Removed loaded

This takes the approach from commit a713e2acaa further. Once we have
remove a candidate from the "closer" set, we still don't need to
recalculate everything that follows. Any candidates that were in the
closer set before still only need to be compared with any new
candidates that we have added.

CHANGELOG.md

@@ -1,7 +1,7 @@
-## 0.5.1 (unreleased)
+## 0.5.1 (2023-10-10)
 
-- Added check for MVCC-compliant snapshot for HNSW index scans
-- Improved performance of index scans for IVFFlat after updates and deletes
+- Improved performance of HNSW index builds
+- Added check for MVCC-compliant snapshot for index scans
 
 ## 0.5.0 (2023-08-28)
 

META.json

@@ -2,7 +2,7 @@
 	"name": "vector",
 	"abstract": "Open-source vector similarity search for Postgres",
 	"description": "Supports L2 distance, inner product, and cosine distance",
-	"version": "0.5.0",
+	"version": "0.5.1",
 	"maintainer": [
 		"Andrew Kane <andrew@ankane.org>"
 	],
@@ -20,7 +20,7 @@
 		"vector": {
 			"file": "sql/vector.sql",
 			"docfile": "README.md",
-			"version": "0.5.0",
+			"version": "0.5.1",
 			"abstract": "Open-source vector similarity search for Postgres"
 		}
 	},

Makefile

@@ -1,10 +1,10 @@
 EXTENSION = vector
-EXTVERSION = 0.5.0
+EXTVERSION = 0.5.1
 
 MODULE_big = vector
 DATA = $(wildcard sql/*--*.sql)
-OBJS = src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/tinyint.o src/vector.o
-HEADERS = src/tinyint.h src/vector.h
+OBJS = src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/vector.o
+HEADERS = src/vector.h
 
 TESTS = $(wildcard test/sql/*.sql)
 REGRESS = $(patsubst test/sql/%.sql,%,$(TESTS))

Makefile.win

@@ -1,8 +1,8 @@
 EXTENSION = vector
-EXTVERSION = 0.5.0
+EXTVERSION = 0.5.1
 
-OBJS = src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\tinyint.obj src\vector.obj
-HEADERS = src\tinyint.h src\vector.h
+OBJS = src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\vector.obj
+HEADERS = src\vector.h
 
 REGRESS = btree cast copy functions input ivfflat_cosine ivfflat_ip ivfflat_l2 ivfflat_options ivfflat_unlogged
 REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)
@@ -56,7 +56,7 @@ install:
 	copy $(EXTENSION).control "$(SHAREDIR)\extension"
 	copy sql\$(EXTENSION)--*.sql "$(SHAREDIR)\extension"
 	mkdir "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
-	for %f in ($(HEADERS)) do copy %f "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
+	copy $(HEADERS) "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
 
 installcheck:
 	"$(BINDIR)\pg_regress" --bindir="$(BINDIR)" $(REGRESS_OPTS) $(REGRESS)

README.md

@@ -18,7 +18,7 @@ Compile and install the extension (supports Postgres 11+)
 
 ```sh
 cd /tmp
-git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
+git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
 cd pgvector
 make
 make install # may need sudo
@@ -509,7 +509,7 @@ Then use `nmake` to build:
 
 ```cmd
 set "PGROOT=C:\Program Files\PostgreSQL\15"
-git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
+git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
 cd pgvector
 nmake /F Makefile.win
 nmake /F Makefile.win install
@@ -530,7 +530,7 @@ This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (r
 You can also build the image manually:
 
 ```sh
-git clone --branch v0.5.0 https://github.com/pgvector/pgvector.git
+git clone --branch v0.5.1 https://github.com/pgvector/pgvector.git
 cd pgvector
 docker build --build-arg PG_MAJOR=15 -t myuser/pgvector .
 ```

sql/vector--0.5.0--0.5.1.sql

@@ -1,67 +1,2 @@
 -- complain if script is sourced in psql, rather than via CREATE EXTENSION
 \echo Use "ALTER EXTENSION vector UPDATE TO '0.5.1'" to load this file. \quit
-
--- tinyint
-
-CREATE TYPE tinyint;
-
-CREATE FUNCTION tinyint_in(cstring, oid, integer) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_out(tinyint) RETURNS cstring
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_recv(internal, oid, integer) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_send(tinyint) RETURNS bytea
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE TYPE tinyint (
-	INPUT     = tinyint_in,
-	OUTPUT    = tinyint_out,
-	RECEIVE   = tinyint_recv,
-	SEND      = tinyint_send,
-	INTERNALLENGTH = 1,
-	PASSEDBYVALUE,
-	ALIGNMENT = char
-);
-
-CREATE FUNCTION integer_to_tinyint(integer, integer, boolean) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION numeric_to_tinyint(numeric, integer, boolean) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE CAST (integer AS tinyint)
-	WITH FUNCTION integer_to_tinyint(integer, integer, boolean) AS IMPLICIT;
-
-CREATE CAST (numeric AS tinyint)
-	WITH FUNCTION numeric_to_tinyint(numeric, integer, boolean) AS IMPLICIT;
-
-CREATE FUNCTION l2_distance(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME', 'tinyint_l2_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION inner_product(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME', 'tinyint_inner_product' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION cosine_distance(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME', 'tinyint_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_negative_inner_product(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE OPERATOR <-> (
-	LEFTARG = tinyint[], RIGHTARG = tinyint[], PROCEDURE = l2_distance,
-	COMMUTATOR = '<->'
-);
-
-CREATE OPERATOR <#> (
-	LEFTARG = tinyint[], RIGHTARG = tinyint[], PROCEDURE = tinyint_negative_inner_product,
-	COMMUTATOR = '<#>'
-);
-
-CREATE OPERATOR <=> (
-	LEFTARG = tinyint[], RIGHTARG = tinyint[], PROCEDURE = cosine_distance,
-	COMMUTATOR = '<=>'
-);

sql/vector.sql

@@ -290,68 +290,3 @@ CREATE OPERATOR CLASS vector_cosine_ops
 	OPERATOR 1 <=> (vector, vector) FOR ORDER BY float_ops,
 	FUNCTION 1 vector_negative_inner_product(vector, vector),
 	FUNCTION 2 vector_norm(vector);
-
--- tinyint
-
-CREATE TYPE tinyint;
-
-CREATE FUNCTION tinyint_in(cstring, oid, integer) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_out(tinyint) RETURNS cstring
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_recv(internal, oid, integer) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_send(tinyint) RETURNS bytea
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE TYPE tinyint (
-	INPUT     = tinyint_in,
-	OUTPUT    = tinyint_out,
-	RECEIVE   = tinyint_recv,
-	SEND      = tinyint_send,
-	INTERNALLENGTH = 1,
-	PASSEDBYVALUE,
-	ALIGNMENT = char
-);
-
-CREATE FUNCTION integer_to_tinyint(integer, integer, boolean) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION numeric_to_tinyint(numeric, integer, boolean) RETURNS tinyint
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE CAST (integer AS tinyint)
-	WITH FUNCTION integer_to_tinyint(integer, integer, boolean) AS IMPLICIT;
-
-CREATE CAST (numeric AS tinyint)
-	WITH FUNCTION numeric_to_tinyint(numeric, integer, boolean) AS IMPLICIT;
-
-CREATE FUNCTION l2_distance(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME', 'tinyint_l2_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION inner_product(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME', 'tinyint_inner_product' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION cosine_distance(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME', 'tinyint_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION tinyint_negative_inner_product(tinyint[], tinyint[]) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE OPERATOR <-> (
-	LEFTARG = tinyint[], RIGHTARG = tinyint[], PROCEDURE = l2_distance,
-	COMMUTATOR = '<->'
-);
-
-CREATE OPERATOR <#> (
-	LEFTARG = tinyint[], RIGHTARG = tinyint[], PROCEDURE = tinyint_negative_inner_product,
-	COMMUTATOR = '<#>'
-);
-
-CREATE OPERATOR <=> (
-	LEFTARG = tinyint[], RIGHTARG = tinyint[], PROCEDURE = cosine_distance,
-	COMMUTATOR = '<=>'
-);

src/hnsw.c

@@ -33,6 +33,12 @@ HnswInit(void)
 					  HNSW_DEFAULT_EF_CONSTRUCTION, HNSW_MIN_EF_CONSTRUCTION, HNSW_MAX_EF_CONSTRUCTION
 #if PG_VERSION_NUM >= 130000
 					  ,AccessExclusiveLock
+#endif
+		);
+	add_int_reloption(hnsw_relopt_kind, "dimensions", "Number of dimensions",
+					  HNSW_DEFAULT_DIMENSIONS, HNSW_MIN_DIMENSIONS, HNSW_MAX_DIMENSIONS
+#if PG_VERSION_NUM >= 130000
+					  ,AccessExclusiveLock
 #endif
 		);
 
