Leverage on dpctl.tensor implementation in dpnp.diff

dpnp/dpnp_iface.py

@@ -654,7 +654,7 @@ def get_result_array(a, out=None, casting="safe"):
 
     Parameters
     ----------
-    a : {dpnp_array}
+    a : {dpnp.ndarray, usm_ndarray}
         Input array.
     out : {dpnp.ndarray, usm_ndarray}
         If provided, value of `a` array will be copied into it
@@ -671,6 +671,8 @@ def get_result_array(a, out=None, casting="safe"):
     """
 
     if out is None:
+        if isinstance(a, dpt.usm_ndarray):
+            return dpnp_array._create_from_usm_ndarray(a)
         return a
 
     if isinstance(out, dpt.usm_ndarray):

dpnp/dpnp_iface_counting.py

@@ -44,25 +44,38 @@
 __all__ = ["count_nonzero"]
 
 
-def count_nonzero(a, axis=None, *, keepdims=False):
+def count_nonzero(a, axis=None, *, keepdims=False, out=None):
     """
     Counts the number of non-zero values in the array `a`.
 
     For full documentation refer to :obj:`numpy.count_nonzero`.
 
+    Parameters
+    ----------
+    a : {dpnp.ndarray, usm_ndarray}
+        The array for which to count non-zeros.
+    axis : {None, int, tuple}, optional
+        Axis or tuple of axes along which to count non-zeros.
+        Default value means that non-zeros will be counted along a flattened
+        version of `a`.
+        Default: ``None``.
+    keepdims : bool, optional
+        If this is set to ``True``, the axes that are counted are left in the
+        result as dimensions with size one. With this option, the result will
+        broadcast correctly against the input array.
+        Default: ``False``.
+    out : {None, dpnp.ndarray, usm_ndarray}, optional
+        The array into which the result is written. The data type of `out` must
+        match the expected shape and the expected data type of the result.
+        If ``None`` then a new array is returned.
+        Default: ``None``.
+
     Returns
     -------
     out : dpnp.ndarray
         Number of non-zero values in the array along a given axis.
-        Otherwise, a zero-dimensional array with the total number of
-        non-zero values in the array is returned.
-
-    Limitations
-    -----------
-    Parameters `a` is supported as either :class:`dpnp.ndarray`
-    or :class:`dpctl.tensor.usm_ndarray`.
-    Otherwise ``TypeError`` exception will be raised.
-    Input array data types are limited by supported DPNP :ref:`Data types`.
+        Otherwise, a zero-dimensional array with the total number of non-zero
+        values in the array is returned.
 
     See Also
     --------
@@ -87,8 +100,10 @@ def count_nonzero(a, axis=None, *, keepdims=False):
 
     """
 
-    # TODO: might be improved by implementing an extension
-    # with `count_nonzero` kernel
     usm_a = dpnp.get_usm_ndarray(a)
-    usm_a = dpt.astype(usm_a, dpnp.bool, copy=False)
-    return dpnp.sum(usm_a, axis=axis, dtype=dpnp.intp, keepdims=keepdims)
+    usm_out = None if out is None else dpnp.get_usm_ndarray(out)
+
+    usm_res = dpt.count_nonzero(
+        usm_a, axis=axis, keepdims=keepdims, out=usm_out
+    )
+    return dpnp.get_result_array(usm_res, out)

dpnp/dpnp_iface_logic.py

@@ -51,7 +51,6 @@
 
 import dpnp
 from dpnp.dpnp_algo.dpnp_elementwise_common import DPNPBinaryFunc, DPNPUnaryFunc
-from dpnp.dpnp_array import dpnp_array
 
 __all__ = [
     "all",
@@ -167,13 +166,11 @@ def all(a, /, axis=None, out=None, keepdims=False, *, where=True):
 
     dpnp.check_limitations(where=where)
 
