Fixed warning

.github/workflows/build.yml

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

CHANGELOG.md

@@ -7,7 +7,6 @@
 - Added `jaccard_distance` function
 - Added `quantize_binary` function
 - Added `subvector` function
-- Added CPU dispatching for distance functions on Linux x86-64
 - Updated comparison operators to support vectors with different dimensions
 
 ## 0.6.2 (2024-03-18)

Makefile

@@ -3,7 +3,7 @@ EXTVERSION = 0.6.2
 
 MODULE_big = vector
 DATA = $(wildcard sql/*--*.sql)
-OBJS = src/bitvector.o src/halfutils.o src/halfvec.o 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/sparsevec.o src/vector.o
+OBJS = src/bitvector.o src/halfvec.o 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/sparsevec.o src/vector.o
 HEADERS = src/halfvec.h src/sparsevec.h src/vector.h
 
 TESTS = $(wildcard test/sql/*.sql)

Makefile.win

@@ -1,7 +1,7 @@
 EXTENSION = vector
 EXTVERSION = 0.6.2
 
-OBJS = src\bitvector.obj src\halfutils.obj src\halfvec.obj 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\sparsevec.obj src\vector.obj
+OBJS = src\bitvector.obj src\halfvec.obj 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\sparsevec.obj src\vector.obj
 HEADERS = src\halfvec.h src\sparsevec.h src\vector.h
 
 REGRESS = bit_functions btree cast copy halfvec_functions halfvec_input hnsw_bit_hamming hnsw_bit_jaccard hnsw_halfvec_cosine hnsw_halfvec_ip hnsw_halfvec_l2 hnsw_options hnsw_sparsevec_cosine hnsw_sparsevec_ip hnsw_sparsevec_l2 hnsw_unlogged hnsw_vector_cosine hnsw_vector_ip hnsw_vector_l2 ivfflat_options ivfflat_unlogged ivfflat_vector_cosine ivfflat_vector_ip ivfflat_vector_l2 sparsevec_functions sparsevec_input vector_functions vector_input

README.md

@@ -346,7 +346,6 @@ CREATE INDEX ON items USING ivfflat (embedding vector_cosine_ops) WITH (lists =
 Supported types are:
 
 - `vector` - up to 2,000 dimensions
-- `halfvec` - up to 4,000 dimensions (unreleased)
 
 ### Query Options
 
@@ -420,103 +419,6 @@ 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.
@@ -528,30 +430,6 @@ SELECT id, content FROM items, plainto_tsquery('hello search') query
 
 You can use [Reciprocal Rank Fusion](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search_rrf.py) or a [cross-encoder](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search.py) to combine results.
 
-## Subvector Indexing
-
-*Unreleased*
-
-Use expression indexing to index subvectors
-
-```sql
-CREATE INDEX ON items USING hnsw ((subvector(embedding, 1, 3)::vector(3)) vector_cosine_ops);
-```
-
-Get the nearest neighbors by cosine distance
-
-```sql
-SELECT * FROM items ORDER BY subvector(embedding, 1, 3)::vector(3) <=> subvector('[1,2,3,4,5]'::vector, 1, 3) LIMIT 5;
-```
-
-Re-rank by the full vectors for better recall
-
-```sql
-SELECT * FROM (
-    SELECT * FROM items ORDER BY subvector(embedding, 1, 3)::vector(3) <=> subvector('[1,2,3,4,5]'::vector, 1, 3) LIMIT 20
-) ORDER BY embedding <=> '[1,2,3,4,5]' LIMIT 5;
-```
-
 ## Performance
 
 ### Tuning
@@ -757,6 +635,18 @@ 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:
@@ -911,7 +801,7 @@ 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 16,000 non-zero elements.
+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.
 
 ### Sparsevec Operators
 

sql/vector--0.6.2--0.7.0.sql

@@ -86,9 +86,6 @@ CREATE FUNCTION halfvec_l2_squared_distance(halfvec, halfvec) RETURNS float8
 CREATE FUNCTION halfvec_negative_inner_product(halfvec, halfvec) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION halfvec_spherical_distance(halfvec, halfvec) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
 CREATE FUNCTION halfvec(halfvec, integer, boolean) RETURNS halfvec
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
@@ -140,27 +137,6 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
-CREATE OPERATOR CLASS halfvec_l2_ops
-	FOR TYPE halfvec USING ivfflat AS
-	OPERATOR 1 <-> (halfvec, halfvec) FOR ORDER BY float_ops,
-	FUNCTION 1 halfvec_l2_squared_distance(halfvec, halfvec),
-	FUNCTION 3 l2_distance(halfvec, halfvec);
-
-CREATE OPERATOR CLASS halfvec_ip_ops
-	FOR TYPE halfvec USING ivfflat AS
-	OPERATOR 1 <#> (halfvec, halfvec) FOR ORDER BY float_ops,
-	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
-	FUNCTION 3 halfvec_spherical_distance(halfvec, halfvec),
-	FUNCTION 4 halfvec_norm(halfvec);
-
-CREATE OPERATOR CLASS halfvec_cosine_ops
-	FOR TYPE halfvec USING ivfflat AS
-	OPERATOR 1 <=> (halfvec, halfvec) FOR ORDER BY float_ops,
-	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
-	FUNCTION 2 halfvec_norm(halfvec),
-	FUNCTION 3 halfvec_spherical_distance(halfvec, halfvec),
-	FUNCTION 4 halfvec_norm(halfvec);
-
 CREATE OPERATOR CLASS halfvec_l2_ops
 	FOR TYPE halfvec USING hnsw AS
 	OPERATOR 1 <-> (halfvec, halfvec) FOR ORDER BY float_ops,
@@ -227,6 +203,9 @@ 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

@@ -381,9 +381,6 @@ CREATE FUNCTION halfvec_l2_squared_distance(halfvec, halfvec) RETURNS float8
 CREATE FUNCTION halfvec_negative_inner_product(halfvec, halfvec) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION halfvec_spherical_distance(halfvec, halfvec) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
 -- halfvec cast functions
 
 CREATE FUNCTION halfvec(halfvec, integer, boolean) RETURNS halfvec
@@ -443,27 +440,6 @@ CREATE OPERATOR <=> (
 
 -- halfvec opclasses
 
-CREATE OPERATOR CLASS halfvec_l2_ops
-	FOR TYPE halfvec USING ivfflat AS
-	OPERATOR 1 <-> (halfvec, halfvec) FOR ORDER BY float_ops,
-	FUNCTION 1 halfvec_l2_squared_distance(halfvec, halfvec),
-	FUNCTION 3 l2_distance(halfvec, halfvec);
-
-CREATE OPERATOR CLASS halfvec_ip_ops
-	FOR TYPE halfvec USING ivfflat AS
-	OPERATOR 1 <#> (halfvec, halfvec) FOR ORDER BY float_ops,
-	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
-	FUNCTION 3 halfvec_spherical_distance(halfvec, halfvec),
-	FUNCTION 4 halfvec_norm(halfvec);
-
-CREATE OPERATOR CLASS halfvec_cosine_ops
-	FOR TYPE halfvec USING ivfflat AS
-	OPERATOR 1 <=> (halfvec, halfvec) FOR ORDER BY float_ops,
-	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
-	FUNCTION 2 halfvec_norm(halfvec),
-	FUNCTION 3 halfvec_spherical_distance(halfvec, halfvec),
-	FUNCTION 4 halfvec_norm(halfvec);
-
 CREATE OPERATOR CLASS halfvec_l2_ops
 	FOR TYPE halfvec USING hnsw AS
 	OPERATOR 1 <-> (halfvec, halfvec) FOR ORDER BY float_ops,
@@ -536,6 +512,9 @@ 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/halfutils.c

@@ -1,156 +0,0 @@
-#include "postgres.h"
-
-#include "halfutils.h"
-#include "halfvec.h"
-
-#ifdef HALFVEC_DISPATCH
-#include <immintrin.h>
-
-#if defined(HAVE__GET_CPUID)
-#include <cpuid.h>
-#elif defined(HAVE__CPUID)
-#include <intrin.h>
-#endif
-
-#ifdef _MSC_VER
-#define TARGET_F16C_FMA
-#else
-#define TARGET_F16C_FMA __attribute__((target("f16c,fma")))
-#endif
-#endif
-
-float		(*HalfvecL2SquaredDistance) (int dim, half * ax, half * bx);
-float		(*HalfvecInnerProduct) (int dim, half * ax, half * bx);
-
-static float
-HalfvecL2SquaredDistanceDefault(int dim, half * ax, half * bx)
-{
-	float		distance = 0.0;
-
-	/* Auto-vectorized */
-	for (int i = 0; i < dim; i++)
-	{
-		float		diff = HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]);
-
-		distance += diff * diff;
-	}
-
-	return distance;
-}
-
-#ifdef HALFVEC_DISPATCH
-TARGET_F16C_FMA static float
-HalfvecL2SquaredDistanceF16cFma(int dim, half * ax, half * bx)
-{
-	float		distance;
-	int			i;
-	float		s[8];
-	int			count = (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 < dim; i++)
-	{
-		float		diff = HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]);
-
-		distance += diff * diff;
-	}
-
-	return distance;
-}
-#endif
-
-static float
-HalfvecInnerProductDefault(int dim, half * ax, half * bx)
-{
-	float		distance = 0.0;
-
-	/* Auto-vectorized */
-	for (int i = 0; i < dim; i++)
-		distance += HalfToFloat4(ax[i]) * HalfToFloat4(bx[i]);
-
-	return distance;
-}
-
-#ifdef HALFVEC_DISPATCH
-TARGET_F16C_FMA static float
-HalfvecInnerProductF16cFma(int dim, half * ax, half * bx)
-{
-	float		distance;
-	int			i;
-	float		s[8];
-	int			count = (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);
-
-		dist = _mm256_fmadd_ps(axs, bxs, 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 < dim; i++)
-		distance += HalfToFloat4(ax[i]) * HalfToFloat4(bx[i]);
-
-	return distance;
-}
-#endif
-
-#ifdef HALFVEC_DISPATCH
-#define CPU_FEATURE_FMA  (1 << 12)
-#define CPU_FEATURE_F16C (1 << 29)
-
-static bool
-SupportsCpuFeature(unsigned int feature)
-{
-	unsigned int exx[4] = {0, 0, 0, 0};
-
-#if defined(HAVE__GET_CPUID)
-	__get_cpuid(1, &exx[0], &exx[1], &exx[2], &exx[3]);
-#elif defined(HAVE__CPUID)
-	__cpuid(exx, 1);
-#endif
-
-	return (exx[2] & feature) == feature;
-}
-#endif
-
-void
-HalfvecInit(void)
-{
-	/*
-	 * Could skip pointer when single function, but no difference in
-	 * performance
-	 */
-	HalfvecL2SquaredDistance = HalfvecL2SquaredDistanceDefault;
-	HalfvecInnerProduct = HalfvecInnerProductDefault;
-
-#ifdef HALFVEC_DISPATCH
-	if (SupportsCpuFeature(CPU_FEATURE_FMA | CPU_FEATURE_F16C))
-	{
-		HalfvecL2SquaredDistance = HalfvecL2SquaredDistanceF16cFma;
-		HalfvecInnerProduct = HalfvecInnerProductF16cFma;
-	}
-#endif
-}

src/halfutils.h

@@ -1,254 +0,0 @@
-#ifndef HALFUTILS_H
-#define HALFUTILS_H
-
-#include <math.h>
-
-#include "common/shortest_dec.h"
-#include "halfvec.h"
-
-#ifdef F16C_SUPPORT
-#include <immintrin.h>
-#endif
-
-extern float (*HalfvecL2SquaredDistance) (int dim, half * ax, half * bx);
-extern float (*HalfvecInnerProduct) (int dim, half * ax, half * bx);
-
-void		HalfvecInit(void);
-
-/*
- * Check if half is NaN
- */
-static inline bool
-HalfIsNan(half num)
-{
-#ifdef FLT16_SUPPORT
-	return isnan(num);
-#else
-	return (num & 0x7C00) == 0x7C00 && (num & 0x7FFF) != 0x7C00;
-#endif
-}
-
-/*
- * Check if half is infinite
- */
-static inline bool
-HalfIsInf(half num)
-{
-#ifdef FLT16_SUPPORT
-	return isinf(num);
-#else
-	return (num & 0x7FFF) == 0x7C00;
-#endif
-}
-
-/*
- * Convert a half to a float4
- */
-static inline float
-HalfToFloat4(half num)
-{
-#if defined(F16C_SUPPORT)
-	return _cvtsh_ss(num);
-#elif defined(FLT16_SUPPORT)
-	return (float) num;
-#else
-	/* TODO Improve performance */
-
-	/* Assumes same endianness for floats and integers */
-	union
-	{
-		float		f;
-		uint32		i;
-	}			swapfloat;
-
-	union
-	{
-		half		h;
-		uint16		i;
-	}			swaphalf;
-
-	uint16		bin;
-	uint32		exponent;
-	uint32		mantissa;
-	uint32		result;
-
-	swaphalf.h = num;
-	bin = swaphalf.i;
-	exponent = (bin & 0x7C00) >> 10;
-	mantissa = bin & 0x03FF;
-
-	/* Sign */
-	result = (bin & 0x8000) << 16;
-
-	if (unlikely(exponent == 31))
-	{
-		if (mantissa == 0)
-		{
-			/* Infinite */
-			result |= 0x7F800000;
-		}
-		else
-		{
-			/* NaN */
-			result |= 0x7FC00000;
-		}
-	}
-	else if (unlikely(exponent == 0))
-	{
-		/* Subnormal */
-		if (mantissa != 0)
-		{
-			exponent = -14;
-
-			for (int i = 0; i < 10; i++)
-			{
-				mantissa <<= 1;
-				exponent -= 1;
-
-				if ((mantissa >> 10) % 2 == 1)
-				{
-					mantissa &= 0x03ff;
-					break;
-				}
-			}
-
-			result |= (exponent + 127) << 23;
-		}
-	}
-	else
-	{
-		/* Normal */
-		result |= (exponent - 15 + 127) << 23;
-	}
-
-	result |= mantissa << 13;
-
-	swapfloat.i = result;
-	return swapfloat.f;
-#endif
-}
-
-/*
- * Convert a float4 to a half
- */
-static inline half
-Float4ToHalfUnchecked(float num)
-{
-#if defined(F16C_SUPPORT)
-	return _cvtss_sh(num, 0);
-#elif defined(FLT16_SUPPORT)
-	return (_Float16) num;
-#else
-	/* TODO Improve performance */
-
-	/* Assumes same endianness for floats and integers */
-	union
-	{
-		float		f;
-		uint32		i;
-	}			swapfloat;
-
-	union
-	{
-		half		h;
-		uint16		i;
-	}			swaphalf;
-
-	uint32		bin;
-	int			exponent;
-	int			mantissa;
-	uint16		result;
-
-	swapfloat.f = num;
-	bin = swapfloat.i;
-	exponent = (bin & 0x7F800000) >> 23;
-	mantissa = bin & 0x007FFFFF;
-
-	/* Sign */
-	result = (bin & 0x80000000) >> 16;
-
-	if (isinf(num))
-	{
-		/* Infinite */
-		result |= 0x7C00;
-	}
-	else if (isnan(num))
-	{
-		/* NaN */
-		result |= 0x7E00;
-		result |= mantissa >> 13;
-	}
-	else if (exponent > 98)
-	{
-		int			m;
-		int			gr;
-		int			s;
-
-		exponent -= 127;
-		s = mantissa & 0x00000FFF;
-
-		/* Subnormal */
-		if (exponent < -14)
-		{
-			int			diff = -exponent - 14;
-
-			mantissa >>= diff;
-			mantissa += 1 << (23 - diff);
-			s |= mantissa & 0x00000FFF;
-		}
-
-		m = mantissa >> 13;
-
-		/* Round */
-		gr = (mantissa >> 12) % 4;
-		if (gr == 3 || (gr == 1 && s != 0))
-			m += 1;
-
-		if (m == 1024)
-		{
-			m = 0;
-			exponent += 1;
-		}
-
-		if (exponent > 15)
-		{
-			/* Infinite */
-			result |= 0x7C00;
-		}
-		else
-		{
-			if (exponent >= -14)
-				result |= (exponent + 15) << 10;
-
-			result |= m;
-		}
-	}
-
-	swaphalf.i = result;
-	return swaphalf.h;
-#endif
-}
-
-/*
- * Convert a float4 to a half
- */
-static inline half
-Float4ToHalf(float num)
-{
-	half		result = Float4ToHalfUnchecked(num);
-
-	if (unlikely(HalfIsInf(result)) && !isinf(num))
-	{
-		char	   *buf = palloc(FLOAT_SHORTEST_DECIMAL_LEN);
-
-		float_to_shortest_decimal_buf(num, buf);
-
-		ereport(ERROR,
-				(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
-				 errmsg("\"%s\" is out of range for type halfvec", buf)));
-	}
-
-	return result;
-}
-
-#endif

src/halfvec.c

@@ -6,7 +6,6 @@
 #include "catalog/pg_type.h"
 #include "common/shortest_dec.h"
 #include "fmgr.h"
-#include "halfutils.h"
 #include "halfvec.h"
 #include "lib/stringinfo.h"
 #include "libpq/pqformat.h"
@@ -23,6 +22,45 @@
 #define TYPALIGN_INT 'i'
 #endif
 
+/*
+ * Check if half is NaN
+ */
+static inline bool
+HalfIsNan(half num)
+{
+#ifdef FLT16_SUPPORT
+	return isnan(num);
+#else
+	return (num & 0x7C00) == 0x7C00 && (num & 0x7FFF) != 0x7C00;
+#endif
+}
+
+/*
+ * Check if half is infinite
+ */
+static inline bool
+HalfIsInf(half num)
+{
+#ifdef FLT16_SUPPORT
+	return isinf(num);
+#else
+	return (num & 0x7FFF) == 0x7C00;
+#endif
+}
+
+/*
+ * Check if half is zero
+ */
+static inline bool
+HalfIsZero(half num)
+{
+#ifdef FLT16_SUPPORT
+	return num == 0;
+#else
+	return (num & 0x7FFF) == 0x0000;
+#endif
+}
+
 /*
  * Get a half from a message buffer
  */
@@ -55,6 +93,207 @@ pq_sendhalf(StringInfo buf, half h)
 	pq_sendint16(buf, swap.i);
 }
 
+/*
+ * Convert a half to a float4
+ */
+float
+HalfToFloat4(half num)
+{
+#ifdef FLT16_SUPPORT
+	return (float) num;
+#else
+	/* TODO Improve performance */
+
+	/* Assumes same endianness for floats and integers */
+	union
+	{
+		float		f;
+		uint32		i;
+	}			swapfloat;
+
+	union
+	{
+		half		h;
+		uint16		i;
+	}			swaphalf;
+
+	uint16		bin;
+	uint32		exponent;
+	uint32		mantissa;
+	uint32		result;
+
+	swaphalf.h = num;
+	bin = swaphalf.i;
+	exponent = (bin & 0x7C00) >> 10;
+	mantissa = bin & 0x03FF;
+
+	/* Sign */
+	result = (bin & 0x8000) << 16;
+
+	if (exponent == 31)
+	{
+		if (mantissa == 0)
+		{
+			/* Infinite */
+			result |= 0x7F800000;
+		}
+		else
+		{
+			/* NaN */
+			result |= 0x7FC00000;
+			result |= mantissa << 13;
+		}
+	}
+	else if (exponent == 0)
+	{
+		/* Subnormal */
+		if (mantissa != 0)
+		{
+			exponent = -14;
+
+			for (int i = 0; i < 10; i++)
+			{
+				mantissa <<= 1;
+				exponent -= 1;
+
+				if ((mantissa >> 10) % 2 == 1)
+				{
+					mantissa &= 0x03ff;
+					break;
+				}
+			}
+
+			result |= (exponent + 127) << 23;
+			result |= mantissa << 13;
+		}
+	}
+	else
+	{
+		/* Normal */
+		result |= (exponent - 15 + 127) << 23;
+		result |= mantissa << 13;
+	}
+
+	swapfloat.i = result;
+	return swapfloat.f;
+#endif
+}
+
+/*
+ * Convert a float4 to a half
+ */
+half
+Float4ToHalfUnchecked(float num)
+{
+#ifdef FLT16_SUPPORT
+	return (_Float16) num;
+#else
+	/* TODO Improve performance */
+
+	/* Assumes same endianness for floats and integers */
+	union
+	{
+		float		f;
+		uint32		i;
+	}			swapfloat;
+
+	union
+	{
+		half		h;
+		uint16		i;
+	}			swaphalf;
+
+	uint32		bin;
+	int			exponent;
+	int			mantissa;
+	uint16		result;
+
+	swapfloat.f = num;
+	bin = swapfloat.i;
+	exponent = (bin & 0x7F800000) >> 23;
+	mantissa = bin & 0x007FFFFF;
+
+	/* Sign */
+	result = (bin & 0x80000000) >> 16;
+
+	if (isinf(num))
+	{
+		/* Infinite */
+		result |= 0x7C00;
+	}
+	else if (isnan(num))
+	{
+		/* NaN */
+		result |= 0x7E00;
+		result |= mantissa >> 13;
+	}
+	else if (exponent > 98)
+	{
+		int			m;
+		int			gr;
+		int			s;
+
+		exponent -= 127;
+		s = mantissa & 0x00000FFF;
+
+		/* Subnormal */
+		if (exponent < -14)
+		{
+			int			diff = -exponent - 14;
+
+			mantissa >>= diff;
+			mantissa += 1 << (23 - diff);
+			s |= mantissa & 0x00000FFF;
+		}
+
+		m = mantissa >> 13;
+
+		/* Round */
+		gr = (mantissa >> 12) % 4;
+		if (gr == 3 || (gr == 1 && s != 0))
+			m += 1;
+
+		if (m == 1024)
+		{
+			m = 0;
+			exponent += 1;
+		}
+
+		if (exponent > 15)
+		{
+			/* Infinite */
+			result |= 0x7C00;
+		}
+		else
+		{
+			if (exponent >= -14)
+				result |= (exponent + 15) << 10;
+
+			result |= m;
+		}
+	}
+
+	swaphalf.i = result;
+	return swaphalf.h;
+#endif
+}
+
+/*
+ * Convert a float4 to a half
+ */
+half
+Float4ToHalf(float num)
+{
+	half		result = Float4ToHalfUnchecked(num);
+
+	if (unlikely(HalfIsInf(result)) && !isinf(num))
+		float_overflow_error();
+	if (unlikely(HalfIsZero(result)) && num != 0.0)
+		float_underflow_error();
+
+	return result;
+}
+
 /*
  * Ensure same dimensions
  */
@@ -146,6 +385,24 @@ halfvec_isspace(char ch)
 	return false;
 }
 