@@ -125,6 +131,7 @@ hnswoptions(Datum reloptions, bool validate)
 	static const relopt_parse_elt tab[] = {
 		{"m", RELOPT_TYPE_INT, offsetof(HnswOptions, m)},
 		{"ef_construction", RELOPT_TYPE_INT, offsetof(HnswOptions, efConstruction)},
+		{"dimensions", RELOPT_TYPE_INT, offsetof(HnswOptions, dimensions)},
 	};
 
 #if PG_VERSION_NUM >= 130000

src/hnsw.h

@@ -42,6 +42,9 @@
 #define HNSW_DEFAULT_EF_SEARCH	40
 #define HNSW_MIN_EF_SEARCH		1
 #define HNSW_MAX_EF_SEARCH		1000
+#define HNSW_DEFAULT_DIMENSIONS	-1
+#define HNSW_MIN_DIMENSIONS		1
+#define HNSW_MAX_DIMENSIONS		HNSW_MAX_DIM
 
 /* Tuple types */
 #define HNSW_ELEMENT_TUPLE_TYPE  1
@@ -59,7 +62,7 @@
 
 #define HNSW_MAX_SIZE (BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - sizeof(ItemIdData))
 
-#define HNSW_ELEMENT_TUPLE_SIZE(_dim)	MAXALIGN(offsetof(HnswElementTupleData, vec) + VECTOR_SIZE(_dim))
+#define HNSW_ELEMENT_TUPLE_SIZE(_datum)	MAXALIGN(offsetof(HnswElementTupleData, value) + VARSIZE_ANY(_datum))
 #define HNSW_NEIGHBOR_TUPLE_SIZE(level, m)	MAXALIGN(offsetof(HnswNeighborTupleData, indextids) + ((level) + 2) * (m) * sizeof(ItemPointerData))
 
 #define HnswPageGetOpaque(page)	((HnswPageOpaque) PageGetSpecialPointer(page))
@@ -103,7 +106,7 @@ typedef struct HnswElementData
 	OffsetNumber offno;
 	OffsetNumber neighborOffno;
 	BlockNumber neighborPage;
-	Vector	   *vec;
+	Datum		value;
 }			HnswElementData;
 
 typedef HnswElementData * HnswElement;
@@ -112,11 +115,13 @@ typedef struct HnswCandidate
 {
 	HnswElement element;
 	float		distance;
+	bool		closer;
 }			HnswCandidate;
 
 typedef struct HnswNeighborArray
 {
 	int			length;
+	bool		closerSet;
 	HnswCandidate *items;
 }			HnswNeighborArray;
 
@@ -132,6 +137,7 @@ typedef struct HnswOptions
 	int32		vl_len_;		/* varlena header (do not touch directly!) */
 	int			m;				/* number of connections */
 	int			efConstruction; /* size of dynamic candidate list */
+	int			dimensions;
 }			HnswOptions;
 
 typedef struct HnswBuildState
@@ -202,7 +208,7 @@ typedef struct HnswElementTupleData
 	ItemPointerData heaptids[HNSW_HEAPTIDS];
 	ItemPointerData neighbortid;
 	uint16		unused2;
-	Vector		vec;
+	char		value[FLEXIBLE_ARRAY_MEMBER];
 }			HnswElementTupleData;
 
 typedef HnswElementTupleData * HnswElementTuple;
@@ -260,6 +266,7 @@ typedef struct HnswVacuumState
 /* Methods */
 int			HnswGetM(Relation index);
 int			HnswGetEfConstruction(Relation index);
+int			HnswGetDimensions(Relation index);
 FmgrInfo   *HnswOptionalProcInfo(Relation index, uint16 procnum);
 bool		HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
 void		HnswCommitBuffer(Buffer buf, GenericXLogState *state);

src/hnswbuild.c

@@ -8,6 +8,7 @@
 #include "lib/pairingheap.h"
 #include "nodes/pg_list.h"
 #include "storage/bufmgr.h"
+#include "utils/datum.h"
 #include "utils/memutils.h"
 
 #if PG_VERSION_NUM >= 140000
@@ -105,8 +106,6 @@ CreateElementPages(HnswBuildState * buildstate)
 {
 	Relation	index = buildstate->index;
 	ForkNumber	forkNum = buildstate->forkNum;
-	int			dimensions = buildstate->dimensions;
-	Size		etupSize;
 	Size		maxSize;
 	HnswElementTuple etup;
 	HnswNeighborTuple ntup;
@@ -118,10 +117,9 @@ CreateElementPages(HnswBuildState * buildstate)
 
 	/* Calculate sizes */
 	maxSize = HNSW_MAX_SIZE;
-	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
 
 	/* Allocate once */
-	etup = palloc0(etupSize);
+	etup = palloc0(BLCKSZ);
 	ntup = palloc0(BLCKSZ);
 
 	/* Prepare first page */
@@ -133,12 +131,14 @@ CreateElementPages(HnswBuildState * buildstate)
 	foreach(lc, buildstate->elements)
 	{
 		HnswElement element = lfirst(lc);
+		Size		etupSize;
 		Size		ntupSize;
 		Size		combinedSize;
 
 		HnswSetElementTuple(etup, element);
 
 		/* Calculate sizes */
+		etupSize = HNSW_ELEMENT_TUPLE_SIZE(element->value);
 		ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, buildstate->m);
 		combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
 
@@ -273,18 +273,15 @@ InsertTuple(Relation index, Datum *values, HnswElement element, HnswBuildState *
 	int			m = buildstate->m;
 
 	/* Detoast once for all calls */
-	Datum		value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
+	element->value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
 	/* Normalize if needed */
 	if (buildstate->normprocinfo != NULL)
 	{
-		if (!HnswNormValue(buildstate->normprocinfo, collation, &value, buildstate->normvec))
+		if (!HnswNormValue(buildstate->normprocinfo, collation, &element->value, buildstate->normvec))
 			return false;
 	}
 
-	/* Copy value to element so accessible outside of memory context */
-	memcpy(element->vec, DatumGetVector(value), VECTOR_SIZE(buildstate->dimensions));
-
 	/* Insert element in graph */
 	HnswInsertElement(element, entryPoint, NULL, procinfo, collation, m, efConstruction, false);
 
@@ -360,7 +357,6 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 
 	/* Allocate necessary memory outside of memory context */
 	element = HnswInitElement(tid, buildstate->m, buildstate->ml, buildstate->maxLevel);
-	element->vec = palloc(VECTOR_SIZE(buildstate->dimensions));
 
 	/* Use memory context since detoast can allocate */
 	oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
@@ -368,9 +364,8 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 	/* Insert tuple */
 	inserted = InsertTuple(index, values, element, buildstate, &dup);
 
