3 * Teodor Sigaev <teodor@sigaev.ru>
9 #include "access/gist.h"
10 #include "access/itup.h"
11 #include "catalog/namespace.h"
12 #include "commands/trigger.h"
13 #include "executor/spi.h"
16 #include "nodes/pg_list.h"
17 #include "storage/bufpage.h"
18 #include "utils/array.h"
19 #include "utils/builtins.h"
25 PG_FUNCTION_INFO_V1(rank);
26 Datum rank(PG_FUNCTION_ARGS);
28 PG_FUNCTION_INFO_V1(rank_def);
29 Datum rank_def(PG_FUNCTION_ARGS);
31 PG_FUNCTION_INFO_V1(rank_cd);
32 Datum rank_cd(PG_FUNCTION_ARGS);
34 PG_FUNCTION_INFO_V1(rank_cd_def);
35 Datum rank_cd_def(PG_FUNCTION_ARGS);
37 PG_FUNCTION_INFO_V1(get_covers);
38 Datum get_covers(PG_FUNCTION_ARGS);
40 static float weights[] = {0.1, 0.2, 0.4, 1.0};
42 #define wpos(wep) ( w[ WEP_GETWEIGHT(wep) ] )
44 #define RANK_NO_NORM 0x00
45 #define RANK_NORM_LOGLENGTH 0x01
46 #define RANK_NORM_LENGTH 0x02
47 #define RANK_NORM_EXTDIST 0x04
48 #define RANK_NORM_UNIQ 0x08
49 #define RANK_NORM_LOGUNIQ 0x10
50 #define DEF_NORM_METHOD RANK_NO_NORM
52 static float calc_rank_or(float *w, tsvector * t, QUERYTYPE * q);
53 static float calc_rank_and(float *w, tsvector * t, QUERYTYPE * q);
56 * Returns a weight of a word collocation
64 return 1.0 / (1.005 + 0.05 * exp(((float4) w) / 1.5 - 2));
68 cnt_length(tsvector * t)
70 WordEntry *ptr = ARRPTR(t),
71 *end = (WordEntry *) STRPTR(t);
77 if ((clen = POSDATALEN(t, ptr)) == 0)
88 WordECompareITEM(char *eval, char *qval, WordEntry * ptr, ITEM * item)
90 if (ptr->len == item->length)
93 qval + item->distance,
96 return (ptr->len > item->length) ? 1 : -1;
100 find_wordentry(tsvector * t, QUERYTYPE * q, ITEM * item)
102 WordEntry *StopLow = ARRPTR(t);
103 WordEntry *StopHigh = (WordEntry *) STRPTR(t);
104 WordEntry *StopMiddle;
107 /* Loop invariant: StopLow <= item < StopHigh */
109 while (StopLow < StopHigh)
111 StopMiddle = StopLow + (StopHigh - StopLow) / 2;
112 difference = WordECompareITEM(STRPTR(t), GETOPERAND(q), StopMiddle, item);
115 else if (difference < 0)
116 StopLow = StopMiddle + 1;
118 StopHigh = StopMiddle;
125 static char *SortAndUniqOperand = NULL;
128 compareITEM(const void *a, const void *b)
130 if ((*(ITEM **) a)->length == (*(ITEM **) b)->length)
131 return strncmp(SortAndUniqOperand + (*(ITEM **) a)->distance,
132 SortAndUniqOperand + (*(ITEM **) b)->distance,
133 (*(ITEM **) b)->length);
135 return ((*(ITEM **) a)->length > (*(ITEM **) b)->length) ? 1 : -1;
139 SortAndUniqItems(char *operand, ITEM * item, int *size)
145 ptr = res = (ITEM **) palloc(sizeof(ITEM *) * *size);
149 if (item->type == VAL)
161 SortAndUniqOperand = operand;
162 qsort(res, *size, sizeof(ITEM **), compareITEM);
167 while (ptr - res < *size)
169 if (compareITEM((void *) ptr, (void *) prevptr) != 0)
177 *size = prevptr + 1 - res;
181 static WordEntryPos POSNULL[] = {
187 calc_rank_and(float *w, tsvector * t, QUERYTYPE * q)
