Implement NA.__array_ufunc__ (#30245)

doc/source/getting_started/dsintro.rst

@@ -676,11 +676,11 @@ similar to an ndarray:
    # only show the first 5 rows
    df[:5].T
 
+.. _dsintro.numpy_interop:
+
 DataFrame interoperability with NumPy functions
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
-.. _dsintro.numpy_interop:
-
 Elementwise NumPy ufuncs (log, exp, sqrt, ...) and various other NumPy functions
 can be used with no issues on Series and DataFrame, assuming the data within
 are numeric:

doc/source/user_guide/missing_data.rst

@@ -920,3 +920,29 @@ filling missing values beforehand.
 
 A similar situation occurs when using Series or DataFrame objects in ``if``
 statements, see :ref:`gotchas.truth`.
+
+NumPy ufuncs
+------------
+
+:attr:`pandas.NA` implements NumPy's ``__array_ufunc__`` protocol. Most ufuncs
+work with ``NA``, and generally return ``NA``:
+
+.. ipython:: python
+
+   np.log(pd.NA)
+   np.add(pd.NA, 1)
+
+.. warning::
+
+   Currently, ufuncs involving an ndarray and ``NA`` will return an
+   object-dtype filled with NA values.
+
+   .. ipython:: python
+
+      a = np.array([1, 2, 3])
+      np.greater(a, pd.NA)
+
+   The return type here may change to return a different array type
+   in the future.
+
+See :ref:`dsintro.numpy_interop` for more on ufuncs.

pandas/_libs/missing.pyx

@@ -14,6 +14,7 @@ from pandas._libs.tslibs.np_datetime cimport (
     get_timedelta64_value, get_datetime64_value)
 from pandas._libs.tslibs.nattype cimport (
     checknull_with_nat, c_NaT as NaT, is_null_datetimelike)
+from pandas._libs.ops_dispatch import maybe_dispatch_ufunc_to_dunder_op
 
 from pandas.compat import is_platform_32bit
 
@@ -290,16 +291,29 @@ cdef inline bint is_null_period(v):
 # Implementation of NA singleton
 
 
-def _create_binary_propagating_op(name, divmod=False):
+def _create_binary_propagating_op(name, is_divmod=False):
 
     def method(self, other):
         if (other is C_NA or isinstance(other, str)
-                or isinstance(other, (numbers.Number, np.bool_))):
-            if divmod:
+                or isinstance(other, (numbers.Number, np.bool_))
+                or isinstance(other, np.ndarray) and not other.shape):
+            # Need the other.shape clause to handle NumPy scalars,
+            # since we do a setitem on `out` below, which
+            # won't work for NumPy scalars.
+            if is_divmod:
                 return NA, NA
             else:
                 return NA
 
+        elif isinstance(other, np.ndarray):
+            out = np.empty(other.shape, dtype=object)
+            out[:] = NA
+
+            if is_divmod:
+                return out, out.copy()
+            else:
+                return out
+
         return NotImplemented
 
     method.__name__ = name
@@ -369,8 +383,8 @@ class NAType(C_NAType):
     __rfloordiv__ = _create_binary_propagating_op("__rfloordiv__")
     __mod__ = _create_binary_propagating_op("__mod__")
     __rmod__ = _create_binary_propagating_op("__rmod__")
-    __divmod__ = _create_binary_propagating_op("__divmod__", divmod=True)
-    __rdivmod__ = _create_binary_propagating_op("__rdivmod__", divmod=True)
+    __divmod__ = _create_binary_propagating_op("__divmod__", is_divmod=True)
+    __rdivmod__ = _create_binary_propagating_op("__rdivmod__", is_divmod=True)
     # __lshift__ and __rshift__ are not implemented
 
     __eq__ = _create_binary_propagating_op("__eq__")
@@ -397,6 +411,8 @@ class NAType(C_NAType):
                 return type(other)(1)
             else:
                 return NA
+        elif isinstance(other, np.ndarray):
+            return np.where(other == 0, other.dtype.type(1), NA)
 
         return NotImplemented
 
@@ -408,6 +424,8 @@ class NAType(C_NAType):
                 return other
             else:
                 return NA
+        elif isinstance(other, np.ndarray):
+            return np.where((other == 1) | (other == -1), other, NA)
 
         return NotImplemented
 
@@ -440,6 +458,31 @@ class NAType(C_NAType):
 
