Merge pull request #20722 from madphysicist/dtype-checking-1

ENH: Removed requirement for C-contiguity when changing to dtype of different size

5 files changed, 102 insertions, 52 deletions

diff --git a/numpy/core/_add_newdocs.py b/numpy/core/_add_newdocs.py
index dcec768f0..a8d73af3f 100644
--- a/numpy/core/_add_newdocs.py
+++ b/numpy/core/_add_newdocs.py
@@ -4469,14 +4469,13 @@ add_newdoc('numpy.core.multiarray', 'ndarray', ('view',
     memory.
 
     For ``a.view(some_dtype)``, if ``some_dtype`` has a different number of
-    bytes per entry than the previous dtype (for example, converting a
-    regular array to a structured array), then the behavior of the view
-    cannot be predicted just from the superficial appearance of ``a`` (shown
-    by ``print(a)``). It also depends on exactly how ``a`` is stored in
-    memory. Therefore if ``a`` is C-ordered versus fortran-ordered, versus
-    defined as a slice or transpose, etc., the view may give different
-    results.
+    bytes per entry than the previous dtype (for example, converting a regular
+    array to a structured array), then the last axis of ``a`` must be
+    contiguous. This axis will be resized in the result.
 
+    .. versionchanged:: 1.23.0
+       Only the last axis needs to be contiguous. Previously, the entire array
+       had to be C-contiguous.
 
     Examples
     --------
@@ -4521,19 +4520,34 @@ add_newdoc('numpy.core.multiarray', 'ndarray', ('view',
     Views that change the dtype size (bytes per entry) should normally be
     avoided on arrays defined by slices, transposes, fortran-ordering, etc.:
 
-    >>> x = np.array([[1,2,3],[4,5,6]], dtype=np.int16)
-    >>> y = x[:, 0:2]
+    >>> x = np.array([[1, 2, 3], [4, 5, 6]], dtype=np.int16)
+    >>> y = x[:, ::2]
     >>> y
-    array([[1, 2],
-           [4, 5]], dtype=int16)
+    array([[1, 3],
+           [4, 6]], dtype=int16)
     >>> y.view(dtype=[('width', np.int16), ('length', np.int16)])
     Traceback (most recent call last):
         ...
-    ValueError: To change to a dtype of a different size, the array must be C-contiguous
+    ValueError: To change to a dtype of a different size, the last axis must be contiguous
     >>> z = y.copy()
     >>> z.view(dtype=[('width', np.int16), ('length', np.int16)])
-    array([[(1, 2)],
-           [(4, 5)]], dtype=[('width', '<i2'), ('length', '<i2')])
+    array([[(1, 3)],
+           [(4, 6)]], dtype=[('width', '<i2'), ('length', '<i2')])
+
+    However, views that change dtype are totally fine for arrays with a
+    contiguous last axis, even if the rest of the axes are not C-contiguous:
+
+    >>> x = np.arange(2 * 3 * 4, dtype=np.int8).reshape(2, 3, 4)
+    >>> x.transpose(1, 0, 2).view(np.int16)
+    array([[[ 256,  770],
+            [3340, 3854]],
+    <BLANKLINE>
+           [[1284, 1798],
+            [4368, 4882]],
+    <BLANKLINE>
+           [[2312, 2826],
+            [5396, 5910]]], dtype=int16)
+
     """))
 
 
diff --git a/numpy/core/src/multiarray/getset.c b/numpy/core/src/multiarray/getset.c
index ce21e948e..ac6465acd 100644
--- a/numpy/core/src/multiarray/getset.c
+++ b/numpy/core/src/multiarray/getset.c
@@ -496,9 +496,6 @@ array_descr_set(PyArrayObject *self, PyObject *arg, void *NPY_UNUSED(ignored))
 