+#if PG_VERSION_NUM < 120003
+static pg_noinline void
+float_overflow_error(void)
+{
+	ereport(ERROR,
+			(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
+			 errmsg("value out of range: overflow")));
+}
+
+static pg_noinline void
+float_underflow_error(void)
+{
+	ereport(ERROR,
+			(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
+			 errmsg("value out of range: underflow")));
+}
+#endif
+
 /*
  * Convert textual representation to internal representation
  */
@@ -157,33 +414,27 @@ halfvec_in(PG_FUNCTION_ARGS)
 	int32		typmod = PG_GETARG_INT32(2);
 	half		x[HALFVEC_MAX_DIM];
 	int			dim = 0;
-	char	   *pt = lit;
+	char	   *pt;
+	char	   *stringEnd;
 	HalfVector *result;
+	char	   *litcopy = pstrdup(lit);
+	char	   *str = litcopy;
 
-	while (halfvec_isspace(*pt))
-		pt++;
+	while (halfvec_isspace(*str))
+		str++;
 
-	if (*pt != '[')
+	if (*str != '[')
 		ereport(ERROR,
 				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-				 errmsg("invalid input syntax for type halfvec: \"%s\"", lit),
+				 errmsg("malformed halfvec literal: \"%s\"", lit),
 				 errdetail("Vector contents must start with \"[\".")));
 
-	pt++;
+	str++;
+	pt = strtok(str, ",");
+	stringEnd = pt;
 
-	while (halfvec_isspace(*pt))
-		pt++;
-
-	if (*pt == ']')
-		ereport(ERROR,
-				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("halfvec must have at least 1 dimension")));
-
-	for (;;)
+	while (pt != NULL && *stringEnd != ']')
 	{
-		float		val;
-		char	   *stringEnd;
-
 		if (dim == HALFVEC_MAX_DIM)
 			ereport(ERROR,
 					(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
@@ -198,56 +449,61 @@ halfvec_in(PG_FUNCTION_ARGS)
 					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
 					 errmsg("invalid input syntax for type halfvec: \"%s\"", lit)));
 
-		errno = 0;
-
-		/* Postgres sets LC_NUMERIC to C on startup */
-		val = strtof(pt, &stringEnd);
+		/* Use strtof like float4in to avoid a double-rounding problem */
+		x[dim] = Float4ToHalf(strtof(pt, &stringEnd));
+		CheckElement(x[dim]);
+		dim++;
 
 		if (stringEnd == pt)
 			ereport(ERROR,
 					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
 					 errmsg("invalid input syntax for type halfvec: \"%s\"", lit)));
 
-		x[dim] = Float4ToHalfUnchecked(val);
+		while (halfvec_isspace(*stringEnd))
+			stringEnd++;
 
-		/* Check for range error like float4in */
-		if ((errno == ERANGE && isinf(val)) || (HalfIsInf(x[dim]) && !isinf(val)))
-			ereport(ERROR,
-					(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
-					 errmsg("\"%s\" is out of range for type halfvec", pnstrdup(pt, stringEnd - pt))));
-
-		CheckElement(x[dim]);
-		dim++;
-
-		pt = stringEnd;
-
-		while (halfvec_isspace(*pt))
-			pt++;
-
-		if (*pt == ',')
-			pt++;
-		else if (*pt == ']')
-		{
-			pt++;
-			break;
-		}
-		else
+		if (*stringEnd != '\0' && *stringEnd != ']')
 			ereport(ERROR,
 					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
 					 errmsg("invalid input syntax for type halfvec: \"%s\"", lit)));
+
+		pt = strtok(NULL, ",");
 	}
 
-	/* Only whitespace is allowed after the closing brace */
-	while (halfvec_isspace(*pt))
-		pt++;
-
-	if (*pt != '\0')
+	if (stringEnd == NULL || *stringEnd != ']')
 		ereport(ERROR,
 				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-				 errmsg("invalid input syntax for type halfvec: \"%s\"", lit),
+				 errmsg("malformed halfvec literal: \"%s\"", lit),
+				 errdetail("Unexpected end of input.")));
+
+	stringEnd++;
+
+	/* Only whitespace is allowed after the closing brace */
+	while (halfvec_isspace(*stringEnd))
+		stringEnd++;
+
+	if (*stringEnd != '\0')
+		ereport(ERROR,
+				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+				 errmsg("malformed halfvec literal: \"%s\"", lit),
 				 errdetail("Junk after closing right brace.")));
 
-	CheckDim(dim);
+	/* Ensure no consecutive delimiters since strtok skips */
+	for (pt = lit + 1; *pt != '\0'; pt++)
+	{
+		if (pt[-1] == ',' && *pt == ',')
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("malformed halfvec literal: \"%s\"", lit)));
+	}
+
+	if (dim < 1)
+		ereport(ERROR,
+				(errcode(ERRCODE_DATA_EXCEPTION),
+				 errmsg("halfvec must have at least 1 dimension")));
+
+	pfree(litcopy);
+
 	CheckExpectedDim(typmod, dim);
 
 	result = InitHalfVector(dim);
@@ -517,7 +773,11 @@ vector_to_halfvec(PG_FUNCTION_ARGS)
 	result = InitHalfVector(vec->dim);
 
 	for (int i = 0; i < vec->dim; i++)
-		result->x[i] = Float4ToHalf(vec->x[i]);
+	{
+		result->x[i] = Float4ToHalfUnchecked(vec->x[i]);
+		/* TODO Better error for overflow */
+		CheckElement(result->x[i]);
+	}
 
 	PG_RETURN_POINTER(result);
 }
@@ -531,10 +791,21 @@ 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);
 
-	PG_RETURN_FLOAT8(sqrt((double) HalfvecL2SquaredDistance(a->dim, a->x, b->x)));
+	/* 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));
 }
 
 /*
@@ -546,10 +817,21 @@ halfvec_l2_squared_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);
 
-	PG_RETURN_FLOAT8((double) HalfvecL2SquaredDistance(a->dim, a->x, b->x));
+	/* Auto-vectorized */
+	for (int i = 0; i < a->dim; i++)
+	{
+		float		diff = HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]);
+
+		distance += diff * diff;
+	}
+
+	PG_RETURN_FLOAT8((double) distance);
 }
 
 /*
@@ -561,10 +843,17 @@ 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);
 
-	PG_RETURN_FLOAT8((double) HalfvecInnerProduct(a->dim, a->x, b->x));
+	/* Auto-vectorized */
+	for (int i = 0; i < a->dim; i++)
+		distance += HalfToFloat4(ax[i]) * HalfToFloat4(bx[i]);
+
+	PG_RETURN_FLOAT8((double) distance);
 }
 
 /*
@@ -576,10 +865,17 @@ 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);
 
-	PG_RETURN_FLOAT8((double) -HalfvecInnerProduct(a->dim, a->x, b->x));
+	/* Auto-vectorized */
+	for (int i = 0; i < a->dim; i++)
+		distance += HalfToFloat4(ax[i]) * HalfToFloat4(bx[i]);
+
+	PG_RETURN_FLOAT8((double) distance * -1);
 }
 
 /*
@@ -629,32 +925,6 @@ halfvec_cosine_distance(PG_FUNCTION_ARGS)
 	PG_RETURN_FLOAT8(1 - similarity);
 }
 
-/*
- * Get the distance for spherical k-means
- * Currently uses angular distance since needs to satisfy triangle inequality
- * Assumes inputs are unit vectors (skips norm)
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_spherical_distance);
-Datum
-halfvec_spherical_distance(PG_FUNCTION_ARGS)
-{
-	HalfVector *a = PG_GETARG_HALFVEC_P(0);
-	HalfVector *b = PG_GETARG_HALFVEC_P(1);
-	double		distance;
-
-	CheckDims(a, b);
-
-	distance = (double) HalfvecInnerProduct(a->dim, a->x, b->x);
-
-	/* Prevent NaN with acos with loss of precision */
-	if (distance > 1)
-		distance = 1;
-	else if (distance < -1)
-		distance = -1;
-
-	PG_RETURN_FLOAT8(acos(distance) / M_PI);
-}
-
 /*
  * Get the L1 distance between two half vectors
  */
@@ -748,30 +1018,3 @@ halfvec_subvector(PG_FUNCTION_ARGS)
 
 	PG_RETURN_POINTER(result);
 }
-
-/*
- * Internal helper to compare half vectors
- */
-int
-halfvec_cmp_internal(HalfVector * a, HalfVector * b)
-{
-	int			dim = Min(a->dim, b->dim);
-
-	/* Check values before dimensions to be consistent with Postgres arrays */
-	for (int i = 0; i < dim; i++)
-	{
-		if (HalfToFloat4(a->x[i]) < HalfToFloat4(b->x[i]))
-			return -1;
-
-		if (HalfToFloat4(a->x[i]) > HalfToFloat4(b->x[i]))
-			return 1;
-	}
-
-	if (a->dim < b->dim)
-		return -1;
-
-	if (a->dim > b->dim)
-		return 1;
-
-	return 0;
-}

src/halfvec.h

@@ -7,14 +7,7 @@
 
 #include "vector.h"
 
-#if defined(__x86_64__) || defined(_M_AMD64)
-#define HALFVEC_DISPATCH
-#endif
-
-/* F16C has better performance than _Float16 (on x86-64) */
-#if defined(__F16C__)
-#define F16C_SUPPORT
-#elif defined(__FLT16_MAX__) && !defined(HALFVEC_DISPATCH)
+#ifdef __FLT16_MAX__
 #define FLT16_SUPPORT
 #endif
 
@@ -22,6 +15,7 @@
 #define half _Float16
 #define HALF_MAX FLT16_MAX
 #else
+/* TODO #pragma message("")? */
 #define half uint16
 #define HALF_MAX 65504
 #endif
@@ -42,6 +36,8 @@ typedef struct HalfVector
 }			HalfVector;
 
 HalfVector *InitHalfVector(int dim);
-int			halfvec_cmp_internal(HalfVector * a, HalfVector * b);
+float		HalfToFloat4(half num);
+half		Float4ToHalf(float num);
+half		Float4ToHalfUnchecked(float num);
 
 #endif

src/hnswscan.c

@@ -158,10 +158,6 @@ 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

@@ -5,7 +5,6 @@
 #include "access/generic_xlog.h"
 #include "catalog/pg_type.h"
 #include "catalog/pg_type_d.h"
-#include "halfutils.h"
 #include "halfvec.h"
 #include "hnsw.h"
 #include "lib/pairingheap.h"
@@ -164,9 +163,9 @@ HnswGetType(Relation index)
 	Oid			typid = TupleDescAttr(index->rd_att, 0)->atttypid;
 	HeapTuple	tuple;
 	Form_pg_type type;
-	HnswType	result;
+	int			result;
 
-	if (typid == BITOID)
+	if (typid == BITOID || typid == VARBITOID)
 		return HNSW_TYPE_BIT;
 
 	tuple = SearchSysCache1(TYPEOID, ObjectIdGetDatum(typid));
@@ -181,10 +180,7 @@ HnswGetType(Relation index)
 	else if (strcmp(NameStr(type->typname), "sparsevec") == 0)
 		result = HNSW_TYPE_SPARSEVEC;
 	else
-	{
-		ReleaseSysCache(tuple);
-		elog(ERROR, "type not supported for hnsw index");
-	}
+		elog(ERROR, "Unsupported type");
 
 	ReleaseSysCache(tuple);
 
@@ -223,7 +219,10 @@ HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, HnswType type)
 			HalfVector *result = InitHalfVector(v->dim);
 
 			for (int i = 0; i < v->dim; i++)
+			{
+				/* TODO Fix */
 				result->x[i] = Float4ToHalfUnchecked(HalfToFloat4(v->x[i]) / norm);
+			}
 
 			*value = PointerGetDatum(result);
 		}

src/ivfbuild.c

@@ -10,14 +10,12 @@
 #include "catalog/pg_operator_d.h"
 #include "catalog/pg_type_d.h"
 #include "commands/progress.h"
-#include "halfvec.h"
 #include "ivfflat.h"
 #include "miscadmin.h"
 #include "optimizer/optimizer.h"
 #include "storage/bufmgr.h"
 #include "tcop/tcopprot.h"
 #include "utils/memutils.h"
-#include "vector.h"
 
 #if PG_VERSION_NUM >= 140000
 #include "utils/backend_progress.h"
@@ -59,13 +57,13 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 	 */
 	if (buildstate->kmeansnormprocinfo != NULL)
 	{
-		if (!IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->collation, &value, buildstate->type))
+		if (!IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->collation, &value))
 			return;
 	}
 
 	if (samples->length < targsamples)
 	{
-		VectorArraySet(samples, samples->length, DatumGetPointer(value));
+		VectorArraySet(samples, samples->length, DatumGetVector(value));
 		samples->length++;
 	}
 	else
@@ -82,7 +80,7 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 #endif
 
 			Assert(k >= 0 && k < targsamples);
-			VectorArraySet(samples, k, DatumGetPointer(value));
+			VectorArraySet(samples, k, DatumGetVector(value));
 		}
 
 		buildstate->rowstoskip -= 1;
@@ -155,7 +153,7 @@ AddTupleToSort(Relation index, ItemPointer tid, Datum *values, IvfflatBuildState
 	/* Normalize if needed */
 	if (buildstate->normprocinfo != NULL)
 	{
-		if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->type))
+		if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value))
 			return;
 	}
 
@@ -314,58 +312,25 @@ InsertTuples(Relation index, IvfflatBuildState * buildstate, ForkNumber forkNum)
 	}
 }
 
-/*
- * Get max dimensions
- */
-static int
-GetMaxDimensions(IvfflatType type)
-{
-	int			maxDimensions = IVFFLAT_MAX_DIM;
-
-	if (type == IVFFLAT_TYPE_HALFVEC)
-		maxDimensions *= 2;
-
-	return maxDimensions;
-}
-
-/*
- * Get item size
- */
-static Size
-GetItemSize(IvfflatType type, int dimensions)
-{
-	if (type == IVFFLAT_TYPE_VECTOR)
-		return VECTOR_SIZE(dimensions);
-	else if (type == IVFFLAT_TYPE_HALFVEC)
-		return HALFVEC_SIZE(dimensions);
-	else
-		elog(ERROR, "Unsupported type");
-}
-
 /*
  * 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 > maxDimensions)
-		elog(ERROR, "column cannot have more than %d dimensions for ivfflat index", maxDimensions);
+	if (buildstate->dimensions > IVFFLAT_MAX_DIM)
+		elog(ERROR, "column cannot have more than %d dimensions for ivfflat index", IVFFLAT_MAX_DIM);
 
 	buildstate->reltuples = 0;
 	buildstate->indtuples = 0;
@@ -388,7 +353,7 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 
 	buildstate->slot = MakeSingleTupleTableSlot(buildstate->tupdesc, &TTSOpsVirtual);
 
-	buildstate->centers = VectorArrayInit(buildstate->lists, buildstate->dimensions, GetItemSize(buildstate->type, buildstate->dimensions));
+	buildstate->centers = VectorArrayInit(buildstate->lists, buildstate->dimensions);
 	buildstate->listInfo = palloc(sizeof(ListInfo) * buildstate->lists);
 
 	buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
@@ -443,7 +408,7 @@ ComputeCenters(IvfflatBuildState * buildstate)
 
 	/* Sample rows */
 	/* TODO Ensure within maintenance_work_mem */
-	buildstate->samples = VectorArrayInit(numSamples, buildstate->dimensions, buildstate->centers->itemsize);
+	buildstate->samples = VectorArrayInit(numSamples, buildstate->dimensions);
 	if (buildstate->heap != NULL)
 	{
 		SampleRows(buildstate);
@@ -458,7 +423,7 @@ ComputeCenters(IvfflatBuildState * buildstate)
 	}
 
 	/* Calculate centers */
-	IvfflatBench("k-means", IvfflatKmeans(buildstate->index, buildstate->samples, buildstate->centers, buildstate->type));
+	IvfflatBench("k-means", IvfflatKmeans(buildstate->index, buildstate->samples, buildstate->centers));
 
 	/* Free samples before we allocate more memory */
 	VectorArrayFree(buildstate->samples);
@@ -503,7 +468,7 @@ CreateListPages(Relation index, VectorArray centers, int dimensions,
 	Size		listSize;
 	IvfflatList list;
 
-	listSize = MAXALIGN(IVFFLAT_LIST_SIZE(centers->itemsize));
+	listSize = MAXALIGN(IVFFLAT_LIST_SIZE(dimensions));
 	list = palloc0(listSize);
 
 	buf = IvfflatNewBuffer(index, forkNum);
@@ -516,7 +481,7 @@ CreateListPages(Relation index, VectorArray centers, int dimensions,
 		/* Load list */
 		list->startPage = InvalidBlockNumber;
 		list->insertPage = InvalidBlockNumber;
-		memcpy(&list->center, VectorArrayGet(centers, i), centers->itemsize);
+		memcpy(&list->center, VectorArrayGet(centers, i), VECTOR_SIZE(dimensions));
 
 		/* Ensure free space */
 		if (PageGetFreeSpace(page) < listSize)
@@ -621,7 +586,7 @@ ParallelHeapScan(IvfflatBuildState * buildstate)
  * Perform a worker's portion of a parallel sort
  */
 static void
-IvfflatParallelScanAndSort(IvfflatSpool * ivfspool, IvfflatShared * ivfshared, Sharedsort *sharedsort, char *ivfcenters, int sortmem, bool progress)
+IvfflatParallelScanAndSort(IvfflatSpool * ivfspool, IvfflatShared * ivfshared, Sharedsort *sharedsort, Vector * ivfcenters, int sortmem, bool progress)
 {
 	SortCoordinate coordinate;
 	IvfflatBuildState buildstate;
@@ -645,7 +610,7 @@ IvfflatParallelScanAndSort(IvfflatSpool * ivfspool, IvfflatShared * ivfshared, S
 	indexInfo = BuildIndexInfo(ivfspool->index);
 	indexInfo->ii_Concurrent = ivfshared->isconcurrent;
 	InitBuildState(&buildstate, ivfspool->heap, ivfspool->index, indexInfo);
-	memcpy(buildstate.centers->items, ivfcenters, buildstate.centers->itemsize * buildstate.centers->maxlen);
+	memcpy(buildstate.centers->items, ivfcenters, VECTOR_SIZE(buildstate.centers->dim) * buildstate.centers->maxlen);
 	buildstate.centers->length = buildstate.centers->maxlen;
 	ivfspool->sortstate = tuplesort_begin_heap(buildstate.tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, sortmem, coordinate, false);
 	buildstate.sortstate = ivfspool->sortstate;
@@ -693,7 +658,7 @@ IvfflatParallelBuildMain(dsm_segment *seg, shm_toc *toc)
 	IvfflatSpool *ivfspool;
 	IvfflatShared *ivfshared;
 	Sharedsort *sharedsort;
-	char	   *ivfcenters;
+	Vector	   *ivfcenters;
 	Relation	heapRel;
 	Relation	indexRel;
 	LOCKMODE	heapLockmode;
@@ -807,7 +772,7 @@ IvfflatBeginParallel(IvfflatBuildState * buildstate, bool isconcurrent, int requ
 	Size		estcenters;
 	IvfflatShared *ivfshared;
 	Sharedsort *sharedsort;
-	char	   *ivfcenters;
+	Vector	   *ivfcenters;
 	IvfflatLeader *ivfleader = (IvfflatLeader *) palloc0(sizeof(IvfflatLeader));
 	bool		leaderparticipates = true;
 	int			querylen;
@@ -834,7 +799,7 @@ IvfflatBeginParallel(IvfflatBuildState * buildstate, bool isconcurrent, int requ
 	shm_toc_estimate_chunk(&pcxt->estimator, estivfshared);
 	estsort = tuplesort_estimate_shared(scantuplesortstates);
 	shm_toc_estimate_chunk(&pcxt->estimator, estsort);
-	estcenters = buildstate->centers->itemsize * buildstate->centers->maxlen;
+	estcenters = VECTOR_SIZE(buildstate->dimensions) * buildstate->lists;
 	shm_toc_estimate_chunk(&pcxt->estimator, estcenters);
 	shm_toc_estimate_keys(&pcxt->estimator, 3);
 
