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.
38 #include <sys/types.h>
41 #include <sys/fcntl.h>
46 #define VERSION "(unknown-version)"
52 * Each input match contributes, or does not contribute, to each
53 * output match; we do not need to consider multiple fragments
54 * relating to the same input/output pair this gives an n*m adjacency
55 * matrix (bitmap). Given such an adjacency matrix, the problem of
56 * finding the best sizes for the fragments can be expressed as a
57 * linear programming problem.
59 * We search all possible adjacency matrices, and for each one we run
60 * GLPK's simplex solver. We represent the adjacency matrix as an
63 * However, there are a couple of wrinkles:
65 * To best represent the problem as a standard LP problem, we separate
66 * out the size of each fragment into a common minimum size variable,
67 * plus a fragment-specific extra size variable. This reduces the LP
68 * problem size at the cost of making the problem construction, and
69 * interpretation of the results, a bit fiddly.
71 * Many of the adjacency matrices are equivalent. In particular,
72 * permutations of the columns, or of the rows, do not change the
73 * meaning. It is only necessasry to consider any one permutation.
74 * We make use of this by considering only adjacency matrices whose
75 * bitmap array contains bitmap words whose numerical values are
76 * nondecreasing in array order.
78 * Once we have a solution, we also avoid considering any candidate
79 * which involves dividing one of the output sticks into so many
80 * fragment that the smallest fragment would necessarily be no bigger
81 * than our best solution. That is, we reject candidates where any of
82 * the hamming weights of the adjacency bitmap words are too large.
84 * And, we want to do the search in order of increasing maximum
85 * hamming weight. This is because in practice optimal solutions tend
86 * to have low hamming weight, and having found a reasonable solution
87 * early allows us to eliminate a lot of candidates without doing the
91 typedef uint32_t AdjWord;
92 #define PRADJ "08"PRIx32
94 static int n, m, maxhamweight;
95 static AdjWord *adjmatrix;
96 static AdjWord adjall;
99 static glp_prob *best_prob;
100 static AdjWord *best_adjmatrix;
102 static int n_over_best;
105 static unsigned printcounter;
107 static void iterate(void);
108 static void iterate_recurse(int i, AdjWord min);
109 static bool preconsider_ok(int nwords, bool doprint);
110 static bool maxhamweight_ok(void);
111 static void optimise(bool doprint);
113 static void progress_eol(void) {
114 fprintf(stderr," \r");
118 static void set_best(double new_best) {
120 n_over_best = floor(n / best);
123 /*----- multicore support -----*/
134 * - one pipe ("work") from generator to workers
135 * - ever-extending file ("bus") containing new "best" values
136 * - one file for each worker giving maxhamweight and adjmatrix for best
138 * generator runs iterate_recurse to a certain depth and writes the
139 * candidates to a pipe
141 * workers read candidates from the pipe and resume iterate_recurse
142 * halfway through the recursion
144 * whenever a worker does a doprint, it checks the bus for new best
145 * value; actual best values are appended
147 * master waits for generator and all workers to finish and then
148 * runs optimise() for each worker's best, then prints
151 static int ncpus = 0, multicore_iteration_boundary = INT_MAX;
153 static int mc_bus, mc_work[2];
154 static off_t mc_bus_read;
161 static Worker *mc_us;
162 static bool mc_am_generator;
164 static void multicore_check_for_new_best(void);
167 static AdjWord mc_iter_min;
169 static size_t mc_iovlen;
170 static struct iovec mc_iov[MAX_NIOVS];
172 #define IOV0 (mc_niovs = mc_iovlen = 0)
174 #define IOV(obj, count) ({ \
175 assert(mc_niovs < MAX_NIOVS); \
176 mc_iov[mc_niovs].iov_base = &(obj); \
177 mc_iov[mc_niovs].iov_len = sizeof(obj) * (count); \
178 mc_iovlen += mc_iov[mc_niovs].iov_len; \
182 static void mc_rwvsetup_outer(void) {
184 IOV(maxhamweight, 1);
186 IOV(*adjmatrix, multicore_iteration_boundary);
190 static void mc_rwvsetup_full(void) {
195 static void vlprintf(const char *fmt, va_list al) {
196 vfprintf(stderr,fmt,al);
200 static void LPRINTF(const char *fmt, ...) {
207 static void mc_awaitpid(int wnum, pid_t pid) {
208 LPRINTF("master awaiting %2d [%ld]",wnum,(long)pid);
210 pid_t got = waitpid(pid, &status, 0);
