* GNU General Public License for more details.
*/
+#define _GNU_SOURCE
+
+#include <publib.h>
+
#include <stdio.h>
#include <stdint.h>
#include <stdlib.h>
#include <unistd.h>
#include <stdbool.h>
#include <inttypes.h>
+#include <math.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <sys/uio.h>
+#include <sys/fcntl.h>
-#include <publib.h>
#include <glpk.h>
+#ifndef VERSION
+#define VERSION "(unknown-version)"
+#endif
+
/*
* Algorithm.
*
typedef uint32_t AdjWord;
#define PRADJ "08"PRIx32
+#define FOR_BITS(j,m) for (j=0, j##bit=1; j < (m); j++, j##bit<<=1)
+
static int n, m, maxhamweight;
static AdjWord *adjmatrix;
static AdjWord adjall;
static glp_prob *best_prob;
static AdjWord *best_adjmatrix;
+static int n_max_frags, m_max_frags;
+static int *weight;
+
static unsigned printcounter;
-static int ncpus = 1;
+static void iterate(void);
+static void iterate_recurse(int i, AdjWord min);
+static bool preconsider_ok(int nwords, bool doprint);
+static bool maxhamweight_ok(void);
+static void optimise(bool doprint);
static void progress_eol(void) {
fprintf(stderr," \r");
fflush(stderr);
}
+static void set_best(double new_best) {
+ best = new_best;
+ /*
+ * When computing n_max_frags, we want to set a value that will skip
+ * anything that won't provide strictly better solutions. So we
+ * want
+ * frags < n / best
+ * _ _
+ * <=> frags < | n / best |
+ * _ _
+ * <=> frags <= | n / best | - 1
+ */
+ n_max_frags = ceil(n / best) - 1;
+ m_max_frags = ceil(m / best) - 1;
+}
+
+/*----- multicore support -----*/
+
+/*
+ * Multicore protocol
+ *
+ * We fork into:
+ * - master (parent)
+ * - generator
+ * - ncpu workers
+ *
+ * ipc facilities:
+ * - one pipe ("work") from generator to workers
+ * - ever-extending file ("bus") containing new "best" values
+ * - one file for each worker giving maxhamweight and adjmatrix for best
+ *
+ * generator runs iterate_recurse to a certain depth and writes the
+ * candidates to a pipe
+ *
+ * workers read candidates from the pipe and resume iterate_recurse
+ * halfway through the recursion
+ *
+ * whenever a worker does a doprint, it checks the bus for new best
+ * value; actual best values are appended
+ *
+ * master waits for generator and all workers to finish and then
+ * runs optimise() for each worker's best, then prints
+ */
+
+static int ncpus = 0, multicore_iteration_boundary = INT_MAX;
+
+static int mc_bus, mc_work[2];
+static off_t mc_bus_read;
+
+typedef struct {
+ int w;
+ FILE *results;
+ pid_t pid;
+} Worker;
+static Worker *mc_us;
+static bool mc_am_generator;
+
+static void multicore_check_for_new_best(void);
+
+#define MAX_NIOVS 4
+static AdjWord mc_iter_min;
+static int mc_niovs;
+static size_t mc_iovlen;
+static struct iovec mc_iov[MAX_NIOVS];
+
+#define IOV0 (mc_niovs = mc_iovlen = 0)
+
+#define IOV(obj, count) ({ \
+ assert(mc_niovs < MAX_NIOVS); \
+ mc_iov[mc_niovs].iov_base = &(obj); \
+ mc_iov[mc_niovs].iov_len = sizeof(obj) * (count); \
+ mc_iovlen += mc_iov[mc_niovs].iov_len; \
+ mc_niovs++; \
+ })
+
+static void mc_rwvsetup_outer(void) {
+ IOV0;
+ IOV(maxhamweight, 1);
+ IOV(mc_iter_min, 1);
+ IOV(*adjmatrix, multicore_iteration_boundary);
+ IOV(*weight, m);
+}
+
+static void mc_rwvsetup_full(void) {
+ IOV0;
+ IOV(*adjmatrix, n);
+}
+
+static void vlprintf(const char *fmt, va_list al) {
+ vfprintf(stderr,fmt,al);
+ progress_eol();
+}
+
