/src/igraph/src/paths/floyd_warshall.c
Line | Count | Source |
1 | | /* |
2 | | igraph library. |
3 | | Copyright (C) 2022-2025 The igraph development team <igraph@igraph.org> |
4 | | |
5 | | This program is free software; you can redistribute it and/or modify |
6 | | it under the terms of the GNU General Public License as published by |
7 | | the Free Software Foundation; either version 2 of the License, or |
8 | | (at your option) any later version. |
9 | | |
10 | | This program is distributed in the hope that it will be useful, |
11 | | but WITHOUT ANY WARRANTY; without even the implied warranty of |
12 | | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
13 | | GNU General Public License for more details. |
14 | | |
15 | | You should have received a copy of the GNU General Public License |
16 | | along with this program. If not, see <https://www.gnu.org/licenses/>. |
17 | | */ |
18 | | |
19 | | #include "igraph_paths.h" |
20 | | #include "igraph_interface.h" |
21 | | #include "igraph_stack.h" |
22 | | |
23 | | #include "core/interruption.h" |
24 | | #include "internal/utils.h" |
25 | | #include "paths/paths_internal.h" |
26 | | |
27 | 0 | static igraph_error_t distances_floyd_warshall_original(igraph_matrix_t *res) { |
28 | |
|
29 | 0 | igraph_int_t no_of_nodes = igraph_matrix_nrow(res); |
30 | |
|
31 | 0 | for (igraph_int_t k = 0; k < no_of_nodes; k++) { |
32 | 0 | IGRAPH_ALLOW_INTERRUPTION(); |
33 | | |
34 | | /* Iteration order matters for performance! |
35 | | * First j, then i, because matrices are stored as column-major. */ |
36 | 0 | for (igraph_int_t j = 0; j < no_of_nodes; j++) { |
37 | 0 | igraph_real_t dkj = MATRIX(*res, k, j); |
38 | 0 | if (dkj == IGRAPH_INFINITY) { |
39 | 0 | continue; |
40 | 0 | } |
41 | | |
42 | 0 | for (igraph_int_t i = 0; i < no_of_nodes; i++) { |
43 | 0 | igraph_real_t di = MATRIX(*res, i, k) + dkj; |
44 | 0 | igraph_real_t dd = MATRIX(*res, i, j); |
45 | 0 | if (di < dd) { |
46 | 0 | MATRIX(*res, i, j) = di; |
47 | 0 | } |
48 | 0 | if (i == j && MATRIX(*res, i, i) < 0) { |
49 | 0 | IGRAPH_ERROR("Negative cycle found while calculating distances with Floyd-Warshall.", |
50 | 0 | IGRAPH_ENEGCYCLE); |
51 | 0 | } |
52 | 0 | } |
53 | 0 | } |
54 | 0 | } |
55 | | |
56 | 0 | return IGRAPH_SUCCESS; |
57 | 0 | } |
58 | | |
59 | | |
60 | 0 | static igraph_error_t distances_floyd_warshall_tree(igraph_matrix_t *res) { |
61 | | |
62 | | /* This is the "Tree" algorithm of Brodnik et al. |
63 | | * A difference from the paper is that instead of using the OUT_k tree of shortest |
64 | | * paths _starting_ in k, we use the IN_k tree of shortest paths _ending_ in k. |
65 | | * This makes it easier to iterate through matrices in column-major order, |
66 | | * i.e. storage order, thus increasing performance. */ |
67 | |
|
68 | 0 | igraph_int_t no_of_nodes = igraph_matrix_nrow(res); |
69 | | |
70 | | /* successors[v][u] is the second vertex on the shortest path from v to u, |
71 | | i.e. the parent of v in the IN_u tree. */ |
72 | 0 | igraph_matrix_int_t successors; |
73 | 0 | IGRAPH_MATRIX_INT_INIT_FINALLY(&successors, no_of_nodes, no_of_nodes); |
74 | | |
75 | | /* children[children_start[u] + i] is the i-th child of u in a tree of shortest paths |
76 | | rooted at k, and ending in k, in the main loop below (IN_k). There are no_of_nodes-1 |
