[XLA:Python] Fix bug where bfloat16 dtype extension's hash value was not initialized correctly.

We were (implicitly) initializing the hash value to 0, when -1 is the correct initialization for a PyArray_Descr's hash value.

Will fix https://github.com/google/jax/issues/4651 when incorporated into a jaxlib release.

PiperOrigin-RevId: 338088352
Change-Id: Ia9d72cbf88b8301b11c7a3bacbdba0848883d0ce

tensorflow/compiler/xla/python/bfloat16.cc

@@ -396,25 +396,30 @@ PyTypeObject PyBfloat16_Type = {
 PyArray_ArrFuncs NPyBfloat16_ArrFuncs;
 
 PyArray_Descr NPyBfloat16_Descr = {
-    PyObject_HEAD_INIT(nullptr) & PyBfloat16_Type,  // typeobj
+    PyObject_HEAD_INIT(nullptr)  //
+                                 /*typeobj=*/
+    (&PyBfloat16_Type),
     // We must register bfloat16 with a kind other than "f", because numpy
     // considers two types with the same kind and size to be equal, but
     // float16 != bfloat16.
     // The downside of this is that NumPy scalar promotion does not work with
     // bfloat16 values.
-    'V',  // kind
+    /*kind=*/'V',
     // TODO(phawkins): there doesn't seem to be a way of guaranteeing a type
     // character is unique.
-    'E',                                                  // type
+    /*type=*/'E',
-    '=',                                                  // byteorder
+    /*byteorder=*/'=',
-    NPY_NEEDS_PYAPI | NPY_USE_GETITEM | NPY_USE_SETITEM,  // hasobject
+    /*flags=*/NPY_NEEDS_PYAPI | NPY_USE_GETITEM | NPY_USE_SETITEM,
-    0,                                                    // type_num
+    /*type_num=*/0,
-    sizeof(bfloat16),                                     // elsize
+    /*elsize=*/sizeof(bfloat16),
-    alignof(bfloat16),                                    // alignment
+    /*alignment=*/alignof(bfloat16),
-    nullptr,                                              // subarray
+    /*subarray=*/nullptr,
-    nullptr,                                              // fields
+    /*fields=*/nullptr,
-    nullptr,                                              // names
+    /*names=*/nullptr,
-    &NPyBfloat16_ArrFuncs,                                // f
+    /*f=*/&NPyBfloat16_ArrFuncs,
+    /*metadata=*/nullptr,
+    /*c_metadata=*/nullptr,
+    /*hash=*/-1,  // -1 means "not computed yet".
 };
 
 // Implementations of NumPy array methods.

tensorflow/compiler/xla/python/bfloat16_test.py

@@ -19,6 +19,7 @@ from __future__ import division
 from __future__ import print_function
 
 import collections
+import copy
 import itertools
 import math
 
@@ -254,6 +255,15 @@ class Bfloat16NumPyTest(parameterized.TestCase):
   def testDtype(self):
     self.assertEqual(bfloat16, np.dtype(bfloat16))
 
+  def testDeepCopyDoesNotAlterHash(self):
+    # For context, see https://github.com/google/jax/issues/4651. If the hash
+    # value of the type descriptor is not initialized correctly, a deep copy
+    # can change the type hash.
+    dtype = np.dtype(bfloat16)
+    h = hash(dtype)
+    _ = copy.deepcopy(dtype)
+    self.assertEqual(h, hash(dtype))
+
   def testArray(self):
     x = np.array([[1, 2, 3]], dtype=bfloat16)
     self.assertEqual(bfloat16, x.dtype)

tensorflow/python/lib/core/bfloat16.cc

@@ -387,6 +387,9 @@ PyArray_Descr NPyBfloat16_Descr = {
     nullptr,                                              // fields
     nullptr,                                              // names
     &NPyBfloat16_ArrFuncs,                                // f
+    nullptr,                                              // metadata
+    nullptr,                                              // c_metadata
+    -1,                                                   // hash
 };
 
 // Registered numpy type ID. Global variable populated by the registration code.