liuyu/pgvector
1
0
Fork 0
mirror of https://github.com/pgvector/pgvector.git synced 2026-06-06 14:01:31 +08:00
Code Issues Packages Projects Releases Wiki Activity

Changed HnswPairingHeapNode to HnswSearchCandidate to reduce allocations and improve code

Browse Source
This commit is contained in:
Andrew Kane
2024-09-21 12:07:44 -07:00
parent 6d2af6d3f9
commit d5e8fc96a5
3 changed files with 41 additions and 46 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/hnsw.h
View File

@@ -155,12 +155,13 @@ struct HnswNeighborArray
HnswCandidate items[FLEXIBLE_ARRAY_MEMBER];
};
typedef struct HnswPairingHeapNode
typedef struct HnswSearchCandidate
{
HnswCandidate *inner;
pairingheap_node c_node;
pairingheap_node w_node;
} HnswPairingHeapNode;
HnswElementPtr element;
float distance;
} HnswSearchCandidate;
/* HNSW index options */
typedef struct HnswOptions
@@ -381,7 +382,7 @@ void *HnswAlloc(HnswAllocator * allocator, Size size);
HnswElement HnswInitElement(char *base, ItemPointer tid, int m, double ml, int maxLevel, HnswAllocator * alloc);
HnswElement HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno);
void HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, bool existing);
HnswCandidate *HnswEntryCandidate(char *base, HnswElement em, Datum q, Relation rel, FmgrInfo *procinfo, Oid collation, bool loadVec);
HnswSearchCandidate *HnswEntryCandidate(char *base, HnswElement em, Datum q, Relation rel, FmgrInfo *procinfo, Oid collation, bool loadVec);
void HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum, bool building);
void HnswSetNeighborTuple(char *base, HnswNeighborTuple ntup, HnswElement e, int m);
void HnswAddHeapTid(HnswElement element, ItemPointer heaptid);

4
src/hnswscan.c
View File

@@ -161,14 +161,14 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
so->first = false;
#if defined(HNSW_MEMORY)
elog(INFO, "memory: %zu MB", MemoryContextMemAllocated(so->tmpCtx, false) / (1024 * 1024));
elog(INFO, "memory: %zu KB", MemoryContextMemAllocated(so->tmpCtx, false) / 1024);
#endif
}
while (list_length(so->w) > 0)
{
char *base = NULL;
HnswCandidate *hc = llast(so->w);
HnswSearchCandidate *hc = llast(so->w);
HnswElement element = HnswPtrAccess(base, hc->element);
ItemPointer heaptid;

