Moved type support to support functions [skip ci]

The text input function simply left out any zero values, but the
binary input function did not. That's problematic because you end up
with an "unnormalized" sparse vector, which behaves in weird ways.  At
least sparsevec_cmp_internal() expects both inputs to not contain
zeros.

The binary send function never produces such zero values, but an
external tool could. Or to test, you can use COPY TO (FORMAT BINARY),
use a hex editor to edit one of the values to be zero, and copy it
back with COPY FROM (FORMAT BINARY).

.github/workflows/build.yml

@@ -115,6 +115,8 @@ jobs:
           sudo -u postgres make prove_installcheck
         env:
           PG_CFLAGS: -DUSE_ASSERT_CHECKING -Wall -Wextra -Werror -Wno-unused-parameter -Wno-sign-compare
+      - if: ${{ failure() }}
+        run: cat pgvector/regression.diffs
   valgrind:
     if: ${{ !startsWith(github.ref_name, 'mac') && !startsWith(github.ref_name, 'windows') }}
     runs-on: ubuntu-latest
@@ -123,6 +125,7 @@ jobs:
       - uses: ankane/setup-postgres-valgrind@v1
         with:
           postgres-version: 16
+          check-ub: yes
       - run: make OPTFLAGS=""
       - run: sudo --preserve-env=PG_CONFIG make install
       - run: make installcheck

CHANGELOG.md

@@ -2,11 +2,14 @@
 
 - Added `halfvec` type
 - Added `sparsevec` type
-- Added support for `bit` vectors to HNSW
+- Added support for indexing `bit` type
+- Added support for indexing L1 distance with HNSW
+- Added `binary_quantize` function
 - Added `hamming_distance` function
 - Added `jaccard_distance` function
-- Added `quantize_binary` function
+- Added `l2_normalize` function
 - Added `subvector` function
+- Added concatenate operator for vectors
 - Added CPU dispatching for distance functions on Linux x86-64
 - Updated comparison operators to support vectors with different dimensions
 

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/bitutils.o src/bitvec.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
 HEADERS = src/halfvec.h src/sparsevec.h src/vector.h
 
 TESTS = $(wildcard test/sql/*.sql)

Makefile.win

@@ -1,10 +1,10 @@
 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\bitutils.obj src\bitvec.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
 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
+REGRESS = bit_functions btree_halfvec btree_sparsevec btree_vector 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_halfvec_cosine ivfflat_halfvec_ip ivfflat_halfvec_l2 ivfflat_options ivfflat_unlogged ivfflat_vector_cosine ivfflat_vector_ip ivfflat_vector_l2 sparsevec_functions sparsevec_input vector_functions vector_input
 REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)
 
 # For /arch flags

README.md

@@ -81,7 +81,7 @@ Get the nearest neighbors by L2 distance
 SELECT * FROM items ORDER BY embedding <-> '[3,1,2]' LIMIT 5;
 ```
 
-Also supports inner product (`<#>`) and cosine distance (`<=>`)
+Also supports inner product (`<#>`), cosine distance (`<=>`), and L1 distance (`<+>`, unreleased)
 
 Note: `<#>` returns the negative inner product since Postgres only supports `ASC` order index scans on operators
 
@@ -138,6 +138,13 @@ Get the nearest neighbors to a vector
 SELECT * FROM items ORDER BY embedding <-> '[3,1,2]' LIMIT 5;
 ```
 
+Supported distance functions are:
+
+- `<->` - L2 distance
+- `<#>` - (negative) inner product
+- `<=>` - cosine distance
+- `<+>` - L1 distance (unreleased)
+
 Get the nearest neighbors to a row
 
 ```sql
@@ -221,6 +228,12 @@ Cosine distance
 CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops);
 ```
 
+L1 distance - unreleased
+
+```sql
+CREATE INDEX ON items USING hnsw (embedding vector_l1_ops);
+```
+
 Hamming distance - unreleased
 
 ```sql
@@ -343,10 +356,17 @@ Cosine distance
 CREATE INDEX ON items USING ivfflat (embedding vector_cosine_ops) WITH (lists = 100);
 ```
 
+Hamming distance - unreleased
+
+```sql
+CREATE INDEX ON items USING ivfflat (embedding bit_hamming_ops) WITH (lists = 100);
+```
+
 Supported types are:
 
 - `vector` - up to 2,000 dimensions
 - `halfvec` - up to 4,000 dimensions (unreleased)
+- `bit` - up to 64,000 dimensions (unreleased)
 
 ### Query Options
 
@@ -476,20 +496,20 @@ Also supports Jaccard distance (`<%>`)
 Use expression indexing for binary quantization
 
 ```sql
-CREATE INDEX ON items USING hnsw ((quantize_binary(embedding)::bit(3)) bit_hamming_ops);
+CREATE INDEX ON items USING hnsw ((binary_quantize(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;
+SELECT * FROM items ORDER BY binary_quantize(embedding)::bit(3) <~> binary_quantize('[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
+    SELECT * FROM items ORDER BY binary_quantize(embedding)::bit(3) <~> binary_quantize('[1,-2,3]') LIMIT 20
 ) ORDER BY embedding <=> '[1,-2,3]' LIMIT 5;
 ```
 
@@ -509,7 +529,7 @@ Insert vectors
 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
+The format is `{index1:value1,index2:value2}/dimensions` and indices start at 1 like SQL arrays
 
 Get the nearest neighbors by L2 distance
 
@@ -528,7 +548,7 @@ 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
+## Indexing Subvectors
 
 *Unreleased*
 
@@ -844,19 +864,22 @@ Operator | Description | Added
 \+ | element-wise addition |
 \- | element-wise subtraction |
 \* | element-wise multiplication | 0.5.0
+\|\| | concatenate | unreleased
 <-> | Euclidean distance |
 <#> | negative inner product |
 <=> | cosine distance |
+<+> | taxicab distance | unreleased
 
 ### Vector Functions
 
 Function | Description | Added
 --- | --- | ---
+binary_quantize(vector) → bit | binary quantize | unreleased
 cosine_distance(vector, vector) → double precision | cosine distance |
 inner_product(vector, vector) → double precision | inner product |
-l2_distance(vector, vector) → double precision | Euclidean distance |
 l1_distance(vector, vector) → double precision | taxicab distance | 0.5.0
-quantize_binary(vector) → bit | quantize | unreleased
+l2_distance(vector, vector) → double precision | Euclidean distance |
+l2_normalize(vector) → vector | Normalize with Euclidean norm | unreleased
 subvector(vector, integer, integer) → vector | subvector | unreleased
 vector_dims(vector) → integer | number of dimensions |
 vector_norm(vector) → double precision | Euclidean norm |
@@ -876,20 +899,35 @@ Each half vector takes `2 * dimensions + 8` bytes of storage. Each element is a
 
 Operator | Description | Added
 --- | --- | ---
+\+ | element-wise addition | unreleased
+\- | element-wise subtraction | unreleased
+\* | element-wise multiplication | unreleased
+\|\| | concatenate | unreleased
 <-> | Euclidean distance | unreleased
 <#> | negative inner product | unreleased
 <=> | cosine distance | unreleased
+<+> | taxicab distance | unreleased
 
 ### Halfvec Functions
 
 Function | Description | Added
 --- | --- | ---
+binary_quantize(halfvec) → bit | binary quantize | unreleased
 cosine_distance(halfvec, halfvec) → double precision | cosine distance | unreleased
 inner_product(halfvec, halfvec) → double precision | inner product | unreleased
-l2_distance(halfvec, halfvec) → double precision | Euclidean distance | unreleased
 l1_distance(halfvec, halfvec) → double precision | taxicab distance | unreleased
-quantize_binary(halfvec) → bit | quantize | unreleased
+l2_distance(halfvec, halfvec) → double precision | Euclidean distance | unreleased
+l2_norm(halfvec) → double precision | Euclidean norm | unreleased
+l2_normalize(halfvec) → halfvec | Normalize with Euclidean norm | unreleased
 subvector(halfvec, integer, integer) → halfvec | subvector | unreleased
+vector_dims(halfvec) → integer | number of dimensions | unreleased
+
+### Halfvec Aggregate Functions
+
+Function | Description | Added
+--- | --- | ---
+avg(halfvec) → halfvec | average | unreleased
+sum(halfvec) → halfvec | sum | unreleased
 
 ### Bit Type
 
@@ -920,6 +958,7 @@ Operator | Description | Added
 <-> | Euclidean distance | unreleased
 <#> | negative inner product | unreleased
 <=> | cosine distance | unreleased
+<+> | taxicab distance | unreleased
 
 ### Sparsevec Functions
 
@@ -927,8 +966,10 @@ Function | Description | Added
 --- | --- | ---
 cosine_distance(sparsevec, sparsevec) → double precision | cosine distance | unreleased
 inner_product(sparsevec, sparsevec) → double precision | inner product | unreleased
-l2_distance(sparsevec, sparsevec) → double precision | Euclidean distance | unreleased
 l1_distance(sparsevec, sparsevec) → double precision | taxicab distance | unreleased
+l2_distance(sparsevec, sparsevec) → double precision | Euclidean distance | unreleased
+l2_norm(sparsevec) → double precision | Euclidean norm | unreleased
+l2_normalize(sparsevec) → sparsevec | Normalize with Euclidean norm | unreleased
 
 ## Installation Notes - Linux and Mac
 
@@ -1055,7 +1096,7 @@ Note: Replace `16` with your Postgres server version
 Install the FreeBSD package with:
 
 ```sh
-pkg install postgresql15-pg_vector
+pkg install postgresql15-pgvector
 ```
 
 or the port with:

sql/vector--0.6.2--0.7.0.sql

@@ -1,18 +1,44 @@
 -- complain if script is sourced in psql, rather than via CREATE EXTENSION
 \echo Use "ALTER EXTENSION vector UPDATE TO '0.7.0'" to load this file. \quit
 
-CREATE FUNCTION quantize_binary(vector) RETURNS bit
+CREATE FUNCTION l2_normalize(vector) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION binary_quantize(vector) RETURNS bit
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
 CREATE FUNCTION subvector(vector, int, int) RETURNS vector
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+CREATE FUNCTION vector_concat(vector, vector) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE OPERATOR <+> (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = l1_distance,
+	COMMUTATOR = '<+>'
+);
+
+CREATE OPERATOR || (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_concat
+);
+
+CREATE OPERATOR CLASS vector_l1_ops
+	FOR TYPE vector USING hnsw AS
+	OPERATOR 1 <+> (vector, vector) FOR ORDER BY float_ops,
+	FUNCTION 1 l1_distance(vector, vector);
+
 CREATE FUNCTION hamming_distance(bit, bit) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
 CREATE FUNCTION jaccard_distance(bit, bit) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+CREATE FUNCTION bit_ivfflat_support(internal) RETURNS internal
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION bit_hnsw_support(internal) RETURNS internal
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
 CREATE OPERATOR <~> (
 	LEFTARG = bit, RIGHTARG = bit, PROCEDURE = hamming_distance,
 	COMMUTATOR = '<~>'
@@ -23,15 +49,24 @@ CREATE OPERATOR <%> (
 	COMMUTATOR = '<%>'
 );
 
+CREATE OPERATOR CLASS bit_hamming_ops
+	FOR TYPE bit USING ivfflat AS
+	OPERATOR 1 <~> (bit, bit) FOR ORDER BY float_ops,
+	FUNCTION 1 hamming_distance(bit, bit),
+	FUNCTION 3 hamming_distance(bit, bit),
+	FUNCTION 6 bit_ivfflat_support(internal);
+
 CREATE OPERATOR CLASS bit_hamming_ops
 	FOR TYPE bit USING hnsw AS
 	OPERATOR 1 <~> (bit, bit) FOR ORDER BY float_ops,
-	FUNCTION 1 hamming_distance(bit, bit);
+	FUNCTION 1 hamming_distance(bit, bit),
+	FUNCTION 4 bit_hnsw_support(internal);
 
 CREATE OPERATOR CLASS bit_jaccard_ops
 	FOR TYPE bit USING hnsw AS
 	OPERATOR 1 <%> (bit, bit) FOR ORDER BY float_ops,
-	FUNCTION 1 jaccard_distance(bit, bit);
+	FUNCTION 1 jaccard_distance(bit, bit),
+	FUNCTION 4 bit_hnsw_support(internal);
 
 CREATE TYPE halfvec;
 
@@ -71,15 +106,54 @@ CREATE FUNCTION cosine_distance(halfvec, halfvec) RETURNS float8
 CREATE FUNCTION l1_distance(halfvec, halfvec) RETURNS float8
 	AS 'MODULE_PATHNAME', 'halfvec_l1_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION halfvec_norm(halfvec) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+CREATE FUNCTION vector_dims(halfvec) RETURNS integer
+	AS 'MODULE_PATHNAME', 'halfvec_vector_dims' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION quantize_binary(halfvec) RETURNS bit
-	AS 'MODULE_PATHNAME', 'halfvec_quantize_binary' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+CREATE FUNCTION l2_norm(halfvec) RETURNS float8
+	AS 'MODULE_PATHNAME', 'halfvec_l2_norm' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION l2_normalize(halfvec) RETURNS halfvec
+	AS 'MODULE_PATHNAME', 'halfvec_l2_normalize' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION binary_quantize(halfvec) RETURNS bit
+	AS 'MODULE_PATHNAME', 'halfvec_binary_quantize' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
 CREATE FUNCTION subvector(halfvec, int, int) RETURNS halfvec
 	AS 'MODULE_PATHNAME', 'halfvec_subvector' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+CREATE FUNCTION halfvec_add(halfvec, halfvec) RETURNS halfvec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_sub(halfvec, halfvec) RETURNS halfvec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_mul(halfvec, halfvec) RETURNS halfvec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_concat(halfvec, halfvec) RETURNS halfvec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_lt(halfvec, halfvec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_le(halfvec, halfvec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_eq(halfvec, halfvec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_ne(halfvec, halfvec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_ge(halfvec, halfvec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_gt(halfvec, halfvec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_cmp(halfvec, halfvec) RETURNS int4
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
 CREATE FUNCTION halfvec_l2_squared_distance(halfvec, halfvec) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
@@ -89,9 +163,43 @@ CREATE FUNCTION halfvec_negative_inner_product(halfvec, halfvec) RETURNS float8
 CREATE FUNCTION halfvec_spherical_distance(halfvec, halfvec) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+CREATE FUNCTION halfvec_accum(double precision[], halfvec) RETURNS double precision[]
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_avg(double precision[]) RETURNS halfvec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_ivfflat_support(internal) RETURNS internal
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_hnsw_support(internal) RETURNS internal
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE AGGREGATE avg(halfvec) (
+	SFUNC = halfvec_accum,
+	STYPE = double precision[],
+	FINALFUNC = halfvec_avg,
+	COMBINEFUNC = vector_combine,
+	INITCOND = '{0}',
+	PARALLEL = SAFE
+);
+
+CREATE AGGREGATE sum(halfvec) (
+	SFUNC = halfvec_add,
+	STYPE = halfvec,
+	COMBINEFUNC = halfvec_add,
+	PARALLEL = SAFE
+);
+
 CREATE FUNCTION halfvec(halfvec, integer, boolean) RETURNS halfvec
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+CREATE FUNCTION halfvec_to_vector(halfvec, integer, boolean) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_to_halfvec(vector, integer, boolean) RETURNS halfvec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
 CREATE FUNCTION array_to_halfvec(integer[], integer, boolean) RETURNS halfvec
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
@@ -110,8 +218,14 @@ CREATE FUNCTION halfvec_to_float4(halfvec, integer, boolean) RETURNS real[]
 CREATE CAST (halfvec AS halfvec)
 	WITH FUNCTION halfvec(halfvec, integer, boolean) AS IMPLICIT;
 
+CREATE CAST (halfvec AS vector)
+	WITH FUNCTION halfvec_to_vector(halfvec, integer, boolean) AS ASSIGNMENT;
+
+CREATE CAST (vector AS halfvec)
+	WITH FUNCTION vector_to_halfvec(vector, integer, boolean) AS IMPLICIT;
+
 CREATE CAST (halfvec AS real[])
-	WITH FUNCTION halfvec_to_float4(halfvec, integer, boolean) AS IMPLICIT;
+	WITH FUNCTION halfvec_to_float4(halfvec, integer, boolean) AS ASSIGNMENT;
 
 CREATE CAST (integer[] AS halfvec)
 	WITH FUNCTION array_to_halfvec(integer[], integer, boolean) AS ASSIGNMENT;
@@ -140,54 +254,125 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
+CREATE OPERATOR <+> (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = l1_distance,
+	COMMUTATOR = '<+>'
+);
+
+CREATE OPERATOR + (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_add,
+	COMMUTATOR = +
+);
+
+CREATE OPERATOR - (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_sub
+);
+
+CREATE OPERATOR * (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_mul,
+	COMMUTATOR = *
+);
+
+CREATE OPERATOR || (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_concat
+);
+
+CREATE OPERATOR < (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_lt,
+	COMMUTATOR = > , NEGATOR = >= ,
+	RESTRICT = scalarltsel, JOIN = scalarltjoinsel
+);
+
+CREATE OPERATOR <= (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_le,
+	COMMUTATOR = >= , NEGATOR = > ,
+	RESTRICT = scalarlesel, JOIN = scalarlejoinsel
+);
+
+CREATE OPERATOR = (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_eq,
+	COMMUTATOR = = , NEGATOR = <> ,
+	RESTRICT = eqsel, JOIN = eqjoinsel
+);
+
+CREATE OPERATOR <> (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_ne,
+	COMMUTATOR = <> , NEGATOR = = ,
+	RESTRICT = eqsel, JOIN = eqjoinsel
+);
+
+CREATE OPERATOR >= (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_ge,
+	COMMUTATOR = <= , NEGATOR = < ,
+	RESTRICT = scalargesel, JOIN = scalargejoinsel
+);
+
+CREATE OPERATOR > (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_gt,
+	COMMUTATOR = < , NEGATOR = <= ,
+	RESTRICT = scalargtsel, JOIN = scalargtjoinsel
+);
+
+CREATE OPERATOR CLASS halfvec_ops
+	DEFAULT FOR TYPE halfvec USING btree AS
+	OPERATOR 1 < ,
+	OPERATOR 2 <= ,
+	OPERATOR 3 = ,
+	OPERATOR 4 >= ,
+	OPERATOR 5 > ,
+	FUNCTION 1 halfvec_cmp(halfvec, halfvec);
+
 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);
+	FUNCTION 3 l2_distance(halfvec, halfvec),
+	FUNCTION 6 halfvec_ivfflat_support(internal);
 
 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);
+	FUNCTION 4 l2_norm(halfvec),
+	FUNCTION 5 l2_normalize(halfvec),
+	FUNCTION 6 halfvec_ivfflat_support(internal);
 
 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 2 l2_norm(halfvec),
 	FUNCTION 3 halfvec_spherical_distance(halfvec, halfvec),
-	FUNCTION 4 halfvec_norm(halfvec);
+	FUNCTION 4 l2_norm(halfvec),
+	FUNCTION 5 l2_normalize(halfvec),
+	FUNCTION 6 halfvec_ivfflat_support(internal);
 
 CREATE OPERATOR CLASS halfvec_l2_ops
 	FOR TYPE halfvec USING hnsw AS
 	OPERATOR 1 <-> (halfvec, halfvec) FOR ORDER BY float_ops,
-	FUNCTION 1 halfvec_l2_squared_distance(halfvec, halfvec);
+	FUNCTION 1 halfvec_l2_squared_distance(halfvec, halfvec),
+	FUNCTION 4 halfvec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS halfvec_ip_ops
 	FOR TYPE halfvec USING hnsw AS
 	OPERATOR 1 <#> (halfvec, halfvec) FOR ORDER BY float_ops,
-	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec);
+	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
+	FUNCTION 4 halfvec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS halfvec_cosine_ops
 	FOR TYPE halfvec USING hnsw AS
 	OPERATOR 1 <=> (halfvec, halfvec) FOR ORDER BY float_ops,
 	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
-	FUNCTION 2 halfvec_norm(halfvec);
+	FUNCTION 2 l2_norm(halfvec),
+	FUNCTION 3 l2_normalize(halfvec),
+	FUNCTION 4 halfvec_hnsw_support(internal);
 
-CREATE FUNCTION halfvec_to_vector(halfvec, integer, boolean) RETURNS vector
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION vector_to_halfvec(vector, integer, boolean) RETURNS halfvec
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE CAST (halfvec AS vector)
-	WITH FUNCTION halfvec_to_vector(halfvec, integer, boolean) AS IMPLICIT;
-
-CREATE CAST (vector AS halfvec)
-	WITH FUNCTION vector_to_halfvec(vector, integer, boolean) AS IMPLICIT;
+CREATE OPERATOR CLASS halfvec_l1_ops
+	FOR TYPE halfvec USING hnsw AS
+	OPERATOR 1 <+> (halfvec, halfvec) FOR ORDER BY float_ops,
+	FUNCTION 1 l1_distance(halfvec, halfvec),
+	FUNCTION 4 halfvec_hnsw_support(internal);
 
 CREATE TYPE sparsevec;
 
@@ -224,7 +409,34 @@ CREATE FUNCTION inner_product(sparsevec, sparsevec) RETURNS float8
 CREATE FUNCTION cosine_distance(sparsevec, sparsevec) RETURNS float8
 	AS 'MODULE_PATHNAME', 'sparsevec_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION sparsevec_norm(sparsevec) RETURNS float8
+CREATE FUNCTION l1_distance(sparsevec, sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME', 'sparsevec_l1_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION l2_norm(sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME', 'sparsevec_l2_norm' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION l2_normalize(sparsevec) RETURNS sparsevec
+	AS 'MODULE_PATHNAME', 'sparsevec_l2_normalize' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_lt(sparsevec, sparsevec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_le(sparsevec, sparsevec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_eq(sparsevec, sparsevec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_ne(sparsevec, sparsevec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_ge(sparsevec, sparsevec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_gt(sparsevec, sparsevec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_cmp(sparsevec, sparsevec) RETURNS int4
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
 CREATE FUNCTION sparsevec_l2_squared_distance(sparsevec, sparsevec) RETURNS float8
@@ -242,15 +454,30 @@ CREATE FUNCTION vector_to_sparsevec(vector, integer, boolean) RETURNS sparsevec
 CREATE FUNCTION sparsevec_to_vector(sparsevec, integer, boolean) RETURNS vector
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+CREATE FUNCTION halfvec_to_sparsevec(halfvec, integer, boolean) RETURNS sparsevec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_to_halfvec(sparsevec, integer, boolean) RETURNS halfvec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_hnsw_support(internal) RETURNS internal
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
 CREATE CAST (sparsevec AS sparsevec)
 	WITH FUNCTION sparsevec(sparsevec, integer, boolean) AS IMPLICIT;
 
 CREATE CAST (sparsevec AS vector)
-	WITH FUNCTION sparsevec_to_vector(sparsevec, integer, boolean) AS IMPLICIT;
+	WITH FUNCTION sparsevec_to_vector(sparsevec, integer, boolean) AS ASSIGNMENT;
 
 CREATE CAST (vector AS sparsevec)
 	WITH FUNCTION vector_to_sparsevec(vector, integer, boolean) AS IMPLICIT;
 
