/** returns the count for this quintuple if there, -1 otherwise */
s64 table_lookup(Sentence sent,
int lw, int rw, Connector *le, Connector *re, int cost)
{
Table_connector *t = find_table_pointer(sent->count_ctxt, lw, rw, le, re, cost);
if (t == NULL) return -1; else return t->count;
}
/** returns the count for this quintuple if there, -1 otherwise */
Count_bin* table_lookup(count_context_t * ctxt,
int lw, int rw, Connector *le, Connector *re,
unsigned int null_count)
{
Table_connector *t = find_table_pointer(ctxt, lw, rw, le, re, null_count);
if (t == NULL) return NULL; else return &t->count;
}
/**
* Stores the value in the table. Unlike table_store, it assumes
* it's already there
*/
static void table_update(count_context_t *ctxt, int lw, int rw,
Connector *le, Connector *re,
int cost, s64 count)
{
Table_connector *t = find_table_pointer(ctxt, lw, rw, le, re, cost);
assert(t != NULL, "This entry is supposed to be in the table.");
t->count = count;
}
static s64 do_count(Sentence sent, int lw, int rw,
Connector *le, Connector *re, int cost)
{
Disjunct * d;
s64 total, pseudototal;
int start_word, end_word, w;
s64 leftcount, rightcount;
int lcost, rcost, Lmatch, Rmatch;
Match_node * m, *m1;
Table_connector *t;
count_context_t *ctxt = sent->count_ctxt;
if (cost < 0) return 0; /* will we ever call it with cost<0 ? */
t = find_table_pointer(ctxt, lw, rw, le, re, cost);
if (t == NULL) {
/* Create the table entry with a tentative cost of 0.
* This cost must be updated before we return. */
t = table_store(ctxt, lw, rw, le, re, cost, 0);
} else {
return t->count;
}
if (rw == 1+lw)
{
/* lw and rw are neighboring words */
/* You can't have a linkage here with cost > 0 */
if ((le == NULL) && (re == NULL) && (cost == 0))
{
t->count = 1;
}
else
{
t->count = 0;
}
return t->count;
}
if ((le == NULL) && (re == NULL))
{
if (!ctxt->islands_ok && (lw != -1))
{
/* If we don't allow islands (a set of words linked together
* but separate from the rest of the sentence) then the cost
* of skipping n words is just n */
if (cost == ((rw-lw-1) + ctxt->null_block-1)/ctxt->null_block)
{
/* If null_block=4 then the cost of
1,2,3,4 nulls is 1; and 5,6,7,8 is 2 etc. */
t->count = 1;
}
else
{
t->count = 0;
}
return t->count;
}
if (cost == 0)
{
/* There is no zero-cost solution in this case. There is
* a slight efficiency hack to separate this cost=0 case
* out, but not necessary for correctness */
t->count = 0;
}
else
{
total = 0;
w = lw+1;
for (d = ctxt->local_sent[w].d; d != NULL; d = d->next)
{
if (d->left == NULL)
{
total += do_count(sent, w, rw, d->right, NULL, cost-1);
}
}
total += do_count(sent, w, rw, NULL, NULL, cost-1);
t->count = total;
}
return t->count;
}
if (le == NULL)
{
start_word = lw+1;
}
else
{
start_word = le->word;
}
if (re == NULL)
{
end_word = rw-1;
}
else
{
end_word = re->word;
}
total = 0;
for (w = start_word; w < end_word+1; w++)
{
m1 = m = form_match_list(sent, w, le, lw, re, rw);
for (; m!=NULL; m=m->next)
{
d = m->d;
for (lcost = 0; lcost <= cost; lcost++)
{
rcost = cost-lcost;
/* Now lcost and rcost are the costs we're assigning
* to those parts respectively */
/* Now, we determine if (based on table only) we can see that
the current range is not parsable. */
Lmatch = (le != NULL) && (d->left != NULL) &&
do_match(sent, le, d->left, lw, w);
Rmatch = (d->right != NULL) && (re != NULL) &&
do_match(sent, d->right, re, w, rw);
rightcount = leftcount = 0;
if (Lmatch)
{
leftcount = pseudocount(sent, lw, w, le->next, d->left->next, lcost);
if (le->multi) leftcount += pseudocount(sent, lw, w, le, d->left->next, lcost);
if (d->left->multi) leftcount += pseudocount(sent, lw, w, le->next, d->left, lcost);
if (le->multi && d->left->multi) leftcount += pseudocount(sent, lw, w, le, d->left, lcost);
}
if (Rmatch)
{
rightcount = pseudocount(sent, w, rw, d->right->next, re->next, rcost);
if (d->right->multi) rightcount += pseudocount(sent, w,rw,d->right,re->next, rcost);
if (re->multi) rightcount += pseudocount(sent, w, rw, d->right->next, re, rcost);
if (d->right->multi && re->multi) rightcount += pseudocount(sent, w, rw, d->right, re, rcost);
}
/* total number where links are used on both sides */
pseudototal = leftcount*rightcount;
if (leftcount > 0) {
/* evaluate using the left match, but not the right */
pseudototal += leftcount * pseudocount(sent, w, rw, d->right, re, rcost);
