2 * Searches for "good" ways to divide n matchsticks up and reassemble them
3 * into m matchsticks. "Good" means the smallest fragment is as big
8 * The algorithm is faster if the arguments are ordered so that n > m.
12 * matchsticks/main.c Copyright 2014 Ian Jackson
14 * This program is free software: you can redistribute it and/or modify
15 * it under the terms of the GNU General Public License as published by
16 * the Free Software Foundation, either version 3 of the License, or
17 * (at your option) any later version.
19 * This program is distributed in the hope that it will be useful,
20 * but WITHOUT ANY WARRANTY; without even the implied warranty of
21 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
22 * GNU General Public License for more details.
33 #include <sys/types.h>
36 #include <sys/fcntl.h>
44 * Each input match contributes, or does not contribute, to each
45 * output match; we do not need to consider multiple fragments
46 * relating to the same input/output pair this gives an n*m adjacency
47 * matrix (bitmap). Given such an adjacency matrix, the problem of
48 * finding the best sizes for the fragments can be expressed as a
49 * linear programming problem.
51 * We search all possible adjacency matrices, and for each one we run
52 * GLPK's simplex solver. We represent the adjacency matrix as an
55 * However, there are a couple of wrinkles:
57 * To best represent the problem as a standard LP problem, we separate
58 * out the size of each fragment into a common minimum size variable,
59 * plus a fragment-specific extra size variable. This reduces the LP
60 * problem size at the cost of making the problem construction, and
61 * interpretation of the results, a bit fiddly.
63 * Many of the adjacency matrices are equivalent. In particular,
64 * permutations of the columns, or of the rows, do not change the
65 * meaning. It is only necessasry to consider any one permutation.
66 * We make use of this by considering only adjacency matrices whose
67 * bitmap array contains bitmap words whose numerical values are
68 * nondecreasing in array order.
70 * Once we have a solution, we also avoid considering any candidate
71 * which involves dividing one of the output sticks into so many
72 * fragment that the smallest fragment would necessarily be no bigger
73 * than our best solution. That is, we reject candidates where any of
74 * the hamming weights of the adjacency bitmap words are too large.
76 * And, we want to do the search in order of increasing maximum
77 * hamming weight. This is because in practice optimal solutions tend
78 * to have low hamming weight, and having found a reasonable solution
79 * early allows us to eliminate a lot of candidates without doing the
83 typedef uint32_t AdjWord;
84 #define PRADJ "08"PRIx32
86 static int n, m, maxhamweight;
87 static AdjWord *adjmatrix;
88 static AdjWord adjall;
91 static glp_prob *best_prob;
92 static AdjWord *best_adjmatrix;
94 static unsigned printcounter;
96 static void iterate(void);
97 static void iterate_recurse(int i, AdjWord min);
98 static void optimise(bool doprint);
100 static void progress_eol(void) {
101 fprintf(stderr," \r");
105 /*----- multicore support -----*/
116 * - one pipe ("work") from generator to workers
117 * - ever-extending file ("bus") containing new "best" values
118 * - one file for each worker giving maxhamweight and adjmatrix for best
120 * generator runs iterate_recurse to a certain depth and writes the
121 * candidates to a pipe
123 * workers read candidates from the pipe and resume iterate_recurse
