Merge master into workaround_qr_cuda

dpnp/dpnp_array.py

@@ -605,6 +605,25 @@ def __truediv__(self, other):
         """Return ``self/value``."""
         return dpnp.true_divide(self, other)
 
+    @property
+    def __usm_ndarray__(self):
+        """
+        Property to support `__usm_ndarray__` protocol.
+
+        It assumes to return :class:`dpctl.tensor.usm_ndarray` instance
+        corresponding to the content of the object.
+
+        This property is intended to speed-up conversion from
+        :class:`dpnp.ndarray` to :class:`dpctl.tensor.usm_ndarray` passed
+        into  `dpctl.tensor.asarray` function. The input object that implements
+        `__usm_ndarray__` protocol is recognized as owner of USM allocation
+        that is managed by a smart pointer, and asynchronous deallocation
+        will not involve GIL.
+
+        """
+
+        return self._array_obj
+
     def __xor__(self, other):
         """Return ``self^value``."""
         return dpnp.bitwise_xor(self, other)

dpnp/tests/test_ndarray.py

@@ -176,6 +176,18 @@ def test_error(self):
             ia.item()
 
 
+class TestUsmNdarrayProtocol:
+    def test_basic(self):
+        a = dpnp.arange(256, dtype=dpnp.int64)
+        usm_a = dpt.asarray(a)
+
+        assert a.sycl_queue == usm_a.sycl_queue
+        assert a.usm_type == usm_a.usm_type
+        assert a.dtype == usm_a.dtype
+        assert usm_a.usm_data.reference_obj is None
+        assert (a == usm_a).all()
+
+
 def test_print_dpnp_int():
     result = repr(dpnp.array([1, 0, 2, -3, -1, 2, 21, -9], dtype="i4"))
     expected = "array([ 1,  0,  2, -3, -1,  2, 21, -9], dtype=int32)"