Rework transpose methods to use dpctl.tensor functions (#1389)

* Rework transpose methods to call dpctl.tensor functions

* Applied review comments & added more tests

Author: antonwolfy

dpnp/backend/include/dpnp_iface_fptr.hpp

@@ -360,7 +360,6 @@ enum class DPNPFuncName : size_t
     DPNP_FN_TANH,                         /**< Used in numpy.tanh() impl  */
     DPNP_FN_TANH_EXT,                     /**< Used in numpy.tanh() impl, requires extra parameters */
     DPNP_FN_TRANSPOSE,                    /**< Used in numpy.transpose() impl  */
-    DPNP_FN_TRANSPOSE_EXT,                /**< Used in numpy.transpose() impl, requires extra parameters */
     DPNP_FN_TRACE,                        /**< Used in numpy.trace() impl  */
     DPNP_FN_TRACE_EXT,                    /**< Used in numpy.trace() impl, requires extra parameters */
     DPNP_FN_TRAPZ,                        /**< Used in numpy.trapz() impl  */

dpnp/backend/kernels/dpnp_krnl_manipulation.cpp

@@ -1,5 +1,5 @@
 //*****************************************************************************
-// Copyright (c) 2016-2020, Intel Corporation
+// Copyright (c) 2016-2023, Intel Corporation
 // All rights reserved.
 //
 // Redistribution and use in source and binary forms, with or without
@@ -211,6 +211,7 @@ void dpnp_elemwise_transpose_c(void* array1_in,
                                                                        size,
                                                                        dep_event_vec_ref);
     DPCTLEvent_WaitAndThrow(event_ref);
+    DPCTLEvent_Delete(event_ref);
 }
 
 template <typename _DataType>
@@ -222,17 +223,6 @@ void (*dpnp_elemwise_transpose_default_c)(void*,
                                           void*,
                                           size_t) = dpnp_elemwise_transpose_c<_DataType>;
 
-template <typename _DataType>
-DPCTLSyclEventRef (*dpnp_elemwise_transpose_ext_c)(DPCTLSyclQueueRef,
-                                                   void*,
-                                                   const shape_elem_type*,
-                                                   const shape_elem_type*,
-                                                   const shape_elem_type*,
-                                                   size_t,
-                                                   void*,
-                                                   size_t,
-                                                   const DPCTLEventVectorRef) = dpnp_elemwise_transpose_c<_DataType>;
-
 void func_map_init_manipulation(func_map_t& fmap)
 {
     fmap[DPNPFuncName::DPNP_FN_REPEAT][eft_INT][eft_INT] = {eft_INT, (void*)dpnp_repeat_default_c<int32_t>};
@@ -253,15 +243,5 @@ void func_map_init_manipulation(func_map_t& fmap)
                                                                (void*)dpnp_elemwise_transpose_default_c<float>};
     fmap[DPNPFuncName::DPNP_FN_TRANSPOSE][eft_DBL][eft_DBL] = {eft_DBL,
                                                                (void*)dpnp_elemwise_transpose_default_c<double>};
-
-    fmap[DPNPFuncName::DPNP_FN_TRANSPOSE_EXT][eft_INT][eft_INT] = {eft_INT,
-                                                                   (void*)dpnp_elemwise_transpose_ext_c<int32_t>};
-    fmap[DPNPFuncName::DPNP_FN_TRANSPOSE_EXT][eft_LNG][eft_LNG] = {eft_LNG,
-                                                                   (void*)dpnp_elemwise_transpose_ext_c<int64_t>};
-    fmap[DPNPFuncName::DPNP_FN_TRANSPOSE_EXT][eft_FLT][eft_FLT] = {eft_FLT,
-                                                                   (void*)dpnp_elemwise_transpose_ext_c<float>};
-    fmap[DPNPFuncName::DPNP_FN_TRANSPOSE_EXT][eft_DBL][eft_DBL] = {eft_DBL,
-                                                                   (void*)dpnp_elemwise_transpose_ext_c<double>};
-
     return;
 }

