/*
* The GiST PickSplit method
*
* New linear algorithm, see 'New Linear node_n Splitting Algorithm for R-tree',
* C.H.Ang and T.C.Tan
*
* This is used for both boxes and points.
*/
datum_t gist_box_picksplit(PG_FUNC_ARGS)
{
struct gist_entry_vector *entryvec = (struct gist_entry_vector *)ARG_POINTER(0);
struct gist_splitvec *v = (struct gist_splitvec *)ARG_POINTER(1);
item_id_t i;
item_id_t *listL;
item_id_t *listR;
item_id_t *listB;
item_id_t *listT;
BOX *unionL;
BOX *unionR;
BOX *unionB;
BOX *unionT;
int posL, posR, posB, posT;
BOX pageunion;
BOX *cur;
char direction = ' ';
bool allisequal = true;
item_id_t maxoff;
int nbytes;
posL = posR = posB = posT = 0;
maxoff = entryvec->n - 1;
cur = D_TO_BOX_P(entryvec->vector[FIRST_ITEM_ID].key);
memcpy((void*) &pageunion, (void*) cur, sizeof(BOX));
/* find MBR */
for (i = ITEM_ID_NEXT(FIRST_ITEM_ID); i <= maxoff; i = ITEM_ID_NEXT(i)) {
cur = D_TO_BOX_P(entryvec->vector[i].key);
if (allisequal
&& (pageunion.high.x != cur->high.x
|| pageunion.high.y != cur->high.y
|| pageunion.low.x != cur->low.x
|| pageunion.low.y != cur->low.y))
allisequal = false;
adjustBox(&pageunion, cur);
}
if (allisequal) {
/*
* All entries are the same
*/
fallbackSplit(entryvec, v);
RET_POINTER(v);
}
nbytes = (maxoff + 2) * sizeof(item_id_t);
listL = (item_id_t *) palloc(nbytes);
listR = (item_id_t *) palloc(nbytes);
listB = (item_id_t *) palloc(nbytes);
listT = (item_id_t *) palloc(nbytes);
unionL = (BOX *) palloc(sizeof(BOX));
unionR = (BOX *) palloc(sizeof(BOX));
unionB = (BOX *) palloc(sizeof(BOX));
unionT = (BOX *) palloc(sizeof(BOX));
#define ADDLIST( list, unionD, pos, num ) do { \
if ( pos ) { \
if ( (unionD)->high.x < cur->high.x ) (unionD)->high.x = cur->high.x; \
if ( (unionD)->low.x > cur->low.x ) (unionD)->low.x = cur->low.x; \
if ( (unionD)->high.y < cur->high.y ) (unionD)->high.y = cur->high.y; \
if ( (unionD)->low.y > cur->low.y ) (unionD)->low.y = cur->low.y; \
} else { \
memcpy( (void*)(unionD), (void*) cur, sizeof( BOX ) ); \
} \
\
(list)[pos] = num; \
(pos)++; \
} while(0)
for (i = FIRST_ITEM_ID; i <= maxoff; i = ITEM_ID_NEXT(i)) {
cur = D_TO_BOX_P(entryvec->vector[i].key);
if (cur->low.x - pageunion.low.x < pageunion.high.x - cur->high.x)
ADDLIST(listL, unionL, posL, i);
else
ADDLIST(listR, unionR, posR, i);
if (cur->low.y - pageunion.low.y < pageunion.high.y - cur->high.y)
ADDLIST(listB, unionB, posB, i);
else
ADDLIST(listT, unionT, posT, i);
}
#define LIMIT_RATIO 0.1
#define _IS_BADRATIO(x,y) ( (y) == 0 || (float)(x)/(float)(y) < LIMIT_RATIO )
#define IS_BADRATIO(x,y) ( _IS_BADRATIO((x),(y)) || _IS_BADRATIO((y),(x)) )
/* bad disposition, try to split by centers of boxes */
if (IS_BADRATIO(posR, posL) && IS_BADRATIO(posT, posB)) {
double avgCenterX = 0.0;
double avgCenterY = 0.0;
double CenterX;
double CenterY;
