Abstractions for read-only arrays

Author: crusaderky

array_api_compat/__init__.py

@@ -1,9 +1,9 @@
 """
 NumPy Array API compatibility library
 
-This is a small wrapper around NumPy and CuPy that is compatible with the
-Array API standard https://data-apis.org/array-api/latest/. See also NEP 47
-https://numpy.org/neps/nep-0047-array-api-standard.html.
+This is a small wrapper around NumPy, CuPy, JAX and others that is compatible
+with the Array API standard https://data-apis.org/array-api/latest/.
+See also NEP 47 https://numpy.org/neps/nep-0047-array-api-standard.html.
 
 Unlike array_api_strict, this is not a strict minimal implementation of the
 Array API, but rather just an extension of the main NumPy namespace with

array_api_compat/common/_helpers.py

@@ -7,10 +7,11 @@
 """
 from __future__ import annotations
 
+import operator
 from typing import TYPE_CHECKING
 
 if TYPE_CHECKING:
-    from typing import Optional, Union, Any
+    from typing import Callable, Literal, Optional, Union, Any
     from ._typing import Array, Device
 
 import sys
@@ -91,7 +92,7 @@ def is_cupy_array(x):
     import cupy as cp
 
     # TODO: Should we reject ndarray subclasses?
-    return isinstance(x, (cp.ndarray, cp.generic))
+    return isinstance(x, cp.ndarray)
 
 def is_torch_array(x):
     """
@@ -787,6 +788,7 @@ def to_device(x: Array, device: Device, /, *, stream: Optional[Union[int, Any]]
         return x
     return x.to_device(device, stream=stream)
 
+
 def size(x):
     """
     Return the total number of elements of x.
@@ -801,6 +803,261 @@ def size(x):
         return None
     return math.prod(x.shape)
 
