From: Ian Jackson <ijackson@chiark.greenend.org.uk>
Date: Sat, 8 Mar 2014 22:11:59 +0000 (+0000)
Subject: Merge remote-tracking branch 'sgt/master'
X-Git-Url: http://www.chiark.greenend.org.uk/ucgi/~ianmdlvl/git?a=commitdiff_plain;h=8e7519302ad8da8ecb172cd863e8835abfa7b9e1;hp=-c;p=matchsticks-search.git

Merge remote-tracking branch 'sgt/master'

Changes made to the new code in sgt/master:
  - use "-std=gnu99" not "--std=gnu99"
  - n_over_best calculation is in set_best
  - n_over_best initialised properly
 style:
  - use calloc, do not cast return value from int, better sizeof
  - extra blank line

Conflicts:
	Makefile
	main.c
---

8e7519302ad8da8ecb172cd863e8835abfa7b9e1
diff --combined main.c
index fd0569b,3c1b123..2fc3a51
--- a/main.c
+++ b/main.c
@@@ -22,23 -22,15 +22,23 @@@
   * GNU General Public License for more details.
   */
  
 +#define _GNU_SOURCE
 +
 +#include <publib.h>
 +
  #include <stdio.h>
  #include <stdint.h>
  #include <stdlib.h>
  #include <string.h>
  #include <assert.h>
 +#include <unistd.h>
  #include <stdbool.h>
  #include <inttypes.h>
 +#include <sys/types.h>
 +#include <sys/wait.h>
 +#include <sys/uio.h>
 +#include <sys/fcntl.h>
  
 -#include <publib.h>
  #include <glpk.h>
  
  /*
@@@ -94,240 -86,11 +94,244 @@@ static double best
  static glp_prob *best_prob;
  static AdjWord *best_adjmatrix;
  
+ static int n_over_best;
+ static int *weight;
+ 
  static unsigned printcounter;
  
 +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;
++  n_over_best = n / best;
 +}
 +
 +/*----- 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 void multicore_check_for_new_best(void);
 +
 +#define MAX_NIOVS 3
 +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);
 +}
 +
 +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;
 +  }
 +  LPRINTF("worker %2d reporting",mc_us->w);
 +  if (best_adjmatrix) {
 +    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) {
 +    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;
 +    maxhamweight = 0;
 +    optimise(1);
 +  }
 +}
 +
 +static void multicore_check_for_new_best(void) {
 +  if (!ncpus) 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 (!ncpus) 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);
  }
@@@ -336,7 -99,9 +340,9 @@@ static void prep(void) 
    adjall = ~((~(AdjWord)0) << m);
    adjmatrix = xalloc_adjmatrix();
    glp_term_out(GLP_OFF);
 -  n_over_best = n / best;
 -  weight = (int *)malloc(m*sizeof(int));
 -  for (int j = 0; j < m; j++) weight[j] = 0;
 +  setlinebuf(stderr);
++  weight = calloc(sizeof(*weight), m);  assert(weight);
++  n_over_best = INT_MAX;
  }
  
  static AdjWord one_adj_bit(int bitnum) {
@@@ -350,24 -115,35 +356,24 @@@ static int count_set_adj_bits(AdjWord w
    return total;
  }
  
 -static void optimise(int doprint) {
 -  /* Consider the best answer (if any) for a given adjacency matrix */
 -  glp_prob *prob = 0;
 -  int i, j, totalfrags;
 +#define PRINTF(...) if (!doprint) ; else fprintf(stderr, __VA_ARGS__)
  
 -  /*
 -   * Up to a certain point, optimise() can be restarted.  We use this
 -   * to go back and print the debugging output if it turns out that we
 -   * have an interesting case.  The HAVE_PRINTED macro does this: its
 -   * semantics are to go back in time and make sure that we have
 -   * printed the description of the search case.
 -   */
 -#define HAVE_PRINTED ({						\
 -      if (!doprint) { doprint = 1; goto retry_with_print; }	\
 -    })
 - retry_with_print:
 -  if (prob) {
 -    glp_delete_prob(prob);
 -    prob = 0;
 -  }
 +static int totalfrags;
  
 -#define PRINTF(...) if (!doprint) ; else fprintf(stderr, __VA_ARGS__) /* bodgy */
 +static bool maxhamweight_ok(void) {
 +  double maxminsize = (double)m / maxhamweight;
 +  return maxminsize > best;
 +}
 +
 +static bool preconsider_ok(int nwords, bool doprint) {
 +  int i;
  
