|
| 1 | +#pragma once |
| 2 | + |
| 3 | +#include "bioimage_cpp/array_view.hxx" |
| 4 | +#include "bioimage_cpp/graph/undirected_graph.hxx" |
| 5 | + |
| 6 | +#include <algorithm> |
| 7 | +#include <cstddef> |
| 8 | +#include <cstdint> |
| 9 | +#include <limits> |
| 10 | +#include <numeric> |
| 11 | +#include <stdexcept> |
| 12 | +#include <string> |
| 13 | +#include <thread> |
| 14 | +#include <type_traits> |
| 15 | +#include <unordered_set> |
| 16 | +#include <utility> |
| 17 | +#include <vector> |
| 18 | + |
| 19 | +namespace bioimage_cpp::graph { |
| 20 | + |
| 21 | +class RegionAdjacencyGraph : public UndirectedGraph { |
| 22 | +public: |
| 23 | + using UndirectedGraph::UndirectedGraph; |
| 24 | + |
| 25 | + explicit RegionAdjacencyGraph( |
| 26 | + const std::uint64_t number_of_nodes, |
| 27 | + std::vector<std::uint64_t> shape |
| 28 | + ) |
| 29 | + : UndirectedGraph(number_of_nodes), |
| 30 | + shape_(std::move(shape)) { |
| 31 | + } |
| 32 | + |
| 33 | + [[nodiscard]] const std::vector<std::uint64_t> &shape() const { |
| 34 | + return shape_; |
| 35 | + } |
| 36 | + |
| 37 | +private: |
| 38 | + std::vector<std::uint64_t> shape_; |
| 39 | +}; |
| 40 | + |
| 41 | +namespace detail { |
| 42 | + |
| 43 | +using Edge = UndirectedGraph::Edge; |
| 44 | + |
| 45 | +struct EdgeHash { |
| 46 | + std::size_t operator()(const Edge &edge) const { |
| 47 | + const auto first = static_cast<std::size_t>(edge.first); |
| 48 | + const auto second = static_cast<std::size_t>(edge.second); |
| 49 | + return first ^ (second + 0x9e3779b97f4a7c15ULL + (first << 6U) + (first >> 2U)); |
| 50 | + } |
| 51 | +}; |
| 52 | + |
| 53 | +inline Edge edge_key(std::uint64_t u, std::uint64_t v) { |
| 54 | + if (v < u) { |
| 55 | + std::swap(u, v); |
| 56 | + } |
| 57 | + return {u, v}; |
| 58 | +} |
| 59 | + |
| 60 | +template <class T> |
| 61 | +std::uint64_t checked_label_to_node(const T value) { |
| 62 | + if constexpr (std::is_signed_v<T>) { |
| 63 | + if (value < 0) { |
| 64 | + throw std::invalid_argument("labels must not contain negative values"); |
| 65 | + } |
| 66 | + } |
| 67 | + return static_cast<std::uint64_t>(value); |
| 68 | +} |
| 69 | + |
| 70 | +template <class T> |
| 71 | +std::uint64_t max_label(const ConstArrayView<T> &labels) { |
| 72 | + const auto number_of_pixels = static_cast<std::size_t>(std::accumulate( |
| 73 | + labels.shape.begin(), |
| 74 | + labels.shape.end(), |
| 75 | + std::ptrdiff_t{1}, |
| 76 | + [](const std::ptrdiff_t a, const std::ptrdiff_t b) { return a * b; } |
| 77 | + )); |
| 78 | + std::uint64_t max_value = 0; |
| 79 | + for (std::size_t index = 0; index < number_of_pixels; ++index) { |
| 80 | + max_value = std::max(max_value, checked_label_to_node(labels.data[index])); |
| 81 | + } |
| 82 | + return max_value; |
| 83 | +} |
| 84 | + |
| 85 | +inline std::size_t normalize_thread_count( |
| 86 | + const std::size_t requested, |
| 87 | + const std::size_t number_of_work_items |
| 88 | +) { |
| 89 | + if (number_of_work_items == 0) { |
| 90 | + return 1; |
| 91 | + } |
| 92 | + std::size_t n_threads = requested; |
| 93 | + if (n_threads == 0) { |
