Started half indexing

sql/vector--0.5.1--0.6.0.sql

@@ -46,45 +46,21 @@ CREATE FUNCTION half_negative_inner_product(half[], half[]) RETURNS float8
 CREATE FUNCTION float4_to_half(real, integer, boolean) RETURNS half
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION half_to_float4(half, integer, boolean) RETURNS real
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION float8_to_half(float8, integer, boolean) RETURNS half
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION half_to_float8(half, integer, boolean) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
 CREATE FUNCTION integer_to_half(integer, integer, boolean) RETURNS half
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
 CREATE FUNCTION numeric_to_half(numeric, integer, boolean) RETURNS half
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION half_to_numeric(half, integer, boolean) RETURNS numeric
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
 CREATE CAST (real AS half)
 	WITH FUNCTION float4_to_half(real, integer, boolean) AS IMPLICIT;
 
-CREATE CAST (half AS real)
-	WITH FUNCTION half_to_float4(half, integer, boolean) AS IMPLICIT;
-
-CREATE CAST (float8 AS half)
-	WITH FUNCTION float8_to_half(float8, integer, boolean) AS IMPLICIT;
-
-CREATE CAST (half AS float8)
-	WITH FUNCTION half_to_float8(half, integer, boolean) AS IMPLICIT;
-
 CREATE CAST (integer AS half)
 	WITH FUNCTION integer_to_half(integer, integer, boolean) AS IMPLICIT;
 
 CREATE CAST (numeric AS half)
 	WITH FUNCTION numeric_to_half(numeric, integer, boolean) AS IMPLICIT;
 
-CREATE CAST (half AS numeric)
-	WITH FUNCTION half_to_numeric(half, integer, boolean) AS IMPLICIT;
-
 CREATE OPERATOR <-> (
 	LEFTARG = half[], RIGHTARG = half[], PROCEDURE = l2_distance,
 	COMMUTATOR = '<->'
@@ -99,3 +75,8 @@ CREATE OPERATOR <=> (
 	LEFTARG = half[], RIGHTARG = half[], PROCEDURE = cosine_distance,
 	COMMUTATOR = '<=>'
 );
+
+CREATE OPERATOR CLASS half_l2_ops
+	FOR TYPE half[] USING hnsw AS
+	OPERATOR 1 <-> (half[], half[]) FOR ORDER BY float_ops,
+	FUNCTION 1 half_l2_squared_distance(half[], half[]);

sql/vector.sql

@@ -344,47 +344,23 @@ CREATE FUNCTION half_negative_inner_product(half[], half[]) RETURNS float8
 CREATE FUNCTION float4_to_half(real, integer, boolean) RETURNS half
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION half_to_float4(half, integer, boolean) RETURNS real
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION float8_to_half(float8, integer, boolean) RETURNS half
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION half_to_float8(half, integer, boolean) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
 CREATE FUNCTION integer_to_half(integer, integer, boolean) RETURNS half
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
 CREATE FUNCTION numeric_to_half(numeric, integer, boolean) RETURNS half
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION half_to_numeric(half, integer, boolean) RETURNS numeric
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
 -- half casts
 
 CREATE CAST (real AS half)
 	WITH FUNCTION float4_to_half(real, integer, boolean) AS IMPLICIT;
 
-CREATE CAST (half AS real)
-	WITH FUNCTION half_to_float4(half, integer, boolean) AS IMPLICIT;
-
-CREATE CAST (float8 AS half)
-	WITH FUNCTION float8_to_half(float8, integer, boolean) AS IMPLICIT;
-
-CREATE CAST (half AS float8)
-	WITH FUNCTION half_to_float8(half, integer, boolean) AS IMPLICIT;
-
 CREATE CAST (integer AS half)
 	WITH FUNCTION integer_to_half(integer, integer, boolean) AS IMPLICIT;
 
 CREATE CAST (numeric AS half)
 	WITH FUNCTION numeric_to_half(numeric, integer, boolean) AS IMPLICIT;
 
-CREATE CAST (half AS numeric)
-	WITH FUNCTION half_to_numeric(half, integer, boolean) AS IMPLICIT;
-
 -- half operators
 
