Test always

.github/workflows/build.yml

@@ -123,6 +123,6 @@ jobs:
       - uses: ankane/setup-postgres-valgrind@v1
         with:
           postgres-version: 16
-      - run: make
+      - run: make OPTFLAGS=""
       - run: sudo --preserve-env=PG_CONFIG make install
       - run: make installcheck

README.md

@@ -419,6 +419,103 @@ Use [partitioning](https://www.postgresql.org/docs/current/ddl-partitioning.html
 CREATE TABLE items (embedding vector(3), category_id int) PARTITION BY LIST(category_id);
 ```
 
+## Half Vectors
+
+*Unreleased*
+
+Use the `halfvec` type to store half-precision vectors
+
+```sql
+CREATE TABLE items (id bigserial PRIMARY KEY, embedding halfvec(3));
+```
+
+## Half Indexing
+
+*Unreleased*
+
+Index vectors at half precision for smaller indexes and faster build times
+
+```sql
+CREATE INDEX ON items USING hnsw ((embedding::halfvec(3)) halfvec_l2_ops);
+```
+
+Get the nearest neighbors
+
+```sql
+SELECT * FROM items ORDER BY embedding::halfvec(3) <-> '[1,2,3]' LIMIT 5;
+```
+
+## Binary Vectors
+
+Use the `bit` type to store binary vectors ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/hash_image_search.py))
+
+```sql
+CREATE TABLE items (id bigserial PRIMARY KEY, embedding bit(3));
+INSERT INTO items (embedding) VALUES ('000'), ('111');
+```
+
+Get the nearest neighbors by Hamming distance
+
+```sql
+SELECT * FROM items ORDER BY bit_count(embedding # '101') LIMIT 5;
+```
+
+Or (unreleased)
+
+```sql
+SELECT * FROM items ORDER BY embedding <~> '101' LIMIT 5;
+```
+
+Also supports Jaccard distance (`<%>`)
+
+## Binary Quantization
+
+*Unreleased*
+
+Use expression indexing for binary quantization
+
+```sql
+CREATE INDEX ON items USING hnsw ((quantize_binary(embedding)::bit(3)) bit_hamming_ops);
+```
+
+Get the nearest neighbors by Hamming distance
+
+```sql
+SELECT * FROM items ORDER BY quantize_binary(embedding)::bit(3) <~> quantize_binary('[1,-2,3]') LIMIT 5;
+```
+
+Re-rank by the original vectors for better recall
+
+```sql
+SELECT * FROM (
+    SELECT * FROM items ORDER BY quantize_binary(embedding)::bit(3) <~> quantize_binary('[1,-2,3]') LIMIT 20
+) ORDER BY embedding <=> '[1,-2,3]' LIMIT 5;
+```
+
+## Sparse Vectors
+
+*Unreleased*
+
+Use the `sparsevec` type to store sparse vectors
+
+```sql
+CREATE TABLE items (id bigserial PRIMARY KEY, embedding sparsevec(5));
+```
+
+Insert vectors
+
+```sql
+INSERT INTO items (embedding) VALUES ('{1:1,3:2,5:3}/5'), ('{1:4,3:5,5:6}/5');
+```
+
+Note: The format is `{index1:value1,index2:value2,...}/dimensions` and indices start at 1 like SQL arrays
+
+Get the nearest neighbors by L2 distance
+
+```sql
+SELECT * FROM items ORDER BY embedding <-> '{1:3,3:1,5:2}/5' LIMIT 5;
+```
+
 ## Hybrid Search
 
 Use together with Postgres [full-text search](https://www.postgresql.org/docs/current/textsearch-intro.html) for hybrid search.
@@ -635,18 +732,6 @@ and query with:
 SELECT * FROM items ORDER BY embedding::vector(3) <-> '[3,1,2]' LIMIT 5;
 ```
 
-#### Are binary vectors supported?
-
-You can store binary vectors and perform exact nearest neighbor search by Hamming distance in Postgres without an extension ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/hash_image_search.py)).
-
-```tsql
-CREATE TABLE items (id bigserial PRIMARY KEY, embedding bit(3));
-INSERT INTO items (embedding) VALUES (B'000'), (B'111');
-SELECT * FROM items ORDER BY bit_count(embedding # B'101') LIMIT 5;
-```
-
-Indexing is not currently supported.
-
 #### Do indexes need to fit into memory?
 
 No, but like other index types, you’ll likely see better performance if they do. You can get the size of an index with:
@@ -801,7 +886,7 @@ jaccard_distance(bit, bit) → double precision | Jaccard distance | unreleased
 
 ### Sparsevec Type
 
-Each sparse vector takes `8 * non-zero elements + 16` bytes of storage. Each element is a single-precision floating-point number, and all elements must be finite (no `NaN`, `Infinity` or `-Infinity`). Sparse vectors can have up to 100,000 dimensions.
+Each sparse vector takes `8 * non-zero elements + 16` bytes of storage. Each element is a single-precision floating-point number, and all elements must be finite (no `NaN`, `Infinity` or `-Infinity`). Sparse vectors can have up to 16,000 non-zero elements.
 
 ### Sparsevec Operators
 

sql/vector--0.6.2--0.7.0.sql

@@ -203,9 +203,6 @@ CREATE FUNCTION cosine_distance(sparsevec, sparsevec) RETURNS float8
 CREATE FUNCTION sparsevec_norm(sparsevec) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION subvector(sparsevec, int, int) RETURNS sparsevec
-	AS 'MODULE_PATHNAME', 'sparsevec_subvector' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
 CREATE FUNCTION sparsevec_l2_squared_distance(sparsevec, sparsevec) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 

sql/vector.sql

@@ -512,9 +512,6 @@ CREATE FUNCTION cosine_distance(sparsevec, sparsevec) RETURNS float8
 CREATE FUNCTION sparsevec_norm(sparsevec) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION subvector(sparsevec, int, int) RETURNS sparsevec
-	AS 'MODULE_PATHNAME', 'sparsevec_subvector' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
 -- sparsevec private functions
 
 CREATE FUNCTION sparsevec_l2_squared_distance(sparsevec, sparsevec) RETURNS float8

src/halfvec.c

@@ -22,6 +22,10 @@
 #define TYPALIGN_INT 'i'
 #endif
 
+#ifdef F16C_SUPPORT
+#include <immintrin.h>
+#endif
+
 /*
  * Check if half is NaN
  */
@@ -99,7 +103,10 @@ pq_sendhalf(StringInfo buf, half h)
 float
 HalfToFloat4(half num)
 {
-#ifdef FLT16_SUPPORT
+#if defined(F16C_SUPPORT) && !defined(_MSC_VER)
+	/* TODO Use instrinsics for Windows */
+	return _cvtsh_ss(num);
+#elif defined(FLT16_SUPPORT)
 	return (float) num;
 #else
 	/* TODO Improve performance */
@@ -130,7 +137,7 @@ HalfToFloat4(half num)
 	/* Sign */
 	result = (bin & 0x8000) << 16;
 