| 94 | + n_threads = std::thread::hardware_concurrency(); |
| 95 | + if (n_threads == 0) { |
| 96 | + n_threads = 1; |
| 97 | + } |
| 98 | + } |
| 99 | + return std::max<std::size_t>(1, std::min(n_threads, number_of_work_items)); |
| 100 | +} |
| 101 | + |
| 102 | +template <class T> |
| 103 | +void add_edge_if_different( |
| 104 | + const T *data, |
| 105 | + const std::size_t first, |
| 106 | + const std::size_t second, |
| 107 | + std::unordered_set<Edge, EdgeHash> &edges |
| 108 | +) { |
| 109 | + const auto u = checked_label_to_node(data[first]); |
| 110 | + const auto v = checked_label_to_node(data[second]); |
| 111 | + if (u != v) { |
| 112 | + edges.insert(edge_key(u, v)); |
| 113 | + } |
| 114 | +} |
| 115 | + |
| 116 | +template <class T> |
| 117 | +void scan_2d_chunk( |
| 118 | + const T *data, |
| 119 | + const std::size_t height, |
| 120 | + const std::size_t width, |
| 121 | + const std::size_t y_begin, |
| 122 | + const std::size_t y_end, |
| 123 | + std::unordered_set<Edge, EdgeHash> &edges |
| 124 | +) { |
| 125 | + for (std::size_t y = y_begin; y < y_end; ++y) { |
| 126 | + const auto row_offset = y * width; |
| 127 | + for (std::size_t x = 0; x < width; ++x) { |
| 128 | + const auto pixel = row_offset + x; |
| 129 | + if (x + 1 < width) { |
| 130 | + add_edge_if_different(data, pixel, pixel + 1, edges); |
| 131 | + } |
| 132 | + if (y + 1 < height) { |
| 133 | + add_edge_if_different(data, pixel, pixel + width, edges); |
| 134 | + } |
| 135 | + } |
| 136 | + } |
| 137 | +} |
| 138 | + |
| 139 | +template <class T> |
| 140 | +void scan_3d_chunk( |
| 141 | + const T *data, |
| 142 | + const std::size_t depth, |
| 143 | + const std::size_t height, |
| 144 | + const std::size_t width, |
| 145 | + const std::size_t z_begin, |
| 146 | + const std::size_t z_end, |
| 147 | + std::unordered_set<Edge, EdgeHash> &edges |
| 148 | +) { |
| 149 | + const auto slice_size = height * width; |
| 150 | + for (std::size_t z = z_begin; z < z_end; ++z) { |
| 151 | + const auto slice_offset = z * slice_size; |
| 152 | + for (std::size_t y = 0; y < height; ++y) { |
| 153 | + const auto row_offset = slice_offset + y * width; |
| 154 | + for (std::size_t x = 0; x < width; ++x) { |
| 155 | + const auto pixel = row_offset + x; |
| 156 | + if (x + 1 < width) { |
| 157 | + add_edge_if_different(data, pixel, pixel + 1, edges); |
| 158 | + } |
| 159 | + if (y + 1 < height) { |
| 160 | + add_edge_if_different(data, pixel, pixel + width, edges); |
| 161 | + } |
| 162 | + if (z + 1 < depth) { |
| 163 | + add_edge_if_different(data, pixel, pixel + slice_size, edges); |
| 164 | + } |
| 165 | + } |
| 166 | + } |
| 167 | + } |
| 168 | +} |
| 169 | + |
| 170 | +inline std::vector<Edge> merge_edge_sets( |
| 171 | + const std::vector<std::unordered_set<Edge, EdgeHash>> &per_thread_edges |
| 172 | +) { |
| 173 | + std::unordered_set<Edge, EdgeHash> merged; |
| 174 | + for (const auto &edges : per_thread_edges) { |
| 175 | + merged.insert(edges.begin(), edges.end()); |
| 176 | + } |
| 177 | + |
| 178 | + std::vector<Edge> sorted_edges(merged.begin(), merged.end()); |
| 179 | + std::sort(sorted_edges.begin(), sorted_edges.end()); |