-	/* Reset memory context */
+	/* Switch memory context */
 	MemoryContextSwitchTo(oldCtx);
-	MemoryContextReset(buildstate->tmpCtx);
 
 	/* Add outside memory context */
 	if (dup != NULL)
@@ -378,9 +373,18 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
 
 	/* Add to buildstate or free */
 	if (inserted)
+	{
+		element->value = datumCopy(element->value, false, -1);
 		buildstate->elements = lappend(buildstate->elements, element);
+	}
 	else
+	{
+		element->value = PointerGetDatum(NULL);
 		HnswFreeElement(element);
+	}
+
+	/* Reset memory context */
+	MemoryContextReset(buildstate->tmpCtx);
 }
 
 /*
@@ -395,6 +399,7 @@ HnswGetMaxInMemoryElements(int m, double ml, int dimensions)
 	elementSize += sizeof(HnswNeighborArray) * (avgLevel + 1);
 	elementSize += sizeof(HnswCandidate) * (m * (avgLevel + 2));
 	elementSize += sizeof(ItemPointerData);
+	/* TODO Handle non-vector types */
 	elementSize += VECTOR_SIZE(dimensions);
 	return (maintenance_work_mem * 1024L) / elementSize;
 }
@@ -412,7 +417,10 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 
 	buildstate->m = HnswGetM(index);
 	buildstate->efConstruction = HnswGetEfConstruction(index);
-	buildstate->dimensions = TupleDescAttr(index->rd_att, 0)->atttypmod;
+	buildstate->dimensions = HnswGetDimensions(index);
+
+	if (buildstate->dimensions < 0)
+		buildstate->dimensions = TupleDescAttr(index->rd_att, 0)->atttypmod;
 
 	/* Require column to have dimensions to be indexed */
 	if (buildstate->dimensions < 0)

src/hnswinsert.c

@@ -123,7 +123,6 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	Size		minCombinedSize;
 	HnswElementTuple etup;
 	BlockNumber currentPage = insertPage;
-	int			dimensions = e->vec->dim;
 	HnswNeighborTuple ntup;
 	Buffer		nbuf;
 	Page		npage;
@@ -132,7 +131,7 @@ WriteNewElementPages(Relation index, HnswElement e, int m, BlockNumber insertPag
 	BlockNumber newInsertPage = InvalidBlockNumber;
 
 	/* Calculate sizes */
-	etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
+	etupSize = HNSW_ELEMENT_TUPLE_SIZE(e->value);
 	ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
 	combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
 	maxSize = HNSW_MAX_SIZE;
@@ -405,7 +404,7 @@ HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
 	Buffer		buf;
 	Page		page;
 	GenericXLogState *state;
-	Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(dup->vec->dim);
+	Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(dup->value);
 	HnswElementTuple etup;
 	int			i;
 
@@ -515,7 +514,7 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 
 	/* Create an element */
 	element = HnswInitElement(heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m));
-	element->vec = DatumGetVector(value);
+	element->value = value;
 
 	/* Prevent concurrent inserts when likely updating entry point */
 	if (entryPoint == NULL || element->level > entryPoint->level)

src/hnswutils.c

@@ -4,6 +4,7 @@
 
 #include "hnsw.h"
 #include "storage/bufmgr.h"
+#include "utils/datum.h"
 #include "vector.h"
 
 /*
@@ -34,6 +35,20 @@ HnswGetEfConstruction(Relation index)
 	return HNSW_DEFAULT_EF_CONSTRUCTION;
 }
 
+/*
+ * Get the number of dimensions in the index
+ */
+int
+HnswGetDimensions(Relation index)
+{
+	HnswOptions *opts = (HnswOptions *) index->rd_options;
+
+	if (opts)
+		return opts->dimensions;
+
+	return HNSW_DEFAULT_DIMENSIONS;
+}
+
 /*
  * Get proc
  */
@@ -139,6 +154,7 @@ HnswInitNeighbors(HnswElement element, int m)
 		a = &element->neighbors[lc];
 		a->length = 0;
 		a->items = palloc(sizeof(HnswCandidate) * lm);
+		a->closerSet = false;
 	}
 }
 
@@ -186,7 +202,8 @@ HnswFreeElement(HnswElement element)
 {
 	HnswFreeNeighbors(element);
 	list_free_deep(element->heaptids);
-	pfree(element->vec);
+	if (DatumGetPointer(element->value))
+		pfree(DatumGetPointer(element->value));
 	pfree(element);
 }
 
@@ -213,7 +230,7 @@ HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
 	element->blkno = blkno;
 	element->offno = offno;
 	element->neighbors = NULL;
-	element->vec = NULL;
+	element->value = PointerGetDatum(NULL);
 	return element;
 }
 
@@ -323,7 +340,7 @@ HnswSetElementTuple(HnswElementTuple etup, HnswElement element)
 		else
 			ItemPointerSetInvalid(&etup->heaptids[i]);
 	}
-	memcpy(&etup->vec, element->vec, VECTOR_SIZE(element->vec->dim));
+	memcpy(&etup->value, DatumGetPointer(element->value), VARSIZE_ANY(element->value));
 }
 
 /*
@@ -446,8 +463,9 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
 
 	if (loadVec)
 	{
-		element->vec = palloc(VECTOR_SIZE(etup->vec.dim));
-		memcpy(element->vec, &etup->vec, VECTOR_SIZE(etup->vec.dim));
+		Datum		value = PointerGetDatum(&etup->value);
+
+		element->value = datumCopy(value, false, -1);
 	}
 }
 
@@ -475,7 +493,7 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
 
 	/* Calculate distance */
 	if (distance != NULL)
-		*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->vec)));
+		*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->value)));
 
 	UnlockReleaseBuffer(buf);
 }
@@ -486,7 +504,7 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
 static float
 GetCandidateDistance(HnswCandidate * hc, Datum q, FmgrInfo *procinfo, Oid collation)
 {
-	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, PointerGetDatum(hc->element->vec)));
+	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, hc->element->value));
 }
 
 /*
@@ -692,6 +710,34 @@ HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *pro
 	return w;
 }
 
+/*
+ * Compare candidate distances
+ */
+static int
+#if PG_VERSION_NUM >= 130000
+CompareCandidateDistances(const ListCell *a, const ListCell *b)
+#else
+CompareCandidateDistances(const void *a, const void *b)
+#endif
+{
+	HnswCandidate *hca = lfirst((ListCell *) a);
+	HnswCandidate *hcb = lfirst((ListCell *) b);
+
+	if (hca->distance < hcb->distance)
+		return 1;
+
+	if (hca->distance > hcb->distance)
+		return -1;
+
+	if (hca->element < hcb->element)
+		return 1;
+
+	if (hca->element > hcb->element)
+		return -1;
+
+	return 0;
+}
+
 /*
  * Calculate the distance between elements
  */
@@ -721,7 +767,7 @@ HnswGetDistance(HnswElement a, HnswElement b, int lc, FmgrInfo *procinfo, Oid co
 		}
 	}
 
-	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(a->vec), PointerGetDatum(b->vec)));
+	return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, a->value, b->value));
 }
 
 /*
@@ -748,33 +794,77 @@ CheckElementCloser(HnswCandidate * e, List *r, int lc, FmgrInfo *procinfo, Oid c
  * Algorithm 4 from paper
  */
 static List *
-SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswCandidate * *pruned)
+SelectNeighbors(List *c, int m, int lc, FmgrInfo *procinfo, Oid collation, HnswElement e2, HnswCandidate * newCandidate, HnswCandidate * *pruned, bool sortCandidates)
 {
 	List	   *r = NIL;
 	List	   *w = list_copy(c);
 	pairingheap *wd;
+	bool		mustCalculate = !e2->neighbors[lc].closerSet;
+	List	   *added = NIL;
+	bool		removedAny = false;
 
 	if (list_length(w) <= m)
 		return w;
 
 	wd = pairingheap_allocate(CompareNearestCandidates, NULL);
 