@@ -886,7 +851,7 @@ IvfflatBeginParallel(IvfflatBuildState * buildstate, bool isconcurrent, int requ
 	tuplesort_initialize_shared(sharedsort, scantuplesortstates,
 								pcxt->seg);
 
-	ivfcenters = shm_toc_allocate(pcxt->toc, estcenters);
+	ivfcenters = (Vector *) shm_toc_allocate(pcxt->toc, estcenters);
 	memcpy(ivfcenters, buildstate->centers->items, estcenters);
 
 	shm_toc_insert(pcxt->toc, PARALLEL_KEY_IVFFLAT_SHARED, ivfshared);

src/ivfflat.h

@@ -43,19 +43,13 @@
 #define IVFFLAT_MAX_LISTS		32768
 #define IVFFLAT_DEFAULT_PROBES	1
 
-typedef enum IvfflatType
-{
-	IVFFLAT_TYPE_VECTOR,
-	IVFFLAT_TYPE_HALFVEC
-}			IvfflatType;
-
 /* Build phases */
 /* PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE is 1 */
 #define PROGRESS_IVFFLAT_PHASE_KMEANS	2
 #define PROGRESS_IVFFLAT_PHASE_ASSIGN	3
 #define PROGRESS_IVFFLAT_PHASE_LOAD		4
 
-#define IVFFLAT_LIST_SIZE(size)	(offsetof(IvfflatListData, center) + size)
+#define IVFFLAT_LIST_SIZE(_dim)	(offsetof(IvfflatListData, center) + VECTOR_SIZE(_dim))
 
 #define IvfflatPageGetOpaque(page)	((IvfflatPageOpaque) PageGetSpecialPointer(page))
 #define IvfflatPageGetMeta(page)	((IvfflatMetaPageData *) PageGetContents(page))
@@ -91,8 +85,7 @@ typedef struct VectorArrayData
 	int			length;
 	int			maxlen;
 	int			dim;
-	Size		itemsize;
-	char	   *items;
+	Vector	   *items;
 }			VectorArrayData;
 
 typedef VectorArrayData * VectorArray;
@@ -151,7 +144,7 @@ typedef struct IvfflatLeader
 	IvfflatShared *ivfshared;
 	Sharedsort *sharedsort;
 	Snapshot	snapshot;
-	char	   *ivfcenters;
+	Vector	   *ivfcenters;
 }			IvfflatLeader;
 
 typedef struct IvfflatBuildState
@@ -160,7 +153,6 @@ typedef struct IvfflatBuildState
 	Relation	heap;
 	Relation	index;
 	IndexInfo  *indexInfo;
-	IvfflatType type;
 
 	/* Settings */
 	int			dimensions;
@@ -263,18 +255,18 @@ typedef struct IvfflatScanOpaqueData
 
 typedef IvfflatScanOpaqueData * IvfflatScanOpaque;
 
-#define VECTOR_ARRAY_SIZE(_length, _size) (sizeof(VectorArrayData) + (_length) * _size)
-#define VECTOR_ARRAY_OFFSET(_arr, _offset) ((char*) (_arr)->items + (_offset) * (_arr)->itemsize)
-#define VectorArrayGet(_arr, _offset) VECTOR_ARRAY_OFFSET(_arr, _offset)
-#define VectorArraySet(_arr, _offset, _val) memcpy(VECTOR_ARRAY_OFFSET(_arr, _offset), _val, (_arr)->itemsize)
+#define VECTOR_ARRAY_SIZE(_length, _dim) (sizeof(VectorArrayData) + (_length) * VECTOR_SIZE(_dim))
+#define VECTOR_ARRAY_OFFSET(_arr, _offset) ((char*) (_arr)->items + (_offset) * VECTOR_SIZE((_arr)->dim))
+#define VectorArrayGet(_arr, _offset) ((Vector *) VECTOR_ARRAY_OFFSET(_arr, _offset))
+#define VectorArraySet(_arr, _offset, _val) memcpy(VECTOR_ARRAY_OFFSET(_arr, _offset), _val, VECTOR_SIZE((_arr)->dim))
 
 /* Methods */
-VectorArray VectorArrayInit(int maxlen, int dimensions, Size itemsize);
+VectorArray VectorArrayInit(int maxlen, int dimensions);
 void		VectorArrayFree(VectorArray arr);
-void		IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatType type);
+void		PrintVectorArray(char *msg, VectorArray arr);
+void		IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers);
 FmgrInfo   *IvfflatOptionalProcInfo(Relation index, uint16 procnum);
-IvfflatType IvfflatGetType(Relation index);
-bool		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, IvfflatType type);
+bool		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value);
 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, IvfflatGetType(index)))
+		if (!IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value))
 			return;
 	}
 

src/ivfkmeans.c

@@ -3,13 +3,12 @@
 #include <float.h>
 #include <math.h>
 
-#include "halfutils.h"
-#include "halfvec.h"
 #include "ivfflat.h"
 #include "miscadmin.h"
-#include "utils/datum.h"
+
+#ifdef IVFFLAT_MEMORY
 #include "utils/memutils.h"
-#include "vector.h"
+#endif
 
 /*
  * Initialize with kmeans++
@@ -47,12 +46,12 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
 
 		for (j = 0; j < numSamples; j++)
 		{
-			Datum		vec = PointerGetDatum(VectorArrayGet(samples, j));
+			Vector	   *vec = VectorArrayGet(samples, j);
 			double		distance;
 
 			/* Only need to compute distance for new center */
 			/* TODO Use triangle inequality to reduce distance calculations */
-			distance = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, vec, PointerGetDatum(VectorArrayGet(centers, i))));
+			distance = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, i))));
 
 			/* Set lower bound */
 			lowerBound[j * numCenters + i] = distance;
@@ -90,29 +89,15 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
  * Apply norm to vector
  */
 static inline void
-ApplyNorm(FmgrInfo *normprocinfo, Oid collation, Datum value, IvfflatType type)
+ApplyNorm(FmgrInfo *normprocinfo, Oid collation, Vector * vec)
 {
-	double		norm = DatumGetFloat8(FunctionCall1Coll(normprocinfo, collation, value));
+	double		norm = DatumGetFloat8(FunctionCall1Coll(normprocinfo, collation, PointerGetDatum(vec)));
 
 	/* TODO Handle zero norm */
 	if (norm > 0)
 	{
-		if (type == IVFFLAT_TYPE_VECTOR)
-		{
-			Vector	   *vec = DatumGetVector(value);
-
-			for (int i = 0; i < vec->dim; i++)
-				vec->x[i] /= norm;
-		}
-		else if (type == IVFFLAT_TYPE_HALFVEC)
-		{
-			HalfVector *vec = DatumGetHalfVector(value);
-
-			for (int i = 0; i < vec->dim; i++)
-				vec->x[i] = Float4ToHalfUnchecked(HalfToFloat4(vec->x[i]) / norm);
-		}
-		else
-			elog(ERROR, "Unsupported type");
+		for (int i = 0; i < vec->dim; i++)
+			vec->x[i] /= norm;
 	}
 }
 
@@ -125,20 +110,11 @@ CompareVectors(const void *a, const void *b)
 	return vector_cmp_internal((Vector *) a, (Vector *) b);
 }
 
-/*
- * Compare half vectors
- */
-static int
-CompareHalfVectors(const void *a, const void *b)
-{
-	return halfvec_cmp_internal((HalfVector *) a, (HalfVector *) b);
-}
-
 /*
  * Quick approach if we have little data
  */
 static void
-QuickCenters(Relation index, VectorArray samples, VectorArray centers, IvfflatType type)
+QuickCenters(Relation index, VectorArray samples, VectorArray centers)
 {
 	int			dimensions = centers->dim;
 	Oid			collation = index->rd_indcollation[0];
@@ -147,20 +123,14 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers, IvfflatTy
 	/* Copy existing vectors while avoiding duplicates */
 	if (samples->length > 0)
 	{
-		if (type == IVFFLAT_TYPE_VECTOR)
-			qsort(samples->items, samples->length, samples->itemsize, CompareVectors);
-		else if (type == IVFFLAT_TYPE_HALFVEC)
-			qsort(samples->items, samples->length, samples->itemsize, CompareHalfVectors);
-		else
-			elog(ERROR, "Unsupported type");
-
+		qsort(samples->items, samples->length, VECTOR_SIZE(samples->dim), CompareVectors);
 		for (int i = 0; i < samples->length; i++)
 		{
-			Datum		vec = PointerGetDatum(VectorArrayGet(samples, i));
+			Vector	   *vec = VectorArrayGet(samples, i);
 
-			if (i == 0 || !datumIsEqual(vec, PointerGetDatum(VectorArrayGet(samples, i - 1)), false, -1))
+			if (i == 0 || CompareVectors(vec, VectorArrayGet(samples, i - 1)) != 0)
 			{
-				VectorArraySet(centers, centers->length, DatumGetPointer(vec));
+				VectorArraySet(centers, centers->length, vec);
 				centers->length++;
 			}
 		}
@@ -169,34 +139,17 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers, IvfflatTy
 	/* Fill remaining with random data */
 	while (centers->length < centers->maxlen)
 	{
-		Datum		center = PointerGetDatum(VectorArrayGet(centers, centers->length));
+		Vector	   *vec = VectorArrayGet(centers, centers->length);
 
-		if (type == IVFFLAT_TYPE_VECTOR)
-		{
-			Vector	   *vec = DatumGetVector(center);
+		SET_VARSIZE(vec, VECTOR_SIZE(dimensions));
+		vec->dim = dimensions;
 
-			SET_VARSIZE(vec, VECTOR_SIZE(dimensions));
-			vec->dim = dimensions;
-
-			for (int j = 0; j < dimensions; j++)
-				vec->x[j] = RandomDouble();
-		}
-		else if (type == IVFFLAT_TYPE_HALFVEC)
-		{
-			HalfVector *vec = DatumGetHalfVector(center);
-
-			SET_VARSIZE(vec, HALFVEC_SIZE(dimensions));
-			vec->dim = dimensions;
-
-			for (int j = 0; j < dimensions; j++)
-				vec->x[j] = Float4ToHalfUnchecked((float) RandomDouble());
-		}
-		else
-			elog(ERROR, "Unsupported type");
+		for (int j = 0; j < dimensions; j++)
+			vec->x[j] = RandomDouble();
 
 		/* Normalize if needed (only needed for random centers) */
 		if (normprocinfo != NULL)
-			ApplyNorm(normprocinfo, collation, center, type);
+			ApplyNorm(normprocinfo, collation, vec);
 
 		centers->length++;
 	}
@@ -207,120 +160,18 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers, IvfflatTy
  * Show memory usage
  */
 static void
-ShowMemoryUsage(MemoryContext context, Size estimatedSize)
+ShowMemoryUsage(Size estimatedSize)
 {
 #if PG_VERSION_NUM >= 130000
 	elog(INFO, "total memory: %zu MB",
-		 MemoryContextMemAllocated(context, true) / (1024 * 1024));
+		 MemoryContextMemAllocated(CurrentMemoryContext, true) / (1024 * 1024));
 #else
-	MemoryContextStats(context);
+	MemoryContextStats(CurrentMemoryContext);
 #endif
 	elog(INFO, "estimated memory: %zu MB", estimatedSize / (1024 * 1024));
 }
 #endif
 
-/*
- * Compute new centers
- */
-static void
-ComputeNewCenters(VectorArray samples, VectorArray aggCenters, VectorArray newCenters, int *centerCounts, int *closestCenters, FmgrInfo *normprocinfo, Oid collation, IvfflatType type)
-{
-	int			dimensions = aggCenters->dim;
-	int			numCenters = aggCenters->maxlen;
-	int			numSamples = samples->length;
-
-	/* Reset sum and count */
-	for (int j = 0; j < numCenters; j++)
-	{
-		Vector	   *vec = (Vector *) VectorArrayGet(aggCenters, j);
-
-		for (int k = 0; k < dimensions; k++)
-			vec->x[k] = 0.0;
-
-		centerCounts[j] = 0;
-	}
-
-	/* Increment sum of closest center */
-	if (type == IVFFLAT_TYPE_VECTOR)
-	{
-		for (int j = 0; j < numSamples; j++)
-		{
-			Vector	   *aggCenter = (Vector *) VectorArrayGet(aggCenters, closestCenters[j]);
-			Vector	   *vec = (Vector *) VectorArrayGet(samples, j);
-
-			for (int k = 0; k < dimensions; k++)
-				aggCenter->x[k] += vec->x[k];
-		}
-	}
-	else if (type == IVFFLAT_TYPE_HALFVEC)
-	{
-		for (int j = 0; j < numSamples; j++)
-		{
-			Vector	   *aggCenter = (Vector *) VectorArrayGet(aggCenters, closestCenters[j]);
-			HalfVector *vec = (HalfVector *) VectorArrayGet(samples, j);
-
-			for (int k = 0; k < dimensions; k++)
-				aggCenter->x[k] += HalfToFloat4(vec->x[k]);
-		}
-	}
-	else
-		elog(ERROR, "Unsupported type");
-
-	/* Increment count of closest center */
-	for (int j = 0; j < numSamples; j++)
-		centerCounts[closestCenters[j]] += 1;
-
-	/* Divide sum by count */
-	for (int j = 0; j < numCenters; j++)
-	{
-		Vector	   *vec = (Vector *) VectorArrayGet(aggCenters, j);
-
-		if (centerCounts[j] > 0)
-		{
-			/* Double avoids overflow, but requires more memory */
-			/* TODO Update bounds */
-			for (int k = 0; k < dimensions; k++)
-			{
-				if (isinf(vec->x[k]))
-					vec->x[k] = vec->x[k] > 0 ? FLT_MAX : -FLT_MAX;
-			}
-
-			for (int k = 0; k < dimensions; k++)
-				vec->x[k] /= centerCounts[j];
-		}
-		else
-		{
-			/* TODO Handle empty centers properly */
-			for (int k = 0; k < dimensions; k++)
-				vec->x[k] = RandomDouble();
-		}
-	}
-
-	/* Set new centers if different from agg centers */
-	if (type == IVFFLAT_TYPE_HALFVEC)
-	{
-		for (int j = 0; j < numCenters; j++)
-		{
-			Vector	   *aggCenter = (Vector *) VectorArrayGet(aggCenters, j);
-			HalfVector *newCenter = (HalfVector *) VectorArrayGet(newCenters, j);
-
-			for (int k = 0; k < dimensions; k++)
-				newCenter->x[k] = Float4ToHalfUnchecked(aggCenter->x[k]);
-		}
-	}
-
-	/* Normalize if needed */
-	if (normprocinfo != NULL)
-	{
-		for (int j = 0; j < numCenters; j++)
-		{
-			Datum		newCenter = PointerGetDatum(VectorArrayGet(newCenters, j));
-
-			ApplyNorm(normprocinfo, collation, newCenter, type);
-		}
-	}
-}
-
 /*
  * Use Elkan for performance. This requires distance function to satisfy triangle inequality.
  *
@@ -330,16 +181,19 @@ ComputeNewCenters(VectorArray samples, VectorArray aggCenters, VectorArray newCe
  * https://www.aaai.org/Papers/ICML/2003/ICML03-022.pdf
  */
 static void
-ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatType type)
+ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
 {
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
 	Oid			collation;
+	Vector	   *vec;
+	Vector	   *newCenter;
+	int64		j;
+	int64		k;
 	int			dimensions = centers->dim;
 	int			numCenters = centers->maxlen;
 	int			numSamples = samples->length;
 	VectorArray newCenters;
-	VectorArray aggCenters;
 	int		   *centerCounts;
 	int		   *closestCenters;
 	float	   *lowerBound;
@@ -347,14 +201,11 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 	float	   *s;
 	float	   *halfcdist;
 	float	   *newcdist;
-	MemoryContext kmeansCtx;
-	MemoryContext oldCtx;
 
 	/* Calculate allocation sizes */
-	Size		samplesSize = VECTOR_ARRAY_SIZE(samples->maxlen, samples->itemsize);
-	Size		centersSize = VECTOR_ARRAY_SIZE(centers->maxlen, centers->itemsize);
-	Size		newCentersSize = VECTOR_ARRAY_SIZE(numCenters, centers->itemsize);
-	Size		aggCentersSize = type == IVFFLAT_TYPE_VECTOR ? 0 : VECTOR_ARRAY_SIZE(numCenters, VECTOR_SIZE(dimensions));
+	Size		samplesSize = VECTOR_ARRAY_SIZE(samples->maxlen, samples->dim);
+	Size		centersSize = VECTOR_ARRAY_SIZE(centers->maxlen, centers->dim);
+	Size		newCentersSize = VECTOR_ARRAY_SIZE(numCenters, dimensions);
 	Size		centerCountsSize = sizeof(int) * numCenters;
 	Size		closestCentersSize = sizeof(int) * numSamples;
 	Size		lowerBoundSize = sizeof(float) * numSamples * numCenters;
@@ -364,7 +215,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 	Size		newcdistSize = sizeof(float) * numCenters;
 
 	/* Calculate total size */
-	Size		totalSize = samplesSize + centersSize + newCentersSize + aggCentersSize + centerCountsSize + closestCentersSize + lowerBoundSize + upperBoundSize + sSize + halfcdistSize + newcdistSize;
+	Size		totalSize = samplesSize + centersSize + newCentersSize + centerCountsSize + closestCentersSize + lowerBoundSize + upperBoundSize + sSize + halfcdistSize + newcdistSize;
 
 	/* Check memory requirements */
 	/* Add one to error message to ceil */
@@ -383,12 +234,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 	normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
 	collation = index->rd_indcollation[0];
 
-	/* Use memory context */
-	kmeansCtx = AllocSetContextCreate(CurrentMemoryContext,
-									  "Ivfflat kmeans temporary context",
-									  ALLOCSET_DEFAULT_SIZES);
-	oldCtx = MemoryContextSwitchTo(kmeansCtx);
-
 	/* Allocate space */
 	/* Use float instead of double to save memory */
 	centerCounts = palloc(centerCountsSize);
@@ -399,50 +244,29 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 	halfcdist = palloc_extended(halfcdistSize, MCXT_ALLOC_HUGE);
 	newcdist = palloc(newcdistSize);
 
-	aggCenters = VectorArrayInit(numCenters, dimensions, VECTOR_SIZE(dimensions));
-	for (int j = 0; j < numCenters; j++)
+	newCenters = VectorArrayInit(numCenters, dimensions);
+	for (j = 0; j < numCenters; j++)
 	{
-		Vector	   *vec = (Vector *) VectorArrayGet(aggCenters, j);
-
+		vec = VectorArrayGet(newCenters, j);
 		SET_VARSIZE(vec, VECTOR_SIZE(dimensions));
 		vec->dim = dimensions;
 	}
 
-	if (type == IVFFLAT_TYPE_VECTOR)
-	{
-		/* Use same centers to save memory */
-		newCenters = aggCenters;
-	}
-	else if (type == IVFFLAT_TYPE_HALFVEC)
-	{
-		newCenters = VectorArrayInit(numCenters, dimensions, centers->itemsize);
-
-		for (int j = 0; j < numCenters; j++)
-		{
-			HalfVector *vec = (HalfVector *) VectorArrayGet(newCenters, j);
-
-			SET_VARSIZE(vec, HALFVEC_SIZE(dimensions));
-			vec->dim = dimensions;
-		}
-	}
-	else
-		elog(ERROR, "Unsupported type");
-
 #ifdef IVFFLAT_MEMORY
-	ShowMemoryUsage(oldCtx, totalSize);
+	ShowMemoryUsage(totalSize);
 #endif
 
 	/* Pick initial centers */
 	InitCenters(index, samples, centers, lowerBound);
 
 	/* Assign each x to its closest initial center c(x) = argmin d(x,c) */
-	for (int64 j = 0; j < numSamples; j++)
+	for (j = 0; j < numSamples; j++)
 	{
 		float		minDistance = FLT_MAX;
 		int			closestCenter = 0;
 
 		/* Find closest center */
-		for (int64 k = 0; k < numCenters; k++)
+		for (k = 0; k < numCenters; k++)
 		{
 			/* TODO Use Lemma 1 in k-means++ initialization */
 			float		distance = lowerBound[j * numCenters + k];
@@ -468,13 +292,13 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 		CHECK_FOR_INTERRUPTS();
 