+CREATE CAST (sparsevec AS halfvec)
+	WITH FUNCTION sparsevec_to_halfvec(sparsevec, integer, boolean) AS ASSIGNMENT;
+
+CREATE CAST (halfvec AS sparsevec)
+	WITH FUNCTION halfvec_to_sparsevec(halfvec, integer, boolean) AS IMPLICIT;
+
 CREATE OPERATOR <-> (
 	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = l2_distance,
 	COMMUTATOR = '<->'
@@ -266,18 +493,78 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
+CREATE OPERATOR <+> (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = l1_distance,
+	COMMUTATOR = '<+>'
+);
+
+CREATE OPERATOR < (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_lt,
+	COMMUTATOR = > , NEGATOR = >= ,
+	RESTRICT = scalarltsel, JOIN = scalarltjoinsel
+);
+
+CREATE OPERATOR <= (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_le,
+	COMMUTATOR = >= , NEGATOR = > ,
+	RESTRICT = scalarlesel, JOIN = scalarlejoinsel
+);
+
+CREATE OPERATOR = (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_eq,
+	COMMUTATOR = = , NEGATOR = <> ,
+	RESTRICT = eqsel, JOIN = eqjoinsel
+);
+
+CREATE OPERATOR <> (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_ne,
+	COMMUTATOR = <> , NEGATOR = = ,
+	RESTRICT = eqsel, JOIN = eqjoinsel
+);
+
+CREATE OPERATOR >= (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_ge,
+	COMMUTATOR = <= , NEGATOR = < ,
+	RESTRICT = scalargesel, JOIN = scalargejoinsel
+);
+
+CREATE OPERATOR > (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_gt,
+	COMMUTATOR = < , NEGATOR = <= ,
+	RESTRICT = scalargtsel, JOIN = scalargtjoinsel
+);
+
+CREATE OPERATOR CLASS sparsevec_ops
+	DEFAULT FOR TYPE sparsevec USING btree AS
+	OPERATOR 1 < ,
+	OPERATOR 2 <= ,
+	OPERATOR 3 = ,
+	OPERATOR 4 >= ,
+	OPERATOR 5 > ,
+	FUNCTION 1 sparsevec_cmp(sparsevec, sparsevec);
+
 CREATE OPERATOR CLASS sparsevec_l2_ops
 	FOR TYPE sparsevec USING hnsw AS
 	OPERATOR 1 <-> (sparsevec, sparsevec) FOR ORDER BY float_ops,
-	FUNCTION 1 sparsevec_l2_squared_distance(sparsevec, sparsevec);
+	FUNCTION 1 sparsevec_l2_squared_distance(sparsevec, sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS sparsevec_ip_ops
 	FOR TYPE sparsevec USING hnsw AS
 	OPERATOR 1 <#> (sparsevec, sparsevec) FOR ORDER BY float_ops,
-	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec);
+	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS sparsevec_cosine_ops
 	FOR TYPE sparsevec USING hnsw AS
 	OPERATOR 1 <=> (sparsevec, sparsevec) FOR ORDER BY float_ops,
 	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec),
-	FUNCTION 2 sparsevec_norm(sparsevec);
+	FUNCTION 2 l2_norm(sparsevec),
+	FUNCTION 3 l2_normalize(sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);
+
+CREATE OPERATOR CLASS sparsevec_l1_ops
+	FOR TYPE sparsevec USING hnsw AS
+	OPERATOR 1 <+> (sparsevec, sparsevec) FOR ORDER BY float_ops,
+	FUNCTION 1 l1_distance(sparsevec, sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);

sql/vector.sql

@@ -49,6 +49,17 @@ CREATE FUNCTION vector_dims(vector) RETURNS integer
 CREATE FUNCTION vector_norm(vector) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+CREATE FUNCTION l2_normalize(vector) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION binary_quantize(vector) RETURNS bit
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION subvector(vector, int, int) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+-- vector private functions
+
 CREATE FUNCTION vector_add(vector, vector) RETURNS vector
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
@@ -58,14 +69,9 @@ CREATE FUNCTION vector_sub(vector, vector) RETURNS vector
 CREATE FUNCTION vector_mul(vector, vector) RETURNS vector
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION quantize_binary(vector) RETURNS bit
+CREATE FUNCTION vector_concat(vector, vector) RETURNS vector
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION subvector(vector, int, int) RETURNS vector
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
--- vector private functions
-
 CREATE FUNCTION vector_lt(vector, vector) RETURNS bool
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
@@ -180,6 +186,11 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
+CREATE OPERATOR <+> (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = l1_distance,
+	COMMUTATOR = '<+>'
+);
+
 CREATE OPERATOR + (
 	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_add,
 	COMMUTATOR = +
@@ -194,6 +205,10 @@ CREATE OPERATOR * (
 	COMMUTATOR = *
 );
 
+CREATE OPERATOR || (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_concat
+);
+
 CREATE OPERATOR < (
 	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_lt,
 	COMMUTATOR = > , NEGATOR = >= ,
@@ -294,6 +309,11 @@ CREATE OPERATOR CLASS vector_cosine_ops
 	FUNCTION 1 vector_negative_inner_product(vector, vector),
 	FUNCTION 2 vector_norm(vector);
 
+CREATE OPERATOR CLASS vector_l1_ops
+	FOR TYPE vector USING hnsw AS
+	OPERATOR 1 <+> (vector, vector) FOR ORDER BY float_ops,
+	FUNCTION 1 l1_distance(vector, vector);
+
 -- bit functions
 
 CREATE FUNCTION hamming_distance(bit, bit) RETURNS float8
@@ -302,6 +322,16 @@ CREATE FUNCTION hamming_distance(bit, bit) RETURNS float8
 CREATE FUNCTION jaccard_distance(bit, bit) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+-- bit private functions
+
+CREATE FUNCTION bit_ivfflat_support(internal) RETURNS internal
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION bit_hnsw_support(internal) RETURNS internal
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+-- bit operators
+
 CREATE OPERATOR <~> (
 	LEFTARG = bit, RIGHTARG = bit, PROCEDURE = hamming_distance,
 	COMMUTATOR = '<~>'
@@ -312,15 +342,26 @@ CREATE OPERATOR <%> (
 	COMMUTATOR = '<%>'
 );
 
+-- bit opclasses
+
+CREATE OPERATOR CLASS bit_hamming_ops
+	FOR TYPE bit USING ivfflat AS
+	OPERATOR 1 <~> (bit, bit) FOR ORDER BY float_ops,
+	FUNCTION 1 hamming_distance(bit, bit),
+	FUNCTION 3 hamming_distance(bit, bit),
+	FUNCTION 6 bit_ivfflat_support(internal);
+
 CREATE OPERATOR CLASS bit_hamming_ops
 	FOR TYPE bit USING hnsw AS
 	OPERATOR 1 <~> (bit, bit) FOR ORDER BY float_ops,
-	FUNCTION 1 hamming_distance(bit, bit);
+	FUNCTION 1 hamming_distance(bit, bit),
+	FUNCTION 4 bit_hnsw_support(internal);
 
 CREATE OPERATOR CLASS bit_jaccard_ops
 	FOR TYPE bit USING hnsw AS
 	OPERATOR 1 <%> (bit, bit) FOR ORDER BY float_ops,
-	FUNCTION 1 jaccard_distance(bit, bit);
+	FUNCTION 1 jaccard_distance(bit, bit),
+	FUNCTION 4 bit_hnsw_support(internal);
 
 -- halfvec type
 
@@ -364,17 +405,56 @@ CREATE FUNCTION cosine_distance(halfvec, halfvec) RETURNS float8
 CREATE FUNCTION l1_distance(halfvec, halfvec) RETURNS float8
 	AS 'MODULE_PATHNAME', 'halfvec_l1_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION halfvec_norm(halfvec) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+CREATE FUNCTION vector_dims(halfvec) RETURNS integer
+	AS 'MODULE_PATHNAME', 'halfvec_vector_dims' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION quantize_binary(halfvec) RETURNS bit
-	AS 'MODULE_PATHNAME', 'halfvec_quantize_binary' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+CREATE FUNCTION l2_norm(halfvec) RETURNS float8
+	AS 'MODULE_PATHNAME', 'halfvec_l2_norm' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION l2_normalize(halfvec) RETURNS halfvec
+	AS 'MODULE_PATHNAME', 'halfvec_l2_normalize' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION binary_quantize(halfvec) RETURNS bit
+	AS 'MODULE_PATHNAME', 'halfvec_binary_quantize' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
 CREATE FUNCTION subvector(halfvec, int, int) RETURNS halfvec
 	AS 'MODULE_PATHNAME', 'halfvec_subvector' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
 -- halfvec private functions
 
+CREATE FUNCTION halfvec_add(halfvec, halfvec) RETURNS halfvec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_sub(halfvec, halfvec) RETURNS halfvec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_mul(halfvec, halfvec) RETURNS halfvec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_concat(halfvec, halfvec) RETURNS halfvec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_lt(halfvec, halfvec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_le(halfvec, halfvec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_eq(halfvec, halfvec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_ne(halfvec, halfvec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_ge(halfvec, halfvec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_gt(halfvec, halfvec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_cmp(halfvec, halfvec) RETURNS int4
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
 CREATE FUNCTION halfvec_l2_squared_distance(halfvec, halfvec) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
@@ -384,11 +464,47 @@ CREATE FUNCTION halfvec_negative_inner_product(halfvec, halfvec) RETURNS float8
 CREATE FUNCTION halfvec_spherical_distance(halfvec, halfvec) RETURNS float8
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+CREATE FUNCTION halfvec_accum(double precision[], halfvec) RETURNS double precision[]
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_avg(double precision[]) RETURNS halfvec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_ivfflat_support(internal) RETURNS internal
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION halfvec_hnsw_support(internal) RETURNS internal
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+-- halfvec aggregates
+
+CREATE AGGREGATE avg(halfvec) (
+	SFUNC = halfvec_accum,
+	STYPE = double precision[],
+	FINALFUNC = halfvec_avg,
+	COMBINEFUNC = vector_combine,
+	INITCOND = '{0}',
+	PARALLEL = SAFE
+);
+
+CREATE AGGREGATE sum(halfvec) (
+	SFUNC = halfvec_add,
+	STYPE = halfvec,
+	COMBINEFUNC = halfvec_add,
+	PARALLEL = SAFE
+);
+
 -- halfvec cast functions
 
 CREATE FUNCTION halfvec(halfvec, integer, boolean) RETURNS halfvec
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+CREATE FUNCTION halfvec_to_vector(halfvec, integer, boolean) RETURNS vector
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION vector_to_halfvec(vector, integer, boolean) RETURNS halfvec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
 CREATE FUNCTION array_to_halfvec(integer[], integer, boolean) RETURNS halfvec
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
@@ -409,8 +525,14 @@ CREATE FUNCTION halfvec_to_float4(halfvec, integer, boolean) RETURNS real[]
 CREATE CAST (halfvec AS halfvec)
 	WITH FUNCTION halfvec(halfvec, integer, boolean) AS IMPLICIT;
 
+CREATE CAST (halfvec AS vector)
+	WITH FUNCTION halfvec_to_vector(halfvec, integer, boolean) AS ASSIGNMENT;
+
+CREATE CAST (vector AS halfvec)
+	WITH FUNCTION vector_to_halfvec(vector, integer, boolean) AS IMPLICIT;
+
 CREATE CAST (halfvec AS real[])
-	WITH FUNCTION halfvec_to_float4(halfvec, integer, boolean) AS IMPLICIT;
+	WITH FUNCTION halfvec_to_float4(halfvec, integer, boolean) AS ASSIGNMENT;
 
 CREATE CAST (integer[] AS halfvec)
 	WITH FUNCTION array_to_halfvec(integer[], integer, boolean) AS ASSIGNMENT;
@@ -441,58 +563,127 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
+CREATE OPERATOR <+> (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = l1_distance,
+	COMMUTATOR = '<+>'
+);
+
+CREATE OPERATOR + (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_add,
+	COMMUTATOR = +
+);
+
+CREATE OPERATOR - (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_sub
+);
+
+CREATE OPERATOR * (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_mul,
+	COMMUTATOR = *
+);
+
+CREATE OPERATOR || (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_concat
+);
+
+CREATE OPERATOR < (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_lt,
+	COMMUTATOR = > , NEGATOR = >= ,
+	RESTRICT = scalarltsel, JOIN = scalarltjoinsel
+);
+
+CREATE OPERATOR <= (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_le,
+	COMMUTATOR = >= , NEGATOR = > ,
+	RESTRICT = scalarlesel, JOIN = scalarlejoinsel
+);
+
+CREATE OPERATOR = (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_eq,
+	COMMUTATOR = = , NEGATOR = <> ,
+	RESTRICT = eqsel, JOIN = eqjoinsel
+);
+
+CREATE OPERATOR <> (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_ne,
+	COMMUTATOR = <> , NEGATOR = = ,
+	RESTRICT = eqsel, JOIN = eqjoinsel
+);
+
+CREATE OPERATOR >= (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_ge,
+	COMMUTATOR = <= , NEGATOR = < ,
+	RESTRICT = scalargesel, JOIN = scalargejoinsel
+);
+
+CREATE OPERATOR > (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_gt,
+	COMMUTATOR = < , NEGATOR = <= ,
+	RESTRICT = scalargtsel, JOIN = scalargtjoinsel
+);
+
 -- halfvec opclasses
 
+CREATE OPERATOR CLASS halfvec_ops
+	DEFAULT FOR TYPE halfvec USING btree AS
+	OPERATOR 1 < ,
+	OPERATOR 2 <= ,
+	OPERATOR 3 = ,
+	OPERATOR 4 >= ,
+	OPERATOR 5 > ,
+	FUNCTION 1 halfvec_cmp(halfvec, halfvec);
+
 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);
+	FUNCTION 3 l2_distance(halfvec, halfvec),
+	FUNCTION 6 halfvec_ivfflat_support(internal);
 
 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);
+	FUNCTION 4 l2_norm(halfvec),
+	FUNCTION 5 l2_normalize(halfvec),
+	FUNCTION 6 halfvec_ivfflat_support(internal);
 
 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 2 l2_norm(halfvec),
 	FUNCTION 3 halfvec_spherical_distance(halfvec, halfvec),
-	FUNCTION 4 halfvec_norm(halfvec);
+	FUNCTION 4 l2_norm(halfvec),
+	FUNCTION 5 l2_normalize(halfvec),
+	FUNCTION 6 halfvec_ivfflat_support(internal);
 
 CREATE OPERATOR CLASS halfvec_l2_ops
 	FOR TYPE halfvec USING hnsw AS
 	OPERATOR 1 <-> (halfvec, halfvec) FOR ORDER BY float_ops,
-	FUNCTION 1 halfvec_l2_squared_distance(halfvec, halfvec);
+	FUNCTION 1 halfvec_l2_squared_distance(halfvec, halfvec),
+	FUNCTION 4 halfvec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS halfvec_ip_ops
 	FOR TYPE halfvec USING hnsw AS
 	OPERATOR 1 <#> (halfvec, halfvec) FOR ORDER BY float_ops,
-	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec);
+	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
+	FUNCTION 4 halfvec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS halfvec_cosine_ops
 	FOR TYPE halfvec USING hnsw AS
 	OPERATOR 1 <=> (halfvec, halfvec) FOR ORDER BY float_ops,
 	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
-	FUNCTION 2 halfvec_norm(halfvec);
+	FUNCTION 2 l2_norm(halfvec),
+	FUNCTION 3 l2_normalize(halfvec),
+	FUNCTION 4 halfvec_hnsw_support(internal);
 
--- extension casts
-
-CREATE FUNCTION halfvec_to_vector(halfvec, integer, boolean) RETURNS vector
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE FUNCTION vector_to_halfvec(vector, integer, boolean) RETURNS halfvec
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-
-CREATE CAST (halfvec AS vector)
-	WITH FUNCTION halfvec_to_vector(halfvec, integer, boolean) AS IMPLICIT;
-
-CREATE CAST (vector AS halfvec)
-	WITH FUNCTION vector_to_halfvec(vector, integer, boolean) AS IMPLICIT;
+CREATE OPERATOR CLASS halfvec_l1_ops
+	FOR TYPE halfvec USING hnsw AS
+	OPERATOR 1 <+> (halfvec, halfvec) FOR ORDER BY float_ops,
+	FUNCTION 1 l1_distance(halfvec, halfvec),
+	FUNCTION 4 halfvec_hnsw_support(internal);
 
 --- sparsevec type
 
@@ -533,11 +724,38 @@ CREATE FUNCTION inner_product(sparsevec, sparsevec) RETURNS float8
 CREATE FUNCTION cosine_distance(sparsevec, sparsevec) RETURNS float8
 	AS 'MODULE_PATHNAME', 'sparsevec_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
-CREATE FUNCTION sparsevec_norm(sparsevec) RETURNS float8
-	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+CREATE FUNCTION l1_distance(sparsevec, sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME', 'sparsevec_l1_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION l2_norm(sparsevec) RETURNS float8
+	AS 'MODULE_PATHNAME', 'sparsevec_l2_norm' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION l2_normalize(sparsevec) RETURNS sparsevec
+	AS 'MODULE_PATHNAME', 'sparsevec_l2_normalize' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
 -- sparsevec private functions
 
+CREATE FUNCTION sparsevec_lt(sparsevec, sparsevec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_le(sparsevec, sparsevec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_eq(sparsevec, sparsevec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_ne(sparsevec, sparsevec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_ge(sparsevec, sparsevec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_gt(sparsevec, sparsevec) RETURNS bool
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_cmp(sparsevec, sparsevec) RETURNS int4
+	AS 'MODULE_PATHNAME' 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;
 
@@ -555,17 +773,32 @@ CREATE FUNCTION vector_to_sparsevec(vector, integer, boolean) RETURNS sparsevec
 CREATE FUNCTION sparsevec_to_vector(sparsevec, integer, boolean) RETURNS vector
 	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
 
+CREATE FUNCTION halfvec_to_sparsevec(halfvec, integer, boolean) RETURNS sparsevec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_to_halfvec(sparsevec, integer, boolean) RETURNS halfvec
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
+CREATE FUNCTION sparsevec_hnsw_support(internal) RETURNS internal
+	AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
+
 -- sparsevec casts
 
 CREATE CAST (sparsevec AS sparsevec)
 	WITH FUNCTION sparsevec(sparsevec, integer, boolean) AS IMPLICIT;
 
 CREATE CAST (sparsevec AS vector)
-	WITH FUNCTION sparsevec_to_vector(sparsevec, integer, boolean) AS IMPLICIT;
+	WITH FUNCTION sparsevec_to_vector(sparsevec, integer, boolean) AS ASSIGNMENT;
 
 CREATE CAST (vector AS sparsevec)
 	WITH FUNCTION vector_to_sparsevec(vector, integer, boolean) AS IMPLICIT;
 
+CREATE CAST (sparsevec AS halfvec)
+	WITH FUNCTION sparsevec_to_halfvec(sparsevec, integer, boolean) AS ASSIGNMENT;
+
+CREATE CAST (halfvec AS sparsevec)
+	WITH FUNCTION halfvec_to_sparsevec(halfvec, integer, boolean) AS IMPLICIT;
+
 -- sparsevec operators
 
 CREATE OPERATOR <-> (
@@ -583,20 +816,80 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
+CREATE OPERATOR <+> (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = l1_distance,
+	COMMUTATOR = '<+>'
+);
+
+CREATE OPERATOR < (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_lt,
+	COMMUTATOR = > , NEGATOR = >= ,
+	RESTRICT = scalarltsel, JOIN = scalarltjoinsel
+);
+
+CREATE OPERATOR <= (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_le,
+	COMMUTATOR = >= , NEGATOR = > ,
+	RESTRICT = scalarlesel, JOIN = scalarlejoinsel
+);
+
+CREATE OPERATOR = (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_eq,
+	COMMUTATOR = = , NEGATOR = <> ,
+	RESTRICT = eqsel, JOIN = eqjoinsel
+);
+
+CREATE OPERATOR <> (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_ne,
+	COMMUTATOR = <> , NEGATOR = = ,
+	RESTRICT = eqsel, JOIN = eqjoinsel
+);
+
+CREATE OPERATOR >= (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_ge,
+	COMMUTATOR = <= , NEGATOR = < ,
+	RESTRICT = scalargesel, JOIN = scalargejoinsel
+);
+
+CREATE OPERATOR > (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_gt,
+	COMMUTATOR = < , NEGATOR = <= ,
+	RESTRICT = scalargtsel, JOIN = scalargtjoinsel
+);
+
 -- sparsevec opclasses
 