+	/* Ensure order of candidates is deterministic for closer caching */
+	if (sortCandidates)
+		list_sort(w, CompareCandidateDistances);
+
 	while (list_length(w) > 0 && list_length(r) < m)
 	{
 		/* Assumes w is already ordered desc */
 		HnswCandidate *e = llast(w);
-		bool		closer;
 
 		w = list_delete_last(w);
 
-		closer = CheckElementCloser(e, r, lc, procinfo, collation);
+		/* Use previous state of r and wd to skip work when possible */
+		if (mustCalculate)
+			e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
+		else if (list_length(added) > 0)
+		{
+			/*
+			 * If the current candidate was closer, we only need to compare it
+			 * with the other candidates that we have added.
+			 */
+			if (e->closer)
+			{
+				e->closer = CheckElementCloser(e, added, lc, procinfo, collation);
 
-		if (closer)
+				if (!e->closer)
+					removedAny = true;
+			}
+			else
+			{
+				/*
+				 * If we have removed any candidates from closer, a candidate
+				 * that was not closer earlier might now be.
+				 */
+				if (removedAny)
+				{
+					e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
+					if (e->closer)
+						added = lappend(added, e);
+				}
+			}
+		}
+		else if (e == newCandidate)
+		{
+			e->closer = CheckElementCloser(e, r, lc, procinfo, collation);
+			if (e->closer)
+				added = lappend(added, e);
+		}
+
+		if (e->closer)
 			r = lappend(r, e);
 		else
 			pairingheap_add(wd, &(CreatePairingHeapNode(e)->ph_node));
 	}
 
+	/* Cached value can only be used in future if sorted deterministically */
+	e2->neighbors[lc].closerSet = sortCandidates;
+
 	/* Keep pruned connections */
 	while (!pairingheap_is_empty(wd) && list_length(r) < m)
 		r = lappend(r, ((HnswPairingHeapNode *) pairingheap_remove_first(wd))->inner);
@@ -804,7 +894,7 @@ HnswFindDuplicate(HnswElement e)
 		HnswCandidate *neighbor = &neighbors->items[i];
 
 		/* Exit early since ordered by distance */
-		if (vector_cmp_internal(e->vec, neighbor->element->vec) != 0)
+		if (!datumIsEqual(e->value, neighbor->element->value, false, -1))
 			break;
 
 		/* Check for space */
@@ -828,28 +918,6 @@ AddConnections(HnswElement element, List *neighbors, int m, int lc)
 		a->items[a->length++] = *((HnswCandidate *) lfirst(lc2));
 }
 
-/*
- * Compare candidate distances
- */
-static int
-#if PG_VERSION_NUM >= 130000
-CompareCandidateDistances(const ListCell *a, const ListCell *b)
-#else
-CompareCandidateDistances(const void *a, const void *b)
-#endif
-{
-	HnswCandidate *hca = lfirst((ListCell *) a);
-	HnswCandidate *hcb = lfirst((ListCell *) b);
-
-	if (hca->distance < hcb->distance)
-		return 1;
-
-	if (hca->distance > hcb->distance)
-		return -1;
-
-	return 0;
-}
-
 /*
  * Update connections
  */
@@ -879,13 +947,13 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
 		/* Load elements on insert */
 		if (index != NULL)
 		{
-			Datum		q = PointerGetDatum(hc->element->vec);
+			Datum		q = hc->element->value;
 
 			for (int i = 0; i < currentNeighbors->length; i++)
 			{
 				HnswCandidate *hc3 = &currentNeighbors->items[i];
 
-				if (hc3->element->vec == NULL)
+				if (!DatumGetPointer(hc3->element->value))
 					HnswLoadElement(hc3->element, &hc3->distance, &q, index, procinfo, collation, true);
 				else
 					hc3->distance = GetCandidateDistance(hc3, q, procinfo, collation);
@@ -903,13 +971,12 @@ HnswUpdateConnection(HnswElement element, HnswCandidate * hc, int m, int lc, int
 		{
 			List	   *c = NIL;
 
-			/* Add and sort candidates */
+			/* Add candidates */
 			for (int i = 0; i < currentNeighbors->length; i++)
 				c = lappend(c, &currentNeighbors->items[i]);
 			c = lappend(c, &hc2);
-			list_sort(c, CompareCandidateDistances);
 
-			SelectNeighbors(c, m, lc, procinfo, collation, &pruned);
+			SelectNeighbors(c, m, lc, procinfo, collation, hc->element, &hc2, &pruned, true);
 
 			/* Should not happen */
 			if (pruned == NULL)
@@ -967,7 +1034,7 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
 	List	   *w;
 	int			level = element->level;
 	int			entryLevel;
-	Datum		q = PointerGetDatum(element->vec);
+	Datum		q = element->value;
 	HnswElement skipElement = existing ? element : NULL;
 
 	/* No neighbors if no entry point */
@@ -1008,7 +1075,12 @@ HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, F
 		else
 			lw = w;
 
-		neighbors = SelectNeighbors(lw, lm, lc, procinfo, collation, NULL);
+		/*
+		 * Candidates are sorted, but not deterministically. Could set
+		 * sortCandidates to true for in-memory builds to enable closer
+		 * caching, but there does not seem to be a difference in performance.
+		 */
+		neighbors = SelectNeighbors(lw, lm, lc, procinfo, collation, element, NULL, NULL, false);
 
 		AddConnections(element, neighbors, lm, lc);
 

src/hnswvacuum.c

@@ -93,7 +93,7 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
 
 				if (itemUpdated)
 				{
-					Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
+					Size		etupSize = HNSW_ELEMENT_TUPLE_SIZE(PointerGetDatum(&etup->value));
 
 					/* Mark rest as invalid */
 					for (int i = idx; i < HNSW_HEAPTIDS; i++)
@@ -481,6 +481,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 			HnswNeighborTuple ntup;
 			Size		etupSize;
 			Size		ntupSize;
+			Datum		value;
 			Buffer		nbuf;
 			Page		npage;
 			BlockNumber neighborPage;
@@ -504,8 +505,11 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 			if (ItemPointerIsValid(&etup->heaptids[0]))
 				continue;
 
+			/* Get datum */
+			value = PointerGetDatum(&etup->value);
+
 			/* Calculate sizes */
-			etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
+			etupSize = HNSW_ELEMENT_TUPLE_SIZE(value);
 			ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(etup->level, vacuumstate->m);
 
 			/* Get neighbor page */
@@ -528,7 +532,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
 
 			/* Overwrite element */
 			etup->deleted = 1;
-			MemSet(&etup->vec.x, 0, etup->vec.dim * sizeof(float));
+			MemSet(&etup->value, 0, VARSIZE_ANY(value));
 
 			/* Overwrite neighbors */
 			for (int i = 0; i < ntup->count; i++)

src/ivfflat.h

@@ -246,8 +246,6 @@ typedef struct IvfflatScanOpaqueData
 	int			probes;
 	int			dimensions;
 	bool		first;
-	Buffer		buf;
-	ItemPointerData heaptid;
 
 	/* Sorting */
 	Tuplesortstate *sortstate;

src/ivfscan.c

@@ -143,10 +143,6 @@ GetScanItems(IndexScanDesc scan, Datum value)
 				bool		isnull;
 				ItemId		itemid = PageGetItemId(page, offno);
 
-				/* Skip dead tuples */
-				if (scan->ignore_killed_tuples && ItemIdIsDead(itemid))
-					continue;
-
 				itup = (IndexTuple) PageGetItem(page, itemid);
 				datum = index_getattr(itup, 1, tupdesc, &isnull);
 