204 item = SortAndUniqItems(GETOPERAND(q), GETQUERY(q), &size);
208 return calc_rank_or(w, t, q);
210 pos = (uint16 **) palloc(sizeof(uint16 *) * q->size);
211 memset(pos, 0, sizeof(uint16 *) * q->size);
212 *(uint16 *) POSNULL = lengthof(POSNULL) - 1;
213 WEP_SETPOS(POSNULL[1], MAXENTRYPOS - 1);
215 for (i = 0; i < size; i++)
217 entry = find_wordentry(t, q, item[i]);
222 pos[i] = (uint16 *) _POSDATAPTR(t, entry);
224 pos[i] = (uint16 *) POSNULL;
227 dimt = *(uint16 *) (pos[i]);
228 post = (WordEntryPos *) (pos[i] + 1);
229 for (k = 0; k < i; k++)
233 lenct = *(uint16 *) (pos[k]);
234 ct = (WordEntryPos *) (pos[k] + 1);
235 for (l = 0; l < dimt; l++)
237 for (p = 0; p < lenct; p++)
239 dist = Abs((int) WEP_GETPOS(post[l]) - (int) WEP_GETPOS(ct[p]));
240 if (dist || (dist == 0 && (pos[i] == (uint16 *) POSNULL || pos[k] == (uint16 *) POSNULL)))
246 curw = sqrt(wpos(post[l]) * wpos(ct[p]) * word_distance(dist));
247 res = (res < 0) ? curw : 1.0 - (1.0 - res) * (1.0 - curw);
259 calc_rank_or(float *w, tsvector * t, QUERYTYPE * q)
270 *(uint16 *) POSNULL = lengthof(POSNULL) - 1;
271 item = SortAndUniqItems(GETOPERAND(q), GETQUERY(q), &size);
273 for (i = 0; i < size; i++)
279 entry = find_wordentry(t, q, item[i]);
285 dimt = POSDATALEN(t, entry);
286 post = POSDATAPTR(t, entry);
290 dimt = *(uint16 *) POSNULL;
297 for (j = 0; j < dimt; j++)
299 resj = resj + wpos(post[j]) / ((j + 1) * (j + 1));
300 if (wpos(post[j]) > wjm)
307 limit (sum(i/i^2),i->inf) = pi^2/6
308 resj = sum(wi/i^2),i=1,noccurence,
309 wi - should be sorted desc,
310 don't sort for now, just choose maximum weight. This should be corrected
313 res = res + (wjm + resj - wjm / ((jm + 1) * (jm + 1))) / 1.64493406685;
322 calc_rank(float *w, tsvector * t, QUERYTYPE * q, int4 method)
324 ITEM *item = GETQUERY(q);
328 if (!t->size || !q->size)
331 res = (item->type != VAL && item->val == (int4) '&') ?
332 calc_rank_and(w, t, q) : calc_rank_or(w, t, q);
337 if ( (method & RANK_NORM_LOGLENGTH) && t->size>0 )
338 res /= log((double) (cnt_length(t) + 1)) / log(2.0);
340 if ( method & RANK_NORM_LENGTH ) {
346 if ( (method & RANK_NORM_UNIQ) && t->size > 0 )
347 res /= (float)( t->size );
349 if ( (method & RANK_NORM_LOGUNIQ) && t->size > 0 )
350 res /= log((double) (t->size + 1)) / log(2.0);
356 rank(PG_FUNCTION_ARGS)
358 ArrayType *win = (ArrayType *) PG_DETOAST_DATUM(PG_GETARG_DATUM(0));
359 tsvector *txt = (tsvector *) PG_DETOAST_DATUM(PG_GETARG_DATUM(1));
360 QUERYTYPE *query = (QUERYTYPE *) PG_DETOAST_DATUM(PG_GETARG_DATUM(2));
361 int method = DEF_NORM_METHOD;
363 float ws[lengthof(weights)];
367 if (ARR_NDIM(win) != 1)
369 (errcode(ERRCODE_ARRAY_SUBSCRIPT_ERROR),
370 errmsg("array of weight must be one-dimensional")));
372 if (ARRNELEMS(win) < lengthof(weights))
374 (errcode(ERRCODE_ARRAY_SUBSCRIPT_ERROR),
375 errmsg("array of weight is too short")));