-	if (exponent == 31)
+	if (unlikely(exponent == 31))
 	{
 		if (mantissa == 0)
 		{
@@ -141,10 +148,9 @@ HalfToFloat4(half num)
 		{
 			/* NaN */
 			result |= 0x7FC00000;
-			result |= mantissa << 13;
 		}
 	}
-	else if (exponent == 0)
+	else if (unlikely(exponent == 0))
 	{
 		/* Subnormal */
 		if (mantissa != 0)
@@ -164,16 +170,16 @@ HalfToFloat4(half num)
 			}
 
 			result |= (exponent + 127) << 23;
-			result |= mantissa << 13;
 		}
 	}
 	else
 	{
 		/* Normal */
 		result |= (exponent - 15 + 127) << 23;
-		result |= mantissa << 13;
 	}
 
+	result |= mantissa << 13;
+
 	swapfloat.i = result;
 	return swapfloat.f;
 #endif
@@ -185,7 +191,10 @@ HalfToFloat4(half num)
 half
 Float4ToHalfUnchecked(float num)
 {
-#ifdef FLT16_SUPPORT
+#if defined(F16C_SUPPORT) && !defined(_MSC_VER)
+	/* TODO Use instrinsics for Windows */
+	return _cvtss_sh(num, 0);
+#elif defined(FLT16_SUPPORT)
 	return (_Float16) num;
 #else
 	/* TODO Improve performance */
@@ -435,6 +444,8 @@ halfvec_in(PG_FUNCTION_ARGS)
 
 	while (pt != NULL && *stringEnd != ']')
 	{
+		float		val;
+
 		if (dim == HALFVEC_MAX_DIM)
 			ereport(ERROR,
 					(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
@@ -450,15 +461,24 @@ halfvec_in(PG_FUNCTION_ARGS)
 					 errmsg("invalid input syntax for type halfvec: \"%s\"", lit)));
 
 		/* Use strtof like float4in to avoid a double-rounding problem */
-		x[dim] = Float4ToHalf(strtof(pt, &stringEnd));
-		CheckElement(x[dim]);
-		dim++;
+		errno = 0;
+		val = strtof(pt, &stringEnd);
 
 		if (stringEnd == pt)
 			ereport(ERROR,
 					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
 					 errmsg("invalid input syntax for type halfvec: \"%s\"", lit)));
 
+		x[dim] = Float4ToHalfUnchecked(val);
+
+		if ((errno == ERANGE && (isinf(val) || val == 0)) || (HalfIsInf(x[dim]) && !isinf(val)) || (HalfIsZero(x[dim]) && val != 0))
+			ereport(ERROR,
+					(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
+					 errmsg("\"%s\" is out of range for type halfvec", pt)));
+
+		CheckElement(x[dim]);
+		dim++;
+
 		while (halfvec_isspace(*stringEnd))
 			stringEnd++;
 
@@ -782,6 +802,56 @@ vector_to_halfvec(PG_FUNCTION_ARGS)
 	PG_RETURN_POINTER(result);
 }
 
+/*
+ * Get the L2 squared distance between half vectors
+ */
+static double
+l2_distance_squared_internal(HalfVector * a, HalfVector * b)
+{
+	half	   *ax = a->x;
+	half	   *bx = b->x;
+	float		distance = 0.0;
+
+#if defined(F16C_SUPPORT) && defined(__FMA__)
+	int			i;
+	float		s[8];
+	int			count = (a->dim / 8) * 8;
+	__m256		dist = _mm256_setzero_ps();
+
+	for (i = 0; i < count; i += 8)
+	{
+		__m128i		axi = _mm_loadu_si128((__m128i *) (ax + i));
+		__m128i		bxi = _mm_loadu_si128((__m128i *) (bx + i));
+		__m256		axs = _mm256_cvtph_ps(axi);
+		__m256		bxs = _mm256_cvtph_ps(bxi);
+		__m256		diff = _mm256_sub_ps(axs, bxs);
+
+		dist = _mm256_fmadd_ps(diff, diff, dist);
+	}
+
+	_mm256_storeu_ps(s, dist);
+
+	distance = s[0] + s[1] + s[2] + s[3] + s[4] + s[5] + s[6] + s[7];
+
+	for (; i < a->dim; i++)
+	{
+		float		diff = HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]);
+
+		distance += diff * diff;
+	}
+#else
+	/* Auto-vectorized */
+	for (int i = 0; i < a->dim; i++)
+	{
+		float		diff = HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]);
+
+		distance += diff * diff;
+	}
+#endif
+
+	return (double) distance;
+}
+
 /*
  * Get the L2 distance between half vectors
  */
@@ -791,21 +861,10 @@ halfvec_l2_distance(PG_FUNCTION_ARGS)
 {
 	HalfVector *a = PG_GETARG_HALFVEC_P(0);
 	HalfVector *b = PG_GETARG_HALFVEC_P(1);
-	half	   *ax = a->x;
-	half	   *bx = b->x;
-	float		distance = 0.0;
 
 	CheckDims(a, b);
 
-	/* Auto-vectorized */
-	for (int i = 0; i < a->dim; i++)
-	{
-		float		diff = HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]);
-
-		distance += diff * diff;
-	}
-
-	PG_RETURN_FLOAT8(sqrt((double) distance));
+	PG_RETURN_FLOAT8(sqrt(l2_distance_squared_internal(a, b)));
 }
 
 /*
@@ -817,21 +876,51 @@ halfvec_l2_squared_distance(PG_FUNCTION_ARGS)
 {
 	HalfVector *a = PG_GETARG_HALFVEC_P(0);
 	HalfVector *b = PG_GETARG_HALFVEC_P(1);
+
+	CheckDims(a, b);
+
+	PG_RETURN_FLOAT8(l2_distance_squared_internal(a, b));
+}
+
+/*
+ * Get the inner product of two half vectors
+ */
+static double
+inner_product_internal(HalfVector * a, HalfVector * b)
+{
 	half	   *ax = a->x;
 	half	   *bx = b->x;
 	float		distance = 0.0;
 
-	CheckDims(a, b);
+#if defined(F16C_SUPPORT) && defined(__FMA__)
+	int			i;
+	float		s[8];
+	int			count = (a->dim / 8) * 8;
+	__m256		dist = _mm256_setzero_ps();
 
-	/* Auto-vectorized */
-	for (int i = 0; i < a->dim; i++)
+	for (i = 0; i < count; i += 8)
 	{
-		float		diff = HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]);
+		__m128i		axi = _mm_loadu_si128((__m128i *) (ax + i));
+		__m128i		bxi = _mm_loadu_si128((__m128i *) (bx + i));
+		__m256		axs = _mm256_cvtph_ps(axi);
+		__m256		bxs = _mm256_cvtph_ps(bxi);
 
-		distance += diff * diff;
+		dist = _mm256_fmadd_ps(axs, bxs, dist);
 	}
 
-	PG_RETURN_FLOAT8((double) distance);
+	_mm256_storeu_ps(s, dist);
+
+	distance = s[0] + s[1] + s[2] + s[3] + s[4] + s[5] + s[6] + s[7];
+
+	for (; i < a->dim; i++)
+		distance += HalfToFloat4(ax[i]) * HalfToFloat4(bx[i]);
+#else
+	/* Auto-vectorized */
+	for (int i = 0; i < a->dim; i++)
+		distance += HalfToFloat4(ax[i]) * HalfToFloat4(bx[i]);
+#endif
+
+	return (double) distance;
 }
 