+static void LPRINTF(const char *fmt, ...) {
+ va_list al;
+ va_start(al,fmt);
+ vlprintf(fmt,al);
+ va_end(al);
+}
+
+static void mc_awaitpid(int wnum, pid_t pid) {
+ LPRINTF("master awaiting %2d [%ld]",wnum,(long)pid);
+ int status;
+ pid_t got = waitpid(pid, &status, 0);
+ assert(got == pid);
+ if (status) {
+ fprintf(stderr,"\nFAILED SUBPROC %2d [%ld] %d\n",
+ wnum, (long)pid, status);
+ exit(-1);
+ }
+}
+
+static void multicore_outer_iteration(int i, AdjWord min) {
+ static unsigned check_counter;
+
+ assert(i == multicore_iteration_boundary);
+ mc_iter_min = min;
+ mc_rwvsetup_outer();
+ ssize_t r = writev(mc_work[1], mc_iov, mc_niovs);
+ assert(r == mc_iovlen);
+ /* effectively, this writev arranges to transfers control
+ * to some worker's instance of iterate_recurse via mc_iterate_worker */
+
+ if (!(check_counter++ & 0xff))
+ multicore_check_for_new_best();
+}
+
+static void mc_iterate_worker(void) {
+ for (;;) {
+ mc_rwvsetup_outer();
+ ssize_t r = readv(mc_work[0], mc_iov, mc_niovs);
+ if (r == 0) break;
+ assert(r == mc_iovlen);
+
+ bool ok = maxhamweight_ok();
+ if (!ok) continue;
+
+ ok = preconsider_ok(multicore_iteration_boundary, 1);
+ progress_eol();
+ if (!ok) continue;
+
+ /* stop iterate_recurse from trying to run multicore_outer_iteration */
+ int mc_org_it_bound = multicore_iteration_boundary;
+ multicore_iteration_boundary = INT_MAX;
+ iterate_recurse(mc_org_it_bound, mc_iter_min);
+ multicore_iteration_boundary = mc_org_it_bound;
+ }
+ if (best_adjmatrix) {
+ LPRINTF("worker %2d reporting",mc_us->w);
+ adjmatrix = best_adjmatrix;
+ mc_rwvsetup_full();
+ ssize_t r = writev(fileno(mc_us->results), mc_iov, mc_niovs);
+ assert(r == mc_iovlen);
+ }
+ LPRINTF("worker %2d ending",mc_us->w);
+ exit(0);
+}
+
+static void multicore(void) {
+ Worker *mc_workers;
+ int w;
+ pid_t genpid;
+
+ multicore_iteration_boundary = n / 2;
+
+ FILE *busf = tmpfile(); assert(busf);
+ mc_bus = fileno(busf);
+ int r = fcntl(mc_bus, F_GETFL); assert(r >= 0);
+ r |= O_APPEND;
+ r = fcntl(mc_bus, F_SETFL, r); assert(r >= 0);
+
+ r = pipe(mc_work); assert(!r);
+
+ mc_workers = xmalloc(sizeof(*mc_workers) * ncpus);
+ for (w=0; w<ncpus; w++) {
+ mc_workers[w].w = w;
+ mc_workers[w].results = tmpfile(); assert(mc_workers[w].results);
+ mc_workers[w].pid = fork(); assert(mc_workers[w].pid >= 0);
+ if (!mc_workers[w].pid) {
+ mc_us = &mc_workers[w];
+ close(mc_work[1]);
+ LPRINTF("worker %2d running", w);
+ mc_iterate_worker();
+ exit(0);
+ }
+ }
+
+ close(mc_work[0]);
+
+ genpid = fork(); assert(genpid >= 0);
+ if (!genpid) {
+ mc_am_generator = 1;
+ LPRINTF("generator running");
+ iterate();
+ exit(0);
+ }
+
+ close(mc_work[1]);
+ mc_awaitpid(-1, genpid);
+ for (w=0; w<ncpus; w++)
+ mc_awaitpid(w, mc_workers[w].pid);
+
+ for (w=0; w<ncpus; w++) {
+ mc_rwvsetup_full();
+ LPRINTF("reading report from %2d",w);
+ ssize_t sr = preadv(fileno(mc_workers[w].results), mc_iov, mc_niovs, 0);
+ if (!sr) continue;
+ LPRINTF("got report from %2d",w);
+ maxhamweight = 0;
+ optimise(1);
+ }
+}
+
+static void multicore_check_for_new_best(void) {
+ if (!(mc_us || mc_am_generator))
+ return;
+
+ for (;;) {
+ double msg;
+ ssize_t got = pread(mc_bus, &msg, sizeof(msg), mc_bus_read);
+ if (!got) break;
+ assert(got == sizeof(msg));
+ if (msg > best)
+ set_best(msg);
+ mc_bus_read += sizeof(msg);
+ }
+}
+