+CREATE OPERATOR CLASS sparsevec_ops
+	DEFAULT FOR TYPE sparsevec USING btree AS
+	OPERATOR 1 < ,
+	OPERATOR 2 <= ,
+	OPERATOR 3 = ,
+	OPERATOR 4 >= ,
+	OPERATOR 5 > ,
+	FUNCTION 1 sparsevec_cmp(sparsevec, sparsevec);
+
 CREATE OPERATOR CLASS sparsevec_l2_ops
 	FOR TYPE sparsevec USING hnsw AS
 	OPERATOR 1 <-> (sparsevec, sparsevec) FOR ORDER BY float_ops,
-	FUNCTION 1 sparsevec_l2_squared_distance(sparsevec, sparsevec);
+	FUNCTION 1 sparsevec_l2_squared_distance(sparsevec, sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS sparsevec_ip_ops
 	FOR TYPE sparsevec USING hnsw AS
 	OPERATOR 1 <#> (sparsevec, sparsevec) FOR ORDER BY float_ops,
-	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec);
+	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS sparsevec_cosine_ops
 	FOR TYPE sparsevec USING hnsw AS
 	OPERATOR 1 <=> (sparsevec, sparsevec) FOR ORDER BY float_ops,
 	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec),
-	FUNCTION 2 sparsevec_norm(sparsevec);
+	FUNCTION 2 l2_norm(sparsevec),
+	FUNCTION 3 l2_normalize(sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);
+
+CREATE OPERATOR CLASS sparsevec_l1_ops
+	FOR TYPE sparsevec USING hnsw AS
+	OPERATOR 1 <+> (sparsevec, sparsevec) FOR ORDER BY float_ops,
+	FUNCTION 1 l1_distance(sparsevec, sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);

src/bitutils.c

@@ -0,0 +1,217 @@
+#include "postgres.h"
+
+#include "bitutils.h"
+#include "port/pg_bitutils.h"
+
+#ifdef BIT_DISPATCH
+#include <immintrin.h>
+
+#if defined(HAVE__GET_CPUID)
+#include <cpuid.h>
+#else
+#include <intrin.h>
+#endif
+
+#ifdef _MSC_VER
+#define TARGET_AVX512_POPCOUNT
+#else
+#define TARGET_AVX512_POPCOUNT __attribute__((target("avx512f,avx512vpopcntdq")))
+#endif
+#endif
+
+/* Disable for LLVM due to crash with bitcode generation */
+#if defined(USE_TARGET_CLONES) && !defined(__POPCNT__) && !defined(__llvm__)
+#define BIT_TARGET_CLONES __attribute__((target_clones("default", "popcnt")))
+#else
+#define BIT_TARGET_CLONES
+#endif
+
+/* Use built-ins when possible for inlining */
+#if defined(HAVE__BUILTIN_POPCOUNT) && defined(HAVE_LONG_INT_64)
+#define popcount64(x) __builtin_popcountl(x)
+#elif defined(HAVE__BUILTIN_POPCOUNT) && defined(HAVE_LONG_LONG_INT_64)
+#define popcount64(x) __builtin_popcountll(x)
+#elif !defined(_MSC_VER)
+/* Fails to resolve with MSVC */
+#define popcount64(x) pg_popcount64(x)
+#endif
+
+uint64		(*BitHammingDistance) (uint32 bytes, unsigned char *ax, unsigned char *bx, uint64 distance);
+double		(*BitJaccardDistance) (uint32 bytes, unsigned char *ax, unsigned char *bx, uint64 ab, uint64 aa, uint64 bb);
+
+BIT_TARGET_CLONES static uint64
+BitHammingDistanceDefault(uint32 bytes, unsigned char *ax, unsigned char *bx, uint64 distance)
+{
+#ifdef popcount64
+	for (; bytes >= sizeof(uint64); bytes -= sizeof(uint64))
+	{
+		uint64		axs;
+		uint64		bxs;
+
+		/* Ensure aligned */
+		memcpy(&axs, ax, sizeof(uint64));
+		memcpy(&bxs, bx, sizeof(uint64));
+
+		distance += popcount64(axs ^ bxs);
+
+		ax += sizeof(uint64);
+		bx += sizeof(uint64);
+	}
+#endif
+
+	for (uint32 i = 0; i < bytes; i++)
+		distance += pg_number_of_ones[ax[i] ^ bx[i]];
+
+	return distance;
+}
+
+#ifdef BIT_DISPATCH
+TARGET_AVX512_POPCOUNT static uint64
+BitHammingDistanceAvx512Popcount(uint32 bytes, unsigned char *ax, unsigned char *bx, uint64 distance)
+{
+	__m512i		dist = _mm512_setzero_si512();
+
+	for (; bytes >= sizeof(__m512i); bytes -= sizeof(__m512i))
+	{
+		__m512i		axs = _mm512_loadu_si512((const __m512i *) ax);
+		__m512i		bxs = _mm512_loadu_si512((const __m512i *) bx);
+
+		dist = _mm512_add_epi64(dist, _mm512_popcnt_epi64(_mm512_xor_si512(axs, bxs)));
+
+		ax += sizeof(__m512i);
+		bx += sizeof(__m512i);
+	}
+
+	distance += _mm512_reduce_add_epi64(dist);
+
+	return BitHammingDistanceDefault(bytes, ax, bx, distance);
+}
+#endif
+
+BIT_TARGET_CLONES static double
+BitJaccardDistanceDefault(uint32 bytes, unsigned char *ax, unsigned char *bx, uint64 ab, uint64 aa, uint64 bb)
+{
+#ifdef popcount64
+	for (; bytes >= sizeof(uint64); bytes -= sizeof(uint64))
+	{
+		uint64		axs;
+		uint64		bxs;
+
+		/* Ensure aligned */
+		memcpy(&axs, ax, sizeof(uint64));
+		memcpy(&bxs, bx, sizeof(uint64));
+
+		ab += popcount64(axs & bxs);
+		aa += popcount64(axs);
+		bb += popcount64(bxs);
+
+		ax += sizeof(uint64);
+		bx += sizeof(uint64);
+	}
+#endif
+
+	for (uint32 i = 0; i < bytes; i++)
+	{
+		ab += pg_number_of_ones[ax[i] & bx[i]];
+		aa += pg_number_of_ones[ax[i]];
+		bb += pg_number_of_ones[bx[i]];
+	}
+
+	if (ab == 0)
+		return 1;
+	else
+		return 1 - (ab / ((double) (aa + bb - ab)));
+}
+
+#ifdef BIT_DISPATCH
+TARGET_AVX512_POPCOUNT static double
+BitJaccardDistanceAvx512Popcount(uint32 bytes, unsigned char *ax, unsigned char *bx, uint64 ab, uint64 aa, uint64 bb)
+{
+	__m512i		abx = _mm512_setzero_si512();
+	__m512i		aax = _mm512_setzero_si512();
+	__m512i		bbx = _mm512_setzero_si512();
+
+	for (; bytes >= sizeof(__m512i); bytes -= sizeof(__m512i))
+	{
+		__m512i		axs = _mm512_loadu_si512((const __m512i *) ax);
+		__m512i		bxs = _mm512_loadu_si512((const __m512i *) bx);
+
+		abx = _mm512_add_epi64(abx, _mm512_popcnt_epi64(_mm512_and_si512(axs, bxs)));
+		aax = _mm512_add_epi64(aax, _mm512_popcnt_epi64(axs));
+		bbx = _mm512_add_epi64(bbx, _mm512_popcnt_epi64(bxs));
+
+		ax += sizeof(__m512i);
+		bx += sizeof(__m512i);
+	}
+
+	ab += _mm512_reduce_add_epi64(abx);
+	aa += _mm512_reduce_add_epi64(aax);
+	bb += _mm512_reduce_add_epi64(bbx);
+
+	return BitJaccardDistanceDefault(bytes, ax, bx, ab, aa, bb);
+}
+#endif
+
+#ifdef BIT_DISPATCH
+#define CPU_FEATURE_OSXSAVE         (1 << 27)	/* F1 ECX */
+#define CPU_FEATURE_AVX512F         (1 << 16)	/* F7,0 EBX */
+#define CPU_FEATURE_AVX512VPOPCNTDQ (1 << 14)	/* F7,0 ECX */
+
+#ifdef _MSC_VER
+#define TARGET_XSAVE
+#else
+#define TARGET_XSAVE __attribute__((target("xsave")))
+#endif
+
+TARGET_XSAVE static bool
+SupportsAvx512Popcount()
+{
+	unsigned int exx[4] = {0, 0, 0, 0};
+
+#if defined(HAVE__GET_CPUID)
+	__get_cpuid(1, &exx[0], &exx[1], &exx[2], &exx[3]);
+#else
+	__cpuid(exx, 1);
+#endif
+
+	/* Check OS supports XSAVE */
+	if ((exx[2] & CPU_FEATURE_OSXSAVE) != CPU_FEATURE_OSXSAVE)
+		return false;
+
+	/* Check XMM, YMM, and ZMM registers are enabled */
+	if ((_xgetbv(0) & 0xe6) != 0xe6)
+		return false;
+
+#if defined(HAVE__GET_CPUID)
+	__get_cpuid_count(7, 0, &exx[0], &exx[1], &exx[2], &exx[3]);
+#else
+	__cpuidex(exx, 7, 0);
+#endif
+
+	/* Check AVX512F */
+	if ((exx[1] & CPU_FEATURE_AVX512F) != CPU_FEATURE_AVX512F)
+		return false;
+
+	/* Check AVX512VPOPCNTDQ */
+	return (exx[2] & CPU_FEATURE_AVX512VPOPCNTDQ) == CPU_FEATURE_AVX512VPOPCNTDQ;
+}
+#endif
+
+void
+BitvecInit(void)
+{
+	/*
+	 * Could skip pointer when single function, but no difference in
+	 * performance
+	 */
+	BitHammingDistance = BitHammingDistanceDefault;
+	BitJaccardDistance = BitJaccardDistanceDefault;
+
+#ifdef BIT_DISPATCH
+	if (SupportsAvx512Popcount())
+	{
+		BitHammingDistance = BitHammingDistanceAvx512Popcount;
+		BitJaccardDistance = BitJaccardDistanceAvx512Popcount;
+	}
+#endif
+}

src/bitutils.h

@@ -0,0 +1,31 @@
+#ifndef BITUTILS_H
+#define BITUTILS_H
+
+/* We use two types of dispatching: intrinsics and target_clones */
+/* TODO Move to better place */
+#ifndef DISABLE_DISPATCH
+/* Only enable for more recent compilers to keep build process simple */
+#if defined(__x86_64__) && defined(__GNUC__) && __GNUC__ >= 8
+#define USE_DISPATCH
+#elif defined(__x86_64__) && defined(__clang_major__) && __clang_major__ >= 7
+#define USE_DISPATCH
+#elif defined(_M_AMD64) && defined(_MSC_VER) && _MSC_VER >= 1920
+#define USE_DISPATCH
+#endif
+#endif
+
+/* target_clones requires glibc */
+#if defined(USE_DISPATCH) && defined(__gnu_linux__)
+#define USE_TARGET_CLONES
+#endif
+
+#if defined(USE_DISPATCH)
+#define BIT_DISPATCH
+#endif
+
+extern uint64 (*BitHammingDistance) (uint32 bytes, unsigned char *ax, unsigned char *bx, uint64 distance);
+extern double (*BitJaccardDistance) (uint32 bytes, unsigned char *ax, unsigned char *bx, uint64 ab, uint64 aa, uint64 bb);
+
+void		BitvecInit(void);
+
+#endif

src/bitvec.c

@@ -1,7 +1,7 @@
 #include "postgres.h"
 
-#include "bitvector.h"
-#include "port/pg_bitutils.h"
+#include "bitutils.h"
+#include "bitvec.h"
 #include "utils/varbit.h"
 
 #if PG_VERSION_NUM >= 160000
@@ -46,17 +46,10 @@ hamming_distance(PG_FUNCTION_ARGS)
 {
 	VarBit	   *a = PG_GETARG_VARBIT_P(0);
 	VarBit	   *b = PG_GETARG_VARBIT_P(1);
-	unsigned char *ax = VARBITS(a);
-	unsigned char *bx = VARBITS(b);
-	uint64		distance = 0;
 
 	CheckDims(a, b);
 
-	/* TODO Improve performance */
-	for (uint32 i = 0; i < VARBITBYTES(a); i++)
-		distance += pg_number_of_ones[ax[i] ^ bx[i]];
-
-	PG_RETURN_FLOAT8((double) distance);
+	PG_RETURN_FLOAT8((double) BitHammingDistance(VARBITBYTES(a), VARBITS(a), VARBITS(b), 0));
 }
 
 /*
@@ -68,23 +61,8 @@ jaccard_distance(PG_FUNCTION_ARGS)
 {
 	VarBit	   *a = PG_GETARG_VARBIT_P(0);
 	VarBit	   *b = PG_GETARG_VARBIT_P(1);
-	unsigned char *ax = VARBITS(a);
-	unsigned char *bx = VARBITS(b);
-	uint64		ab = 0;
-	uint64		aa;
-	uint64		bb;
 
 	CheckDims(a, b);
 
-	/* TODO Improve performance */
-	for (uint32 i = 0; i < VARBITBYTES(a); i++)
-		ab += pg_number_of_ones[ax[i] & bx[i]];
-
-	if (ab == 0)
-		PG_RETURN_FLOAT8(1);
-
-	aa = pg_popcount((char *) ax, VARBITBYTES(a));
-	bb = pg_popcount((char *) bx, VARBITBYTES(b));
-
-	PG_RETURN_FLOAT8(1 - (ab / ((double) (aa + bb - ab))));
+	PG_RETURN_FLOAT8(BitJaccardDistance(VARBITBYTES(a), VARBITS(a), VARBITS(b), 0, 0, 0));
 }

src/halfutils.c

@@ -8,19 +8,21 @@
 
 #if defined(HAVE__GET_CPUID)
 #include <cpuid.h>
-#elif defined(HAVE__CPUID)
+#else
 #include <intrin.h>
 #endif
 
 #ifdef _MSC_VER
-#define TARGET_F16C_FMA
+#define TARGET_F16C
 #else
-#define TARGET_F16C_FMA __attribute__((target("f16c,fma")))
+#define TARGET_F16C __attribute__((target("avx,f16c,fma")))
 #endif
 #endif
 
 float		(*HalfvecL2SquaredDistance) (int dim, half * ax, half * bx);
 float		(*HalfvecInnerProduct) (int dim, half * ax, half * bx);
+double		(*HalfvecCosineSimilarity) (int dim, half * ax, half * bx);
+float		(*HalfvecL1Distance) (int dim, half * ax, half * bx);
 
 static float
 HalfvecL2SquaredDistanceDefault(int dim, half * ax, half * bx)
@@ -39,8 +41,8 @@ HalfvecL2SquaredDistanceDefault(int dim, half * ax, half * bx)
 }
 
 #ifdef HALFVEC_DISPATCH
-TARGET_F16C_FMA static float
-HalfvecL2SquaredDistanceF16cFma(int dim, half * ax, half * bx)
+TARGET_F16C static float
+HalfvecL2SquaredDistanceF16c(int dim, half * ax, half * bx)
 {
 	float		distance;
 	int			i;
@@ -87,8 +89,8 @@ HalfvecInnerProductDefault(int dim, half * ax, half * bx)
 }
 
 #ifdef HALFVEC_DISPATCH
-TARGET_F16C_FMA static float
-HalfvecInnerProductF16cFma(int dim, half * ax, half * bx)
+TARGET_F16C static float
+HalfvecInnerProductF16c(int dim, half * ax, half * bx)
 {
 	float		distance;
 	int			i;
@@ -117,21 +119,156 @@ HalfvecInnerProductF16cFma(int dim, half * ax, half * bx)
 }
 #endif
 
-#ifdef HALFVEC_DISPATCH
-#define CPU_FEATURE_FMA  (1 << 12)
-#define CPU_FEATURE_F16C (1 << 29)
+static double
+HalfvecCosineSimilarityDefault(int dim, half * ax, half * bx)
+{
+	float		similarity = 0.0;
+	float		norma = 0.0;
+	float		normb = 0.0;
 
-static bool
+	/* Auto-vectorized */
+	for (int i = 0; i < dim; i++)
+	{
+		float		axi = HalfToFloat4(ax[i]);
+		float		bxi = HalfToFloat4(bx[i]);
+
+		similarity += axi * bxi;
+		norma += axi * axi;
+		normb += bxi * bxi;
+	}
+
+	/* Use sqrt(a * b) over sqrt(a) * sqrt(b) */
+	return (double) similarity / sqrt((double) norma * (double) normb);
+}
+
+#ifdef HALFVEC_DISPATCH
+TARGET_F16C static double
+HalfvecCosineSimilarityF16c(int dim, half * ax, half * bx)
+{
+	float		similarity;
+	float		norma;
+	float		normb;
+	int			i;
+	float		s[8];
+	int			count = (dim / 8) * 8;
+	__m256		sim = _mm256_setzero_ps();
+	__m256		na = _mm256_setzero_ps();
+	__m256		nb = _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);
+
+		sim = _mm256_fmadd_ps(axs, bxs, sim);
+		na = _mm256_fmadd_ps(axs, axs, na);
+		nb = _mm256_fmadd_ps(bxs, bxs, nb);
+	}
+
+	_mm256_storeu_ps(s, sim);
+	similarity = s[0] + s[1] + s[2] + s[3] + s[4] + s[5] + s[6] + s[7];
+
+	_mm256_storeu_ps(s, na);
+	norma = s[0] + s[1] + s[2] + s[3] + s[4] + s[5] + s[6] + s[7];
+
+	_mm256_storeu_ps(s, nb);
+	normb = s[0] + s[1] + s[2] + s[3] + s[4] + s[5] + s[6] + s[7];
+
+	/* Auto-vectorized */
+	for (; i < dim; i++)
+	{
+		float		axi = HalfToFloat4(ax[i]);
+		float		bxi = HalfToFloat4(bx[i]);
+
+		similarity += axi * bxi;
+		norma += axi * axi;
+		normb += bxi * bxi;
+	}
+
+	/* Use sqrt(a * b) over sqrt(a) * sqrt(b) */
+	return (double) similarity / sqrt((double) norma * (double) normb);
+}
+#endif
+
+static float
+HalfvecL1DistanceDefault(int dim, half * ax, half * bx)
+{
+	float		distance = 0.0;
+
+	/* Auto-vectorized */
+	for (int i = 0; i < dim; i++)
+		distance += fabsf(HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]));
+
+	return distance;
+}
+
+#ifdef HALFVEC_DISPATCH
+/* Does not require FMA, but keep logic simple */
+TARGET_F16C static float
+HalfvecL1DistanceF16c(int dim, half * ax, half * bx)
+{
+	float		distance;
+	int			i;
+	float		s[8];
+	int			count = (dim / 8) * 8;
+	__m256		dist = _mm256_setzero_ps();
+	__m256		sign = _mm256_set1_ps(-0.0);
+
+	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_add_ps(dist, _mm256_andnot_ps(sign, _mm256_sub_ps(axs, bxs)));
+	}
+
+	_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 += fabsf(HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]));
+
+	return distance;
+}
+#endif
+
+#ifdef HALFVEC_DISPATCH
+#define CPU_FEATURE_FMA     (1 << 12)
+#define CPU_FEATURE_OSXSAVE (1 << 27)
+#define CPU_FEATURE_AVX     (1 << 28)
+#define CPU_FEATURE_F16C    (1 << 29)
+
+#ifdef _MSC_VER
+#define TARGET_XSAVE
+#else
+#define TARGET_XSAVE __attribute__((target("xsave")))
+#endif
+
+TARGET_XSAVE 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)
+#else
 	__cpuid(exx, 1);
 #endif
 
+	/* Check OS supports XSAVE */
+	if ((exx[2] & CPU_FEATURE_OSXSAVE) != CPU_FEATURE_OSXSAVE)
+		return false;
+
+	/* Check XMM and YMM registers are enabled */
+	if ((_xgetbv(0) & 6) != 6)
+		return false;
+
+	/* Now check features */
 	return (exx[2] & feature) == feature;
 }
 #endif
@@ -145,12 +282,17 @@ HalfvecInit(void)
 	 */
 	HalfvecL2SquaredDistance = HalfvecL2SquaredDistanceDefault;
 	HalfvecInnerProduct = HalfvecInnerProductDefault;
+	HalfvecCosineSimilarity = HalfvecCosineSimilarityDefault;
+	HalfvecL1Distance = HalfvecL1DistanceDefault;
 
 #ifdef HALFVEC_DISPATCH
-	if (SupportsCpuFeature(CPU_FEATURE_FMA | CPU_FEATURE_F16C))
+	if (SupportsCpuFeature(CPU_FEATURE_AVX | CPU_FEATURE_F16C | CPU_FEATURE_FMA))
 	{
-		HalfvecL2SquaredDistance = HalfvecL2SquaredDistanceF16cFma;
-		HalfvecInnerProduct = HalfvecInnerProductF16cFma;
+		HalfvecL2SquaredDistance = HalfvecL2SquaredDistanceF16c;
+		HalfvecInnerProduct = HalfvecInnerProductF16c;
+		HalfvecCosineSimilarity = HalfvecCosineSimilarityF16c;
+		/* Does not require FMA, but keep logic simple */
+		HalfvecL1Distance = HalfvecL1DistanceF16c;
 	}
 #endif
 }

src/halfutils.h

@@ -12,6 +12,8 @@
 
 extern float (*HalfvecL2SquaredDistance) (int dim, half * ax, half * bx);
 extern float (*HalfvecInnerProduct) (int dim, half * ax, half * bx);
+extern double (*HalfvecCosineSimilarity) (int dim, half * ax, half * bx);
+extern float (*HalfvecL1Distance) (int dim, half * ax, half * bx);
 
 void		HalfvecInit(void);
 
@@ -41,6 +43,19 @@ HalfIsInf(half num)
 #endif
 }
 
+/*
+ * Check if half is zero
+ */
+static inline bool
+HalfIsZero(half num)
+{
+#ifdef FLT16_SUPPORT
+	return num == 0;
+#else
+	return (num & 0x7FFF) == 0x0000;
+#endif
+}
+
 /*
  * Convert a half to a float4
  */

src/halfvec.c

@@ -2,7 +2,7 @@
 
 #include <math.h>
 
-#include "bitvector.h"
+#include "bitvec.h"
 #include "catalog/pg_type.h"
 #include "common/shortest_dec.h"
 #include "fmgr.h"
@@ -11,6 +11,7 @@
 #include "lib/stringinfo.h"
 #include "libpq/pqformat.h"
 #include "port.h"				/* for strtof() */
+#include "sparsevec.h"
 #include "utils/array.h"
 #include "utils/builtins.h"
 #include "utils/float.h"
@@ -23,6 +24,9 @@
 #define TYPALIGN_INT 'i'
 #endif
 
+#define STATE_DIMS(x) (ARR_DIMS(x)[0] - 1)
+#define CreateStateDatums(dim) palloc(sizeof(Datum) * (dim + 1))
+
 /*
  * Get a half from a message buffer
  */
@@ -146,6 +150,38 @@ halfvec_isspace(char ch)
 	return false;
 }
 
+/*
+ * Check state array
+ */
+static float8 *
+CheckStateArray(ArrayType *statearray, const char *caller)
+{
+	if (ARR_NDIM(statearray) != 1 ||
+		ARR_DIMS(statearray)[0] < 1 ||
+		ARR_HASNULL(statearray) ||
+		ARR_ELEMTYPE(statearray) != FLOAT8OID)
+		elog(ERROR, "%s: expected state array", caller);
+	return (float8 *) ARR_DATA_PTR(statearray);
+}
+
+#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
  */
@@ -591,28 +627,11 @@ halfvec_cosine_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;
-	float		norma = 0.0;
-	float		normb = 0.0;
 	double		similarity;
 
 	CheckDims(a, b);
 
-	/* Auto-vectorized */
-	for (int i = 0; i < a->dim; i++)
-	{
-		float		axi = HalfToFloat4(ax[i]);
-		float		bxi = HalfToFloat4(bx[i]);
-
-		distance += axi * bxi;
-		norma += axi * axi;
-		normb += bxi * bxi;
-	}
-
-	/* Use sqrt(a * b) over sqrt(a) * sqrt(b) */
-	similarity = (double) distance / sqrt((double) norma * (double) normb);
+	similarity = HalfvecCosineSimilarity(a->dim, a->x, b->x);
 
 #ifdef _MSC_VER
 	/* /fp:fast may not propagate NaN */
@@ -664,25 +683,30 @@ halfvec_l1_distance(PG_FUNCTION_ARGS)
 {
 	HalfVector *a = PG_GETARG_HALFVEC_P(0);
 	HalfVector *b = PG_GETARG_HALFVEC_P(1);
-	half	   *ax = a->x;
-	half	   *bx = b->x;
-	float		distance = 0.0;
 
 	CheckDims(a, b);
 
-	/* Auto-vectorized */
-	for (int i = 0; i < a->dim; i++)
-		distance += fabsf(HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]));
+	PG_RETURN_FLOAT8((double) HalfvecL1Distance(a->dim, a->x, b->x));
+}
 
-	PG_RETURN_FLOAT8((double) distance);
+/*
+ * Get the dimensions of a half vector
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_vector_dims);
+Datum
+halfvec_vector_dims(PG_FUNCTION_ARGS)
+{
+	HalfVector *a = PG_GETARG_HALFVEC_P(0);
+
+	PG_RETURN_INT32(a->dim);
 }
 
 /*
  * Get the L2 norm of a half vector
  */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_norm);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_l2_norm);
 Datum
-halfvec_norm(PG_FUNCTION_ARGS)
+halfvec_l2_norm(PG_FUNCTION_ARGS)
 {
 	HalfVector *a = PG_GETARG_HALFVEC_P(0);
 	half	   *ax = a->x;
@@ -699,12 +723,195 @@ halfvec_norm(PG_FUNCTION_ARGS)
 	PG_RETURN_FLOAT8(sqrt(norm));
 }
 