 CREATE OPERATOR <-> (
@@ -401,3 +377,10 @@ CREATE OPERATOR <=> (
 	LEFTARG = half[], RIGHTARG = half[], PROCEDURE = cosine_distance,
 	COMMUTATOR = '<=>'
 );
+
+-- half opclasses
+
+CREATE OPERATOR CLASS half_l2_ops
+	FOR TYPE half[] USING hnsw AS
+	OPERATOR 1 <-> (half[], half[]) FOR ORDER BY float_ops,
+	FUNCTION 1 half_l2_squared_distance(half[], half[]);

src/half.c

@@ -103,7 +103,7 @@ DatumGetHalf(Datum X)
 }
 
 /*
- * Get a half from a message buffer
+ * Append a half to a StringInfo buffer
  */
 static half
 pq_getmsghalf(StringInfo msg)
@@ -114,12 +114,13 @@ pq_getmsghalf(StringInfo msg)
 		uint16		i;
 	}			swap;
 
+	/* TODO Likely use float4 for clients */
 	swap.i = pq_getmsgint(msg, 2);
 	return swap.h;
 }
 
 /*
- * Append a half to a StringInfo buffer
+ * Get a half from a message buffer
  */
 static void
 pq_sendhalf(StringInfo buf, half h)
@@ -130,6 +131,7 @@ pq_sendhalf(StringInfo buf, half h)
 		uint16		i;
 	}			swap;
 
+	/* TODO Likely use float4 for clients */
 	swap.h = h;
 	pq_sendint16(buf, swap.i);
 }
@@ -279,32 +281,6 @@ Float4ToHalfUnchecked(float num)
 #endif
 }
 
-/*
- * Check if half is infinite
- */
-static inline bool
-HalfIsInf(half num)
-{
-#ifdef FLT16_SUPPORT
-	return isinf(num);
-#else
-	return (num << 1) == 0xF800;
-#endif
-}
-
-/*
- * Check if half is zero
- */
-static inline bool
-HalfIsZero(half num)
-{
-#ifdef FLT16_SUPPORT
-	return num == 0;
-#else
-	return (num << 1) == 0x0000;
-#endif
-}
-
 /*
  * Convert a float4 to a half
  */
@@ -313,26 +289,10 @@ Float4ToHalf(float num)
 {
 	half		result = Float4ToHalfUnchecked(num);
 
-	if (unlikely(HalfIsInf(result)) && !isinf(num))
+	/* TODO Perform checks without HalfToFloat4 */
+	if (unlikely(isinf(HalfToFloat4(result))) && !isinf(num))
 		float_overflow_error();
-	if (unlikely(HalfIsZero(result)) && num != 0.0)
-		float_underflow_error();
-
-	return result;
-}
-
-/*
- * Convert a float8 to a half
- */
-static half
-Float8ToHalf(double num)
-{
-	/* TODO Convert directly for greater accuracy */
-	half		result = Float4ToHalfUnchecked((float) num);
-
-	if (unlikely(HalfIsInf(result)) && !isinf(num))
-		float_overflow_error();
-	if (unlikely(HalfIsZero(result)) && num != 0.0)
+	if (unlikely(HalfToFloat4(result) == 0.0f) && num != 0.0)
 		float_underflow_error();
 
 	return result;
@@ -454,8 +414,6 @@ integer_to_half(PG_FUNCTION_ARGS)
 
 /*
  * Convert numeric to half
- *
- * TODO Improve error message if out of range
  */
 PGDLLEXPORT PG_FUNCTION_INFO_V1(numeric_to_half);
 Datum
@@ -468,20 +426,6 @@ numeric_to_half(PG_FUNCTION_ARGS)
 	PG_RETURN_HALF(h);
 }
 
-/*
- * Convert half to numeric
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(half_to_numeric);
-Datum
-half_to_numeric(PG_FUNCTION_ARGS)
-{
-	half		h = PG_GETARG_HALF(0);
-	float		f = HalfToFloat4(h);
-	Numeric		num = DatumGetNumeric(DirectFunctionCall1(float4_numeric, Float4GetDatum(f)));
-
-	PG_RETURN_NUMERIC(num);
-}
-
 /*
  * Convert float4 to half
  */
@@ -495,45 +439,6 @@ float4_to_half(PG_FUNCTION_ARGS)
 	PG_RETURN_HALF(h);
 }
 
