|
| 1 | +//***************************************************************************** |
| 2 | +// Copyright (c) 2025, Intel Corporation |
| 3 | +// All rights reserved. |
| 4 | +// |
| 5 | +// Redistribution and use in source and binary forms, with or without |
| 6 | +// modification, are permitted provided that the following conditions are met: |
| 7 | +// - Redistributions of source code must retain the above copyright notice, |
| 8 | +// this list of conditions and the following disclaimer. |
| 9 | +// - Redistributions in binary form must reproduce the above copyright notice, |
| 10 | +// this list of conditions and the following disclaimer in the documentation |
| 11 | +// and/or other materials provided with the distribution. |
| 12 | +// |
| 13 | +// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" |
| 14 | +// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 15 | +// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 16 | +// ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE |
| 17 | +// LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR |
| 18 | +// CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF |
| 19 | +// SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS |
| 20 | +// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN |
| 21 | +// CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) |
| 22 | +// ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF |
| 23 | +// THE POSSIBILITY OF SUCH DAMAGE. |
| 24 | +//***************************************************************************** |
| 25 | + |
| 26 | +#include "piecewise.hpp" |
| 27 | + |
| 28 | +#include "utils/output_validation.hpp" |
| 29 | +#include "utils/type_dispatch.hpp" |
| 30 | +#include "utils/type_utils.hpp" |
| 31 | + |
| 32 | +#include <pybind11/numpy.h> |
| 33 | +#include <sycl/sycl.hpp> |
| 34 | + |
| 35 | +namespace dpnp::extensions::functional |
| 36 | +{ |
| 37 | +namespace dpctl_td_ns = dpctl::tensor::type_dispatch; |
| 38 | + |
| 39 | +typedef sycl::event (*piecewise_fn_ptr_t)(sycl::queue &, |
| 40 | + const py::object &, |
| 41 | + const std::size_t, |
| 42 | + const char *, |
| 43 | + char *, |
| 44 | + const std::vector<sycl::event> &); |
| 45 | + |
| 46 | +static piecewise_fn_ptr_t piecewise_dispatch_vector[dpctl_td_ns::num_types]; |
| 47 | + |
| 48 | +template <typename T> |
| 49 | +class PiecewiseFunctor |
| 50 | +{ |
| 51 | +private: |
| 52 | + const T val; |
| 53 | + const bool *cond = nullptr; |
| 54 | + T *res = nullptr; |
| 55 | + |
| 56 | +public: |
| 57 | + PiecewiseFunctor(const T val, const bool *cond, T *res) |
| 58 | + : val(val), cond(cond), res(res) |
| 59 | + { |
| 60 | + } |
| 61 | + |
| 62 | + void operator()(sycl::id<1> id) const |
| 63 | + { |
| 64 | + const auto i = id.get(0); |
| 65 | + if (cond[i]) { |
| 66 | + res[i] = val; |
| 67 | + } |
| 68 | + } |
| 69 | +}; |
| 70 | + |
| 71 | +template <typename T> |
| 72 | +sycl::event piecewise_impl(sycl::queue &exec_q, |
| 73 | + const py::object &value, |
| 74 | + const std::size_t nelems, |
| 75 | + const char *condition, |
| 76 | + char *result, |
| 77 | + const std::vector<sycl::event> &depends) |
| 78 | +{ |
| 79 | + dpctl::tensor::type_utils::validate_type_for_device<T>(exec_q); |
| 80 | + |
| 81 | + py::object type_obj = py::type::of(value); |
| 82 | + std::string type_name = py::str(type_obj.attr("__name__")); |
| 83 | + |
| 84 | + T *res = reinterpret_cast<T *>(result); |
| 85 | + const bool *cond = reinterpret_cast<const bool *>(condition); |
| 86 | + T val = py::cast<const T>(value); |
| 87 | + |
| 88 | + sycl::event piecewise_ev = exec_q.submit([&](sycl::handler &cgh) { |
| 89 | + cgh.depends_on(depends); |
| 90 | + |
| 91 | + using PiecewiseKernel = PiecewiseFunctor<T>; |
| 92 | + cgh. parallel_for<PiecewiseKernel>(sycl::range< 1>(nelems), |
| 93 | + PiecewiseKernel(val, cond, res)); |
| 94 | + }); |
| 95 | + |
| 96 | + return piecewise_ev; |
| 97 | +} |
| 98 | + |
| 99 | +/** |
| 100 | + * @brief A factory to define pairs of supported types for which |
| 101 | + * piecewise function is available. |
| 102 | + * |
| 103 | + * @tparam T Type of input vector `a` and of result vector `y`. |
| 104 | + */ |
| 105 | +template <typename T> |
| 106 | +struct PiecewiseOutputType |
| 107 | +{ |
| 108 | + using value_type = typename std::disjunction< |
| 109 | + dpctl_td_ns::TypeMapResultEntry<T, bool>, |
