(x1, x2):
    out = xp.logaddexp(x1, x2)
    _assert_correctness_binary(
        "logaddexp",
        logaddexp_refimpl,
        in_dtypes=[x1.dtype, x2.dtype],
        in_shapes=[x1.shape, x2.shape],
        in_arrs=[x1, x2],
        out=out
    )


@given(hh.arrays(dtype=xp.bool, shape=hh.shapes()))
def test_logical_not(x):
    out = xp.logical_not(x)
    ph.assert_dtype("logical_not", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("logical_not", out_shape=out.shape, expected=x.shape)
    unary_assert_against_refimpl(
        "logical_not", x, out, operator.not_, expr_template="(not {})={}"
    )


@given(*hh.two_mutual_arrays([xp.bool]))
def test_logical_and(x1, x2):
    out = xp.logical_and(x1, x2)
    _assert_correctness_binary(
        "logical_and",
        operator.and_,
        in_dtypes=[x1.dtype, x2.dtype],
        in_shapes=[x1.shape, x2.shape],
        in_arrs=[x1, x2],
        out=out,
        expr_template="({} and {})={}"
    )


@given(*hh.two_mutual_arrays([xp.bool]))
def test_logical_or(x1, x2):
    out = xp.logical_or(x1, x2)
    _assert_correctness_binary(
        "logical_or",
        operator.or_,
        in_dtypes=[x1.dtype, x2.dtype],
        in_shapes=[x1.shape, x2.shape],
        in_arrs=[x1, x2],
        out=out,
        expr_template="({} or {})={}"
    )


@given(*hh.two_mutual_arrays([xp.bool]))
def test_logical_xor(x1, x2):
    out = xp.logical_xor(x1, x2)
    _assert_correctness_binary(
        "logical_xor",
        operator.xor,
        in_dtypes=[x1.dtype, x2.dtype],
        in_shapes=[x1.shape, x2.shape],
        in_arrs=[x1, x2],
        out=out,
        expr_template="({} ^ {})={}"
    )


@pytest.mark.min_version("2023.12")
@given(*hh.two_mutual_arrays(dh.real_float_dtypes))
def test_maximum(x1, x2):
    out = xp.maximum(x1, x2)
    _assert_correctness_binary(
        "maximum", max, [x1.dtype, x2.dtype], [x1.shape, x2.shape], (x1, x2), out, strict_check=True
    )


@pytest.mark.min_version("2023.12")
@given(*hh.two_mutual_arrays(dh.real_float_dtypes))
def test_minimum(x1, x2):
    out = xp.minimum(x1, x2)
    _assert_correctness_binary(
        "minimum", min, [x1.dtype, x2.dtype], [x1.shape, x2.shape], (x1, x2), out, strict_check=True
    )


@pytest.mark.parametrize("ctx", make_binary_params("multiply", dh.numeric_dtypes))
@given(data=st.data())
def test_multiply(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)

    res = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, res)
    binary_param_assert_shape(ctx, left, right, res)
    binary_param_assert_against_refimpl(ctx, left, right, res, "*", operator.mul)


# TODO: clarify if uints are acceptable, adjust accordingly
@pytest.mark.parametrize("ctx", make_unary_params("negative", dh.numeric_dtypes))
@given(data=st.data())
def test_negative(ctx, data):
    x = data.draw(ctx.strat, label="x")
    # negative of the smallest negative integer is out-of-scope
    if x.dtype in dh.int_dtypes:
        assume(xp.all(x > dh.dtype_ranges[x.dtype].min))

    out = ctx.func(x)

    ph.assert_dtype(ctx.func_name, in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape(ctx.func_name, out_shape=out.shape, expected=x.shape)
    unary_assert_against_refimpl(
        ctx.func_name, x, out, operator.neg, expr_template="-({})={}"  # type: ignore
    )