+static void multicore_found_new_best(void) {
+ if (!mc_us)
+ return;
+
+ if (mc_us /* might be master */) fprintf(stderr," w%-2d ",mc_us->w);
+ ssize_t wrote = write(mc_bus, &best, sizeof(best));
+ assert(wrote == sizeof(best));
+}
+
+/*----- end of multicore support -----*/
+
static AdjWord *xalloc_adjmatrix(void) {
return xmalloc(sizeof(*adjmatrix)*n);
}
adjmatrix = xalloc_adjmatrix();
glp_term_out(GLP_OFF);
setlinebuf(stderr);
+ weight = calloc(sizeof(*weight), m); assert(weight);
+ n_max_frags = INT_MAX;
+ m_max_frags = INT_MAX;
}
+#if 0
static AdjWord one_adj_bit(int bitnum) {
return (AdjWord)1 << bitnum;
}
+#endif
static int count_set_adj_bits(AdjWord w) {
- int j, total;
- for (j=0, total=0; j<m; j++)
- total += !!(w & one_adj_bit(j));
+ int j, total = 0;
+ AdjWord jbit;
+ FOR_BITS(j,m)
+ total += !!(w & jbit);
return total;
}
static int totalfrags;
+static bool maxhamweight_ok(void) {
+ return maxhamweight <= m_max_frags;
+}
+
static bool preconsider_ok(int nwords, bool doprint) {
int i;
bool had_max = 0;
for (i=0, totalfrags=0; i<nwords; i++) {
int frags = count_set_adj_bits(adjmatrix[i]);
- had_max += (frags >= maxhamweight);
- totalfrags += frags;
PRINTF("%"PRADJ" ", adjmatrix[i]);
- double maxminsize = (double)m / frags;
- if (maxminsize <= best) {
+ if (frags > m_max_frags) {
PRINTF(" too fine");
goto out;
}
+ had_max += (frags >= maxhamweight);
+ totalfrags += frags;
}
if (!had_max) {
/* Skip this candidate as its max hamming weight is lower than
/* Consider the best answer (if any) for a given adjacency matrix */
glp_prob *prob = 0;
int i, j;
+ AdjWord jbit;
/*
* Up to a certain point, optimise() can be restarted. We use this
glp_set_obj_coef(prob, X_minimum, 1);
for (i=0; i<n; i++) {
- for (j=0; j<m; j++) {
- if (!(adjmatrix[i] & one_adj_bit(j)))
+ FOR_BITS(j,m) {
+ if (!(adjmatrix[i] & jbit))
continue;
/* x_total_i += x_minimum */
/* x_total_j += x_minimum */
HAVE_PRINTED;
- best = got;
+ set_best(got);
+ multicore_found_new_best();
if (best_prob) glp_delete_prob(best_prob);
best_prob = prob;
if (prob)
glp_delete_prob(prob);
if (doprint) progress_eol();
+ if (doprint) multicore_check_for_new_best();
}
static void iterate_recurse(int i, AdjWord min) {
+ int j;
+ AdjWord jbit;
+
if (i >= n) {
printcounter++;
optimise(!(printcounter & 0xfff));
return;
}
+ if (i >= multicore_iteration_boundary) {
+ multicore_outer_iteration(i, min);
+ return;
+ }
for (adjmatrix[i] = min;
;
adjmatrix[i]++) {
if (i == 0 && (adjmatrix[i] & (1+adjmatrix[i])))
goto again;
+ FOR_BITS(j,m)
+ if (adjmatrix[i] & jbit)
+ weight[j]++;
+ for (int j = 0; j < m; j++)
+ if (weight[j] >= n_max_frags)
+ goto takeout;
+
iterate_recurse(i+1, adjmatrix[i]);
+ takeout:
+ FOR_BITS(j,m)
+ if (adjmatrix[i] & jbit)
+ weight[j]--;
+
again:
if (adjmatrix[i] == adjall)
return;
static void iterate(void) {
for (maxhamweight=1; maxhamweight<=m; maxhamweight++) {
- double maxminsize = (double)m / maxhamweight;
- if (maxminsize <= best)
+ if (!maxhamweight_ok())
continue;
iterate_recurse(0, 1);
continue;
a[x][y] = min + glp_get_col_prim(best_prob, i);
}
- printf("%d into %d: min fragment %g\n", n, m, min);
+ printf("%d into %d: min fragment %g [%s]\n", n, m, min, VERSION);
for (i = 0; i < n; i++) {
for (j = 0; j < m; j++) {
if (a[i][j])
m = atoi(argv[2]);
prep();
- iterate();
+
+ if (ncpus) multicore();
+ else iterate();
+
report();
return 0;
}