     __rxor__ = __xor__
 
+    __array_priority__ = 1000
+    _HANDLED_TYPES = (np.ndarray, numbers.Number, str, np.bool_)
+
+    def __array_ufunc__(self, ufunc, method, *inputs, **kwargs):
+        types = self._HANDLED_TYPES + (NAType,)
+        for x in inputs:
+            if not isinstance(x, types):
+                return NotImplemented
+
+        if method != "__call__":
+            raise ValueError(f"ufunc method '{method}' not supported for NA")
+        result = maybe_dispatch_ufunc_to_dunder_op(
+            self, ufunc, method, *inputs, **kwargs
+        )
+        if result is NotImplemented:
+            # For a NumPy ufunc that's not a binop, like np.logaddexp
+            index = [i for i, x in enumerate(inputs) if x is NA][0]
+            result = np.broadcast_arrays(*inputs)[index]
+            if result.ndim == 0:
+                result = result.item()
+            if ufunc.nout > 1:
+                result = (NA,) * ufunc.nout
+
+        return result
+
 
 C_NA = NAType()   # C-visible
 NA = C_NA         # Python-visible

pandas/_libs/ops_dispatch.pyx

@@ -0,0 +1,94 @@
+DISPATCHED_UFUNCS = {
+    "add",
+    "sub",
+    "mul",
+    "pow",
+    "mod",
+    "floordiv",
+    "truediv",
+    "divmod",
+    "eq",
+    "ne",
+    "lt",
+    "gt",
+    "le",
+    "ge",
+    "remainder",
+    "matmul",
+    "or",
+    "xor",
+    "and",
+}
+UFUNC_ALIASES = {
+    "subtract": "sub",
+    "multiply": "mul",
+    "floor_divide": "floordiv",
+    "true_divide": "truediv",
+    "power": "pow",
+    "remainder": "mod",
+    "divide": "div",
+    "equal": "eq",
+    "not_equal": "ne",
+    "less": "lt",
+    "less_equal": "le",
+    "greater": "gt",
+    "greater_equal": "ge",
+    "bitwise_or": "or",
+    "bitwise_and": "and",
+    "bitwise_xor": "xor",
+}
+
+# For op(., Array) -> Array.__r{op}__
+REVERSED_NAMES = {
+    "lt": "__gt__",
+    "le": "__ge__",
+    "gt": "__lt__",
+    "ge": "__le__",
+    "eq": "__eq__",
+    "ne": "__ne__",
+}
+
+
+def maybe_dispatch_ufunc_to_dunder_op(
+    object self, object ufunc, str method, *inputs, **kwargs
+):
+    """
+    Dispatch a ufunc to the equivalent dunder method.
+
+    Parameters
+    ----------
+    self : ArrayLike
+        The array whose dunder method we dispatch to
+    ufunc : Callable
+        A NumPy ufunc
+    method : {'reduce', 'accumulate', 'reduceat', 'outer', 'at', '__call__'}
+    inputs : ArrayLike
+        The input arrays.
+    kwargs : Any
+        The additional keyword arguments, e.g. ``out``.
+
+    Returns
+    -------
+    result : Any
+        The result of applying the ufunc
+    """
+    # special has the ufuncs we dispatch to the dunder op on
+
+    op_name = ufunc.__name__
+    op_name = UFUNC_ALIASES.get(op_name, op_name)
+
+    def not_implemented(*args, **kwargs):
+        return NotImplemented
+
+    if (method == "__call__"
+            and op_name in DISPATCHED_UFUNCS
+            and kwargs.get("out") is None):
+        if isinstance(inputs[0], type(self)):
+            name = f"__{op_name}__"
+            return getattr(self, name, not_implemented)(inputs[1])
+        else:
+            name = REVERSED_NAMES.get(op_name, f"__r{op_name}__")
+            result = getattr(self, name, not_implemented)(inputs[0])
+            return result
+    else:
+        return NotImplemented

