Add pow call from OneMKL by pybind11 extension

Author: vlad-perevezentsev

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/dpnp_algo/dpnp_elementwise_common.py

@@ -869,15 +869,31 @@ def dpnp_not_equal(x1, x2, out=None, order="K"):
 
 
 def dpnp_power(x1, x2, out=None, order="K"):
-    """Invokes pow() from dpctl.tensor implementation for power() function."""
+    """
+    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, ti._pow, _power_docstring_
+        "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)