-    dpt_array = dpnp.get_usm_ndarray(a)
-    result = dpnp_array._create_from_usm_ndarray(
-        dpt.all(dpt_array, axis=axis, keepdims=keepdims)
-    )
+    usm_a = dpnp.get_usm_ndarray(a)
+    usm_res = dpt.all(usm_a, axis=axis, keepdims=keepdims)
+
     # TODO: temporary solution until dpt.all supports out parameter
-    result = dpnp.get_result_array(result, out)
-    return result
+    return dpnp.get_result_array(usm_res, out)
 
 
 def allclose(a, b, rtol=1.0e-5, atol=1.0e-8, equal_nan=False):
@@ -333,13 +330,11 @@ def any(a, /, axis=None, out=None, keepdims=False, *, where=True):
 
     dpnp.check_limitations(where=where)
 
-    dpt_array = dpnp.get_usm_ndarray(a)
-    result = dpnp_array._create_from_usm_ndarray(
-        dpt.any(dpt_array, axis=axis, keepdims=keepdims)
-    )
+    usm_a = dpnp.get_usm_ndarray(a)
+    usm_res = dpt.any(usm_a, axis=axis, keepdims=keepdims)
+
     # TODO: temporary solution until dpt.any supports out parameter
-    result = dpnp.get_result_array(result, out)
-    return result
+    return dpnp.get_result_array(usm_res, out)
 
 
 _EQUAL_DOCSTRING = """

dpnp/dpnp_iface_mathematical.py

@@ -130,31 +130,6 @@
 ]
 
 
-def _append_to_diff_array(a, axis, combined, values):
-    """
-    Append `values` to `combined` list based on data of array `a`.
-
-    Scalar value (including case with 0d array) is expanded to an array
-    with length=1 in the direction of axis and the shape of the input array `a`
-    along all other axes.
-    Note, if `values` is a scalar, then it is converted to 0d array allocating
-    on the same SYCL queue as the input array `a` and with the same USM type.
-
-    """
-
-    dpnp.check_supported_arrays_type(values, scalar_type=True, all_scalars=True)
-    if dpnp.isscalar(values):
-        values = dpnp.asarray(
-            values, sycl_queue=a.sycl_queue, usm_type=a.usm_type
-        )
-
-    if values.ndim == 0:
-        shape = list(a.shape)
-        shape[axis] = 1
-        values = dpnp.broadcast_to(values, tuple(shape))
-    combined.append(values)
-
-
 def _get_reduction_res_dt(a, dtype, _out):
     """Get a data type used by dpctl for result array in reduction function."""
 
@@ -1206,39 +1181,18 @@ def diff(a, n=1, axis=-1, prepend=None, append=None):
 
     """
 
-    dpnp.check_supported_arrays_type(a)
-    if n == 0:
-        return a
+    usm_a = dpnp.get_usm_ndarray(a)
+    usm_pre = (
+        None if prepend is None else dpnp.get_usm_ndarray_or_scalar(prepend)
+    )
+    usm_app = None if append is None else dpnp.get_usm_ndarray_or_scalar(append)
+
     if n < 0:
+        # TODO: remove once dpctl-1779 is resolved
         raise ValueError(f"order must be non-negative but got {n}")
 
-    nd = a.ndim
-    if nd == 0:
-        raise ValueError("diff requires input that is at least one dimensional")
-    axis = normalize_axis_index(axis, nd)
-
-    combined = []
-    if prepend is not None:
-        _append_to_diff_array(a, axis, combined, prepend)
-
-    combined.append(a)
-    if append is not None:
-        _append_to_diff_array(a, axis, combined, append)
-
-    if len(combined) > 1:
-        a = dpnp.concatenate(combined, axis=axis)
-
-    slice1 = [slice(None)] * nd
-    slice2 = [slice(None)] * nd
-    slice1[axis] = slice(1, None)
-    slice2[axis] = slice(None, -1)
-    slice1 = tuple(slice1)
-    slice2 = tuple(slice2)
-
-    op = dpnp.not_equal if a.dtype == numpy.bool_ else dpnp.subtract
-    for _ in range(n):
-        a = op(a[slice1], a[slice2])
-    return a
+    usm_res = dpt.diff(usm_a, axis=axis, n=n, prepend=usm_pre, append=usm_app)
+    return dpnp_array._create_from_usm_ndarray(usm_res)
 
