Support of alternative array classes: ndarray-like

docs/release.rst

@@ -10,11 +10,14 @@ Unreleased
   By :user:`Haiying Xu <halehawk>`, `John Kirkham <jakirkham>`, `Ryan Abernathey <rabernat>` :
   issue:`303`.
 
-.. _release_0.9.1:
+* Add support of alternative array classes (other than NumPy arrays)
+  By :user:`Mads R. B. Kristensen <madsbk>`, :issue:`305`.
 
 * Add ability to find codecs via entrypoints
   By :user:`Martin Durant <martindurant>`, :issue:`290`.
 
+.. _release_0.9.1:
+
 0.9.1
 -----
 

numcodecs/compat.py

@@ -1,103 +1,120 @@
 # flake8: noqa
+import functools
 import sys
 import codecs
 import array
 from functools import reduce
 
 import numpy as np
 
+from .ndarray_like import NDArrayLike, is_ndarray_like
 
-def ensure_ndarray(buf):
-    """Convenience function to coerce `buf` to a numpy array, if it is not already a
-    numpy array.
+
+def ensure_ndarray_like(buf) -> NDArrayLike:
+    """Convenience function to coerce `buf` to ndarray-like array.
 
     Parameters
     ----------
-    buf : array-like or bytes-like
-        A numpy array or any object exporting a buffer interface.
+    buf : ndarray-like, array-like, or bytes-like
+        A numpy array like object such as numpy.ndarray, cupy.ndarray, or
+        any object exporting a buffer interface.
 
     Returns
     -------
-    arr : ndarray
-        A numpy array, sharing memory with `buf`.
+    arr : NDArrayLike
+        A ndarray-like, sharing memory with `buf`.
 
     Notes
     -----
     This function will not create a copy under any circumstances, it is guaranteed to
     return a view on memory exported by `buf`.
-
     """
 
-    if isinstance(buf, np.ndarray):
-        # already a numpy array
-        arr = buf
+    if not is_ndarray_like(buf):
+        if isinstance(buf, array.array) and buf.typecode in "cu":
+            # Guard condition, do not support array.array with unicode type, this is
+            # problematic because numpy does not support it on all platforms. Also do not
+            # support char as it was removed in Python 3.
+            raise TypeError("array.array with char or unicode type is not supported")
+        else:
+            # N.B., first take a memoryview to make sure that we subsequently create a
+            # numpy array from a memory buffer with no copy
+            mem = memoryview(buf)
+            # instantiate array from memoryview, ensures no copy
+            buf = np.array(mem, copy=False)
+    return buf
 
-    elif isinstance(buf, array.array) and buf.typecode in 'cu':
-        # Guard condition, do not support array.array with unicode type, this is
-        # problematic because numpy does not support it on all platforms. Also do not
-        # support char as it was removed in Python 3.
-        raise TypeError('array.array with char or unicode type is not supported')
 
-    else:
+def ensure_ndarray(buf) -> np.ndarray:
+    """Convenience function to coerce `buf` to a numpy array, if it is not already a
+    numpy array.
 
-        # N.B., first take a memoryview to make sure that we subsequently create a
-        # numpy array from a memory buffer with no copy
-        mem = memoryview(buf)
+    Parameters
+    ----------
+    buf : array-like or bytes-like
+        A numpy array or any object exporting a buffer interface.
 
-        # instantiate array from memoryview, ensures no copy
-        arr = np.array(mem, copy=False)
+    Returns
+    -------
+    arr : ndarray
+        A numpy array, sharing memory with `buf`.
 
-    return arr
+    Notes
+    -----
+    This function will not create a copy under any circumstances, it is guaranteed to
+    return a view on memory exported by `buf`.
+    """
+    return np.array(ensure_ndarray_like(buf), copy=False)
 
 
-def ensure_contiguous_ndarray(buf, max_buffer_size=None, flatten=True):
-    """Convenience function to coerce `buf` to a numpy array, if it is not already a
-    numpy array. Also ensures that the returned value exports fully contiguous memory,
+def ensure_contiguous_ndarray_like(
+    buf, max_buffer_size=None, flatten=True
+) -> NDArrayLike:
+    """Convenience function to coerce `buf` to ndarray-like array.
+    Also ensures that the returned value exports fully contiguous memory,
     and supports the new-style buffer interface. If the optional max_buffer_size is
     provided, raise a ValueError if the number of bytes consumed by the returned
     array exceeds this value.
 
