Add implementation of dpctl.tensor.not_equal

Author: vlad-perevezentsev

dpctl/tensor/__init__.py

@@ -100,6 +100,7 @@
     isfinite,
     isinf,
     isnan,
+    not_equal,
     sqrt,
 )
 
@@ -187,4 +188,5 @@
     "sqrt",
     "divide",
     "equal",
+    "not_equal",
 ]

dpctl/tensor/_elementwise_funcs.py

@@ -237,7 +237,27 @@
 # FIXME: implement U25
 
 # B20: ==== NOT_EQUAL   (x1, x2)
-# FIXME: implement B20
+_not_equal_docstring_ = """
+not_equal(x1, x2, order='K')
+
+Calculates inequality test results for each element `x1_i` of the
+input array `x1` with the respective element `x2_i` of the input array `x2`.
+
+Args:
+    x1 (usm_ndarray):
+        First input array, expected to have numeric data type.
+    x2 (usm_ndarray):
+        Second input array, also expected to have numeric data type.
+Returns:
+    usm_narray:
+        an array containing the result of element-wise inequality comparison.
+        The data type of the returned array is determined by the
+        Type Promotion Rules.
+"""
+
+not_equal = BinaryElementwiseFunc(
+    "not_equal", ti._not_equal_result_type, ti._not_equal, _not_equal_docstring_
+)
 