@@ -161,8 +157,6 @@ GetScanItems(IndexScanDesc scan, Datum value)
 				slot->tts_isnull[0] = false;
 				slot->tts_values[1] = PointerGetDatum(&itup->t_tid);
 				slot->tts_isnull[1] = false;
-				slot->tts_values[2] = Int32GetDatum((int) searchPage);
-				slot->tts_isnull[2] = false;
 				ExecStoreVirtualTuple(slot);
 
 				tuplesort_puttupleslot(so->sortstate, slot);
@@ -187,55 +181,6 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	tuplesort_performsort(so->sortstate);
 }
 
-/*
- * Mark prior tuple as dead
- */
-static void
-MarkPriorTupleDead(IndexScanDesc scan)
-{
-	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
-	Buffer		buf = so->buf;
-	Page		page;
-	OffsetNumber maxoffno;
-
-	/* Safety check */
-	if (!BufferIsValid(so->buf) || !ItemPointerIsValid(&so->heaptid))
-		return;
-
-	/* Only a shared locked is needed for ItemIdMarkDead */
-	LockBuffer(buf, BUFFER_LOCK_SHARE);
-	page = BufferGetPage(buf);
-	maxoffno = PageGetMaxOffsetNumber(page);
-
-	for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
-	{
-		ItemId		itemid = PageGetItemId(page, offno);
-		IndexTuple	itup = (IndexTuple) PageGetItem(page, itemid);
-
-		/*
-		 * Find tuple. Since buffer has been pinned, tuple cannot have been
-		 * vacuumed (and heap TID reused).
-		 */
-		if (ItemPointerEquals(&itup->t_tid, &so->heaptid))
-		{
-			/*
-			 * Make sure tuple has not already been marked dead to avoid extra
-			 * WAL if wal_log_hints or data checksums enabled
-			 */
-			if (!ItemIdIsDead(itemid))
-			{
-				ItemIdMarkDead(itemid);
-				MarkBufferDirtyHint(buf, true);
-			}
-
-			break;
-		}
-	}
-
-	/* Unlock buffer */
-	LockBuffer(buf, BUFFER_LOCK_UNLOCK);
-}
-
 /*
  * Prepare for an index scan
  */
@@ -261,9 +206,7 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 		probes = lists;
 
 	so = (IvfflatScanOpaque) palloc(offsetof(IvfflatScanOpaqueData, lists) + probes * sizeof(IvfflatScanList));
-	so->buf = InvalidBuffer;
 	so->first = true;
-	ItemPointerSetInvalid(&so->heaptid);
 	so->probes = probes;
 	so->dimensions = dimensions;
 
@@ -274,13 +217,12 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 
 	/* Create tuple description for sorting */
 #if PG_VERSION_NUM >= 120000
-	so->tupdesc = CreateTemplateTupleDesc(3);
+	so->tupdesc = CreateTemplateTupleDesc(2);
 #else
-	so->tupdesc = CreateTemplateTupleDesc(3, false);
+	so->tupdesc = CreateTemplateTupleDesc(2, false);
 #endif
 	TupleDescInitEntry(so->tupdesc, (AttrNumber) 1, "distance", FLOAT8OID, -1, 0);
 	TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "heaptid", TIDOID, -1, 0);
-	TupleDescInitEntry(so->tupdesc, (AttrNumber) 3, "indexblkno", INT4OID, -1, 0);
 
 	/* Prep sort */
 	so->sortstate = tuplesort_begin_heap(so->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, work_mem, NULL, false);
@@ -312,7 +254,6 @@ ivfflatrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int
 #endif
 
 	so->first = true;
-	ItemPointerSetInvalid(&so->heaptid);
 	pairingheap_reset(so->listQueue);
 
 	if (keys && scan->numberOfKeys > 0)
@@ -347,6 +288,11 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		if (scan->orderByData == NULL)
 			elog(ERROR, "cannot scan ivfflat index without order");
 
+		/* Requires MVCC-compliant snapshot as not able to pin during sorting */
+		/* https://www.postgresql.org/docs/current/index-locking.html */
+		if (!IsMVCCSnapshot(scan->xs_snapshot))
+			elog(ERROR, "non-MVCC snapshots are not supported with ivfflat");
+
 		if (scan->orderByData->sk_flags & SK_ISNULL)
 			value = PointerGetDatum(InitVector(so->dimensions));
 		else
@@ -370,17 +316,10 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		if (value != scan->orderByData->sk_argument)
 			pfree(DatumGetPointer(value));
 	}
-	else
-	{
-		/* Mark prior tuple as dead */
-		if (scan->kill_prior_tuple)
-			MarkPriorTupleDead(scan);
-	}
 
 	if (tuplesort_gettupleslot(so->sortstate, true, false, so->slot, NULL))
 	{
 		ItemPointer heaptid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
-		BlockNumber indexblkno = DatumGetInt32(slot_getattr(so->slot, 3, &so->isnull));
 
 #if PG_VERSION_NUM >= 120000
 		scan->xs_heaptid = *heaptid;
@@ -388,21 +327,6 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		scan->xs_ctup.t_self = *heaptid;
 #endif
 
-		/* Keep track of info needed to mark tuple as dead */
-		so->heaptid = *heaptid;
-
-		/* Unpin buffer */
-		if (BufferIsValid(so->buf))
-			ReleaseBuffer(so->buf);
-
-		/*
-		 * An index scan must maintain a pin on the index page holding the
-		 * item last returned by amgettuple
-		 *
-		 * https://www.postgresql.org/docs/current/index-locking.html
-		 */
-		so->buf = ReadBuffer(scan->indexRelation, indexblkno);
-
 		scan->xs_recheckorderby = false;
 		return true;
 	}
@@ -418,10 +342,6 @@ ivfflatendscan(IndexScanDesc scan)
 {
 	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
 
-	/* Release pin */
-	if (BufferIsValid(so->buf))
-		ReleaseBuffer(so->buf);
-
 	pairingheap_free(so->listQueue);
 	tuplesort_end(so->sortstate);
 