for (i = FIRST_ITEM_ID; i <= maxoff; i = ITEM_ID_NEXT(i)) {
cur = D_TO_BOX_P(entryvec->vector[i].key);
avgCenterX += ((double)cur->high.x + (double)cur->low.x) / 2.0;
avgCenterY += ((double)cur->high.y + (double)cur->low.y) / 2.0;
}
avgCenterX /= maxoff;
avgCenterY /= maxoff;
posL = posR = posB = posT = 0;
for (i = FIRST_ITEM_ID; i <= maxoff; i = ITEM_ID_NEXT(i)) {
cur = D_TO_BOX_P(entryvec->vector[i].key);
CenterX = ((double)cur->high.x + (double)cur->low.x) / 2.0;
CenterY = ((double)cur->high.y + (double)cur->low.y) / 2.0;
if (CenterX < avgCenterX)
ADDLIST(listL, unionL, posL, i);
else if (CenterX == avgCenterX) {
if (posL > posR)
ADDLIST(listR, unionR, posR, i);
else
ADDLIST(listL, unionL, posL, i);
} else
ADDLIST(listR, unionR, posR, i);
if (CenterY < avgCenterY)
ADDLIST(listB, unionB, posB, i);
else if (CenterY == avgCenterY) {
if (posB > posT)
ADDLIST(listT, unionT, posT, i);
else
ADDLIST(listB, unionB, posB, i);
} else
ADDLIST(listT, unionT, posT, i);
}
if (IS_BADRATIO(posR, posL) && IS_BADRATIO(posT, posB)) {
fallbackSplit(entryvec, v);
RET_POINTER(v);
}
}
/* which split more optimal? */
if (Max(posL, posR) < Max(posB, posT))
direction = 'x';
else if (Max(posL, posR) > Max(posB, posT))
direction = 'y';
else {
datum_t interLR = DIRECT_FC2(rt_box_inter, BOX_P_TO_D(unionL), BOX_P_TO_D(unionR));
datum_t interBT = DIRECT_FC2(rt_box_inter, BOX_P_TO_D(unionB), BOX_P_TO_D(unionT));
double sizeLR;
double sizeBT;
sizeLR = size_box(interLR);
sizeBT = size_box(interBT);
if (sizeLR < sizeBT)
direction = 'x';
else
direction = 'y';
}
if (direction == 'x')
chooseLR(v, listL, posL, unionL, listR, posR, unionR);
else
chooseLR(v, listB, posB, unionB, listT, posT, unionT);
RET_POINTER(v);
}
/*
* The GiST PickSplit method
*
* New linear algorithm, see 'New Linear Node Splitting Algorithm for R-tree',
* C.H.Ang and T.C.Tan
*
* This is used for both boxes and points.
*/
Datum
gist_box_picksplit(PG_FUNCTION_ARGS)
{
GistEntryVector *entryvec = (GistEntryVector *) PG_GETARG_POINTER(0);
GIST_SPLITVEC *v = (GIST_SPLITVEC *) PG_GETARG_POINTER(1);
OffsetNumber i;
OffsetNumber *listL,
*listR,
*listB,
*listT;
BOX *unionL,
*unionR,
*unionB,
*unionT;
int posL,
posR,
posB,
posT;
BOX pageunion;
BOX *cur;
char direction = ' ';
bool allisequal = true;
OffsetNumber maxoff;
int nbytes;
posL = posR = posB = posT = 0;
maxoff = entryvec->n - 1;
cur = DatumGetBoxP(entryvec->vector[FirstOffsetNumber].key);
memcpy((void *) &pageunion, (void *) cur, sizeof(BOX));
/* find MBR */
for (i = OffsetNumberNext(FirstOffsetNumber); i <= maxoff; i = OffsetNumberNext(i))
{
cur = DatumGetBoxP(entryvec->vector[i].key);
if (allisequal && (
pageunion.high.x != cur->high.x ||
pageunion.high.y != cur->high.y ||
pageunion.low.x != cur->low.x ||
pageunion.low.y != cur->low.y
))
allisequal = false;
adjustBox(&pageunion, cur);
}
if (allisequal)