 /*
@@ -843,17 +932,10 @@ halfvec_inner_product(PG_FUNCTION_ARGS)
 {
 	HalfVector *a = PG_GETARG_HALFVEC_P(0);
 	HalfVector *b = PG_GETARG_HALFVEC_P(1);
-	half	   *ax = a->x;
-	half	   *bx = b->x;
-	float		distance = 0.0;
 
 	CheckDims(a, b);
 
-	/* Auto-vectorized */
-	for (int i = 0; i < a->dim; i++)
-		distance += HalfToFloat4(ax[i]) * HalfToFloat4(bx[i]);
-
-	PG_RETURN_FLOAT8((double) distance);
+	PG_RETURN_FLOAT8(inner_product_internal(a, b));
 }
 
 /*
@@ -865,17 +947,10 @@ halfvec_negative_inner_product(PG_FUNCTION_ARGS)
 {
 	HalfVector *a = PG_GETARG_HALFVEC_P(0);
 	HalfVector *b = PG_GETARG_HALFVEC_P(1);
-	half	   *ax = a->x;
-	half	   *bx = b->x;
-	float		distance = 0.0;
 
 	CheckDims(a, b);
 
-	/* Auto-vectorized */
-	for (int i = 0; i < a->dim; i++)
-		distance += HalfToFloat4(ax[i]) * HalfToFloat4(bx[i]);
-
-	PG_RETURN_FLOAT8((double) distance * -1);
+	PG_RETURN_FLOAT8(-inner_product_internal(a, b));
 }
 
 /*

src/halfvec.h

@@ -7,7 +7,10 @@
 
 #include "vector.h"
 
-#ifdef __FLT16_MAX__
+/* F16C has better performance than _Float16 (on x86-64) */
+#if defined(__F16C__) || defined(_MSC_VER)
+#define F16C_SUPPORT
+#elif defined(__FLT16_MAX__)
 #define FLT16_SUPPORT
 #endif
 
@@ -15,7 +18,6 @@
 #define half _Float16
 #define HALF_MAX FLT16_MAX
 #else
-/* TODO #pragma message("")? */
 #define half uint16
 #define HALF_MAX 65504
 #endif

src/hnswscan.c

@@ -158,6 +158,10 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
 		UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
 
 		so->first = false;
+
+#if defined(HNSW_MEMORY) && PG_VERSION_NUM >= 130000
+		elog(INFO, "memory: %zu MB", MemoryContextMemAllocated(so->tmpCtx, false) / (1024 * 1024));
+#endif
 	}
 
 	while (list_length(so->w) > 0)

src/hnswutils.c

@@ -163,9 +163,9 @@ HnswGetType(Relation index)
 	Oid			typid = TupleDescAttr(index->rd_att, 0)->atttypid;
 	HeapTuple	tuple;
 	Form_pg_type type;
-	int			result;
+	HnswType	result;
 
-	if (typid == BITOID || typid == VARBITOID)
+	if (typid == BITOID)
 		return HNSW_TYPE_BIT;
 
 	tuple = SearchSysCache1(TYPEOID, ObjectIdGetDatum(typid));
@@ -180,7 +180,10 @@ HnswGetType(Relation index)
 	else if (strcmp(NameStr(type->typname), "sparsevec") == 0)
 		result = HNSW_TYPE_SPARSEVEC;
 	else
-		elog(ERROR, "Unsupported type");
+	{
+		ReleaseSysCache(tuple);
+		elog(ERROR, "type not supported for hnsw index");
+	}
 
 	ReleaseSysCache(tuple);
 

src/ivfbuild.c

@@ -57,7 +57,7 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 	 */
 	if (buildstate->kmeansnormprocinfo != NULL)
 	{
-		if (!IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->collation, &value))
+		if (!IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->collation, &value, buildstate->type))
 			return;
 	}
 
@@ -153,7 +153,7 @@ AddTupleToSort(Relation index, ItemPointer tid, Datum *values, IvfflatBuildState
 	/* Normalize if needed */
 	if (buildstate->normprocinfo != NULL)
 	{
-		if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value))
+		if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->type))
 			return;
 	}
 
@@ -312,25 +312,39 @@ InsertTuples(Relation index, IvfflatBuildState * buildstate, ForkNumber forkNum)
 	}
 }
 
+/*
+ * Get max dimensions
+ */
+static int
+GetMaxDimensions(IvfflatType type)
+{
+	return IVFFLAT_MAX_DIM;
+}
+
 /*
  * Initialize the build state
  */
 static void
 InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, IndexInfo *indexInfo)
 {
+	int			maxDimensions;
+
 	buildstate->heap = heap;
 	buildstate->index = index;
 	buildstate->indexInfo = indexInfo;
+	buildstate->type = IvfflatGetType(index);
 
 	buildstate->lists = IvfflatGetLists(index);
 	buildstate->dimensions = TupleDescAttr(index->rd_att, 0)->atttypmod;
 
+	maxDimensions = GetMaxDimensions(buildstate->type);
+
 	/* Require column to have dimensions to be indexed */
 	if (buildstate->dimensions < 0)
 		elog(ERROR, "column does not have dimensions");
 
-	if (buildstate->dimensions > IVFFLAT_MAX_DIM)
-		elog(ERROR, "column cannot have more than %d dimensions for ivfflat index", IVFFLAT_MAX_DIM);
+	if (buildstate->dimensions > maxDimensions)
+		elog(ERROR, "column cannot have more than %d dimensions for ivfflat index", maxDimensions);
 
 	buildstate->reltuples = 0;
 	buildstate->indtuples = 0;

src/ivfflat.h

@@ -43,6 +43,11 @@
 #define IVFFLAT_MAX_LISTS		32768
 #define IVFFLAT_DEFAULT_PROBES	1
 
+typedef enum IvfflatType
+{
+	IVFFLAT_TYPE_VECTOR
+}			IvfflatType;
+
 /* Build phases */
 /* PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE is 1 */
 #define PROGRESS_IVFFLAT_PHASE_KMEANS	2
@@ -153,6 +158,7 @@ typedef struct IvfflatBuildState
 	Relation	heap;
 	Relation	index;
 	IndexInfo  *indexInfo;
+	IvfflatType type;
 