377 if (ARR_HASNULL(win))
379 (errcode(ERRCODE_NULL_VALUE_NOT_ALLOWED),
380 errmsg("array of weight must not contain nulls")));
382 arrdata = (float4 *) ARR_DATA_PTR(win);
383 for (i = 0; i < lengthof(weights); i++)
385 ws[i] = (arrdata[i] >= 0) ? arrdata[i] : weights[i];
388 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
389 errmsg("weight out of range")));
393 method = PG_GETARG_INT32(3);
395 res = calc_rank(ws, txt, query, method);
397 PG_FREE_IF_COPY(win, 0);
398 PG_FREE_IF_COPY(txt, 1);
399 PG_FREE_IF_COPY(query, 2);
400 PG_RETURN_FLOAT4(res);
404 rank_def(PG_FUNCTION_ARGS)
406 tsvector *txt = (tsvector *) PG_DETOAST_DATUM(PG_GETARG_DATUM(0));
407 QUERYTYPE *query = (QUERYTYPE *) PG_DETOAST_DATUM(PG_GETARG_DATUM(1));
409 int method = DEF_NORM_METHOD;
412 method = PG_GETARG_INT32(2);
414 res = calc_rank(weights, txt, query, method);
416 PG_FREE_IF_COPY(txt, 0);
417 PG_FREE_IF_COPY(query, 1);
418 PG_RETURN_FLOAT4(res);
432 compareDocR(const void *a, const void *b)
434 if (((DocRepresentation *) a)->pos == ((DocRepresentation *) b)->pos)
436 return (((DocRepresentation *) a)->pos > ((DocRepresentation *) b)->pos) ? 1 : -1;
440 checkcondition_ITEM(void *checkval, ITEM * val)
442 return (bool) (val->istrue);
446 reset_istrue_flag(QUERYTYPE * query)
448 ITEM *item = GETQUERY(query);
451 /* reset istrue flag */
452 for (i = 0; i < query->size; i++)
454 if (item->type == VAL)
464 DocRepresentation *begin;
465 DocRepresentation *end;
470 Cover(DocRepresentation * doc, int len, QUERYTYPE * query, Extention *ext)
472 DocRepresentation *ptr;
473 int lastpos = ext->pos;
477 reset_istrue_flag(query);
481 ptr = doc + ext->pos;
483 /* find upper bound of cover from current position, move up */
484 while (ptr - doc < len)
486 for (i = 0; i < ptr->nitem; i++)
487 ptr->item[i]->istrue = 1;
488 if (TS_execute(GETQUERY(query), NULL, false, checkcondition_ITEM))
490 if (ptr->pos > ext->q)
505 reset_istrue_flag(query);
509 /* find lower bound of cover from founded upper bound, move down */
512 for (i = 0; i < ptr->nitem; i++)
513 ptr->item[i]->istrue = 1;
514 if (TS_execute(GETQUERY(query), NULL, true, checkcondition_ITEM))
516 if (ptr->pos < ext->p) {
525 if (ext->p <= ext->q)
528 * set position for next try to next lexeme after begining of founded
531 ext->pos = (ptr - doc) + 1;
536 return Cover(doc, len, query, ext);
539 static DocRepresentation *
540 get_docrep(tsvector * txt, QUERYTYPE * query, int *doclen)
542 ITEM *item = GETQUERY(query);
548 int len = query->size * 4,
550 DocRepresentation *doc;
552 *(uint16 *) POSNULL = lengthof(POSNULL) - 1;
553 doc = (DocRepresentation *) palloc(sizeof(DocRepresentation) * len);
554 SortAndUniqOperand = GETOPERAND(query);
555 reset_istrue_flag(query);
557 for (i = 0; i < query->size; i++)
559 if (item[i].type != VAL || item[i].istrue)
562 entry = find_wordentry(txt, query, &(item[i]));