+/*
+ * Normalize a half vector with the L2 norm
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_l2_normalize);
+Datum
+halfvec_l2_normalize(PG_FUNCTION_ARGS)
+{
+	HalfVector *a = PG_GETARG_HALFVEC_P(0);
+	half	   *ax = a->x;
+	double		norm = 0;
+	HalfVector *result;
+	half	   *rx;
+
+	result = InitHalfVector(a->dim);
+	rx = result->x;
+
+	/* Auto-vectorized */
+	for (int i = 0; i < a->dim; i++)
+		norm += (double) HalfToFloat4(ax[i]) * (double) HalfToFloat4(ax[i]);
+
+	norm = sqrt(norm);
+
+	/* Return zero vector for zero norm */
+	if (norm > 0)
+	{
+		for (int i = 0; i < a->dim; i++)
+			rx[i] = Float4ToHalfUnchecked(HalfToFloat4(ax[i]) / norm);
+
+		/* Check for overflow */
+		for (int i = 0; i < a->dim; i++)
+		{
+			if (HalfIsInf(rx[i]))
+				float_overflow_error();
+		}
+	}
+
+	PG_RETURN_POINTER(result);
+}
+
+/*
+ * Add half vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_add);
+Datum
+halfvec_add(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;
+	HalfVector *result;
+	half	   *rx;
+
+	CheckDims(a, b);
+
+	result = InitHalfVector(a->dim);
+	rx = result->x;
+
+	/* Auto-vectorized */
+	for (int i = 0, imax = a->dim; i < imax; i++)
+	{
+#ifdef FLT16_SUPPORT
+		rx[i] = ax[i] + bx[i];
+#else
+		rx[i] = Float4ToHalfUnchecked(HalfToFloat4(ax[i]) + HalfToFloat4(bx[i]));
+#endif
+	}
+
+	/* Check for overflow */
+	for (int i = 0, imax = a->dim; i < imax; i++)
+	{
+		if (HalfIsInf(rx[i]))
+			float_overflow_error();
+	}
+
+	PG_RETURN_POINTER(result);
+}
+
+/*
+ * Subtract half vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_sub);
+Datum
+halfvec_sub(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;
+	HalfVector *result;
+	half	   *rx;
+
+	CheckDims(a, b);
+
+	result = InitHalfVector(a->dim);
+	rx = result->x;
+
+	/* Auto-vectorized */
+	for (int i = 0, imax = a->dim; i < imax; i++)
+	{
+#ifdef FLT16_SUPPORT
+		rx[i] = ax[i] - bx[i];
+#else
+		rx[i] = Float4ToHalfUnchecked(HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]));
+#endif
+	}
+
+	/* Check for overflow */
+	for (int i = 0, imax = a->dim; i < imax; i++)
+	{
+		if (HalfIsInf(rx[i]))
+			float_overflow_error();
+	}
+
+	PG_RETURN_POINTER(result);
+}
+
+/*
+ * Multiply half vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_mul);
+Datum
+halfvec_mul(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;
+	HalfVector *result;
+	half	   *rx;
+
+	CheckDims(a, b);
+
+	result = InitHalfVector(a->dim);
+	rx = result->x;
+
+	/* Auto-vectorized */
+	for (int i = 0, imax = a->dim; i < imax; i++)
+	{
+#ifdef FLT16_SUPPORT
+		rx[i] = ax[i] * bx[i];
+#else
+		rx[i] = Float4ToHalfUnchecked(HalfToFloat4(ax[i]) * HalfToFloat4(bx[i]));
+#endif
+	}
+
+	/* Check for overflow and underflow */
+	for (int i = 0, imax = a->dim; i < imax; i++)
+	{
+		if (HalfIsInf(rx[i]))
+			float_overflow_error();
+
+		if (HalfIsZero(rx[i]) && !(HalfIsZero(ax[i]) || HalfIsZero(bx[i])))
+			float_underflow_error();
+	}
+
+	PG_RETURN_POINTER(result);
+}
+
+/*
+ * Concatenate half vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_concat);
+Datum
+halfvec_concat(PG_FUNCTION_ARGS)
+{
+	HalfVector *a = PG_GETARG_HALFVEC_P(0);
+	HalfVector *b = PG_GETARG_HALFVEC_P(1);
+	HalfVector *result;
+	int			dim = a->dim + b->dim;
+
+	CheckDim(dim);
+	result = InitHalfVector(dim);
+
+	for (int i = 0; i < a->dim; i++)
+		result->x[i] = a->x[i];
+
+	for (int i = 0; i < b->dim; i++)
+		result->x[i + a->dim] = b->x[i];
+
+	PG_RETURN_POINTER(result);
+}
+
 /*
  * Quantize a half vector
  */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_quantize_binary);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_binary_quantize);
 Datum
-halfvec_quantize_binary(PG_FUNCTION_ARGS)
+halfvec_binary_quantize(PG_FUNCTION_ARGS)
 {
 	HalfVector *a = PG_GETARG_HALFVEC_P(0);
 	half	   *ax = a->x;
@@ -775,3 +982,212 @@ halfvec_cmp_internal(HalfVector * a, HalfVector * b)
 
 	return 0;
 }
+
+/*
+ * Less than
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_lt);
+Datum
+halfvec_lt(PG_FUNCTION_ARGS)
+{
+	HalfVector *a = PG_GETARG_HALFVEC_P(0);
+	HalfVector *b = PG_GETARG_HALFVEC_P(1);
+
+	PG_RETURN_BOOL(halfvec_cmp_internal(a, b) < 0);
+}
+
+/*
+ * Less than or equal
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_le);
+Datum
+halfvec_le(PG_FUNCTION_ARGS)
+{
+	HalfVector *a = PG_GETARG_HALFVEC_P(0);
+	HalfVector *b = PG_GETARG_HALFVEC_P(1);
+
+	PG_RETURN_BOOL(halfvec_cmp_internal(a, b) <= 0);
+}
+
+/*
+ * Equal
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_eq);
+Datum
+halfvec_eq(PG_FUNCTION_ARGS)
+{
+	HalfVector *a = PG_GETARG_HALFVEC_P(0);
+	HalfVector *b = PG_GETARG_HALFVEC_P(1);
+
+	PG_RETURN_BOOL(halfvec_cmp_internal(a, b) == 0);
+}
+
+/*
+ * Not equal
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_ne);
+Datum
+halfvec_ne(PG_FUNCTION_ARGS)
+{
+	HalfVector *a = PG_GETARG_HALFVEC_P(0);
+	HalfVector *b = PG_GETARG_HALFVEC_P(1);
+
+	PG_RETURN_BOOL(halfvec_cmp_internal(a, b) != 0);
+}
+
+/*
+ * Greater than or equal
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_ge);
+Datum
+halfvec_ge(PG_FUNCTION_ARGS)
+{
+	HalfVector *a = PG_GETARG_HALFVEC_P(0);
+	HalfVector *b = PG_GETARG_HALFVEC_P(1);
+
+	PG_RETURN_BOOL(halfvec_cmp_internal(a, b) >= 0);
+}
+
+/*
+ * Greater than
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_gt);
+Datum
+halfvec_gt(PG_FUNCTION_ARGS)
+{
+	HalfVector *a = PG_GETARG_HALFVEC_P(0);
+	HalfVector *b = PG_GETARG_HALFVEC_P(1);
+
+	PG_RETURN_BOOL(halfvec_cmp_internal(a, b) > 0);
+}
+
+/*
+ * Compare half vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_cmp);
+Datum
+halfvec_cmp(PG_FUNCTION_ARGS)
+{
+	HalfVector *a = PG_GETARG_HALFVEC_P(0);
+	HalfVector *b = PG_GETARG_HALFVEC_P(1);
+
+	PG_RETURN_INT32(halfvec_cmp_internal(a, b));
+}
+
+/*
+ * Accumulate half vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_accum);
+Datum
+halfvec_accum(PG_FUNCTION_ARGS)
+{
+	ArrayType  *statearray = PG_GETARG_ARRAYTYPE_P(0);
+	HalfVector *newval = PG_GETARG_HALFVEC_P(1);
+	float8	   *statevalues;
+	int16		dim;
+	bool		newarr;
+	float8		n;
+	Datum	   *statedatums;
+	half	   *x = newval->x;
+	ArrayType  *result;
+
+	/* Check array before using */
+	statevalues = CheckStateArray(statearray, "halfvec_accum");
+	dim = STATE_DIMS(statearray);
+	newarr = dim == 0;
+
+	if (newarr)
+		dim = newval->dim;
+	else
+		CheckExpectedDim(dim, newval->dim);
+
+	n = statevalues[0] + 1.0;
+
+	statedatums = CreateStateDatums(dim);
+	statedatums[0] = Float8GetDatum(n);
+
+	if (newarr)
+	{
+		for (int i = 0; i < dim; i++)
+			statedatums[i + 1] = Float8GetDatum((double) HalfToFloat4(x[i]));
+	}
+	else
+	{
+		for (int i = 0; i < dim; i++)
+		{
+			double		v = statevalues[i + 1] + (double) HalfToFloat4(x[i]);
+
+			/* Check for overflow */
+			if (isinf(v))
+				float_overflow_error();
+
+			statedatums[i + 1] = Float8GetDatum(v);
+		}
+	}
+
+	/* Use float8 array like float4_accum */
+	result = construct_array(statedatums, dim + 1,
+							 FLOAT8OID,
+							 sizeof(float8), FLOAT8PASSBYVAL, TYPALIGN_DOUBLE);
+
+	pfree(statedatums);
+
+	PG_RETURN_ARRAYTYPE_P(result);
+}
+
+/*
+ * Average half vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_avg);
+Datum
+halfvec_avg(PG_FUNCTION_ARGS)
+{
+	ArrayType  *statearray = PG_GETARG_ARRAYTYPE_P(0);
+	float8	   *statevalues;
+	float8		n;
+	uint16		dim;
+	HalfVector *result;
+
+	/* Check array before using */
+	statevalues = CheckStateArray(statearray, "halfvec_avg");
+	n = statevalues[0];
+
+	/* SQL defines AVG of no values to be NULL */
+	if (n == 0.0)
+		PG_RETURN_NULL();
+
+	/* Create half vector */
+	dim = STATE_DIMS(statearray);
+	CheckDim(dim);
+	result = InitHalfVector(dim);
+	for (int i = 0; i < dim; i++)
+	{
+		result->x[i] = Float4ToHalf(statevalues[i + 1] / n);
+		CheckElement(result->x[i]);
+	}
+
+	PG_RETURN_POINTER(result);
+}
+
+/*
+ * Convert sparse vector to half vector
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_to_halfvec);
+Datum
+sparsevec_to_halfvec(PG_FUNCTION_ARGS)
+{
+	SparseVector *svec = PG_GETARG_SPARSEVEC_P(0);
+	int32		typmod = PG_GETARG_INT32(1);
+	HalfVector *result;
+	int			dim = svec->dim;
+	float	   *values = SPARSEVEC_VALUES(svec);
+
+	CheckDim(dim);
+	CheckExpectedDim(typmod, dim);
+
+	result = InitHalfVector(dim);
+	for (int i = 0; i < svec->nnz; i++)
+		result->x[svec->indices[i] - 1] = Float4ToHalf(values[i]);
+
+	PG_RETURN_POINTER(result);
+}

src/halfvec.h

@@ -5,9 +5,11 @@
 
 #include <float.h>
 
+#include "bitutils.h"
+#include "fmgr.h"
 #include "vector.h"
 
-#if defined(__x86_64__) || defined(_M_AMD64)
+#if defined(USE_DISPATCH)
 #define HALFVEC_DISPATCH
 #endif
 

src/hnsw.c

@@ -194,7 +194,7 @@ hnswhandler(PG_FUNCTION_ARGS)
 	IndexAmRoutine *amroutine = makeNode(IndexAmRoutine);
 
 	amroutine->amstrategies = 0;
-	amroutine->amsupport = 2;
+	amroutine->amsupport = 4;
 #if PG_VERSION_NUM >= 130000
 	amroutine->amoptsprocnum = 0;
 #endif

src/hnsw.h

@@ -22,6 +22,8 @@
 /* Support functions */
 #define HNSW_DISTANCE_PROC 1
 #define HNSW_NORM_PROC 2
+#define HNSW_NORMALIZE_PROC 3
+#define HNSW_TYPE_SUPPORT_PROC 4
 
 #define HNSW_VERSION	1
 #define HNSW_MAGIC_NUMBER 0xA953A953
@@ -61,7 +63,8 @@ typedef enum HnswType
 	HNSW_TYPE_VECTOR,
 	HNSW_TYPE_HALFVEC,
 	HNSW_TYPE_BIT,
-	HNSW_TYPE_SPARSEVEC
+	HNSW_TYPE_SPARSEVEC,
+	HNSW_TYPE_UNSUPPORTED
 }			HnswType;
 
 /* Build phases */
@@ -265,6 +268,7 @@ typedef struct HnswBuildState
 	/* Support functions */
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
+	FmgrInfo   *normalizeprocinfo;
 	Oid			collation;
 
 	/* Variables */
@@ -341,6 +345,7 @@ typedef struct HnswScanOpaqueData
 	/* Support functions */
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
+	FmgrInfo   *normalizeprocinfo;
 	Oid			collation;
 }			HnswScanOpaqueData;
 
@@ -377,7 +382,8 @@ int			HnswGetM(Relation index);
 int			HnswGetEfConstruction(Relation index);
 FmgrInfo   *HnswOptionalProcInfo(Relation index, uint16 procnum);
 HnswType	HnswGetType(Relation index);
-bool		HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, HnswType type);
+Datum		HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum value);
+bool		HnswCheckNorm(FmgrInfo *procinfo, Oid collation, Datum value);
 void		HnswCheckValue(Datum value, HnswType type);
 Buffer		HnswNewBuffer(Relation index, ForkNumber forkNum);
 void		HnswInitPage(Buffer buf, Page page);

src/hnswbuild.c

@@ -493,8 +493,10 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 	/* Normalize if needed */
 	if (buildstate->normprocinfo != NULL)
 	{
-		if (!HnswNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->type))
+		if (!HnswCheckNorm(buildstate->normprocinfo, buildstate->collation, value))
 			return false;
+
+		value = HnswNormValue(buildstate->normalizeprocinfo, buildstate->collation, value);
 	}
 
 	/* Get datum size */
@@ -723,6 +725,7 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	/* Get support functions */
 	buildstate->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
 	buildstate->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
+	buildstate->normalizeprocinfo = HnswOptionalProcInfo(index, HNSW_NORMALIZE_PROC);
 	buildstate->collation = index->rd_indcollation[0];
 
 	InitGraph(&buildstate->graphData, NULL, maintenance_work_mem * 1024L);

src/hnswinsert.c

@@ -614,20 +614,21 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 	Datum		value;
 	FmgrInfo   *normprocinfo;
 	Oid			collation = index->rd_indcollation[0];
-	HnswType	type = HnswGetType(index);
 
 	/* Detoast once for all calls */
 	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
 	/* Check value */
-	HnswCheckValue(value, type);
+	HnswCheckValue(value, HnswGetType(index));
 
 	/* Normalize if needed */
 	normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
 	if (normprocinfo != NULL)
 	{
-		if (!HnswNormValue(normprocinfo, collation, &value, type))
+		if (!HnswCheckNorm(normprocinfo, collation, value))
 			return;
+
+		value = HnswNormValue(HnswOptionalProcInfo(index, HNSW_NORMALIZE_PROC), collation, value);
 	}
 
 	HnswInsertTupleOnDisk(index, value, values, isnull, heap_tid, false);

src/hnswscan.c

@@ -61,7 +61,7 @@ GetScanValue(IndexScanDesc scan)
 
 		/* Fine if normalization fails */
 		if (so->normprocinfo != NULL)
-			HnswNormValue(so->normprocinfo, so->collation, &value, HnswGetType(scan->indexRelation));
+			value = HnswNormValue(so->normalizeprocinfo, so->collation, value);
 	}
 
 	return value;
@@ -87,6 +87,7 @@ hnswbeginscan(Relation index, int nkeys, int norderbys)
 	/* Set support functions */
 	so->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
 	so->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
+	so->normalizeprocinfo = HnswOptionalProcInfo(index, HNSW_NORMALIZE_PROC);
 	so->collation = index->rd_indcollation[0];
 
 	scan->opaque = so;

src/hnswutils.c

@@ -5,8 +5,7 @@
 #include "access/generic_xlog.h"
 #include "catalog/pg_type.h"
 #include "catalog/pg_type_d.h"
-#include "halfutils.h"
-#include "halfvec.h"
+#include "fmgr.h"
 #include "hnsw.h"
 #include "lib/pairingheap.h"
 #include "sparsevec.h"
@@ -14,8 +13,6 @@
 #include "utils/datum.h"
 #include "utils/memdebug.h"
 #include "utils/rel.h"
-#include "utils/syscache.h"
-#include "vector.h"
 
 #if PG_VERSION_NUM >= 130000
 #include "common/hashfn.h"
@@ -161,94 +158,40 @@ HnswOptionalProcInfo(Relation index, uint16 procnum)
 HnswType
 HnswGetType(Relation index)
 {
+	FmgrInfo   *procinfo = HnswOptionalProcInfo(index, HNSW_TYPE_SUPPORT_PROC);
 	Oid			typid = TupleDescAttr(index->rd_att, 0)->atttypid;
-	HeapTuple	tuple;
-	Form_pg_type type;
 	HnswType	result;
 
-	if (typid == BITOID)
-		return HNSW_TYPE_BIT;
+	if (procinfo == NULL)
+		return HNSW_TYPE_VECTOR;
 
-	tuple = SearchSysCache1(TYPEOID, ObjectIdGetDatum(typid));
-	if (!HeapTupleIsValid(tuple))
-		elog(ERROR, "cache lookup failed for type %u", typid);
+	result = (HnswType) DatumGetInt32(FunctionCall1(procinfo, ObjectIdGetDatum(typid)));
 
-	type = (Form_pg_type) GETSTRUCT(tuple);
-	if (strcmp(NameStr(type->typname), "vector") == 0)
-		result = HNSW_TYPE_VECTOR;
-	else if (strcmp(NameStr(type->typname), "halfvec") == 0)
-		result = HNSW_TYPE_HALFVEC;
-	else if (strcmp(NameStr(type->typname), "sparsevec") == 0)
-		result = HNSW_TYPE_SPARSEVEC;
-	else
-	{
-		ReleaseSysCache(tuple);
+	if (result == HNSW_TYPE_UNSUPPORTED)
 		elog(ERROR, "type not supported for hnsw index");
-	}
-
-	ReleaseSysCache(tuple);
 
 	return result;
 }
 
 /*
- * Divide by the norm
- *
- * Returns false if value should not be indexed
- *
- * The caller needs to free the pointer stored in value
- * if it's different than the original value
+ * Normalize value
+ */
+Datum
+HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum value)
+{
+	if (procinfo == NULL)
+		return DirectFunctionCall1(l2_normalize, value);
+
+	return FunctionCall1Coll(procinfo, collation, value);
+}
+
+/*
+ * Check if non-zero norm
  */
 bool
-HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, HnswType type)
+HnswCheckNorm(FmgrInfo *procinfo, Oid collation, Datum value)
 {
-	double		norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
-
-	if (norm > 0)
-	{
-		/* TODO Remove vector-specific code */
-		if (type == HNSW_TYPE_VECTOR)
-		{
-			Vector	   *v = DatumGetVector(*value);
-			Vector	   *result = InitVector(v->dim);
-
-			for (int i = 0; i < v->dim; i++)
-				result->x[i] = v->x[i] / norm;
-
-			*value = PointerGetDatum(result);
-		}
-		else if (type == HNSW_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 if (type == HNSW_TYPE_SPARSEVEC)
-		{
-			SparseVector *v = DatumGetSparseVector(*value);
-			SparseVector *result = InitSparseVector(v->dim, v->nnz);
-			float	   *vx = SPARSEVEC_VALUES(v);
-			float	   *rx = SPARSEVEC_VALUES(result);
-
-			for (int i = 0; i < v->nnz; i++)
-			{
-				result->indices[i] = v->indices[i];
-				rx[i] = vx[i] / norm;
-			}
-
-			*value = PointerGetDatum(result);
-		}
-		else
-			elog(ERROR, "Unsupported type");
-
-		return true;
-	}
-
-	return false;
+	return DatumGetFloat8(FunctionCall1Coll(procinfo, collation, value)) > 0;
 }
 
 /*
@@ -1359,3 +1302,29 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 		ep = w;
 	}
 }
+
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_hnsw_support);
+Datum
+halfvec_hnsw_support(PG_FUNCTION_ARGS)
+{
+	PG_RETURN_INT32(HNSW_TYPE_HALFVEC);
+};
+
+PGDLLEXPORT PG_FUNCTION_INFO_V1(bit_hnsw_support);
+Datum
+bit_hnsw_support(PG_FUNCTION_ARGS)
+{
+	Oid			typid = PG_GETARG_OID(0);
+
+	if (typid == BITOID)
+		PG_RETURN_INT32(HNSW_TYPE_BIT);
+	else
+		PG_RETURN_INT32(HNSW_TYPE_UNSUPPORTED);
+};
+
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_hnsw_support);
+Datum
+sparsevec_hnsw_support(PG_FUNCTION_ARGS)
+{
+	PG_RETURN_INT32(HNSW_TYPE_SPARSEVEC);
+};

src/ivfbuild.c

@@ -6,6 +6,7 @@
 #include "access/tableam.h"
 #include "access/parallel.h"
 #include "access/xact.h"
+#include "bitvec.h"
 #include "catalog/index.h"
 #include "catalog/pg_operator_d.h"
 #include "catalog/pg_type_d.h"
@@ -59,8 +60,10 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 	 */
 	if (buildstate->kmeansnormprocinfo != NULL)
 	{
-		if (!IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->collation, &value, buildstate->type))
+		if (!IvfflatCheckNorm(buildstate->kmeansnormprocinfo, buildstate->collation, value))
 			return;
+
+		value = IvfflatNormValue(buildstate->normalizeprocinfo, buildstate->collation, value);
 	}
 
 	if (samples->length < targsamples)
@@ -155,8 +158,10 @@ 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 (!IvfflatCheckNorm(buildstate->normprocinfo, buildstate->collation, value))
 			return;
+
+		value = IvfflatNormValue(buildstate->normalizeprocinfo, buildstate->collation, value);
 	}
 
 	/* Find the list that minimizes the distance */
@@ -324,6 +329,8 @@ GetMaxDimensions(IvfflatType type)
 
 	if (type == IVFFLAT_TYPE_HALFVEC)
 		maxDimensions *= 2;
+	else if (type == IVFFLAT_TYPE_BIT)
+		maxDimensions *= 32;
 
 	return maxDimensions;
 }
@@ -338,6 +345,8 @@ GetItemSize(IvfflatType type, int dimensions)
 		return VECTOR_SIZE(dimensions);
 	else if (type == IVFFLAT_TYPE_HALFVEC)
 		return HALFVEC_SIZE(dimensions);
+	else if (type == IVFFLAT_TYPE_BIT)
+		return VARBITTOTALLEN(dimensions);
 	else
 		elog(ERROR, "Unsupported type");
 }
@@ -374,6 +383,7 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 	buildstate->procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
 	buildstate->normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
 	buildstate->kmeansnormprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
+	buildstate->normalizeprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORMALIZE_PROC);
 	buildstate->collation = index->rd_indcollation[0];
 
 	/* Require more than one dimension for spherical k-means */

src/ivfflat.c

@@ -188,7 +188,7 @@ ivfflathandler(PG_FUNCTION_ARGS)
 	IndexAmRoutine *amroutine = makeNode(IndexAmRoutine);
 
 	amroutine->amstrategies = 0;
-	amroutine->amsupport = 4;
+	amroutine->amsupport = 6;
 #if PG_VERSION_NUM >= 130000
 	amroutine->amoptsprocnum = 0;
 #endif

src/ivfflat.h

@@ -28,6 +28,8 @@
 #define IVFFLAT_NORM_PROC 2
 #define IVFFLAT_KMEANS_DISTANCE_PROC 3
 #define IVFFLAT_KMEANS_NORM_PROC 4
+#define IVFFLAT_NORMALIZE_PROC 5
+#define IVFFLAT_TYPE_SUPPORT_PROC 6
 