+
+def is_writeable_array(x) -> bool:
+    """
+    Return False if x.__setitem__ is expected to raise; True otherwise
+    """
+    if is_numpy_array(x):
+        return x.flags.writeable
+    if is_jax_array(x) or is_pydata_sparse_array(x):
+        return False
+    return True
+
+
+def _is_fancy_index(idx) -> bool:
+    if not isinstance(idx, tuple):
+        idx = (idx,)
+    return any(
+        isinstance(i, (list, tuple)) or is_array_api_obj(i)
+        for i in idx
+    )
+
+
+_undef = object()
+
+
+class at:
+    """
+    Update operations for read-only arrays.
+
+    This implements ``jax.numpy.ndarray.at`` for all backends.
+    Writeable arrays may be updated in place; you should not rely on it.
+
+    Keyword arguments (e.g. ``indices_are_sorted``) are passed to JAX and are
+    quietly ignored for backends that don't support them.
+
+    Additionally, this introduces support for the `copy` keyword for all backends:
+
+    None
+        x *may* be modified in place if it is possible and beneficial
+        for performance. You should not use x after calling this function.
+    True
+        Ensure that the inputs are not modified. This is the default.
+    False
+        Raise ValueError if a copy cannot be avoided.
+
+    Examples
+    --------
+    Given either of these equivalent expressions::
+
+      x = at(x)[1].add(2, copy=None)
+      x = at(x, 1).add(2, copy=None)
+
+    If x is a JAX array, they are the same as::
+
+      x = x.at[1].add(2)
+
+    If x is a read-only numpy array, they are the same as::
+
+      x = x.copy()
+      x[1] += 2
+
+    Otherwise, they are the same as::
+
+      x[1] += 2
+
+    Warning
+    -------
+    When you use copy=None, you should always immediately overwrite
+    the parameter array::
+
+        x = at(x, 0).set(2, copy=None)
+
+    The anti-pattern below must be avoided, as it will result in different behaviour
+    on read-only versus writeable arrays:
+
+        x = xp.asarray([0, 0, 0])
+        y = at(x, 0).set(2, copy=None)
+        z = at(x, 1).set(3, copy=None)
+
+    In the above example, y == [2, 0, 0] and z == [0, 3, 0] when x is read-only,
+    whereas y == z == [2, 3, 0] when x is writeable!
+
+    Caveat
+    ------
+    The behaviour of methods other than `get()` when the index is an array of
+    integers which contains multiple occurrences of the same index is undefined.
+
+    **Undefined behaviour:** ``at(x, [0, 0]).set(2)``
+
+    See Also
+    --------
+    https://jax.readthedocs.io/en/latest/_autosummary/jax.numpy.ndarray.at.html
+    """
+
+    __slots__ = ("x", "idx")
+
+    def __init__(self, x, idx=_undef):
+        self.x = x
+        self.idx = idx
+
+    def __getitem__(self, idx):
+        """
+        Allow for the alternate syntax ``at(x)[start:stop:step]``,
+        which looks prettier than ``at(x, slice(start, stop, step))``
+        and feels more intuitive coming from the JAX documentation.
+        """
+        if self.idx is not _undef:
+            raise ValueError("Index has already been set")
+        self.idx = idx
+        return self
+
+    def _common(
+        self,
+        at_op: str,
+        y=_undef,
+        copy: bool | None | Literal["_force_false"] = True,
+        **kwargs,
+    ):
+        """Perform common prepocessing.
+
+        Returns
+        -------
+        If the operation can be resolved by at[], (return value, None)
+        Otherwise, (None, preprocessed x)
+        """
+        if self.idx is _undef:
+            raise TypeError(
+                "Index has not been set.\n"
+                "Usage: either\n"
+                "    at(x, idx).set(value)\n"
+                "or\n"
+                "    at(x)[idx].set(value)\n"
+                "(same for all other methods)."
+            )
+        
+        x = self.x
+
+        if copy is False:
+            if not is_writeable_array(x) or is_dask_array(x):
+                raise ValueError("Cannot modify parameter in place")
+        elif copy is None:
+            copy = not is_writeable_array(x)
+        elif copy == "_force_false":
+            copy = False
+        elif copy is not True:
+            raise ValueError(f"Invalid value for copy: {copy!r}")
+
+        if is_jax_array(x):
+            # Use JAX's at[]
+            at_ = x.at[self.idx]
+            args = (y,) if y is not _undef else ()
+            return getattr(at_, at_op)(*args, **kwargs), None
+
+        # Emulate at[] behaviour for non-JAX arrays
+        if copy:
+            # FIXME We blindly expect the output of x.copy() to be always writeable.
+            # This holds true for read-only numpy arrays, but not necessarily for
+            # other backends.
+            xp = get_namespace(x)
+            x = xp.asarray(x, copy=True)
+
+        return None, x
+
+    def get(self, copy: bool | None = True, **kwargs):
+        """
+        Return x[idx]. In addition to plain __getitem__, this allows ensuring
+        that the output is (not) a copy and kwargs are passed to the backend.
+        """
+        # __getitem__ with a fancy index always returns a copy.
+        # Avoid an unnecessary double copy.
+        # If copy is forced to False, raise.
+        if _is_fancy_index(self.idx):
+            if copy is False:
+                raise ValueError(
+                    "Indexing a numpy array with a fancy index always "
+                    "results in a copy"
+                )
+            # Skip copy inside _common, even if array is not writeable
+            copy = "_force_false"  # type: ignore
+
+        res, x = self._common("get", copy=copy, **kwargs)
+        if res is not None:
+            return res
+        return x[self.idx]
+
+    def set(self, y, /, **kwargs):
+        """x[idx] = y"""
+        res, x = self._common("set", y, **kwargs)
+        if res is not None:
+            return res
+        x[self.idx] = y
+        return x
+
+    def apply(self, ufunc, /, **kwargs):
+        """ufunc.at(x, idx)"""
+        if is_cupy_array(self.x) or is_torch_array(self.x) or is_dask_array(self.x):
+            # ufunc.at not implemented
+            return self.set(ufunc(self.x[self.idx]), **kwargs)
+
+        res, x = self._common("apply", ufunc, **kwargs)
+        if res is not None:
+            return res
+        ufunc.at(x, self.idx)
+        return x
+
+    def _iop(
+        self, at_op: str, elwise_op: Callable[[Array, Array], Array], y: Array, **kwargs
+    ):
+        """x[idx] += y or equivalent in-place operation on a subset of x
+
+        which is the same as saying
+            x[idx] = x[idx] + y
+        Note that this is not the same as
+            operator.iadd(x[idx], y)
+        Consider for example when x is a numpy array and idx is a fancy index, which
+        triggers a deep copy on __getitem__.
+        """
+        res, x = self._common(at_op, y, **kwargs)
+        if res is not None:
+            return res
+        x[self.idx] = elwise_op(x[self.idx], y)
+        return x
+
+    def add(self, y, /, **kwargs):
+        """x[idx] += y"""
+        return self._iop("add", operator.add, y, **kwargs)
+
+    def subtract(self, y, /, **kwargs):
+        """x[idx] -= y"""
+        return self._iop("subtract", operator.sub, y, **kwargs)
+
+    def multiply(self, y, /, **kwargs):
+        """x[idx] *= y"""
+        return self._iop("multiply", operator.mul, y, **kwargs)
+
+    def divide(self, y, /, **kwargs):
+        """x[idx] /= y"""
+        return self._iop("divide", operator.truediv, y, **kwargs)
+
+    def power(self, y, /, **kwargs):
+        """x[idx] **= y"""
+        return self._iop("power", operator.pow, y, **kwargs)
+
+    def min(self, y, /, **kwargs):
+        """x[idx] = minimum(x[idx], y)"""
+        import numpy as np
+
+        return self._iop("min", np.minimum, y, **kwargs)
+
+    def max(self, y, /, **kwargs):
+        """x[idx] = maximum(x[idx], y)"""
+        import numpy as np
+
+        return self._iop("max", np.maximum, y, **kwargs)
+
+
 __all__ = [
     "array_namespace",
     "device",
@@ -821,8 +1078,10 @@ def size(x):
     "is_ndonnx_namespace",
     "is_pydata_sparse_array",
     "is_pydata_sparse_namespace",
+    "is_writeable_array",
     "size",
     "to_device",
+    "at",
 ]
 
-_all_ignore = ['sys', 'math', 'inspect', 'warnings']
+_all_ignore = ['inspect', 'math', 'operator', 'warnings', 'sys']

docs/helper-functions.rst

@@ -36,6 +36,7 @@ instead, which would be wrapped.
 .. autofunction:: device
 .. autofunction:: to_device
 .. autofunction:: size
+.. autofunction:: at
 
 Inspection Helpers
 ------------------
@@ -51,6 +52,7 @@ yet.
 .. autofunction:: is_jax_array
 .. autofunction:: is_pydata_sparse_array
 .. autofunction:: is_ndonnx_array
+.. autofunction:: is_writeable_array
 .. autofunction:: is_numpy_namespace
 .. autofunction:: is_cupy_namespace
 .. autofunction:: is_torch_namespace