     /* Changing the size of the dtype results in a shape change */
     if (newtype->elsize != PyArray_DESCR(self)->elsize) {
-        int axis;
-        npy_intp newdim;
-
         /* forbidden cases */
         if (PyArray_NDIM(self) == 0) {
             PyErr_SetString(PyExc_ValueError,
@@ -513,31 +510,20 @@ array_descr_set(PyArrayObject *self, PyObject *arg, void *NPY_UNUSED(ignored))
             goto fail;
         }
 
-        /* determine which axis to resize */
-        if (PyArray_IS_C_CONTIGUOUS(self)) {
-            axis = PyArray_NDIM(self) - 1;
-        }
-        else if (PyArray_IS_F_CONTIGUOUS(self)) {
-            /* 2015-11-27 1.11.0, gh-6747 */
-            if (DEPRECATE(
-                        "Changing the shape of an F-contiguous array by "
-                        "descriptor assignment is deprecated. To maintain the "
-                        "Fortran contiguity of a multidimensional Fortran "
-                        "array, use 'a.T.view(...).T' instead") < 0) {
-                goto fail;
-            }
-            axis = 0;
-        }
-        else {
-            /* Don't mention the deprecated F-contiguous support */
+        /* resize on last axis only */
+        int axis = PyArray_NDIM(self) - 1;
+        if (PyArray_DIMS(self)[axis] != 1 &&
+                PyArray_STRIDES(self)[axis] != PyArray_DESCR(self)->elsize) {
             PyErr_SetString(PyExc_ValueError,
-                    "To change to a dtype of a different size, the array must "
-                    "be C-contiguous");
+                    "To change to a dtype of a different size, the last axis "
+                    "must be contiguous");
             goto fail;
         }
 
+        npy_intp newdim;
+
         if (newtype->elsize < PyArray_DESCR(self)->elsize) {
-            /* if it is compatible, increase the size of the relevant axis */
+            /* if it is compatible, increase the size of the last axis */
             if (newtype->elsize == 0 ||
                     PyArray_DESCR(self)->elsize % newtype->elsize != 0) {
                 PyErr_SetString(PyExc_ValueError,
@@ -549,7 +535,7 @@ array_descr_set(PyArrayObject *self, PyObject *arg, void *NPY_UNUSED(ignored))
             PyArray_DIMS(self)[axis] *= newdim;
             PyArray_STRIDES(self)[axis] = newtype->elsize;
         }
-        else if (newtype->elsize > PyArray_DESCR(self)->elsize) {
+        else /* newtype->elsize > PyArray_DESCR(self)->elsize */ {
             /* if it is compatible, decrease the size of the relevant axis */
             newdim = PyArray_DIMS(self)[axis] * PyArray_DESCR(self)->elsize;
             if ((newdim % newtype->elsize) != 0) {
diff --git a/numpy/core/tests/test_deprecations.py b/numpy/core/tests/test_deprecations.py
index d148c89f5..76486f755 100644
--- a/numpy/core/tests/test_deprecations.py
+++ b/numpy/core/tests/test_deprecations.py
@@ -257,20 +257,6 @@ class TestDatetime64Timezone(_DeprecationTestCase):
         self.assert_deprecated(np.datetime64, args=(dt,))
 