{
/*
* All entries are the same
*/
fallbackSplit(entryvec, v);
PG_RETURN_POINTER(v);
}
nbytes = (maxoff + 2) * sizeof(OffsetNumber);
listL = (OffsetNumber *) palloc(nbytes);
listR = (OffsetNumber *) palloc(nbytes);
listB = (OffsetNumber *) palloc(nbytes);
listT = (OffsetNumber *) palloc(nbytes);
unionL = (BOX *) palloc(sizeof(BOX));
unionR = (BOX *) palloc(sizeof(BOX));
unionB = (BOX *) palloc(sizeof(BOX));
unionT = (BOX *) palloc(sizeof(BOX));
#define ADDLIST( list, unionD, pos, num ) do { \
if ( pos ) { \
if ( (unionD)->high.x < cur->high.x ) (unionD)->high.x = cur->high.x; \
if ( (unionD)->low.x > cur->low.x ) (unionD)->low.x = cur->low.x; \
if ( (unionD)->high.y < cur->high.y ) (unionD)->high.y = cur->high.y; \
if ( (unionD)->low.y > cur->low.y ) (unionD)->low.y = cur->low.y; \
} else { \
memcpy( (void*)(unionD), (void*) cur, sizeof( BOX ) ); \
} \
(list)[pos] = num; \
(pos)++; \
} while(0)
for (i = FirstOffsetNumber; i <= maxoff; i = OffsetNumberNext(i))
{
cur = DatumGetBoxP(entryvec->vector[i].key);
if (cur->low.x - pageunion.low.x < pageunion.high.x - cur->high.x)
ADDLIST(listL, unionL, posL, i);
else
ADDLIST(listR, unionR, posR, i);
if (cur->low.y - pageunion.low.y < pageunion.high.y - cur->high.y)
ADDLIST(listB, unionB, posB, i);
else
ADDLIST(listT, unionT, posT, i);
}
#define LIMIT_RATIO 0.1
#define _IS_BADRATIO(x,y) ( (y) == 0 || (float)(x)/(float)(y) < LIMIT_RATIO )
#define IS_BADRATIO(x,y) ( _IS_BADRATIO((x),(y)) || _IS_BADRATIO((y),(x)) )
/* bad disposition, try to split by centers of boxes */
if (IS_BADRATIO(posR, posL) && IS_BADRATIO(posT, posB))
{
double avgCenterX = 0.0,
avgCenterY = 0.0;
double CenterX,
CenterY;
for (i = FirstOffsetNumber; i <= maxoff; i = OffsetNumberNext(i))
{
cur = DatumGetBoxP(entryvec->vector[i].key);
avgCenterX += ((double) cur->high.x + (double) cur->low.x) / 2.0;
avgCenterY += ((double) cur->high.y + (double) cur->low.y) / 2.0;
}
avgCenterX /= maxoff;
avgCenterY /= maxoff;
posL = posR = posB = posT = 0;
for (i = FirstOffsetNumber; i <= maxoff; i = OffsetNumberNext(i))
{
cur = DatumGetBoxP(entryvec->vector[i].key);
CenterX = ((double) cur->high.x + (double) cur->low.x) / 2.0;
CenterY = ((double) cur->high.y + (double) cur->low.y) / 2.0;
if (CenterX < avgCenterX)
ADDLIST(listL, unionL, posL, i);
else if (CenterX == avgCenterX)
{
if (posL > posR)
ADDLIST(listR, unionR, posR, i);
else
ADDLIST(listL, unionL, posL, i);
}
else
ADDLIST(listR, unionR, posR, i);
if (CenterY < avgCenterY)
ADDLIST(listB, unionB, posB, i);
else if (CenterY == avgCenterY)
{
if (posB > posT)
ADDLIST(listT, unionT, posT, i);
else
ADDLIST(listB, unionB, posB, i);
}
else
ADDLIST(listT, unionT, posT, i);
}
if (IS_BADRATIO(posR, posL) && IS_BADRATIO(posT, posB))
{
fallbackSplit(entryvec, v);
PG_RETURN_POINTER(v);
}
}
/* which split more optimal? */