 
 _DIVIDE_DOCSTRING = """

dpnp/dpnp_iface_searching.py

@@ -391,7 +391,5 @@ def where(condition, x=None, y=None, /, *, order="K", out=None):
     usm_condition = dpnp.get_usm_ndarray(condition)
 
     usm_out = None if out is None else dpnp.get_usm_ndarray(out)
-    result = dpnp_array._create_from_usm_ndarray(
-        dpt.where(usm_condition, usm_x, usm_y, order=order, out=usm_out)
-    )
-    return dpnp.get_result_array(result, out)
+    usm_res = dpt.where(usm_condition, usm_x, usm_y, order=order, out=usm_out)
+    return dpnp.get_result_array(usm_res, out)

dpnp/dpnp_iface_statistics.py

@@ -904,11 +904,9 @@ def std(
         )
         dpnp.sqrt(result, out=result)
     else:
-        dpt_array = dpnp.get_usm_ndarray(a)
-        result = dpnp_array._create_from_usm_ndarray(
-            dpt.std(dpt_array, axis=axis, correction=ddof, keepdims=keepdims)
-        )
-        result = dpnp.get_result_array(result, out)
+        usm_a = dpnp.get_usm_ndarray(a)
+        usm_res = dpt.std(usm_a, axis=axis, correction=ddof, keepdims=keepdims)
+        result = dpnp.get_result_array(usm_res, out)
 
     if dtype is not None and out is None:
         result = result.astype(dtype, casting="same_kind")
@@ -1028,11 +1026,9 @@ def var(
 
         dpnp.divide(result, cnt, out=result)
     else:
-        dpt_array = dpnp.get_usm_ndarray(a)
-        result = dpnp_array._create_from_usm_ndarray(
-            dpt.var(dpt_array, axis=axis, correction=ddof, keepdims=keepdims)
-        )
-        result = dpnp.get_result_array(result, out)
+        usm_a = dpnp.get_usm_ndarray(a)
+        usm_res = dpt.var(usm_a, axis=axis, correction=ddof, keepdims=keepdims)
+        result = dpnp.get_result_array(usm_res, out)
 
     if out is None and dtype is not None:
         result = result.astype(dtype, casting="same_kind")

dpnp/dpnp_utils/dpnp_utils_reduction.py

@@ -25,7 +25,6 @@
 
 
 import dpnp
-from dpnp.dpnp_array import dpnp_array
 
 __all__ = ["dpnp_wrap_reduction_call"]
 
@@ -53,5 +52,4 @@ def dpnp_wrap_reduction_call(
 
     kwargs["out"] = usm_out
     res_usm = _reduction_fn(*args, **kwargs)
-    res = dpnp_array._create_from_usm_ndarray(res_usm)
-    return dpnp.get_result_array(res, input_out, casting="unsafe")
+    return dpnp.get_result_array(res_usm, input_out, casting="unsafe")

tests/third_party/cupy/math_tests/test_sumprod.py

@@ -687,7 +687,11 @@ def test_diff_2dim_with_append(self, xp, dtype):
     @testing.numpy_cupy_allclose(type_check=has_support_aspect64())
     def test_diff_2dim_with_scalar_append(self, xp, dtype):
         a = testing.shaped_arange((4, 5), xp, dtype)
-        return xp.diff(a, prepend=1, append=0)
+        res = xp.diff(a, prepend=1, append=0)
+        if xp is numpy and a.dtype != res.dtype:
+            # res.dtype must follow a.dtype
+            res = res.astype(a.dtype)
+        return res
 
     @testing.with_requires("numpy>=1.16")
     def test_diff_invalid_axis(self):