124 * halfway through the recursion
126 * whenever a worker does a doprint, it checks the bus for new best
127 * value; actual best values are appended
129 * master waits for generator and all workers to finish and then
130 * runs optimise() for each worker's best, then prints
133 static int ncpus = 0, multicore_iteration_boundary = INT_MAX;
135 static int mc_bus, mc_work[2];
136 static off_t mc_bus_read;
143 static Worker *mc_us;
146 static AdjWord mc_iter_min;
148 static size_t mc_iovlen;
149 static struct iovec mc_iov[MAX_NIOVS];
151 #define IOV0 (mc_niovs = mc_iovlen = 0)
153 #define IOV(obj, count) ({ \
154 assert(mc_niovs < MAX_NIOVS); \
155 mc_iov[mc_niovs].iov_base = &(obj); \
156 mc_iov[mc_niovs].iov_len = sizeof(obj) * (count); \
157 mc_iovlen += mc_iov[mc_niovs].iov_len; \
161 static void mc_rwvsetup_outer(void) {
163 IOV(maxhamweight, 1);
165 IOV(*adjmatrix, multicore_iteration_boundary);
168 static void mc_rwvsetup_full(void) {
173 static void vlprintf(const char *fmt, va_list al) {
174 vfprintf(stderr,fmt,al);
178 static void LPRINTF(const char *fmt, ...) {
185 static void mc_awaitpid(int wnum, pid_t pid) {
186 LPRINTF("master awaiting %2d [%ld]",wnum,(long)pid);
188 pid_t got = waitpid(pid, &status, 0);
191 fprintf(stderr,"\nFAILED SUBPROC %2d [%ld] %d\n",
192 wnum, (long)pid, status);
197 static void multicore_outer_iteration(int i, AdjWord min) {
198 assert(i == multicore_iteration_boundary);
201 ssize_t r = writev(mc_work[1], mc_iov, mc_niovs);
202 assert(r == mc_iovlen);
203 /* effectively, this writev arranges to transfers control
204 * to some worker's instance of iterate_recurse via mc_iterate_worker */
207 static void mc_iterate_worker(void) {
210 ssize_t r = readv(mc_work[0], mc_iov, mc_niovs);
212 assert(r == mc_iovlen);
214 /* stop iterate_recurse from trying to run multicore_outer_iteration */
215 int mc_org_it_bound = multicore_iteration_boundary;
216 multicore_iteration_boundary = INT_MAX;
217 iterate_recurse(mc_org_it_bound, mc_iter_min);
218 multicore_iteration_boundary = mc_org_it_bound;
220 LPRINTF("worker %2d reporting",mc_us->w);
221 if (best_adjmatrix) {
222 adjmatrix = best_adjmatrix;
224 ssize_t r = writev(fileno(mc_us->results), mc_iov, mc_niovs);
225 assert(r == mc_iovlen);
227 LPRINTF("worker %2d ending",mc_us->w);
231 static void multicore(void) {
236 multicore_iteration_boundary = n / 2;
238 FILE *busf = tmpfile(); assert(busf);
239 mc_bus = fileno(busf);
240 int r = fcntl(mc_bus, F_GETFL); assert(r >= 0);
242 r = fcntl(mc_bus, F_SETFL, r); assert(r >= 0);
244 r = pipe(mc_work); assert(!r);
246 mc_workers = xmalloc(sizeof(*mc_workers) * ncpus);
247 for (w=0; w<ncpus; w++) {
249 mc_workers[w].results = tmpfile(); assert(mc_workers[w].results);
250 mc_workers[w].pid = fork(); assert(mc_workers[w].pid >= 0);
251 if (!mc_workers[w].pid) {
252 mc_us = &mc_workers[w];
254 LPRINTF("worker %2d running", w);
262 genpid = fork(); assert(genpid >= 0);
264 LPRINTF("generator running");
270 mc_awaitpid(-1, genpid);
271 for (w=0; w<ncpus; w++)
272 mc_awaitpid(w, mc_workers[w].pid);
274 for (w=0; w<ncpus; w++) {
276 LPRINTF("reading report from %2d",w);
277 ssize_t sr = preadv(fileno(mc_workers[w].results), mc_iov, mc_niovs, 0);
284 static void multicore_check_for_new_best(void) {
289 ssize_t got = pread(mc_bus, &msg, sizeof(msg), mc_bus_read);
291 assert(got == sizeof(msg));
294 mc_bus_read += sizeof(msg);
298 static void multicore_found_new_best(void) {
301 if (mc_us /* might be master */) fprintf(stderr," w%-2d ",mc_us->w);