 	/* Settings */
 	int			dimensions;
@@ -266,7 +272,8 @@ void		VectorArrayFree(VectorArray arr);
 void		PrintVectorArray(char *msg, VectorArray arr);
 void		IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers);
 FmgrInfo   *IvfflatOptionalProcInfo(Relation index, uint16 procnum);
-bool		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value);
+IvfflatType IvfflatGetType(Relation index);
+bool		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, IvfflatType type);
 int			IvfflatGetLists(Relation index);
 void		IvfflatGetMetaPageInfo(Relation index, int *lists, int *dimensions);
 void		IvfflatUpdateList(Relation index, ListInfo listInfo, BlockNumber insertPage, BlockNumber originalInsertPage, BlockNumber startPage, ForkNumber forkNum);

src/ivfinsert.c

@@ -85,7 +85,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
 	normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
 	if (normprocinfo != NULL)
 	{
-		if (!IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value))
+		if (!IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value, IvfflatGetType(index)))
 			return;
 	}
 

src/ivfscan.c

@@ -268,6 +268,7 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 	if (so->first)
 	{
 		Datum		value;
+		IvfflatType type = IvfflatGetType(scan->indexRelation);
 
 		/* Count index scan for stats */
 		pgstat_count_index_scan(scan->indexRelation);
@@ -282,7 +283,12 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 			elog(ERROR, "non-MVCC snapshots are not supported with ivfflat");
 
 		if (scan->orderByData->sk_flags & SK_ISNULL)
+		{
+			if (type == IVFFLAT_TYPE_VECTOR)
 				value = PointerGetDatum(InitVector(so->dimensions));
+			else
+				elog(ERROR, "Unsupported type");
+		}
 		else
 		{
 			value = scan->orderByData->sk_argument;
@@ -293,7 +299,7 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 
 			/* Fine if normalization fails */
 			if (so->normprocinfo != NULL)
-				IvfflatNormValue(so->normprocinfo, so->collation, &value);
+				IvfflatNormValue(so->normprocinfo, so->collation, &value, type);
 		}
 
 		IvfflatBench("GetScanLists", GetScanLists(scan, value));

src/ivfutils.c

@@ -66,6 +66,15 @@ IvfflatOptionalProcInfo(Relation index, uint16 procnum)
 	return index_getprocinfo(index, 1, procnum);
 }
 
+/*
+ * Get type
+ */
+IvfflatType
+IvfflatGetType(Relation index)
+{
+	return IVFFLAT_TYPE_VECTOR;
+}
+
 /*
  * Divide by the norm
  *
@@ -75,11 +84,13 @@ IvfflatOptionalProcInfo(Relation index, uint16 procnum)
  * if it's different than the original value
  */
 bool
-IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value)
+IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, IvfflatType type)
 {
 	double		norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
 
 	if (norm > 0)
+	{
+		if (type == IVFFLAT_TYPE_VECTOR)
 		{
 			Vector	   *v = DatumGetVector(*value);
 			Vector	   *result = InitVector(v->dim);
@@ -88,6 +99,9 @@ IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value)
 				result->x[i] = v->x[i] / norm;
 
 			*value = PointerGetDatum(result);
+		}
+		else
+			elog(ERROR, "Unsupported type");
 
 		return true;
 	}

src/sparsevec.c

@@ -68,7 +68,12 @@ CheckNnz(int nnz, int dim)
 	if (nnz < 0)
 		ereport(ERROR,
 				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("sparsevec must have at least one element")));
+				 errmsg("sparsevec cannot have negative number of elements")));
+
+	if (nnz > SPARSEVEC_MAX_NNZ)
+		ereport(ERROR,
+				(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
+				 errmsg("sparsevec cannot have more than %d non-zero elements", SPARSEVEC_MAX_NNZ)));
 
 	if (nnz > dim)
 		ereport(ERROR,
@@ -84,15 +89,15 @@ CheckIndex(int32 *indices, int i, int dim)
 {
 	int32		index = indices[i];
 
-	if (index < 0)
+	if (index < 1)
 		ereport(ERROR,
 				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("index must not be negative")));
+				 errmsg("index must be greater than zero")));
 
-	if (index >= dim)
+	if (index > dim)
 		ereport(ERROR,
 				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("index must be less than dimensions")));
+				 errmsg("index must be less than or equal to dimensions")));
 
 	if (i > 0)
 	{
@@ -190,6 +195,11 @@ sparsevec_in(PG_FUNCTION_ARGS)
 		pt++;
 	}
 
+	if (maxNnz > SPARSEVEC_MAX_NNZ)
+		ereport(ERROR,
+				(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
+				 errmsg("sparsevec cannot have more than %d non-zero elements", SPARSEVEC_MAX_NNZ)));
+
 	indices = palloc(maxNnz * sizeof(int32));
 	values = palloc(maxNnz * sizeof(float));
 
@@ -235,7 +245,7 @@ sparsevec_in(PG_FUNCTION_ARGS)
 					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
 					 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
 
-		if (errno == ERANGE || index < 0 || index > INT_MAX)
+		if (errno == ERANGE || index < 1 || index > INT_MAX)
 			ereport(ERROR,
 					(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
 					 errmsg("index \"%ld\" is out of range for type sparsevec", index)));
@@ -560,7 +570,7 @@ vector_to_sparsevec(PG_FUNCTION_ARGS)
 			if (j == nnz)
 				elog(ERROR, "safety check failed");
 
-			result->indices[j] = i;
+			result->indices[j] = i + 1;
 			values[j] = vec->x[i];
 			j++;
 		}
@@ -776,55 +786,3 @@ sparsevec_norm(PG_FUNCTION_ARGS)
 
 	PG_RETURN_FLOAT8(sqrt(norm));
 }
-
-/*
- * Get a subvector
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_subvector);
-Datum
-sparsevec_subvector(PG_FUNCTION_ARGS)
-{
-	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
-	int32		start = PG_GETARG_INT32(1);
-	int32		count = PG_GETARG_INT32(2);
-	int32		end = start + count;
-	float	   *ax = SPARSEVEC_VALUES(a);
-	SparseVector *result;
-	float	   *rx;
-	int			dim;
-	int			nnz = 0;
-	int			startIndex;
-
-	/* Indexing starts at 1, like substring */
-	if (start < 1)
-		start = 1;
-
-	if (end > a->dim)
-		end = a->dim + 1;
-
-	dim = end - start;
-	CheckDim(dim);
-
-	for (startIndex = 0; startIndex < a->nnz; startIndex++)
-	{
-		if (a->indices[startIndex] >= start - 1)
-			break;
-	}
-
-	for (int i = startIndex; i < a->nnz; i++)
-	{
-		if (a->indices[i] < end - 1)
-			nnz++;
-	}
-
-	result = InitSparseVector(dim, nnz);
-	rx = SPARSEVEC_VALUES(result);
-
-	for (int i = 0; i < nnz; i++)
-	{
-		result->indices[i] = a->indices[startIndex + i];
-		rx[i] = ax[startIndex + i];
-	}
-
-	PG_RETURN_POINTER(result);
-}

src/sparsevec.h

@@ -2,6 +2,7 @@
 #define SPARSEVEC_H
 
 #define SPARSEVEC_MAX_DIM 100000
+#define SPARSEVEC_MAX_NNZ 16000
 
 /* Ensure values are aligned */
 #define SPARSEVEC_SIZE(_nnz)		(offsetof(SparseVector, indices) + MAXALIGN((_nnz) * sizeof(int32)) + (_nnz * sizeof(float)))

src/vector.c

@@ -1244,7 +1244,7 @@ sparsevec_to_vector(PG_FUNCTION_ARGS)
 
 	result = InitVector(dim);
 	for (int i = 0; i < svec->nnz; i++)
-		result->x[svec->indices[i]] = values[i];
+		result->x[svec->indices[i] - 1] = values[i];
 