| 110 | + dpctl_td_ns::TypeMapResultEntry<T, std::uint8_t>, |
| 111 | + dpctl_td_ns::TypeMapResultEntry<T, std::int8_t>, |
| 112 | + dpctl_td_ns::TypeMapResultEntry<T, std::uint16_t>, |
| 113 | + dpctl_td_ns::TypeMapResultEntry<T, std::int16_t>, |
| 114 | + dpctl_td_ns::TypeMapResultEntry<T, std::uint32_t>, |
| 115 | + dpctl_td_ns::TypeMapResultEntry<T, std::int32_t>, |
| 116 | + dpctl_td_ns::TypeMapResultEntry<T, std::uint64_t>, |
| 117 | + dpctl_td_ns::TypeMapResultEntry<T, std::int64_t>, |
| 118 | + dpctl_td_ns::TypeMapResultEntry<T, sycl::half>, |
| 119 | + dpctl_td_ns::TypeMapResultEntry<T, float>, |
| 120 | + dpctl_td_ns::TypeMapResultEntry<T, double>, |
| 121 | + dpctl_td_ns::TypeMapResultEntry<T, std::complex<float>>, |
| 122 | + dpctl_td_ns::TypeMapResultEntry<T, std::complex<double>>, |
| 123 | + dpctl_td_ns::DefaultResultEntry<void>>::result_type; |
| 124 | +}; |
| 125 | + |
| 126 | +template <typename fnT, typename T> |
| 127 | +struct PiecewiseFactory |
| 128 | +{ |
| 129 | + fnT get() |
| 130 | + { |
| 131 | + if constexpr (std::is_same_v< |
| 132 | + typename PiecewiseOutputType<T>::value_type, void>) { |
| 133 | + return nullptr; |
| 134 | + } |
| 135 | + else { |
| 136 | + return piecewise_impl<T>; |
| 137 | + } |
| 138 | + } |
| 139 | +}; |
| 140 | + |
| 141 | +std::pair<sycl::event, sycl::event> |
| 142 | + py_piecewise(sycl::queue &exec_q, |
| 143 | + const py::object &value, |
| 144 | + const dpctl::tensor::usm_ndarray &condition, |
| 145 | + const dpctl::tensor::usm_ndarray &result, |
| 146 | + const std::vector<sycl::event> &depends) |
| 147 | +{ |
| 148 | + dpctl::tensor::validation::CheckWritable::throw_if_not_writable(result); |
| 149 | + |
| 150 | + const int res_nd = result.get_ndim(); |
| 151 | + const int cond_nd = condition.get_ndim(); |
| 152 | + if (res_nd != cond_nd) { |
| 153 | + throw py::value_error( |
| 154 | + "Condition and result arrays must have the same dimension."); |
| 155 | + } |
| 156 | + |
| 157 | + if (!dpctl::utils::queues_are_compatible( |
| 158 | + exec_q, {condition.get_queue(), result.get_queue()})) |
| 159 | + { |
| 160 | + throw py::value_error( |
| 161 | + "Execution queue is not compatible with allocation queue."); |
| 162 | + } |
| 163 | + |
| 164 | + const bool is_result_c_contig = result.is_c_contiguous(); |
| 165 | + if (!is_result_c_contig) { |
| 166 | + throw py::value_error("The result array is not c-contiguous."); |
| 167 | + } |
| 168 | + |
| 169 | + const py::ssize_t *res_shape = result.get_shape_raw(); |
| 170 | + const py::ssize_t *cond_shape = condition.get_shape_raw(); |
| 171 | + |
| 172 | + const bool shapes_equal = |
| 173 | + std::equal(res_shape, res_shape + res_nd, cond_shape); |
| 174 | + if (!shapes_equal) { |
| 175 | + throw py::value_error( |
| 176 | + "Condition and result arrays must have the same shape."); |
| 177 | + } |
| 178 | + |
| 179 | + const std::size_t nelems = result.get_size(); |
| 180 | + if (nelems == 0) { |
| 181 | + return std::make_pair(sycl::event{}, sycl::event{}); |
| 182 | + } |
| 183 | + |
| 184 | + const int result_typenum = result.get_typenum(); |
| 185 | + auto array_types = dpctl_td_ns::usm_ndarray_types(); |
| 186 | + const int result_type_id = array_types.typenum_to_lookup_id(result_typenum); |
| 187 | + auto piecewise_fn = piecewise_dispatch_vector[result_type_id]; |
| 188 | + |
| 189 | + if (piecewise_fn == nullptr) { |
| 190 | + throw std::runtime_error("Type of given array is not supported"); |
| 191 | + } |
| 192 | + |
| 193 | + const char *condition_typeless_ptr = condition.get_data(); |
| 194 | + char *result_typeless_ptr = result.get_data(); |
| 195 | + |
| 196 | + sycl::event piecewise_ev = |
| 197 | + piecewise_fn(exec_q, value, nelems, condition_typeless_ptr, |
| 198 | + result_typeless_ptr, depends); |
| 199 | + sycl::event args_ev = |
| 200 | + dpctl::utils::keep_args_alive(exec_q, {result}, {piecewise_ev}); |
| 201 | + |
| 202 | + return std::make_pair(args_ev, piecewise_ev); |
| 203 | +} |
| 204 | + |
| 205 | +void init_piecewise_dispatch_vectors(void) |
| 206 | +{ |
| 207 | + dpctl_td_ns::DispatchVectorBuilder<piecewise_fn_ptr_t, PiecewiseFactory, |
| 208 | + dpctl_td_ns::num_types> |
| 209 | + contig; |
| 210 | + contig.populate_dispatch_vector(piecewise_dispatch_vector); |
| 211 | + |
| 212 | + return; |
| 213 | +} |
| 214 | + |
| 215 | +} // namespace dpnp::extensions::functional |
0 commit comments