 		/* Step 1: For all centers, compute distance */
-		for (int64 j = 0; j < numCenters; j++)
+		for (j = 0; j < numCenters; j++)
 		{
-			Datum		vec = PointerGetDatum(VectorArrayGet(centers, j));
+			vec = VectorArrayGet(centers, j);
 
-			for (int64 k = j + 1; k < numCenters; k++)
+			for (k = j + 1; k < numCenters; k++)
 			{
-				float		distance = 0.5 * DatumGetFloat8(FunctionCall2Coll(procinfo, collation, vec, PointerGetDatum(VectorArrayGet(centers, k))));
+				float		distance = 0.5 * DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, k))));
 
 				halfcdist[j * numCenters + k] = distance;
 				halfcdist[k * numCenters + j] = distance;
@@ -482,11 +306,11 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 		}
 
 		/* For all centers c, compute s(c) */
-		for (int64 j = 0; j < numCenters; j++)
+		for (j = 0; j < numCenters; j++)
 		{
 			float		minDistance = FLT_MAX;
 
-			for (int64 k = 0; k < numCenters; k++)
+			for (k = 0; k < numCenters; k++)
 			{
 				float		distance;
 
@@ -503,7 +327,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 
 		rjreset = iteration != 0;
 
-		for (int64 j = 0; j < numSamples; j++)
+		for (j = 0; j < numSamples; j++)
 		{
 			bool		rj;
 
@@ -513,9 +337,8 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 
 			rj = rjreset;
 
-			for (int64 k = 0; k < numCenters; k++)
+			for (k = 0; k < numCenters; k++)
 			{
-				Datum		vec;
 				float		dxcx;
 
 				/* Step 3: For all remaining points x and centers c */
@@ -528,12 +351,12 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 				if (upperBound[j] <= halfcdist[closestCenters[j] * numCenters + k])
 					continue;
 
-				vec = PointerGetDatum(VectorArrayGet(samples, j));
+				vec = VectorArrayGet(samples, j);
 
 				/* Step 3a */
 				if (rj)
 				{
-					dxcx = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, vec, PointerGetDatum(VectorArrayGet(centers, closestCenters[j]))));
+					dxcx = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, closestCenters[j]))));
 
 					/* d(x,c(x)) computed, which is a form of d(x,c) */
 					lowerBound[j * numCenters + closestCenters[j]] = dxcx;
@@ -547,7 +370,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 				/* Step 3b */
 				if (dxcx > lowerBound[j * numCenters + k] || dxcx > halfcdist[closestCenters[j] * numCenters + k])
 				{
-					float		dxc = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, vec, PointerGetDatum(VectorArrayGet(centers, k))));
+					float		dxc = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, k))));
 
 					/* d(x,c) calculated */
 					lowerBound[j * numCenters + k] = dxc;
@@ -566,15 +389,66 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 		}
 
 		/* Step 4: For each center c, let m(c) be mean of all points assigned */
-		ComputeNewCenters(samples, aggCenters, newCenters, centerCounts, closestCenters, normprocinfo, collation, type);
+		for (j = 0; j < numCenters; j++)
+		{
+			vec = VectorArrayGet(newCenters, j);
+			for (k = 0; k < dimensions; k++)
+				vec->x[k] = 0.0;
+
+			centerCounts[j] = 0;
+		}
+
+		for (j = 0; j < numSamples; j++)
+		{
+			int			closestCenter;
+
+			vec = VectorArrayGet(samples, j);
+			closestCenter = closestCenters[j];
+
+			/* Increment sum and count of closest center */
+			newCenter = VectorArrayGet(newCenters, closestCenter);
+			for (k = 0; k < dimensions; k++)
+				newCenter->x[k] += vec->x[k];
+
+			centerCounts[closestCenter] += 1;
+		}
+
+		for (j = 0; j < numCenters; j++)
+		{
+			vec = VectorArrayGet(newCenters, j);
+
+			if (centerCounts[j] > 0)
+			{
+				/* Double avoids overflow, but requires more memory */
+				/* TODO Update bounds */
+				for (k = 0; k < dimensions; k++)
+				{
+					if (isinf(vec->x[k]))
+						vec->x[k] = vec->x[k] > 0 ? FLT_MAX : -FLT_MAX;
+				}
+
+				for (k = 0; k < dimensions; k++)
+					vec->x[k] /= centerCounts[j];
+			}
+			else
+			{
+				/* TODO Handle empty centers properly */
+				for (k = 0; k < dimensions; k++)
+					vec->x[k] = RandomDouble();
+			}
+
+			/* Normalize if needed */
+			if (normprocinfo != NULL)
+				ApplyNorm(normprocinfo, collation, vec);
+		}
 
 		/* Step 5 */
-		for (int j = 0; j < numCenters; j++)
+		for (j = 0; j < numCenters; j++)
 			newcdist[j] = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(VectorArrayGet(centers, j)), PointerGetDatum(VectorArrayGet(newCenters, j))));
 
-		for (int64 j = 0; j < numSamples; j++)
+		for (j = 0; j < numSamples; j++)
 		{
-			for (int64 k = 0; k < numCenters; k++)
+			for (k = 0; k < numCenters; k++)
 			{
 				float		distance = lowerBound[j * numCenters + k] - newcdist[k];
 
@@ -587,26 +461,32 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 
 		/* Step 6 */
 		/* We reset r(x) before Step 3 in the next iteration */
-		for (int j = 0; j < numSamples; j++)
+		for (j = 0; j < numSamples; j++)
 			upperBound[j] += newcdist[closestCenters[j]];
 
 		/* Step 7 */
-		for (int j = 0; j < numCenters; j++)
+		for (j = 0; j < numCenters; j++)
 			VectorArraySet(centers, j, VectorArrayGet(newCenters, j));
 
 		if (changes == 0 && iteration != 0)
 			break;
 	}
 
-	MemoryContextSwitchTo(oldCtx);
-	MemoryContextDelete(kmeansCtx);
+	VectorArrayFree(newCenters);
+	pfree(centerCounts);
+	pfree(closestCenters);
+	pfree(lowerBound);
+	pfree(upperBound);
+	pfree(s);
+	pfree(halfcdist);
+	pfree(newcdist);
 }
 
 /*
  * Detect issues with centers
  */
 static void
-CheckCenters(Relation index, VectorArray centers, IvfflatType type)
+CheckCenters(Relation index, VectorArray centers)
 {
 	FmgrInfo   *normprocinfo;
 
@@ -616,48 +496,24 @@ CheckCenters(Relation index, VectorArray centers, IvfflatType type)
 	/* Ensure no NaN or infinite values */
 	for (int i = 0; i < centers->length; i++)
 	{
-		if (type == IVFFLAT_TYPE_VECTOR)
+		Vector	   *vec = VectorArrayGet(centers, i);
+
+		for (int j = 0; j < vec->dim; j++)
 		{
-			Vector	   *vec = (Vector *) VectorArrayGet(centers, i);
+			if (isnan(vec->x[j]))
+				elog(ERROR, "NaN detected. Please report a bug.");
 
-			for (int j = 0; j < vec->dim; j++)
-			{
-				if (isnan(vec->x[j]))
-					elog(ERROR, "NaN detected. Please report a bug.");
-
-				if (isinf(vec->x[j]))
-					elog(ERROR, "Infinite value detected. Please report a bug.");
-			}
+			if (isinf(vec->x[j]))
+				elog(ERROR, "Infinite value detected. Please report a bug.");
 		}
-		else if (type == IVFFLAT_TYPE_HALFVEC)
-		{
-			HalfVector *vec = (HalfVector *) VectorArrayGet(centers, i);
-
-			for (int j = 0; j < vec->dim; j++)
-			{
-				if (HalfIsNan(vec->x[j]))
-					elog(ERROR, "NaN detected. Please report a bug.");
-
-				if (HalfIsInf(vec->x[j]))
-					elog(ERROR, "Infinite value detected. Please report a bug.");
-			}
-		}
-		else
-			elog(ERROR, "Unsupported type");
 	}
 
 	/* Ensure no duplicate centers */
 	/* Fine to sort in-place */
-	if (type == IVFFLAT_TYPE_VECTOR)
-		qsort(centers->items, centers->length, centers->itemsize, CompareVectors);
-	else if (type == IVFFLAT_TYPE_HALFVEC)
-		qsort(centers->items, centers->length, centers->itemsize, CompareHalfVectors);
-	else
-		elog(ERROR, "Unsupported type");
-
+	qsort(centers->items, centers->length, VECTOR_SIZE(centers->dim), CompareVectors);
 	for (int i = 1; i < centers->length; i++)
 	{
-		if (datumIsEqual(PointerGetDatum(VectorArrayGet(centers, i)), PointerGetDatum(VectorArrayGet(centers, i - 1)), false, -1))
+		if (CompareVectors(VectorArrayGet(centers, i), VectorArrayGet(centers, i - 1)) == 0)
 			elog(ERROR, "Duplicate centers detected. Please report a bug.");
 	}
 
@@ -683,12 +539,12 @@ CheckCenters(Relation index, VectorArray centers, IvfflatType type)
  * We use spherical k-means for inner product and cosine
  */
 void
-IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatType type)
+IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers)
 {
 	if (samples->length <= centers->maxlen)
-		QuickCenters(index, samples, centers, type);
+		QuickCenters(index, samples, centers);
 	else
-		ElkanKmeans(index, samples, centers, type);
+		ElkanKmeans(index, samples, centers);
 
-	CheckCenters(index, centers, type);
+	CheckCenters(index, centers);
 }

src/ivfscan.c

@@ -5,7 +5,6 @@
 #include "access/relscan.h"
 #include "catalog/pg_operator_d.h"
 #include "catalog/pg_type_d.h"
-#include "halfvec.h"
 #include "lib/pairingheap.h"
 #include "ivfflat.h"
 #include "miscadmin.h"
@@ -178,42 +177,6 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	tuplesort_performsort(so->sortstate);
 }
 
-/*
- * Get scan value
- */
-static Datum
-GetScanValue(IndexScanDesc scan)
-{
-	IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
-	Datum		value;
-
-	if (scan->orderByData->sk_flags & SK_ISNULL)
-	{
-		IvfflatType type = IvfflatGetType(scan->indexRelation);
-
-		if (type == IVFFLAT_TYPE_VECTOR)
-			value = PointerGetDatum(InitVector(so->dimensions));
-		else if (type == IVFFLAT_TYPE_HALFVEC)
-			value = PointerGetDatum(InitHalfVector(so->dimensions));
-		else
-			elog(ERROR, "Unsupported type");
-	}
-	else
-	{
-		value = scan->orderByData->sk_argument;
-
-		/* Value should not be compressed or toasted */
-		Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
-		Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
-
-		/* Fine if normalization fails */
-		if (so->normprocinfo != NULL)
-			IvfflatNormValue(so->normprocinfo, so->collation, &value, IvfflatGetType(scan->indexRelation));
-	}
-
-	return value;
-}
-
 /*
  * Prepare for an index scan
  */
@@ -318,7 +281,21 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
 		if (!IsMVCCSnapshot(scan->xs_snapshot))
 			elog(ERROR, "non-MVCC snapshots are not supported with ivfflat");
 
-		value = GetScanValue(scan);
+		if (scan->orderByData->sk_flags & SK_ISNULL)
+			value = PointerGetDatum(InitVector(so->dimensions));
+		else
+		{
+			value = scan->orderByData->sk_argument;
+
+			/* Value should not be compressed or toasted */
+			Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
+			Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
+
+			/* Fine if normalization fails */
+			if (so->normprocinfo != NULL)
+				IvfflatNormValue(so->normprocinfo, so->collation, &value);
+		}
+
 		IvfflatBench("GetScanLists", GetScanLists(scan, value));
 		IvfflatBench("GetScanItems", GetScanItems(scan, value));
 		so->first = false;

src/ivfutils.c

@@ -1,27 +1,22 @@
 #include "postgres.h"
 
 #include "access/generic_xlog.h"
-#include "catalog/pg_type.h"
-#include "halfutils.h"
-#include "halfvec.h"
 #include "ivfflat.h"
 #include "storage/bufmgr.h"
 #include "vector.h"
-#include "utils/syscache.h"
 
 /*
  * Allocate a vector array
  */
 VectorArray
-VectorArrayInit(int maxlen, int dimensions, Size itemsize)
+VectorArrayInit(int maxlen, int dimensions)
 {
 	VectorArray res = palloc(sizeof(VectorArrayData));
 
 	res->length = 0;
 	res->maxlen = maxlen;
 	res->dim = dimensions;
-	res->itemsize = itemsize;
-	res->items = palloc_extended(maxlen * itemsize, MCXT_ALLOC_ZERO | MCXT_ALLOC_HUGE);
+	res->items = palloc_extended(maxlen * VECTOR_SIZE(dimensions), MCXT_ALLOC_ZERO | MCXT_ALLOC_HUGE);
 	return res;
 }
 
@@ -35,6 +30,16 @@ VectorArrayFree(VectorArray arr)
 	pfree(arr);
 }
 
+/*
+ * Print vector array - useful for debugging
+ */
+void
+PrintVectorArray(char *msg, VectorArray arr)
+{
+	for (int i = 0; i < arr->length; i++)
+		PrintVector(msg, VectorArrayGet(arr, i));
+}
+
 /*
  * Get the number of lists in the index
  */
@@ -61,37 +66,6 @@ IvfflatOptionalProcInfo(Relation index, uint16 procnum)
 	return index_getprocinfo(index, 1, procnum);
 }
 
-/*
- * Get type
- */
-IvfflatType
-IvfflatGetType(Relation index)
-{
-	Oid			typid = TupleDescAttr(index->rd_att, 0)->atttypid;
-	HeapTuple	tuple;
-	Form_pg_type type;
-	IvfflatType result;
-
-	tuple = SearchSysCache1(TYPEOID, ObjectIdGetDatum(typid));
-	if (!HeapTupleIsValid(tuple))
-		elog(ERROR, "cache lookup failed for type %u", typid);
-
-	type = (Form_pg_type) GETSTRUCT(tuple);
-	if (strcmp(NameStr(type->typname), "vector") == 0)
-		result = IVFFLAT_TYPE_VECTOR;
-	else if (strcmp(NameStr(type->typname), "halfvec") == 0)
-		result = IVFFLAT_TYPE_HALFVEC;
-	else
-	{
-		ReleaseSysCache(tuple);
-		elog(ERROR, "type not supported for ivfflat index");
-	}
-
-	ReleaseSysCache(tuple);
-
-	return result;
-}
-
 /*
  * Divide by the norm
  *
@@ -101,34 +75,19 @@ IvfflatGetType(Relation index)
  * if it's different than the original value
  */
 bool
-IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, IvfflatType type)
+IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value)
 {
 	double		norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
 
 	if (norm > 0)
 	{
-		if (type == IVFFLAT_TYPE_VECTOR)
-		{
-			Vector	   *v = DatumGetVector(*value);
-			Vector	   *result = InitVector(v->dim);
+		Vector	   *v = DatumGetVector(*value);
+		Vector	   *result = InitVector(v->dim);
 
-			for (int i = 0; i < v->dim; i++)
-				result->x[i] = v->x[i] / norm;
+		for (int i = 0; i < v->dim; i++)
+			result->x[i] = v->x[i] / norm;
 
-			*value = PointerGetDatum(result);
-		}
-		else if (type == IVFFLAT_TYPE_HALFVEC)
-		{
-			HalfVector *v = DatumGetHalfVector(*value);
-			HalfVector *result = InitHalfVector(v->dim);
-
-			for (int i = 0; i < v->dim; i++)
-				result->x[i] = Float4ToHalfUnchecked(HalfToFloat4(v->x[i]) / norm);
-
-			*value = PointerGetDatum(result);
-		}
-		else
-			elog(ERROR, "Unsupported type");
+		*value = PointerGetDatum(result);
 
 		return true;
 	}

src/sparsevec.c

@@ -68,12 +68,7 @@ CheckNnz(int nnz, int dim)
 	if (nnz < 0)
 		ereport(ERROR,
 				(errcode(ERRCODE_DATA_EXCEPTION),
-				 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)));
+				 errmsg("sparsevec must have at least one element")));
 
 	if (nnz > dim)
 		ereport(ERROR,
@@ -89,15 +84,15 @@ CheckIndex(int32 *indices, int i, int dim)
 {
 	int32		index = indices[i];
 
-	if (index < 1)
+	if (index < 0)
 		ereport(ERROR,
 				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("index must be greater than zero")));
+				 errmsg("index must not be negative")));
 
-	if (index > dim)
+	if (index >= dim)
 		ereport(ERROR,
 				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("index must be less than or equal to dimensions")));
+				 errmsg("index must be less than dimensions")));
 
 	if (i > 0)
 	{
@@ -174,16 +169,19 @@ sparsevec_in(PG_FUNCTION_ARGS)
 	char	   *lit = PG_GETARG_CSTRING(0);
 	int32		typmod = PG_GETARG_INT32(2);
 	int			dim;
-	char	   *pt = lit;
+	char	   *pt;
 	char	   *stringEnd;
 	SparseVector *result;
 	float	   *rvalues;
+	char	   *litcopy = pstrdup(lit);
+	char	   *str = litcopy;
 	int32	   *indices;
 	float	   *values;
 	int			maxNnz;
 	int			nnz = 0;
 
 	maxNnz = 1;
+	pt = str;
 	while (*pt != '\0')
 	{
 		if (*pt == ',')
@@ -192,145 +190,124 @@ 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));
 
-	pt = lit;
+	while (sparsevec_isspace(*str))
+		str++;
 
-	while (sparsevec_isspace(*pt))
-		pt++;
-
-	if (*pt != '{')
+	if (*str != '{')
 		ereport(ERROR,
 				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-				 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit),
+				 errmsg("malformed sparsevec literal: \"%s\"", lit),
 				 errdetail("Vector contents must start with \"{\".")));
 
-	pt++;
+	str++;
+	pt = strtok(str, ",");
+	stringEnd = pt;
 
-	while (sparsevec_isspace(*pt))
-		pt++;
-
-	if (*pt == '}')
-		pt++;
-	else
+	while (pt != NULL && *stringEnd != '}')
 	{
-		for (;;)
-		{
-			long		index;
-			float		value;
+		long		index;
+		float		value;
 
-			/* TODO Better error */
-			if (nnz == maxNnz)
-				ereport(ERROR,
-						(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-						 errmsg("ran out of buffer: \"%s\"", lit)));
-
-			while (sparsevec_isspace(*pt))
-				pt++;
-
-			/* Check for empty string like float4in */
-			if (*pt == '\0')
-				ereport(ERROR,
-						(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-						 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
-
-			/* Use similar logic as int2vectorin */
-			errno = 0;
-			index = strtol(pt, &stringEnd, 10);
-
-			if (stringEnd == pt)
-				ereport(ERROR,
-						(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-						 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
-
-			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)));
-
-			pt = stringEnd;
-
-			while (sparsevec_isspace(*pt))
-				pt++;
-
-			if (*pt != ':')
-				ereport(ERROR,
-						(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-						 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
+		/* TODO Better error */
+		if (nnz == maxNnz)
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("ran out of buffer: \"%s\"", lit)));
 
+		while (sparsevec_isspace(*pt))
 			pt++;
 
-			while (sparsevec_isspace(*pt))
-				pt++;
+		/* Check for empty string like float4in */
+		if (*pt == '\0')
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
 
-			errno = 0;
+		/* Use similar logic as int2vectorin */
+		errno = 0;
+		index = strtol(pt, &stringEnd, 10);
 
-			/* Use strtof like float4in to avoid a double-rounding problem */
-			/* Postgres sets LC_NUMERIC to C on startup */
-			value = strtof(pt, &stringEnd);
+		if (stringEnd == pt)
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
 
-			if (stringEnd == pt)
-				ereport(ERROR,
-						(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-						 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
+		if (errno == ERANGE || index < 0 || index > INT_MAX)
+			ereport(ERROR,
+					(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
+					 errmsg("index \"%ld\" is out of range for type sparsevec", index)));
 
-			/* Check for range error like float4in */
-			if (errno == ERANGE && (value == 0 || isinf(value)))
-				ereport(ERROR,
-						(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
-						 errmsg("\"%s\" is out of range for type sparsevec", pnstrdup(pt, stringEnd - pt))));
+		if (stringEnd == pt)
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
 
-			CheckElement(value);
+		while (sparsevec_isspace(*stringEnd))
+			stringEnd++;
 
-			/* Do not store zero values */
-			if (value != 0)
-			{
-				indices[nnz] = index;
-				values[nnz] = value;
-				nnz++;
-			}
+		if (*stringEnd != ':')
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
 
-			pt = stringEnd;
+		stringEnd++;
 
-			while (sparsevec_isspace(*pt))
-				pt++;
+		while (sparsevec_isspace(*stringEnd))
+			stringEnd++;
 