-/*
- * Convert half to float4
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(half_to_float4);
-Datum
-half_to_float4(PG_FUNCTION_ARGS)
-{
-	half		h = PG_GETARG_HALF(0);
-	float		f = HalfToFloat4(h);
-
-	PG_RETURN_FLOAT4(f);
-}
-
-/*
- * Convert float8 to half
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(float8_to_half);
-Datum
-float8_to_half(PG_FUNCTION_ARGS)
-{
-	float8		d = PG_GETARG_FLOAT8(0);
-	half		h = Float8ToHalf(d);
-
-	PG_RETURN_HALF(h);
-}
-
-/*
- * Convert half to float8
- */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(half_to_float8);
-Datum
-half_to_float8(PG_FUNCTION_ARGS)
-{
-	half		h = PG_GETARG_HALF(0);
-	float		f = HalfToFloat4(h);
-
-	PG_RETURN_FLOAT8((double) f);
-}
-
 /*
  * Get the L2 distance between half arrays
  */

src/half.h

@@ -5,7 +5,12 @@
 
 #include <float.h>
 
-#ifdef __FLT16_MAX__
+/* _Float16 and __fp16 are not supported on x86_64 with GCC 11 */
+#if defined(__is_identifier)
+#if __is_identifier(_Float16)
+#define FLT16_SUPPORT
+#endif
+#elif defined(FLT16_MAX)
 #define FLT16_SUPPORT
 #endif
 

src/hnswbuild.c

@@ -462,8 +462,8 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	buildstate->dimensions = TupleDescAttr(index->rd_att, 0)->atttypmod;
 
 	/* Require column to have dimensions to be indexed */
-	if (buildstate->dimensions < 0)
-		elog(ERROR, "column does not have dimensions");
+	// if (buildstate->dimensions < 0)
+	// 	elog(ERROR, "column does not have dimensions");
 
 	if (buildstate->dimensions > HNSW_MAX_DIM)
 		elog(ERROR, "column cannot have more than %d dimensions for hnsw index", HNSW_MAX_DIM);

test/expected/half.out

@@ -58,12 +58,6 @@ SELECT '{1,2,3}'::half[];
  {1,2,3}
 (1 row)
 
-SELECT '{1,2,3}'::half[]::real[];
- float4  
----------
- {1,2,3}
-(1 row)
-
 SELECT '65505'::integer::half;
  half  
 -------
@@ -82,44 +76,6 @@ SELECT 'Infinity'::real::half;
  Infinity
 (1 row)
 
-SELECT '1e-38'::real::half;
-ERROR:  value out of range: underflow
-SELECT '1.5'::half::real;
- float4 
---------
-    1.5
-(1 row)
-
-SELECT '1.5'::real::half;
- half 
-------
- 1.5
-(1 row)
-
-SELECT '1.5'::half::double precision;
- float8 
---------
-    1.5
-(1 row)
-
-SELECT '1.5'::double precision::half;
- half 
-------
- 1.5
-(1 row)
-
-SELECT '1.5'::half::numeric;
- numeric 
----------
-     1.5
-(1 row)
-
-SELECT '1.5'::numeric::half;
- half 
-------
- 1.5
-(1 row)
-
 SELECT l2_distance('{0,0}'::half[], '{3,4}'::half[]);
  l2_distance 
 -------------

test/sql/half.sql

@@ -12,21 +12,10 @@ SELECT '1.5 '::half;
 SELECT '1.5a'::half;
 
 SELECT '{1,2,3}'::half[];
-SELECT '{1,2,3}'::half[]::real[];
 
 SELECT '65505'::integer::half;
 SELECT 'NaN'::real::half;
 SELECT 'Infinity'::real::half;
-SELECT '1e-38'::real::half;
-
-SELECT '1.5'::half::real;
-SELECT '1.5'::real::half;
-
-SELECT '1.5'::half::double precision;
-SELECT '1.5'::double precision::half;
-
-SELECT '1.5'::half::numeric;
-SELECT '1.5'::numeric::half;
 
 SELECT l2_distance('{0,0}'::half[], '{3,4}'::half[]);
 SELECT l2_distance('{0,0}'::half[], '{0,1}'::half[]);

test/t/019_hnsw_half.pl

@@ -0,0 +1,93 @@
+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 half[3]);");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()]::numeric[]::half[] FROM generate_series(1, 10000) i;"
+);
+
+# Generate queries
+for (1 .. 20)
+{
+	my $r1 = rand();
+	my $r2 = rand();
+	my $r3 = rand();
+	push(@queries, "{$r1,$r2,$r3}");
+}
+
+# Check each index type
+my @operators = ("<->");
+my @opclasses = ("half_l2_ops");
+
+for my $i (0 .. $#operators)
+{
+	my $operator = $operators[$i];
+	my $opclass = $opclasses[$i];
+
+	# Get exact results
+	@expected = ();
+	foreach (@queries)
+	{
+		my $res = $node->safe_psql("postgres", "SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;");
+		push(@expected, $res);
+	}
+
+	# Add index
+	$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v $opclass);");
+
+	my $min = $operator eq "<#>" ? 0.80 : 0.99;
+	test_recall($min, $operator);
+}
+
+done_testing();