}
if ((le == NULL) && (rightcount > 0)) {
/* evaluate using the right match, but not the left */
pseudototal += rightcount * pseudocount(sent, lw, w, le, d->left, lcost);
}
/* now pseudototal is 0 implies that we know that the true total is 0 */
if (pseudototal != 0) {
rightcount = leftcount = 0;
if (Lmatch) {
leftcount = do_count(sent, lw, w, le->next, d->left->next, lcost);
if (le->multi) leftcount += do_count(sent, lw, w, le, d->left->next, lcost);
if (d->left->multi) leftcount += do_count(sent, lw, w, le->next, d->left, lcost);
if (le->multi && d->left->multi) leftcount += do_count(sent, lw, w, le, d->left, lcost);
}
if (Rmatch) {
rightcount = do_count(sent, w, rw, d->right->next, re->next, rcost);
if (d->right->multi) rightcount += do_count(sent, w,rw,d->right,re->next, rcost);
if (re->multi) rightcount += do_count(sent, w, rw, d->right->next, re, rcost);
if (d->right->multi && re->multi) rightcount += do_count(sent, w, rw, d->right, re, rcost);
}
total += leftcount*rightcount; /* total number where links are used on both sides */
if (leftcount > 0) {
/* evaluate using the left match, but not the right */
total += leftcount * do_count(sent, w, rw, d->right, re, rcost);
}
if ((le == NULL) && (rightcount > 0)) {
/* evaluate using the right match, but not the left */
total += rightcount * do_count(sent, lw, w, le, d->left, lcost);
}
}
}
}
put_match_list(sent, m1);
}
t->count = total;
return total;
}
static Count_bin do_count(fast_matcher_t *mchxt,
count_context_t *ctxt,
int lw, int rw,
Connector *le, Connector *re,
int null_count)
{
Count_bin zero = hist_zero();
Count_bin total;
int start_word, end_word, w;
Table_connector *t;
assert (0 <= null_count, "Bad null count");
t = find_table_pointer(ctxt, lw, rw, le, re, null_count);
if (t) return t->count;
/* Create the table entry with a tentative null count of 0.
* This count must be updated before we return. */
t = table_store(ctxt, lw, rw, le, re, null_count);
if (rw == 1+lw)
{
/* lw and rw are neighboring words */
/* You can't have a linkage here with null_count > 0 */
if ((le == NULL) && (re == NULL) && (null_count == 0))
{
t->count = hist_one();
}
else
{
t->count = zero;
}
return t->count;
}
/* The left and right connectors are null, but the two words are
* NOT next to each-other. */
if ((le == NULL) && (re == NULL))
{
if (!ctxt->islands_ok && (lw != -1))
{
/* If we don't allow islands (a set of words linked together
* but separate from the rest of the sentence) then the
* null_count of skipping n words is just n. */
if (null_count == (rw-lw-1))
{
t->count = hist_one();
}
else
{
t->count = zero;
}
return t->count;
}
if (null_count == 0)
{
/* There is no solution without nulls in this case. There is
* a slight efficiency hack to separate this null_count==0
* case out, but not necessary for correctness */
t->count = zero;
}
else
{
t->count = zero;
Disjunct * d;
int w = lw + 1;
for (d = ctxt->local_sent[w].d; d != NULL; d = d->next)
{
if (d->left == NULL)
{
hist_accumv(&t->count, d->cost,
do_count(mchxt, ctxt, w, rw, d->right, NULL, null_count-1));
}
}
hist_accumv(&t->count, 0.0,
do_count(mchxt, ctxt, w, rw, NULL, NULL, null_count-1));
}
return t->count;
}
if (le == NULL)
{
start_word = lw+1;
}
else
{
start_word = le->word;
}
if (re == NULL)
{
end_word = rw;
}
else
{
end_word = re->word +1;
}
total = zero;
for (w = start_word; w < end_word; w++)
{
Match_node *m, *m1;
m1 = m = form_match_list(mchxt, w, le, lw, re, rw);
for (; m != NULL; m = m->next)
{
unsigned int lnull_cnt, rnull_cnt;
Disjunct * d = m->d;
/* _p1 avoids a gcc warning about unsafe loop opt */
unsigned int null_count_p1 = null_count + 1;
for (lnull_cnt = 0; lnull_cnt < null_count_p1; lnull_cnt++)
{
bool Lmatch, Rmatch;
bool leftpcount = false;
bool rightpcount = false;
bool pseudototal = false;
rnull_cnt = null_count - lnull_cnt;
/* Now lnull_cnt and rnull_cnt are the costs we're assigning
* to those parts respectively */
/* Now, we determine if (based on table only) we can see that
the current range is not parsable. */
Lmatch = (le != NULL) && (d->left != NULL) &&
do_match(le, d->left, lw, w);
Rmatch = (d->right != NULL) && (re != NULL) &&
do_match(d->right, re, w, rw);
/* First, perform pseudocounting as an optimization. If
* the pseudocount is zero, then we know that the true
* count will be zero, and so skip counting entirely,
* in that case.