 #define IVFFLAT_VERSION	1
 #define IVFFLAT_MAGIC_NUMBER 0x14FF1A7
@@ -46,7 +48,9 @@
 typedef enum IvfflatType
 {
 	IVFFLAT_TYPE_VECTOR,
-	IVFFLAT_TYPE_HALFVEC
+	IVFFLAT_TYPE_HALFVEC,
+	IVFFLAT_TYPE_BIT,
+	IVFFLAT_TYPE_UNSUPPORTED
 }			IvfflatType;
 
 /* Build phases */
@@ -174,6 +178,7 @@ typedef struct IvfflatBuildState
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
 	FmgrInfo   *kmeansnormprocinfo;
+	FmgrInfo   *normalizeprocinfo;
 	Oid			collation;
 
 	/* Variables */
@@ -254,7 +259,9 @@ typedef struct IvfflatScanOpaqueData
 	/* Support functions */
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
+	FmgrInfo   *normalizeprocinfo;
 	Oid			collation;
+	Datum		(*distfunc) (FmgrInfo *flinfo, Oid collation, Datum arg1, Datum arg2);
 
 	/* Lists */
 	pairingheap *listQueue;
@@ -263,7 +270,7 @@ typedef struct IvfflatScanOpaqueData
 
 typedef IvfflatScanOpaqueData * IvfflatScanOpaque;
 
-#define VECTOR_ARRAY_SIZE(_length, _size) (sizeof(VectorArrayData) + (_length) * _size)
+#define VECTOR_ARRAY_SIZE(_length, _size) (sizeof(VectorArrayData) + (_length) * MAXALIGN(_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)
@@ -274,7 +281,8 @@ void		VectorArrayFree(VectorArray arr);
 void		IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatType type);
 FmgrInfo   *IvfflatOptionalProcInfo(Relation index, uint16 procnum);
 IvfflatType IvfflatGetType(Relation index);
-bool		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, IvfflatType type);
+Datum		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum value);
+bool		IvfflatCheckNorm(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

@@ -85,8 +85,12 @@ 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)))
+		Oid			collation = index->rd_indcollation[0];
+
+		if (!IvfflatCheckNorm(normprocinfo, collation, value))
 			return;
+
+		value = IvfflatNormValue(IvfflatOptionalProcInfo(index, IVFFLAT_NORMALIZE_PROC), collation, value);
 	}
 
 	/* Find the insert page - sets the page and list info */

src/ivfkmeans.c

@@ -3,10 +3,12 @@
 #include <float.h>
 #include <math.h>
 
+#include "bitvec.h"
 #include "halfutils.h"
 #include "halfvec.h"
 #include "ivfflat.h"
 #include "miscadmin.h"
+#include "utils/builtins.h"
 #include "utils/datum.h"
 #include "utils/memutils.h"
 #include "vector.h"
@@ -134,6 +136,15 @@ CompareHalfVectors(const void *a, const void *b)
 	return halfvec_cmp_internal((HalfVector *) a, (HalfVector *) b);
 }
 
+/*
+ * Compare bit vectors
+ */
+static int
+CompareBitVectors(const void *a, const void *b)
+{
+	return DirectFunctionCall2(bitcmp, VarBitPGetDatum((VarBit *) a), VarBitPGetDatum((VarBit *) b));
+}
+
 /*
  * Quick approach if we have little data
  */
@@ -151,6 +162,8 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers, IvfflatTy
 			qsort(samples->items, samples->length, samples->itemsize, CompareVectors);
 		else if (type == IVFFLAT_TYPE_HALFVEC)
 			qsort(samples->items, samples->length, samples->itemsize, CompareHalfVectors);
+		else if (type == IVFFLAT_TYPE_BIT)
+			qsort(samples->items, samples->length, samples->itemsize, CompareBitVectors);
 		else
 			elog(ERROR, "Unsupported type");
 
@@ -191,6 +204,16 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers, IvfflatTy
 			for (int j = 0; j < dimensions; j++)
 				vec->x[j] = Float4ToHalfUnchecked((float) RandomDouble());
 		}
+		else if (type == IVFFLAT_TYPE_BIT)
+		{
+			VarBit	   *vec = DatumGetVarBitP(center);
+
+			SET_VARSIZE(vec, VARBITTOTALLEN(dimensions));
+			VARBITLEN(vec) = dimensions;
+
+			for (int j = 0; j < dimensions; j++)
+				VARBITS(vec)[j / dimensions] |= (RandomDouble() > 0.5 ? 1 : 0) << (7 - (j % 8));
+		}
 		else
 			elog(ERROR, "Unsupported type");
 
@@ -263,6 +286,17 @@ ComputeNewCenters(VectorArray samples, VectorArray aggCenters, VectorArray newCe
 				aggCenter->x[k] += HalfToFloat4(vec->x[k]);
 		}
 	}
+	else if (type == IVFFLAT_TYPE_BIT)
+	{
+		for (int j = 0; j < numSamples; j++)
+		{
+			Vector	   *aggCenter = (Vector *) VectorArrayGet(aggCenters, closestCenters[j]);
+			VarBit	   *vec = (VarBit *) VectorArrayGet(samples, j);
+
+			for (int k = 0; k < dimensions; k++)
+				aggCenter->x[k] += (float) (((VARBITS(vec)[k / 8]) >> (7 - (k % 8))) & 0x01);
+		}
+	}
 	else
 		elog(ERROR, "Unsupported type");
 
@@ -308,6 +342,21 @@ ComputeNewCenters(VectorArray samples, VectorArray aggCenters, VectorArray newCe
 				newCenter->x[k] = Float4ToHalfUnchecked(aggCenter->x[k]);
 		}
 	}
+	else if (type == IVFFLAT_TYPE_BIT)
+	{
+		for (int j = 0; j < numCenters; j++)
+		{
+			Vector	   *aggCenter = (Vector *) VectorArrayGet(aggCenters, j);
+			VarBit	   *newCenter = (VarBit *) VectorArrayGet(newCenters, j);
+			unsigned char *nx = VARBITS(newCenter);
+
+			for (uint32 k = 0; k < VARBITBYTES(newCenter); k++)
+				nx[k] = 0;
+
+			for (int k = 0; k < dimensions; k++)
+				nx[k / 8] |= (aggCenter->x[k] > 0.5) << (7 - (k % 8));
+		}
+	}
 
 	/* Normalize if needed */
 	if (normprocinfo != NULL)
@@ -425,6 +474,18 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 			vec->dim = dimensions;
 		}
 	}
+	else if (type == IVFFLAT_TYPE_BIT)
+	{
+		newCenters = VectorArrayInit(numCenters, dimensions, centers->itemsize);
+
+		for (int j = 0; j < numCenters; j++)
+		{
+			VarBit	   *vec = (VarBit *) VectorArrayGet(newCenters, j);
+
+			SET_VARSIZE(vec, VARBITTOTALLEN(dimensions));
+			VARBITLEN(vec) = dimensions;
+		}
+	}
 	else
 		elog(ERROR, "Unsupported type");
 
@@ -642,23 +703,26 @@ CheckCenters(Relation index, VectorArray centers, IvfflatType type)
 					elog(ERROR, "Infinite value detected. Please report a bug.");
 			}
 		}
-		else
+		else if (type != IVFFLAT_TYPE_BIT)
 			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");
-
-	for (int i = 1; i < centers->length; i++)
+	if (type != IVFFLAT_TYPE_BIT)
 	{
-		if (datumIsEqual(PointerGetDatum(VectorArrayGet(centers, i)), PointerGetDatum(VectorArrayGet(centers, i - 1)), false, -1))
-			elog(ERROR, "Duplicate centers detected. Please report a bug.");
+		/* 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");
+
+		for (int i = 1; i < centers->length; i++)
+		{
+			if (datumIsEqual(PointerGetDatum(VectorArrayGet(centers, i)), PointerGetDatum(VectorArrayGet(centers, i - 1)), false, -1))
+				elog(ERROR, "Duplicate centers detected. Please report a bug.");
+		}
 	}
 
 	/* Ensure no zero vectors for cosine distance */

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"
@@ -57,7 +56,7 @@ GetScanLists(IndexScanDesc scan, Datum value)
 			double		distance;
 
 			/* Use procinfo from the index instead of scan key for performance */
-			distance = DatumGetFloat8(FunctionCall2Coll(so->procinfo, so->collation, PointerGetDatum(&list->center), value));
+			distance = DatumGetFloat8(so->distfunc(so->procinfo, so->collation, PointerGetDatum(&list->center), value));
 
 			if (listCount < so->probes)
 			{
@@ -150,7 +149,7 @@ GetScanItems(IndexScanDesc scan, Datum value)
 				 * performance
 				 */
 				ExecClearTuple(slot);
-				slot->tts_values[0] = FunctionCall2Coll(so->procinfo, so->collation, datum, value);
+				slot->tts_values[0] = so->distfunc(so->procinfo, so->collation, datum, value);
 				slot->tts_isnull[0] = false;
 				slot->tts_values[1] = PointerGetDatum(&itup->t_tid);
 				slot->tts_isnull[1] = false;
@@ -178,6 +177,15 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	tuplesort_performsort(so->sortstate);
 }
 
+/*
+ * Zero distance
+ */
+static Datum
+ZeroDistance(FmgrInfo *flinfo, Oid collation, Datum arg1, Datum arg2)
+{
+	return Float8GetDatum(0.0);
+}
+
 /*
  * Get scan value
  */
@@ -189,26 +197,21 @@ GetScanValue(IndexScanDesc scan)
 
 	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");
+		value = PointerGetDatum(NULL);
+		so->distfunc = ZeroDistance;
 	}
 	else
 	{
 		value = scan->orderByData->sk_argument;
+		so->distfunc = FunctionCall2Coll;
 
 		/* Value should not be compressed or toasted */
 		Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
 		Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
 
-		/* Fine if normalization fails */
+		/* Check normprocinfo since normalizeprocinfo not set for vector */
 		if (so->normprocinfo != NULL)
-			IvfflatNormValue(so->normprocinfo, so->collation, &value, IvfflatGetType(scan->indexRelation));
+			value = IvfflatNormValue(so->normalizeprocinfo, so->collation, value);
 	}
 
 	return value;
@@ -246,6 +249,7 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	/* Set support functions */
 	so->procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
 	so->normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
+	so->normalizeprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORMALIZE_PROC);
 	so->collation = index->rd_indcollation[0];
 
 	/* Create tuple description for sorting */

src/ivfutils.c

@@ -2,12 +2,9 @@
 
 #include "access/generic_xlog.h"
 #include "catalog/pg_type.h"
-#include "halfutils.h"
-#include "halfvec.h"
+#include "fmgr.h"
 #include "ivfflat.h"
 #include "storage/bufmgr.h"
-#include "vector.h"
-#include "utils/syscache.h"
 
 /*
  * Allocate a vector array
@@ -17,6 +14,9 @@ VectorArrayInit(int maxlen, int dimensions, Size itemsize)
 {
 	VectorArray res = palloc(sizeof(VectorArrayData));
 
+	/* Ensure items are aligned to prevent UB */
+	itemsize = MAXALIGN(itemsize);
+
 	res->length = 0;
 	res->maxlen = maxlen;
 	res->dim = dimensions;
@@ -67,73 +67,40 @@ IvfflatOptionalProcInfo(Relation index, uint16 procnum)
 IvfflatType
 IvfflatGetType(Relation index)
 {
+	FmgrInfo   *procinfo = IvfflatOptionalProcInfo(index, IVFFLAT_TYPE_SUPPORT_PROC);
 	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);
+	if (procinfo == NULL)
+		return IVFFLAT_TYPE_VECTOR;
 
-	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);
+	result = (IvfflatType) DatumGetInt32(FunctionCall1(procinfo, ObjectIdGetDatum(typid)));
+
+	if (result == IVFFLAT_TYPE_UNSUPPORTED)
 		elog(ERROR, "type not supported for ivfflat index");
-	}
-
-	ReleaseSysCache(tuple);
 
 	return result;
 }
 
 /*
- * Divide by the norm
- *
- * Returns false if value should not be indexed
- *
- * The caller needs to free the pointer stored in value
- * if it's different than the original value
+ * Normalize value
+ */
+Datum
+IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum value)
+{
+	if (procinfo == NULL)
+		return DirectFunctionCall1(l2_normalize, value);
+
+	return FunctionCall1Coll(procinfo, collation, value);
+}
+
+/*
+ * Check if non-zero norm
  */
 bool
-IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, IvfflatType type)
+IvfflatCheckNorm(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);
-
-			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");
-
-		return true;
-	}
-
-	return false;
+	return DatumGetFloat8(FunctionCall1Coll(procinfo, collation, value)) > 0;
 }
 
 /*
@@ -278,3 +245,22 @@ IvfflatUpdateList(Relation index, ListInfo listInfo,
 		UnlockReleaseBuffer(buf);
 	}
 }
+
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_ivfflat_support);
+Datum
+halfvec_ivfflat_support(PG_FUNCTION_ARGS)
+{
+	PG_RETURN_INT32(IVFFLAT_TYPE_HALFVEC);
+};
+
+PGDLLEXPORT PG_FUNCTION_INFO_V1(bit_ivfflat_support);
+Datum
+bit_ivfflat_support(PG_FUNCTION_ARGS)
+{
+	Oid			typid = PG_GETARG_OID(0);
+
+	if (typid == BITOID)
+		PG_RETURN_INT32(IVFFLAT_TYPE_BIT);
+	else
+		PG_RETURN_INT32(IVFFLAT_TYPE_UNSUPPORTED);
+};

src/sparsevec.c

@@ -4,6 +4,8 @@
 #include <math.h>
 
 #include "fmgr.h"
+#include "halfutils.h"
+#include "halfvec.h"
 #include "libpq/pqformat.h"
 #include "sparsevec.h"
 #include "utils/array.h"
@@ -18,6 +20,12 @@
 #include "utils/builtins.h"
 #endif
 
+typedef struct SparseInputElement
+{
+	int32		index;
+	float		value;
+}			SparseInputElement;
+
 /*
  * Ensure same dimensions
  */
@@ -164,6 +172,21 @@ sparsevec_isspace(char ch)
 	return false;
 }
 
+/*
+ * Compare indices
+ */
+static int
+CompareIndices(const void *a, const void *b)
+{
+	if (((SparseInputElement *) a)->index < ((SparseInputElement *) b)->index)
+		return -1;
+
+	if (((SparseInputElement *) a)->index > ((SparseInputElement *) b)->index)
+		return 1;
+
+	return 0;
+}
+
 /*
  * Convert textual representation to internal representation
  */
@@ -178,8 +201,7 @@ sparsevec_in(PG_FUNCTION_ARGS)
 	char	   *stringEnd;
 	SparseVector *result;
 	float	   *rvalues;
-	int32	   *indices;
-	float	   *values;
+	SparseInputElement *elements;
 	int			maxNnz;
 	int			nnz = 0;
 
@@ -197,8 +219,7 @@ sparsevec_in(PG_FUNCTION_ARGS)
 				(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));
+	elements = palloc(maxNnz * sizeof(SparseInputElement));
 
 	pt = lit;
 
@@ -291,8 +312,8 @@ sparsevec_in(PG_FUNCTION_ARGS)
 			/* Do not store zero values */
 			if (value != 0)
 			{
-				indices[nnz] = index;
-				values[nnz] = value;
+				elements[nnz].index = index;
+				elements[nnz].value = value;
 				nnz++;
 			}
 
@@ -353,12 +374,14 @@ sparsevec_in(PG_FUNCTION_ARGS)
 	CheckDim(dim);
 	CheckExpectedDim(typmod, dim);
 
+	qsort(elements, nnz, sizeof(SparseInputElement), CompareIndices);
+
 	result = InitSparseVector(dim, nnz);
 	rvalues = SPARSEVEC_VALUES(result);
 	for (int i = 0; i < nnz; i++)
 	{
-		result->indices[i] = indices[i];
-		rvalues[i] = values[i];
+		result->indices[i] = elements[i].index;
+		rvalues[i] = elements[i].value;
 
 		CheckIndex(result->indices, i, dim);
 	}
@@ -504,6 +527,10 @@ sparsevec_recv(PG_FUNCTION_ARGS)
 	{
 		values[i] = pq_getmsgfloat4(buf);
 		CheckElement(values[i]);
+		if (values[i] == 0)
+			ereport(ERROR,
+					(errcode(ERRCODE_DATA_EXCEPTION),
+					 errmsg("binary representation of sparsevec cannot contain zero values")));
 	}
 
 	PG_RETURN_POINTER(result);
@@ -579,7 +606,7 @@ vector_to_sparsevec(PG_FUNCTION_ARGS)
 		if (vec->x[i] != 0)
 		{
 			/* Safety check */
-			if (j == nnz)
+			if (j >= result->nnz)
 				elog(ERROR, "safety check failed");
 
 			result->indices[j] = i + 1;
@@ -591,15 +618,58 @@ vector_to_sparsevec(PG_FUNCTION_ARGS)
 	PG_RETURN_POINTER(result);
 }
 
+/*
+ * Convert half vector to sparse vector
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_to_sparsevec);
+Datum
+halfvec_to_sparsevec(PG_FUNCTION_ARGS)
+{
+	HalfVector *vec = PG_GETARG_HALFVEC_P(0);
+	int32		typmod = PG_GETARG_INT32(1);
+	SparseVector *result;
+	int			dim = vec->dim;
+	int			nnz = 0;
+	float	   *values;
+	int			j = 0;
+
+	CheckDim(dim);
+	CheckExpectedDim(typmod, dim);
+
+	for (int i = 0; i < dim; i++)
+	{
+		if (!HalfIsZero(vec->x[i]))
+			nnz++;
+	}
+
+	result = InitSparseVector(dim, nnz);
+	values = SPARSEVEC_VALUES(result);
+	for (int i = 0; i < dim; i++)
+	{
+		if (!HalfIsZero(vec->x[i]))
+		{
+			/* Safety check */
+			if (j >= result->nnz)
+				elog(ERROR, "safety check failed");
+
+			result->indices[j] = i + 1;
+			values[j] = HalfToFloat4(vec->x[i]);
+			j++;
+		}
+	}
+
+	PG_RETURN_POINTER(result);
+}
+
 /*
  * Get the L2 squared distance between sparse vectors
  */
-static double
+static float
 SparsevecL2SquaredDistance(SparseVector * a, SparseVector * b)
 {
 	float	   *ax = SPARSEVEC_VALUES(a);
 	float	   *bx = SPARSEVEC_VALUES(b);
-	double		distance = 0.0;
+	float		distance = 0.0;
 	int			bpos = 0;
 
 	for (int i = 0; i < a->nnz; i++)
@@ -673,12 +743,12 @@ sparsevec_l2_squared_distance(PG_FUNCTION_ARGS)
 /*
  * Get the inner product of two sparse vectors
  */
-static double
+static float
 SparsevecInnerProduct(SparseVector * a, SparseVector * b)
 {
 	float	   *ax = SPARSEVEC_VALUES(a);
 	float	   *bx = SPARSEVEC_VALUES(b);
-	double		distance = 0.0;
+	float		distance = 0.0;
 	int			bpos = 0;
 
 	for (int i = 0; i < a->nnz; i++)
@@ -781,12 +851,61 @@ sparsevec_cosine_distance(PG_FUNCTION_ARGS)
 	PG_RETURN_FLOAT8(1.0 - similarity);
 }
 
+/*
+ * Get the L1 distance between two sparse vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_l1_distance);
+Datum
+sparsevec_l1_distance(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
+	float	   *ax = SPARSEVEC_VALUES(a);
+	float	   *bx = SPARSEVEC_VALUES(b);
+	double		distance = 0.0;
+	int			bpos = 0;
+
+	CheckDims(a, b);
+
+	for (int i = 0; i < a->nnz; i++)
+	{
+		int			ai = a->indices[i];
+		int			bi = -1;
+
+		for (int j = bpos; j < b->nnz; j++)
+		{
+			bi = b->indices[j];
+
+			if (ai == bi)
+				distance += fabsf(ax[i] - bx[j]);
+			else if (ai > bi)
+				distance += fabsf(bx[j]);
+
+			/* Update start for next iteration */
+			if (ai >= bi)
+				bpos = j + 1;
+
+			/* Found or passed it */
+			if (bi >= ai)
+				break;
+		}
+
+		if (ai != bi)
+			distance += fabsf(ax[i]);
+	}
+
+	for (int j = bpos; j < b->nnz; j++)
+		distance += fabsf(bx[j]);
+
+	PG_RETURN_FLOAT8(distance);
+}
+
 /*
  * Get the L2 norm of a sparse vector
  */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_norm);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_l2_norm);
 Datum
-sparsevec_norm(PG_FUNCTION_ARGS)
+sparsevec_l2_norm(PG_FUNCTION_ARGS)
 {
 	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
 	float	   *ax = SPARSEVEC_VALUES(a);
@@ -798,3 +917,205 @@ sparsevec_norm(PG_FUNCTION_ARGS)
 
