Add casting keyword to numeric array types. (#547)

* Add casting keyowrd to AbstractArray and its children

* Explicitly add  to Array and CArray init

* Add docstring

* Update docstring

* Change requests

* Docstring tweak

* Docstring tweak

* make casting kwarg-onlt for ArrayAbstract

traits/tests/test_array.py

@@ -10,7 +10,7 @@
 
 import unittest
 
-from traits.api import Array, Bool, HasTraits
+from traits.api import Array, Bool, HasTraits, TraitError
 from traits.testing.optional_dependencies import numpy, requires_numpy
 
 
@@ -25,10 +25,10 @@ def _a_changed(self):
             self.event_fired = True
 
 
+@requires_numpy
 class ArrayTestCase(unittest.TestCase):
     """ Test cases for delegated traits. """
 
-    @requires_numpy
     def test_zero_to_one_element(self):
         """ Test that an event fires when an Array trait changes from zero to
         one element.
@@ -43,3 +43,19 @@ def test_zero_to_one_element(self):
 
         # Confirm that the static trait handler was invoked.
         self.assertEqual(f.event_fired, True)
+
+    def test_safe_casting(self):
+        class Bar(HasTraits):
+            unsafe_f32 = Array(dtype="float32")
+            safe_f32 = Array(dtype="float32", casting="safe")
+
+        f64 = numpy.array([1], dtype="float64")
+        f32 = numpy.array([1], dtype="float32")
+
+        b = Bar()
+
+        b.unsafe_f32 = f32
+        b.unsafe_f32 = f64
+        b.safe_f32 = f32
+        with self.assertRaises(TraitError):
+            b.safe_f32 = f64

traits/tests/test_array_or_none.py

@@ -166,3 +166,19 @@ class FooBar(HasTraits):
         foo = foo_bar.foo
         foo += 1729.0
         self.assertFalse((foo_bar.foo == foo_bar.bar).all())
+
+    def test_safe_casting(self):
+        class Bar(HasTraits):
+            unsafe_f32 = ArrayOrNone(dtype="float32")
+            safe_f32 = ArrayOrNone(dtype="float32", casting="safe")
+
+        f64 = numpy.array([1], dtype="float64")
+        f32 = numpy.array([1], dtype="float32")
+
+        b = Bar()
+
+        b.unsafe_f32 = f32
+        b.unsafe_f32 = f64
+        b.safe_f32 = f32
+        with self.assertRaises(TraitError):
+            b.safe_f32 = f64

traits/trait_numeric.py

@@ -55,6 +55,8 @@ def __init__(
         value=None,
         coerce=False,
         typecode=None,
+        *,
+        casting="unsafe",
         **metadata
     ):
         global ndarray, asarray
@@ -130,6 +132,7 @@ def __init__(
         self.dtype = dtype
         self.shape = shape
         self.coerce = coerce
+        self.casting = casting
 
         super(AbstractArray, self).__init__(value, **metadata)
 
@@ -148,11 +151,7 @@ def validate(self, object, name, value):
 
             # Make sure the array is of the right type:
             if (self.dtype is not None) and (value.dtype != self.dtype):
-                if self.coerce:
-                    value = value.astype(self.dtype)
-                else:
-                    # XXX: this also coerces.
-                    value = asarray(value, self.dtype)
+                value = value.astype(self.dtype, casting=self.casting)
 
             # If no shape requirements, then return the value:
             trait_shape = self.shape
@@ -310,13 +309,39 @@ class Array(AbstractArray):
         second dimension must be at least 2.)
     value : numpy array
         A default value for the array.
+    casting : str
+        Casting rule for the array's dtype. If ``dtype`` is set, a value can
+        only be assigned if it passes the casting rule. Values can be:
+
+        - "no": No casting is allowed
+        - "equiv": Only byte-order changes are allowed
+        - "safe": Only allow casting that fully preserves values (e.g.
+          "float32" to "float64")
+        - "same-kind": Only safe casts or casts within a kind (e.g. "float64"
+          to "float32") are allowed
+        - "unsafe": Any casting is allowed
+
+        Default is "unsafe".
     """
 
     def __init__(
-        self, dtype=None, shape=None, value=None, typecode=None, **metadata
+        self,
+        dtype=None,
+        shape=None,
+        value=None,
+        typecode=None,
+        *,
+        casting="unsafe",
+        **metadata
     ):
         super(Array, self).__init__(
-            dtype, shape, value, False, typecode=typecode, **metadata
+            dtype,
+            shape,
+            value,
+            False,
+            typecode=typecode,
+            casting=casting,
+            **metadata
         )
 
 
@@ -351,13 +376,39 @@ class CArray(AbstractArray):
         second dimension must be at least 2.)
     value : numpy array
         A default value for the array.
+    casting : str
+        Casting rule for the array's dtype. If ``dtype`` is set, a value can
+        only be assigned if it passes the casting rule. Values can be:
+
+        - "no": No casting is allowed
+        - "equiv": Only byte-order changes are allowed
+        - "safe": Only allow casting that fully preserves values (e.g.
+          "float32" to "float64")
+        - "same-kind": Only safe casts or casts within a kind (e.g. "float64"
+          to "float32") are allowed
+        - "unsafe": Any casting is allowed
+
+        Default is "unsafe".
     """
 
     def __init__(
-        self, dtype=None, shape=None, value=None, typecode=None, **metadata
+        self,
+        dtype=None,
+        shape=None,
+        value=None,
+        typecode=None,
+        *,
+        casting="unsafe",
+        **metadata
     ):
         super(CArray, self).__init__(
-            dtype, shape, value, True, typecode=typecode, **metadata
+            dtype,
+            shape,
+            value,
+            True,
+            typecode=typecode,
+            casting=casting,
+            **metadata
         )