if (Max(posL, posR) < Max(posB, posT))
direction = 'x';
else if (Max(posL, posR) > Max(posB, posT))
direction = 'y';
else
{
Datum interLR = DirectFunctionCall2(rt_box_inter,
BoxPGetDatum(unionL),
BoxPGetDatum(unionR));
Datum interBT = DirectFunctionCall2(rt_box_inter,
BoxPGetDatum(unionB),
BoxPGetDatum(unionT));
double sizeLR,
sizeBT;
sizeLR = size_box(interLR);
sizeBT = size_box(interBT);
if (sizeLR < sizeBT)
direction = 'x';
else
direction = 'y';
}
if (direction == 'x')
chooseLR(v,
listL, posL, unionL,
listR, posR, unionR);
else
chooseLR(v,
listB, posB, unionB,
listT, posT, unionT);
PG_RETURN_POINTER(v);
}
/*
* The GiST PickSplit method
*
* New linear algorithm, see 'New Linear Node Splitting Algorithm for R-tree',
* C.H.Ang and T.C.Tan
*/
Datum
gist_box_picksplit(PG_FUNCTION_ARGS)
{
GistEntryVector *entryvec = (GistEntryVector *) PG_GETARG_POINTER(0);
GIST_SPLITVEC *v = (GIST_SPLITVEC *) PG_GETARG_POINTER(1);
OffsetNumber i;
OffsetNumber *listL,
*listR,
*listB,
*listT;
BOX *unionL,
*unionR,
*unionB,
*unionT;
int posL,
posR,
posB,
posT;
BOX pageunion;
BOX *cur;
char direction = ' ';
bool allisequal = true;
OffsetNumber maxoff;
int nbytes;
posL = posR = posB = posT = 0;
maxoff = entryvec->n - 1;
cur = DatumGetBoxP(entryvec->vector[FirstOffsetNumber].key);
memcpy((void *) &pageunion, (void *) cur, sizeof(BOX));
/* find MBR */
for (i = OffsetNumberNext(FirstOffsetNumber); i <= maxoff; i = OffsetNumberNext(i))
{
cur = DatumGetBoxP(entryvec->vector[i].key);
if (allisequal == true && (
pageunion.high.x != cur->high.x ||
pageunion.high.y != cur->high.y ||
pageunion.low.x != cur->low.x ||
pageunion.low.y != cur->low.y
))
allisequal = false;
adjustBox(&pageunion, cur);
}
nbytes = (maxoff + 2) * sizeof(OffsetNumber);
listL = (OffsetNumber *) palloc(nbytes);
listR = (OffsetNumber *) palloc(nbytes);
unionL = (BOX *) palloc(sizeof(BOX));
unionR = (BOX *) palloc(sizeof(BOX));
if (allisequal)
{
cur = DatumGetBoxP(entryvec->vector[OffsetNumberNext(FirstOffsetNumber)].key);
if (memcmp((void *) cur, (void *) &pageunion, sizeof(BOX)) == 0)
{
v->spl_left = listL;
v->spl_right = listR;
v->spl_nleft = v->spl_nright = 0;
memcpy((void *) unionL, (void *) &pageunion, sizeof(BOX));
memcpy((void *) unionR, (void *) &pageunion, sizeof(BOX));
for (i = FirstOffsetNumber; i <= maxoff; i = OffsetNumberNext(i))
{
if (i <= (maxoff - FirstOffsetNumber + 1) / 2)
{
v->spl_left[v->spl_nleft] = i;
v->spl_nleft++;
}
else
{
v->spl_right[v->spl_nright] = i;
v->spl_nright++;
}
}
if (v->spl_ldatum_exists)
adjustBox(unionL, DatumGetBoxP(v->spl_ldatum));
v->spl_ldatum = BoxPGetDatum(unionL);
if (v->spl_rdatum_exists)
adjustBox(unionR, DatumGetBoxP(v->spl_rdatum));
v->spl_rdatum = BoxPGetDatum(unionR);
v->spl_ldatum_exists = v->spl_rdatum_exists = false;
PG_RETURN_POINTER(v);
}
}
listB = (OffsetNumber *) palloc(nbytes);
