Update dpnp.power using dpctl and OneMKL implementations

dpnp/backend/extensions/vm/pow.hpp

@@ -0,0 +1,81 @@
+//*****************************************************************************
+// Copyright (c) 2023, Intel Corporation
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are met:
+// - Redistributions of source code must retain the above copyright notice,
+//   this list of conditions and the following disclaimer.
+// - Redistributions in binary form must reproduce the above copyright notice,
+//   this list of conditions and the following disclaimer in the documentation
+//   and/or other materials provided with the distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+// ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+// LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+// CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+// SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+// CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+// ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
+// THE POSSIBILITY OF SUCH DAMAGE.
+//*****************************************************************************
+
+#pragma once
+
+#include <CL/sycl.hpp>
+
+#include "common.hpp"
+#include "types_matrix.hpp"
+
+namespace dpnp
+{
+namespace backend
+{
+namespace ext
+{
+namespace vm
+{
+template <typename T>
+sycl::event pow_contig_impl(sycl::queue exec_q,
+                            const std::int64_t n,
+                            const char *in_a,
+                            const char *in_b,
+                            char *out_y,
+                            const std::vector<sycl::event> &depends)
+{
+    type_utils::validate_type_for_device<T>(exec_q);
+
+    const T *a = reinterpret_cast<const T *>(in_a);
+    const T *b = reinterpret_cast<const T *>(in_b);
+    T *y = reinterpret_cast<T *>(out_y);
+
+    return mkl_vm::pow(exec_q,
+                       n, // number of elements to be calculated
+                       a, // pointer `a` containing 1st input vector of size n
+                       b, // pointer `b` containing 2nd input vector of size n
+                       y, // pointer `y` to the output vector of size n
+                       depends);
+}
+
+template <typename fnT, typename T>
+struct PowContigFactory
+{
+    fnT get()
+    {
+        if constexpr (std::is_same_v<
+                          typename types::PowOutputType<T>::value_type, void>)
+        {
+            return nullptr;
+        }
+        else {
+            return pow_contig_impl<T>;
+        }
+    }
+};
+} // namespace vm
+} // namespace ext
+} // namespace backend
+} // namespace dpnp

dpnp/backend/extensions/vm/types_matrix.hpp

@@ -106,6 +106,31 @@ struct LnOutputType
         dpctl_td_ns::DefaultResultEntry<void>>::result_type;
 };
 
+/**
+ * @brief A factory to define pairs of supported types for which
+ * MKL VM library provides support in oneapi::mkl::vm::pow<T> function.
+ *
+ * @tparam T Type of input vectors `a` and `b` and of result vector `y`.
+ */
+template <typename T>
+struct PowOutputType
+{
+    using value_type = typename std::disjunction<
+        dpctl_td_ns::BinaryTypeMapResultEntry<T,
+                                              std::complex<double>,
+                                              T,
+                                              std::complex<double>,
+                                              std::complex<double>>,
+        dpctl_td_ns::BinaryTypeMapResultEntry<T,
+                                              std::complex<float>,
+                                              T,
+                                              std::complex<float>,
+                                              std::complex<float>>,
+        dpctl_td_ns::BinaryTypeMapResultEntry<T, double, T, double, double>,
+        dpctl_td_ns::BinaryTypeMapResultEntry<T, float, T, float, float>,
+        dpctl_td_ns::DefaultResultEntry<void>>::result_type;
+};
+
 /**
  * @brief A factory to define pairs of supported types for which
  * MKL VM library provides support in oneapi::mkl::vm::sin<T> function.

dpnp/backend/extensions/vm/vm_py.cpp

@@ -34,6 +34,7 @@
 #include "cos.hpp"
 #include "div.hpp"
 #include "ln.hpp"
+#include "pow.hpp"
 #include "sin.hpp"
 #include "sqr.hpp"
 #include "sqrt.hpp"
@@ -46,6 +47,7 @@ using vm_ext::binary_impl_fn_ptr_t;
 using vm_ext::unary_impl_fn_ptr_t;
 
 static binary_impl_fn_ptr_t div_dispatch_vector[dpctl_td_ns::num_types];
+static binary_impl_fn_ptr_t pow_dispatch_vector[dpctl_td_ns::num_types];
 
 static unary_impl_fn_ptr_t cos_dispatch_vector[dpctl_td_ns::num_types];
 static unary_impl_fn_ptr_t ln_dispatch_vector[dpctl_td_ns::num_types];
@@ -144,6 +146,36 @@ PYBIND11_MODULE(_vm_impl, m)
               py::arg("sycl_queue"), py::arg("src"), py::arg("dst"));
     }
 