213 fprintf(stderr,"\nFAILED SUBPROC %2d [%ld] %d\n",
214 wnum, (long)pid, status);
219 static void multicore_outer_iteration(int i, AdjWord min) {
220 static unsigned check_counter;
222 assert(i == multicore_iteration_boundary);
225 ssize_t r = writev(mc_work[1], mc_iov, mc_niovs);
226 assert(r == mc_iovlen);
227 /* effectively, this writev arranges to transfers control
228 * to some worker's instance of iterate_recurse via mc_iterate_worker */
230 if (!(check_counter++ & 0xff))
231 multicore_check_for_new_best();
234 static void mc_iterate_worker(void) {
237 ssize_t r = readv(mc_work[0], mc_iov, mc_niovs);
239 assert(r == mc_iovlen);
241 bool ok = maxhamweight_ok();
244 ok = preconsider_ok(multicore_iteration_boundary, 1);
248 /* stop iterate_recurse from trying to run multicore_outer_iteration */
249 int mc_org_it_bound = multicore_iteration_boundary;
250 multicore_iteration_boundary = INT_MAX;
251 iterate_recurse(mc_org_it_bound, mc_iter_min);
252 multicore_iteration_boundary = mc_org_it_bound;
254 if (best_adjmatrix) {
255 LPRINTF("worker %2d reporting",mc_us->w);
256 adjmatrix = best_adjmatrix;
258 ssize_t r = writev(fileno(mc_us->results), mc_iov, mc_niovs);
259 assert(r == mc_iovlen);
261 LPRINTF("worker %2d ending",mc_us->w);
265 static void multicore(void) {
270 multicore_iteration_boundary = n / 2;
272 FILE *busf = tmpfile(); assert(busf);
273 mc_bus = fileno(busf);
274 int r = fcntl(mc_bus, F_GETFL); assert(r >= 0);
276 r = fcntl(mc_bus, F_SETFL, r); assert(r >= 0);
278 r = pipe(mc_work); assert(!r);
280 mc_workers = xmalloc(sizeof(*mc_workers) * ncpus);
281 for (w=0; w<ncpus; w++) {
283 mc_workers[w].results = tmpfile(); assert(mc_workers[w].results);
284 mc_workers[w].pid = fork(); assert(mc_workers[w].pid >= 0);
285 if (!mc_workers[w].pid) {
286 mc_us = &mc_workers[w];
288 LPRINTF("worker %2d running", w);
296 genpid = fork(); assert(genpid >= 0);
299 LPRINTF("generator running");
305 mc_awaitpid(-1, genpid);
306 for (w=0; w<ncpus; w++)
307 mc_awaitpid(w, mc_workers[w].pid);
309 for (w=0; w<ncpus; w++) {
311 LPRINTF("reading report from %2d",w);
312 ssize_t sr = preadv(fileno(mc_workers[w].results), mc_iov, mc_niovs, 0);
314 LPRINTF("got report from %2d",w);
320 static void multicore_check_for_new_best(void) {
321 if (!(mc_us || mc_am_generator))
326 ssize_t got = pread(mc_bus, &msg, sizeof(msg), mc_bus_read);
328 assert(got == sizeof(msg));
331 mc_bus_read += sizeof(msg);
335 static void multicore_found_new_best(void) {
339 if (mc_us /* might be master */) fprintf(stderr," w%-2d ",mc_us->w);
340 ssize_t wrote = write(mc_bus, &best, sizeof(best));
341 assert(wrote == sizeof(best));
344 /*----- end of multicore support -----*/
346 static AdjWord *xalloc_adjmatrix(void) {
347 return xmalloc(sizeof(*adjmatrix)*n);
350 static void prep(void) {
351 adjall = ~((~(AdjWord)0) << m);
352 adjmatrix = xalloc_adjmatrix();
353 glp_term_out(GLP_OFF);
355 weight = calloc(sizeof(*weight), m); assert(weight);
356 n_over_best = INT_MAX;
359 static AdjWord one_adj_bit(int bitnum) {
360 return (AdjWord)1 << bitnum;
363 static int count_set_adj_bits(AdjWord w) {
365 for (j=0, total=0; j<m; j++)
366 total += !!(w & one_adj_bit(j));
370 #define PRINTF(...) if (!doprint) ; else fprintf(stderr, __VA_ARGS__)
372 static int totalfrags;
374 static bool maxhamweight_ok(void) {
375 double maxminsize = (double)m / maxhamweight;
376 return maxminsize > best;
379 static bool preconsider_ok(int nwords, bool doprint) {
382 PRINTF("%2d ", maxhamweight);
385 for (i=0, totalfrags=0; i<nwords; i++) {
386 int frags = count_set_adj_bits(adjmatrix[i]);
387 had_max += (frags >= maxhamweight);
389 PRINTF("%"PRADJ" ", adjmatrix[i]);
390 double maxminsize = (double)m / frags;
391 if (maxminsize <= best) {
397 /* Skip this candidate as its max hamming weight is lower than
398 * we're currently looking for (which means we must have done it
399 * already). (The recursive iteration ensures that none of the
400 * words have more than the max hamming weight.) */
410 static void optimise(bool doprint) {
411 /* Consider the best answer (if any) for a given adjacency matrix */
416 * Up to a certain point, optimise() can be restarted. We use this
417 * to go back and print the debugging output if it turns out that we
418 * have an interesting case. The HAVE_PRINTED macro does this: its
419 * semantics are to go back in time and make sure that we have
420 * printed the description of the search case.