-			if (*pt == ',')
-				pt++;
-			else if (*pt == '}')
-			{
-				pt++;
-				break;
-			}
-			else
-				ereport(ERROR,
-						(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-						 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
+		errno = 0;
+		pt = stringEnd;
+		value = strtof(pt, &stringEnd);
+
+		if (stringEnd == pt)
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
+
+		/* Check for range error like float4in */
+		if (errno == ERANGE && (value == 0 || isinf(value)))
+			ereport(ERROR,
+					(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
+					 errmsg("\"%s\" is out of range for type sparsevec", pt)));
+
+		/* TODO Decide whether to store zero values */
+		if (value != 0)
+		{
+			indices[nnz] = index;
+			values[nnz] = value;
+			nnz++;
 		}
+
+		if (*stringEnd != '\0' && *stringEnd != '}')
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
+
+		pt = strtok(NULL, ",");
 	}
 
-	while (sparsevec_isspace(*pt))
-		pt++;
-
-	if (*pt != '/')
+	if (stringEnd == NULL || *stringEnd != '}')
 		ereport(ERROR,
 				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-				 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit),
+				 errmsg("malformed sparsevec literal: \"%s\"", lit),
 				 errdetail("Unexpected end of input.")));
 
-	pt++;
+	stringEnd++;
 
-	while (sparsevec_isspace(*pt))
-		pt++;
+	if (*stringEnd != '/')
+		ereport(ERROR,
+				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+				 errmsg("malformed sparsevec literal: \"%s\"", lit),
+				 errdetail("Unexpected end of input.")));
+
+	stringEnd++;
 
 	/* Use similar logic as int2vectorin */
 	errno = 0;
+	pt = stringEnd;
 	dim = strtol(pt, &stringEnd, 10);
 
 	if (stringEnd == pt)
@@ -338,18 +315,18 @@ sparsevec_in(PG_FUNCTION_ARGS)
 				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
 				 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
 
-	pt = stringEnd;
-
 	/* Only whitespace is allowed after the closing brace */
-	while (sparsevec_isspace(*pt))
-		pt++;
+	while (sparsevec_isspace(*stringEnd))
+		stringEnd++;
 
-	if (*pt != '\0')
+	if (*stringEnd != '\0')
 		ereport(ERROR,
 				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-				 errmsg("invalid input syntax for type sparsevec: \"%s\"", lit),
+				 errmsg("malformed sparsevec literal: \"%s\"", lit),
 				 errdetail("Junk after closing.")));
 
+	pfree(litcopy);
+
 	CheckDim(dim);
 	CheckExpectedDim(typmod, dim);
 
@@ -361,6 +338,7 @@ sparsevec_in(PG_FUNCTION_ARGS)
 		rvalues[i] = values[i];
 
 		CheckIndex(result->indices, i, dim);
+		CheckElement(rvalues[i]);
 	}
 
 	PG_RETURN_POINTER(result);
@@ -582,7 +560,7 @@ vector_to_sparsevec(PG_FUNCTION_ARGS)
 			if (j == nnz)
 				elog(ERROR, "safety check failed");
 
-			result->indices[j] = i + 1;
+			result->indices[j] = i;
 			values[j] = vec->x[i];
 			j++;
 		}
@@ -592,10 +570,10 @@ vector_to_sparsevec(PG_FUNCTION_ARGS)
 }
 
 /*
- * Get the L2 squared distance between sparse vectors
+ *  Get the L2 squared distance between sparse vectors
  */
 static double
-SparsevecL2SquaredDistance(SparseVector * a, SparseVector * b)
+l2_distance_squared_internal(SparseVector * a, SparseVector * b)
 {
 	float	   *ax = SPARSEVEC_VALUES(a);
 	float	   *bx = SPARSEVEC_VALUES(b);
@@ -651,7 +629,7 @@ sparsevec_l2_distance(PG_FUNCTION_ARGS)
 
 	CheckDims(a, b);
 
-	PG_RETURN_FLOAT8(sqrt(SparsevecL2SquaredDistance(a, b)));
+	PG_RETURN_FLOAT8(sqrt(l2_distance_squared_internal(a, b)));
 }
 
 /*
@@ -667,14 +645,14 @@ sparsevec_l2_squared_distance(PG_FUNCTION_ARGS)
 
 	CheckDims(a, b);
 
-	PG_RETURN_FLOAT8(SparsevecL2SquaredDistance(a, b));
+	PG_RETURN_FLOAT8(l2_distance_squared_internal(a, b));
 }
 
 /*
  * Get the inner product of two sparse vectors
  */
 static double
-SparsevecInnerProduct(SparseVector * a, SparseVector * b)
+inner_product_internal(SparseVector * a, SparseVector * b)
 {
 	float	   *ax = SPARSEVEC_VALUES(a);
 	float	   *bx = SPARSEVEC_VALUES(b);
@@ -718,7 +696,7 @@ sparsevec_inner_product(PG_FUNCTION_ARGS)
 
 	CheckDims(a, b);
 
-	PG_RETURN_FLOAT8(SparsevecInnerProduct(a, b));
+	PG_RETURN_FLOAT8(inner_product_internal(a, b));
 }
 
 /*
@@ -733,7 +711,7 @@ sparsevec_negative_inner_product(PG_FUNCTION_ARGS)
 
 	CheckDims(a, b);
 
-	PG_RETURN_FLOAT8(-SparsevecInnerProduct(a, b));
+	PG_RETURN_FLOAT8(-inner_product_internal(a, b));
 }
 
 /*
@@ -753,7 +731,7 @@ sparsevec_cosine_distance(PG_FUNCTION_ARGS)
 
 	CheckDims(a, b);
 
-	similarity = SparsevecInnerProduct(a, b);
+	similarity = inner_product_internal(a, b);
 
 	/* Auto-vectorized */
 	for (int i = 0; i < a->nnz; i++)
@@ -798,3 +776,55 @@ 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

@@ -2,7 +2,6 @@
 #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

@@ -6,7 +6,6 @@
 #include "catalog/pg_type.h"
 #include "common/shortest_dec.h"
 #include "fmgr.h"
-#include "halfutils.h"
 #include "halfvec.h"
 #include "hnsw.h"
 #include "ivfflat.h"
@@ -33,13 +32,6 @@
 #define STATE_DIMS(x) (ARR_DIMS(x)[0] - 1)
 #define CreateStateDatums(dim) palloc(sizeof(Datum) * (dim + 1))
 
-/* target_clones requires glibc */
-#if defined(__x86_64__) && defined(__gnu_linux__) && defined(__has_attribute) && __has_attribute(target_clones) && !defined(__FMA__)
-#define VECTOR_DISPATCH __attribute__((target_clones("default", "fma")))
-#else
-#define VECTOR_DISPATCH
-#endif
-
 PG_MODULE_MAGIC;
 
 /*
@@ -49,7 +41,6 @@ PGDLLEXPORT void _PG_init(void);
 void
 _PG_init(void)
 {
-	HalfvecInit();
 	HnswInit();
 	IvfflatInit();
 }
@@ -188,32 +179,28 @@ vector_in(PG_FUNCTION_ARGS)
 	int32		typmod = PG_GETARG_INT32(2);
 	float		x[VECTOR_MAX_DIM];
 	int			dim = 0;
-	char	   *pt = lit;
+	char	   *pt;
+	char	   *stringEnd;
 	Vector	   *result;
+	char	   *litcopy = pstrdup(lit);
+	char	   *str = litcopy;
 
-	while (vector_isspace(*pt))
-		pt++;
+	while (vector_isspace(*str))
+		str++;
 
-	if (*pt != '[')
+	if (*str != '[')
 		ereport(ERROR,
 				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-				 errmsg("invalid input syntax for type vector: \"%s\"", lit),
+				 errmsg("malformed vector literal: \"%s\"", lit),
 				 errdetail("Vector contents must start with \"[\".")));
 
-	pt++;
+	str++;
+	pt = strtok(str, ",");
+	stringEnd = pt;
 
-	while (vector_isspace(*pt))
-		pt++;
-
-	if (*pt == ']')
-		ereport(ERROR,
-				(errcode(ERRCODE_DATA_EXCEPTION),
-				 errmsg("vector must have at least 1 dimension")));
-
-	for (;;)
+	while (pt != NULL && *stringEnd != ']')
 	{
 		float		val;
-		char	   *stringEnd;
 
 		if (dim == VECTOR_MAX_DIM)
 			ereport(ERROR,
@@ -229,10 +216,8 @@ vector_in(PG_FUNCTION_ARGS)
 					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
 					 errmsg("invalid input syntax for type vector: \"%s\"", lit)));
 
-		errno = 0;
-
 		/* Use strtof like float4in to avoid a double-rounding problem */
-		/* Postgres sets LC_NUMERIC to C on startup */
+		errno = 0;
 		val = strtof(pt, &stringEnd);
 
 		if (stringEnd == pt)
@@ -244,40 +229,56 @@ vector_in(PG_FUNCTION_ARGS)
 		if (errno == ERANGE && isinf(val))
 			ereport(ERROR,
 					(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
-					 errmsg("\"%s\" is out of range for type vector", pnstrdup(pt, stringEnd - pt))));
+					 errmsg("\"%s\" is out of range for type vector", pt)));
 
 		CheckElement(val);
 		x[dim++] = val;
 
-		pt = stringEnd;
+		while (vector_isspace(*stringEnd))
+			stringEnd++;
 
-		while (vector_isspace(*pt))
-			pt++;
-
-		if (*pt == ',')
-			pt++;
-		else if (*pt == ']')
-		{
-			pt++;
-			break;
-		}
-		else
+		if (*stringEnd != '\0' && *stringEnd != ']')
 			ereport(ERROR,
 					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
 					 errmsg("invalid input syntax for type vector: \"%s\"", lit)));
+
+		pt = strtok(NULL, ",");
 	}
 
-	/* Only whitespace is allowed after the closing brace */
-	while (vector_isspace(*pt))
-		pt++;
-
-	if (*pt != '\0')
+	if (stringEnd == NULL || *stringEnd != ']')
 		ereport(ERROR,
 				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
-				 errmsg("invalid input syntax for type vector: \"%s\"", lit),
+				 errmsg("malformed vector literal: \"%s\"", lit),
+				 errdetail("Unexpected end of input.")));
+
+	stringEnd++;
+
+	/* Only whitespace is allowed after the closing brace */
+	while (vector_isspace(*stringEnd))
+		stringEnd++;
+
+	if (*stringEnd != '\0')
+		ereport(ERROR,
+				(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+				 errmsg("malformed vector literal: \"%s\"", lit),
 				 errdetail("Junk after closing right brace.")));
 
-	CheckDim(dim);
+	/* Ensure no consecutive delimiters since strtok skips */
+	for (pt = lit + 1; *pt != '\0'; pt++)
+	{
+		if (pt[-1] == ',' && *pt == ',')
+			ereport(ERROR,
+					(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
+					 errmsg("malformed vector literal: \"%s\"", lit)));
+	}
+
+	if (dim < 1)
+		ereport(ERROR,
+				(errcode(ERRCODE_DATA_EXCEPTION),
+				 errmsg("vector must have at least 1 dimension")));
+
+	pfree(litcopy);
+
 	CheckExpectedDim(typmod, dim);
 
 	result = InitVector(dim);
@@ -564,22 +565,6 @@ halfvec_to_vector(PG_FUNCTION_ARGS)
 	PG_RETURN_POINTER(result);
 }
 
-VECTOR_DISPATCH static float
-VectorL2SquaredDistance(int dim, float *ax, float *bx)
-{
-	float		distance = 0.0;
-
-	/* Auto-vectorized */
-	for (int i = 0; i < dim; i++)
-	{
-		float		diff = ax[i] - bx[i];
-
-		distance += diff * diff;
-	}
-
-	return distance;
-}
-
 /*
  * Get the L2 distance between vectors
  */
@@ -589,10 +574,21 @@ l2_distance(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	float	   *ax = a->x;
+	float	   *bx = b->x;
+	float		distance = 0.0;
+	float		diff;
 
 	CheckDims(a, b);
 
-	PG_RETURN_FLOAT8(sqrt((double) VectorL2SquaredDistance(a->dim, a->x, b->x)));
+	/* Auto-vectorized */
+	for (int i = 0; i < a->dim; i++)
+	{
+		diff = ax[i] - bx[i];
+		distance += diff * diff;
+	}
+
+	PG_RETURN_FLOAT8(sqrt((double) distance));
 }
 
 /*
@@ -605,22 +601,21 @@ vector_l2_squared_distance(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	float	   *ax = a->x;
+	float	   *bx = b->x;
+	float		distance = 0.0;
+	float		diff;
 
 	CheckDims(a, b);
 
-	PG_RETURN_FLOAT8((double) VectorL2SquaredDistance(a->dim, a->x, b->x));
-}
-
-VECTOR_DISPATCH static float
-VectorInnerProduct(int dim, float *ax, float *bx)
-{
-	float		distance = 0.0;
-
 	/* Auto-vectorized */
-	for (int i = 0; i < dim; i++)
-		distance += ax[i] * bx[i];
+	for (int i = 0; i < a->dim; i++)
+	{
+		diff = ax[i] - bx[i];
+		distance += diff * diff;
+	}
 
-	return distance;
+	PG_RETURN_FLOAT8((double) distance);
 }
 
 /*
@@ -632,10 +627,17 @@ inner_product(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	float	   *ax = a->x;
+	float	   *bx = b->x;
+	float		distance = 0.0;
 
 	CheckDims(a, b);
 
-	PG_RETURN_FLOAT8((double) VectorInnerProduct(a->dim, a->x, b->x));
+	/* Auto-vectorized */
+	for (int i = 0; i < a->dim; i++)
+		distance += ax[i] * bx[i];
+
+	PG_RETURN_FLOAT8((double) distance);
 }
 
 /*
@@ -647,10 +649,17 @@ vector_negative_inner_product(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	float	   *ax = a->x;
+	float	   *bx = b->x;
+	float		distance = 0.0;
 
 	CheckDims(a, b);
 
-	PG_RETURN_FLOAT8((double) -VectorInnerProduct(a->dim, a->x, b->x));
+	/* Auto-vectorized */
+	for (int i = 0; i < a->dim; i++)
+		distance += ax[i] * bx[i];
+
+	PG_RETURN_FLOAT8((double) distance * -1);
 }
 
 /*
@@ -708,11 +717,18 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	float	   *ax = a->x;
+	float	   *bx = b->x;
+	float		dp = 0.0;
 	double		distance;
 
 	CheckDims(a, b);
 
-	distance = (double) VectorInnerProduct(a->dim, a->x, b->x);
+	/* Auto-vectorized */
+	for (int i = 0; i < a->dim; i++)
+		dp += ax[i] * bx[i];
+
+	distance = (double) dp;
 
 	/* Prevent NaN with acos with loss of precision */
 	if (distance > 1)
@@ -1228,7 +1244,7 @@ sparsevec_to_vector(PG_FUNCTION_ARGS)
 
 	result = InitVector(dim);
 	for (int i = 0; i < svec->nnz; i++)
-		result->x[svec->indices[i] - 1] = values[i];
+		result->x[svec->indices[i]] = values[i];
 
 	PG_RETURN_POINTER(result);
 }

test/expected/bit_functions.out

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

@@ -28,26 +28,6 @@ SELECT ARRAY[1,2,3]::numeric[]::vector;
  [1,2,3]
 (1 row)
 
-SELECT '[1,2,3]'::vector::real[];
- float4  
----------
- {1,2,3}
-(1 row)
-
-SELECT '{1,2,3}'::real[]::vector;
- vector  
----------
- [1,2,3]
-(1 row)
-
-SELECT '{1,2,3}'::real[]::vector(3);
- vector  
----------
- [1,2,3]
-(1 row)
-
-SELECT '{1,2,3}'::real[]::vector(2);
-ERROR:  expected 2 dimensions, not 3
 SELECT '{NULL}'::real[]::vector;
 ERROR:  array must not contain nulls
 SELECT '{NaN}'::real[]::vector;
@@ -60,26 +40,10 @@ SELECT '{}'::real[]::vector;
 ERROR:  vector must have at least 1 dimension
 SELECT '{{1}}'::real[]::vector;
 ERROR:  array must be 1-D
-SELECT '{1,2,3}'::double precision[]::vector;
- vector  
+SELECT '[1,2,3]'::vector::real[];
+ float4  
 ---------
- [1,2,3]
-(1 row)
-
-SELECT '{1,2,3}'::double precision[]::vector(3);
- vector  
----------
- [1,2,3]
-(1 row)
-
-SELECT '{1,2,3}'::double precision[]::vector(2);
-ERROR:  expected 2 dimensions, not 3
-SELECT '{4e38,-4e38}'::double precision[]::vector;
-ERROR:  infinite value not allowed in vector
-SELECT '{1e-46,-1e-46}'::double precision[]::vector;
- vector 
---------
- [0,-0]
+ {1,2,3}
 (1 row)
 
 SELECT '[1,2,3]'::vector::halfvec;
@@ -88,88 +52,24 @@ SELECT '[1,2,3]'::vector::halfvec;
  [1,2,3]
 (1 row)
 
-SELECT '[1,2,3]'::vector::halfvec(3);
- halfvec 
----------
- [1,2,3]
-(1 row)
-
-SELECT '[1,2,3]'::vector::halfvec(2);
-ERROR:  expected 2 dimensions, not 3
-SELECT '[65520]'::vector::halfvec;
-ERROR:  "65520" is out of range for type halfvec
-SELECT '[1e-8]'::vector::halfvec;
- halfvec 
----------
- [0]
-(1 row)
-
 SELECT '[1,2,3]'::halfvec::vector;
  vector  
 ---------
  [1,2,3]
 (1 row)
 
-SELECT '[1,2,3]'::halfvec::vector(3);
- vector  
----------
- [1,2,3]
-(1 row)
-
+SELECT '[1,2,3]'::vector::halfvec(2);
+ERROR:  expected 2 dimensions, not 3
 SELECT '[1,2,3]'::halfvec::vector(2);
 ERROR:  expected 2 dimensions, not 3
-SELECT '{1,2,3}'::real[]::halfvec;
+SELECT '[65520]'::vector::halfvec;
+ERROR:  infinite value not allowed in halfvec
+SELECT '[1e-8]'::vector::halfvec;
  halfvec 
 ---------
- [1,2,3]
+ [0]
 (1 row)
 
-SELECT '{1,2,3}'::real[]::halfvec(3);
- halfvec 
----------
- [1,2,3]
-(1 row)
-
-SELECT '{1,2,3}'::real[]::halfvec(2);
-ERROR:  expected 2 dimensions, not 3
-SELECT '{65520,-65520}'::real[]::halfvec;
-ERROR:  "65520" is out of range for type halfvec
-SELECT '{1e-8,-1e-8}'::real[]::halfvec;
- halfvec 
----------
- [0,-0]
-(1 row)
-
-SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec;
-    sparsevec    
------------------
- {2:1.5,4:3.5}/5
-(1 row)
-
-SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec(5);
-    sparsevec    
------------------
- {2:1.5,4:3.5}/5
-(1 row)
-
-SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec(4);
-ERROR:  expected 4 dimensions, not 5
-SELECT '{2:1.5,4:3.5}/5'::sparsevec::vector;
-     vector      
------------------
- [0,1.5,0,3.5,0]
-(1 row)
-
-SELECT '{2:1.5,4:3.5}/5'::sparsevec::vector(5);
-     vector      
------------------
- [0,1.5,0,3.5,0]
-(1 row)
-
-SELECT '{2:1.5,4:3.5}/5'::sparsevec::vector(4);
-ERROR:  expected 4 dimensions, not 5
-SELECT '{}/16001'::sparsevec::vector;
-ERROR:  vector cannot have more than 16000 dimensions
 SELECT array_agg(n)::vector FROM generate_series(1, 16001) n;
 ERROR:  vector cannot have more than 16000 dimensions
 SELECT array_to_vector(array_agg(n), 16001, false) FROM generate_series(1, 16001) n;

test/expected/copy.out

@@ -1,15 +1,15 @@
-CREATE TABLE t (val vector(3), val2 halfvec(3), val3 sparsevec(3));
-INSERT INTO t (val, val2, val3) VALUES ('[0,0,0]', '[0,0,0]', '{}/3'), ('[1,2,3]', '[1,2,3]', '{1:1,2:2,3:3}/3'), ('[1,1,1]', '[1,1,1]', '{1:1,2:1,3:1}/3'), (NULL, NULL, NULL);
-CREATE TABLE t2 (val vector(3), val2 halfvec(3), val3 sparsevec(3));
+CREATE TABLE t (val vector(3), val2 halfvec(3));
+INSERT INTO t (val, val2) VALUES ('[0,0,0]', '[0,0,0]'), ('[1,2,3]', '[1,2,3]'), ('[1,1,1]', '[1,1,1]'), (NULL, NULL);
+CREATE TABLE t2 (val vector(3), val2 halfvec(3));
 \copy t TO 'results/data.bin' WITH (FORMAT binary)
 \copy t2 FROM 'results/data.bin' WITH (FORMAT binary)
 SELECT * FROM t2 ORDER BY val;
-   val   |  val2   |      val3       
----------+---------+-----------------
- [0,0,0] | [0,0,0] | {}/3
- [1,1,1] | [1,1,1] | {1:1,2:1,3:1}/3
- [1,2,3] | [1,2,3] | {1:1,2:2,3:3}/3
-         |         | 
+   val   |  val2   
+---------+---------
+ [0,0,0] | [0,0,0]
+ [1,1,1] | [1,1,1]
+ [1,2,3] | [1,2,3]
+         | 
 (4 rows)
 