| 180 | + return sorted_edges; |
| 181 | +} |
| 182 | + |
| 183 | +} // namespace detail |
| 184 | + |
| 185 | +template <class T> |
| 186 | +RegionAdjacencyGraph grid_region_adjacency_graph( |
| 187 | + const ConstArrayView<T> &labels, |
| 188 | + const std::size_t number_of_threads |
| 189 | +) { |
| 190 | + if (labels.ndim() != 2 && labels.ndim() != 3) { |
| 191 | + throw std::invalid_argument( |
| 192 | + "labels must be a 2D or 3D array, got ndim=" + |
| 193 | + std::to_string(labels.ndim()) |
| 194 | + ); |
| 195 | + } |
| 196 | + for (const auto axis_size : labels.shape) { |
| 197 | + if (axis_size <= 0) { |
| 198 | + throw std::invalid_argument("labels must not have empty dimensions"); |
| 199 | + } |
| 200 | + } |
| 201 | + |
| 202 | + const auto max_node = detail::max_label(labels); |
| 203 | + if (max_node == std::numeric_limits<std::uint64_t>::max()) { |
| 204 | + throw std::overflow_error("maximum label is too large"); |
| 205 | + } |
| 206 | + |
| 207 | + std::vector<std::uint64_t> shape; |
| 208 | + shape.reserve(labels.shape.size()); |
| 209 | + for (const auto axis_size : labels.shape) { |
| 210 | + shape.push_back(static_cast<std::uint64_t>(axis_size)); |
| 211 | + } |
| 212 | + |
| 213 | + const auto work_items = static_cast<std::size_t>(labels.shape[0]); |
| 214 | + const auto n_threads = detail::normalize_thread_count(number_of_threads, work_items); |
| 215 | + std::vector<std::unordered_set<detail::Edge, detail::EdgeHash>> per_thread_edges(n_threads); |
| 216 | + std::vector<std::thread> threads; |
| 217 | + threads.reserve(n_threads > 0 ? n_threads - 1 : 0); |
| 218 | + |
| 219 | + const auto run_chunk = [&](const std::size_t thread_id) { |
| 220 | + const auto begin = thread_id * work_items / n_threads; |
| 221 | + const auto end = (thread_id + 1) * work_items / n_threads; |
| 222 | + if (labels.ndim() == 2) { |
| 223 | + detail::scan_2d_chunk( |
| 224 | + labels.data, |
| 225 | + static_cast<std::size_t>(labels.shape[0]), |
| 226 | + static_cast<std::size_t>(labels.shape[1]), |
| 227 | + begin, |
| 228 | + end, |
| 229 | + per_thread_edges[thread_id] |
| 230 | + ); |
| 231 | + } else { |
| 232 | + detail::scan_3d_chunk( |
| 233 | + labels.data, |
| 234 | + static_cast<std::size_t>(labels.shape[0]), |
| 235 | + static_cast<std::size_t>(labels.shape[1]), |
| 236 | + static_cast<std::size_t>(labels.shape[2]), |
| 237 | + begin, |
| 238 | + end, |
| 239 | + per_thread_edges[thread_id] |
| 240 | + ); |
| 241 | + } |
| 242 | + }; |
| 243 | + |
| 244 | + for (std::size_t thread_id = 1; thread_id < n_threads; ++thread_id) { |
| 245 | + threads.emplace_back(run_chunk, thread_id); |
| 246 | + } |
| 247 | + run_chunk(0); |
| 248 | + for (auto &thread : threads) { |
| 249 | + thread.join(); |
| 250 | + } |
| 251 | + |
| 252 | + auto graph = RegionAdjacencyGraph(max_node + 1, std::move(shape)); |
| 253 | + for (const auto edge : detail::merge_edge_sets(per_thread_edges)) { |
| 254 | + graph.insert_edge(edge.first, edge.second); |
| 255 | + } |
| 256 | + return graph; |
| 257 | +} |
| 258 | + |
| 259 | +} // namespace bioimage_cpp::graph |
0 commit comments