 	PG_RETURN_FLOAT8(sqrt(norm));
 }
+
+/*
+ * Normalize a sparse vector with the L2 norm
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_l2_normalize);
+Datum
+sparsevec_l2_normalize(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	float	   *ax = SPARSEVEC_VALUES(a);
+	double		norm = 0;
+	SparseVector *result;
+	float	   *rx;
+
+	result = InitSparseVector(a->dim, a->nnz);
+	rx = SPARSEVEC_VALUES(result);
+
+	/* Auto-vectorized */
+	for (int i = 0; i < a->nnz; i++)
+		norm += (double) ax[i] * (double) ax[i];
+
+	norm = sqrt(norm);
+
+	/* Return zero vector for zero norm */
+	if (norm > 0)
+	{
+		int			zeros = 0;
+
+		for (int i = 0; i < a->nnz; i++)
+		{
+			result->indices[i] = a->indices[i];
+			rx[i] = ax[i] / norm;
+
+			if (isinf(rx[i]))
+				float_overflow_error();
+
+			if (rx[i] == 0)
+				zeros++;
+		}
+
+		/* Allocate a new vector in the unlikely event there are zeros */
+		if (zeros > 0)
+		{
+			SparseVector *newResult = InitSparseVector(result->dim, result->nnz - zeros);
+			float	   *nx = SPARSEVEC_VALUES(newResult);
+			int			j = 0;
+
+			for (int i = 0; i < result->nnz; i++)
+			{
+				if (rx[i] == 0)
+					continue;
+
+				/* Safety check */
+				if (j >= newResult->nnz)
+					elog(ERROR, "safety check failed");
+
+				newResult->indices[j] = result->indices[i];
+				nx[j] = rx[i];
+				j++;
+			}
+
+			pfree(result);
+
+			PG_RETURN_POINTER(newResult);
+		}
+	}
+
+	PG_RETURN_POINTER(result);
+}
+
+/*
+ * Internal helper to compare sparse vectors
+ */
+static int
+sparsevec_cmp_internal(SparseVector * a, SparseVector * b)
+{
+	float	   *ax = SPARSEVEC_VALUES(a);
+	float	   *bx = SPARSEVEC_VALUES(b);
+	int			nnz = Min(a->nnz, b->nnz);
+
+	/* Check values before dimensions to be consistent with Postgres arrays */
+	for (int i = 0; i < nnz; i++)
+	{
+		if (a->indices[i] < b->indices[i])
+			return ax[i] < 0 ? -1 : 1;
+
+		if (a->indices[i] > b->indices[i])
+			return bx[i] < 0 ? 1 : -1;
+
+		if (ax[i] < bx[i])
+			return -1;
+
+		if (ax[i] > bx[i])
+			return 1;
+	}
+
+	/* Check <= dim since indices start at 1 */
+	if (a->nnz < b->nnz && b->indices[nnz] <= a->dim)
+		return bx[nnz] < 0 ? 1 : -1;
+
+	if (a->nnz > b->nnz && a->indices[nnz] <= b->dim)
+		return ax[nnz] < 0 ? -1 : 1;
+
+	if (a->dim < b->dim)
+		return -1;
+
+	if (a->dim > b->dim)
+		return 1;
+
+	return 0;
+}
+
+/*
+ * Less than
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_lt);
+Datum
+sparsevec_lt(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
+
+	PG_RETURN_BOOL(sparsevec_cmp_internal(a, b) < 0);
+}
+
+/*
+ * Less than or equal
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_le);
+Datum
+sparsevec_le(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
+
+	PG_RETURN_BOOL(sparsevec_cmp_internal(a, b) <= 0);
+}
+
+/*
+ * Equal
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_eq);
+Datum
+sparsevec_eq(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
+
+	PG_RETURN_BOOL(sparsevec_cmp_internal(a, b) == 0);
+}
+
+/*
+ * Not equal
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_ne);
+Datum
+sparsevec_ne(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
+
+	PG_RETURN_BOOL(sparsevec_cmp_internal(a, b) != 0);
+}
+
+/*
+ * Greater than or equal
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_ge);
+Datum
+sparsevec_ge(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
+
+	PG_RETURN_BOOL(sparsevec_cmp_internal(a, b) >= 0);
+}
+
+/*
+ * Greater than
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_gt);
+Datum
+sparsevec_gt(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
+
+	PG_RETURN_BOOL(sparsevec_cmp_internal(a, b) > 0);
+}
+
+/*
+ * Compare sparse vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_cmp);
+Datum
+sparsevec_cmp(PG_FUNCTION_ARGS)
+{
+	SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
+	SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
+
+	PG_RETURN_INT32(sparsevec_cmp_internal(a, b));
+}

src/sparsevec.h

@@ -1,6 +1,8 @@
 #ifndef SPARSEVEC_H
 #define SPARSEVEC_H
 
+#include "fmgr.h"
+
 #define SPARSEVEC_MAX_DIM 100000
 #define SPARSEVEC_MAX_NNZ 16000
 

src/vector.c

@@ -2,7 +2,8 @@
 
 #include <math.h>
 
-#include "bitvector.h"
+#include "bitutils.h"
+#include "bitvec.h"
 #include "catalog/pg_type.h"
 #include "common/shortest_dec.h"
 #include "fmgr.h"
@@ -33,11 +34,10 @@
 #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")))
+#if defined(USE_TARGET_CLONES) && !defined(__FMA__)
+#define VECTOR_TARGET_CLONES __attribute__((target_clones("default", "fma")))
 #else
-#define VECTOR_DISPATCH
+#define VECTOR_TARGET_CLONES
 #endif
 
 PG_MODULE_MAGIC;
@@ -49,6 +49,7 @@ PGDLLEXPORT void _PG_init(void);
 void
 _PG_init(void)
 {
+	BitvecInit();
 	HalfvecInit();
 	HnswInit();
 	IvfflatInit();
@@ -564,7 +565,7 @@ halfvec_to_vector(PG_FUNCTION_ARGS)
 	PG_RETURN_POINTER(result);
 }
 
-VECTOR_DISPATCH static float
+VECTOR_TARGET_CLONES static float
 VectorL2SquaredDistance(int dim, float *ax, float *bx)
 {
 	float		distance = 0.0;
@@ -611,7 +612,7 @@ vector_l2_squared_distance(PG_FUNCTION_ARGS)
 	PG_RETURN_FLOAT8((double) VectorL2SquaredDistance(a->dim, a->x, b->x));
 }
 
-VECTOR_DISPATCH static float
+VECTOR_TARGET_CLONES static float
 VectorInnerProduct(int dim, float *ax, float *bx)
 {
 	float		distance = 0.0;
@@ -653,6 +654,25 @@ vector_negative_inner_product(PG_FUNCTION_ARGS)
 	PG_RETURN_FLOAT8((double) -VectorInnerProduct(a->dim, a->x, b->x));
 }
 
+VECTOR_TARGET_CLONES static double
+VectorCosineSimilarity(int dim, float *ax, float *bx)
+{
+	float		similarity = 0.0;
+	float		norma = 0.0;
+	float		normb = 0.0;
+
+	/* Auto-vectorized */
+	for (int i = 0; i < dim; i++)
+	{
+		similarity += ax[i] * bx[i];
+		norma += ax[i] * ax[i];
+		normb += bx[i] * bx[i];
+	}
+
+	/* Use sqrt(a * b) over sqrt(a) * sqrt(b) */
+	return (double) similarity / sqrt((double) norma * (double) normb);
+}
+
 /*
  * Get the cosine distance between two vectors
  */
@@ -662,25 +682,11 @@ cosine_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		norma = 0.0;
-	float		normb = 0.0;
 	double		similarity;
 
 	CheckDims(a, b);
 
-	/* Auto-vectorized */
-	for (int i = 0; i < a->dim; i++)
-	{
-		distance += ax[i] * bx[i];
-		norma += ax[i] * ax[i];
-		normb += bx[i] * bx[i];
-	}
-
-	/* Use sqrt(a * b) over sqrt(a) * sqrt(b) */
-	similarity = (double) distance / sqrt((double) norma * (double) normb);
+	similarity = VectorCosineSimilarity(a->dim, a->x, b->x);
 
 #ifdef _MSC_VER
 	/* /fp:fast may not propagate NaN */
@@ -723,6 +729,19 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
 	PG_RETURN_FLOAT8(acos(distance) / M_PI);
 }
 
+/* Does not require FMA, but keep logic simple */
+VECTOR_TARGET_CLONES static float
+VectorL1Distance(int dim, float *ax, float *bx)
+{
+	float		distance = 0.0;
+
+	/* Auto-vectorized */
+	for (int i = 0; i < dim; i++)
+		distance += fabsf(ax[i] - bx[i]);
+
+	return distance;
+}
+
 /*
  * Get the L1 distance between two vectors
  */
@@ -732,17 +751,10 @@ l1_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;
 
 	CheckDims(a, b);
 
-	/* Auto-vectorized */
-	for (int i = 0; i < a->dim; i++)
-		distance += fabsf(ax[i] - bx[i]);
-
-	PG_RETURN_FLOAT8((double) distance);
+	PG_RETURN_FLOAT8((double) VectorL1Distance(a->dim, a->x, b->x));
 }
 
 /*
@@ -775,6 +787,45 @@ vector_norm(PG_FUNCTION_ARGS)
 	PG_RETURN_FLOAT8(sqrt(norm));
 }
 
+/*
+ * Normalize a vector with the L2 norm
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(l2_normalize);
+Datum
+l2_normalize(PG_FUNCTION_ARGS)
+{
+	Vector	   *a = PG_GETARG_VECTOR_P(0);
+	float	   *ax = a->x;
+	double		norm = 0;
+	Vector	   *result;
+	float	   *rx;
+
+	result = InitVector(a->dim);
+	rx = result->x;
+
+	/* Auto-vectorized */
+	for (int i = 0; i < a->dim; i++)
+		norm += (double) ax[i] * (double) ax[i];
+
+	norm = sqrt(norm);
+
+	/* Return zero vector for zero norm */
+	if (norm > 0)
+	{
+		for (int i = 0; i < a->dim; i++)
+			rx[i] = ax[i] / norm;
+
+		/* Check for overflow */
+		for (int i = 0; i < a->dim; i++)
+		{
+			if (isinf(rx[i]))
+				float_overflow_error();
+		}
+	}
+
+	PG_RETURN_POINTER(result);
+}
+
 /*
  * Add vectors
  */
@@ -877,12 +928,36 @@ vector_mul(PG_FUNCTION_ARGS)
 	PG_RETURN_POINTER(result);
 }
 
+/*
+ * Concatenate vectors
+ */
+PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_concat);
+Datum
+vector_concat(PG_FUNCTION_ARGS)
+{
+	Vector	   *a = PG_GETARG_VECTOR_P(0);
+	Vector	   *b = PG_GETARG_VECTOR_P(1);
+	Vector	   *result;
+	int			dim = a->dim + b->dim;
+
+	CheckDim(dim);
+	result = InitVector(dim);
+
+	for (int i = 0; i < a->dim; i++)
+		result->x[i] = a->x[i];
+
+	for (int i = 0; i < b->dim; i++)
+		result->x[i + a->dim] = b->x[i];
+
+	PG_RETURN_POINTER(result);
+}
+
 /*
  * Quantize a vector
  */
-PGDLLEXPORT PG_FUNCTION_INFO_V1(quantize_binary);
+PGDLLEXPORT PG_FUNCTION_INFO_V1(binary_quantize);
 Datum
-quantize_binary(PG_FUNCTION_ARGS)
+binary_quantize(PG_FUNCTION_ARGS)
 {
 	Vector	   *a = PG_GETARG_VECTOR_P(0);
 	float	   *ax = a->x;
@@ -1107,7 +1182,7 @@ vector_accum(PG_FUNCTION_ARGS)
 }
 
 /*
- * Combine vectors
+ * Combine vectors or half vectors
  */
 PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_combine);
 Datum

src/vector.h

@@ -1,6 +1,8 @@
 #ifndef VECTOR_H
 #define VECTOR_H
 
+#include "fmgr.h"
+
 #define VECTOR_MAX_DIM 16000
 
 #define VECTOR_SIZE(_dim)		(offsetof(Vector, x) + sizeof(float)*(_dim))
@@ -19,5 +21,6 @@ typedef struct Vector
 Vector	   *InitVector(int dim);
 void		PrintVector(char *msg, Vector * vector);
 int			vector_cmp_internal(Vector * a, Vector * b);
+PGDLLEXPORT Datum l2_normalize(PG_FUNCTION_ARGS);
 
 #endif

test/expected/bit_functions.out

@@ -28,6 +28,24 @@ SELECT hamming_distance('10101010101010101010', '01010101010101010101');
                20
 (1 row)
 
+SELECT hamming_distance('101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101', '101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101');
+ hamming_distance 
+------------------
+                0
+(1 row)
+
+SELECT hamming_distance('101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101', '010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010');
+ hamming_distance 
+------------------
+              513
+(1 row)
+
+SELECT hamming_distance('110000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000011', '100000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001');
+ hamming_distance 
+------------------
+                2
+(1 row)
+
 SELECT hamming_distance('', '');
  hamming_distance 
 ------------------
@@ -86,6 +104,24 @@ SELECT jaccard_distance('10101010101010101010', '01010101010101010101');
                 1
 (1 row)
 
+SELECT jaccard_distance('101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101', '101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101');
+ jaccard_distance 
+------------------
+                0
+(1 row)
+
+SELECT jaccard_distance('101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101', '010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010');
+ jaccard_distance 
+------------------
+                1
+(1 row)
+
+SELECT jaccard_distance('110000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000011', '100000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001');
+ jaccard_distance 
+------------------
+              0.5
+(1 row)
+
 SELECT jaccard_distance('', '');
  jaccard_distance 
 ------------------

test/expected/btree_halfvec.out

@@ -0,0 +1,20 @@
+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 (val);
+SELECT * FROM t WHERE val = '[1,2,3]';
+   val   
+---------
+ [1,2,3]
+(1 row)
+
+SELECT * FROM t ORDER BY val;
+   val   
+---------
+ [0,0,0]
+ [1,1,1]
+ [1,2,3]
+ 
+(4 rows)
+
+DROP TABLE t;

test/expected/btree_sparsevec.out

@@ -0,0 +1,20 @@
+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);
+CREATE INDEX ON t (val);
+SELECT * FROM t WHERE val = '{1:1,2:2,3:3}/3';
+       val       
+-----------------
+ {1:1,2:2,3:3}/3
+(1 row)
+
+SELECT * FROM t ORDER BY val;
+       val       
+-----------------
+ {}/3
+ {1:1,2:1,3:1}/3
+ {1:1,2:2,3:3}/3
+ 
+(4 rows)
+
+DROP TABLE t;

test/expected/btree_vector.out

@@ -8,10 +8,13 @@ SELECT * FROM t WHERE val = '[1,2,3]';
  [1,2,3]
 (1 row)
 
-SELECT * FROM t ORDER BY val LIMIT 1;
+SELECT * FROM t ORDER BY val;
    val   
 ---------
  [0,0,0]
-(1 row)
+ [1,1,1]
+ [1,2,3]
+ 
+(4 rows)
 
 DROP TABLE t;

test/expected/cast.out

@@ -170,6 +170,44 @@ 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 '[0,1.5,0,3.5,0]'::halfvec::sparsevec;
+    sparsevec    
+-----------------
+ {2:1.5,4:3.5}/5
+(1 row)
+
+SELECT '[0,1.5,0,3.5,0]'::halfvec::sparsevec(5);
+    sparsevec    
+-----------------
+ {2:1.5,4:3.5}/5
+(1 row)
+
+SELECT '[0,1.5,0,3.5,0]'::halfvec::sparsevec(4);
+ERROR:  expected 4 dimensions, not 5
+SELECT '{2:1.5,4:3.5}/5'::sparsevec::halfvec;
+     halfvec     
+-----------------
+ [0,1.5,0,3.5,0]
+(1 row)
+
+SELECT '{2:1.5,4:3.5}/5'::sparsevec::halfvec(5);
+     halfvec     
+-----------------
+ [0,1.5,0,3.5,0]
+(1 row)
+
+SELECT '{2:1.5,4:3.5}/5'::sparsevec::halfvec(4);
+ERROR:  expected 4 dimensions, not 5
+SELECT '{}/16001'::sparsevec::halfvec;
+ERROR:  halfvec cannot have more than 16000 dimensions
+SELECT '{1:65520}/1'::sparsevec::halfvec;
+ERROR:  "65520" is out of range for type halfvec
+SELECT '{1:1e-8}/1'::sparsevec::halfvec;
+ halfvec 
+---------
+ [0]
+(1 row)
+
 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/halfvec_functions.out

@@ -1,19 +1,179 @@
-SELECT round(halfvec_norm('[1,1]')::numeric, 5);
+SELECT '[1,2,3]'::halfvec + '[4,5,6]';
+ ?column? 
+----------
+ [5,7,9]
+(1 row)
+
+SELECT '[65519]'::halfvec + '[65519]';
+ERROR:  value out of range: overflow
+SELECT '[1,2]'::halfvec + '[3]';
+ERROR:  different halfvec dimensions 2 and 1
+SELECT '[1,2,3]'::halfvec - '[4,5,6]';
+  ?column?  
+------------
+ [-3,-3,-3]
+(1 row)
+
+SELECT '[-65519]'::halfvec - '[65519]';
+ERROR:  value out of range: overflow
+SELECT '[1,2]'::halfvec - '[3]';
+ERROR:  different halfvec dimensions 2 and 1
+SELECT '[1,2,3]'::halfvec * '[4,5,6]';
+ ?column?  
+-----------
+ [4,10,18]
+(1 row)
+
+SELECT '[65519]'::halfvec * '[65519]';
+ERROR:  value out of range: overflow
+SELECT '[1e-7]'::halfvec * '[1e-7]';
+ERROR:  value out of range: underflow
+SELECT '[1,2]'::halfvec * '[3]';
+ERROR:  different halfvec dimensions 2 and 1
+SELECT '[1,2,3]'::halfvec || '[4,5]';
+  ?column?   
+-------------
+ [1,2,3,4,5]
+(1 row)
+
+SELECT array_fill(0, ARRAY[16000])::halfvec || '[1]';
+ERROR:  halfvec cannot have more than 16000 dimensions
+SELECT '[1,2,3]'::halfvec < '[1,2,3]';
+ ?column? 
+----------
+ f
+(1 row)
+
+SELECT '[1,2,3]'::halfvec < '[1,2]';
+ ?column? 
+----------
+ f
+(1 row)
+
+SELECT '[1,2,3]'::halfvec <= '[1,2,3]';
+ ?column? 
+----------
+ t
+(1 row)
+
+SELECT '[1,2,3]'::halfvec <= '[1,2]';
+ ?column? 
+----------
+ f
+(1 row)
+
+SELECT '[1,2,3]'::halfvec = '[1,2,3]';
+ ?column? 
+----------
+ t
+(1 row)
+
+SELECT '[1,2,3]'::halfvec = '[1,2]';
+ ?column? 
+----------
+ f
+(1 row)
+
+SELECT '[1,2,3]'::halfvec != '[1,2,3]';
+ ?column? 
+----------
+ f
+(1 row)
+
+SELECT '[1,2,3]'::halfvec != '[1,2]';
+ ?column? 
+----------
+ t
+(1 row)
+
+SELECT '[1,2,3]'::halfvec >= '[1,2,3]';
+ ?column? 
+----------
+ t
+(1 row)
+
+SELECT '[1,2,3]'::halfvec >= '[1,2]';
+ ?column? 
+----------
+ t
+(1 row)
+
+SELECT '[1,2,3]'::halfvec > '[1,2,3]';
+ ?column? 
+----------
+ f
+(1 row)
+
+SELECT '[1,2,3]'::halfvec > '[1,2]';
+ ?column? 
+----------
+ t
+(1 row)
+
+SELECT halfvec_cmp('[1,2,3]', '[1,2,3]');
+ halfvec_cmp 
+-------------
+           0
+(1 row)
+
+SELECT halfvec_cmp('[1,2,3]', '[0,0,0]');
+ halfvec_cmp 
+-------------
+           1
+(1 row)
+
+SELECT halfvec_cmp('[0,0,0]', '[1,2,3]');
+ halfvec_cmp 
+-------------
+          -1
+(1 row)
+
+SELECT halfvec_cmp('[1,2]', '[1,2,3]');
+ halfvec_cmp 
+-------------
+          -1
+(1 row)
+
+SELECT halfvec_cmp('[1,2,3]', '[1,2]');
+ halfvec_cmp 
+-------------
+           1
+(1 row)
+
+SELECT halfvec_cmp('[1,2]', '[2,3,4]');
+ halfvec_cmp 
+-------------
+          -1
+(1 row)
+
+SELECT halfvec_cmp('[2,3]', '[1,2,3]');
+ halfvec_cmp 
+-------------
+           1
+(1 row)
+
+SELECT vector_dims('[1,2,3]'::halfvec);
+ vector_dims 
+-------------
+           3
+(1 row)
+
+SELECT round(l2_norm('[1,1]'::halfvec)::numeric, 5);
   round  
 ---------
  1.41421
 (1 row)
 
-SELECT halfvec_norm('[3,4]');
- halfvec_norm 
---------------
-            5
+SELECT l2_norm('[3,4]'::halfvec);
+ l2_norm 
+---------
+       5
 (1 row)
 
-SELECT halfvec_norm('[0,1]');
- halfvec_norm 
---------------
-            1
+SELECT l2_norm('[0,1]'::halfvec);
+ l2_norm 
+---------
+       1
 (1 row)
 
 SELECT l2_distance('[0,0]'::halfvec, '[3,4]');
@@ -112,6 +272,18 @@ SELECT cosine_distance('[1,1]'::halfvec, '[-1.1,-1.1]');
                2
 (1 row)
 
+SELECT cosine_distance('[1,2,3,4,5,6,7,8,9]'::halfvec, '[1,2,3,4,5,6,7,8,9]');
+ cosine_distance 
+-----------------
+               0
+(1 row)
+
+SELECT cosine_distance('[1,2,3,4,5,6,7,8,9]'::halfvec, '[-1,-2,-3,-4,-5,-6,-7,-8,-9]');
+ cosine_distance 
+-----------------
+               2
+(1 row)
+
 SELECT '[1,2]'::halfvec <=> '[2,4]';
  ?column? 
 ----------
@@ -132,14 +304,62 @@ SELECT l1_distance('[0,0]'::halfvec, '[0,1]');
 
 SELECT l1_distance('[1,2]'::halfvec, '[3]');
 ERROR:  different halfvec dimensions 2 and 1
-SELECT quantize_binary('[1,0,-1]'::halfvec);
- quantize_binary 
+SELECT l1_distance('[1,2,3,4,5,6,7,8,9]'::halfvec, '[1,2,3,4,5,6,7,8,9]');
+ l1_distance 
+-------------
+           0
+(1 row)
+
+SELECT l1_distance('[1,2,3,4,5,6,7,8,9]'::halfvec, '[0,3,2,5,4,7,6,9,8]');
+ l1_distance 
+-------------
+           9
+(1 row)
+
+SELECT '[0,0]'::halfvec <+> '[3,4]';
+ ?column? 
+----------
+        7
+(1 row)
+
+SELECT l2_normalize('[3,4]'::halfvec);
+      l2_normalize      
+------------------------
+ [0.60009766,0.7998047]
+(1 row)
+
+SELECT l2_normalize('[3,0]'::halfvec);
+ l2_normalize 
+--------------
+ [1,0]
+(1 row)
+
+SELECT l2_normalize('[0,0.1]'::halfvec);
+ l2_normalize 
+--------------
+ [0,1]
+(1 row)
+
+SELECT l2_normalize('[0,0]'::halfvec);
+ l2_normalize 
+--------------
+ [0,0]
+(1 row)
+
+SELECT l2_normalize('[65504]'::halfvec);
+ l2_normalize 
+--------------
+ [1]
+(1 row)
+
+SELECT binary_quantize('[1,0,-1]'::halfvec);
+ binary_quantize 
 -----------------
  100
 (1 row)
 
