{n,m}_max_frags: fix fencepost efficiency problem

[matchsticks-search.git] / main.c

diff --git a/main.c b/main.c
index fd0569be80e712602988a5f79207b77bcb45ff5b..11bdf0fba64e7605e11e54511142be5d20bbf689 100644 (file)
--- a/main.c
+++ b/main.c
@@ -34,6 +34,7 @@
 #include <unistd.h>
 #include <stdbool.h>
 #include <inttypes.h>
+#include <math.h>
 #include <sys/types.h>
 #include <sys/wait.h>
 #include <sys/uio.h>
@@ -41,6 +42,10 @@
 
 #include <glpk.h>
 
+#ifndef VERSION
+#define VERSION "(unknown-version)"
+#endif
+
 /*
  * Algorithm.
  *
@@ -86,6 +91,8 @@
 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;
@@ -94,6 +101,9 @@ static double best;
 static glp_prob *best_prob;
 static AdjWord *best_adjmatrix;
 
+static int n_max_frags, m_max_frags;
+static int *weight;
+
 static unsigned printcounter;
 
 static void iterate(void);
@@ -109,6 +119,18 @@ static void progress_eol(void) {
 
 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 -----*/
@@ -150,10 +172,11 @@ typedef struct {
   pid_t pid;
 } Worker;
 static Worker *mc_us;
+static bool mc_am_generator;
 
 static void multicore_check_for_new_best(void);
 
-#define MAX_NIOVS 3
+#define MAX_NIOVS 4
 static AdjWord mc_iter_min;
 static int mc_niovs;
 static size_t mc_iovlen;
@@ -174,6 +197,7 @@ static void mc_rwvsetup_outer(void) {
   IOV(maxhamweight, 1);
   IOV(mc_iter_min, 1);
   IOV(*adjmatrix, multicore_iteration_boundary);
+  IOV(*weight, m);
 }
 
 static void mc_rwvsetup_full(void) {
@@ -240,8 +264,8 @@ static void mc_iterate_worker(void) {
     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) {
+    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);
@@ -284,6 +308,7 @@ static void multicore(void) {
 
   genpid = fork();  assert(genpid >= 0);
   if (!genpid) {
+    mc_am_generator = 1;
     LPRINTF("generator running");
     iterate();
     exit(0);
@@ -299,13 +324,15 @@ static void multicore(void) {
     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 (!ncpus) return;
+  if (!(mc_us || mc_am_generator))
+    return;
 
   for (;;) {
     double msg;
@@ -319,7 +346,8 @@ static void multicore_check_for_new_best(void) {
 }
 
 static void multicore_found_new_best(void) {
-  if (!ncpus) return;
+  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));
@@ -337,16 +365,22 @@ static void prep(void) {
   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;
 }
 
@@ -355,8 +389,7 @@ static int count_set_adj_bits(AdjWord w) {
 static int totalfrags;
 
 static bool maxhamweight_ok(void) {
-  double maxminsize = (double)m / maxhamweight;
-  return maxminsize > best;
+  return maxhamweight <= m_max_frags;
 }
 
 static bool preconsider_ok(int nwords, bool doprint) {
@@ -367,14 +400,13 @@ static bool preconsider_ok(int nwords, bool doprint) {
   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
@@ -394,6 +426,7 @@ static void optimise(bool doprint) {
   /* 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
@@ -479,8 +512,8 @@ static void optimise(bool doprint) {
   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 (j=0, jbit=1; j<m; j++, jbit<<=1) {
+      if (!(adjmatrix[i] & jbit))
        continue;
       /* x_total_i += x_minimum */
       /* x_total_j += x_minimum */
@@ -581,6 +614,9 @@ static void optimise(bool doprint) {
 }
 
 static void iterate_recurse(int i, AdjWord min) {
+  int j;
+  AdjWord jbit;
+
   if (i >= n) {
     printcounter++;
     optimise(!(printcounter & 0xfff));
@@ -598,8 +634,20 @@ static void iterate_recurse(int i, AdjWord min) {
     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;
@@ -631,7 +679,7 @@ static void report(void) {
         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])