pandas/core/ops/__init__.py

@@ -10,6 +10,7 @@
 import numpy as np
 
 from pandas._libs import Timedelta, Timestamp, lib
+from pandas._libs.ops_dispatch import maybe_dispatch_ufunc_to_dunder_op  # noqa:F401
 from pandas.util._decorators import Appender
 
 from pandas.core.dtypes.common import is_list_like, is_timedelta64_dtype
@@ -31,7 +32,6 @@
 )
 from pandas.core.ops.array_ops import comp_method_OBJECT_ARRAY  # noqa:F401
 from pandas.core.ops.common import unpack_zerodim_and_defer
-from pandas.core.ops.dispatch import maybe_dispatch_ufunc_to_dunder_op  # noqa:F401
 from pandas.core.ops.dispatch import should_series_dispatch
 from pandas.core.ops.docstrings import (
     _arith_doc_FRAME,

pandas/core/ops/dispatch.py

@@ -1,12 +1,10 @@
 """
 Functions for defining unary operations.
 """
-from typing import Any, Callable, Union
+from typing import Any, Union
 
 import numpy as np
 
-from pandas._typing import ArrayLike
-
 from pandas.core.dtypes.common import (
     is_datetime64_dtype,
     is_extension_array_dtype,
@@ -126,94 +124,3 @@ def dispatch_to_extension_op(
     # on the ExtensionArray
     res_values = op(left, right)
     return res_values
-
-
-def maybe_dispatch_ufunc_to_dunder_op(
-    self: ArrayLike, ufunc: Callable, method: str, *inputs: ArrayLike, **kwargs: Any
-):
-    """
-    Dispatch a ufunc to the equivalent dunder method.
-
-    Parameters
-    ----------
-    self : ArrayLike
-        The array whose dunder method we dispatch to
-    ufunc : Callable
-        A NumPy ufunc
-    method : {'reduce', 'accumulate', 'reduceat', 'outer', 'at', '__call__'}
-    inputs : ArrayLike
-        The input arrays.
-    kwargs : Any
-        The additional keyword arguments, e.g. ``out``.
-
-    Returns
-    -------
-    result : Any
-        The result of applying the ufunc
-    """
-    # special has the ufuncs we dispatch to the dunder op on
-    special = {
-        "add",
-        "sub",
-        "mul",
-        "pow",
-        "mod",
-        "floordiv",
-        "truediv",
-        "divmod",
-        "eq",
-        "ne",
-        "lt",
-        "gt",
-        "le",
-        "ge",
-        "remainder",
-        "matmul",
-        "or",
-        "xor",
-        "and",
-    }
-    aliases = {
-        "subtract": "sub",
-        "multiply": "mul",
-        "floor_divide": "floordiv",
-        "true_divide": "truediv",
-        "power": "pow",
-        "remainder": "mod",
-        "divide": "div",
-        "equal": "eq",
-        "not_equal": "ne",
-        "less": "lt",
-        "less_equal": "le",
-        "greater": "gt",
-        "greater_equal": "ge",
-        "bitwise_or": "or",
-        "bitwise_and": "and",
-        "bitwise_xor": "xor",
-    }
-
-    # For op(., Array) -> Array.__r{op}__
-    flipped = {
-        "lt": "__gt__",
-        "le": "__ge__",
-        "gt": "__lt__",
-        "ge": "__le__",
-        "eq": "__eq__",
-        "ne": "__ne__",
-    }
-
-    op_name = ufunc.__name__
-    op_name = aliases.get(op_name, op_name)
-
-    def not_implemented(*args, **kwargs):
-        return NotImplemented
-
-    if method == "__call__" and op_name in special and kwargs.get("out") is None:
-        if isinstance(inputs[0], type(self)):
-            name = f"__{op_name}__"
-            return getattr(self, name, not_implemented)(inputs[1])
-        else:
-            name = flipped.get(op_name, f"__r{op_name}__")
-            return getattr(self, name, not_implemented)(inputs[0])
-    else:
-        return NotImplemented