+    // BinaryUfunc: ==== Pow(x1, x2) ====
+    {
+        vm_ext::init_ufunc_dispatch_vector<binary_impl_fn_ptr_t,
+                                           vm_ext::PowContigFactory>(
+            pow_dispatch_vector);
+
+        auto pow_pyapi = [&](sycl::queue exec_q, arrayT src1, arrayT src2,
+                             arrayT dst, const event_vecT &depends = {}) {
+            return vm_ext::binary_ufunc(exec_q, src1, src2, dst, depends,
+                                        pow_dispatch_vector);
+        };
+        m.def("_pow", pow_pyapi,
+              "Call `pow` function from OneMKL VM library to performs element "
+              "by element exponentiation of vector `src1` raised to the power "
+              "of vector `src2` to resulting vector `dst`",
+              py::arg("sycl_queue"), py::arg("src1"), py::arg("src2"),
+              py::arg("dst"), py::arg("depends") = py::list());
+
+        auto pow_need_to_call_pyapi = [&](sycl::queue exec_q, arrayT src1,
+                                          arrayT src2, arrayT dst) {
+            return vm_ext::need_to_call_binary_ufunc(exec_q, src1, src2, dst,
+                                                     pow_dispatch_vector);
+        };
+        m.def("_mkl_pow_to_call", pow_need_to_call_pyapi,
+              "Check input arguments to answer if `pow` function from "
+              "OneMKL VM library can be used",
+              py::arg("sycl_queue"), py::arg("src1"), py::arg("src2"),
+              py::arg("dst"));
+    }
+
     // UnaryUfunc: ==== Sin(x) ====
     {
         vm_ext::init_ufunc_dispatch_vector<unary_impl_fn_ptr_t,

dpnp/backend/include/dpnp_iface_fptr.hpp

@@ -312,8 +312,6 @@ enum class DPNPFuncName : size_t
                               parameters */
     DPNP_FN_PLACE,         /**< Used in numpy.place() impl  */
     DPNP_FN_POWER,         /**< Used in numpy.power() impl  */
-    DPNP_FN_POWER_EXT,     /**< Used in numpy.power() impl, requires extra
-                              parameters */
     DPNP_FN_PROD,          /**< Used in numpy.prod() impl  */
     DPNP_FN_PROD_EXT, /**< Used in numpy.prod() impl, requires extra parameters
                        */

dpnp/backend/kernels/dpnp_krnl_elemwise.cpp

@@ -1628,13 +1628,6 @@ static void func_map_elemwise_2arg_3type_core(func_map_t &fmap)
                func_type_map_t::find_type<FT1>,
                func_type_map_t::find_type<FTs>>}),
      ...);
-    ((fmap[DPNPFuncName::DPNP_FN_POWER_EXT][FT1][FTs] =
-          {populate_func_types<FT1, FTs>(),
-           (void *)dpnp_power_c_ext<
-               func_type_map_t::find_type<populate_func_types<FT1, FTs>()>,
-               func_type_map_t::find_type<FT1>,
-               func_type_map_t::find_type<FTs>>}),
-     ...);
     ((fmap[DPNPFuncName::DPNP_FN_SUBTRACT_EXT][FT1][FTs] =
           {populate_func_types<FT1, FTs>(),
            (void *)dpnp_subtract_c_ext<

dpnp/dpnp_algo/dpnp_algo.pxd

@@ -190,8 +190,6 @@ cdef extern from "dpnp_iface_fptr.hpp" namespace "DPNPFuncName":  # need this na
         DPNP_FN_PARTITION
         DPNP_FN_PARTITION_EXT
         DPNP_FN_PLACE
-        DPNP_FN_POWER
-        DPNP_FN_POWER_EXT
         DPNP_FN_PROD
         DPNP_FN_PROD_EXT
         DPNP_FN_PTP
@@ -490,8 +488,6 @@ cpdef dpnp_descriptor dpnp_maximum(dpnp_descriptor x1_obj, dpnp_descriptor x2_ob
 cpdef dpnp_descriptor dpnp_minimum(dpnp_descriptor x1_obj, dpnp_descriptor x2_obj, object dtype=*,
                                    dpnp_descriptor out=*, object where=*)
 cpdef dpnp_descriptor dpnp_negative(dpnp_descriptor array1)
-cpdef dpnp_descriptor dpnp_power(dpnp_descriptor x1_obj, dpnp_descriptor x2_obj, object dtype=*,
-                                 dpnp_descriptor out=*, object where=*)
 cpdef dpnp_descriptor dpnp_remainder(dpnp_descriptor x1_obj, dpnp_descriptor x2_obj, object dtype=*,
                                      dpnp_descriptor out=*, object where=*)
 

dpnp/dpnp_algo/dpnp_algo_mathematical.pxi

@@ -60,7 +60,6 @@ __all__ += [
     "dpnp_nanprod",
     "dpnp_nansum",
     "dpnp_negative",
-    "dpnp_power",
     "dpnp_prod",
     "dpnp_remainder",
     "dpnp_sign",
@@ -476,14 +475,6 @@ cpdef utils.dpnp_descriptor dpnp_negative(dpnp_descriptor x1):
     return call_fptr_1in_1out_strides(DPNP_FN_NEGATIVE_EXT, x1)
 