 # U26: ==== POSITIVE    (x)
 # FIXME: implement U26

dpctl/tensor/libtensor/include/kernels/elementwise_functions/not_equal.hpp

@@ -0,0 +1,287 @@
+//=== not_equal.hpp -   Binary function NOT_EQUAL        ------  *-C++-*--/===//
+//
+//                      Data Parallel Control (dpctl)
+//
+// Copyright 2020-2023 Intel Corporation
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//    http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+//===---------------------------------------------------------------------===//
+///
+/// \file
+/// This file defines kernels for elementwise evaluation of inequality of
+/// tensor elements.
+//===---------------------------------------------------------------------===//
+
+#pragma once
+#include <CL/sycl.hpp>
+#include <cstddef>
+#include <cstdint>
+#include <type_traits>
+
+#include "utils/offset_utils.hpp"
+#include "utils/type_dispatch.hpp"
+#include "utils/type_utils.hpp"
+
+#include "kernels/elementwise_functions/common.hpp"
+#include <pybind11/pybind11.h>
+
+namespace dpctl
+{
+namespace tensor
+{
+namespace kernels
+{
+namespace not_equal
+{
+
+namespace py = pybind11;
+namespace td_ns = dpctl::tensor::type_dispatch;
+namespace tu_ns = dpctl::tensor::type_utils;
+
+template <typename argT1, typename argT2, typename resT> struct NotEqualFunctor
+{
+    static_assert(std::is_same_v<resT, bool>);
+
+    using supports_sg_loadstore = std::negation<
+        std::disjunction<tu_ns::is_complex<argT1>, tu_ns::is_complex<argT2>>>;
+    using supports_vec = std::conjunction<
+        std::is_same<argT1, argT2>,
+        std::negation<std::disjunction<tu_ns::is_complex<argT1>,
+                                       tu_ns::is_complex<argT2>>>>;
+
+    resT operator()(const argT1 &in1, const argT2 &in2)
+    {
+        return (in1 != in2);
+    }
+
+    template <int vec_sz>
+    sycl::vec<resT, vec_sz> operator()(const sycl::vec<argT1, vec_sz> &in1,
+                                       const sycl::vec<argT2, vec_sz> &in2)
+    {
+        auto tmp = (in1 != in2);
+        if constexpr (std::is_same_v<resT,
+                                     typename decltype(tmp)::element_type>) {
+            return tmp;
+        }
+        else {
+            using dpctl::tensor::type_utils::vec_cast;
+
+            return vec_cast<resT, typename decltype(tmp)::element_type, vec_sz>(
+                tmp);
+        }
+    }
+};
+
+template <typename argT1,
+          typename argT2,
+          typename resT,
+          unsigned int vec_sz = 4,
+          unsigned int n_vecs = 2>
+using NotEqualContigFunctor =
+    elementwise_common::BinaryContigFunctor<argT1,
+                                            argT2,
+                                            resT,
+                                            NotEqualFunctor<argT1, argT2, resT>,
+                                            vec_sz,
+                                            n_vecs>;
+
+template <typename argT1, typename argT2, typename resT, typename IndexerT>
+using NotEqualStridedFunctor =
+    elementwise_common::BinaryStridedFunctor<argT1,
+                                             argT2,
+                                             resT,
+                                             IndexerT,
+                                             NotEqualFunctor<argT1, argT2, resT>>;
+
+template <typename T1, typename T2> struct NotEqualOutputType
+{
+    using value_type = typename std::disjunction< // disjunction is C++17
+                                                  // feature, supported by DPC++
+        td_ns::BinaryTypeMapResultEntry<T1, bool, T2, bool, bool>,
+        td_ns::
+            BinaryTypeMapResultEntry<T1, std::uint8_t, T2, std::uint8_t, bool>,
+        td_ns::BinaryTypeMapResultEntry<T1, std::int8_t, T2, std::int8_t, bool>,
+        td_ns::BinaryTypeMapResultEntry<T1,
+                                        std::uint16_t,
+                                        T2,
+                                        std::uint16_t,
+                                        bool>,
+        td_ns::
+            BinaryTypeMapResultEntry<T1, std::int16_t, T2, std::int16_t, bool>,
+        td_ns::BinaryTypeMapResultEntry<T1,
+                                        std::uint32_t,
+                                        T2,
+                                        std::uint32_t,
+                                        bool>,
+        td_ns::
+            BinaryTypeMapResultEntry<T1, std::int32_t, T2, std::int32_t, bool>,
+        td_ns::BinaryTypeMapResultEntry<T1,
+                                        std::uint64_t,
+                                        T2,
+                                        std::uint64_t,
+                                        bool>,
+        td_ns::
+            BinaryTypeMapResultEntry<T1, std::int64_t, T2, std::int64_t, bool>,
+        td_ns::BinaryTypeMapResultEntry<T1, sycl::half, T2, sycl::half, bool>,
+        td_ns::BinaryTypeMapResultEntry<T1, float, T2, float, bool>,
+        td_ns::BinaryTypeMapResultEntry<T1, double, T2, double, bool>,
+        td_ns::BinaryTypeMapResultEntry<T1,
+                                        std::complex<float>,
+                                        T2,