dpnp/dpnp_algo/dpnp_algo.pxd

@@ -338,7 +338,6 @@ cdef extern from "dpnp_iface_fptr.hpp" namespace "DPNPFuncName":  # need this na
         DPNP_FN_TRACE
         DPNP_FN_TRACE_EXT
         DPNP_FN_TRANSPOSE
-        DPNP_FN_TRANSPOSE_EXT
         DPNP_FN_TRAPZ
         DPNP_FN_TRAPZ_EXT
         DPNP_FN_TRI
@@ -551,7 +550,6 @@ cpdef dpnp_descriptor dpnp_subtract(dpnp_descriptor x1_obj, dpnp_descriptor x2_o
 Array manipulation routines
 """
 cpdef dpnp_descriptor dpnp_repeat(dpnp_descriptor array1, repeats, axes=*)
-cpdef dpnp_descriptor dpnp_transpose(dpnp_descriptor array1, axes=*)
 
 
 """

dpnp/dpnp_algo/dpnp_algo_manipulation.pxi

@@ -42,20 +42,10 @@ __all__ += [
     "dpnp_expand_dims",
     "dpnp_repeat",
     "dpnp_reshape",
-    "dpnp_transpose",
 ]
 
 
 # C function pointer to the C library template functions
-ctypedef c_dpctl.DPCTLSyclEventRef(*fptr_custom_elemwise_transpose_1in_1out_t)(c_dpctl.DPCTLSyclQueueRef,
-                                                                               void * ,
-                                                                               shape_elem_type * ,
-                                                                               shape_elem_type * ,
-                                                                               shape_elem_type * ,
-                                                                               size_t,
-                                                                               void * ,
-                                                                               size_t,
-                                                                               const c_dpctl.DPCTLEventVectorRef)
 ctypedef c_dpctl.DPCTLSyclEventRef(*fptr_dpnp_repeat_t)(c_dpctl.DPCTLSyclQueueRef,
                                                         const void *, void * , const size_t , const size_t,
                                                         const c_dpctl.DPCTLEventVectorRef)
@@ -229,67 +219,3 @@ cpdef utils.dpnp_descriptor dpnp_reshape(utils.dpnp_descriptor array1, newshape,
                                                usm_type=array1_obj.usm_type,
                                                sycl_queue=array1_obj.sycl_queue),
                                     copy_when_nondefault_queue=False)
-
-
-cpdef utils.dpnp_descriptor dpnp_transpose(utils.dpnp_descriptor array1, axes=None):
-    cdef shape_type_c input_shape = array1.shape
-    cdef size_t input_shape_size = array1.ndim
-    cdef shape_type_c result_shape = shape_type_c(input_shape_size, 1)
-
-    cdef shape_type_c permute_axes
-    if axes is None:
-        """
-        template to do transpose a tensor
-        input_shape=[2, 3, 4]
-        permute_axes=[2, 1, 0]
-        after application `permute_axes` to `input_shape` result:
-        result_shape=[4, 3, 2]
-
-        'do nothing' axes variable is `permute_axes=[0, 1, 2]`
-
-        test: pytest tests/third_party/cupy/manipulation_tests/test_transpose.py::TestTranspose::test_external_transpose_all
-        """
-        permute_axes = list(reversed([i for i in range(input_shape_size)]))
-    else:
-        permute_axes = utils.normalize_axis(axes, input_shape_size)
-
-    for i in range(input_shape_size):
-        """ construct output shape """
-        result_shape[i] = input_shape[permute_axes[i]]
-
-    # convert string type names (array.dtype) to C enum DPNPFuncType
-    cdef DPNPFuncType param1_type = dpnp_dtype_to_DPNPFuncType(array1.dtype)
-
-    # get the FPTR data structure
-    cdef DPNPFuncData kernel_data = get_dpnp_function_ptr(DPNP_FN_TRANSPOSE_EXT, param1_type, param1_type)
-
-    array1_obj = array1.get_array()
-
-    # ceate result array with type given by FPTR data
-    cdef utils.dpnp_descriptor result = utils.create_output_descriptor(result_shape,
-                                                                       kernel_data.return_type,
-                                                                       None,
-                                                                       device=array1_obj.sycl_device,
-                                                                       usm_type=array1_obj.usm_type,
-                                                                       sycl_queue=array1_obj.sycl_queue)
-    result_sycl_queue = result.get_array().sycl_queue
-
-    cdef c_dpctl.SyclQueue q = <c_dpctl.SyclQueue> result_sycl_queue
-    cdef c_dpctl.DPCTLSyclQueueRef q_ref = q.get_queue_ref()
-
-    cdef fptr_custom_elemwise_transpose_1in_1out_t func = <fptr_custom_elemwise_transpose_1in_1out_t > kernel_data.ptr
-    # call FPTR function
-    cdef c_dpctl.DPCTLSyclEventRef event_ref = func(q_ref,
-                                                    array1.get_data(),
-                                                    input_shape.data(),
-                                                    result_shape.data(),
-                                                    permute_axes.data(),
-                                                    input_shape_size,
-                                                    result.get_data(),
-                                                    array1.size,
-                                                    NULL)  # dep_events_ref
-
-    with nogil: c_dpctl.DPCTLEvent_WaitAndThrow(event_ref)
-    c_dpctl.DPCTLEvent_Delete(event_ref)
-
-    return result