@pytest.mark.parametrize("ctx", make_binary_params("not_equal", dh.all_dtypes))
@given(data=st.data())
def test_not_equal(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)

    out = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, out, xp.bool)
    binary_param_assert_shape(ctx, left, right, out)
    if not ctx.right_is_scalar:
        # See test_equal note
        promoted_dtype = dh.promotion_table[left.dtype, right.dtype]
        left = xp.astype(left, promoted_dtype)
        right = xp.astype(right, promoted_dtype)
    binary_param_assert_against_refimpl(
        ctx, left, right, out, "!=", operator.ne, res_stype=bool
    )


@pytest.mark.parametrize("ctx", make_unary_params("positive", dh.numeric_dtypes))
@given(data=st.data())
def test_positive(ctx, data):
    x = data.draw(ctx.strat, label="x")

    out = ctx.func(x)

    ph.assert_dtype(ctx.func_name, in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape(ctx.func_name, out_shape=out.shape, expected=x.shape)
    ph.assert_array_elements(ctx.func_name, out=out, expected=x)


@pytest.mark.parametrize("ctx", make_binary_params("pow", dh.numeric_dtypes))
@given(data=st.data())
def test_pow(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)
    if ctx.right_is_scalar:
        if isinstance(right, int):
            assume(right >= 0)
    else:
        if dh.is_int_dtype(right.dtype):
            assume(xp.all(right >= 0))

    with hh.reject_overflow():
        res = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, res)
    binary_param_assert_shape(ctx, left, right, res)
    # Values testing pow is too finicky


@pytest.mark.min_version("2022.12")
@pytest.mark.skipif(hh.complex_dtypes.is_empty, reason="no complex data types to draw from")
@given(hh.arrays(dtype=hh.complex_dtypes, shape=hh.shapes()))
def test_real(x):
    out = xp.real(x)
    ph.assert_dtype("real", in_dtype=x.dtype, out_dtype=out.dtype, expected=dh.dtype_components[x.dtype])
    ph.assert_shape("real", out_shape=out.shape, expected=x.shape)
    unary_assert_against_refimpl("real", x, out, operator.attrgetter("real"))


@pytest.mark.min_version("2024.12")
@given(hh.arrays(dtype=hh.floating_dtypes, shape=hh.shapes(), elements=finite_kw))
def test_reciprocal(x):
    out = xp.reciprocal(x)
    ph.assert_dtype("reciprocal", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("reciprocal", out_shape=out.shape, expected=x.shape)
    refimpl = lambda x: 1.0 / x
    unary_assert_against_refimpl(
        "reciprocal",
        x,
        out,
        refimpl,
        strict_check=True,
    )


@pytest.mark.skip(reason="flaky")
@pytest.mark.parametrize("ctx", make_binary_params("remainder", dh.real_dtypes))
@given(data=st.data())
def test_remainder(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)
    if ctx.right_is_scalar:
        assume(right != 0)
    else:
        assume(not xp.any(right == 0))

    res = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, res)
    binary_param_assert_shape(ctx, left, right, res)
    binary_param_assert_against_refimpl(ctx, left, right, res, "%", operator.mod)


@given(hh.arrays(dtype=hh.numeric_dtypes, shape=hh.shapes()))
def test_round(x):
    out = xp.round(x)
    ph.assert_dtype("round", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("round", out_shape=out.shape, expected=x.shape)
    if x.dtype in dh.complex_dtypes:
        refimpl = lambda z: complex(round(z.real), round(z.imag))
    else:
        refimpl = round
    unary_assert_against_refimpl("round", x, out, refimpl, strict_check=True)


@pytest.mark.min_version("2023.12")
@given(hh.arrays(dtype=hh.real_floating_dtypes, shape=hh.shapes()))
def test_signbit(x):
    out = xp.signbit(x)
    ph.assert_dtype("signbit", in_dtype=x.dtype, out_dtype=out.dtype, expected=xp.bool)
    ph.assert_shape("signbit", out_shape=out.shape, expected=x.shape)
    refimpl = lambda x: math.copysign(1.0, x) < 0
    unary_assert_against_refimpl("round", x, out, refimpl, strict_check=True)