302 ssize_t wrote = write(mc_bus, &best, sizeof(best));
303 assert(wrote == sizeof(best));
306 /*----- end of multicore support -----*/
308 static AdjWord *xalloc_adjmatrix(void) {
309 return xmalloc(sizeof(*adjmatrix)*n);
312 static void prep(void) {
313 adjall = ~((~(AdjWord)0) << m);
314 adjmatrix = xalloc_adjmatrix();
315 glp_term_out(GLP_OFF);
319 static AdjWord one_adj_bit(int bitnum) {
320 return (AdjWord)1 << bitnum;
323 static int count_set_adj_bits(AdjWord w) {
325 for (j=0, total=0; j<m; j++)
326 total += !!(w & one_adj_bit(j));
330 #define PRINTF(...) if (!doprint) ; else fprintf(stderr, __VA_ARGS__)
332 static int totalfrags;
334 static bool preconsider_ok(int nwords, bool doprint) {
337 PRINTF("%2d ", maxhamweight);
340 for (i=0, totalfrags=0; i<nwords; i++) {
341 int frags = count_set_adj_bits(adjmatrix[i]);
342 had_max += (frags >= maxhamweight);
344 PRINTF("%"PRADJ" ", adjmatrix[i]);
345 double maxminsize = (double)m / frags;
346 if (maxminsize <= best) {
352 /* Skip this candidate as its max hamming weight is lower than
353 * we're currently looking for (which means we must have done it
354 * already). (The recursive iteration ensures that none of the
355 * words have more than the max hamming weight.) */
365 static void optimise(bool doprint) {
366 /* Consider the best answer (if any) for a given adjacency matrix */
371 * Up to a certain point, optimise() can be restarted. We use this
372 * to go back and print the debugging output if it turns out that we
373 * have an interesting case. The HAVE_PRINTED macro does this: its
374 * semantics are to go back in time and make sure that we have
375 * printed the description of the search case.
377 #define HAVE_PRINTED ({ \
378 if (!doprint) { doprint = 1; goto retry_with_print; } \
382 glp_delete_prob(prob);
386 bool ok = preconsider_ok(n, doprint);
391 * We formulate our problem as an LP problem as follows.
392 * In this file "n" and "m" are the matchstick numbers.
394 * Each set bit in the adjacency matrix corresponds to taking a
395 * fragment from old match i and making it part of new match j.
397 * The structural variables (columns) are:
398 * x_minimum minimum size of any fragment (bounded below by 0)
399 * x_morefrag_i_j the amount by which the size of the fragment
400 * i,j exceeds the minimum size (bounded below by 0)
402 * The auxiliary variables (rows) are:
403 * x_total_i total length for each input match (fixed variable)
404 * x_total_j total length for each output match (fixed variable)
406 * The objective function is simply
409 * We use X_ and Y_ to refer to GLPK's (1-based) column and row indices.
410 * ME_ refers to entries in the list of constraint matrix elements
411 * which we build up as we go.
414 prob = glp_create_prob();
416 int Y_totals_i = glp_add_rows(prob, n);
417 int Y_totals_j = glp_add_rows(prob, m);
418 int X_minimum = glp_add_cols(prob, 1);
421 int next_matrix_entry = 1; /* wtf GLPK! */
422 int matrix_entries_size = next_matrix_entry + n + m + totalfrags*2;
423 double matrix_entries[matrix_entries_size];
424 int matrix_entries_XY[2][matrix_entries_size];
426 #define ADD_MATRIX_ENTRY(Y,X) ({ \
427 assert(next_matrix_entry < matrix_entries_size); \
428 matrix_entries_XY[0][next_matrix_entry] = (X); \
429 matrix_entries_XY[1][next_matrix_entry] = (Y); \
430 matrix_entries[next_matrix_entry] = 0; \
431 next_matrix_entry++; \
434 int ME_totals_i__minimum = next_matrix_entry;
435 for (i=0; i<n; i++) ADD_MATRIX_ENTRY(Y_totals_i+i, X_minimum);