77 | | child vertices in total, as the root vertex is excluded. This is essentially a contiguously |
78 | | stored adjacency list representation of IN_k. */ |
79 | 0 | igraph_vector_int_t children; |
80 | 0 | IGRAPH_VECTOR_INT_INIT_FINALLY(&children, no_of_nodes-1); |
81 | | |
82 | | /* children_start[u] indicates where the children of u are stored in children[]. |
83 | | These are effectively the cumulative sums of no_of_children[], with the first |
84 | | element being 0. The last element, children_start[no_of_nodes], is equal to the |
85 | | total number of children in the tree, i.e. no_of_nodes-1. */ |
86 | 0 | igraph_vector_int_t children_start; |
87 | 0 | IGRAPH_VECTOR_INT_INIT_FINALLY(&children_start, no_of_nodes+1); |
88 | | |
89 | | /* no_of_children[u] is the number of children that u has in IN_k in the main loop below. */ |
90 | 0 | igraph_vector_int_t no_of_children; |
91 | 0 | IGRAPH_VECTOR_INT_INIT_FINALLY(&no_of_children, no_of_nodes); |
92 | | |
93 | | /* dfs_traversal and dfs_skip arrays for running time optimization, |
94 | | see "Practical improvement" in Section 3.1 of the paper */ |
95 | 0 | igraph_vector_int_t dfs_traversal; |
96 | 0 | IGRAPH_VECTOR_INT_INIT_FINALLY(&dfs_traversal, no_of_nodes); |
97 | 0 | igraph_vector_int_t dfs_skip; |
98 | 0 | IGRAPH_VECTOR_INT_INIT_FINALLY(&dfs_skip, no_of_nodes); |
99 | | |
100 | 0 | igraph_stack_int_t stack; |
101 | 0 | IGRAPH_STACK_INT_INIT_FINALLY(&stack, no_of_nodes); |
102 | | |
103 | 0 | for (igraph_int_t u = 0; u < no_of_nodes; u++) { |
104 | 0 | for (igraph_int_t v = 0; v < no_of_nodes; v++) { |
105 | 0 | MATRIX(successors, v, u) = u; |
106 | 0 | } |
107 | 0 | } |
108 | |
|
109 | 0 | for (igraph_int_t k = 0; k < no_of_nodes; k++) { |
110 | 0 | IGRAPH_ALLOW_INTERRUPTION(); |
111 | | |
112 | | /* Count the children of each node in the shortest path tree, assuming that at |
113 | | this point all elements of no_of_children[] are zeros. */ |
114 | 0 | for (igraph_int_t v = 0; v < no_of_nodes; v++) { |
115 | 0 | if (v == k) continue; |
116 | 0 | igraph_int_t parent = MATRIX(successors, v, k); |
117 | 0 | VECTOR(no_of_children)[parent]++; |
118 | 0 | } |
119 | | |
120 | | /* Note: we do not use igraph_vector_int_cumsum() here as that function produces |
121 | | an output vector of the same length as the input vector. Here we need an output |
122 | | one longer, with a 0 being prepended to what vector_cumsum() would produce. */ |
123 | 0 | igraph_int_t cumsum = 0; |
124 | 0 | for (igraph_int_t v = 0; v < no_of_nodes; v++) { |
125 | 0 | VECTOR(children_start)[v] = cumsum; |
126 | 0 | cumsum += VECTOR(no_of_children)[v]; |
127 | 0 | } |
128 | 0 | VECTOR(children_start)[no_of_nodes] = cumsum; |
129 | | |
130 | | /* Constructing the tree IN_k (as in the paper) and representing it |
131 | | as a contiguously stored adjacency list. The entries of the no_of_children |
132 | | vector as re-used as an index of where to insert child node indices. |
133 | | At the end of the calculation, all elements of no_of_children[] will be zeros, |
134 | | making this vector ready for the next iteration of the outer loop. */ |
135 | 0 | for (igraph_int_t v = 0; v < no_of_nodes; v++) { |
136 | 0 | if (v == k) continue; |
137 | 0 | igraph_int_t parent = MATRIX(successors, v, k); |
138 | 0 | VECTOR(no_of_children)[parent]--; |