     Parameters
     ----------
-    buf : array-like or bytes-like
-        A numpy array or any object exporting a buffer interface.
+    buf : ndarray-like, array-like, or bytes-like
+        A numpy array like object such as numpy.ndarray, cupy.ndarray, or
+        any object exporting a buffer interface.
     max_buffer_size : int
         If specified, the largest allowable value of arr.nbytes, where arr
         is the returned array.
+    flatten : bool
+        If True, the array are flatten.
 
     Returns
     -------
-    arr : ndarray
-        A numpy array, sharing memory with `buf`.
+    arr : NDArrayLike
+        A ndarray-like, sharing memory with `buf`.
 
     Notes
     -----
     This function will not create a copy under any circumstances, it is guaranteed to
     return a view on memory exported by `buf`.
-
     """
-
-    # ensure input is a numpy array
-    arr = ensure_ndarray(buf)
+    arr = ensure_ndarray_like(buf)
 
     # check for object arrays, these are just memory pointers, actual memory holding
     # item data is scattered elsewhere
     if arr.dtype == object:
-        raise TypeError('object arrays are not supported')
+        raise TypeError("object arrays are not supported")
 
     # check for datetime or timedelta ndarray, the buffer interface doesn't support those
-    if arr.dtype.kind in 'Mm':
+    if arr.dtype.kind in "Mm":
         arr = arr.view(np.int64)
 
     # check memory is contiguous, if so flatten
     if arr.flags.c_contiguous or arr.flags.f_contiguous:
-        # check if flatten flag is on or not
         if flatten:
             # can flatten without copy
-            arr = arr.reshape(-1, order='A')
-
+            arr = arr.reshape(-1, order="A")
     else:
-        raise ValueError('an array with contiguous memory is required')
+        raise ValueError("an array with contiguous memory is required")
 
     if max_buffer_size is not None and arr.nbytes > max_buffer_size:
         msg = "Codec does not support buffers of > {} bytes".format(max_buffer_size)
@@ -106,45 +123,78 @@ def ensure_contiguous_ndarray(buf, max_buffer_size=None, flatten=True):
     return arr
 
 
-def ensure_bytes(buf):
+def ensure_contiguous_ndarray(buf, max_buffer_size=None, flatten=True) -> np.array:
+    """Convenience function to coerce `buf` to a numpy array, if it is not already a
+    numpy array. Also ensures that the returned value exports fully contiguous memory,
+    and supports the new-style buffer interface. If the optional max_buffer_size is
+    provided, raise a ValueError if the number of bytes consumed by the returned
+    array exceeds this value.
+
+    Parameters
+    ----------
+    buf : array-like or bytes-like
+        A numpy array or any object exporting a buffer interface.
+    max_buffer_size : int
+        If specified, the largest allowable value of arr.nbytes, where arr
+        is the returned array.
+    flatten : bool
+        If True, the array are flatten.
+
+    Returns
+    -------
+    arr : ndarray
+        A numpy array, sharing memory with `buf`.
+
+    Notes
+    -----
+    This function will not create a copy under any circumstances, it is guaranteed to
+    return a view on memory exported by `buf`.
+    """
+
+    return ensure_ndarray(
+        ensure_contiguous_ndarray_like(
+            buf, max_buffer_size=max_buffer_size, flatten=flatten
+        )
+    )
+
+
+def ensure_bytes(buf) -> bytes:
     """Obtain a bytes object from memory exposed by `buf`."""
 
     if not isinstance(buf, bytes):
-
-        # go via numpy, for convenience
-        arr = ensure_ndarray(buf)
+        arr = ensure_ndarray_like(buf)
 
         # check for object arrays, these are just memory pointers,
         # actual memory holding item data is scattered elsewhere
         if arr.dtype == object:
-            raise TypeError('object arrays are not supported')
+            raise TypeError("object arrays are not supported")
 