+                                        std::complex<float>,
+                                        bool>,
+        td_ns::BinaryTypeMapResultEntry<T1,
+                                        std::complex<double>,
+                                        T2,
+                                        std::complex<double>,
+                                        bool>,
+        td_ns::DefaultResultEntry<void>>::result_type;
+};
+
+template <typename argT1,
+          typename argT2,
+          typename resT,
+          unsigned int vec_sz,
+          unsigned int n_vecs>
+class not_equal_contig_kernel;
+
+template <typename argTy1, typename argTy2>
+sycl::event not_equal_contig_impl(sycl::queue exec_q,
+                              size_t nelems,
+                              const char *arg1_p,
+                              py::ssize_t arg1_offset,
+                              const char *arg2_p,
+                              py::ssize_t arg2_offset,
+                              char *res_p,
+                              py::ssize_t res_offset,
+                              const std::vector<sycl::event> &depends = {})
+{
+    sycl::event comp_ev = exec_q.submit([&](sycl::handler &cgh) {
+        cgh.depends_on(depends);
+
+        size_t lws = 64;
+        constexpr unsigned int vec_sz = 4;
+        constexpr unsigned int n_vecs = 2;
+        const size_t n_groups =
+            ((nelems + lws * n_vecs * vec_sz - 1) / (lws * n_vecs * vec_sz));
+        const auto gws_range = sycl::range<1>(n_groups * lws);
+        const auto lws_range = sycl::range<1>(lws);
+
+        using resTy = typename NotEqualOutputType<argTy1, argTy2>::value_type;
+
+        const argTy1 *arg1_tp =
+            reinterpret_cast<const argTy1 *>(arg1_p) + arg1_offset;
+        const argTy2 *arg2_tp =
+            reinterpret_cast<const argTy2 *>(arg2_p) + arg2_offset;
+        resTy *res_tp = reinterpret_cast<resTy *>(res_p) + res_offset;
+
+        cgh.parallel_for<
+            not_equal_contig_kernel<argTy1, argTy2, resTy, vec_sz, n_vecs>>(
+            sycl::nd_range<1>(gws_range, lws_range),
+            NotEqualContigFunctor<argTy1, argTy2, resTy, vec_sz, n_vecs>(
+                arg1_tp, arg2_tp, res_tp, nelems));
+    });
+    return comp_ev;
+}
+
+template <typename fnT, typename T1, typename T2> struct NotEqualContigFactory
+{
+    fnT get()
+    {
+        if constexpr (std::is_same_v<
+                         typename NotEqualOutputType<T1, T2>::value_type, void>)
+        {
+            fnT fn = nullptr;
+            return fn;
+        }
+        else {
+            fnT fn = not_equal_contig_impl<T1, T2>;
+            return fn;
+        }
+    }
+};
+
+template <typename fnT, typename T1, typename T2> struct NotEqualTypeMapFactory
+{
+    /*! @brief get typeid for output type of operator()==(x, y), always bool */
+    std::enable_if_t<std::is_same<fnT, int>::value, int> get()
+    {
+        using rT = typename NotEqualOutputType<T1, T2>::value_type;
+        return td_ns::GetTypeid<rT>{}.get();
+    }
+};
+
+template <typename T1, typename T2, typename resT, typename IndexerT>
+class not_equal_strided_strided_kernel;
+
+template <typename argTy1, typename argTy2>
+sycl::event
+not_equal_strided_impl(sycl::queue exec_q,
+                   size_t nelems,
+                   int nd,
+                   const py::ssize_t *shape_and_strides,
+                   const char *arg1_p,
+                   py::ssize_t arg1_offset,
+                   const char *arg2_p,
+                   py::ssize_t arg2_offset,
+                   char *res_p,
+                   py::ssize_t res_offset,
+                   const std::vector<sycl::event> &depends,
+                   const std::vector<sycl::event> &additional_depends)
+{
+    sycl::event comp_ev = exec_q.submit([&](sycl::handler &cgh) {
+        cgh.depends_on(depends);
+        cgh.depends_on(additional_depends);
+
+        using resTy = typename NotEqualOutputType<argTy1, argTy2>::value_type;
+
+        using IndexerT =
+            typename dpctl::tensor::offset_utils::ThreeOffsets_StridedIndexer;
+
+        IndexerT indexer{nd, arg1_offset, arg2_offset, res_offset,
+                         shape_and_strides};
+
+        const argTy1 *arg1_tp = reinterpret_cast<const argTy1 *>(arg1_p);
+        const argTy2 *arg2_tp = reinterpret_cast<const argTy2 *>(arg2_p);
+        resTy *res_tp = reinterpret_cast<resTy *>(res_p);
+
+        cgh.parallel_for<
+            not_equal_strided_strided_kernel<argTy1, argTy2, resTy, IndexerT>>(
+            {nelems}, NotEqualStridedFunctor<argTy1, argTy2, resTy, IndexerT>(
+                          arg1_tp, arg2_tp, res_tp, indexer));
+    });
+    return comp_ev;
+}
+
+template <typename fnT, typename T1, typename T2> struct NotEqualStridedFactory
+{
+    fnT get()
+    {
+        if constexpr (std::is_same_v<
+                         typename NotEqualOutputType<T1, T2>::value_type, void>)
+        {
+            fnT fn = nullptr;
+            return fn;
+        }
+        else {
+            fnT fn = not_equal_strided_impl<T1, T2>;
+            return fn;
+        }
+    }
+};
+
+} // namespace not_equal
+} // namespace kernels
+} // namespace tensor
+} // namespace dpctl