Apply fallback to numpy for all unsupported functions on cuda device.

Author: npolina4

dpnp/dpnp_iface.py

@@ -69,10 +69,10 @@
     "get_normalized_queue_device",
     "get_result_array",
     "get_usm_ndarray",
+    "is_cuda_backend",
     "get_usm_ndarray_or_scalar",
     "is_supported_array_or_scalar",
     "is_supported_array_type",
-    "not_implemented_for_cuda_backend",
     "synchronize_array_data",
 ]
 
@@ -758,6 +758,37 @@ def get_usm_ndarray_or_scalar(a):
     return a if dpnp.isscalar(a) else get_usm_ndarray(a)
 
 
+def is_cuda_backend(obj=None):
+    """
+    Checks that object has a cuda backend.
+
+    Parameters
+    ----------
+    obj : {Device, SyclDevice, SyclQueue, dpnp.ndarray, usm_ndarray, None},
+          optional
+        An input object with sycl_device property to check device backend.
+        If obj is  ``None``, device backend will be checked for the default
+        queue.
+        Default: ``None``.
+
+    Returns
+    -------
+    out : bool
+        Return ``True`` if object has a cuda backend, otherwise``False``.
+
+    """
+
+    if obj is None:
+        sycl_device = dpctl.SyclQueue().sycl_device
+    elif isinstance(obj, dpctl.SyclDevice):
+        sycl_device = obj
+    else:
+        sycl_device = getattr(obj, "sycl_device", None)
+    if sycl_device is not None and "cuda" in sycl_device.backend.name:
+        return True
+    return False
+
+
 def is_supported_array_or_scalar(a):
     """
     Return ``True`` if `a` is a scalar or an array of either
@@ -801,21 +832,6 @@ def is_supported_array_type(a):
     return isinstance(a, (dpnp_array, dpt.usm_ndarray))
 
 
-def not_implemented_for_cuda_backend(obj):
-    """
-    Raise NotImplementedError for cuda devices.
-
-    Parameters
-    ----------
-    obj : {SyclDevice, SyclQueue, dpnp.ndarray, usm_ndarray}
-        An input object with sycl_device property to check device backend.
-
-    """
-    sycl_device = getattr(obj, "sycl_device", None)
-    if sycl_device is not None and "cuda" in sycl_device.backend.name:
-        raise NotImplementedError("function not implemented for cuda backend")
-
-
 def synchronize_array_data(a):
     """
     The dpctl interface was reworked to make asynchronous execution.

dpnp/dpnp_iface_indexing.py

@@ -174,8 +174,6 @@ def choose(x1, choices, out=None, mode="raise"):
 
     """
 
-    dpnp.not_implemented_for_cuda_backend(x1)
-
     x1_desc = dpnp.get_dpnp_descriptor(x1, copy_when_nondefault_queue=False)
 
     choices_list = []
@@ -195,6 +193,8 @@ def choose(x1, choices, out=None, mode="raise"):
             pass
         elif not choices_list:
             pass
+        elif dpnp.is_cuda_backend(x1):
+            pass
         else:
             size = x1_desc.size
             choices_size = choices_list[0].size

dpnp/dpnp_iface_libmath.py

@@ -78,12 +78,10 @@ def erf(in_array1):
 
     """
 
-    dpnp.not_implemented_for_cuda_backend(in_array1)
-
     x1_desc = dpnp.get_dpnp_descriptor(
         in_array1, copy_when_strides=False, copy_when_nondefault_queue=False
     )
-    if x1_desc:
+    if x1_desc and dpnp.is_cuda_backend(in_array1):
         return dpnp_erf(x1_desc).get_pyobj()
 
     result = create_output_descriptor_py(

dpnp/dpnp_iface_mathematical.py

@@ -2477,11 +2477,14 @@ def modf(x1, **kwargs):
 
     """
 
-    dpnp.not_implemented_for_cuda_backend(x1)
-
     x1_desc = dpnp.get_dpnp_descriptor(x1, copy_when_nondefault_queue=False)
-    if x1_desc and not kwargs:
-        return dpnp_modf(x1_desc)
+    if x1_desc:
+        if not kwargs:
+            pass
+        elif dpnp.is_cuda_backend(x1):
+            pass
+        else:
+            return dpnp_modf(x1_desc)
 
     return call_origin(numpy.modf, x1, **kwargs)
 

dpnp/dpnp_iface_sorting.py

@@ -174,8 +174,6 @@ def partition(x1, kth, axis=-1, kind="introselect", order=None):
 
     """
 
-    dpnp.not_implemented_for_cuda_backend(x1)
-
     x1_desc = dpnp.get_dpnp_descriptor(x1, copy_when_nondefault_queue=False)
     if x1_desc:
         if not isinstance(kth, int):
@@ -190,6 +188,8 @@ def partition(x1, kth, axis=-1, kind="introselect", order=None):
             pass
         elif order is not None:
             pass
+        elif dpnp.is_cuda_backend(x1):
+            pass
         else:
             return dpnp_partition(x1_desc, kth, axis, kind, order).get_pyobj()
 

dpnp/dpnp_iface_statistics.py

@@ -371,8 +371,6 @@ def correlate(x1, x2, mode="valid"):
 
     """
 
-    dpnp.not_implemented_for_cuda_backend(x1)
-
     x1_desc = dpnp.get_dpnp_descriptor(x1, copy_when_nondefault_queue=False)
     x2_desc = dpnp.get_dpnp_descriptor(x2, copy_when_nondefault_queue=False)
     if x1_desc and x2_desc:
@@ -382,6 +380,8 @@ def correlate(x1, x2, mode="valid"):
             pass
         elif mode != "valid":
             pass
+        elif dpnp.is_cuda_backend(x1) or dpnp.is_cuda_backend(x2):
+            pass
         else:
             return dpnp_correlate(x1_desc, x2_desc).get_pyobj()
 
@@ -657,8 +657,6 @@ def median(x1, axis=None, out=None, overwrite_input=False, keepdims=False):
 
     """
 
-    dpnp.not_implemented_for_cuda_backend(x1)
-
     x1_desc = dpnp.get_dpnp_descriptor(x1, copy_when_nondefault_queue=False)
     if x1_desc:
         if axis is not None:
@@ -669,6 +667,8 @@ def median(x1, axis=None, out=None, overwrite_input=False, keepdims=False):
             pass
         elif keepdims:
             pass
+        elif dpnp.is_cuda_backend(x1):
+            pass
         else:
             result_obj = dpnp_median(x1_desc).get_pyobj()
             result = dpnp.convert_single_elem_array_to_scalar(result_obj)