ENH: delegate broadcast_shapes

Author: Cyril-36

src/array_api_extra/__init__.py

@@ -3,6 +3,7 @@
 from ._delegation import (
     argpartition,
     atleast_nd,
+    broadcast_shapes,
     cov,
     create_diagonal,
     expand_dims,
@@ -20,7 +21,6 @@
 from ._lib._at import at
 from ._lib._funcs import (
     apply_where,
-    broadcast_shapes,
     default_dtype,
     kron,
     nunique,

src/array_api_extra/_delegation.py

@@ -2,7 +2,7 @@
 
 from collections.abc import Sequence
 from types import ModuleType
-from typing import Literal
+from typing import Literal, cast
 
 from ._lib import _funcs
 from ._lib._utils._compat import (
@@ -20,6 +20,7 @@
 
 __all__ = [
     "atleast_nd",
+    "broadcast_shapes",
     "cov",
     "create_diagonal",
     "expand_dims",
@@ -81,6 +82,64 @@ def atleast_nd(x: Array, /, *, ndim: int, xp: ModuleType | None = None) -> Array
     return _funcs.atleast_nd(x, ndim=ndim, xp=xp)
 
 
+def _numpy_broadcast_shapes(*shapes: tuple[int, ...]) -> tuple[int, ...] | None:
+    try:
+        import numpy as np
+    except ImportError:
+        return None
+
+    return np.broadcast_shapes(*shapes)
+
+
+def broadcast_shapes(*shapes: tuple[float | None, ...]) -> tuple[int | None, ...]:
+    """
+    Compute the shape of the broadcasted arrays.
+
+    Duplicates :func:`numpy.broadcast_shapes`, with additional support for
+    None and NaN sizes.
+
+    This is equivalent to ``xp.broadcast_arrays(arr1, arr2, ...)[0].shape``
+    without needing to worry about the backend potentially deep copying
+    the arrays.
+
+    Parameters
+    ----------
+    *shapes : tuple[int | None, ...]
+        Shapes of the arrays to broadcast.
+
+    Returns
+    -------
+    tuple[int | None, ...]
+        The shape of the broadcasted arrays.
+
+    See Also
+    --------
+    numpy.broadcast_shapes : Equivalent NumPy function.
+    array_api.broadcast_arrays : Function to broadcast actual arrays.
+
+    Notes
+    -----
+    This function accepts the Array API's ``None`` for unknown sizes,
+    as well as Dask's non-standard ``math.nan``.
+    Regardless of input, the output always contains ``None`` for unknown sizes.
+
+    Examples
+    --------
+    >>> import array_api_extra as xpx
+    >>> xpx.broadcast_shapes((2, 3), (2, 1))
+    (2, 3)
+    >>> xpx.broadcast_shapes((4, 2, 3), (2, 1), (1, 3))
+    (4, 2, 3)
+    """
+    if all(isinstance(size, int) for shape in shapes for size in shape):
+        int_shapes = cast(tuple[tuple[int, ...], ...], shapes)
+        out = _numpy_broadcast_shapes(*int_shapes)
+        if out is not None:
+            return cast(tuple[int | None, ...], out)
+
+    return _funcs.broadcast_shapes(*shapes)
+
+
 def cov(m: Array, /, *, xp: ModuleType | None = None) -> Array:
     """
     Estimate a covariance matrix (or a stack of covariance matrices).

src/array_api_extra/_lib/_funcs.py

@@ -220,46 +220,10 @@ def atleast_nd(x: Array, /, *, ndim: int, xp: ModuleType) -> Array:
 
 # `float` in signature to accept `math.nan` for Dask.
 # `int`s are still accepted as `float` is a superclass of `int` in typing
-def broadcast_shapes(*shapes: tuple[float | None, ...]) -> tuple[int | None, ...]:
-    """
-    Compute the shape of the broadcasted arrays.
-
-    Duplicates :func:`numpy.broadcast_shapes`, with additional support for
-    None and NaN sizes.
-
-    This is equivalent to ``xp.broadcast_arrays(arr1, arr2, ...)[0].shape``
-    without needing to worry about the backend potentially deep copying
-    the arrays.
-
-    Parameters
-    ----------
-    *shapes : tuple[int | None, ...]
-        Shapes of the arrays to broadcast.
-
-    Returns
-    -------
-    tuple[int | None, ...]
-        The shape of the broadcasted arrays.
-
-    See Also
-    --------
-    numpy.broadcast_shapes : Equivalent NumPy function.
-    array_api.broadcast_arrays : Function to broadcast actual arrays.
-
-    Notes
-    -----
-    This function accepts the Array API's ``None`` for unknown sizes,
-    as well as Dask's non-standard ``math.nan``.
-    Regardless of input, the output always contains ``None`` for unknown sizes.
-
-    Examples
-    --------
-    >>> import array_api_extra as xpx
-    >>> xpx.broadcast_shapes((2, 3), (2, 1))
-    (2, 3)
-    >>> xpx.broadcast_shapes((4, 2, 3), (2, 1), (1, 3))
-    (4, 2, 3)
-    """
+def broadcast_shapes(  # numpydoc ignore=PR01,RT01
+    *shapes: tuple[float | None, ...],
+) -> tuple[int | None, ...]:
+    """See docstring in array_api_extra._delegation."""
     if not shapes:
         return ()  # Match NumPy output
 

tests/test_funcs.py

@@ -12,6 +12,7 @@
 from hypothesis import strategies as st
 from typing_extensions import override
 
+import array_api_extra._delegation as delegation
 from array_api_extra import (
     apply_where,
     argpartition,
@@ -489,6 +490,42 @@ def test_5D_values(self, xp: ModuleType):
 
 
 class TestBroadcastShapes:
+    def test_delegates_known_integer_shapes(self, monkeypatch: pytest.MonkeyPatch):
+        calls = []
+
+        def mock_broadcast_shapes(*shapes: tuple[int, ...]) -> tuple[int, ...]:
+            calls.append(shapes)
+            return (99,)
+
+        monkeypatch.setattr(
+            delegation, "_numpy_broadcast_shapes", mock_broadcast_shapes
+        )
+
+        assert broadcast_shapes((2,), (1,)) == (99,)
+        assert calls == [((2,), (1,))]
+
+    def test_fallback_for_unknown_sizes(self, monkeypatch: pytest.MonkeyPatch):
+        def mock_broadcast_shapes(*_shapes: tuple[int, ...]) -> tuple[int, ...]:
+            msg = "NumPy delegation should not handle unknown sizes"
+            raise AssertionError(msg)
+
+        monkeypatch.setattr(
+            delegation, "_numpy_broadcast_shapes", mock_broadcast_shapes
+        )
+
+        assert broadcast_shapes((None,), (1,)) == (None,)
+        assert broadcast_shapes((math.nan,), (1,)) == (None,)
+
+    def test_fallback_without_numpy(self, monkeypatch: pytest.MonkeyPatch):
+        def mock_broadcast_shapes(*_shapes: tuple[int, ...]) -> tuple[int, ...] | None:
+            return None
+
+        monkeypatch.setattr(
+            delegation, "_numpy_broadcast_shapes", mock_broadcast_shapes
+        )
+
+        assert broadcast_shapes((2,), (1,)) == (2,)
+
     @pytest.mark.parametrize(
         "args",
         [