568 dimt = POSDATALEN(txt, entry);
569 post = POSDATAPTR(txt, entry);
573 dimt = *(uint16 *) POSNULL;
577 while (cur + dimt >= len)
580 doc = (DocRepresentation *) repalloc(doc, sizeof(DocRepresentation) * len);
583 for (j = 0; j < dimt; j++)
591 doc[cur].needfree = false;
593 doc[cur].item = (ITEM **) palloc(sizeof(ITEM *) * query->size);
595 for (k = 0; k < query->size; k++)
598 if (k == i || (item[k].type == VAL && compareITEM(&kptr, &iptr) == 0))
600 doc[cur].item[doc[cur].nitem] = item + k;
608 doc[cur].needfree = false;
609 doc[cur].nitem = doc[cur - 1].nitem;
610 doc[cur].item = doc[cur - 1].item;
612 doc[cur].pos = WEP_GETPOS(post[j]);
613 doc[cur].wclass = WEP_GETWEIGHT(post[j]);
623 qsort((void *) doc, cur, sizeof(DocRepresentation), compareDocR);
632 calc_rank_cd(float4 *arrdata, tsvector *txt, QUERYTYPE *query, int method) {
633 DocRepresentation *doc;
639 double invws[lengthof(weights)];
640 double SumDist=0.0, PrevExtPos=0.0, CurExtPos=0.0;
643 for (i = 0; i < lengthof(weights); i++)
645 invws[i] = ((double)((arrdata[i] >= 0) ? arrdata[i] : weights[i]));
648 (errcode(ERRCODE_INVALID_PARAMETER_VALUE),
649 errmsg("weight out of range")));
650 invws[i] = 1.0/invws[i];
653 doc = get_docrep(txt, query, &doclen);
657 MemSet( &ext, 0, sizeof(Extention) );
658 while (Cover(doc, doclen, query, &ext)) {
661 DocRepresentation *ptr = ext.begin;
663 while ( ptr<=ext.end ) {
664 InvSum += invws[ ptr->wclass ];
668 Cpos = ((double)( ext.end-ext.begin+1 )) / InvSum;
669 Wdoc += Cpos / ( (double)(( 1 + (ext.q - ext.p) - (ext.end - ext.begin) )) );
671 CurExtPos = ((double)(ext.q + ext.p))/2.0;
672 if ( NExtent>0 && CurExtPos > PrevExtPos /* prevent devision by zero in a case of multiple lexize */ )
673 SumDist += 1.0/( CurExtPos - PrevExtPos );
675 PrevExtPos = CurExtPos;
679 if ( (method & RANK_NORM_LOGLENGTH) && txt->size > 0 )
680 Wdoc /= log((double) (cnt_length(txt) + 1));
682 if ( method & RANK_NORM_LENGTH ) {
683 len = cnt_length(txt);
685 Wdoc /= (double) len;
688 if ( (method & RANK_NORM_EXTDIST) && SumDist > 0 )
689 Wdoc /= ((double)NExtent) / SumDist;
691 if ( (method & RANK_NORM_UNIQ) && txt->size > 0 )
692 Wdoc /= (double)( txt->size );
694 if ( (method & RANK_NORM_LOGUNIQ) && txt->size > 0 )
695 Wdoc /= log((double) (txt->size + 1)) / log(2.0);
697 for (i = 0; i < doclen; i++)
706 rank_cd(PG_FUNCTION_ARGS)
708 ArrayType *win = (ArrayType *) PG_DETOAST_DATUM(PG_GETARG_DATUM(0));
709 tsvector *txt = (tsvector *) PG_DETOAST_DATUM(PG_GETARG_DATUM(1));
710 QUERYTYPE *query = (QUERYTYPE *) PG_DETOAST_DATUM_COPY(PG_GETARG_DATUM(2));
711 int method = DEF_NORM_METHOD;
714 if (ARR_NDIM(win) != 1)
716 (errcode(ERRCODE_ARRAY_SUBSCRIPT_ERROR),
717 errmsg("array of weight must be one-dimensional")));
719 if (ARRNELEMS(win) < lengthof(weights))
721 (errcode(ERRCODE_ARRAY_SUBSCRIPT_ERROR),