         # create bytes
-        buf = arr.tobytes(order='A')
+        buf = arr.tobytes(order="A")
 
     return buf
 
 
-def ensure_text(s, encoding='utf-8'):
+def ensure_text(s, encoding="utf-8"):
     if not isinstance(s, str):
         s = ensure_contiguous_ndarray(s)
         s = codecs.decode(s, encoding)
     return s
 
 
-def ndarray_copy(src, dst):
+def ndarray_copy(src, dst) -> NDArrayLike:
     """Copy the contents of the array from `src` to `dst`."""
 
     if dst is None:
         # no-op
         return src
 
-    # ensure ndarrays
-    src = ensure_ndarray(src)
-    dst = ensure_ndarray(dst)
+    # ensure ndarray like
+    src = ensure_ndarray_like(src)
+    dst = ensure_ndarray_like(dst)
 
     # flatten source array
-    src = src.reshape(-1, order='A')
+    src = src.reshape(-1, order="A")
 
     # ensure same data type
     if dst.dtype != object:
@@ -153,9 +203,9 @@ def ndarray_copy(src, dst):
     # reshape source to match destination
     if src.shape != dst.shape:
         if dst.flags.f_contiguous:
-            order = 'F'
+            order = "F"
         else:
-            order = 'C'
+            order = "C"
         src = src.reshape(dst.shape, order=order)
 
     # copy via numpy

numcodecs/ndarray_like.py

@@ -0,0 +1,56 @@
+import sys
+from typing import Any, Optional, Tuple
+
+if sys.version_info >= (3, 8):
+    from typing import Protocol, runtime_checkable
+else:
+    from typing_extensions import Protocol, runtime_checkable
+
+
+@runtime_checkable
+class DType(Protocol):
+    itemsize: int
+    name: str
+    kind: str
+
+
+@runtime_checkable
+class FlagsObj(Protocol):
+    c_contiguous: bool
+    f_contiguous: bool
+    owndata: bool
+
+
+@runtime_checkable
+class NDArrayLike(Protocol):
+    dtype: DType
+    shape: Tuple[int, ...]
+    strides: Tuple[int, ...]
+    ndim: int
+    size: int
+    itemsize: int
+    nbytes: int
+    flags: FlagsObj
+
+    def __len__(self) -> int:
+        ...
+
+    def __getitem__(self, key) -> Any:
+        ...
+
+    def __setitem__(self, key, value):
+        ...
+
+    def tobytes(self, order: Optional[str] = ...) -> bytes:
+        ...
+
+    def reshape(self, *shape: int, order: str = ...) -> "NDArrayLike":
+        ...
+
+    def view(self, dtype: DType = ...) -> "NDArrayLike":
+        ...
+
+
+def is_ndarray_like(obj: object) -> bool:
+    """Return True when `obj` is ndarray-like"""
+    return isinstance(obj, NDArrayLike)

numcodecs/tests/test_ndarray_like.py

@@ -0,0 +1,15 @@
+import pytest
+
+from numcodecs.ndarray_like import NDArrayLike
+
+
+@pytest.mark.parametrize("module", ["numpy", "cupy"])
+def test_is_ndarray_like(module):
+    m = pytest.importorskip(module)
+    a = m.arange(10)
+    assert isinstance(a, NDArrayLike)
+
+
+def test_is_not_ndarray_like():
+    assert not isinstance([1, 2, 3], NDArrayLike)
+    assert not isinstance(b"1,2,3", NDArrayLike)

requirements.txt

@@ -3,3 +3,4 @@ numpy
 msgpack
 pytest
 zfpy
+typing-extensions

requirements_dev.txt

@@ -2,5 +2,4 @@ Cython==0.29.21
 msgpack==1.0.2
 numpy==1.21.0
 zfpy==0.5.5
-
-
+typing-extensions>=3.7.4

requirements_rtfd.txt

@@ -7,3 +7,4 @@ mock
 numpy
 cython
 zfpy==0.5.5
+typing-extensions

setup.py

@@ -338,6 +338,7 @@ def run_setup(with_extensions):
         ],
         install_requires=[
             'numpy>=1.7',
+            'typing-extensions>=3.7.4',
         ],
         extras_require={
             'msgpack':  ["msgpack"],