 
-class TestNonCContiguousViewDeprecation(_DeprecationTestCase):
-    """View of non-C-contiguous arrays deprecated in 1.11.0.
-
-    The deprecation will not be raised for arrays that are both C and F
-    contiguous, as C contiguous is dominant. There are more such arrays
-    with relaxed stride checking than without so the deprecation is not
-    as visible with relaxed stride checking in force.
-    """
-
-    def test_fortran_contiguous(self):
-        self.assert_deprecated(np.ones((2,2)).T.view, args=(complex,))
-        self.assert_deprecated(np.ones((2,2)).T.view, args=(np.int8,))
-
-
 class TestArrayDataAttributeAssignmentDeprecation(_DeprecationTestCase):
     """Assigning the 'data' attribute of an ndarray is unsafe as pointed
      out in gh-7093. Eventually, such assignment should NOT be allowed, but
diff --git a/numpy/core/tests/test_multiarray.py b/numpy/core/tests/test_multiarray.py
index 0c611abb5..2529705d5 100644
--- a/numpy/core/tests/test_multiarray.py
+++ b/numpy/core/tests/test_multiarray.py
@@ -9191,3 +9191,66 @@ def test_getfield():
     pytest.raises(ValueError, a.getfield, 'uint8', -1)
     pytest.raises(ValueError, a.getfield, 'uint8', 16)
     pytest.raises(ValueError, a.getfield, 'uint64', 0)
+
+
+class TestViewDtype:
+    """
+    Verify that making a view of a non-contiguous array works as expected.
+    """
+    def test_smaller_dtype_multiple(self):
+        # x is non-contiguous
+        x = np.arange(10, dtype='<i4')[::2]
+        with pytest.raises(ValueError,
+                           match='the last axis must be contiguous'): 
+            x.view('<i2')
+        expected = [[0, 0], [2, 0], [4, 0], [6, 0], [8, 0]]
+        assert_array_equal(x[:, np.newaxis].view('<i2'), expected)
+
+    def test_smaller_dtype_not_multiple(self):
+        # x is non-contiguous
+        x = np.arange(5, dtype='<i4')[::2]
+
+        with pytest.raises(ValueError,
+                           match='the last axis must be contiguous'):
+            x.view('S3')
+        with pytest.raises(ValueError,
+                           match='When changing to a smaller dtype'):
+            x[:, np.newaxis].view('S3')
+
+        # Make sure the problem is because of the dtype size
+        expected = [[b''], [b'\x02'], [b'\x04']]
+        assert_array_equal(x[:, np.newaxis].view('S4'), expected)
+
+    def test_larger_dtype_multiple(self):
+        # x is non-contiguous in the first dimension, contiguous in the last
+        x = np.arange(20, dtype='<i2').reshape(10, 2)[::2, :]
+        expected = np.array([[65536], [327684], [589832],
+                             [851980], [1114128]], dtype='<i4')
+        assert_array_equal(x.view('<i4'), expected)
+
+    def test_larger_dtype_not_multiple(self):
+        # x is non-contiguous in the first dimension, contiguous in the last
+        x = np.arange(20, dtype='<i2').reshape(10, 2)[::2, :]
+        with pytest.raises(ValueError,
+                           match='When changing to a larger dtype'):
+            x.view('S3')
+        # Make sure the problem is because of the dtype size
+        expected = [[b'\x00\x00\x01'], [b'\x04\x00\x05'], [b'\x08\x00\t'],
+                    [b'\x0c\x00\r'], [b'\x10\x00\x11']]
+        assert_array_equal(x.view('S4'), expected)
+
+    def test_f_contiguous(self):
+        # x is F-contiguous
+        x = np.arange(4 * 3, dtype='<i4').reshape(4, 3).T
+        with pytest.raises(ValueError,
+                           match='the last axis must be contiguous'):
+            x.view('<i2')
+
+    def test_non_c_contiguous(self):
+        # x is contiguous in axis=-1, but not C-contiguous in other axes
+        x = np.arange(2 * 3 * 4, dtype='i1').\
+                    reshape(2, 3, 4).transpose(1, 0, 2)
+        expected = [[[256, 770], [3340, 3854]],
+                    [[1284, 1798], [4368, 4882]],
+                    [[2312, 2826], [5396, 5910]]]
+        assert_array_equal(x.view('<i2'), expected)
diff --git a/numpy/lib/stride_tricks.py b/numpy/lib/stride_tricks.py
index 5093993a9..6794ad557 100644
--- a/numpy/lib/stride_tricks.py
+++ b/numpy/lib/stride_tricks.py
@@ -86,6 +86,7 @@ def as_strided(x, shape=None, strides=None, subok=False, writeable=True):
     Vectorized write operations on such arrays will typically be
     unpredictable. They may even give different results for small, large,
     or transposed arrays.
+
     Since writing to these arrays has to be tested and done with great
     care, you may want to use ``writeable=False`` to avoid accidental write
     operations.