722 errmsg("array of weight is too short")));
724 if (ARR_HASNULL(win))
726 (errcode(ERRCODE_NULL_VALUE_NOT_ALLOWED),
727 errmsg("array of weight must not contain nulls")));
730 method = PG_GETARG_INT32(3);
732 res = calc_rank_cd( (float4 *) ARR_DATA_PTR(win), txt, query, method);
734 PG_FREE_IF_COPY(win, 0);
735 PG_FREE_IF_COPY(txt, 1);
736 PG_FREE_IF_COPY(query, 2);
738 PG_RETURN_FLOAT4(res);
743 rank_cd_def(PG_FUNCTION_ARGS)
745 tsvector *txt = (tsvector *) PG_DETOAST_DATUM(PG_GETARG_DATUM(0));
746 QUERYTYPE *query = (QUERYTYPE *) PG_DETOAST_DATUM_COPY(PG_GETARG_DATUM(1));
749 res = calc_rank_cd( weights, txt, query, (PG_NARGS() == 3) ? PG_GETARG_DATUM(2) : DEF_NORM_METHOD);
751 PG_FREE_IF_COPY(txt, 0);
752 PG_FREE_IF_COPY(query, 1);
754 PG_RETURN_FLOAT4(res);
757 /**************debug*************/
769 compareDocWord(const void *a, const void *b)
771 if (((DocWord *) a)->pos == ((DocWord *) b)->pos)
773 return (((DocWord *) a)->pos > ((DocWord *) b)->pos) ? 1 : -1;
778 get_covers(PG_FUNCTION_ARGS)
780 tsvector *txt = (tsvector *) PG_DETOAST_DATUM(PG_GETARG_DATUM(0));
781 QUERYTYPE *query = (QUERYTYPE *) PG_DETOAST_DATUM_COPY(PG_GETARG_DATUM(1));
782 WordEntry *pptr = ARRPTR(txt);
793 DocRepresentation *doc;
798 doc = get_docrep(txt, query, &rlen);
802 out = palloc(VARHDRSZ);
803 VARATT_SIZEP(out) = VARHDRSZ;
804 PG_FREE_IF_COPY(txt, 0);
805 PG_FREE_IF_COPY(query, 1);
806 PG_RETURN_POINTER(out);
809 for (i = 0; i < txt->size; i++)
813 (errcode(ERRCODE_SYNTAX_ERROR),
814 errmsg("no pos info")));
815 dlen += POSDATALEN(txt, &(pptr[i]));
818 dwptr = dw = palloc(sizeof(DocWord) * dlen);
819 memset(dw, 0, sizeof(DocWord) * dlen);
821 for (i = 0; i < txt->size; i++)
823 WordEntryPos *posdata = POSDATAPTR(txt, &(pptr[i]));
825 for (j = 0; j < POSDATALEN(txt, &(pptr[i])); j++)
827 dw[cur].w = STRPTR(txt) + pptr[i].pos;
828 dw[cur].len = pptr[i].len;
829 dw[cur].pos = WEP_GETPOS(posdata[j]);
832 len += (pptr[i].len + 1) * (int) POSDATALEN(txt, &(pptr[i]));
834 qsort((void *) dw, dlen, sizeof(DocWord), compareDocWord);
836 MemSet( &ext, 0, sizeof(Extention) );
837 while (Cover(doc, rlen, query, &ext))
839 dwptr = dw + olddwpos;
840 while (dwptr->pos < ext.p && dwptr - dw < dlen)
842 olddwpos = dwptr - dw;
843 dwptr->start = ncover;
844 while (dwptr->pos < ext.q + 1 && dwptr - dw < dlen)
846 (dwptr - 1)->finish = ncover;
847 len += 4 /* {}+two spaces */ + 2 * 16 /* numbers */ ;
851 out = palloc(VARHDRSZ + len);
852 cptr = ((char *) out) + VARHDRSZ;
855 while (dwptr - dw < dlen)
859 sprintf(cptr, "{%d ", dwptr->start);
860 cptr = strchr(cptr, '\0');
862 memcpy(cptr, dwptr->w, dwptr->len);
868 sprintf(cptr, "}%d ", dwptr->finish);
869 cptr = strchr(cptr, '\0');
874 VARATT_SIZEP(out) = cptr - ((char *) out);
877 for (i = 0; i < rlen; i++)
882 PG_FREE_IF_COPY(txt, 0);
883 PG_FREE_IF_COPY(query, 1);
884 PG_RETURN_POINTER(out);