 DROP TABLE t;

test/expected/halfvec_functions.out

@@ -1,21 +1,3 @@
-SELECT round(halfvec_norm('[1,1]')::numeric, 5);
-  round  
----------
- 1.41421
-(1 row)
-
-SELECT halfvec_norm('[3,4]');
- halfvec_norm 
---------------
-            5
-(1 row)
-
-SELECT halfvec_norm('[0,1]');
- halfvec_norm 
---------------
-            1
-(1 row)
-
 SELECT l2_distance('[0,0]'::halfvec, '[3,4]');
  l2_distance 
 -------------
@@ -30,12 +12,6 @@ SELECT l2_distance('[0,0]'::halfvec, '[0,1]');
 
 SELECT l2_distance('[1,2]'::halfvec, '[3]');
 ERROR:  different halfvec dimensions 2 and 1
-SELECT l2_distance('[1,1,1,1,1,1,1,1,1]'::halfvec, '[1,1,1,1,1,1,1,4,5]');
- l2_distance 
--------------
-           5
-(1 row)
-
 SELECT '[0,0]'::halfvec <-> '[3,4]';
  ?column? 
 ----------
@@ -56,12 +32,6 @@ SELECT inner_product('[65504]'::halfvec, '[65504]');
     4290774016
 (1 row)
 
-SELECT inner_product('[1,1,1,1,1,1,1,1,1]'::halfvec, '[1,2,3,4,5,6,7,8,9]');
- inner_product 
----------------
-            45
-(1 row)
-
 SELECT '[1,2]'::halfvec <#> '[3,4]';
  ?column? 
 ----------

test/expected/halfvec_input.out

@@ -51,68 +51,51 @@ SELECT '[65519,-65519]'::halfvec;
 (1 row)
 
 SELECT '[65520,-65520]'::halfvec;
-ERROR:  "65520" is out of range for type halfvec
+ERROR:  value out of range: overflow
 LINE 1: SELECT '[65520,-65520]'::halfvec;
                ^
 SELECT '[1e-8,-1e-8]'::halfvec;
- halfvec 
----------
- [0,-0]
-(1 row)
-
+ERROR:  value out of range: underflow
+LINE 1: SELECT '[1e-8,-1e-8]'::halfvec;
+               ^
 SELECT '[4e38,1]'::halfvec;
-ERROR:  "4e38" is out of range for type halfvec
+ERROR:  infinite value not allowed in halfvec
 LINE 1: SELECT '[4e38,1]'::halfvec;
                ^
-SELECT '[1e-46,1]'::halfvec;
- halfvec 
----------
- [0,1]
-(1 row)
-
 SELECT '[1,2,3'::halfvec;
-ERROR:  invalid input syntax for type halfvec: "[1,2,3"
+ERROR:  malformed halfvec literal: "[1,2,3"
 LINE 1: SELECT '[1,2,3'::halfvec;
                ^
+DETAIL:  Unexpected end of input.
 SELECT '[1,2,3]9'::halfvec;
-ERROR:  invalid input syntax for type halfvec: "[1,2,3]9"
+ERROR:  malformed halfvec literal: "[1,2,3]9"
 LINE 1: SELECT '[1,2,3]9'::halfvec;
                ^
 DETAIL:  Junk after closing right brace.
 SELECT '1,2,3'::halfvec;
-ERROR:  invalid input syntax for type halfvec: "1,2,3"
+ERROR:  malformed halfvec literal: "1,2,3"
 LINE 1: SELECT '1,2,3'::halfvec;
                ^
 DETAIL:  Vector contents must start with "[".
 SELECT ''::halfvec;
-ERROR:  invalid input syntax for type halfvec: ""
+ERROR:  malformed halfvec literal: ""
 LINE 1: SELECT ''::halfvec;
                ^
 DETAIL:  Vector contents must start with "[".
 SELECT '['::halfvec;
-ERROR:  invalid input syntax for type halfvec: "["
+ERROR:  malformed halfvec literal: "["
 LINE 1: SELECT '['::halfvec;
                ^
-SELECT '[ '::halfvec;
-ERROR:  invalid input syntax for type halfvec: "[ "
-LINE 1: SELECT '[ '::halfvec;
-               ^
+DETAIL:  Unexpected end of input.
 SELECT '[,'::halfvec;
-ERROR:  invalid input syntax for type halfvec: "[,"
+ERROR:  malformed halfvec literal: "[,"
 LINE 1: SELECT '[,'::halfvec;
                ^
+DETAIL:  Unexpected end of input.
 SELECT '[]'::halfvec;
 ERROR:  halfvec must have at least 1 dimension
 LINE 1: SELECT '[]'::halfvec;
                ^
-SELECT '[ ]'::halfvec;
-ERROR:  halfvec must have at least 1 dimension
-LINE 1: SELECT '[ ]'::halfvec;
-               ^
-SELECT '[,]'::halfvec;
-ERROR:  invalid input syntax for type halfvec: "[,]"
-LINE 1: SELECT '[,]'::halfvec;
-               ^
 SELECT '[1,]'::halfvec;
 ERROR:  invalid input syntax for type halfvec: "[1,]"
 LINE 1: SELECT '[1,]'::halfvec;
@@ -122,7 +105,7 @@ ERROR:  invalid input syntax for type halfvec: "[1a]"
 LINE 1: SELECT '[1a]'::halfvec;
                ^
 SELECT '[1,,3]'::halfvec;
-ERROR:  invalid input syntax for type halfvec: "[1,,3]"
+ERROR:  malformed halfvec literal: "[1,,3]"
 LINE 1: SELECT '[1,,3]'::halfvec;
                ^
 SELECT '[1, ,3]'::halfvec;

test/expected/hnsw_bit_hamming.out

@@ -19,11 +19,3 @@ 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);
-CREATE INDEX ON t USING hnsw ((val::bit(64001)) bit_hamming_ops);
-ERROR:  column cannot have more than 64000 dimensions for hnsw index
-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'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
+INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
 CREATE INDEX ON t USING hnsw (val sparsevec_cosine_ops);
-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';
+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';
        val       
 -----------------
- {1:1,2:1,3:1}/3
- {1:1,2:2,3:3}/3
- {1:1,2:2,3:4}/3
+ {0:1,1:1,2:1}/3
+ {0:1,1:2,2:3}/3
+ {0:1,1:2,2:4}/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'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
+INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
 CREATE INDEX ON t USING hnsw (val sparsevec_ip_ops);
-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';
+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';
        val       
 -----------------
- {1:1,2:2,3:4}/3
- {1:1,2:2,3:3}/3
- {1:1,2:1,3:1}/3
+ {0:1,1:2,2:4}/3
+ {0:1,1:2,2:3}/3
+ {0:1,1:1,2:1}/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'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
+INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
 CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
-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';
+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';
        val       
 -----------------
- {1:1,2:2,3:3}/3
- {1:1,2:2,3:4}/3
- {1:1,2:1,3:1}/3
+ {0:1,1:2,2:3}/3
+ {0:1,1:2,2:4}/3
+ {0:1,1:1,2:1}/3
  {}/3
 (4 rows)
 
@@ -25,7 +25,7 @@ SELECT COUNT(*) FROM t;
 (1 row)
 
 TRUNCATE t;
-SELECT * FROM t ORDER BY val <-> '{1:3,2:3,3:3}/3';
+SELECT * FROM t ORDER BY val <-> '{0:3,1:3,2:3}/3';
  val 
 -----
 (0 rows)

test/expected/ivfflat_halfvec_cosine.out

@@ -1,26 +0,0 @@
-SET enable_seqscan = off;
-CREATE TABLE t (val halfvec(3));
-INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
-CREATE INDEX ON t USING ivfflat (val halfvec_cosine_ops) WITH (lists = 1);
-INSERT INTO t (val) VALUES ('[1,2,4]');
-SELECT * FROM t ORDER BY val <=> '[3,3,3]';
-   val   
----------
- [1,1,1]
- [1,2,3]
- [1,2,4]
-(3 rows)
-
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
- count 
--------
-     3
-(1 row)
-
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::halfvec)) t2;
- count 
--------
-     3
-(1 row)
-
-DROP TABLE t;

test/expected/ivfflat_halfvec_ip.out

@@ -1,21 +0,0 @@
-SET enable_seqscan = off;
-CREATE TABLE t (val halfvec(3));
-INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
-CREATE INDEX ON t USING ivfflat (val halfvec_ip_ops) WITH (lists = 1);
-INSERT INTO t (val) VALUES ('[1,2,4]');
-SELECT * FROM t ORDER BY val <#> '[3,3,3]';
-   val   
----------
- [1,2,4]
- [1,2,3]
- [1,1,1]
- [0,0,0]
-(4 rows)
-
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::halfvec)) t2;
- count 
--------
-     4
-(1 row)
-
-DROP TABLE t;

test/expected/ivfflat_halfvec_l2.out

@@ -1,36 +0,0 @@
-SET enable_seqscan = off;
-CREATE TABLE t (val halfvec(3));
-INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
-CREATE INDEX ON t USING ivfflat (val halfvec_l2_ops) WITH (lists = 1);
-INSERT INTO t (val) VALUES ('[1,2,4]');
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
-   val   
----------
- [1,2,3]
- [1,2,4]
- [1,1,1]
- [0,0,0]
-(4 rows)
-
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::halfvec)) t2;
- count 
--------
-     4
-(1 row)
-
-SELECT COUNT(*) FROM t;
- count 
--------
-     5
-(1 row)
-
-TRUNCATE t;
-NOTICE:  ivfflat index created with little data
-DETAIL:  This will cause low recall.
-HINT:  Drop the index until the table has more data.
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
- val 
------
-(0 rows)
-
-DROP TABLE t;

test/expected/ivfflat_vector_l2.out

@@ -12,11 +12,14 @@ SELECT * FROM t ORDER BY val <-> '[3,3,3]';
  [0,0,0]
 (4 rows)
 
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector)) t2;
- count 
--------
-     4
-(1 row)
+SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
+   val   
+---------
+ [0,0,0]
+ [1,1,1]
+ [1,2,3]
+ [1,2,4]
+(4 rows)
 
 SELECT COUNT(*) FROM t;
  count 

test/expected/sparsevec_functions.out

@@ -1,76 +1,52 @@
-SELECT round(sparsevec_norm('{1:1,2:1}/2')::numeric, 5);
-  round  
----------
- 1.41421
-(1 row)
-
-SELECT sparsevec_norm('{1:3,2:4}/2');
- sparsevec_norm 
-----------------
-              5
-(1 row)
-
-SELECT sparsevec_norm('{2:1}/2');
- sparsevec_norm 
-----------------
-              1
-(1 row)
-
-SELECT sparsevec_norm('{1:3e37,2:4e37}/2')::real;
- sparsevec_norm 
-----------------
-          5e+37
-(1 row)
-
-SELECT l2_distance('{}/2'::sparsevec, '{1:3,2:4}/2');
+SELECT l2_distance('{}/2'::sparsevec, '{0:3,1:4}/2');
  l2_distance 
 -------------
            5
 (1 row)
 
-SELECT l2_distance('{}/2'::sparsevec, '{2:1}/2');
+SELECT l2_distance('{}/2'::sparsevec, '{1:1}/2');
  l2_distance 
 -------------
            1
 (1 row)
 
-SELECT '{}/2'::sparsevec <-> '{1:3,2:4}/2';
+SELECT '{}/2'::sparsevec <-> '{0:3,1:4}/2';
  ?column? 
 ----------
         5
 (1 row)
 
-SELECT inner_product('{1:1,2:2}/2'::sparsevec, '{1:2,2:4}/2');
+SELECT inner_product('{0:1,1:2}/2'::sparsevec, '{0:2,1:4}/2');
  inner_product 
 ---------------
             10
 (1 row)
 
-SELECT sparsevec_negative_inner_product('{1:1,2:2}/2', '{1:2,2:4}/2');
+SELECT sparsevec_negative_inner_product('{0:1,1:2}/2', '{0:2,1:4}/2');
  sparsevec_negative_inner_product 
 ----------------------------------
                               -10
 (1 row)
 
-SELECT cosine_distance('{1:1,2:2}/2'::sparsevec, '{1:2,2:4}/2');
+SELECT cosine_distance('{0:1,1:2}/2'::sparsevec, '{0:2,1:4}/2');
  cosine_distance 
 -----------------
                0
 (1 row)
 
-SELECT cosine_distance('{1:1,2:2}/2'::sparsevec, '{}/2');
+SELECT cosine_distance('{0:1,1:2}/2'::sparsevec, '{}/2');
  cosine_distance 
 -----------------
              NaN
 (1 row)
 
-SELECT cosine_distance('{1:1,2:1}/2'::sparsevec, '{1:-1,2:-1}/2');
+SELECT cosine_distance('{0:1,1:1}/2'::sparsevec, '{0:-1,1:-1}/2');
  cosine_distance 
 -----------------
                2
 (1 row)
 
-SELECT cosine_distance('{1:2}/2'::sparsevec, '{2:2}/2');
+SELECT cosine_distance('{0:1}/2'::sparsevec, '{1:2}/2');
  cosine_distance 
 -----------------
                1
@@ -82,5 +58,35 @@ SELECT cosine_distance('{}/1'::sparsevec, '{}/1');
              NaN
 (1 row)
 
-SELECT cosine_distance('{1:2}/2'::sparsevec, '{1:1}/3');
+SELECT cosine_distance('{0:1}/2'::sparsevec, '{0:1}/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

@@ -1,171 +1,38 @@
-SELECT '{1:1.5,3:3.5}/5'::sparsevec;
+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;
     sparsevec    
 -----------------
  {1:1.5,3:3.5}/5
 (1 row)
 
-SELECT '{1:-2,3:-4}/5'::sparsevec;
-   sparsevec   
----------------
- {1:-2,3:-4}/5
-(1 row)
-
-SELECT '{1:2.,3:4.}/5'::sparsevec;
-  sparsevec  
--------------
- {1:2,3:4}/5
-(1 row)
-
-SELECT ' { 1 : 1.5 ,  3  :  3.5  } / 5 '::sparsevec;
-    sparsevec    
------------------
- {1:1.5,3:3.5}/5
-(1 row)
-
-SELECT '{1:1.23456}/1'::sparsevec;
-   sparsevec   
----------------
- {1:1.23456}/1
-(1 row)
-
-SELECT '{1:hello,2:1}/2'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{1:hello,2:1}/2"
-LINE 1: SELECT '{1:hello,2:1}/2'::sparsevec;
-               ^
-SELECT '{1:NaN,2:1}/2'::sparsevec;
-ERROR:  NaN not allowed in sparsevec
-LINE 1: SELECT '{1:NaN,2:1}/2'::sparsevec;
-               ^
-SELECT '{1:Infinity,2:1}/2'::sparsevec;
-ERROR:  infinite value not allowed in sparsevec
-LINE 1: SELECT '{1:Infinity,2:1}/2'::sparsevec;
-               ^
-SELECT '{1:-Infinity,2:1}/2'::sparsevec;
-ERROR:  infinite value not allowed in sparsevec
-LINE 1: SELECT '{1:-Infinity,2:1}/2'::sparsevec;
-               ^
-SELECT '{1:1.5e38,2:-1.5e38}/2'::sparsevec;
-        sparsevec         
---------------------------
- {1:1.5e+38,2:-1.5e+38}/2
-(1 row)
-
-SELECT '{1:1.5e+38,2:-1.5e+38}/2'::sparsevec;
-        sparsevec         
---------------------------
- {1:1.5e+38,2:-1.5e+38}/2
-(1 row)
-
-SELECT '{1:1.5e-38,2:-1.5e-38}/2'::sparsevec;
-        sparsevec         
---------------------------
- {1:1.5e-38,2:-1.5e-38}/2
-(1 row)
-
-SELECT '{1:4e38,2:1}/2'::sparsevec;
-ERROR:  "4e38" is out of range for type sparsevec
-LINE 1: SELECT '{1:4e38,2:1}/2'::sparsevec;
-               ^
-SELECT '{1:-4e38,2:1}/2'::sparsevec;
-ERROR:  "-4e38" is out of range for type sparsevec
-LINE 1: SELECT '{1:-4e38,2:1}/2'::sparsevec;
-               ^
-SELECT '{1:1e-46,2:1}/2'::sparsevec;
-ERROR:  "1e-46" is out of range for type sparsevec
-LINE 1: SELECT '{1:1e-46,2:1}/2'::sparsevec;
-               ^
-SELECT '{1:-1e-46,2:1}/2'::sparsevec;
-ERROR:  "-1e-46" is out of range for type sparsevec
-LINE 1: SELECT '{1:-1e-46,2:1}/2'::sparsevec;
-               ^
-SELECT ''::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: ""
-LINE 1: SELECT ''::sparsevec;
-               ^
-DETAIL:  Vector contents must start with "{".
-SELECT '{'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{"
-LINE 1: SELECT '{'::sparsevec;
-               ^
-SELECT '{ '::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{ "
-LINE 1: SELECT '{ '::sparsevec;
-               ^
-SELECT '{:'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{:"
-LINE 1: SELECT '{:'::sparsevec;
-               ^
-SELECT '{,'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{,"
-LINE 1: SELECT '{,'::sparsevec;
-               ^
-SELECT '{}'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{}"
-LINE 1: SELECT '{}'::sparsevec;
-               ^
-DETAIL:  Unexpected end of input.
-SELECT '{}/'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{}/"
-LINE 1: SELECT '{}/'::sparsevec;
-               ^
-SELECT '{}/1'::sparsevec;
+SELECT '{0:0,1:1,2:0}/3'::sparsevec;
  sparsevec 
 -----------
- {}/1
+ {1:1}/3
 (1 row)
 
-SELECT '{}/1a'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{}/1a"
-LINE 1: SELECT '{}/1a'::sparsevec;
-               ^
-DETAIL:  Junk after closing.
-SELECT '{ }/1'::sparsevec;
- sparsevec 
------------
- {}/1
-(1 row)
-
-SELECT '{:}/1'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{:}/1"
-LINE 1: SELECT '{:}/1'::sparsevec;
-               ^
-SELECT '{,}/1'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{,}/1"
-LINE 1: SELECT '{,}/1'::sparsevec;
-               ^
-SELECT '{1,}/1'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{1,}/1"
-LINE 1: SELECT '{1,}/1'::sparsevec;
-               ^
-SELECT '{:1}/1'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{:1}/1"
-LINE 1: SELECT '{:1}/1'::sparsevec;
-               ^
-SELECT '{1:}/1'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{1:}/1"
-LINE 1: SELECT '{1:}/1'::sparsevec;
-               ^
-SELECT '{1a:1}/1'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{1a:1}/1"
-LINE 1: SELECT '{1a:1}/1'::sparsevec;
-               ^
-SELECT '{1:1a}/1'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{1:1a}/1"
-LINE 1: SELECT '{1:1a}/1'::sparsevec;
-               ^
-SELECT '{1:1,}/1'::sparsevec;
-ERROR:  invalid input syntax for type sparsevec: "{1:1,}/1"
-LINE 1: SELECT '{1:1,}/1'::sparsevec;
-               ^
-SELECT '{1:0,2:1,3:0}/3'::sparsevec;
- sparsevec 
------------
- {2:1}/3
-(1 row)
-
-SELECT '{2:1,1:1}/2'::sparsevec;
+SELECT '{1:1,0:1}/2'::sparsevec;
 ERROR:  indexes must be in ascending order
-LINE 1: SELECT '{2:1,1:1}/2'::sparsevec;
+LINE 1: SELECT '{1:1,0:1}/2'::sparsevec;
                ^
 SELECT '{}/5'::sparsevec;
  sparsevec 
@@ -181,35 +48,15 @@ SELECT '{}/100001'::sparsevec;
 ERROR:  sparsevec cannot have more than 100000 dimensions
 LINE 1: SELECT '{}/100001'::sparsevec;
                ^
-SELECT '{0:1}/1'::sparsevec;
-ERROR:  index "0" is out of range for type sparsevec
-LINE 1: SELECT '{0:1}/1'::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 '{2:1}/1'::sparsevec;
-ERROR:  index must be less than or equal to dimensions
-LINE 1: SELECT '{2:1}/1'::sparsevec;
+SELECT '{1:1}/1'::sparsevec;
+ERROR:  index must be less than dimensions
+LINE 1: SELECT '{1:1}/1'::sparsevec;
                ^
-SELECT '{}/3'::sparsevec(3);
- sparsevec 
------------
- {}/3
-(1 row)
-
-SELECT '{}/3'::sparsevec(2);
-ERROR:  expected 2 dimensions, not 3
-SELECT '{}/3'::sparsevec(3, 2);
-ERROR:  invalid type modifier
-LINE 1: SELECT '{}/3'::sparsevec(3, 2);
-                       ^
-SELECT '{}/3'::sparsevec('a');
-ERROR:  invalid input syntax for type integer: "a"
-LINE 1: SELECT '{}/3'::sparsevec('a');
-                       ^
-SELECT '{}/3'::sparsevec(0);
-ERROR:  dimensions for type sparsevec must be at least 1
-LINE 1: SELECT '{}/3'::sparsevec(0);
-                       ^
-SELECT '{}/3'::sparsevec(100001);
-ERROR:  dimensions for type sparsevec cannot exceed 100000
-LINE 1: SELECT '{}/3'::sparsevec(100001);
-                       ^
+SELECT '{}/1'::sparsevec(2);
+ERROR:  expected 2 dimensions, not 1

test/expected/vector_input.out

@@ -83,48 +83,39 @@ SELECT '[-1e-46,1]'::vector;
 (1 row)
 