dpnp/dpnp_array.py

@@ -99,15 +99,8 @@ def get_array(self):
 
     @property
     def T(self):
-        """Shape-reversed view of the array.
-
-        If ndim < 2, then this is just a reference to the array itself.
-
-        """
-        if self.ndim < 2:
-            return self
-        else:
-            return dpnp.transpose(self)
+        """View of the transposed array."""
+        return self.transpose()
 
     def to_device(self, target_device):
         """
@@ -1025,15 +1018,62 @@ def take(self, indices, axis=None, out=None, mode='raise'):
 
     def transpose(self, *axes):
         """
-        Returns a view of the array with axes permuted.
+        Returns a view of the array with axes transposed.
 
-        .. seealso::
-           :obj:`dpnp.transpose` for full documentation,
-           :meth:`numpy.ndarray.reshape`
+        For full documentation refer to :obj:`numpy.ndarray.transpose`.
+
+        Returns
+        -------
+        y : dpnp.ndarray
+            View of the array with its axes suitably permuted.
+
+        See Also
+        --------
+            :obj:`dpnp.transpose` : Equivalent function.
+            :obj:`dpnp.ndarray.ndarray.T` : Array property returning the array transposed.
+            :obj:`dpnp.ndarray.reshape` : Give a new shape to an array without changing its data.
 
+        Examples
+        --------
+        >>> import dpnp as dp
+        >>> a = dp.array([[1, 2], [3, 4]])
+        >>> a
+        array([[1, 2],
+               [3, 4]])
+        >>> a.transpose()
+        array([[1, 3],
+               [2, 4]])
+        >>> a.transpose((1, 0))
+        array([[1, 3],
+               [2, 4]])
+
+        >>> a = dp.array([1, 2, 3, 4])
+        >>> a
+        array([1, 2, 3, 4])
+        >>> a.transpose()
+        array([1, 2, 3, 4])
+            
         """
 