139 | 0 | VECTOR(children)[ VECTOR(children_start)[parent] + VECTOR(no_of_children)[parent] ] = v; |
140 | 0 | } |
141 | | |
142 | | /* constructing dfs-traversal and dfs-skip arrays for the IN_k tree */ |
143 | 0 | IGRAPH_CHECK(igraph_stack_int_push(&stack, k)); |
144 | 0 | igraph_int_t counter = 0; |
145 | 0 | while (!igraph_stack_int_empty(&stack)) { |
146 | 0 | igraph_int_t parent = igraph_stack_int_pop(&stack); |
147 | 0 | if (parent >= 0) { |
148 | 0 | VECTOR(dfs_traversal)[counter] = parent; |
149 | 0 | counter++; |
150 | | /* a negative marker -parent - 1 that is popped right after |
151 | | all the descendants of the parent were processed */ |
152 | 0 | IGRAPH_CHECK(igraph_stack_int_push(&stack, -parent - 1)); |
153 | 0 | for (igraph_int_t l = VECTOR(children_start)[parent]; l < VECTOR(children_start)[parent + 1]; l++) { |
154 | 0 | IGRAPH_CHECK(igraph_stack_int_push(&stack, VECTOR(children)[l])); |
155 | 0 | } |
156 | 0 | } else { |
157 | 0 | VECTOR(dfs_skip)[-(parent + 1)] = counter; |
158 | 0 | } |
159 | 0 | } |
160 | | |
161 | | /* main inner loop */ |
162 | 0 | for (igraph_int_t i = 0; i < no_of_nodes; i++) { |
163 | 0 | igraph_real_t dki = MATRIX(*res, k, i); |
164 | 0 | if (dki == IGRAPH_INFINITY || i == k) { |
165 | 0 | continue; |
166 | 0 | } |
167 | 0 | igraph_int_t counter = 1; |
168 | 0 | while (counter < no_of_nodes) { |
169 | 0 | igraph_int_t j = VECTOR(dfs_traversal)[counter]; |
170 | 0 | igraph_real_t di = MATRIX(*res, j, k) + dki; |
171 | 0 | igraph_real_t dd = MATRIX(*res, j, i); |
172 | 0 | if (di < dd) { |
173 | 0 | MATRIX(*res, j, i) = di; |
174 | 0 | MATRIX(successors, j, i) = MATRIX(successors, j, k); |
175 | 0 | counter++; |
176 | 0 | } else { |
177 | 0 | counter = VECTOR(dfs_skip)[j]; |
178 | 0 | } |
179 | 0 | if (i == j && MATRIX(*res, i, i) < 0) { |
180 | 0 | IGRAPH_ERROR("Negative cycle found while calculating distances with Floyd-Warshall.", |
181 | 0 | IGRAPH_ENEGCYCLE); |
182 | 0 | } |
183 | 0 | } |
184 | 0 | } |
185 | 0 | } |
186 | | |
187 | 0 | igraph_stack_int_destroy(&stack); |
188 | 0 | igraph_vector_int_destroy(&dfs_traversal); |
189 | 0 | igraph_vector_int_destroy(&dfs_skip); |
190 | 0 | igraph_vector_int_destroy(&no_of_children); |
191 | 0 | igraph_vector_int_destroy(&children_start); |
192 | 0 | igraph_vector_int_destroy(&children); |
193 | 0 | igraph_matrix_int_destroy(&successors); |
194 | 0 | IGRAPH_FINALLY_CLEAN(7); |
195 | |
|
196 | 0 | return IGRAPH_SUCCESS; |
197 | 0 | } |
198 | | |
199 | | /** |
200 | | * \function igraph_distances_floyd_warshall |
201 | | * \brief Weighted all-pairs shortest path lengths with the Floyd-Warshall algorithm. |
202 | | * |
203 | | * The Floyd-Warshall algorithm computes weighted shortest path lengths between |
204 | | * all pairs of vertices at the same time. It is useful with very dense weighted graphs, |
205 | | * as its running time is primarily determined by the vertex count, and is not sensitive |
206 | | * to the graph density. In sparse graphs, other methods such as the Dijkstra or |
207 | | * Bellman-Ford algorithms will perform significantly better. |
208 | | * |
209 | | * </para><para> |
210 | | * In addition to the original Floyd-Warshall algorithm, igraph contains implementations |
211 | | * of variants that offer better asymptotic complexity as well as better practical |