 SELECT '[1,2,3'::vector;
-ERROR:  invalid input syntax for type vector: "[1,2,3"
+ERROR:  malformed vector literal: "[1,2,3"
 LINE 1: SELECT '[1,2,3'::vector;
                ^
+DETAIL:  Unexpected end of input.
 SELECT '[1,2,3]9'::vector;
-ERROR:  invalid input syntax for type vector: "[1,2,3]9"
+ERROR:  malformed vector literal: "[1,2,3]9"
 LINE 1: SELECT '[1,2,3]9'::vector;
                ^
 DETAIL:  Junk after closing right brace.
 SELECT '1,2,3'::vector;
-ERROR:  invalid input syntax for type vector: "1,2,3"
+ERROR:  malformed vector literal: "1,2,3"
 LINE 1: SELECT '1,2,3'::vector;
                ^
 DETAIL:  Vector contents must start with "[".
 SELECT ''::vector;
-ERROR:  invalid input syntax for type vector: ""
+ERROR:  malformed vector literal: ""
 LINE 1: SELECT ''::vector;
                ^
 DETAIL:  Vector contents must start with "[".
 SELECT '['::vector;
-ERROR:  invalid input syntax for type vector: "["
+ERROR:  malformed vector literal: "["
 LINE 1: SELECT '['::vector;
                ^
-SELECT '[ '::vector;
-ERROR:  invalid input syntax for type vector: "[ "
-LINE 1: SELECT '[ '::vector;
-               ^
+DETAIL:  Unexpected end of input.
 SELECT '[,'::vector;
-ERROR:  invalid input syntax for type vector: "[,"
+ERROR:  malformed vector literal: "[,"
 LINE 1: SELECT '[,'::vector;
                ^
+DETAIL:  Unexpected end of input.
 SELECT '[]'::vector;
 ERROR:  vector must have at least 1 dimension
 LINE 1: SELECT '[]'::vector;
                ^
-SELECT '[ ]'::vector;
-ERROR:  vector must have at least 1 dimension
-LINE 1: SELECT '[ ]'::vector;
-               ^
-SELECT '[,]'::vector;
-ERROR:  invalid input syntax for type vector: "[,]"
-LINE 1: SELECT '[,]'::vector;
-               ^
 SELECT '[1,]'::vector;
 ERROR:  invalid input syntax for type vector: "[1,]"
 LINE 1: SELECT '[1,]'::vector;
@@ -134,7 +125,7 @@ ERROR:  invalid input syntax for type vector: "[1a]"
 LINE 1: SELECT '[1a]'::vector;
                ^
 SELECT '[1,,3]'::vector;
-ERROR:  invalid input syntax for type vector: "[1,,3]"
+ERROR:  malformed vector literal: "[1,,3]"
 LINE 1: SELECT '[1,,3]'::vector;
                ^
 SELECT '[1, ,3]'::vector;

test/sql/bit_functions.sql

@@ -1,21 +1,13 @@
-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 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 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));
+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');

test/sql/cast.sql

@@ -3,50 +3,19 @@ SELECT ARRAY[1.0,2.0,3.0]::vector;
 SELECT ARRAY[1,2,3]::float4[]::vector;
 SELECT ARRAY[1,2,3]::float8[]::vector;
 SELECT ARRAY[1,2,3]::numeric[]::vector;
-
-SELECT '[1,2,3]'::vector::real[];
-
-SELECT '{1,2,3}'::real[]::vector;
-SELECT '{1,2,3}'::real[]::vector(3);
-SELECT '{1,2,3}'::real[]::vector(2);
 SELECT '{NULL}'::real[]::vector;
 SELECT '{NaN}'::real[]::vector;
 SELECT '{Infinity}'::real[]::vector;
 SELECT '{-Infinity}'::real[]::vector;
 SELECT '{}'::real[]::vector;
 SELECT '{{1}}'::real[]::vector;
-
-SELECT '{1,2,3}'::double precision[]::vector;
-SELECT '{1,2,3}'::double precision[]::vector(3);
-SELECT '{1,2,3}'::double precision[]::vector(2);
-SELECT '{4e38,-4e38}'::double precision[]::vector;
-SELECT '{1e-46,-1e-46}'::double precision[]::vector;
-
+SELECT '[1,2,3]'::vector::real[];
 SELECT '[1,2,3]'::vector::halfvec;
-SELECT '[1,2,3]'::vector::halfvec(3);
+SELECT '[1,2,3]'::halfvec::vector;
 SELECT '[1,2,3]'::vector::halfvec(2);
+SELECT '[1,2,3]'::halfvec::vector(2);
 SELECT '[65520]'::vector::halfvec;
 SELECT '[1e-8]'::vector::halfvec;
-
-SELECT '[1,2,3]'::halfvec::vector;
-SELECT '[1,2,3]'::halfvec::vector(3);
-SELECT '[1,2,3]'::halfvec::vector(2);
-
-SELECT '{1,2,3}'::real[]::halfvec;
-SELECT '{1,2,3}'::real[]::halfvec(3);
-SELECT '{1,2,3}'::real[]::halfvec(2);
-SELECT '{65520,-65520}'::real[]::halfvec;
-SELECT '{1e-8,-1e-8}'::real[]::halfvec;
-
-SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec;
-SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec(5);
-SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec(4);
-
-SELECT '{2:1.5,4:3.5}/5'::sparsevec::vector;
-SELECT '{2:1.5,4:3.5}/5'::sparsevec::vector(5);
-SELECT '{2:1.5,4:3.5}/5'::sparsevec::vector(4);
-SELECT '{}/16001'::sparsevec::vector;
-
 SELECT array_agg(n)::vector FROM generate_series(1, 16001) n;
 SELECT array_to_vector(array_agg(n), 16001, false) FROM generate_series(1, 16001) n;
 

test/sql/copy.sql

@@ -1,7 +1,7 @@
-CREATE TABLE t (val vector(3), val2 halfvec(3), val3 sparsevec(3));
-INSERT INTO t (val, val2, val3) VALUES ('[0,0,0]', '[0,0,0]', '{}/3'), ('[1,2,3]', '[1,2,3]', '{1:1,2:2,3:3}/3'), ('[1,1,1]', '[1,1,1]', '{1:1,2:1,3:1}/3'), (NULL, NULL, NULL);
+CREATE TABLE t (val vector(3), val2 halfvec(3));
+INSERT INTO t (val, val2) VALUES ('[0,0,0]', '[0,0,0]'), ('[1,2,3]', '[1,2,3]'), ('[1,1,1]', '[1,1,1]'), (NULL, NULL);
 
-CREATE TABLE t2 (val vector(3), val2 halfvec(3), val3 sparsevec(3));
+CREATE TABLE t2 (val vector(3), val2 halfvec(3));
 
 \copy t TO 'results/data.bin' WITH (FORMAT binary)
 \copy t2 FROM 'results/data.bin' WITH (FORMAT binary)

test/sql/halfvec_functions.sql

@@ -1,17 +1,11 @@
-SELECT round(halfvec_norm('[1,1]')::numeric, 5);
-SELECT halfvec_norm('[3,4]');
-SELECT halfvec_norm('[0,1]');
-
 SELECT l2_distance('[0,0]'::halfvec, '[3,4]');
 SELECT l2_distance('[0,0]'::halfvec, '[0,1]');
 SELECT l2_distance('[1,2]'::halfvec, '[3]');
-SELECT l2_distance('[1,1,1,1,1,1,1,1,1]'::halfvec, '[1,1,1,1,1,1,1,4,5]');
 SELECT '[0,0]'::halfvec <-> '[3,4]';
 
 SELECT inner_product('[1,2]'::halfvec, '[3,4]');
 SELECT inner_product('[1,2]'::halfvec, '[3]');
 SELECT inner_product('[65504]'::halfvec, '[65504]');
-SELECT inner_product('[1,1,1,1,1,1,1,1,1]'::halfvec, '[1,2,3,4,5,6,7,8,9]');
 SELECT '[1,2]'::halfvec <#> '[3,4]';
 
 SELECT cosine_distance('[1,2]'::halfvec, '[2,4]');

test/sql/halfvec_input.sql

@@ -11,17 +11,13 @@ 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;
 SELECT ''::halfvec;
 SELECT '['::halfvec;
-SELECT '[ '::halfvec;
 SELECT '[,'::halfvec;
 SELECT '[]'::halfvec;
-SELECT '[ ]'::halfvec;
-SELECT '[,]'::halfvec;
 SELECT '[1,]'::halfvec;
 SELECT '[1a]'::halfvec;
 SELECT '[1,,3]'::halfvec;

test/sql/hnsw_bit_hamming.sql

@@ -10,10 +10,3 @@ 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);
-CREATE INDEX ON t USING hnsw ((val::bit(64001)) 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'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
+INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
 CREATE INDEX ON t USING hnsw (val sparsevec_cosine_ops);
 
-INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
+INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
 
-SELECT * FROM t ORDER BY val <=> '{1:3,2:3,3:3}/3';
+SELECT * FROM t ORDER BY val <=> '{0:3,1:3,2:3}/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'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
+INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
 CREATE INDEX ON t USING hnsw (val sparsevec_ip_ops);
 
-INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
+INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
 
-SELECT * FROM t ORDER BY val <#> '{1:3,2:3,3:3}/3';
+SELECT * FROM t ORDER BY val <#> '{0:3,1:3,2:3}/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'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
+INSERT INTO t (val) VALUES ('{}/3'), ('{0:1,1:2,2:3}/3'), ('{0:1,1:1,2:1}/3'), (NULL);
 CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
 
-INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
+INSERT INTO t (val) VALUES ('{0:1,1:2,2:4}/3');
 
-SELECT * FROM t ORDER BY val <-> '{1:3,2:3,3:3}/3';
+SELECT * FROM t ORDER BY val <-> '{0:3,1:3,2:3}/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 <-> '{1:3,2:3,3:3}/3';
+SELECT * FROM t ORDER BY val <-> '{0:3,1:3,2:3}/3';
 
 DROP TABLE t;
 

test/sql/ivfflat_halfvec_cosine.sql

@@ -1,13 +0,0 @@
-SET enable_seqscan = off;
-
-CREATE TABLE t (val halfvec(3));
-INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
-CREATE INDEX ON t USING ivfflat (val halfvec_cosine_ops) WITH (lists = 1);
-
-INSERT INTO t (val) VALUES ('[1,2,4]');
-
-SELECT * FROM t ORDER BY val <=> '[3,3,3]';
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::halfvec)) t2;
-
-DROP TABLE t;

test/sql/ivfflat_halfvec_ip.sql

@@ -1,12 +0,0 @@
-SET enable_seqscan = off;
-
-CREATE TABLE t (val halfvec(3));
-INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
-CREATE INDEX ON t USING ivfflat (val halfvec_ip_ops) WITH (lists = 1);
-
-INSERT INTO t (val) VALUES ('[1,2,4]');
-
-SELECT * FROM t ORDER BY val <#> '[3,3,3]';
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::halfvec)) t2;
-
-DROP TABLE t;

test/sql/ivfflat_halfvec_l2.sql

@@ -1,16 +0,0 @@
-SET enable_seqscan = off;
-
-CREATE TABLE t (val halfvec(3));
-INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
-CREATE INDEX ON t USING ivfflat (val halfvec_l2_ops) WITH (lists = 1);
-
-INSERT INTO t (val) VALUES ('[1,2,4]');
-
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::halfvec)) t2;
-SELECT COUNT(*) FROM t;
-
-TRUNCATE t;
-SELECT * FROM t ORDER BY val <-> '[3,3,3]';
-
-DROP TABLE t;

test/sql/ivfflat_vector_l2.sql

@@ -7,7 +7,7 @@ CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 1);
 INSERT INTO t (val) VALUES ('[1,2,4]');
 
 SELECT * FROM t ORDER BY val <-> '[3,3,3]';
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector)) t2;
+SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
 SELECT COUNT(*) FROM t;
 
 TRUNCATE t;

test/sql/sparsevec_functions.sql

@@ -1,18 +1,21 @@
-SELECT round(sparsevec_norm('{1:1,2:1}/2')::numeric, 5);
-SELECT sparsevec_norm('{1:3,2:4}/2');
-SELECT sparsevec_norm('{2:1}/2');
-SELECT sparsevec_norm('{1:3e37,2:4e37}/2')::real;
+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('{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('{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'::sparsevec, '{}/1');
-SELECT cosine_distance('{1:2}/2'::sparsevec, '{1:1}/3');
+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);

test/sql/sparsevec_input.sql

@@ -1,48 +1,19 @@
-SELECT '{1:1.5,3:3.5}/5'::sparsevec;
-SELECT '{1:-2,3:-4}/5'::sparsevec;
-SELECT '{1:2.,3:4.}/5'::sparsevec;
-SELECT ' { 1 : 1.5 ,  3  :  3.5  } / 5 '::sparsevec;
-SELECT '{1:1.23456}/1'::sparsevec;
-SELECT '{1:hello,2:1}/2'::sparsevec;
-SELECT '{1:NaN,2:1}/2'::sparsevec;
-SELECT '{1:Infinity,2:1}/2'::sparsevec;
-SELECT '{1:-Infinity,2:1}/2'::sparsevec;
-SELECT '{1:1.5e38,2:-1.5e38}/2'::sparsevec;
-SELECT '{1:1.5e+38,2:-1.5e+38}/2'::sparsevec;
-SELECT '{1:1.5e-38,2:-1.5e-38}/2'::sparsevec;
-SELECT '{1:4e38,2:1}/2'::sparsevec;
-SELECT '{1:-4e38,2:1}/2'::sparsevec;
-SELECT '{1:1e-46,2:1}/2'::sparsevec;
-SELECT '{1:-1e-46,2:1}/2'::sparsevec;
-SELECT ''::sparsevec;
-SELECT '{'::sparsevec;
-SELECT '{ '::sparsevec;
-SELECT '{:'::sparsevec;
-SELECT '{,'::sparsevec;
-SELECT '{}'::sparsevec;
-SELECT '{}/'::sparsevec;
-SELECT '{}/1'::sparsevec;
-SELECT '{}/1a'::sparsevec;
-SELECT '{ }/1'::sparsevec;
-SELECT '{:}/1'::sparsevec;
-SELECT '{,}/1'::sparsevec;
-SELECT '{1,}/1'::sparsevec;
-SELECT '{:1}/1'::sparsevec;
-SELECT '{1:}/1'::sparsevec;
-SELECT '{1a:1}/1'::sparsevec;
-SELECT '{1:1a}/1'::sparsevec;
-SELECT '{1:1,}/1'::sparsevec;
-SELECT '{1:0,2:1,3:0}/3'::sparsevec;
-SELECT '{2:1,1:1}/2'::sparsevec;
+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 '[0,1.5,0,3.5,0]'::vector::sparsevec;
+
+SELECT '{0:0,1:1,2:0}/3'::sparsevec;
+
+SELECT '{1:1,0:1}/2'::sparsevec;
+
 SELECT '{}/5'::sparsevec;
 SELECT '{}/-1'::sparsevec;
 SELECT '{}/100001'::sparsevec;
-SELECT '{0:1}/1'::sparsevec;
-SELECT '{2:1}/1'::sparsevec;
+SELECT '{}/16001'::sparsevec::vector;
 
-SELECT '{}/3'::sparsevec(3);
-SELECT '{}/3'::sparsevec(2);
-SELECT '{}/3'::sparsevec(3, 2);
-SELECT '{}/3'::sparsevec('a');
-SELECT '{}/3'::sparsevec(0);
-SELECT '{}/3'::sparsevec(100001);
+SELECT '{-1:1}/1'::sparsevec;
+SELECT '{1:1}/1'::sparsevec;
+
+SELECT '{}/1'::sparsevec(2);

test/sql/vector_input.sql

@@ -19,11 +19,8 @@ SELECT '[1,2,3]9'::vector;
 SELECT '1,2,3'::vector;
 SELECT ''::vector;
 SELECT '['::vector;
-SELECT '[ '::vector;
 SELECT '[,'::vector;
 SELECT '[]'::vector;
-SELECT '[ ]'::vector;
-SELECT '[,]'::vector;
 SELECT '[1,]'::vector;
 SELECT '[1a]'::vector;
 SELECT '[1,,3]'::vector;

test/t/003_ivfflat_build_recall.pl

@@ -30,19 +30,18 @@ sub test_recall
 			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]);
-		my %expected_set = map { $_ => 1 } @expected_ids;
 
-		foreach (@actual_ids)
+		foreach (@expected_ids)
 		{
-			if (exists($expected_set{$_}))
+			if (exists($actual_set{$_}))
 			{
 				$correct++;
 			}
+			$total++;
 		}
-
-		$total += $limit;
 	}
 
 	cmp_ok($correct / $total, ">=", $min, $operator);
@@ -82,12 +81,7 @@ for my $i (0 .. $#operators)
 	@expected = ();
 	foreach (@queries)
 	{
-		my $res = $node->safe_psql("postgres", qq(
-			WITH top AS (
-				SELECT v $operator '$_' AS distance FROM tst ORDER BY distance LIMIT $limit
-			)
-			SELECT i FROM tst WHERE (v $operator '$_') <= (SELECT MAX(distance) FROM top)
-		));
+		my $res = $node->safe_psql("postgres", "SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;");
 		push(@expected, $res);
 	}
 
@@ -104,16 +98,8 @@ for my $i (0 .. $#operators)
 		test_recall(1, 0.71, $operator);
 		test_recall(10, 0.95, $operator);
 	}
-
-	# Test probes equals lists
-	if ($operator eq "<=>")
-	{
-		test_recall(100, 0.9925, $operator);
-	}
-	else
-	{
-		test_recall(100, 1.00, $operator);
-	}
+	# Account for equal distances
+	test_recall(100, 0.9925, $operator);
 
 	$node->safe_psql("postgres", "DROP INDEX idx;");
 
@@ -133,16 +119,8 @@ for my $i (0 .. $#operators)
 		test_recall(1, 0.71, $operator);
 		test_recall(10, 0.95, $operator);
 	}
-
-	# Test probes equals lists
-	if ($operator eq "<=>")
-	{
-		test_recall(100, 0.9925, $operator);
-	}
-	else
-	{
-		test_recall(100, 1.00, $operator);
-	}
+	# Account for equal distances
+	test_recall(100, 0.9925, $operator);
 
 	$node->safe_psql("postgres", "DROP INDEX idx;");
 }

test/t/013_hnsw_insert_recall.pl

@@ -98,7 +98,6 @@ for my $i (0 .. $#operators)
 		push(@expected, $res);
 	}
 
-	# Test approximate results
 	my $min = $operator eq "<#>" ? 0.80 : 0.99;
 	test_recall($min, $operator);
 

test/t/016_hnsw_vacuum_recall.pl

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

test/t/020_hnsw_bit_build_recall.pl

@@ -85,7 +85,7 @@ for my $i (0 .. $#operators)
 		# Handle ties
 		my $res = $node->safe_psql("postgres", qq(
 			WITH top AS (
-				SELECT v $operator $_ AS distance FROM tst ORDER BY distance LIMIT $limit
+				SELECT v $operator $_ AS distance FROM tst ORDER BY v $operator $_ LIMIT $limit
 			)
 			SELECT i FROM tst WHERE (v $operator $_) <= (SELECT MAX(distance) FROM top)
 		));
@@ -99,7 +99,7 @@ for my $i (0 .. $#operators)
 	));
 
 	# Test approximate results
-	my $min = $operator eq "<\%>" ? 0.95 : 0.98;
+	my $min = $operator eq "<\%>" ? 0.96 : 0.98;
 	test_recall($min, $operator);
 
 	$node->safe_psql("postgres", "DROP INDEX idx;");

test/t/021_hnsw_halfvec_build_recall.pl

@@ -94,7 +94,7 @@ for my $i (0 .. $#operators)
 	));
 