 	PG_RETURN_POINTER(result);
 }

test/expected/bit_functions.out

@@ -1,64 +1,104 @@
-SELECT hamming_distance(B'111', B'111');
+SELECT hamming_distance('111', '111');
  hamming_distance 
 ------------------
                 0
 (1 row)
 
-SELECT hamming_distance(B'111', B'110');
+SELECT hamming_distance('111', '110');
  hamming_distance 
 ------------------
                 1
 (1 row)
 
-SELECT hamming_distance(B'111', B'100');
+SELECT hamming_distance('111', '100');
  hamming_distance 
 ------------------
                 2
 (1 row)
 
-SELECT hamming_distance(B'111', B'000');
+SELECT hamming_distance('111', '000');
  hamming_distance 
 ------------------
                 3
 (1 row)
 
-SELECT hamming_distance(B'111', B'00');
+SELECT hamming_distance('10101010101010101010', '01010101010101010101');
+ hamming_distance 
+------------------
+               20
+(1 row)
+
+SELECT hamming_distance('', '');
+ hamming_distance 
+------------------
+                0
+(1 row)
+
+SELECT hamming_distance('111', '00');
 ERROR:  different bit lengths 3 and 2
-SELECT jaccard_distance(B'1111', B'1111');
+SELECT hamming_distance('111', '000'::varbit(4));
+ hamming_distance 
+------------------
+                3
+(1 row)
+
+SELECT hamming_distance('111', '0000'::varbit(4));
+ERROR:  different bit lengths 3 and 4
+SELECT jaccard_distance('1111', '1111');
  jaccard_distance 
 ------------------
                 0
 (1 row)
 
-SELECT jaccard_distance(B'1111', B'1110');
+SELECT jaccard_distance('1111', '1110');
  jaccard_distance 
 ------------------
              0.25
 (1 row)
 
-SELECT jaccard_distance(B'1111', B'1100');
+SELECT jaccard_distance('1111', '1100');
  jaccard_distance 
 ------------------
               0.5
 (1 row)
 
-SELECT jaccard_distance(B'1111', B'1000');
+SELECT jaccard_distance('1111', '1000');
  jaccard_distance 
 ------------------
              0.75
 (1 row)
 
-SELECT jaccard_distance(B'1111', B'0000');
+SELECT jaccard_distance('1111', '0000');
  jaccard_distance 
 ------------------
                 1
 (1 row)
 
-SELECT jaccard_distance(B'1100', B'1000');
+SELECT jaccard_distance('1100', '1000');
  jaccard_distance 
 ------------------
               0.5
 (1 row)
 
-SELECT jaccard_distance(B'1111', B'000');
+SELECT jaccard_distance('10101010101010101010', '01010101010101010101');
+ jaccard_distance 
+------------------
+                1
+(1 row)
+
+SELECT jaccard_distance('', '');
+ jaccard_distance 
+------------------
+                1
+(1 row)
+
+SELECT jaccard_distance('1111', '000');
 ERROR:  different bit lengths 4 and 3
+SELECT jaccard_distance('1111', '0000'::varbit(5));
+ jaccard_distance 
+------------------
+                1
+(1 row)
+
+SELECT jaccard_distance('1111', '00000'::varbit(5));
+ERROR:  different bit lengths 4 and 5

test/expected/halfvec_input.out

@@ -51,17 +51,21 @@ SELECT '[65519,-65519]'::halfvec;
 (1 row)
 
 SELECT '[65520,-65520]'::halfvec;
-ERROR:  value out of range: overflow
+ERROR:  "65520" is out of range for type halfvec
 LINE 1: SELECT '[65520,-65520]'::halfvec;
                ^
 SELECT '[1e-8,-1e-8]'::halfvec;
-ERROR:  value out of range: underflow
+ERROR:  "1e-8" is out of range for type halfvec
 LINE 1: SELECT '[1e-8,-1e-8]'::halfvec;
                ^
 SELECT '[4e38,1]'::halfvec;
-ERROR:  infinite value not allowed in halfvec
+ERROR:  "4e38" is out of range for type halfvec
 LINE 1: SELECT '[4e38,1]'::halfvec;
                ^
+SELECT '[1e-46,1]'::halfvec;
+ERROR:  "1e-46" is out of range for type halfvec
+LINE 1: SELECT '[1e-46,1]'::halfvec;
+               ^
 SELECT '[1,2,3'::halfvec;
 ERROR:  malformed halfvec literal: "[1,2,3"
 LINE 1: SELECT '[1,2,3'::halfvec;

test/expected/hnsw_bit_hamming.out

@@ -19,3 +19,9 @@ SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <~> (SELECT NULL::bit)) t2;
 (1 row)
 
 DROP TABLE t;
+-- TODO move
+CREATE TABLE t (val varbit(3));
+CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
+ERROR:  type not supported for hnsw index
+CREATE INDEX ON t USING hnsw ((val::bit(3)) bit_hamming_ops);
+DROP TABLE t;

test/expected/hnsw_sparsevec_cosine.out

@@ -1,14 +1,14 @@
 SET enable_seqscan = off;
 CREATE TABLE t (val sparsevec(3));
-INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
+INSERT INTO t (val) VALUES ('{}/3'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
 CREATE INDEX ON t USING hnsw (val sparsevec_cosine_ops);
-INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
-SELECT * FROM t ORDER BY val <=> '{0:3,1:3,2:3}/3';
+INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
+SELECT * FROM t ORDER BY val <=> '{1:3,2:3,3:3}/3';
        val       
 -----------------
- {0:1,1:1,2:1}/3
- {0:1,1:2,2:3}/3
- {0:1,1:2,2:4}/3
+ {1:1,2:1,3:1}/3
+ {1:1,2:2,3:3}/3
+ {1:1,2:2,3:4}/3
 (3 rows)
 
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '{}/3') t2;

test/expected/hnsw_sparsevec_ip.out

@@ -1,14 +1,14 @@
 SET enable_seqscan = off;
 CREATE TABLE t (val sparsevec(3));
-INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
+INSERT INTO t (val) VALUES ('{}/3'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
 CREATE INDEX ON t USING hnsw (val sparsevec_ip_ops);
-INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
-SELECT * FROM t ORDER BY val <#> '{0:3,1:3,2:3}/3';
+INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
+SELECT * FROM t ORDER BY val <#> '{1:3,2:3,3:3}/3';
        val       
 -----------------
- {0:1,1:2,2:4}/3
- {0:1,1:2,2:3}/3
- {0:1,1:1,2:1}/3
+ {1:1,2:2,3:4}/3
+ {1:1,2:2,3:3}/3
+ {1:1,2:1,3:1}/3
  {}/3
 (4 rows)
 

test/expected/hnsw_sparsevec_l2.out

@@ -1,14 +1,14 @@
 SET enable_seqscan = off;
 CREATE TABLE t (val sparsevec(3));
-INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
+INSERT INTO t (val) VALUES ('{}/3'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
 CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
-INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
-SELECT * FROM t ORDER BY val <-> '{0:3,1:3,2:3}/3';
+INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
+SELECT * FROM t ORDER BY val <-> '{1:3,2:3,3:3}/3';
        val       
 -----------------
- {0:1,1:2,2:3}/3
- {0:1,1:2,2:4}/3
- {0:1,1:1,2:1}/3
+ {1:1,2:2,3:3}/3
+ {1:1,2:2,3:4}/3
+ {1:1,2:1,3:1}/3
  {}/3
 (4 rows)
 