212 | | * running times for most instances. See the reference below for more information. |
213 | | * |
214 | | * </para><para> |
215 | | * Note that internally this function always computes the distance matrix |
216 | | * for all pairs of vertices. The \p from and \p to parameters only serve |
217 | | * to subset this matrix, but do not affect the time or memory taken by the |
218 | | * calculation. |
219 | | * |
220 | | * </para><para> |
221 | | * Reference: |
222 | | * |
223 | | * </para><para> |
224 | | * Brodnik, A., Grgurovič, M., Požar, R.: |
225 | | * Modifications of the Floyd-Warshall algorithm with nearly quadratic expected-time, |
226 | | * Ars Mathematica Contemporanea, vol. 22, issue 1, p. #P1.01 (2021). |
227 | | * https://doi.org/10.26493/1855-3974.2467.497 |
228 | | * |
229 | | * \param graph The graph object. |
230 | | * \param res An intialized matrix, the distances will be stored here. |
231 | | * \param from The source vertices. |
232 | | * \param to The target vertices. |
233 | | * \param weights The edge weights. If \c NULL, all weights are assumed to be 1. |
234 | | * Negative weights are allowed, but the graph must not contain negative cycles. |
235 | | * Edges with positive infinite weights are ignored. |
236 | | * \param mode The type of shortest paths to be use for the |
237 | | * calculation in directed graphs. Possible values: |
238 | | * \clist |
239 | | * \cli IGRAPH_OUT |
240 | | * the outgoing paths are calculated. |
241 | | * \cli IGRAPH_IN |
242 | | * the incoming paths are calculated. |
243 | | * \cli IGRAPH_ALL |
244 | | * the directed graph is considered as an |
245 | | * undirected one for the computation. |
246 | | * \endclist |
247 | | * \param method The type of the algorithm used. |
248 | | * \clist |
249 | | * \cli IGRAPH_FLOYD_WARSHALL_AUTOMATIC |
250 | | * tries to select the best performing variant for the current graph; |
251 | | * presently this option always uses the "Tree" method. |
252 | | * \cli IGRAPH_FLOYD_WARSHALL_ORIGINAL |
253 | | * the basic Floyd-Warshall algorithm. |
254 | | * \cli IGRAPH_FLOYD_WARSHALL_TREE |
255 | | * the "Tree" speedup of Brodnik et al., faster than the original algorithm |
256 | | * in most cases. |
257 | | * \endclist |
258 | | * \return Error code. \c IGRAPH_ENEGCYCLE is returned if a negative-weight |
259 | | * cycle is found. |
260 | | * |
261 | | * \sa \ref igraph_distances(), \ref igraph_distances_dijkstra(), |
262 | | * \ref igraph_distances_bellman_ford(), \ref igraph_distances_johnson() |
263 | | * |
264 | | * Time complexity: |
265 | | * The original variant has complexity O(|V|^3 + |E|). |
266 | | * The "Tree" variant has expected-case complexity of O(|V|^2 log^2 |V|) |
267 | | * according to Brodnik et al., while its worst-time complexity remains O(|V|^3). |
268 | | * Here |V| denotes the number of vertices and |E| is the number of edges. |
269 | | */ |
270 | | igraph_error_t igraph_distances_floyd_warshall( |
271 | | const igraph_t *graph, igraph_matrix_t *res, |
272 | | igraph_vs_t from, igraph_vs_t to, |
273 | | const igraph_vector_t *weights, igraph_neimode_t mode, |
274 | 0 | const igraph_floyd_warshall_algorithm_t method) { |
275 | |
|
276 | 0 | igraph_bool_t negative_weights; |
277 | 0 | IGRAPH_CHECK(igraph_i_validate_distance_weights(graph, weights, &negative_weights)); |