 	# Test approximate results
-	my $min = $operator eq "<#>" ? 0.93 : 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

@@ -1,128 +0,0 @@
-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();

test/t/023_hnsw_bit_insert_recall.pl

@@ -1,116 +0,0 @@
-use strict;
-use warnings;
-use PostgresNode;
-use TestLib;
-use Test::More;
-
-my $node;
-my @queries = ();
-my @expected;
-my $limit = 20;
-my $dim = 52;
-my $max = 2**$dim;
-
-sub test_recall
-{
-	my ($min, $operator) = @_;
-	my $correct = 0;
-	my $total = 0;
-
-	my $explain = $node->safe_psql("postgres", qq(
-		SET enable_seqscan = off;
-		SET hnsw.ef_search = 100;
-		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;
-			SET hnsw.ef_search = 100;
-			SELECT i FROM tst ORDER BY v $operator $queries[$i] LIMIT $limit;
-		));
-		my @actual_ids = split("\n", $actual);
-
-		my @expected_ids = split("\n", $expected[$i]);
-		my %expected_set = map { $_ => 1 } @expected_ids;
-
-		foreach (@actual_ids)
-		{
-			if (exists($expected_set{$_}))
-			{
-				$correct++;
-			}
-		}
-
-		$total += $limit;
-	}
-
-	cmp_ok($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 serial, v bit($dim));");
-
-# Generate queries
-for (1 .. 20)
-{
-	my $r = int(rand() * $max);
-	push(@queries, "${r}::bigint::bit($dim)");
-}
-
-# Check each index type
-my @operators = ("<~>", "<\%>");
-my @opclasses = ("bit_hamming_ops", "bit_jaccard_ops");
-
-for my $i (0 .. $#operators)
-{
-	my $operator = $operators[$i];
-	my $opclass = $opclasses[$i];
-
-	# Add index
-	$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING hnsw (v $opclass);");
-
-	# Use concurrent inserts
-	$node->pgbench(
-		"--no-vacuum --client=10 --transactions=1000",
-		0,
-		[qr{actually processed}],
-		[qr{^$}],
-		"concurrent INSERTs",
-		{
-			"023_hnsw_bit_insert_recall_$opclass" => "INSERT INTO tst (v) VALUES ((random() * $max)::bigint::bit($dim));"
-		}
-	);
-
-	# Get exact results
-	@expected = ();
-	foreach (@queries)
-	{
-		# Handle ties
-		my $res = $node->safe_psql("postgres", qq(
-			SET enable_indexscan = off;
-			WITH top AS (
-				SELECT v $operator $_ AS distance FROM tst ORDER BY distance LIMIT $limit
-			)
-			SELECT i FROM tst WHERE (v $operator $_) <= (SELECT MAX(distance) FROM top)
-		));
-		push(@expected, $res);
-	}
-
-	# Test approximate results
-	my $min = $operator eq "<\%>" ? 0.95 : 0.98;
-	test_recall($min, $operator);
-
-	$node->safe_psql("postgres", "DROP INDEX idx;");
-	$node->safe_psql("postgres", "TRUNCATE tst;");
-}
-
-done_testing();

test/t/024_hnsw_halfvec_insert_recall.pl

@@ -1,113 +0,0 @@
-use strict;
-use warnings;
-use PostgresNode;
-use TestLib;
-use Test::More;
-
-my $node;
-my @queries = ();
-my @expected;
-my $limit = 20;
-my $dim = 10;
-my $array_sql = join(",", ('random()') x $dim);
-
-sub test_recall
-{
-	my ($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 serial, v halfvec($dim));");
-
-# Generate queries
-for (1 .. 20)
-{
-	my @r = ();
-	for (1 .. $dim)
-	{
-		push(@r, rand());
-	}
-	push(@queries, "[" . join(",", @r) . "]");
-}
-
-# Check each index type
-my @operators = ("<->", "<#>", "<=>");
-my @opclasses = ("halfvec_l2_ops", "halfvec_ip_ops", "halfvec_cosine_ops");
-
-for my $i (0 .. $#operators)
-{
-	my $operator = $operators[$i];
-	my $opclass = $opclasses[$i];
-
-	# Add index
-	$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING hnsw (v $opclass);");
-
-	# Use concurrent inserts
-	$node->pgbench(
-		"--no-vacuum --client=10 --transactions=1000",
-		0,
-		[qr{actually processed}],
-		[qr{^$}],
-		"concurrent INSERTs",
-		{
-			"024_hnsw_halfvec_insert_recall_$opclass" => "INSERT INTO tst (v) VALUES (ARRAY[$array_sql]);"
-		}
-	);
-
-	# Get exact results
-	@expected = ();
-	foreach (@queries)
-	{
-		my $res = $node->safe_psql("postgres", qq(
-			SET enable_indexscan = off;
-			SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;
-		));
-		push(@expected, $res);
-	}
-
-	# Test approximate results
-	my $min = $operator eq "<#>" ? 0.94 : 0.98;
-	test_recall($min, $operator);
-
-	$node->safe_psql("postgres", "DROP INDEX idx;");
-	$node->safe_psql("postgres", "TRUNCATE tst;");
-}
-
-done_testing();

test/t/025_hnsw_sparsevec_insert_recall.pl

@@ -1,109 +0,0 @@
-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 serial, v sparsevec(3));");
-
-# 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];
-
-	# Add index
-	$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING hnsw (v $opclass);");
-
-	# Use concurrent inserts
-	$node->pgbench(
-		"--no-vacuum --client=10 --transactions=1000",
-		0,
-		[qr{actually processed}],
-		[qr{^$}],
-		"concurrent INSERTs",
-		{
-			"025_hnsw_sparsevec_insert_recall_$opclass" => "INSERT INTO tst (v) VALUES (ARRAY[random(), random(), random()]::vector::sparsevec);"
-		}
-	);
-
-	# Get exact results
-	@expected = ();
-	foreach (@queries)
-	{
-		my $res = $node->safe_psql("postgres", qq(
-			SET enable_indexscan = off;
-			SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;
-		));
-		push(@expected, $res);
-	}
-
-	# Test approximate results
-	my $min = $operator eq "<#>" ? 0.80 : 0.99;
-	test_recall($min, $operator);
-
-	$node->safe_psql("postgres", "DROP INDEX idx;");
-	$node->safe_psql("postgres", "TRUNCATE tst;");
-}
-
-done_testing();

test/t/026_hnsw_bit_duplicates.pl

@@ -1,58 +0,0 @@
-use strict;
-use warnings;
-use PostgresNode;
-use TestLib;
-use Test::More;
-
-# Initialize node
-my $node = get_new_node('node');
-$node->init;
-$node->start;
-
-# Create table
-$node->safe_psql("postgres", "CREATE EXTENSION vector;");
-$node->safe_psql("postgres", "CREATE TABLE tst (v bit(3));");
-
-sub insert_vectors
-{
-	for my $i (1 .. 20)
-	{
-		$node->safe_psql("postgres", "INSERT INTO tst VALUES ('111');");
-	}
-}
-
-sub test_duplicates
-{
-	my $res = $node->safe_psql("postgres", qq(
-		SET enable_seqscan = off;
-		SET hnsw.ef_search = 1;
-		SELECT COUNT(*) FROM (SELECT * FROM tst ORDER BY v <~> '111') t;
-	));
-	is($res, 10);
-}
-
-# Test duplicates with build
-insert_vectors();
-$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING hnsw (v bit_hamming_ops);");
-test_duplicates();
-
-# Reset
-$node->safe_psql("postgres", "TRUNCATE tst;");
-
-# Test duplicates with inserts
-insert_vectors();
-test_duplicates();
-
-# Test fallback path for inserts
-$node->pgbench(
-	"--no-vacuum --client=5 --transactions=100",
-	0,
-	[qr{actually processed}],
-	[qr{^$}],
-	"concurrent INSERTs",
-	{
-		"026_hnsw_bit_duplicates" => "INSERT INTO tst VALUES ('111');"
-	}
-);
-
-done_testing();

test/t/027_hnsw_halfvec_duplicates.pl

@@ -1,58 +0,0 @@
-use strict;
-use warnings;
-use PostgresNode;
-use TestLib;
-use Test::More;
-
-# Initialize node
-my $node = get_new_node('node');
-$node->init;
-$node->start;
-
-# Create table
-$node->safe_psql("postgres", "CREATE EXTENSION vector;");
-$node->safe_psql("postgres", "CREATE TABLE tst (v halfvec(3));");
-
-sub insert_vectors
-{
-	for my $i (1 .. 20)
-	{
-		$node->safe_psql("postgres", "INSERT INTO tst VALUES ('[1,1,1]');");
-	}
-}
-
-sub test_duplicates
-{
-	my $res = $node->safe_psql("postgres", qq(
-		SET enable_seqscan = off;
-		SET hnsw.ef_search = 1;
-		SELECT COUNT(*) FROM (SELECT * FROM tst ORDER BY v <-> '[1,1,1]') t;
-	));
-	is($res, 10);
-}
-
-# Test duplicates with build
-insert_vectors();
-$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING hnsw (v halfvec_l2_ops);");
-test_duplicates();
-
-# Reset
-$node->safe_psql("postgres", "TRUNCATE tst;");
-
-# Test duplicates with inserts
-insert_vectors();
-test_duplicates();
-
-# Test fallback path for inserts
-$node->pgbench(
-	"--no-vacuum --client=5 --transactions=100",
-	0,
-	[qr{actually processed}],
-	[qr{^$}],
-	"concurrent INSERTs",
-	{
-		"027_hnsw_halfvec_duplicates" => "INSERT INTO tst VALUES ('[1,1,1]');"
-	}
-);
-
-done_testing();

test/t/028_hnsw_sparsevec_duplicates.pl

@@ -1,58 +0,0 @@
-use strict;
-use warnings;
-use PostgresNode;
-use TestLib;
-use Test::More;
-
-# Initialize node
-my $node = get_new_node('node');
-$node->init;
-$node->start;
-
-# Create table
-$node->safe_psql("postgres", "CREATE EXTENSION vector;");
-$node->safe_psql("postgres", "CREATE TABLE tst (v sparsevec(3));");
-
-sub insert_vectors
-{
-	for my $i (1 .. 20)
-	{
-		$node->safe_psql("postgres", "INSERT INTO tst VALUES ('{1:1,2:1,3:1}/3');");
-	}
-}
-
-sub test_duplicates
-{
-	my $res = $node->safe_psql("postgres", qq(
-		SET enable_seqscan = off;
-		SET hnsw.ef_search = 1;
-		SELECT COUNT(*) FROM (SELECT * FROM tst ORDER BY v <-> '{1:1,2:1,3:1}/3') t;
-	));
-	is($res, 10);
-}
-
-# Test duplicates with build
-insert_vectors();
-$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING hnsw (v sparsevec_l2_ops);");
-test_duplicates();
-
-# Reset
-$node->safe_psql("postgres", "TRUNCATE tst;");
-
-# Test duplicates with inserts
-insert_vectors();
-test_duplicates();
-
-# Test fallback path for inserts
-$node->pgbench(
-	"--no-vacuum --client=5 --transactions=100",
-	0,
-	[qr{actually processed}],
-	[qr{^$}],
-	"concurrent INSERTs",
-	{
-		"028_hnsw_sparsevec_duplicates" => "INSERT INTO tst VALUES ('{1:1,2:1,3:1}/3');"
-	}
-);
-
-done_testing();

test/t/029_hnsw_bit_vacuum_recall.pl

@@ -1,100 +0,0 @@
-use strict;
-use warnings;
-use PostgresNode;
-use TestLib;
-use Test::More;
-
-my $node;
-my @queries = ();
-my @expected;
-my $limit = 20;
-my $dim = 52;
-my $max = 2**$dim;
-
-sub test_recall
-{
-	my ($min, $ef_search, $test_name) = @_;
-	my $correct = 0;
-	my $total = 0;
-
-	my $explain = $node->safe_psql("postgres", qq(
-		SET enable_seqscan = off;
-		SET hnsw.ef_search = $ef_search;
-		EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v <~> $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;
-			SET hnsw.ef_search = $ef_search;
-			SELECT i FROM tst ORDER BY v <~> $queries[$i] LIMIT $limit;
-		));
-		my @actual_ids = split("\n", $actual);
-
-		my @expected_ids = split("\n", $expected[$i]);
-		my %expected_set = map { $_ => 1 } @expected_ids;
-
-		foreach (@actual_ids)
-		{
-			if (exists($expected_set{$_}))
-			{
-				$correct++;
-			}
-		}
-
-		$total += $limit;
-	}
-
-	cmp_ok($correct / $total, ">=", $min, $test_name);
-}
-
-# Initialize node
-$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 bit($dim));");
-$node->safe_psql("postgres", "ALTER TABLE tst SET (autovacuum_enabled = false);");
-$node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i, (random() * $max)::bigint::bit($dim) FROM generate_series(1, 10000) i;"
-);
-
-# Add index
-$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v bit_hamming_ops) WITH (m = 4, ef_construction = 8);");
-
-# Delete data
-$node->safe_psql("postgres", "DELETE FROM tst WHERE i > 2500;");
-
-# Generate queries
-for (1 .. 20)
-{
-	my $r = int(rand() * $max);
-	push(@queries, "${r}::bigint::bit($dim)");
-}
-
-# Get exact results
-@expected = ();
-foreach (@queries)
-{
-	my $res = $node->safe_psql("postgres", qq(
-		SET enable_indexscan = off;
-		WITH top AS (
-			SELECT v <~> $_ AS distance FROM tst ORDER BY distance LIMIT $limit
-		)
-		SELECT i FROM tst WHERE (v <~> $_) <= (SELECT MAX(distance) FROM top)
-	));
-	push(@expected, $res);
-}
-
-test_recall(0.35, 100, "before vacuum");
-
-# TODO Test concurrent inserts with vacuum
-$node->safe_psql("postgres", "VACUUM tst;");
-
-test_recall(0.80, 100, "after vacuum");
-
-done_testing();

test/t/030_hnsw_halfvec_vacuum_recall.pl

@@ -1,97 +0,0 @@
-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, $ef_search, $test_name) = @_;
-	my $correct = 0;
-	my $total = 0;
-
-	my $explain = $node->safe_psql("postgres", qq(
-		SET enable_seqscan = off;
-		SET hnsw.ef_search = $ef_search;
-		EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v <-> '$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;
-			SET hnsw.ef_search = $ef_search;
-			SELECT i FROM tst ORDER BY v <-> '$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, $test_name);
-}
-
-# Initialize node
-$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 halfvec(3));");
-$node->safe_psql("postgres", "ALTER TABLE tst SET (autovacuum_enabled = false);");
-$node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()] FROM generate_series(1, 10000) i;"
-);
-
-# Add index
-$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v halfvec_l2_ops) WITH (m = 4, ef_construction = 8);");
-
-# Delete data
-$node->safe_psql("postgres", "DELETE FROM tst WHERE i > 2500;");
-
-# Generate queries
-for (1 .. 20)
-{
-	my $r1 = rand();
-	my $r2 = rand();
-	my $r3 = rand();
-	push(@queries, "[$r1,$r2,$r3]");
-}
-
-# Get exact results
-@expected = ();
-foreach (@queries)
-{
-	my $res = $node->safe_psql("postgres", qq(
-		SET enable_indexscan = off;
-		SELECT i FROM tst ORDER BY v <-> '$_' LIMIT $limit;
-	));
-	push(@expected, $res);
-}
-
-test_recall(0.18, $limit, "before vacuum");
-test_recall(0.95, 100, "before vacuum");
-
-# TODO Test concurrent inserts with vacuum
-$node->safe_psql("postgres", "VACUUM tst;");
-
-test_recall(0.95, $limit, "after vacuum");
-
-done_testing();

test/t/031_hnsw_sparsevec_vacuum_recall.pl

@@ -1,97 +0,0 @@
-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, $ef_search, $test_name) = @_;
-	my $correct = 0;
-	my $total = 0;
-
-	my $explain = $node->safe_psql("postgres", qq(
-		SET enable_seqscan = off;
-		SET hnsw.ef_search = $ef_search;
-		EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v <-> '$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;
-			SET hnsw.ef_search = $ef_search;
-			SELECT i FROM tst ORDER BY v <-> '$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, $test_name);
-}
-
-# Initialize node
-$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));");
-$node->safe_psql("postgres", "ALTER TABLE tst SET (autovacuum_enabled = false);");
-$node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()]::vector::sparsevec(3) FROM generate_series(1, 10000) i;"
-);
-
-# Add index
-$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v sparsevec_l2_ops) WITH (m = 4, ef_construction = 8);");
-
-# Delete data
-$node->safe_psql("postgres", "DELETE FROM tst WHERE i > 2500;");
-
-# Generate queries
-for (1 .. 20)
-{
-	my $r1 = rand();
-	my $r2 = rand();
-	my $r3 = rand();
-	push(@queries, "{1:$r1,2:$r2,3:$r3}/3");
-}
-
-# Get exact results
-@expected = ();
-foreach (@queries)
-{
-	my $res = $node->safe_psql("postgres", qq(
-		SET enable_indexscan = off;
-		SELECT i FROM tst ORDER BY v <-> '$_' LIMIT $limit;
-	));
-	push(@expected, $res);
-}
-
-test_recall(0.18, $limit, "before vacuum");
-test_recall(0.95, 100, "before vacuum");
-
-# TODO Test concurrent inserts with vacuum
-$node->safe_psql("postgres", "VACUUM tst;");
-
-test_recall(0.95, $limit, "after vacuum");
-
-done_testing();

test/t/032_ivfflat_halfvec_build_recall.pl

@@ -1,154 +0,0 @@
-use strict;
-use warnings;
-use PostgresNode;
-use TestLib;
-use Test::More;
-
-my $node;
-my @queries = ();
-my @expected;
-my $limit = 20;
-my $dim = 10;
-my $array_sql = join(",", ('random()') x $dim);
-
-sub test_recall
-{
-	my ($probes, $min, $operator) = @_;
-	my $correct = 0;
-	my $total = 0;
-
-	my $explain = $node->safe_psql("postgres", qq(
-		SET enable_seqscan = off;
-		SET ivfflat.probes = $probes;
-		EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v $operator '$queries[0]' LIMIT $limit;
-	));
-	like($explain, qr/Index Scan using idx on tst/);
-
-	for my $i (0 .. $#queries)
-	{
-		my $actual = $node->safe_psql("postgres", qq(
-			SET enable_seqscan = off;
-			SET ivfflat.probes = $probes;
-			SELECT i FROM tst ORDER BY v $operator '$queries[$i]' LIMIT $limit;
-		));
-		my @actual_ids = split("\n", $actual);
-
-		my @expected_ids = split("\n", $expected[$i]);
-		my %expected_set = map { $_ => 1 } @expected_ids;
-
-		foreach (@actual_ids)
-		{
-			if (exists($expected_set{$_}))
-			{
-				$correct++;
-			}
-		}
-
-		$total += $limit;
-	}
-
-	cmp_ok($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 halfvec($dim));");
-$node->safe_psql("postgres",
-	"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
-);
-
-# Generate queries
-for (1 .. 20)
-{
-	my @r = ();
-	for (1 .. $dim)
-	{
-		push(@r, rand());
-	}
-	push(@queries, "[" . join(",", @r) . "]");
-}
-
-# Check each index type
-my @operators = ("<->", "<#>", "<=>");
-my @opclasses = ("halfvec_l2_ops", "halfvec_ip_ops", "halfvec_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", qq(
-			WITH top AS (
-				SELECT v $operator '$_' AS distance FROM tst ORDER BY distance LIMIT $limit
-			)
-			SELECT i FROM tst WHERE (v $operator '$_') <= (SELECT MAX(distance) FROM top)
-		));
-		push(@expected, $res);
-	}
-
-	# Build index serially
-	$node->safe_psql("postgres", qq(
-		SET max_parallel_maintenance_workers = 0;
-		CREATE INDEX idx ON tst USING ivfflat (v $opclass);
-	));
-
-	# Test approximate results
-	if ($operator ne "<#>")
-	{
-		# TODO Fix test (uniform random vectors all have similar inner product)
-		test_recall(1, 0.35, $operator);
-		test_recall(10, 0.95, $operator);
-	}
-
-	# Test probes equals lists
-	if ($operator eq "<=>")
-	{
-		test_recall(100, 0.98, $operator);
-	}
-	else
-	{
-		test_recall(100, 1.00, $operator);
-	}
-
-	$node->safe_psql("postgres", "DROP INDEX idx;");
-
-	# Build index in parallel
-	my ($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 ivfflat (v $opclass);
-	));
-	is($ret, 0, $stderr);
-	like($stderr, qr/using \d+ parallel workers/);
-
-	# Test approximate results
-	if ($operator ne "<#>")
-	{
-		# TODO Fix test (uniform random vectors all have similar inner product)
-		test_recall(1, 0.35, $operator);
-		test_recall(10, 0.95, $operator);
-	}
-
-	# Test probes equals lists
-	if ($operator eq "<=>")
-	{
-		test_recall(100, 0.98, $operator);
-	}
-	else
-	{
-		test_recall(100, 1.00, $operator);
-	}
-
-	$node->safe_psql("postgres", "DROP INDEX idx;");
-}
-
-done_testing();