422 #define HAVE_PRINTED ({ \
423 if (!doprint) { doprint = 1; goto retry_with_print; } \
427 glp_delete_prob(prob);
431 bool ok = preconsider_ok(n, doprint);
436 * We formulate our problem as an LP problem as follows.
437 * In this file "n" and "m" are the matchstick numbers.
439 * Each set bit in the adjacency matrix corresponds to taking a
440 * fragment from old match i and making it part of new match j.
442 * The structural variables (columns) are:
443 * x_minimum minimum size of any fragment (bounded below by 0)
444 * x_morefrag_i_j the amount by which the size of the fragment
445 * i,j exceeds the minimum size (bounded below by 0)
447 * The auxiliary variables (rows) are:
448 * x_total_i total length for each input match (fixed variable)
449 * x_total_j total length for each output match (fixed variable)
451 * The objective function is simply
454 * We use X_ and Y_ to refer to GLPK's (1-based) column and row indices.
455 * ME_ refers to entries in the list of constraint matrix elements
456 * which we build up as we go.
459 prob = glp_create_prob();
461 int Y_totals_i = glp_add_rows(prob, n);
462 int Y_totals_j = glp_add_rows(prob, m);
463 int X_minimum = glp_add_cols(prob, 1);
466 int next_matrix_entry = 1; /* wtf GLPK! */
467 int matrix_entries_size = next_matrix_entry + n + m + totalfrags*2;
468 double matrix_entries[matrix_entries_size];
469 int matrix_entries_XY[2][matrix_entries_size];
471 #define ADD_MATRIX_ENTRY(Y,X) ({ \
472 assert(next_matrix_entry < matrix_entries_size); \
473 matrix_entries_XY[0][next_matrix_entry] = (X); \
474 matrix_entries_XY[1][next_matrix_entry] = (Y); \
475 matrix_entries[next_matrix_entry] = 0; \
476 next_matrix_entry++; \
479 int ME_totals_i__minimum = next_matrix_entry;
480 for (i=0; i<n; i++) ADD_MATRIX_ENTRY(Y_totals_i+i, X_minimum);
482 int ME_totals_j__minimum = next_matrix_entry;
483 for (j=0; j<m; j++) ADD_MATRIX_ENTRY(Y_totals_j+j, X_minimum);
485 /* \forall_i x_total_i = m */
486 /* \forall_i x_total_j = n */
487 for (i=0; i<n; i++) glp_set_row_bnds(prob, Y_totals_i+i, GLP_FX, m,m);
488 for (j=0; j<m; j++) glp_set_row_bnds(prob, Y_totals_j+j, GLP_FX, n,n);
491 glp_set_col_bnds(prob, X_minimum, GLP_LO, 0, 0);
492 glp_set_col_name(prob, X_minimum, "minimum");
494 /* objective is maximising x_minimum */
495 glp_set_obj_dir(prob, GLP_MAX);
496 glp_set_obj_coef(prob, X_minimum, 1);
498 for (i=0; i<n; i++) {
499 for (j=0; j<m; j++) {
500 if (!(adjmatrix[i] & one_adj_bit(j)))
502 /* x_total_i += x_minimum */
503 /* x_total_j += x_minimum */
504 matrix_entries[ ME_totals_i__minimum + i ] ++;
505 matrix_entries[ ME_totals_j__minimum + j ] ++;
507 /* x_morefrag_i_j >= 0 */
508 int X_morefrag_i_j = glp_add_cols(prob, 1);
509 glp_set_col_bnds(prob, X_morefrag_i_j, GLP_LO, 0, 0);