278 | 0 | return igraph_i_distances_floyd_warshall(graph, res, from, to, weights, mode, method); |
279 | 0 | } |
280 | | |
281 | | igraph_error_t igraph_i_distances_floyd_warshall( |
282 | | const igraph_t *graph, igraph_matrix_t *res, |
283 | | igraph_vs_t from, igraph_vs_t to, |
284 | | const igraph_vector_t *weights, igraph_neimode_t mode, |
285 | 0 | const igraph_floyd_warshall_algorithm_t method) { |
286 | |
|
287 | 0 | igraph_int_t no_of_nodes = igraph_vcount(graph); |
288 | 0 | igraph_int_t no_of_edges = igraph_ecount(graph); |
289 | 0 | igraph_bool_t in = false, out = false; |
290 | |
|
291 | 0 | if (! igraph_is_directed(graph)) { |
292 | 0 | mode = IGRAPH_ALL; |
293 | 0 | } |
294 | |
|
295 | 0 | switch (mode) { |
296 | 0 | case IGRAPH_ALL: |
297 | 0 | in = out = true; |
298 | 0 | break; |
299 | 0 | case IGRAPH_OUT: |
300 | 0 | out = true; |
301 | 0 | break; |
302 | 0 | case IGRAPH_IN: |
303 | 0 | in = true; |
304 | 0 | break; |
305 | 0 | default: |
306 | 0 | IGRAPH_ERROR("Invalid mode for Floyd-Warshall shortest path calculation.", IGRAPH_EINVMODE); |
307 | 0 | } |
308 | | |
309 | 0 | IGRAPH_CHECK(igraph_matrix_resize(res, no_of_nodes, no_of_nodes)); |
310 | 0 | igraph_matrix_fill(res, IGRAPH_INFINITY); |
311 | |
|
312 | 0 | for (igraph_int_t v = 0; v < no_of_nodes; v++) { |
313 | 0 | MATRIX(*res, v, v) = 0; |
314 | 0 | } |
315 | |
|
316 | 0 | for (igraph_int_t e = 0; e < no_of_edges; e++) { |
317 | 0 | igraph_int_t from = IGRAPH_FROM(graph, e); |
318 | 0 | igraph_int_t to = IGRAPH_TO(graph, e); |
319 | 0 | igraph_real_t w = weights ? VECTOR(*weights)[e] : 1; |
320 | |
|
321 | 0 | if (w < 0) { |
322 | 0 | if (mode == IGRAPH_ALL) { |
323 | 0 | IGRAPH_ERRORF("Negative edge weight (%g) found in undirected graph " |
324 | 0 | "while calculating distances with Floyd-Warshall.", |
325 | 0 | IGRAPH_ENEGCYCLE, w); |
326 | 0 | } else if (to == from) { |
327 | 0 | IGRAPH_ERRORF("Self-loop with negative weight (%g) found " |
328 | 0 | "while calculating distances with Floyd-Warshall.", |
329 | 0 | IGRAPH_ENEGCYCLE, w); |
330 | 0 | } |
331 | 0 | } else if (w == IGRAPH_INFINITY) { |
332 | | /* Ignore edges with infinite weight */ |
333 | 0 | continue; |
334 | 0 | } |
335 | | |
336 | 0 | if (out && MATRIX(*res, from, to) > w) { |
337 | 0 | MATRIX(*res, from, to) = w; |
338 | 0 | } |
339 | 0 | if (in && MATRIX(*res, to, from) > w) { |
340 | 0 | MATRIX(*res, to, from) = w; |
341 | 0 | } |
342 | 0 | } |
343 | | |
344 | | /* If there are zero or one vertices, nothing needs to be done. |
345 | | * This is special-cased so that at later stages we can rely on no_of_nodes - 1 >= 0. */ |
346 | 0 | if (no_of_nodes <= 1) { |
347 | 0 | return IGRAPH_SUCCESS; |
348 | 0 | } |
349 | | |
350 | 0 | switch (method) { |
351 | 0 | case IGRAPH_FLOYD_WARSHALL_ORIGINAL: |
352 | 0 | IGRAPH_CHECK(distances_floyd_warshall_original(res)); |
353 | 0 | break; |
354 | 0 | case IGRAPH_FLOYD_WARSHALL_AUTOMATIC: |
355 | 0 | case IGRAPH_FLOYD_WARSHALL_TREE: |
356 | 0 | IGRAPH_CHECK(distances_floyd_warshall_tree(res)); |
357 | 0 | break; |
358 | 0 | default: |
359 | 0 | IGRAPH_ERROR("Invalid method.", IGRAPH_EINVAL); |
360 | 0 | } |
361 | | |
362 | 0 | IGRAPH_CHECK(igraph_i_matrix_subset_vertices(res, graph, from, to)); |
363 | | |
364 | 0 | return IGRAPH_SUCCESS; |
365 | 0 | } |