-        return dpnp.transpose(self, axes)
+        ndim = self.ndim
+        if ndim < 2:
+            return self
+
+        axes_len = len(axes)
+        if axes_len == 1 and isinstance(axes[0], tuple):
+            axes = axes[0]
+
+        res = self.__new__(dpnp_array)
+        if ndim == 2 and axes_len == 0:
+            res._array_obj = self._array_obj.T
+        else:
+            if len(axes) == 0 or axes[0] is None:
+                # self.transpose().shape == self.shape[::-1]
+                # self.transpose(None).shape == self.shape[::-1]
+                axes = tuple((ndim - x - 1) for x in range(ndim))
+
+            res._array_obj = dpt.permute_dims(self._array_obj, axes)
+        return res
 
     def var(self, axis=None, dtype=None, out=None, ddof=0, keepdims=False):
         """

dpnp/dpnp_iface_manipulation.py

@@ -673,54 +673,65 @@ def swapaxes(x1, axis1, axis2):
     return call_origin(numpy.swapaxes, x1, axis1, axis2)
 
 
-def transpose(x1, axes=None):
+def transpose(a, axes=None):
     """
-    Reverse or permute the axes of an array; returns the modified array.
+    Returns an array with axes transposed.
 
     For full documentation refer to :obj:`numpy.transpose`.
 
+    Returns
+    -------
+    y : dpnp.ndarray
+        `a` with its axes permuted. A view is returned whenever possible.
+
     Limitations
     -----------
-    Input array is supported as :obj:`dpnp.ndarray`.
-    Otherwise the function will be executed sequentially on CPU.
-    Value of the parameter ``axes`` likely to be replaced with ``None``.
-    Input array data types are limited by supported DPNP :ref:`Data types`.
+    Input array is supported as either :class:`dpnp.ndarray`
+    or :class:`dpctl.tensor.usm_ndarray`.
 
     See Also
     --------
+    :obj:`dpnp.ndarray.transpose` : Equivalent method.
     :obj:`dpnp.moveaxis` : Move array axes to new positions.
     :obj:`dpnp.argsort` : Returns the indices that would sort an array.
 
     Examples
     --------
-    >>> import dpnp as np
-    >>> x = np.arange(4).reshape((2,2))
-    >>> x.shape
-    (2, 2)
-    >>> [i for i in x]
-    [0, 1, 2, 3]
-    >>> out = np.transpose(x)
-    >>> out.shape
-    (2, 2)
-    >>> [i for i in out]
-    [0, 2, 1, 3]
-
-    """
-
-    x1_desc = dpnp.get_dpnp_descriptor(x1, copy_when_nondefault_queue=False)
-    if x1_desc:
-        if axes is not None:
-            if not any(axes):
-                """
-                pytest tests/third_party/cupy/manipulation_tests/test_transpose.py
-                """
-                axes = None
-
-        result = dpnp_transpose(x1_desc, axes).get_pyobj()
-
-        return result
-
-    return call_origin(numpy.transpose, x1, axes=axes)
+    >>> import dpnp as dp
+    >>> a = dp.array([[1, 2], [3, 4]])
+    >>> a
+    array([[1, 2],
+           [3, 4]])
+    >>> dp.transpose(a)
+    array([[1, 3],
+           [2, 4]])
+
+    >>> a = dp.array([1, 2, 3, 4])
+    >>> a
+    array([1, 2, 3, 4])
+    >>> dp.transpose(a)
+    array([1, 2, 3, 4])
+
+    >>> a = dp.ones((1, 2, 3))
+    >>> dp.transpose(a, (1, 0, 2)).shape
+    (2, 1, 3)
+
+    >>> a = dp.ones((2, 3, 4, 5))
+    >>> dp.transpose(a).shape
+    (5, 4, 3, 2)
+
+    """
+
+    if isinstance(a, dpnp_array):
+        array = a
+    elif isinstance(a, dpt.usm_ndarray):
+        array = dpnp_array._create_from_usm_ndarray(a.get_array())
+    else:
+        raise TypeError("An array must be any of supported type, but got {}".format(type(a)))
+
+    if axes is None:
+        return array.transpose()
+    return array.transpose(*axes)
 
 
 def unique(x1, **kwargs):