src/tinyint.c

@@ -1,294 +0,0 @@
-#include "postgres.h"
-
-#include <math.h>
-#include <stdint.h>
-
-#include "fmgr.h"
-#include "lib/stringinfo.h"
-#include "libpq/pqformat.h"
-#include "tinyint.h"
-#include "utils/array.h"
-#include "utils/builtins.h"
-#include "utils/numeric.h"
-
-/*
- * Check if array is a vector
- */
-static bool
-ArrayIsVector(ArrayType *a)
-{
-	return ARR_NDIM(a) == 1 && !array_contains_nulls(a);
-}
-
-/*
- * Check if dimensions are the same
- */
-static int
-CheckDims(ArrayType *a, ArrayType *b)
-{
-	int			dima;
-	int			dimb;
-
-	if (!ArrayIsVector(a) || !ArrayIsVector(b))
-		return 0;
-
-	dima = ARR_DIMS(a)[0];
-	dimb = ARR_DIMS(b)[0];
-
-	if (dima != dimb)
-		return 0;
-
-	return dima;
-}
-
-/*
- * Check range
- */
-static void
-CheckRange(long i)
-{
-	if (i < INT8_MIN || i > INT8_MAX)
-		ereport(ERROR,
-				(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
-				 errmsg("value \"%ld\" is out of range for type tinyint", i)));
-}
-
-/*
- * Convert textual representation to internal representation
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_in);
-Datum
-tinyint_in(PG_FUNCTION_ARGS)
-{
-	char	   *s = PG_GETARG_CSTRING(0);
-	const char *ptr = s;
-	long		i;
-	char	   *end;
-
-	/* skip leading spaces */
-	while (*ptr != '\0' && isspace((unsigned char) *ptr))
-		ptr++;
-
-	if (*ptr == '\0')
-		ereport(ERROR,
-				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-				 errmsg("invalid input syntax for type tinyint: \"%s\"", s)));
-
-	i = strtol(ptr, &end, 10);
-	ptr = end;
-
-	if (i < INT8_MIN || i > INT8_MAX)
-		ereport(ERROR,
-				(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
-				 errmsg("value \"%s\" is out of range for type tinyint", s)));
-
-	/* allow trailing whitespace, but not other trailing chars */
-	while (*ptr != '\0' && isspace((unsigned char) *ptr))
-		ptr++;
-
-	if (*ptr != '\0')
-		ereport(ERROR,
-				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-				 errmsg("invalid input syntax for type tinyint: \"%s\"", s)));
-
-	PG_RETURN_INT8(i);
-}
-
-/*
- * Convert internal representation to textual representation
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_out);
-Datum
-tinyint_out(PG_FUNCTION_ARGS)
-{
-	int8		num = PG_GETARG_INT8(0);
-	char	   *result = (char *) palloc(5);	/* sign, 3 digits, '\0' */
-
-	pg_ltoa((int32) num, result);
-	PG_RETURN_CSTRING(result);
-}
-
-/*
- * Convert external binary representation to internal representation
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_recv);
-Datum
-tinyint_recv(PG_FUNCTION_ARGS)
-{
-	StringInfo	buf = (StringInfo) PG_GETARG_POINTER(0);
-
-	PG_RETURN_INT8((int8) pq_getmsgint(buf, sizeof(int8)));
-}
-
-/*
- * Convert internal representation to the external binary representation
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_send);
-Datum
-tinyint_send(PG_FUNCTION_ARGS)
-{
-	int8		arg1 = PG_GETARG_INT8(0);
-	StringInfoData buf;
-
-	pq_begintypsend(&buf);
-	pq_sendint8(&buf, arg1);
-	PG_RETURN_BYTEA_P(pq_endtypsend(&buf));
-}
-
-/*
- * Convert integer to tinyint
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(integer_to_tinyint);
-Datum
-integer_to_tinyint(PG_FUNCTION_ARGS)
-{
-	int32		i = PG_GETARG_INT32(0);
-
-	CheckRange(i);
-
-	PG_RETURN_INT8(i);
-}
-
-/*
- * Convert numeric to tinyint
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(numeric_to_tinyint);
-Datum
-numeric_to_tinyint(PG_FUNCTION_ARGS)
-{
-	Numeric		num = PG_GETARG_NUMERIC(0);
-	int32		i = numeric_int4_opt_error(num, NULL);
-
-	CheckRange(i);
-
-	PG_RETURN_INT8(i);
-}
-
-/*
- * Get the L2 distance between tinyint arrays
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_l2_distance);
-Datum
-tinyint_l2_distance(PG_FUNCTION_ARGS)
-{
-	ArrayType  *a = PG_GETARG_ARRAYTYPE_P(0);
-	ArrayType  *b = PG_GETARG_ARRAYTYPE_P(1);
-	int8    *ax = (int8 *) ARR_DATA_PTR(a);
-	int8    *bx = (int8 *) ARR_DATA_PTR(b);
-	double		distance = 0.0;
-	int			dim = CheckDims(a, b);
-
-	/* TODO Decide on error or NULL */
-	if (!dim)
-		PG_RETURN_NULL();
-
-	/* Auto-vectorized */
-	for (int i = 0; i < dim; i++)
-	{
-		double		diff = ax[i] - bx[i];
-
-		distance += diff * diff;
-	}
-
-	PG_RETURN_FLOAT8(sqrt(distance));
-}
-
-/*
- * Get the inner product of two tinyint arrays
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_inner_product);
-Datum
-tinyint_inner_product(PG_FUNCTION_ARGS)
-{
-	ArrayType  *a = PG_GETARG_ARRAYTYPE_P(0);
-	ArrayType  *b = PG_GETARG_ARRAYTYPE_P(1);
-	int8	   *ax = (int8 *) ARR_DATA_PTR(a);
-	int8	   *bx = (int8 *) ARR_DATA_PTR(b);
-	double		distance = 0.0;
-	int			dim = CheckDims(a, b);
-
-	/* TODO Decide on error or NULL */
-	if (!dim)
-		PG_RETURN_NULL();
-
-	/* Auto-vectorized */
-	for (int i = 0; i < dim; i++)
-		distance += ax[i] * bx[i];
-
-	PG_RETURN_FLOAT8(distance);
-}
-
-/*
- * Get the negative inner product of two tinyint arrays
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_negative_inner_product);
-Datum
-tinyint_negative_inner_product(PG_FUNCTION_ARGS)
-{
-	ArrayType  *a = PG_GETARG_ARRAYTYPE_P(0);
-	ArrayType  *b = PG_GETARG_ARRAYTYPE_P(1);
-	int8	   *ax = (int8 *) ARR_DATA_PTR(a);
-	int8	   *bx = (int8 *) ARR_DATA_PTR(b);
-	double		distance = 0.0;
-	int			dim = CheckDims(a, b);
-
-	/* TODO Decide on error or NULL */
-	if (!dim)
-		PG_RETURN_NULL();
-
-	/* Auto-vectorized */
-	for (int i = 0; i < dim; i++)
-		distance += ax[i] * bx[i];
-
-	PG_RETURN_FLOAT8(distance * -1);
-}
-
-/*
- * Get the cosine distance between two float2 arrays
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(tinyint_cosine_distance);
-Datum
-tinyint_cosine_distance(PG_FUNCTION_ARGS)
-{
-	ArrayType  *a = PG_GETARG_ARRAYTYPE_P(0);
-	ArrayType  *b = PG_GETARG_ARRAYTYPE_P(1);
-	int8   *ax = (int8 *) ARR_DATA_PTR(a);
-	int8   *bx = (int8 *) ARR_DATA_PTR(b);
-	double		distance = 0.0;
-	double		norma = 0.0;
-	double		normb = 0.0;
-	double		similarity;
-	int			dim = CheckDims(a, b);
-
-	/* TODO Decide on error or NULL */
-	if (!dim)
-		PG_RETURN_NULL();
-
-	/* Auto-vectorized */
-	for (int i = 0; i < dim; i++)
-	{
-		float		axi = ax[i];
-		float		bxi = bx[i];
-
-		distance += axi * bxi;
-		norma += axi * axi;
-		normb += bxi * bxi;
-	}
-
-	/* Use sqrt(a * b) over sqrt(a) * sqrt(b) */
-	similarity = distance / sqrt(norma * normb);
-
-#ifdef _MSC_VER
-	/* /fp:fast may not propagate NaN */
-	if (isnan(similarity))
-		PG_RETURN_FLOAT8(NAN);
-#endif
-
-	/* Keep in range */
-	if (similarity > 1)
-		similarity = 1;
-	else if (similarity < -1)
-		similarity = -1;
-
-	PG_RETURN_FLOAT8(1 - similarity);
-}

src/tinyint.h

@@ -1,8 +0,0 @@
-#ifndef TINYINT_H
-#define TINYINT_H
-
-#define DatumGetInt8(X) ((int8) (X))
-#define PG_GETARG_INT8(n)	 DatumGetInt8(PG_GETARG_DATUM(n))
-#define PG_RETURN_INT8(x)	 return Int8GetDatum(x)
-
-#endif

test/expected/functions.out

@@ -54,85 +54,85 @@ SELECT vector_norm('[3e37,4e37]')::real;
        5e+37
 (1 row)
 
-SELECT l2_distance('[0,0]'::vector, '[3,4]');
+SELECT l2_distance('[0,0]', '[3,4]');
  l2_distance 
 -------------
            5
 (1 row)
 
-SELECT l2_distance('[0,0]'::vector, '[0,1]');
+SELECT l2_distance('[0,0]', '[0,1]');
  l2_distance 
 -------------
            1
 (1 row)
 