74
src/hnswutils.c
View File

@@ -608,10 +608,10 @@ GetElementDistance(char *base, HnswElement element, Datum q, FmgrInfo *procinfo,
/*
* Create a candidate for the entry point
*/
HnswCandidate *
HnswSearchCandidate *
HnswEntryCandidate(char *base, HnswElement entryPoint, Datum q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec)
{
HnswCandidate *hc = palloc(sizeof(HnswCandidate));
HnswSearchCandidate *hc = palloc(sizeof(HnswSearchCandidate));
HnswPtrStore(base, hc->element, entryPoint);
if (index == NULL)
@@ -621,8 +621,8 @@ HnswEntryCandidate(char *base, HnswElement entryPoint, Datum q, Relation index,
return hc;
}
#define HnswGetPairingHeapCandidate(membername, ptr) (pairingheap_container(HnswPairingHeapNode, membername, ptr)->inner)
#define HnswGetPairingHeapCandidateConst(membername, ptr) (pairingheap_const_container(HnswPairingHeapNode, membername, ptr)->inner)
#define HnswGetSearchCandidate(membername, ptr) pairingheap_container(HnswSearchCandidate, membername, ptr)
#define HnswGetSearchCandidateConst(membername, ptr) pairingheap_const_container(HnswSearchCandidate, membername, ptr)
/*
* Compare candidate distances
@@ -630,10 +630,10 @@ HnswEntryCandidate(char *base, HnswElement entryPoint, Datum q, Relation index,
static int
CompareNearestCandidates(const pairingheap_node *a, const pairingheap_node *b, void *arg)
{
if (HnswGetPairingHeapCandidateConst(c_node, a)->distance < HnswGetPairingHeapCandidateConst(c_node, b)->distance)
if (HnswGetSearchCandidateConst(c_node, a)->distance < HnswGetSearchCandidateConst(c_node, b)->distance)
return 1;
if (HnswGetPairingHeapCandidateConst(c_node, a)->distance > HnswGetPairingHeapCandidateConst(c_node, b)->distance)
if (HnswGetSearchCandidateConst(c_node, a)->distance > HnswGetSearchCandidateConst(c_node, b)->distance)
return -1;
return 0;
@@ -645,27 +645,15 @@ CompareNearestCandidates(const pairingheap_node *a, const pairingheap_node *b, v
static int
CompareFurthestCandidates(const pairingheap_node *a, const pairingheap_node *b, void *arg)
{
if (HnswGetPairingHeapCandidateConst(w_node, a)->distance < HnswGetPairingHeapCandidateConst(w_node, b)->distance)
if (HnswGetSearchCandidateConst(w_node, a)->distance < HnswGetSearchCandidateConst(w_node, b)->distance)
return -1;
if (HnswGetPairingHeapCandidateConst(w_node, a)->distance > HnswGetPairingHeapCandidateConst(w_node, b)->distance)
if (HnswGetSearchCandidateConst(w_node, a)->distance > HnswGetSearchCandidateConst(w_node, b)->distance)
return 1;
return 0;
}
/*
* Create a pairing heap node for a candidate
*/
static HnswPairingHeapNode *
CreatePairingHeapNode(HnswCandidate * c)
{
HnswPairingHeapNode *node = palloc(sizeof(HnswPairingHeapNode));
node->inner = c;
return node;
}
/*
* Init visited
*/
@@ -825,15 +813,13 @@ HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, F
/* Add entry points to v, C, and W */
foreach(lc2, ep)
{
HnswCandidate *hc = (HnswCandidate *) lfirst(lc2);
HnswSearchCandidate *hc = (HnswSearchCandidate *) lfirst(lc2);
bool found;
HnswPairingHeapNode *node;
AddToVisited(base, &v, hc->element, index, &found);
node = CreatePairingHeapNode(hc);
pairingheap_add(C, &node->c_node);
pairingheap_add(W, &node->w_node);
pairingheap_add(C, &hc->c_node);
pairingheap_add(W, &hc->w_node);
/*
* Do not count elements being deleted towards ef when vacuuming. It
@@ -846,8 +832,8 @@ HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, F
while (!pairingheap_is_empty(C))
{
HnswCandidate *c = HnswGetPairingHeapCandidate(c_node, pairingheap_remove_first(C));
HnswCandidate *f = HnswGetPairingHeapCandidate(w_node, pairingheap_first(W));
HnswSearchCandidate *c = HnswGetSearchCandidate(c_node, pairingheap_remove_first(C));
HnswSearchCandidate *f = HnswGetSearchCandidate(w_node, pairingheap_first(W));
HnswElement cElement;
if (c->distance > f->distance)
@@ -863,12 +849,11 @@ HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, F
for (int i = 0; i < unvisitedLength; i++)
{
HnswElement eElement;
HnswCandidate *e;
HnswPairingHeapNode *node;
HnswSearchCandidate *e;
float eDistance;
bool alwaysAdd = wlen < ef;
f = HnswGetPairingHeapCandidate(w_node, pairingheap_first(W));
f = HnswGetSearchCandidate(w_node, pairingheap_first(W));
if (index == NULL)
{
@@ -899,13 +884,11 @@ HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, F
continue;
/* Create a new candidate */
e = palloc(sizeof(HnswCandidate));
e = palloc(sizeof(HnswSearchCandidate));
HnswPtrStore(base, e->element, eElement);
e->distance = eDistance;
node = CreatePairingHeapNode(e);
pairingheap_add(C, &node->c_node);
pairingheap_add(W, &node->w_node);
pairingheap_add(C, &e->c_node);
pairingheap_add(W, &e->w_node);
/*
* Do not count elements being deleted towards ef when vacuuming.
@@ -926,7 +909,7 @@ HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, F
/* Add each element of W to w */
while (!pairingheap_is_empty(W))
{
HnswCandidate *hc = HnswGetPairingHeapCandidate(w_node, pairingheap_remove_first(W));
HnswSearchCandidate *hc = HnswGetSearchCandidate(w_node, pairingheap_remove_first(W));
w = lappend(w, hc);
}
@@ -1307,16 +1290,27 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
{
int lm = HnswGetLayerM(m, lc);
List *neighbors;
List *lw;
List *lw = NIL;
ListCell *lc2;
w = HnswSearchLayer(base, q, ep, efConstruction, lc, index, procinfo, collation, m, true, skipElement);
/* Convert search candidates to candidates */
foreach(lc2, w)
{
HnswSearchCandidate *sc = lfirst(lc2);
HnswCandidate *hc = palloc(sizeof(HnswCandidate));
hc->element = sc->element;
hc->distance = sc->distance;
lw = lappend(lw, hc);
}
/* Elements being deleted or skipped can help with search */
/* but should be removed before selecting neighbors */
if (index != NULL)
lw = RemoveElements(base, w, skipElement);
else
lw = w;
lw = RemoveElements(base, lw, skipElement);
/*
* Candidates are sorted, but not deterministically. Could set