dpnp/dpnp_iface_statistics.py

@@ -290,8 +290,7 @@ def cov(x1, y=None, rowvar=True, bias=False, ddof=None, fweights=None, aweights=
         pass
     else:
         if not rowvar and x1.shape[0] != 1:
-            x1 = x1.get_array() if isinstance(x1, dpnp_array) else x1
-            x1 = dpnp_array._create_from_usm_ndarray(x1.mT)
+            x1 = x1.T
 
         if not x1.dtype in (dpnp.float32, dpnp.float64):
             x1 = dpnp.astype(x1, dpnp.default_float_type(sycl_queue=x1.sycl_queue))

tests/skipped_tests.tbl

@@ -232,8 +232,6 @@ tests/third_party/cupy/core_tests/test_ndarray_copy_and_view.py::TestArrayCopyAn
 tests/third_party/cupy/core_tests/test_ndarray_copy_and_view.py::TestArrayCopyAndView::test_astype_strides_swapped
 tests/third_party/cupy/core_tests/test_ndarray_copy_and_view.py::TestArrayCopyAndView::test_astype_type_c_contiguous_no_copy
 tests/third_party/cupy/core_tests/test_ndarray_copy_and_view.py::TestArrayCopyAndView::test_astype_type_f_contiguous_no_copy
-tests/third_party/cupy/core_tests/test_ndarray_copy_and_view.py::TestArrayCopyAndView::test_transposed_fill
-tests/third_party/cupy/core_tests/test_ndarray_copy_and_view.py::TestArrayCopyAndView::test_transposed_flatten
 tests/third_party/cupy/core_tests/test_ndarray_copy_and_view.py::TestArrayCopyAndView::test_flatten
 tests/third_party/cupy/core_tests/test_ndarray_copy_and_view.py::TestArrayCopyAndView::test_flatten_copied
 tests/third_party/cupy/core_tests/test_ndarray_copy_and_view.py::TestArrayCopyAndView::test_isinstance_numpy_copy
@@ -834,7 +832,6 @@ tests/third_party/cupy/math_tests/test_rounding.py::TestRounding::test_trunc
 tests/third_party/cupy/math_tests/test_sumprod.py::TestSumprod::test_sum_all
 tests/third_party/cupy/math_tests/test_sumprod.py::TestSumprod::test_sum_all2
 tests/third_party/cupy/math_tests/test_sumprod.py::TestSumprod::test_sum_all_keepdims
-tests/third_party/cupy/math_tests/test_sumprod.py::TestSumprod::test_sum_all_transposed
 tests/third_party/cupy/math_tests/test_sumprod.py::TestSumprod::test_sum_all_transposed2
 tests/third_party/cupy/math_tests/test_sumprod.py::TestSumprod::test_sum_axes
 tests/third_party/cupy/math_tests/test_sumprod.py::TestSumprod::test_sum_axes2
@@ -883,7 +880,6 @@ tests/third_party/cupy/math_tests/test_sumprod.py::TestNansumNanprodLong_param_1
 tests/third_party/cupy/math_tests/test_sumprod.py::TestNansumNanprodLong_param_9_{axis=0, func='nanprod', keepdims=True, shape=(2, 3, 4), transpose_axes=False}::test_nansum_all
 tests/third_party/cupy/math_tests/test_sumprod.py::TestNansumNanprodLong_param_9_{axis=0, func='nanprod', keepdims=True, shape=(2, 3, 4), transpose_axes=False}::test_nansum_axis_transposed
 tests/third_party/cupy/math_tests/test_sumprod.py::TestSumprod::test_sum_all2
-tests/third_party/cupy/math_tests/test_sumprod.py::TestSumprod::test_sum_all_transposed2
 tests/third_party/cupy/math_tests/test_trigonometric.py::TestUnwrap::test_unwrap_1dim
 tests/third_party/cupy/math_tests/test_trigonometric.py::TestUnwrap::test_unwrap_1dim_with_discont
 tests/third_party/cupy/math_tests/test_trigonometric.py::TestUnwrap::test_unwrap_2dim_with_axis