*/
if (Lmatch)
{
leftpcount = pseudocount(ctxt, lw, w, le->next, d->left->next, lnull_cnt);
if (!leftpcount && le->multi)
leftpcount =
pseudocount(ctxt, lw, w, le, d->left->next, lnull_cnt);
if (!leftpcount && d->left->multi)
leftpcount =
pseudocount(ctxt, lw, w, le->next, d->left, lnull_cnt);
if (!leftpcount && le->multi && d->left->multi)
leftpcount =
pseudocount(ctxt, lw, w, le, d->left, lnull_cnt);
}
if (Rmatch)
{
rightpcount = pseudocount(ctxt, w, rw, d->right->next, re->next, rnull_cnt);
if (!rightpcount && d->right->multi)
rightpcount =
pseudocount(ctxt, w,rw, d->right, re->next, rnull_cnt);
if (!rightpcount && re->multi)
rightpcount =
pseudocount(ctxt, w, rw, d->right->next, re, rnull_cnt);
if (!rightpcount && d->right->multi && re->multi)
rightpcount =
pseudocount(ctxt, w, rw, d->right, re, rnull_cnt);
}
/* Total number where links are used on both sides */
pseudototal = leftpcount && rightpcount;
if (!pseudototal && leftpcount) {
/* Evaluate using the left match, but not the right. */
pseudototal =
pseudocount(ctxt, w, rw, d->right, re, rnull_cnt);
}
if (!pseudototal && (le == NULL) && rightpcount) {
/* Evaluate using the right match, but not the left. */
pseudototal =
pseudocount(ctxt, lw, w, le, d->left, lnull_cnt);
}
/* If pseudototal is zero (false), that implies that
* we know that the true total is zero. So we don't
* bother counting at all, in that case. */
if (pseudototal)
{
Count_bin leftcount = zero;
Count_bin rightcount = zero;
if (Lmatch) {
leftcount = do_count(mchxt, ctxt, lw, w, le->next, d->left->next, lnull_cnt);
if (le->multi)
hist_accumv(&leftcount, d->cost,
do_count(mchxt, ctxt, lw, w, le, d->left->next, lnull_cnt));
if (d->left->multi)
hist_accumv(&leftcount, d->cost,
do_count(mchxt, ctxt, lw, w, le->next, d->left, lnull_cnt));
if (le->multi && d->left->multi)
hist_accumv(&leftcount, d->cost,
do_count(mchxt, ctxt, lw, w, le, d->left, lnull_cnt));
}
if (Rmatch) {
rightcount = do_count(mchxt, ctxt, w, rw, d->right->next, re->next, rnull_cnt);
if (d->right->multi)
hist_accumv(&rightcount, d->cost,
do_count(mchxt, ctxt, w, rw, d->right,re->next, rnull_cnt));
if (re->multi)
hist_accumv(&rightcount, d->cost,
do_count(mchxt, ctxt, w, rw, d->right->next, re, rnull_cnt));
if (d->right->multi && re->multi)
hist_accumv(&rightcount, d->cost,
do_count(mchxt, ctxt, w, rw, d->right, re, rnull_cnt));
}
/* Total number where links are used on both sides */
hist_muladd(&total, &leftcount, 0.0, &rightcount);
if (0 < hist_total(&leftcount))
{
/* Evaluate using the left match, but not the right */
hist_muladdv(&total, &leftcount, d->cost,
do_count(mchxt, ctxt, w, rw, d->right, re, rnull_cnt));
}
if ((le == NULL) && (0 < hist_total(&rightcount)))
{
/* Evaluate using the right match, but not the left */
hist_muladdv(&total, &rightcount, d->cost,
do_count(mchxt, ctxt, lw, w, le, d->left, lnull_cnt));
}
/* Sigh. Overflows can and do occur, esp for the ANY language. */
if (INT_MAX < hist_total(&total))
{
#ifdef PERFORM_COUNT_HISTOGRAMMING
total.total = INT_MAX;
#else
total = INT_MAX;
#endif /* PERFORM_COUNT_HISTOGRAMMING */
t->count = total;
put_match_list(mchxt, m1);
return total;
}
}
}
}
put_match_list(mchxt, m1);
}
t->count = total;
return total;
}