@given(hh.arrays(dtype=hh.numeric_dtypes, shape=hh.shapes(), elements=finite_kw))
def test_sign(x):
    out = xp.sign(x)
    ph.assert_dtype("sign", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("sign", out_shape=out.shape, expected=x.shape)
    refimpl = lambda x: x / abs(x) if x != 0 else 0
    unary_assert_against_refimpl(
        "sign",
        x,
        out,
        refimpl,
        strict_check=True,
    )


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_sin(x):
    out = xp.sin(x)
    ph.assert_dtype("sin", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("sin", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.sin if x.dtype in dh.complex_dtypes else math.sin
    unary_assert_against_refimpl("sin", x, out, refimpl)


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_sinh(x):
    out = xp.sinh(x)
    ph.assert_dtype("sinh", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("sinh", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.sinh if x.dtype in dh.complex_dtypes else math.sinh
    unary_assert_against_refimpl("sinh", x, out, refimpl)


@given(hh.arrays(dtype=hh.numeric_dtypes, shape=hh.shapes()))
def test_square(x):
    out = xp.square(x)
    ph.assert_dtype("square", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("square", out_shape=out.shape, expected=x.shape)
    unary_assert_against_refimpl(
        "square", x, out, lambda s: s*s, expr_template="{}²={}"
    )


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_sqrt(x):
    out = xp.sqrt(x)
    ph.assert_dtype("sqrt", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("sqrt", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.sqrt if x.dtype in dh.complex_dtypes else math.sqrt
    filter_ = default_filter if x.dtype in dh.complex_dtypes else lambda s: default_filter(s) and s >= 0
    unary_assert_against_refimpl(
        "sqrt", x, out, refimpl, filter_=filter_
    )


@pytest.mark.parametrize("ctx", make_binary_params("subtract", dh.numeric_dtypes))
@given(data=st.data())
def test_subtract(ctx, data):
    left = data.draw(ctx.left_strat, label=ctx.left_sym)
    right = data.draw(ctx.right_strat, label=ctx.right_sym)

    with hh.reject_overflow():
        res = ctx.func(left, right)

    binary_param_assert_dtype(ctx, left, right, res)
    binary_param_assert_shape(ctx, left, right, res)
    binary_param_assert_against_refimpl(ctx, left, right, res, "-", operator.sub)


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_tan(x):
    out = xp.tan(x)
    ph.assert_dtype("tan", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("tan", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.tan if x.dtype in dh.complex_dtypes else math.tan
    unary_assert_against_refimpl("tan", x, out, refimpl)


@given(hh.arrays(dtype=hh.all_floating_dtypes(), shape=hh.shapes()))
def test_tanh(x):
    out = xp.tanh(x)
    ph.assert_dtype("tanh", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("tanh", out_shape=out.shape, expected=x.shape)
    refimpl = cmath.tanh if x.dtype in dh.complex_dtypes else math.tanh
    unary_assert_against_refimpl("tanh", x, out, refimpl)


@given(hh.arrays(dtype=hh.real_dtypes, shape=xps.array_shapes()))
def test_trunc(x):
    out = xp.trunc(x)
    ph.assert_dtype("trunc", in_dtype=x.dtype, out_dtype=out.dtype)
    ph.assert_shape("trunc", out_shape=out.shape, expected=x.shape)
    unary_assert_against_refimpl("trunc", x, out, math.trunc, strict_check=True)


def _check_binary_with_scalars(func_data, x1x2):
    x1, x2 = x1x2
    func_name, refimpl, kwds, expected_dtype = func_data
    func = getattr(xp, func_name)
    out = func(x1, x2)
    in_dtypes, in_shapes, (x1a, x2a) = _convert_scalars_helper(x1, x2)
    _assert_correctness_binary(
        func_name, refimpl, in_dtypes, in_shapes, (x1a, x2a), out, expected_dtype, **kwds
    )


def _filter_zero(x):
    return x != 0 if dh.is_scalar(x) else (not xp.any(x == 0))