-SELECT quantize_binary('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]'::halfvec);
- quantize_binary 
+SELECT binary_quantize('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]'::halfvec);
+ binary_quantize 
 -----------------
  01001110101
 (1 row)
@@ -174,3 +394,53 @@ SELECT subvector('[1,2,3,4,5]'::halfvec, 3, -1);
 ERROR:  halfvec must have at least 1 dimension
 SELECT subvector('[1,2,3,4,5]'::halfvec, -1, 2);
 ERROR:  halfvec must have at least 1 dimension
+SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::halfvec, '[3,5,7]']) v;
+    avg    
+-----------
+ [2,3.5,5]
+(1 row)
+
+SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::halfvec, '[3,5,7]', NULL]) v;
+    avg    
+-----------
+ [2,3.5,5]
+(1 row)
+
+SELECT avg(v) FROM unnest(ARRAY[]::halfvec[]) v;
+ avg 
+-----
+ 
+(1 row)
+
+SELECT avg(v) FROM unnest(ARRAY['[1,2]'::halfvec, '[3]']) v;
+ERROR:  expected 2 dimensions, not 1
+SELECT avg(v) FROM unnest(ARRAY['[65504]'::halfvec, '[65504]']) v;
+   avg   
+---------
+ [65504]
+(1 row)
+
+SELECT halfvec_avg(array_agg(n)) FROM generate_series(1, 16002) n;
+ERROR:  halfvec cannot have more than 16000 dimensions
+SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::halfvec, '[3,5,7]']) v;
+   sum    
+----------
+ [4,7,10]
+(1 row)
+
+SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::halfvec, '[3,5,7]', NULL]) v;
+   sum    
+----------
+ [4,7,10]
+(1 row)
+
+SELECT sum(v) FROM unnest(ARRAY[]::halfvec[]) v;
+ sum 
+-----
+ 
+(1 row)
+
+SELECT sum(v) FROM unnest(ARRAY['[1,2]'::halfvec, '[3]']) v;
+ERROR:  different halfvec dimensions 2 and 1
+SELECT sum(v) FROM unnest(ARRAY['[65504]'::halfvec, '[65504]']) v;
+ERROR:  value out of range: overflow

test/expected/hnsw_halfvec_l1.out

@@ -0,0 +1,21 @@
+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 hnsw (val halfvec_l1_ops);
+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)
+
+DROP TABLE t;

test/expected/hnsw_sparsevec_l1.out

@@ -0,0 +1,21 @@
+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);
+CREATE INDEX ON t USING hnsw (val sparsevec_l1_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';
+       val       
+-----------------
+ {1:1,2:2,3:3}/3
+ {1:1,2:2,3:4}/3
+ {1:1,2:1,3:1}/3
+ {}/3
+(4 rows)
+
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <+> (SELECT NULL::sparsevec)) t2;
+ count 
+-------
+     4
+(1 row)
+
+DROP TABLE t;

test/expected/hnsw_vector_l1.out

@@ -0,0 +1,21 @@
+SET enable_seqscan = off;
+CREATE TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING hnsw (val vector_l1_ops);
+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::vector)) t2;
+ count 
+-------
+     4
+(1 row)
+
+DROP TABLE t;

test/expected/ivfflat_bit_hamming.out

@@ -0,0 +1,36 @@
+SET enable_seqscan = off;
+CREATE TABLE t (val bit(3));
+INSERT INTO t (val) VALUES (B'000'), (B'100'), (B'111'), (NULL);
+CREATE INDEX ON t USING ivfflat (val bit_hamming_ops) WITH (lists = 1);
+INSERT INTO t (val) VALUES (B'110');
+SELECT * FROM t ORDER BY val <~> B'111';
+ val 
+-----
+ 111
+ 110
+ 100
+ 000
+(4 rows)
+
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <~> (SELECT NULL::bit)) t2;
+ count 
+-------
+     4
+(1 row)
+
+DROP TABLE t;
+-- TODO move
+CREATE TABLE t (val varbit(3));
+CREATE INDEX ON t USING ivfflat (val bit_hamming_ops) WITH (lists = 1);
+ERROR:  type not supported for ivfflat index
+CREATE INDEX ON t USING ivfflat ((val::bit(3)) bit_hamming_ops) WITH (lists = 1);
+NOTICE:  ivfflat index created with little data
+DETAIL:  This will cause low recall.
+HINT:  Drop the index until the table has more data.
+CREATE INDEX ON t USING ivfflat ((val::bit(64001)) bit_hamming_ops) WITH (lists = 1);
+ERROR:  column cannot have more than 64000 dimensions for ivfflat index
+CREATE INDEX ON t USING ivfflat ((val::bit(2)) bit_hamming_ops) WITH (lists = 5);
+NOTICE:  ivfflat index created with little data
+DETAIL:  This will cause low recall.
+HINT:  Drop the index until the table has more data.
+DROP TABLE t;

test/expected/sparsevec_functions.out

@@ -1,25 +1,139 @@
-SELECT round(sparsevec_norm('{1:1,2:1}/2')::numeric, 5);
+SELECT '{1:1,2:2,3:3}/3'::sparsevec < '{1:1,2:2,3:3}/3';
+ ?column? 
+----------
+ f
+(1 row)
+
+SELECT '{1:1,2:2,3:3}/3'::sparsevec < '{1:1,2:2}/2';
+ ?column? 
+----------
+ f
+(1 row)
+
+SELECT '{1:1,2:2,3:3}/3'::sparsevec <= '{1:1,2:2,3:3}/3';
+ ?column? 
+----------
+ t
+(1 row)
+
+SELECT '{1:1,2:2,3:3}/3'::sparsevec <= '{1:1,2:2}/2';
+ ?column? 
+----------
+ f
+(1 row)
+
+SELECT '{1:1,2:2,3:3}/3'::sparsevec = '{1:1,2:2,3:3}/3';
+ ?column? 
+----------
+ t
+(1 row)
+
+SELECT '{1:1,2:2,3:3}/3'::sparsevec = '{1:1,2:2}/2';
+ ?column? 
+----------
+ f
+(1 row)
+
+SELECT '{1:1,2:2,3:3}/3'::sparsevec != '{1:1,2:2,3:3}/3';
+ ?column? 
+----------
+ f
+(1 row)
+
+SELECT '{1:1,2:2,3:3}/3'::sparsevec != '{1:1,2:2}/2';
+ ?column? 
+----------
+ t
+(1 row)
+
+SELECT '{1:1,2:2,3:3}/3'::sparsevec >= '{1:1,2:2,3:3}/3';
+ ?column? 
+----------
+ t
+(1 row)
+
+SELECT '{1:1,2:2,3:3}/3'::sparsevec >= '{1:1,2:2}/2';
+ ?column? 
+----------
+ t
+(1 row)
+
+SELECT '{1:1,2:2,3:3}/3'::sparsevec > '{1:1,2:2,3:3}/3';
+ ?column? 
+----------
+ f
+(1 row)
+
+SELECT '{1:1,2:2,3:3}/3'::sparsevec > '{1:1,2:2}/2';
+ ?column? 
+----------
+ t
+(1 row)
+
+SELECT sparsevec_cmp('{1:1,2:2,3:3}/3', '{1:1,2:2,3:3}/3');
+ sparsevec_cmp 
+---------------
+             0
+(1 row)
+
+SELECT sparsevec_cmp('{1:1,2:2,3:3}/3', '{}/3');
+ sparsevec_cmp 
+---------------
+             1
+(1 row)
+
+SELECT sparsevec_cmp('{}/3', '{1:1,2:2,3:3}/3');
+ sparsevec_cmp 
+---------------
+            -1
+(1 row)
+
+SELECT sparsevec_cmp('{1:1,2:2}/2', '{1:1,2:2,3:3}/3');
+ sparsevec_cmp 
+---------------
+            -1
+(1 row)
+
+SELECT sparsevec_cmp('{1:1,2:2,3:3}/3', '{1:1,2:2}/2');
+ sparsevec_cmp 
+---------------
+             1
+(1 row)
+
+SELECT sparsevec_cmp('{1:1,2:2}/2', '{1:2,2:3,3:4}/3');
+ sparsevec_cmp 
+---------------
+            -1
+(1 row)
+
+SELECT sparsevec_cmp('{1:2,2:3}/2', '{1:1,2:2,3:3}/3');
+ sparsevec_cmp 
+---------------
+             1
+(1 row)
+
+SELECT round(l2_norm('{1:1,2:1}/2'::sparsevec)::numeric, 5);
   round  
 ---------
  1.41421
 (1 row)
 
-SELECT sparsevec_norm('{1:3,2:4}/2');
- sparsevec_norm 
-----------------
-              5
+SELECT l2_norm('{1:3,2:4}/2'::sparsevec);
+ l2_norm 
+---------
+       5
 (1 row)
 
-SELECT sparsevec_norm('{2:1}/2');
- sparsevec_norm 
-----------------
-              1
+SELECT l2_norm('{2:1}/2'::sparsevec);
+ l2_norm 
+---------
+       1
 (1 row)
 
-SELECT sparsevec_norm('{1:3e37,2:4e37}/2')::real;
- sparsevec_norm 
-----------------
-          5e+37
+SELECT l2_norm('{1:3e37,2:4e37}/2'::sparsevec)::real;
+ l2_norm 
+---------
+   5e+37
 (1 row)
 
 SELECT l2_distance('{}/2'::sparsevec, '{1:3,2:4}/2');
@@ -28,6 +142,18 @@ SELECT l2_distance('{}/2'::sparsevec, '{1:3,2:4}/2');
            5
 (1 row)
 
+SELECT l2_distance('{1:3}/2'::sparsevec, '{2:4}/2');
+ l2_distance 
+-------------
+           5
+(1 row)
+
+SELECT l2_distance('{2:4}/2'::sparsevec, '{1:3}/2');
+ l2_distance 
+-------------
+           5
+(1 row)
+
 SELECT l2_distance('{}/2'::sparsevec, '{2:1}/2');
  l2_distance 
 -------------
@@ -46,10 +172,24 @@ SELECT inner_product('{1:1,2:2}/2'::sparsevec, '{1:2,2:4}/2');
             10
 (1 row)
 
-SELECT sparsevec_negative_inner_product('{1:1,2:2}/2', '{1:2,2:4}/2');
- sparsevec_negative_inner_product 
-----------------------------------
-                              -10
+SELECT inner_product('{1:1,2:2}/2'::sparsevec, '{1:3}/1');
+ERROR:  different sparsevec dimensions 2 and 1
+SELECT inner_product('{1:3e38}/1'::sparsevec, '{1:3e38}/1');
+ inner_product 
+---------------
+      Infinity
+(1 row)
+
+SELECT inner_product('{1:1,3:3,5:5}/5'::sparsevec, '{2:4,3:6,4:8}/5');
+ inner_product 
+---------------
+            18
+(1 row)
+
+SELECT '{1:1,2:2}/2'::sparsevec <#> '{1:3,2:4}/2';
+ ?column? 
+----------
+      -11
 (1 row)
 
 SELECT cosine_distance('{1:1,2:2}/2'::sparsevec, '{1:2,2:4}/2');
@@ -64,6 +204,18 @@ SELECT cosine_distance('{1:1,2:2}/2'::sparsevec, '{}/2');
              NaN
 (1 row)
 
+SELECT cosine_distance('{1:1,2:1}/2'::sparsevec, '{1:1,2:1}/2');
+ cosine_distance 
+-----------------
+               0
+(1 row)
+
+SELECT cosine_distance('{1:1}/2'::sparsevec, '{2:2}/2');
+ cosine_distance 
+-----------------
+               1
+(1 row)
+
 SELECT cosine_distance('{1:1,2:1}/2'::sparsevec, '{1:-1,2:-1}/2');
  cosine_distance 
 -----------------
@@ -76,11 +228,109 @@ SELECT cosine_distance('{1:2}/2'::sparsevec, '{2:2}/2');
                1
 (1 row)
 
+SELECT cosine_distance('{1:1,2:2}/2'::sparsevec, '{1:3}/1');
+ERROR:  different sparsevec dimensions 2 and 1
+SELECT cosine_distance('{1:1,2:1}/2'::sparsevec, '{1:1.1,2:1.1}/2');
+ cosine_distance 
+-----------------
+               0
+(1 row)
+
+SELECT cosine_distance('{1:1,2:1}/2'::sparsevec, '{1:-1.1,2:-1.1}/2');
+ cosine_distance 
+-----------------
+               2
+(1 row)
+
+SELECT cosine_distance('{1:3e38}/1'::sparsevec, '{1:3e38}/1');
+ cosine_distance 
+-----------------
+             NaN
+(1 row)
+
 SELECT cosine_distance('{}/1'::sparsevec, '{}/1');
  cosine_distance 
 -----------------
              NaN
 (1 row)
 
-SELECT cosine_distance('{1:2}/2'::sparsevec, '{1:1}/3');
-ERROR:  different sparsevec dimensions 2 and 3
+SELECT '{1:1,2:2}/2'::sparsevec <=> '{1:2,2:4}/2';
+ ?column? 
+----------
+        0
+(1 row)
+
+SELECT l1_distance('{}/2'::sparsevec, '{1:3,2:4}/2');
+ l1_distance 
+-------------
+           7
+(1 row)
+
+SELECT l1_distance('{}/2'::sparsevec, '{2:1}/2');
+ l1_distance 
+-------------
+           1
+(1 row)
+
+SELECT l1_distance('{1:1,2:2}/2'::sparsevec, '{1:3}/1');
+ERROR:  different sparsevec dimensions 2 and 1
+SELECT l1_distance('{1:3e38}/1'::sparsevec, '{1:-3e38}/1');
+ l1_distance 
+-------------
+    Infinity
+(1 row)
+
+SELECT l1_distance('{1:1,3:3,5:5,7:7}/8'::sparsevec, '{2:2,4:4,6:6,8:8}/8');
+ l1_distance 
+-------------
+          36
+(1 row)
+
+SELECT l1_distance('{1:1,3:3,5:5,7:7,9:9}/9'::sparsevec, '{2:2,4:4,6:6,8:8}/9');
+ l1_distance 
+-------------
+          45
+(1 row)
+
+SELECT l2_normalize('{1:3,2:4}/2'::sparsevec);
+  l2_normalize   
+-----------------
+ {1:0.6,2:0.8}/2
+(1 row)
+
+SELECT l2_normalize('{1:3}/2'::sparsevec);
+ l2_normalize 
+--------------
+ {1:1}/2
+(1 row)
+
+SELECT l2_normalize('{2:0.1}/2'::sparsevec);
+ l2_normalize 
+--------------
+ {2:1}/2
+(1 row)
+
+SELECT l2_normalize('{}/2'::sparsevec);
+ l2_normalize 
+--------------
+ {}/2
+(1 row)
+
+SELECT l2_normalize('{1:3e38}/1'::sparsevec);
+ l2_normalize 
+--------------
+ {1:1}/1
+(1 row)
+
+SELECT l2_normalize('{1:3e38,2:1e-37}/2'::sparsevec);
+ l2_normalize 
+--------------
+ {1:1}/2
+(1 row)
+
+SELECT l2_normalize('{2:3e37,4:3e-37,6:4e37,8:4e-37}/9'::sparsevec);
+  l2_normalize   
+-----------------
+ {2:0.6,6:0.8}/9
+(1 row)
+

test/expected/sparsevec_input.out

@@ -164,8 +164,18 @@ SELECT '{1:0,2:1,3:0}/3'::sparsevec;
 (1 row)
 
 SELECT '{2:1,1:1}/2'::sparsevec;
-ERROR:  indexes must be in ascending order
-LINE 1: SELECT '{2:1,1:1}/2'::sparsevec;
+  sparsevec  
+-------------
+ {1:1,2:1}/2
+(1 row)
+
+SELECT '{1:1,1:1}/2'::sparsevec;
+ERROR:  indexes must not contain duplicates
+LINE 1: SELECT '{1:1,1:1}/2'::sparsevec;
+               ^
+SELECT '{1:1,2:1,1:1}/2'::sparsevec;
+ERROR:  indexes must not contain duplicates
+LINE 1: SELECT '{1:1,2:1,1:1}/2'::sparsevec;
                ^
 SELECT '{}/5'::sparsevec;
  sparsevec 

test/expected/vector_functions.out

@@ -6,6 +6,8 @@ SELECT '[1,2,3]'::vector + '[4,5,6]';
 
 SELECT '[3e38]'::vector + '[3e38]';
 ERROR:  value out of range: overflow
+SELECT '[1,2]'::vector + '[3]';
+ERROR:  different vector dimensions 2 and 1
 SELECT '[1,2,3]'::vector - '[4,5,6]';
   ?column?  
 ------------
@@ -14,6 +16,8 @@ SELECT '[1,2,3]'::vector - '[4,5,6]';
 
 SELECT '[-3e38]'::vector - '[3e38]';
 ERROR:  value out of range: overflow
+SELECT '[1,2]'::vector - '[3]';
+ERROR:  different vector dimensions 2 and 1
 SELECT '[1,2,3]'::vector * '[4,5,6]';
  ?column?  
 -----------
@@ -24,6 +28,16 @@ SELECT '[1e37]'::vector * '[1e37]';
 ERROR:  value out of range: overflow
 SELECT '[1e-37]'::vector * '[1e-37]';
 ERROR:  value out of range: underflow
+SELECT '[1,2]'::vector * '[3]';
+ERROR:  different vector dimensions 2 and 1
+SELECT '[1,2,3]'::vector || '[4,5]';
+  ?column?   
+-------------
+ [1,2,3,4,5]
+(1 row)
+
+SELECT array_fill(0, ARRAY[16000])::vector || '[1]';
+ERROR:  vector cannot have more than 16000 dimensions
 SELECT '[1,2,3]'::vector < '[1,2,3]';
  ?column? 
 ----------
@@ -138,7 +152,7 @@ SELECT vector_cmp('[2,3]', '[1,2,3]');
           1
 (1 row)
 
-SELECT vector_dims('[1,2,3]');
+SELECT vector_dims('[1,2,3]'::vector);
  vector_dims 
 -------------
            3
@@ -188,6 +202,18 @@ SELECT l2_distance('[3e38]'::vector, '[-3e38]');
     Infinity
 (1 row)
 
+SELECT l2_distance('[1,1,1,1,1,1,1,1,1]'::vector, '[1,1,1,1,1,1,1,4,5]');
+ l2_distance 
+-------------
+           5
+(1 row)
+
+SELECT '[0,0]'::vector <-> '[3,4]';
+ ?column? 
+----------
+        5
+(1 row)
+
 SELECT inner_product('[1,2]'::vector, '[3,4]');
  inner_product 
 ---------------
@@ -202,6 +228,18 @@ SELECT inner_product('[3e38]'::vector, '[3e38]');
       Infinity
 (1 row)
 
+SELECT inner_product('[1,1,1,1,1,1,1,1,1]'::vector, '[1,2,3,4,5,6,7,8,9]');
+ inner_product 
+---------------
+            45
+(1 row)
+
+SELECT '[1,2]'::vector <#> '[3,4]';
+ ?column? 
+----------
+      -11
+(1 row)
+
 SELECT cosine_distance('[1,2]'::vector, '[2,4]');
  cosine_distance 
 -----------------
@@ -252,6 +290,24 @@ SELECT cosine_distance('[3e38]'::vector, '[3e38]');
              NaN
 (1 row)
 
+SELECT cosine_distance('[1,2,3,4,5,6,7,8,9]'::vector, '[1,2,3,4,5,6,7,8,9]');
+ cosine_distance 
+-----------------
+               0
+(1 row)
+
+SELECT cosine_distance('[1,2,3,4,5,6,7,8,9]'::vector, '[-1,-2,-3,-4,-5,-6,-7,-8,-9]');
+ cosine_distance 
+-----------------
+               2
+(1 row)
+
+SELECT '[1,2]'::vector <=> '[2,4]';
+ ?column? 
+----------
+        0
+(1 row)
+
 SELECT l1_distance('[0,0]'::vector, '[3,4]');
  l1_distance 
 -------------
@@ -272,14 +328,62 @@ SELECT l1_distance('[3e38]'::vector, '[-3e38]');
     Infinity
 (1 row)
 
-SELECT quantize_binary('[1,0,-1]'::vector);
- quantize_binary 
+SELECT l1_distance('[1,2,3,4,5,6,7,8,9]'::vector, '[1,2,3,4,5,6,7,8,9]');
+ l1_distance 
+-------------
+           0
+(1 row)
+
+SELECT l1_distance('[1,2,3,4,5,6,7,8,9]'::vector, '[0,3,2,5,4,7,6,9,8]');
+ l1_distance 
+-------------
+           9
+(1 row)
+
+SELECT '[0,0]'::vector <+> '[3,4]';
+ ?column? 
+----------
+        7
+(1 row)
+
+SELECT l2_normalize('[3,4]'::vector);
+ l2_normalize 
+--------------
+ [0.6,0.8]
+(1 row)
+
+SELECT l2_normalize('[3,0]'::vector);
+ l2_normalize 
+--------------
+ [1,0]
+(1 row)
+
+SELECT l2_normalize('[0,0.1]'::vector);
+ l2_normalize 
+--------------
+ [0,1]
+(1 row)
+
+SELECT l2_normalize('[0,0]'::vector);
+ l2_normalize 
+--------------
+ [0,0]
+(1 row)
+
+SELECT l2_normalize('[3e38]'::vector);
+ l2_normalize 
+--------------
+ [1]
+(1 row)
+
+SELECT binary_quantize('[1,0,-1]'::vector);
+ binary_quantize 
 -----------------
  100
 (1 row)
 
-SELECT quantize_binary('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]'::vector);
- quantize_binary 
+SELECT binary_quantize('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]'::vector);
+ binary_quantize 
 -----------------
  01001110101
 (1 row)

test/sql/bit_functions.sql

@@ -3,6 +3,9 @@ SELECT hamming_distance('111', '110');
 SELECT hamming_distance('111', '100');
 SELECT hamming_distance('111', '000');
 SELECT hamming_distance('10101010101010101010', '01010101010101010101');
+SELECT hamming_distance('101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101', '101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101');
+SELECT hamming_distance('101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101', '010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010');
+SELECT hamming_distance('110000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000011', '100000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001');
 SELECT hamming_distance('', '');
 SELECT hamming_distance('111', '00');
 SELECT hamming_distance('111', '000'::varbit(4));
@@ -15,6 +18,9 @@ SELECT jaccard_distance('1111', '1000');
 SELECT jaccard_distance('1111', '0000');
 SELECT jaccard_distance('1100', '1000');
 SELECT jaccard_distance('10101010101010101010', '01010101010101010101');
+SELECT jaccard_distance('101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101', '101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101');
+SELECT jaccard_distance('101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101', '010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010101010');
+SELECT jaccard_distance('110000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000011', '100000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000001');
 SELECT jaccard_distance('', '');
 SELECT jaccard_distance('1111', '000');
 SELECT jaccard_distance('1111', '0000'::varbit(5));

test/sql/btree_halfvec.sql

@@ -0,0 +1,10 @@
+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 (val);
+
+SELECT * FROM t WHERE val = '[1,2,3]';
+SELECT * FROM t ORDER BY val;
+
+DROP TABLE t;

test/sql/btree_sparsevec.sql

@@ -0,0 +1,10 @@
+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);
+CREATE INDEX ON t (val);
+
+SELECT * FROM t WHERE val = '{1:1,2:2,3:3}/3';
+SELECT * FROM t ORDER BY val;
+
+DROP TABLE t;

test/sql/btree_vector.sql

@@ -5,6 +5,6 @@ INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
 CREATE INDEX ON t (val);
 
 SELECT * FROM t WHERE val = '[1,2,3]';
-SELECT * FROM t ORDER BY val LIMIT 1;
+SELECT * FROM t ORDER BY val;
 
 DROP TABLE t;

test/sql/cast.sql

@@ -47,6 +47,17 @@ 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 '[0,1.5,0,3.5,0]'::halfvec::sparsevec;
+SELECT '[0,1.5,0,3.5,0]'::halfvec::sparsevec(5);
+SELECT '[0,1.5,0,3.5,0]'::halfvec::sparsevec(4);
+
+SELECT '{2:1.5,4:3.5}/5'::sparsevec::halfvec;
+SELECT '{2:1.5,4:3.5}/5'::sparsevec::halfvec(5);
+SELECT '{2:1.5,4:3.5}/5'::sparsevec::halfvec(4);
+SELECT '{}/16001'::sparsevec::halfvec;
+SELECT '{1:65520}/1'::sparsevec::halfvec;
+SELECT '{1:1e-8}/1'::sparsevec::halfvec;
+
 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/halfvec_functions.sql