listT = (OffsetNumber *) palloc(nbytes);
unionB = (BOX *) palloc(sizeof(BOX));
unionT = (BOX *) palloc(sizeof(BOX));
#define ADDLIST( list, unionD, pos, num ) do { \
if ( pos ) { \
if ( (unionD)->high.x < cur->high.x ) (unionD)->high.x = cur->high.x; \
if ( (unionD)->low.x > cur->low.x ) (unionD)->low.x = cur->low.x; \
if ( (unionD)->high.y < cur->high.y ) (unionD)->high.y = cur->high.y; \
if ( (unionD)->low.y > cur->low.y ) (unionD)->low.y = cur->low.y; \
} else { \
memcpy( (void*)(unionD), (void*) cur, sizeof( BOX ) ); \
} \
(list)[pos] = num; \
(pos)++; \
} while(0)
for (i = FirstOffsetNumber; i <= maxoff; i = OffsetNumberNext(i))
{
cur = DatumGetBoxP(entryvec->vector[i].key);
if (cur->low.x - pageunion.low.x < pageunion.high.x - cur->high.x)
ADDLIST(listL, unionL, posL, i);
else
ADDLIST(listR, unionR, posR, i);
if (cur->low.y - pageunion.low.y < pageunion.high.y - cur->high.y)
ADDLIST(listB, unionB, posB, i);
else
ADDLIST(listT, unionT, posT, i);
}
/* bad disposition, sort by ascending and resplit */
if ((posR == 0 || posL == 0) && (posT == 0 || posB == 0))
{
KBsort *arr = (KBsort *) palloc(sizeof(KBsort) * maxoff);
posL = posR = posB = posT = 0;
for (i = FirstOffsetNumber; i <= maxoff; i = OffsetNumberNext(i))
{
arr[i - 1].key = DatumGetBoxP(entryvec->vector[i].key);
arr[i - 1].pos = i;
}
qsort(arr, maxoff, sizeof(KBsort), compare_KB);
for (i = FirstOffsetNumber; i <= maxoff; i = OffsetNumberNext(i))
{
cur = arr[i - 1].key;
if (cur->low.x - pageunion.low.x < pageunion.high.x - cur->high.x)
ADDLIST(listL, unionL, posL, arr[i - 1].pos);
else if (cur->low.x - pageunion.low.x == pageunion.high.x - cur->high.x)
{
if (posL > posR)
ADDLIST(listR, unionR, posR, arr[i - 1].pos);
else
ADDLIST(listL, unionL, posL, arr[i - 1].pos);
}
else
ADDLIST(listR, unionR, posR, arr[i - 1].pos);
if (cur->low.y - pageunion.low.y < pageunion.high.y - cur->high.y)
ADDLIST(listB, unionB, posB, arr[i - 1].pos);
else if (cur->low.y - pageunion.low.y == pageunion.high.y - cur->high.y)
{
if (posB > posT)
ADDLIST(listT, unionT, posT, arr[i - 1].pos);
else
ADDLIST(listB, unionB, posB, arr[i - 1].pos);
}
else
ADDLIST(listT, unionT, posT, arr[i - 1].pos);
}
}
/* which split more optimal? */
if (Max(posL, posR) < Max(posB, posT))
direction = 'x';
else if (Max(posL, posR) > Max(posB, posT))
direction = 'y';
else
{
Datum interLR = DirectFunctionCall2(rt_box_inter,
BoxPGetDatum(unionL),
BoxPGetDatum(unionR));
Datum interBT = DirectFunctionCall2(rt_box_inter,
BoxPGetDatum(unionB),
BoxPGetDatum(unionT));
double sizeLR,
sizeBT;
sizeLR = size_box(interLR);
sizeBT = size_box(interBT);
if (sizeLR < sizeBT)
direction = 'x';
else
direction = 'y';
}
if (direction == 'x')
chooseLR(v,
listL, posL, unionL,
listR, posR, unionR);
else
chooseLR(v,
listB, posB, unionB,
listT, posT, unionT);
PG_RETURN_POINTER(v);
}