437 int ME_totals_j__minimum = next_matrix_entry;
438 for (j=0; j<m; j++) ADD_MATRIX_ENTRY(Y_totals_j+j, X_minimum);
440 /* \forall_i x_total_i = m */
441 /* \forall_i x_total_j = n */
442 for (i=0; i<n; i++) glp_set_row_bnds(prob, Y_totals_i+i, GLP_FX, m,m);
443 for (j=0; j<m; j++) glp_set_row_bnds(prob, Y_totals_j+j, GLP_FX, n,n);
446 glp_set_col_bnds(prob, X_minimum, GLP_LO, 0, 0);
447 glp_set_col_name(prob, X_minimum, "minimum");
449 /* objective is maximising x_minimum */
450 glp_set_obj_dir(prob, GLP_MAX);
451 glp_set_obj_coef(prob, X_minimum, 1);
453 for (i=0; i<n; i++) {
454 for (j=0; j<m; j++) {
455 if (!(adjmatrix[i] & one_adj_bit(j)))
457 /* x_total_i += x_minimum */
458 /* x_total_j += x_minimum */
459 matrix_entries[ ME_totals_i__minimum + i ] ++;
460 matrix_entries[ ME_totals_j__minimum + j ] ++;
462 /* x_morefrag_i_j >= 0 */
463 int X_morefrag_i_j = glp_add_cols(prob, 1);
464 glp_set_col_bnds(prob, X_morefrag_i_j, GLP_LO, 0, 0);
467 snprintf(buf,sizeof(buf),"mf %d,%d",i,j);
468 glp_set_col_name(prob, X_morefrag_i_j, buf);
471 /* x_total_i += x_morefrag_i_j */
472 /* x_total_j += x_morefrag_i_j */
473 int ME_totals_i__mf_i_j = ADD_MATRIX_ENTRY(Y_totals_i+i, X_morefrag_i_j);
474 int ME_totals_j__mf_i_j = ADD_MATRIX_ENTRY(Y_totals_j+j, X_morefrag_i_j);
475 matrix_entries[ME_totals_i__mf_i_j] = 1;
476 matrix_entries[ME_totals_j__mf_i_j] = 1;
480 assert(next_matrix_entry == matrix_entries_size);
482 glp_load_matrix(prob, matrix_entries_size-1,
483 matrix_entries_XY[1], matrix_entries_XY[0],
486 int r = glp_simplex(prob, NULL);
487 PRINTF(" glp=%d", r);
490 case e: PRINTF(" " #e ); goto out;
492 case e: HAVE_PRINTED; printf(" " #e " CRASHING\n"); exit(-1);
494 default: HAVE_PRINTED; printf(" ! CRASHING\n"); exit(-1);
514 r = glp_get_status(prob);
515 PRINTF(" status=%d", r);
527 double got = glp_get_obj_val(prob);
535 multicore_found_new_best();
537 if (best_prob) glp_delete_prob(best_prob);
540 free(best_adjmatrix);
541 best_adjmatrix = xalloc_adjmatrix();
542 memcpy(best_adjmatrix, adjmatrix, sizeof(*adjmatrix)*n);
550 glp_delete_prob(prob);
551 if (doprint) progress_eol();
552 if (doprint) multicore_check_for_new_best();
555 static void iterate_recurse(int i, AdjWord min) {
558 optimise(!(printcounter & 0xfff));
561 if (i >= multicore_iteration_boundary) {
562 multicore_outer_iteration(i, min);
565 for (adjmatrix[i] = min;
568 if (count_set_adj_bits(adjmatrix[i]) > maxhamweight)
570 if (i == 0 && (adjmatrix[i] & (1+adjmatrix[i])))
573 iterate_recurse(i+1, adjmatrix[i]);
576 if (adjmatrix[i] == adjall)
581 static void iterate(void) {
582 for (maxhamweight=1; maxhamweight<=m; maxhamweight++) {
583 double maxminsize = (double)m / maxhamweight;
584 if (maxminsize <= best)
587 iterate_recurse(0, 1);
591 static void report(void) {
592 fprintf(stderr, "\n");
594 double min = glp_get_obj_val(best_prob);
597 for (i = 0; i < n; i++)
598 for (j = 0; j < m; j++)
600 cols = glp_get_num_cols(best_prob);
601 for (i = 1; i <= cols; i++) {
603 if (2 != sscanf(glp_get_col_name(best_prob, i), "mf %d,%d", &x, &y))
605 a[x][y] = min + glp_get_col_prim(best_prob, i);
607 printf("%d into %d: min fragment %g\n", n, m, min);
608 for (i = 0; i < n; i++) {
609 for (j = 0; j < m; j++) {
611 printf(" %9.3f", a[i][j]);
618 if (ferror(stdout) || fclose(stdout)) { perror("stdout"); exit(-1); }
621 int main(int argc, char **argv) {
623 while ((opt = getopt(argc,argv,"j:")) >= 0) {
625 case 'j': ncpus = atoi(optarg); break;
626 case '+': assert(!"bad option");
638 if (ncpus) multicore();