512 snprintf(buf,sizeof(buf),"mf %d,%d",i,j);
513 glp_set_col_name(prob, X_morefrag_i_j, buf);
516 /* x_total_i += x_morefrag_i_j */
517 /* x_total_j += x_morefrag_i_j */
518 int ME_totals_i__mf_i_j = ADD_MATRIX_ENTRY(Y_totals_i+i, X_morefrag_i_j);
519 int ME_totals_j__mf_i_j = ADD_MATRIX_ENTRY(Y_totals_j+j, X_morefrag_i_j);
520 matrix_entries[ME_totals_i__mf_i_j] = 1;
521 matrix_entries[ME_totals_j__mf_i_j] = 1;
525 assert(next_matrix_entry == matrix_entries_size);
527 glp_load_matrix(prob, matrix_entries_size-1,
528 matrix_entries_XY[1], matrix_entries_XY[0],
531 int r = glp_simplex(prob, NULL);
532 PRINTF(" glp=%d", r);
535 case e: PRINTF(" " #e ); goto out;
537 case e: HAVE_PRINTED; printf(" " #e " CRASHING\n"); exit(-1);
539 default: HAVE_PRINTED; printf(" ! CRASHING\n"); exit(-1);
559 r = glp_get_status(prob);
560 PRINTF(" status=%d", r);
572 double got = glp_get_obj_val(prob);
580 multicore_found_new_best();
582 if (best_prob) glp_delete_prob(best_prob);
585 free(best_adjmatrix);
586 best_adjmatrix = xalloc_adjmatrix();
587 memcpy(best_adjmatrix, adjmatrix, sizeof(*adjmatrix)*n);
595 glp_delete_prob(prob);
596 if (doprint) progress_eol();
597 if (doprint) multicore_check_for_new_best();
600 static void iterate_recurse(int i, AdjWord min) {
603 optimise(!(printcounter & 0xfff));
606 if (i >= multicore_iteration_boundary) {
607 multicore_outer_iteration(i, min);
610 for (adjmatrix[i] = min;
613 if (count_set_adj_bits(adjmatrix[i]) > maxhamweight)
615 if (i == 0 && (adjmatrix[i] & (1+adjmatrix[i])))
618 for (int j = 0; j < m; j++)
619 if (adjmatrix[i] & one_adj_bit(j))
621 for (int j = 0; j < m; j++)
622 if (weight[j] >= n_over_best)
625 iterate_recurse(i+1, adjmatrix[i]);
628 for (int j = 0; j < m; j++)
629 if (adjmatrix[i] & one_adj_bit(j))
633 if (adjmatrix[i] == adjall)
638 static void iterate(void) {
639 for (maxhamweight=1; maxhamweight<=m; maxhamweight++) {
640 if (!maxhamweight_ok())
643 iterate_recurse(0, 1);
647 static void report(void) {
648 fprintf(stderr, "\n");
650 double min = glp_get_obj_val(best_prob);
653 for (i = 0; i < n; i++)
654 for (j = 0; j < m; j++)
656 cols = glp_get_num_cols(best_prob);
657 for (i = 1; i <= cols; i++) {
659 if (2 != sscanf(glp_get_col_name(best_prob, i), "mf %d,%d", &x, &y))
661 a[x][y] = min + glp_get_col_prim(best_prob, i);
663 printf("%d into %d: min fragment %g [%s]\n", n, m, min, VERSION);
664 for (i = 0; i < n; i++) {
665 for (j = 0; j < m; j++) {
667 printf(" %9.3f", a[i][j]);
674 if (ferror(stdout) || fclose(stdout)) { perror("stdout"); exit(-1); }
677 int main(int argc, char **argv) {
679 while ((opt = getopt(argc,argv,"j:")) >= 0) {
681 case 'j': ncpus = atoi(optarg); break;
682 case '+': assert(!"bad option");
694 if (ncpus) multicore();