    PRINTF("%2d ", maxhamweight);
  
    bool had_max = 0;
 -  for (i=0, totalfrags=0; i<n; i++) {
 +  for (i=0, totalfrags=0; i<nwords; i++) {
      int frags = count_set_adj_bits(adjmatrix[i]);
 -    had_max += (frags == maxhamweight);
 +    had_max += (frags >= maxhamweight);
      totalfrags += frags;
      PRINTF("%"PRADJ" ", adjmatrix[i]);
      double maxminsize = (double)m / frags;
@@@ -384,36 -160,6 +390,36 @@@
      PRINTF(" nomaxham");
      goto out;
    }
 +  return 1;
 +
 + out:
 +  return 0;
 +}
 +
 +static void optimise(bool doprint) {
 +  /* Consider the best answer (if any) for a given adjacency matrix */
 +  glp_prob *prob = 0;
 +  int i, j;
 +
 +  /*
 +   * Up to a certain point, optimise() can be restarted.  We use this
 +   * to go back and print the debugging output if it turns out that we
 +   * have an interesting case.  The HAVE_PRINTED macro does this: its
 +   * semantics are to go back in time and make sure that we have
 +   * printed the description of the search case.
 +   */
 +#define HAVE_PRINTED ({						\
 +      if (!doprint) { doprint = 1; goto retry_with_print; }	\
 +    })
 + retry_with_print:
 +  if (prob) {
 +    glp_delete_prob(prob);
 +    prob = 0;
 +  }
 +
 +  bool ok = preconsider_ok(n, doprint);
 +  if (!ok)
 +    goto out;
  
    /*
     * We formulate our problem as an LP problem as follows.
@@@ -559,8 -305,8 +565,8 @@@
  
    HAVE_PRINTED;
  
 -  best = got;
 -  n_over_best = n / best;
 +  set_best(got);
 +  multicore_found_new_best();
  
    if (best_prob) glp_delete_prob(best_prob);
    best_prob = prob;
@@@ -576,8 -322,7 +582,8 @@@
   out:
    if (prob)
      glp_delete_prob(prob);
 -  if (doprint) { PRINTF("        \r"); fflush(stdout); }
 +  if (doprint) progress_eol();
 +  if (doprint) multicore_check_for_new_best();
  }
  
  static void iterate_recurse(int i, AdjWord min) {
@@@ -586,10 -331,6 +592,10 @@@
      optimise(!(printcounter & 0xfff));
      return;
    }
 +  if (i >= multicore_iteration_boundary) {
 +    multicore_outer_iteration(i, min);
 +    return;
 +  }
    for (adjmatrix[i] = min;
         ;
         adjmatrix[i]++) {
@@@ -597,9 -338,20 +603,21 @@@
        goto again;
      if (i == 0 && (adjmatrix[i] & (1+adjmatrix[i])))
        goto again;
 +
+     for (int j = 0; j < m; j++)
 -      if (adjmatrix[i] & (1 << j))
++      if (adjmatrix[i] & one_adj_bit(j))
+         weight[j]++;
+     for (int j = 0; j < m; j++)
+       if (weight[j] >= n_over_best)
+         goto takeout;
+ 
      iterate_recurse(i+1, adjmatrix[i]);
  
 -   takeout:
++  takeout:
+     for (int j = 0; j < m; j++)
 -      if (adjmatrix[i] & (1 << j))
++      if (adjmatrix[i] & one_adj_bit(j))
+         weight[j]--;
+ 
    again:
      if (adjmatrix[i] == adjall)
        return;
@@@ -608,14 -360,20 +626,14 @@@
  
  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);
    }
  }
  
 -int main(int argc, char **argv) {
 -  assert(argc==3);
 -  n = atoi(argv[1]);
 -  m = atoi(argv[2]);
 -  prep();
 -  iterate();
 +static void report(void) {
    fprintf(stderr, "\n");
    if (best_prob) {
      double min = glp_get_obj_val(best_prob);
@@@ -643,28 -401,5 +661,28 @@@
      }
    }
    if (ferror(stdout) || fclose(stdout)) { perror("stdout"); exit(-1); }
 +}
 + 
 +int main(int argc, char **argv) {
 +  int opt;
 +  while ((opt = getopt(argc,argv,"j:")) >= 0) {
 +    switch (opt) {
 +    case 'j': ncpus = atoi(optarg); break;
 +    case '+': assert(!"bad option");
 +    default: abort();
 +    }
 +  }
 +  argc -= optind-1;
 +  argv += optind-1;
 +  assert(argc==3);
 +  n = atoi(argv[1]);
 +  m = atoi(argv[2]);
 +
 +  prep();
 +
 +  if (ncpus) multicore();
 +  else iterate();
 +
 +  report();
    return 0;
  }