@@ -25,7 +25,7 @@ SELECT COUNT(*) FROM t;
 (1 row)
 
 TRUNCATE t;
-SELECT * FROM t ORDER BY val <-> '{0:3,1:3,2:3}/3';
+SELECT * FROM t ORDER BY val <-> '{1:3,2:3,3:3}/3';
  val 
 -----
 (0 rows)

test/expected/sparsevec_functions.out

@@ -1,52 +1,52 @@
-SELECT l2_distance('{}/2'::sparsevec, '{0:3,1:4}/2');
+SELECT l2_distance('{}/2'::sparsevec, '{1:3,2:4}/2');
  l2_distance 
 -------------
            5
 (1 row)
 
-SELECT l2_distance('{}/2'::sparsevec, '{1:1}/2');
+SELECT l2_distance('{}/2'::sparsevec, '{2:1}/2');
  l2_distance 
 -------------
            1
 (1 row)
 
-SELECT '{}/2'::sparsevec <-> '{0:3,1:4}/2';
+SELECT '{}/2'::sparsevec <-> '{1:3,2:4}/2';
  ?column? 
 ----------
         5
 (1 row)
 
-SELECT inner_product('{0:1,1:2}/2'::sparsevec, '{0:2,1:4}/2');
+SELECT inner_product('{1:1,2:2}/2'::sparsevec, '{1:2,2:4}/2');
  inner_product 
 ---------------
             10
 (1 row)
 
-SELECT sparsevec_negative_inner_product('{0:1,1:2}/2', '{0:2,1:4}/2');
+SELECT sparsevec_negative_inner_product('{1:1,2:2}/2', '{1:2,2:4}/2');
  sparsevec_negative_inner_product 
 ----------------------------------
                               -10
 (1 row)
 
-SELECT cosine_distance('{0:1,1:2}/2'::sparsevec, '{0:2,1:4}/2');
+SELECT cosine_distance('{1:1,2:2}/2'::sparsevec, '{1:2,2:4}/2');
  cosine_distance 
 -----------------
                0
 (1 row)
 
-SELECT cosine_distance('{0:1,1:2}/2'::sparsevec, '{}/2');
+SELECT cosine_distance('{1:1,2:2}/2'::sparsevec, '{}/2');
  cosine_distance 
 -----------------
              NaN
 (1 row)
 
-SELECT cosine_distance('{0:1,1:1}/2'::sparsevec, '{0:-1,1:-1}/2');
+SELECT cosine_distance('{1:1,2:1}/2'::sparsevec, '{1:-1,2:-1}/2');
  cosine_distance 
 -----------------
                2
 (1 row)
 
-SELECT cosine_distance('{0:1}/2'::sparsevec, '{1:2}/2');
+SELECT cosine_distance('{1:2}/2'::sparsevec, '{2:2}/2');
  cosine_distance 
 -----------------
                1
@@ -58,35 +58,5 @@ SELECT cosine_distance('{}/1'::sparsevec, '{}/1');
              NaN
 (1 row)
 
-SELECT cosine_distance('{0:1}/2'::sparsevec, '{0:1}/3');
+SELECT cosine_distance('{1:2}/2'::sparsevec, '{1:1}/3');
 ERROR:  different sparsevec dimensions 2 and 3
-SELECT subvector('{0:1,1:2,2:3,3:4,4:5}/5'::sparsevec, 1, 3);
-    subvector    
------------------
- {0:1,1:2,2:3}/3
-(1 row)
-
-SELECT subvector('{0:1,1:2,2:3,3:4,4:5}/5'::sparsevec, 3, 2);
-  subvector  
--------------
- {2:3,3:4}/2
-(1 row)
-
-SELECT subvector('{0:1,1:2,2:3,3:4,4:5}/5'::sparsevec, -1, 3);
- subvector 
------------
- {0:1}/1
-(1 row)
-
-SELECT subvector('{0:1,1:2,2:3,3:4,4:5}/5'::sparsevec, 3, 9);
-    subvector    
------------------
- {2:3,3:4,4:5}/3
-(1 row)
-
-SELECT subvector('{0:1,1:2,2:3,3:4,4:5}/5'::sparsevec, 1, 0);
-ERROR:  sparsevec must have at least 1 dimension
-SELECT subvector('{0:1,1:2,2:3,3:4,4:5}/5'::sparsevec, 3, -1);
-ERROR:  sparsevec must have at least 1 dimension
-SELECT subvector('{0:1,1:2,2:3,3:4,4:5}/5'::sparsevec, -1, 2);
-ERROR:  sparsevec must have at least 1 dimension

test/expected/sparsevec_input.out

@@ -1,38 +1,38 @@
-SELECT '{0:1.5,2:3.5}/5'::sparsevec;
-    sparsevec    
------------------
- {0:1.5,2:3.5}/5
-(1 row)
-
-SELECT '{0:1.5,2:3.5}/5'::sparsevec::vector;
-     vector      
------------------
- [1.5,0,3.5,0,0]
-(1 row)
-
-SELECT '{0:1.5,2:3.5}/5'::sparsevec::vector(5);
-     vector      
------------------
- [1.5,0,3.5,0,0]
-(1 row)
-
-SELECT '{0:1.5,2:3.5}/5'::sparsevec::vector(4);
-ERROR:  expected 4 dimensions, not 5
-SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec;
+SELECT '{1:1.5,3:3.5}/5'::sparsevec;
     sparsevec    
 -----------------
  {1:1.5,3:3.5}/5
 (1 row)
 
-SELECT '{0:0,1:1,2:0}/3'::sparsevec;
- sparsevec 
------------
- {1:1}/3
+SELECT '{1:1.5,3:3.5}/5'::sparsevec::vector;
+     vector      
+-----------------
+ [1.5,0,3.5,0,0]
 (1 row)
 