-SELECT l2_distance('[1,2]'::vector, '[3]');
+SELECT l2_distance('[1,2]', '[3]');
 ERROR:  different vector dimensions 2 and 1
-SELECT l2_distance('[3e38]'::vector, '[-3e38]');
+SELECT l2_distance('[3e38]', '[-3e38]');
  l2_distance 
 -------------
     Infinity
 (1 row)
 
-SELECT inner_product('[1,2]'::vector, '[3,4]');
+SELECT inner_product('[1,2]', '[3,4]');
  inner_product 
 ---------------
             11
 (1 row)
 
-SELECT inner_product('[1,2]'::vector, '[3]');
+SELECT inner_product('[1,2]', '[3]');
 ERROR:  different vector dimensions 2 and 1
-SELECT inner_product('[3e38]'::vector, '[3e38]');
+SELECT inner_product('[3e38]', '[3e38]');
  inner_product 
 ---------------
       Infinity
 (1 row)
 
-SELECT cosine_distance('[1,2]'::vector, '[2,4]');
+SELECT cosine_distance('[1,2]', '[2,4]');
  cosine_distance 
 -----------------
                0
 (1 row)
 
-SELECT cosine_distance('[1,2]'::vector, '[0,0]');
+SELECT cosine_distance('[1,2]', '[0,0]');
  cosine_distance 
 -----------------
              NaN
 (1 row)
 
-SELECT cosine_distance('[1,1]'::vector, '[1,1]');
+SELECT cosine_distance('[1,1]', '[1,1]');
  cosine_distance 
 -----------------
                0
 (1 row)
 
-SELECT cosine_distance('[1,0]'::vector, '[0,2]');
+SELECT cosine_distance('[1,0]', '[0,2]');
  cosine_distance 
 -----------------
                1
 (1 row)
 
-SELECT cosine_distance('[1,1]'::vector, '[-1,-1]');
+SELECT cosine_distance('[1,1]', '[-1,-1]');
  cosine_distance 
 -----------------
                2
 (1 row)
 
-SELECT cosine_distance('[1,2]'::vector, '[3]');
+SELECT cosine_distance('[1,2]', '[3]');
 ERROR:  different vector dimensions 2 and 1
-SELECT cosine_distance('[1,1]'::vector, '[1.1,1.1]');
+SELECT cosine_distance('[1,1]', '[1.1,1.1]');
  cosine_distance 
 -----------------
                0
 (1 row)
 
-SELECT cosine_distance('[1,1]'::vector, '[-1.1,-1.1]');
+SELECT cosine_distance('[1,1]', '[-1.1,-1.1]');
  cosine_distance 
 -----------------
                2
 (1 row)
 
-SELECT cosine_distance('[3e38]'::vector, '[3e38]');
+SELECT cosine_distance('[3e38]', '[3e38]');
  cosine_distance 
 -----------------
              NaN

test/expected/tinyint.out

@@ -1,148 +0,0 @@
-SELECT '127'::tinyint;
- tinyint 
----------
- 127
-(1 row)
-
-SELECT '128'::tinyint;
-ERROR:  value "128" is out of range for type tinyint
-LINE 1: SELECT '128'::tinyint;
-               ^
-SELECT '-128'::tinyint;
- tinyint 
----------
- -128
-(1 row)
-
-SELECT '-129'::tinyint;
-ERROR:  value "-129" is out of range for type tinyint
-LINE 1: SELECT '-129'::tinyint;
-               ^
-SELECT ''::tinyint;
-ERROR:  invalid input syntax for type tinyint: ""
-LINE 1: SELECT ''::tinyint;
-               ^
-SELECT ' 1'::tinyint;
- tinyint 
----------
- 1
-(1 row)
-
-SELECT '1 '::tinyint;
- tinyint 
----------
- 1
-(1 row)
-
-SELECT '1a'::tinyint;
-ERROR:  invalid input syntax for type tinyint: "1a"
-LINE 1: SELECT '1a'::tinyint;
-               ^
-SELECT '{1,2,3}'::tinyint[];
- tinyint 
----------
- {1,2,3}
-(1 row)
-
-SELECT '128'::numeric::tinyint;
-ERROR:  value "128" is out of range for type tinyint
-SELECT 'NaN'::numeric::tinyint;
-ERROR:  cannot convert NaN to integer
-SELECT l2_distance('{0,0}'::tinyint[], '{3,4}'::tinyint[]);
- l2_distance 
--------------
-           5
-(1 row)
-
-SELECT l2_distance('{0,0}'::tinyint[], '{0,1}'::tinyint[]);
- l2_distance 
--------------
-           1
-(1 row)
-
-SELECT l2_distance('{1,2}'::tinyint[], '{3}'::tinyint[]);
- l2_distance 
--------------
-            
-(1 row)
-
-SELECT l2_distance('{3e38}'::tinyint[], '{-3e38}'::tinyint[]);
-ERROR:  invalid input syntax for type tinyint: "3e38"
-LINE 1: SELECT l2_distance('{3e38}'::tinyint[], '{-3e38}'::tinyint[]...
-                           ^
-SELECT '{0,0}'::tinyint[] <-> '{3,4}'::tinyint[];
- ?column? 
-----------
-        5
-(1 row)
-
-SELECT inner_product('{1,2}'::tinyint[], '{3,4}'::tinyint[]);
- inner_product 
----------------
-            11
-(1 row)
-
-SELECT inner_product('{1,2}'::tinyint[], '{3}'::tinyint[]);
- inner_product 
----------------
-              
-(1 row)
-
-SELECT inner_product('{127}'::tinyint[], '{127}'::tinyint[]);
- inner_product 
----------------
-         16129
-(1 row)
-
-SELECT '{1,2}'::tinyint[] <#> '{3,4}'::tinyint[];
- ?column? 
-----------
-      -11
-(1 row)
-
-SELECT cosine_distance('{1,2}'::tinyint[], '{2,4}'::tinyint[]);
- cosine_distance 
------------------
-               0
-(1 row)
-
-SELECT cosine_distance('{1,2}'::tinyint[], '{0,0}'::tinyint[]);
- cosine_distance 
------------------
-             NaN
-(1 row)
-
-SELECT cosine_distance('{1,1}'::tinyint[], '{1,1}'::tinyint[]);
- cosine_distance 
------------------
-               0
-(1 row)
-
-SELECT cosine_distance('{1,0}'::tinyint[], '{0,2}'::tinyint[]);
- cosine_distance 
------------------
-               1
-(1 row)
-
-SELECT cosine_distance('{1,1}'::tinyint[], '{-1,-1}'::tinyint[]);
- cosine_distance 
------------------
-               2
-(1 row)
-
-SELECT cosine_distance('{1,2}'::tinyint[], '{3}'::tinyint[]);
- cosine_distance 
------------------
-                
-(1 row)
-
-SELECT cosine_distance('{3e38}'::tinyint[], '{3e38}'::tinyint[]);
-ERROR:  invalid input syntax for type tinyint: "3e38"
-LINE 1: SELECT cosine_distance('{3e38}'::tinyint[], '{3e38}'::tinyin...
-                               ^
-SELECT '{1,2}'::tinyint[] <=> '{2,4}'::tinyint[];
- ?column? 
-----------
-        0
-(1 row)
-

test/sql/functions.sql

@@ -13,24 +13,24 @@ SELECT vector_norm('[3,4]');
 SELECT vector_norm('[0,1]');
 SELECT vector_norm('[3e37,4e37]')::real;
 
-SELECT l2_distance('[0,0]'::vector, '[3,4]');
-SELECT l2_distance('[0,0]'::vector, '[0,1]');
-SELECT l2_distance('[1,2]'::vector, '[3]');
-SELECT l2_distance('[3e38]'::vector, '[-3e38]');
+SELECT l2_distance('[0,0]', '[3,4]');
+SELECT l2_distance('[0,0]', '[0,1]');
+SELECT l2_distance('[1,2]', '[3]');
+SELECT l2_distance('[3e38]', '[-3e38]');
 