@pytest.mark.min_version("2024.12")
@pytest.mark.parametrize('func_data',
    # func_name, refimpl, kwargs, expected_dtype
    [
        ("add", operator.add, {}, None),
        ("atan2", math.atan2, {}, None),
        ("copysign", math.copysign, {}, None),
        ("divide", operator.truediv, {"filter_": lambda s: s != 0}, None),
        ("hypot", math.hypot, {}, None),
        ("logaddexp", logaddexp_refimpl, {}, None),
        ("maximum", max, {'strict_check': True}, None),
        ("minimum", min, {'strict_check': True}, None),
        ("multiply", operator.mul, {}, None),
        ("subtract", operator.sub, {}, None),

        ("equal", operator.eq, {}, xp.bool),
        ("not_equal", operator.ne, {}, xp.bool),
        ("less", operator.lt, {}, xp.bool),
        ("less_equal", operator.le, {}, xp.bool),
        ("greater", operator.gt, {}, xp.bool),
        ("greater_equal", operator.ge, {}, xp.bool),
    ],
    ids=lambda func_data: func_data[0]  # use names for test IDs
)
@given(x1x2=hh.array_and_py_scalar(dh.real_float_dtypes))
def test_binary_with_scalars_real(func_data, x1x2):
    _check_binary_with_scalars(func_data, x1x2)


@pytest.mark.min_version("2024.12")
@pytest.mark.parametrize('func_data',
    # func_name, refimpl, kwargs, expected_dtype
    [
        ("logical_and", operator.and_, {"expr_template": "({} or {})={}"}, None),
        ("logical_or", operator.or_, {"expr_template": "({} or {})={}"}, None),
        ("logical_xor", operator.xor, {"expr_template": "({} or {})={}"}, None),
    ],
    ids=lambda func_data: func_data[0]  # use names for test IDs
)
@given(x1x2=hh.array_and_py_scalar([xp.bool]))
def test_binary_with_scalars_bool(func_data, x1x2):
    _check_binary_with_scalars(func_data, x1x2)


@pytest.mark.min_version("2024.12")
@pytest.mark.parametrize('func_data',
    # func_name, refimpl, kwargs, expected_dtype
    [
        ("floor_divide", operator.floordiv, {}, None),
        ("remainder", operator.mod, {}, None),
    ],
    ids=lambda func_data: func_data[0]  # use names for test IDs
)
@given(x1x2=hh.array_and_py_scalar([xp.int64]))
def test_binary_with_scalars_int(func_data, x1x2):
    assume(_filter_zero(x1x2[1]))
    assume(_filter_zero(x1x2[0]) and _filter_zero(x1x2[1]))
    _check_binary_with_scalars(func_data, x1x2)


@pytest.mark.min_version("2024.12")
@pytest.mark.parametrize('func_data',
    # func_name, refimpl, kwargs, expected_dtype
    [
        ("bitwise_and", operator.and_, {}, None),
        ("bitwise_or", operator.or_, {}, None),
        ("bitwise_xor", operator.xor, {}, None),
    ],
    ids=lambda func_data: func_data[0]  # use names for test IDs
)
@given(x1x2=hh.array_and_py_scalar([xp.int32]))
def test_binary_with_scalars_bitwise(func_data, x1x2):
    func_name, refimpl, kwargs, expected = func_data
    # repack the refimpl
    refimpl_ = lambda l, r: mock_int_dtype(refimpl(l, r), xp.int32 )
    _check_binary_with_scalars((func_name, refimpl_, kwargs, expected), x1x2)