 
-cpdef utils.dpnp_descriptor dpnp_power(utils.dpnp_descriptor x1_obj,
-                                       utils.dpnp_descriptor x2_obj,
-                                       object dtype=None,
-                                       utils.dpnp_descriptor out=None,
-                                       object where=True):
-    return call_fptr_2in_1out_strides(DPNP_FN_POWER_EXT, x1_obj, x2_obj, dtype, out, where, func_name="power")
-
-
 cpdef utils.dpnp_descriptor dpnp_prod(utils.dpnp_descriptor x1,
                                       object axis=None,
                                       object dtype=None,

dpnp/dpnp_algo/dpnp_elementwise_common.py

@@ -58,6 +58,7 @@
     "dpnp_logical_xor",
     "dpnp_multiply",
     "dpnp_not_equal",
+    "dpnp_power",
     "dpnp_sin",
     "dpnp_sqrt",
     "dpnp_square",
@@ -613,7 +614,7 @@ def _call_log(src, dst, sycl_queue, depends=None):
         Memory layout of the newly output array, if parameter `out` is `None`.
         Default: "K".
 Returns:
-    usm_narray:
+    dpnp.ndarray:
         An array containing the element-wise logical AND results.
 """
 
@@ -649,7 +650,7 @@ def dpnp_logical_and(x1, x2, out=None, order="K"):
         output array, if parameter `out` is `None`.
         Default: "K".
 Return:
-    usm_ndarray:
+    dpnp.ndarray:
         An array containing the element-wise logical NOT results.
 """
 
@@ -689,7 +690,7 @@ def dpnp_logical_not(x, out=None, order="K"):
         Memory layout of the newly output array, if parameter `out` is `None`.
         Default: "K".
 Returns:
-    usm_narray:
+    dpnp.ndarray:
         An array containing the element-wise logical OR results.
 """
 
@@ -730,7 +731,7 @@ def dpnp_logical_or(x1, x2, out=None, order="K"):
         Memory layout of the newly output array, if parameter `out` is `None`.
         Default: "K".
 Returns:
-    usm_narray:
+    dpnp.ndarray:
         An array containing the element-wise logical XOR results.
 """
 
@@ -844,6 +845,60 @@ def dpnp_not_equal(x1, x2, out=None, order="K"):
     return dpnp_array._create_from_usm_ndarray(res_usm)
 
 
+_power_docstring_ = """
+power(x1, x2, out=None, order="K")
+Calculates `x1_i` raised to `x2_i` for each element `x1_i` of the input array
+`x1` with the respective element `x2_i` of the input array `x2`.
+Args:
+    x1 (dpnp.ndarray):
+        First input array, expected to have numeric data type.
+    x2 (dpnp.ndarray):
+        Second input array, also expected to have numeric data type.
+    out ({None, dpnp.ndarray}, optional):
+        Output array to populate. Array must have the correct
+        shape and the expected data type.
+    order ("C","F","A","K", None, optional):
+        Output array, if parameter `out` is `None`.
+        Default: "K".
+Returns:
+    dpnp.ndarray:
+        An array containing the result of element-wise of raising each element
+        to a specified power.
+        The data type of the returned array is determined by the Type Promotion Rules.
+"""
+
+
+def dpnp_power(x1, x2, out=None, order="K"):
+    """
+    Invokes pow() function from pybind11 extension of OneMKL VM if possible.
+
+    Otherwise fully relies on dpctl.tensor implementation for pow() function.
+
+    """
+
+    def _call_pow(src1, src2, dst, sycl_queue, depends=None):
+        """A callback to register in BinaryElementwiseFunc class of dpctl.tensor"""
+
+        if depends is None:
+            depends = []
+
+        if vmi._mkl_pow_to_call(sycl_queue, src1, src2, dst):
+            # call pybind11 extension for pow() function from OneMKL VM
+            return vmi._pow(sycl_queue, src1, src2, dst, depends)
+        return ti._pow(src1, src2, dst, sycl_queue, depends)
+
+    # dpctl.tensor only works with usm_ndarray or scalar
+    x1_usm_or_scalar = dpnp.get_usm_ndarray_or_scalar(x1)
+    x2_usm_or_scalar = dpnp.get_usm_ndarray_or_scalar(x2)
+    out_usm = None if out is None else dpnp.get_usm_ndarray(out)
+
+    func = BinaryElementwiseFunc(
+        "pow", ti._pow_result_type, _call_pow, _power_docstring_
+    )
+    res_usm = func(x1_usm_or_scalar, x2_usm_or_scalar, out=out_usm, order=order)
+    return dpnp_array._create_from_usm_ndarray(res_usm)
+
+
 _sin_docstring = """
 sin(x, out=None, order='K')
 Computes sine for each element `x_i` of input array `x`.