-SELECT inner_product('[1,2]'::vector, '[3,4]');
-SELECT inner_product('[1,2]'::vector, '[3]');
-SELECT inner_product('[3e38]'::vector, '[3e38]');
+SELECT inner_product('[1,2]', '[3,4]');
+SELECT inner_product('[1,2]', '[3]');
+SELECT inner_product('[3e38]', '[3e38]');
 
-SELECT cosine_distance('[1,2]'::vector, '[2,4]');
-SELECT cosine_distance('[1,2]'::vector, '[0,0]');
-SELECT cosine_distance('[1,1]'::vector, '[1,1]');
-SELECT cosine_distance('[1,0]'::vector, '[0,2]');
-SELECT cosine_distance('[1,1]'::vector, '[-1,-1]');
-SELECT cosine_distance('[1,2]'::vector, '[3]');
-SELECT cosine_distance('[1,1]'::vector, '[1.1,1.1]');
-SELECT cosine_distance('[1,1]'::vector, '[-1.1,-1.1]');
-SELECT cosine_distance('[3e38]'::vector, '[3e38]');
+SELECT cosine_distance('[1,2]', '[2,4]');
+SELECT cosine_distance('[1,2]', '[0,0]');
+SELECT cosine_distance('[1,1]', '[1,1]');
+SELECT cosine_distance('[1,0]', '[0,2]');
+SELECT cosine_distance('[1,1]', '[-1,-1]');
+SELECT cosine_distance('[1,2]', '[3]');
+SELECT cosine_distance('[1,1]', '[1.1,1.1]');
+SELECT cosine_distance('[1,1]', '[-1.1,-1.1]');
+SELECT cosine_distance('[3e38]', '[3e38]');
 
 SELECT l1_distance('[0,0]', '[3,4]');
 SELECT l1_distance('[0,0]', '[0,1]');

test/sql/tinyint.sql

@@ -1,34 +0,0 @@
-SELECT '127'::tinyint;
-SELECT '128'::tinyint;
-SELECT '-128'::tinyint;
-SELECT '-129'::tinyint;
-
-SELECT ''::tinyint;
-SELECT ' 1'::tinyint;
-SELECT '1 '::tinyint;
-SELECT '1a'::tinyint;
-
-SELECT '{1,2,3}'::tinyint[];
-
-SELECT '128'::numeric::tinyint;
-SELECT 'NaN'::numeric::tinyint;
-
-SELECT l2_distance('{0,0}'::tinyint[], '{3,4}'::tinyint[]);
-SELECT l2_distance('{0,0}'::tinyint[], '{0,1}'::tinyint[]);
-SELECT l2_distance('{1,2}'::tinyint[], '{3}'::tinyint[]);
-SELECT l2_distance('{3e38}'::tinyint[], '{-3e38}'::tinyint[]);
-SELECT '{0,0}'::tinyint[] <-> '{3,4}'::tinyint[];
-
-SELECT inner_product('{1,2}'::tinyint[], '{3,4}'::tinyint[]);
-SELECT inner_product('{1,2}'::tinyint[], '{3}'::tinyint[]);
-SELECT inner_product('{127}'::tinyint[], '{127}'::tinyint[]);
-SELECT '{1,2}'::tinyint[] <#> '{3,4}'::tinyint[];
-
-SELECT cosine_distance('{1,2}'::tinyint[], '{2,4}'::tinyint[]);
-SELECT cosine_distance('{1,2}'::tinyint[], '{0,0}'::tinyint[]);
-SELECT cosine_distance('{1,1}'::tinyint[], '{1,1}'::tinyint[]);
-SELECT cosine_distance('{1,0}'::tinyint[], '{0,2}'::tinyint[]);
-SELECT cosine_distance('{1,1}'::tinyint[], '{-1,-1}'::tinyint[]);
-SELECT cosine_distance('{1,2}'::tinyint[], '{3}'::tinyint[]);
-SELECT cosine_distance('{3e38}'::tinyint[], '{3e38}'::tinyint[]);
-SELECT '{1,2}'::tinyint[] <=> '{2,4}'::tinyint[];

test/t/019_hnsw_array.pl

@@ -0,0 +1,113 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+my $node;
+my @queries = ();
+my @expected;
+my $limit = 20;
+
+sub test_recall
+{
+	my ($min, $operator) = @_;
+	my $correct = 0;
+	my $total = 0;
+
+	my $explain = $node->safe_psql("postgres", qq(
+		SET enable_seqscan = off;
+		EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v $operator '$queries[0]' LIMIT $limit;
+	));
+	like($explain, qr/Index Scan/);
+
+	for my $i (0 .. $#queries)
+	{
+		my $actual = $node->safe_psql("postgres", qq(
+			SET enable_seqscan = off;
+			SELECT i FROM tst ORDER BY v $operator '$queries[$i]' LIMIT $limit;
+		));
+		my @actual_ids = split("\n", $actual);
+		my %actual_set = map { $_ => 1 } @actual_ids;
+
+		my @expected_ids = split("\n", $expected[$i]);
+
+		foreach (@expected_ids)
+		{
+			if (exists($actual_set{$_}))
+			{
+				$correct++;
+			}
+			$total++;
+		}
+	}
+
+	cmp_ok($correct / $total, ">=", $min, $operator);
+}
+
+# Initialize node
+$node = get_new_node('node');
+$node->init;
+$node->start;
+
+# Create table
+$node->safe_psql("postgres", "CREATE EXTENSION vector;");
+$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v float4[3]);");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()] FROM generate_series(1, 10000) i;"
+);
+
+$node->safe_psql("postgres", qq(
+	CREATE FUNCTION float4_l2_distance(float4[], float4[]) RETURNS float8
+		AS 'BEGIN RETURN l2_distance(\$1::vector, \$2::vector); END;'
+		LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
+
+	CREATE FUNCTION float4_l2_squared_distance(float4[], float4[]) RETURNS float8
+		AS 'BEGIN RETURN vector_l2_squared_distance(\$1::vector, \$2::vector); END;'
+		LANGUAGE plpgsql IMMUTABLE STRICT PARALLEL SAFE;
+
+	CREATE OPERATOR <-> (
+		LEFTARG = float4[], RIGHTARG = float4[], PROCEDURE = float4_l2_distance,
+		COMMUTATOR = '<->'
+	);
+
+	CREATE OPERATOR CLASS float4_l2_ops
+		FOR TYPE float4[] USING hnsw AS
+		OPERATOR 1 <-> (float4[], float4[]) FOR ORDER BY float_ops,
+		FUNCTION 1 float4_l2_squared_distance(float4[], float4[]);
+));
+
+# Generate queries
+for (1 .. 20)
+{
+	my $r1 = rand();
+	my $r2 = rand();
+	my $r3 = rand();
+	push(@queries, "{$r1,$r2,$r3}");
+}
+
+# Check each index type
+my @operators = ("<->");
+my @opclasses = ("float4_l2_ops");
+
+for my $i (0 .. $#operators)
+{
+	my $operator = $operators[$i];
+	my $opclass = $opclasses[$i];
+
+	# Get exact results
+	@expected = ();
+	foreach (@queries)
+	{
+		my $res = $node->safe_psql("postgres", "SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;");
+		push(@expected, $res);
+	}
+
+	# Add index
+	$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v $opclass) WITH (dimensions = 3);");
+
+	my $min = $operator eq "<#>" ? 0.80 : 0.99;
+	test_recall($min, $operator);
+}
+
+done_testing();

vector.control

@@ -1,4 +1,4 @@
 comment = 'vector data type and ivfflat and hnsw access methods'
-default_version = '0.5.0'
+default_version = '0.5.1'
 module_pathname = '$libdir/vector'
 relocatable = true