|
| 1 | +/* |
| 2 | + * Copyright (c) 2025, NVIDIA CORPORATION. |
| 3 | + * |
| 4 | + * Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 | + * you may not use this file except in compliance with the License. |
| 6 | + * You may obtain a copy of the License at |
| 7 | + * |
| 8 | + * http://www.apache.org/licenses/LICENSE-2.0 |
| 9 | + * |
| 10 | + * Unless required by applicable law or agreed to in writing, software |
| 11 | + * distributed under the License is distributed on an "AS IS" BASIS, |
| 12 | + * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 | + * See the License for the specific language governing permissions and |
| 14 | + * limitations under the License. |
| 15 | + */ |
| 16 | +#include <cudf/column/column_factories.hpp> |
| 17 | +#include <cudf/detail/nvtx/ranges.hpp> |
| 18 | +#include <cudf/detail/utilities/algorithm.cuh> |
| 19 | +#include <cudf/hashing/detail/hashing.hpp> |
| 20 | +#include <cudf/hashing/detail/xxhash_32.cuh> |
| 21 | +#include <cudf/table/table_device_view.cuh> |
| 22 | +#include <cudf/utilities/memory_resource.hpp> |
| 23 | +#include <cudf/utilities/span.hpp> |
| 24 | + |
| 25 | +#include <rmm/cuda_stream_view.hpp> |
| 26 | +#include <rmm/exec_policy.hpp> |
| 27 | + |
| 28 | +#include <cuda/std/limits> |
| 29 | +#include <thrust/tabulate.h> |
| 30 | + |
| 31 | +namespace cudf { |
| 32 | +namespace hashing { |
| 33 | +namespace detail { |
| 34 | + |
| 35 | +namespace { |
| 36 | + |
| 37 | +/** |
| 38 | + * @brief Computes the hash value of a row in the given table. |
| 39 | + * |
| 40 | + * @tparam Nullate A cudf::nullate type describing whether to check for nulls. |
| 41 | + */ |
| 42 | +template <typename Nullate> |
| 43 | +class device_row_hasher { |
| 44 | + public: |
| 45 | + device_row_hasher(Nullate nulls, table_device_view const& t, hash_value_type seed) |
| 46 | + : _check_nulls(nulls), _table(t), _seed(seed) |
| 47 | + { |
| 48 | + } |
| 49 | + |
| 50 | + __device__ auto operator()(size_type row_index) const noexcept |
| 51 | + { |
| 52 | + return cudf::detail::accumulate( |
| 53 | + _table.begin(), |
| 54 | + _table.end(), |
| 55 | + _seed, |
| 56 | + [row_index, nulls = _check_nulls] __device__(auto hash, auto column) { |
| 57 | + return cudf::type_dispatcher( |
| 58 | + column.type(), element_hasher_adapter{}, column, row_index, nulls, hash); |
| 59 | + }); |
| 60 | + } |
| 61 | + |
| 62 | + /** |
| 63 | + * @brief Computes the hash value of an element in the given column. |
| 64 | + */ |
| 65 | + class element_hasher_adapter { |
| 66 | + public: |
| 67 | + template <typename T, CUDF_ENABLE_IF(column_device_view::has_element_accessor<T>())> |
| 68 | + __device__ hash_value_type operator()(column_device_view const& col, |
| 69 | + size_type const row_index, |
| 70 | + Nullate const _check_nulls, |
| 71 | + hash_value_type const _seed) const noexcept |
| 72 | + { |
| 73 | + if (_check_nulls && col.is_null(row_index)) { |
| 74 | + return cuda::std::numeric_limits<hash_value_type>::max(); |
| 75 | + } |
| 76 | + auto const hasher = XXHash_32<T>{_seed}; |
| 77 | + return hasher(col.element<T>(row_index)); |
| 78 | + } |
| 79 | + |
| 80 | + template <typename T, CUDF_ENABLE_IF(not column_device_view::has_element_accessor<T>())> |
| 81 | + __device__ hash_value_type operator()(column_device_view const&, |
| 82 | + size_type const, |
| 83 | + Nullate const, |
| 84 | + hash_value_type const) const noexcept |
| 85 | + { |
| 86 | + CUDF_UNREACHABLE("Unsupported type for XXHash_32"); |
| 87 | + } |
| 88 | + }; |
| 89 | + |
| 90 | + Nullate const _check_nulls; |
| 91 | + table_device_view const _table; |
| 92 | + hash_value_type const _seed; |
| 93 | +}; |
| 94 | + |
| 95 | +} // namespace |
| 96 | + |
| 97 | +std::unique_ptr<column> xxhash_32(table_view const& input, |
| 98 | + uint32_t seed, |
| 99 | + rmm::cuda_stream_view stream, |
| 100 | + rmm::device_async_resource_ref mr) |
| 101 | +{ |
| 102 | + auto output = make_numeric_column(data_type(type_to_id<hash_value_type>()), |
| 103 | + input.num_rows(), |
| 104 | + mask_state::UNALLOCATED, |
| 105 | + stream, |
| 106 | + mr); |
| 107 | + |
| 108 | + // Return early if there's nothing to hash |
| 109 | + if (input.num_columns() == 0 || input.num_rows() == 0) { return output; } |
| 110 | + |
| 111 | + bool const nullable = has_nulls(input); |
| 112 | + auto const input_view = table_device_view::create(input, stream); |
| 113 | + auto output_view = output->mutable_view(); |
| 114 | + |
| 115 | + // Compute the hash value for each row |
| 116 | + thrust::tabulate(rmm::exec_policy(stream), |
| 117 | + output_view.begin<hash_value_type>(), |
| 118 | + output_view.end<hash_value_type>(), |
| 119 | + device_row_hasher(nullable, *input_view, seed)); |
| 120 | + |
| 121 | + return output; |
| 122 | +} |
| 123 | + |
| 124 | +} // namespace detail |
| 125 | + |
| 126 | +std::unique_ptr<column> xxhash_32(table_view const& input, |
| 127 | + uint32_t seed, |
| 128 | + rmm::cuda_stream_view stream, |
| 129 | + rmm::device_async_resource_ref mr) |
| 130 | +{ |
| 131 | + CUDF_FUNC_RANGE(); |
| 132 | + return detail::xxhash_32(input, seed, stream, mr); |
| 133 | +} |
| 134 | + |
| 135 | +} // namespace hashing |
| 136 | +} // namespace cudf |
0 commit comments