-SELECT '{1:1,0:1}/2'::sparsevec;
+SELECT '{1:1.5,3:3.5}/5'::sparsevec::vector(5);
+     vector      
+-----------------
+ [1.5,0,3.5,0,0]
+(1 row)
+
+SELECT '{1:1.5,3:3.5}/5'::sparsevec::vector(4);
+ERROR:  expected 4 dimensions, not 5
+SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec;
+    sparsevec    
+-----------------
+ {2:1.5,4:3.5}/5
+(1 row)
+
+SELECT '{1:0,2:1,3:0}/3'::sparsevec;
+ sparsevec 
+-----------
+ {2:1}/3
+(1 row)
+
+SELECT '{2:1,1:1}/2'::sparsevec;
 ERROR:  indexes must be in ascending order
-LINE 1: SELECT '{1:1,0:1}/2'::sparsevec;
+LINE 1: SELECT '{2:1,1:1}/2'::sparsevec;
                ^
 SELECT '{}/5'::sparsevec;
  sparsevec 
@@ -50,13 +50,13 @@ LINE 1: SELECT '{}/100001'::sparsevec;
                ^
 SELECT '{}/16001'::sparsevec::vector;
 ERROR:  vector cannot have more than 16000 dimensions
-SELECT '{-1:1}/1'::sparsevec;
-ERROR:  index "-1" is out of range for type sparsevec
-LINE 1: SELECT '{-1:1}/1'::sparsevec;
+SELECT '{0:1}/1'::sparsevec;
+ERROR:  index "0" is out of range for type sparsevec
+LINE 1: SELECT '{0:1}/1'::sparsevec;
                ^
-SELECT '{1:1}/1'::sparsevec;
-ERROR:  index must be less than dimensions
-LINE 1: SELECT '{1:1}/1'::sparsevec;
+SELECT '{2:1}/1'::sparsevec;
+ERROR:  index must be less than or equal to dimensions
+LINE 1: SELECT '{2:1}/1'::sparsevec;
                ^
 SELECT '{}/1'::sparsevec(2);
 ERROR:  expected 2 dimensions, not 1

test/sql/bit_functions.sql

@@ -1,13 +1,21 @@
-SELECT hamming_distance(B'111', B'111');
-SELECT hamming_distance(B'111', B'110');
-SELECT hamming_distance(B'111', B'100');
-SELECT hamming_distance(B'111', B'000');
-SELECT hamming_distance(B'111', B'00');
+SELECT hamming_distance('111', '111');
+SELECT hamming_distance('111', '110');
+SELECT hamming_distance('111', '100');
+SELECT hamming_distance('111', '000');
+SELECT hamming_distance('10101010101010101010', '01010101010101010101');
+SELECT hamming_distance('', '');
+SELECT hamming_distance('111', '00');
+SELECT hamming_distance('111', '000'::varbit(4));
+SELECT hamming_distance('111', '0000'::varbit(4));
 
-SELECT jaccard_distance(B'1111', B'1111');
-SELECT jaccard_distance(B'1111', B'1110');
-SELECT jaccard_distance(B'1111', B'1100');
-SELECT jaccard_distance(B'1111', B'1000');
-SELECT jaccard_distance(B'1111', B'0000');
-SELECT jaccard_distance(B'1100', B'1000');
-SELECT jaccard_distance(B'1111', B'000');
+SELECT jaccard_distance('1111', '1111');
+SELECT jaccard_distance('1111', '1110');
+SELECT jaccard_distance('1111', '1100');
+SELECT jaccard_distance('1111', '1000');
+SELECT jaccard_distance('1111', '0000');
+SELECT jaccard_distance('1100', '1000');
+SELECT jaccard_distance('10101010101010101010', '01010101010101010101');
+SELECT jaccard_distance('', '');
+SELECT jaccard_distance('1111', '000');
+SELECT jaccard_distance('1111', '0000'::varbit(5));
+SELECT jaccard_distance('1111', '00000'::varbit(5));

test/sql/halfvec_input.sql

@@ -11,6 +11,7 @@ SELECT '[65519,-65519]'::halfvec;
 SELECT '[65520,-65520]'::halfvec;
 SELECT '[1e-8,-1e-8]'::halfvec;
 SELECT '[4e38,1]'::halfvec;
+SELECT '[1e-46,1]'::halfvec;
 SELECT '[1,2,3'::halfvec;
 SELECT '[1,2,3]9'::halfvec;
 SELECT '1,2,3'::halfvec;

test/sql/hnsw_bit_hamming.sql

@@ -10,3 +10,9 @@ SELECT * FROM t ORDER BY val <~> B'111';
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <~> (SELECT NULL::bit)) t2;
 
 DROP TABLE t;
+
+-- TODO move
+CREATE TABLE t (val varbit(3));
+CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
+CREATE INDEX ON t USING hnsw ((val::bit(3)) bit_hamming_ops);
+DROP TABLE t;

test/sql/hnsw_sparsevec_cosine.sql

@@ -1,12 +1,12 @@
 SET enable_seqscan = off;
 
 CREATE TABLE t (val sparsevec(3));
-INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
+INSERT INTO t (val) VALUES ('{}/3'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
 CREATE INDEX ON t USING hnsw (val sparsevec_cosine_ops);
 
-INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
+INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
 
-SELECT * FROM t ORDER BY val <=> '{0:3,1:3,2:3}/3';
+SELECT * FROM t ORDER BY val <=> '{1:3,2:3,3:3}/3';
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '{}/3') t2;
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::sparsevec)) t2;
 

test/sql/hnsw_sparsevec_ip.sql

@@ -1,12 +1,12 @@
 SET enable_seqscan = off;
 
 CREATE TABLE t (val sparsevec(3));
-INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
+INSERT INTO t (val) VALUES ('{}/3'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
 CREATE INDEX ON t USING hnsw (val sparsevec_ip_ops);
 
-INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
+INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
 
-SELECT * FROM t ORDER BY val <#> '{0:3,1:3,2:3}/3';
+SELECT * FROM t ORDER BY val <#> '{1:3,2:3,3:3}/3';
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::sparsevec)) t2;
 
 DROP TABLE t;

test/sql/hnsw_sparsevec_l2.sql

@@ -1,17 +1,17 @@
 SET enable_seqscan = off;
 
 CREATE TABLE t (val sparsevec(3));
-INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
+INSERT INTO t (val) VALUES ('{}/3'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
 CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
 
-INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
+INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
 
-SELECT * FROM t ORDER BY val <-> '{0:3,1:3,2:3}/3';
+SELECT * FROM t ORDER BY val <-> '{1:3,2:3,3:3}/3';
 SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::sparsevec)) t2;
 SELECT COUNT(*) FROM t;
 
 TRUNCATE t;
-SELECT * FROM t ORDER BY val <-> '{0:3,1:3,2:3}/3';
+SELECT * FROM t ORDER BY val <-> '{1:3,2:3,3:3}/3';
 
 DROP TABLE t;
 

test/sql/sparsevec_functions.sql

@@ -1,21 +1,13 @@
-SELECT l2_distance('{}/2'::sparsevec, '{0:3,1:4}/2');
-SELECT l2_distance('{}/2'::sparsevec, '{1:1}/2');
-SELECT '{}/2'::sparsevec <-> '{0:3,1:4}/2';
+SELECT l2_distance('{}/2'::sparsevec, '{1:3,2:4}/2');
+SELECT l2_distance('{}/2'::sparsevec, '{2:1}/2');
+SELECT '{}/2'::sparsevec <-> '{1:3,2:4}/2';
 
-SELECT inner_product('{0:1,1:2}/2'::sparsevec, '{0:2,1:4}/2');
-SELECT sparsevec_negative_inner_product('{0:1,1:2}/2', '{0:2,1:4}/2');
+SELECT inner_product('{1:1,2:2}/2'::sparsevec, '{1:2,2:4}/2');
+SELECT sparsevec_negative_inner_product('{1:1,2:2}/2', '{1:2,2:4}/2');
 
-SELECT cosine_distance('{0:1,1:2}/2'::sparsevec, '{0:2,1:4}/2');
-SELECT cosine_distance('{0:1,1:2}/2'::sparsevec, '{}/2');
-SELECT cosine_distance('{0:1,1:1}/2'::sparsevec, '{0:-1,1:-1}/2');
-SELECT cosine_distance('{0:1}/2'::sparsevec, '{1:2}/2');
+SELECT cosine_distance('{1:1,2:2}/2'::sparsevec, '{1:2,2:4}/2');
+SELECT cosine_distance('{1:1,2:2}/2'::sparsevec, '{}/2');
+SELECT cosine_distance('{1:1,2:1}/2'::sparsevec, '{1:-1,2:-1}/2');
+SELECT cosine_distance('{1:2}/2'::sparsevec, '{2:2}/2');
 SELECT cosine_distance('{}/1'::sparsevec, '{}/1');
-SELECT cosine_distance('{0:1}/2'::sparsevec, '{0:1}/3');
-
-SELECT subvector('{0:1,1:2,2:3,3:4,4:5}/5'::sparsevec, 1, 3);
-SELECT subvector('{0:1,1:2,2:3,3:4,4:5}/5'::sparsevec, 3, 2);
-SELECT subvector('{0:1,1:2,2:3,3:4,4:5}/5'::sparsevec, -1, 3);
-SELECT subvector('{0:1,1:2,2:3,3:4,4:5}/5'::sparsevec, 3, 9);
-SELECT subvector('{0:1,1:2,2:3,3:4,4:5}/5'::sparsevec, 1, 0);
-SELECT subvector('{0:1,1:2,2:3,3:4,4:5}/5'::sparsevec, 3, -1);
-SELECT subvector('{0:1,1:2,2:3,3:4,4:5}/5'::sparsevec, -1, 2);
+SELECT cosine_distance('{1:2}/2'::sparsevec, '{1:1}/3');

test/sql/sparsevec_input.sql

@@ -1,19 +1,19 @@
-SELECT '{0:1.5,2:3.5}/5'::sparsevec;
-SELECT '{0:1.5,2:3.5}/5'::sparsevec::vector;
-SELECT '{0:1.5,2:3.5}/5'::sparsevec::vector(5);
-SELECT '{0:1.5,2:3.5}/5'::sparsevec::vector(4);
+SELECT '{1:1.5,3:3.5}/5'::sparsevec;
+SELECT '{1:1.5,3:3.5}/5'::sparsevec::vector;
+SELECT '{1:1.5,3:3.5}/5'::sparsevec::vector(5);
+SELECT '{1:1.5,3:3.5}/5'::sparsevec::vector(4);
 SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec;
 
-SELECT '{0:0,1:1,2:0}/3'::sparsevec;
+SELECT '{1:0,2:1,3:0}/3'::sparsevec;
 
-SELECT '{1:1,0:1}/2'::sparsevec;
+SELECT '{2:1,1:1}/2'::sparsevec;
 
 SELECT '{}/5'::sparsevec;
 SELECT '{}/-1'::sparsevec;
 SELECT '{}/100001'::sparsevec;
 SELECT '{}/16001'::sparsevec::vector;
 
-SELECT '{-1:1}/1'::sparsevec;
-SELECT '{1:1}/1'::sparsevec;
+SELECT '{0:1}/1'::sparsevec;
+SELECT '{2:1}/1'::sparsevec;
 
 SELECT '{}/1'::sparsevec(2);

test/t/016_hnsw_vacuum_recall.pl

@@ -86,7 +86,7 @@ foreach (@queries)
 	push(@expected, $res);
 }
 
-test_recall(0.19, $limit, "before vacuum");
+test_recall(0.18, $limit, "before vacuum");
 test_recall(0.95, 100, "before vacuum");
 
 # TODO Test concurrent inserts with vacuum

test/t/020_hnsw_bit_build_recall.pl

@@ -99,7 +99,7 @@ for my $i (0 .. $#operators)
 	));
 
 	# Test approximate results
-	my $min = $operator eq "<\%>" ? 0.96 : 0.98;
+	my $min = $operator eq "<\%>" ? 0.95 : 0.98;
 	test_recall($min, $operator);
 
 	$node->safe_psql("postgres", "DROP INDEX idx;");

test/t/022_hnsw_sparsevec_build_recall.pl

@@ -0,0 +1,128 @@
+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 sparsevec(3));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()]::vector::sparsevec FROM generate_series(1, 10000) i;"
+);
+
+# Generate queries
+for (1 .. 20)
+{
+	my $r1 = rand();
+	my $r2 = rand();
+	my $r3 = rand();
+	push(@queries, "{1:$r1,2:$r2,3:$r3}/3");
+}
+
+# Check each index type
+my @operators = ("<->", "<#>", "<=>");
+my @opclasses = ("sparsevec_l2_ops", "sparsevec_ip_ops", "sparsevec_cosine_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);
+	}
+
+	# Build index serially
+	$node->safe_psql("postgres", qq(
+		SET max_parallel_maintenance_workers = 0;
+		CREATE INDEX idx ON tst USING hnsw (v $opclass);
+	));
+
+	# Test approximate results
+	my $min = $operator eq "<#>" ? 0.80 : 0.99;
+	test_recall($min, $operator);
+
+	$node->safe_psql("postgres", "DROP INDEX idx;");
+
+	# Build index in parallel in memory
+	my ($ret, $stdout, $stderr) = $node->psql("postgres", qq(
+		SET client_min_messages = DEBUG;
+		SET min_parallel_table_scan_size = 1;
+		CREATE INDEX idx ON tst USING hnsw (v $opclass);
+	));
+	is($ret, 0, $stderr);
+	like($stderr, qr/using \d+ parallel workers/);
+
+	# Test approximate results
+	test_recall($min, $operator);
+
+	$node->safe_psql("postgres", "DROP INDEX idx;");
+
+	# Build index in parallel on disk
+	# Set parallel_workers on table to use workers with low maintenance_work_mem
+	($ret, $stdout, $stderr) = $node->psql("postgres", qq(
+		ALTER TABLE tst SET (parallel_workers = 2);
+		SET client_min_messages = DEBUG;
+		SET maintenance_work_mem = '4MB';
+		CREATE INDEX idx ON tst USING hnsw (v $opclass);
+		ALTER TABLE tst RESET (parallel_workers);
+	));
+	is($ret, 0, $stderr);
+	like($stderr, qr/using \d+ parallel workers/);
+	like($stderr, qr/hnsw graph no longer fits into maintenance_work_mem/);
+
+	$node->safe_psql("postgres", "DROP INDEX idx;");
+}
+
+done_testing();