@@ -1,6 +1,45 @@
-SELECT round(halfvec_norm('[1,1]')::numeric, 5);
-SELECT halfvec_norm('[3,4]');
-SELECT halfvec_norm('[0,1]');
+SELECT '[1,2,3]'::halfvec + '[4,5,6]';
+SELECT '[65519]'::halfvec + '[65519]';
+SELECT '[1,2]'::halfvec + '[3]';
+
+SELECT '[1,2,3]'::halfvec - '[4,5,6]';
+SELECT '[-65519]'::halfvec - '[65519]';
+SELECT '[1,2]'::halfvec - '[3]';
+
+SELECT '[1,2,3]'::halfvec * '[4,5,6]';
+SELECT '[65519]'::halfvec * '[65519]';
+SELECT '[1e-7]'::halfvec * '[1e-7]';
+SELECT '[1,2]'::halfvec * '[3]';
+
+SELECT '[1,2,3]'::halfvec || '[4,5]';
+SELECT array_fill(0, ARRAY[16000])::halfvec || '[1]';
+
+SELECT '[1,2,3]'::halfvec < '[1,2,3]';
+SELECT '[1,2,3]'::halfvec < '[1,2]';
+SELECT '[1,2,3]'::halfvec <= '[1,2,3]';
+SELECT '[1,2,3]'::halfvec <= '[1,2]';
+SELECT '[1,2,3]'::halfvec = '[1,2,3]';
+SELECT '[1,2,3]'::halfvec = '[1,2]';
+SELECT '[1,2,3]'::halfvec != '[1,2,3]';
+SELECT '[1,2,3]'::halfvec != '[1,2]';
+SELECT '[1,2,3]'::halfvec >= '[1,2,3]';
+SELECT '[1,2,3]'::halfvec >= '[1,2]';
+SELECT '[1,2,3]'::halfvec > '[1,2,3]';
+SELECT '[1,2,3]'::halfvec > '[1,2]';
+
+SELECT halfvec_cmp('[1,2,3]', '[1,2,3]');
+SELECT halfvec_cmp('[1,2,3]', '[0,0,0]');
+SELECT halfvec_cmp('[0,0,0]', '[1,2,3]');
+SELECT halfvec_cmp('[1,2]', '[1,2,3]');
+SELECT halfvec_cmp('[1,2,3]', '[1,2]');
+SELECT halfvec_cmp('[1,2]', '[2,3,4]');
+SELECT halfvec_cmp('[2,3]', '[1,2,3]');
+
+SELECT vector_dims('[1,2,3]'::halfvec);
+
+SELECT round(l2_norm('[1,1]'::halfvec)::numeric, 5);
+SELECT l2_norm('[3,4]'::halfvec);
+SELECT l2_norm('[0,1]'::halfvec);
 
 SELECT l2_distance('[0,0]'::halfvec, '[3,4]');
 SELECT l2_distance('[0,0]'::halfvec, '[0,1]');
@@ -22,14 +61,25 @@ SELECT cosine_distance('[1,1]'::halfvec, '[-1,-1]');
 SELECT cosine_distance('[1,2]'::halfvec, '[3]');
 SELECT cosine_distance('[1,1]'::halfvec, '[1.1,1.1]');
 SELECT cosine_distance('[1,1]'::halfvec, '[-1.1,-1.1]');
+SELECT cosine_distance('[1,2,3,4,5,6,7,8,9]'::halfvec, '[1,2,3,4,5,6,7,8,9]');
+SELECT cosine_distance('[1,2,3,4,5,6,7,8,9]'::halfvec, '[-1,-2,-3,-4,-5,-6,-7,-8,-9]');
 SELECT '[1,2]'::halfvec <=> '[2,4]';
 
 SELECT l1_distance('[0,0]'::halfvec, '[3,4]');
 SELECT l1_distance('[0,0]'::halfvec, '[0,1]');
 SELECT l1_distance('[1,2]'::halfvec, '[3]');
+SELECT l1_distance('[1,2,3,4,5,6,7,8,9]'::halfvec, '[1,2,3,4,5,6,7,8,9]');
+SELECT l1_distance('[1,2,3,4,5,6,7,8,9]'::halfvec, '[0,3,2,5,4,7,6,9,8]');
+SELECT '[0,0]'::halfvec <+> '[3,4]';
 
-SELECT quantize_binary('[1,0,-1]'::halfvec);
-SELECT quantize_binary('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]'::halfvec);
+SELECT l2_normalize('[3,4]'::halfvec);
+SELECT l2_normalize('[3,0]'::halfvec);
+SELECT l2_normalize('[0,0.1]'::halfvec);
+SELECT l2_normalize('[0,0]'::halfvec);
+SELECT l2_normalize('[65504]'::halfvec);
+
+SELECT binary_quantize('[1,0,-1]'::halfvec);
+SELECT binary_quantize('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]'::halfvec);
 
 SELECT subvector('[1,2,3,4,5]'::halfvec, 1, 3);
 SELECT subvector('[1,2,3,4,5]'::halfvec, 3, 2);
@@ -38,3 +88,16 @@ SELECT subvector('[1,2,3,4,5]'::halfvec, 3, 9);
 SELECT subvector('[1,2,3,4,5]'::halfvec, 1, 0);
 SELECT subvector('[1,2,3,4,5]'::halfvec, 3, -1);
 SELECT subvector('[1,2,3,4,5]'::halfvec, -1, 2);
+
+SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::halfvec, '[3,5,7]']) v;
+SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::halfvec, '[3,5,7]', NULL]) v;
+SELECT avg(v) FROM unnest(ARRAY[]::halfvec[]) v;
+SELECT avg(v) FROM unnest(ARRAY['[1,2]'::halfvec, '[3]']) v;
+SELECT avg(v) FROM unnest(ARRAY['[65504]'::halfvec, '[65504]']) v;
+SELECT halfvec_avg(array_agg(n)) FROM generate_series(1, 16002) n;
+
+SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::halfvec, '[3,5,7]']) v;
+SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::halfvec, '[3,5,7]', NULL]) v;
+SELECT sum(v) FROM unnest(ARRAY[]::halfvec[]) v;
+SELECT sum(v) FROM unnest(ARRAY['[1,2]'::halfvec, '[3]']) v;
+SELECT sum(v) FROM unnest(ARRAY['[65504]'::halfvec, '[65504]']) v;

test/sql/hnsw_halfvec_l1.sql

@@ -0,0 +1,12 @@
+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 hnsw (val halfvec_l1_ops);
+
+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/hnsw_sparsevec_l1.sql

@@ -0,0 +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);
+CREATE INDEX ON t USING hnsw (val sparsevec_l1_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';
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <+> (SELECT NULL::sparsevec)) t2;
+
+DROP TABLE t;

test/sql/hnsw_vector_l1.sql

@@ -0,0 +1,12 @@
+SET enable_seqscan = off;
+
+CREATE TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING hnsw (val vector_l1_ops);
+
+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;
+
+DROP TABLE t;

test/sql/ivfflat_bit_hamming.sql

@@ -0,0 +1,20 @@
+SET enable_seqscan = off;
+
+CREATE TABLE t (val bit(3));
+INSERT INTO t (val) VALUES (B'000'), (B'100'), (B'111'), (NULL);
+CREATE INDEX ON t USING ivfflat (val bit_hamming_ops) WITH (lists = 1);
+
+INSERT INTO t (val) VALUES (B'110');
+
+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 ivfflat (val bit_hamming_ops) WITH (lists = 1);
+CREATE INDEX ON t USING ivfflat ((val::bit(3)) bit_hamming_ops) WITH (lists = 1);
+CREATE INDEX ON t USING ivfflat ((val::bit(64001)) bit_hamming_ops) WITH (lists = 1);
+CREATE INDEX ON t USING ivfflat ((val::bit(2)) bit_hamming_ops) WITH (lists = 5);
+DROP TABLE t;

test/sql/sparsevec_functions.sql

@@ -1,18 +1,65 @@
-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 '{1:1,2:2,3:3}/3'::sparsevec < '{1:1,2:2,3:3}/3';
+SELECT '{1:1,2:2,3:3}/3'::sparsevec < '{1:1,2:2}/2';
+SELECT '{1:1,2:2,3:3}/3'::sparsevec <= '{1:1,2:2,3:3}/3';
+SELECT '{1:1,2:2,3:3}/3'::sparsevec <= '{1:1,2:2}/2';
+SELECT '{1:1,2:2,3:3}/3'::sparsevec = '{1:1,2:2,3:3}/3';
+SELECT '{1:1,2:2,3:3}/3'::sparsevec = '{1:1,2:2}/2';
+SELECT '{1:1,2:2,3:3}/3'::sparsevec != '{1:1,2:2,3:3}/3';
+SELECT '{1:1,2:2,3:3}/3'::sparsevec != '{1:1,2:2}/2';
+SELECT '{1:1,2:2,3:3}/3'::sparsevec >= '{1:1,2:2,3:3}/3';
+SELECT '{1:1,2:2,3:3}/3'::sparsevec >= '{1:1,2:2}/2';
+SELECT '{1:1,2:2,3:3}/3'::sparsevec > '{1:1,2:2,3:3}/3';
+SELECT '{1:1,2:2,3:3}/3'::sparsevec > '{1:1,2:2}/2';
+
+SELECT sparsevec_cmp('{1:1,2:2,3:3}/3', '{1:1,2:2,3:3}/3');
+SELECT sparsevec_cmp('{1:1,2:2,3:3}/3', '{}/3');
+SELECT sparsevec_cmp('{}/3', '{1:1,2:2,3:3}/3');
+SELECT sparsevec_cmp('{1:1,2:2}/2', '{1:1,2:2,3:3}/3');
+SELECT sparsevec_cmp('{1:1,2:2,3:3}/3', '{1:1,2:2}/2');
+SELECT sparsevec_cmp('{1:1,2:2}/2', '{1:2,2:3,3:4}/3');
+SELECT sparsevec_cmp('{1:2,2:3}/2', '{1:1,2:2,3:3}/3');
+
+SELECT round(l2_norm('{1:1,2:1}/2'::sparsevec)::numeric, 5);
+SELECT l2_norm('{1:3,2:4}/2'::sparsevec);
+SELECT l2_norm('{2:1}/2'::sparsevec);
+SELECT l2_norm('{1:3e37,2:4e37}/2'::sparsevec)::real;
 
 SELECT l2_distance('{}/2'::sparsevec, '{1:3,2:4}/2');
+SELECT l2_distance('{1:3}/2'::sparsevec, '{2:4}/2');
+SELECT l2_distance('{2:4}/2'::sparsevec, '{1:3}/2');
 SELECT l2_distance('{}/2'::sparsevec, '{2:1}/2');
 SELECT '{}/2'::sparsevec <-> '{1:3,2:4}/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 inner_product('{1:1,2:2}/2'::sparsevec, '{1:3}/1');
+SELECT inner_product('{1:3e38}/1'::sparsevec, '{1:3e38}/1');
+SELECT inner_product('{1:1,3:3,5:5}/5'::sparsevec, '{2:4,3:6,4:8}/5');
+SELECT '{1:1,2:2}/2'::sparsevec <#> '{1:3,2:4}/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('{1:1}/2'::sparsevec, '{2:2}/2');
 SELECT cosine_distance('{1:1,2:1}/2'::sparsevec, '{1:-1,2:-1}/2');
 SELECT cosine_distance('{1:2}/2'::sparsevec, '{2:2}/2');
+SELECT cosine_distance('{1:1,2:2}/2'::sparsevec, '{1:3}/1');
+SELECT cosine_distance('{1:1,2:1}/2'::sparsevec, '{1:1.1,2:1.1}/2');
+SELECT cosine_distance('{1:1,2:1}/2'::sparsevec, '{1:-1.1,2:-1.1}/2');
+SELECT cosine_distance('{1:3e38}/1'::sparsevec, '{1:3e38}/1');
 SELECT cosine_distance('{}/1'::sparsevec, '{}/1');
-SELECT cosine_distance('{1:2}/2'::sparsevec, '{1:1}/3');
+SELECT '{1:1,2:2}/2'::sparsevec <=> '{1:2,2:4}/2';
+
+SELECT l1_distance('{}/2'::sparsevec, '{1:3,2:4}/2');
+SELECT l1_distance('{}/2'::sparsevec, '{2:1}/2');
+SELECT l1_distance('{1:1,2:2}/2'::sparsevec, '{1:3}/1');
+SELECT l1_distance('{1:3e38}/1'::sparsevec, '{1:-3e38}/1');
+SELECT l1_distance('{1:1,3:3,5:5,7:7}/8'::sparsevec, '{2:2,4:4,6:6,8:8}/8');
+SELECT l1_distance('{1:1,3:3,5:5,7:7,9:9}/9'::sparsevec, '{2:2,4:4,6:6,8:8}/9');
+
+SELECT l2_normalize('{1:3,2:4}/2'::sparsevec);
+SELECT l2_normalize('{1:3}/2'::sparsevec);
+SELECT l2_normalize('{2:0.1}/2'::sparsevec);
+SELECT l2_normalize('{}/2'::sparsevec);
+SELECT l2_normalize('{1:3e38}/1'::sparsevec);
+SELECT l2_normalize('{1:3e38,2:1e-37}/2'::sparsevec);
+SELECT l2_normalize('{2:3e37,4:3e-37,6:4e37,8:4e-37}/9'::sparsevec);

test/sql/sparsevec_input.sql

@@ -34,6 +34,8 @@ 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 '{1:1,1:1}/2'::sparsevec;
+SELECT '{1:1,2:1,1:1}/2'::sparsevec;
 SELECT '{}/5'::sparsevec;
 SELECT '{}/-1'::sparsevec;
 SELECT '{}/100001'::sparsevec;

test/sql/vector_functions.sql

@@ -1,10 +1,18 @@
 SELECT '[1,2,3]'::vector + '[4,5,6]';
 SELECT '[3e38]'::vector + '[3e38]';
+SELECT '[1,2]'::vector + '[3]';
+
 SELECT '[1,2,3]'::vector - '[4,5,6]';
 SELECT '[-3e38]'::vector - '[3e38]';
+SELECT '[1,2]'::vector - '[3]';
+
 SELECT '[1,2,3]'::vector * '[4,5,6]';
 SELECT '[1e37]'::vector * '[1e37]';
 SELECT '[1e-37]'::vector * '[1e-37]';
+SELECT '[1,2]'::vector * '[3]';
+
+SELECT '[1,2,3]'::vector || '[4,5]';
+SELECT array_fill(0, ARRAY[16000])::vector || '[1]';
 
 SELECT '[1,2,3]'::vector < '[1,2,3]';
 SELECT '[1,2,3]'::vector < '[1,2]';
@@ -27,7 +35,7 @@ SELECT vector_cmp('[1,2,3]', '[1,2]');
 SELECT vector_cmp('[1,2]', '[2,3,4]');
 SELECT vector_cmp('[2,3]', '[1,2,3]');
 
-SELECT vector_dims('[1,2,3]');
+SELECT vector_dims('[1,2,3]'::vector);
 
 SELECT round(vector_norm('[1,1]')::numeric, 5);
 SELECT vector_norm('[3,4]');
@@ -38,10 +46,14 @@ SELECT l2_distance('[0,0]'::vector, '[3,4]');
 SELECT l2_distance('[0,0]'::vector, '[0,1]');
 SELECT l2_distance('[1,2]'::vector, '[3]');
 SELECT l2_distance('[3e38]'::vector, '[-3e38]');
+SELECT l2_distance('[1,1,1,1,1,1,1,1,1]'::vector, '[1,1,1,1,1,1,1,4,5]');
+SELECT '[0,0]'::vector <-> '[3,4]';
 
 SELECT inner_product('[1,2]'::vector, '[3,4]');
 SELECT inner_product('[1,2]'::vector, '[3]');
 SELECT inner_product('[3e38]'::vector, '[3e38]');
+SELECT inner_product('[1,1,1,1,1,1,1,1,1]'::vector, '[1,2,3,4,5,6,7,8,9]');
+SELECT '[1,2]'::vector <#> '[3,4]';
 
 SELECT cosine_distance('[1,2]'::vector, '[2,4]');
 SELECT cosine_distance('[1,2]'::vector, '[0,0]');
@@ -52,14 +64,26 @@ SELECT cosine_distance('[1,2]'::vector, '[3]');
 SELECT cosine_distance('[1,1]'::vector, '[1.1,1.1]');
 SELECT cosine_distance('[1,1]'::vector, '[-1.1,-1.1]');
 SELECT cosine_distance('[3e38]'::vector, '[3e38]');
+SELECT cosine_distance('[1,2,3,4,5,6,7,8,9]'::vector, '[1,2,3,4,5,6,7,8,9]');
+SELECT cosine_distance('[1,2,3,4,5,6,7,8,9]'::vector, '[-1,-2,-3,-4,-5,-6,-7,-8,-9]');
+SELECT '[1,2]'::vector <=> '[2,4]';
 
 SELECT l1_distance('[0,0]'::vector, '[3,4]');
 SELECT l1_distance('[0,0]'::vector, '[0,1]');
 SELECT l1_distance('[1,2]'::vector, '[3]');
 SELECT l1_distance('[3e38]'::vector, '[-3e38]');
+SELECT l1_distance('[1,2,3,4,5,6,7,8,9]'::vector, '[1,2,3,4,5,6,7,8,9]');
+SELECT l1_distance('[1,2,3,4,5,6,7,8,9]'::vector, '[0,3,2,5,4,7,6,9,8]');
+SELECT '[0,0]'::vector <+> '[3,4]';
 
-SELECT quantize_binary('[1,0,-1]'::vector);
-SELECT quantize_binary('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]'::vector);
+SELECT l2_normalize('[3,4]'::vector);
+SELECT l2_normalize('[3,0]'::vector);
+SELECT l2_normalize('[0,0.1]'::vector);
+SELECT l2_normalize('[0,0]'::vector);
+SELECT l2_normalize('[3e38]'::vector);
+
+SELECT binary_quantize('[1,0,-1]'::vector);
+SELECT binary_quantize('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]'::vector);
 
 SELECT subvector('[1,2,3,4,5]'::vector, 1, 3);
 SELECT subvector('[1,2,3,4,5]'::vector, 3, 2);

test/t/012_hnsw_build_recall.pl

@@ -67,8 +67,8 @@ for (1 .. 20)
 }
 
 # Check each index type
-my @operators = ("<->", "<#>", "<=>");
-my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
+my @operators = ("<->", "<#>", "<=>", "<+>");
+my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops", "vector_l1_ops");
 
 for my $i (0 .. $#operators)
 {

test/t/013_hnsw_insert_recall.pl

@@ -64,8 +64,8 @@ for (1 .. 20)
 }
 
 # Check each index type
-my @operators = ("<->", "<#>", "<=>");
-my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
+my @operators = ("<->", "<#>", "<=>", "<+>");
+my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops", "vector_l1_ops");
 
 for my $i (0 .. $#operators)
 {

test/t/021_hnsw_halfvec_build_recall.pl

@@ -71,8 +71,8 @@ for (1 .. 20)
 }
 
 # Check each index type
-my @operators = ("<->", "<#>", "<=>");
-my @opclasses = ("halfvec_l2_ops", "halfvec_ip_ops", "halfvec_cosine_ops");
+my @operators = ("<->", "<#>", "<=>", "<+>");
+my @opclasses = ("halfvec_l2_ops", "halfvec_ip_ops", "halfvec_cosine_ops", "halfvec_l1_ops");
 
 for my $i (0 .. $#operators)
 {

test/t/022_hnsw_sparsevec_build_recall.pl

@@ -67,8 +67,8 @@ for (1 .. 20)
 }
 
 # Check each index type
-my @operators = ("<->", "<#>", "<=>");
-my @opclasses = ("sparsevec_l2_ops", "sparsevec_ip_ops", "sparsevec_cosine_ops");
+my @operators = ("<->", "<#>", "<=>", "<+>");
+my @opclasses = ("sparsevec_l2_ops", "sparsevec_ip_ops", "sparsevec_cosine_ops", "sparsevec_l1_ops");
 
 for my $i (0 .. $#operators)
 {

test/t/024_hnsw_halfvec_insert_recall.pl

@@ -68,8 +68,8 @@ for (1 .. 20)
 }
 
 # Check each index type
-my @operators = ("<->", "<#>", "<=>");
-my @opclasses = ("halfvec_l2_ops", "halfvec_ip_ops", "halfvec_cosine_ops");
+my @operators = ("<->", "<#>", "<=>", "<+>");
+my @opclasses = ("halfvec_l2_ops", "halfvec_ip_ops", "halfvec_cosine_ops", "halfvec_l1_ops");
 
 for my $i (0 .. $#operators)
 {

test/t/025_hnsw_sparsevec_insert_recall.pl

@@ -64,8 +64,8 @@ for (1 .. 20)
 }
 
 # Check each index type
-my @operators = ("<->", "<#>", "<=>");
-my @opclasses = ("sparsevec_l2_ops", "sparsevec_ip_ops", "sparsevec_cosine_ops");
+my @operators = ("<->", "<#>", "<=>", "<+>");
+my @opclasses = ("sparsevec_l2_ops", "sparsevec_ip_ops", "sparsevec_cosine_ops", "sparsevec_l1_ops");
 
 for my $i (0 .. $#operators)
 {
@@ -99,7 +99,7 @@ for my $i (0 .. $#operators)
 	}
 
 	# Test approximate results
-	my $min = $operator eq "<#>" ? 0.80 : 0.99;
+	my $min = $operator eq "<#>" ? 0.79 : 0.99;
 	test_recall($min, $operator);
 
 	$node->safe_psql("postgres", "DROP INDEX idx;");

test/t/032_ivfflat_halfvec_build_recall.pl

@@ -105,8 +105,8 @@ for my $i (0 .. $#operators)
 	if ($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_recall(1, 0.34, $operator);
+		test_recall(10, 0.93, $operator);
 	}
 
 	# Test probes equals lists
@@ -134,8 +134,8 @@ for my $i (0 .. $#operators)
 	if ($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_recall(1, 0.34, $operator);
+		test_recall(10, 0.93, $operator);
 	}
 
 	# Test probes equals lists

test/t/033_comparison.pl

@@ -0,0 +1,48 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+my $node;
+my $array_sql = join(",", ('floor(random() * 2)::int - 1') x 3);
+
+# Initialize node
+$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 (v real[]);");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
+);
+
+for (1 .. 50)
+{
+	# Generate query
+	my @r = ();
+	for (1 .. (int(rand() * 3) + 2))
+	{
+		push(@r, int(rand() * 2) - 1);
+	}
+	my $query = "{" . join(",", @r) . "}";
+
+	# Get expected result
+	my $expected = $node->safe_psql("postgres", "SELECT btarraycmp(v, '$query') FROM tst");
+
+	# Test vector
+	my $actual = $node->safe_psql("postgres", "SELECT vector_cmp(v::vector, '$query'::real[]::vector) FROM tst");
+	is($expected, $actual);
+
+	# Test halfvec
+	$actual = $node->safe_psql("postgres", "SELECT halfvec_cmp(v::halfvec, '$query'::real[]::halfvec) FROM tst");
+	is($expected, $actual);
+
+	# Test sparsevec
+	$actual = $node->safe_psql("postgres", "SELECT sparsevec_cmp(v::vector::sparsevec, '$query'::real[]::vector::sparsevec) FROM tst");
+	is($expected, $actual);
+}
+
+done_testing();

test/t/034_distance_functions.pl

@@ -0,0 +1,54 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+my $node;
+my @queries = ();
+my $dim = 5;
+my $array_sql = join(",", ('floor(random() * 4)::int - 2') x $dim);
+
+# Initialize node
+$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 (v vector($dim));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
+);
+
+# Generate queries
+for (1 .. 20)
+{
+	my @r = ();
+	for (1 .. $dim)
+	{
+		push(@r, int(rand() * 4) - 2);
+	}
+	push(@queries, "[" . join(",", @r) . "]");
+}
+
+# Check each distance function
+my @functions = ("l2_distance", "inner_product", "cosine_distance", "l1_distance");
+
+for my $function (@functions)
+{
+	for my $query (@queries)
+	{
+		my $expected = $node->safe_psql("postgres", "SELECT $function(v, '$query') FROM tst");
+
+		# Test halfvec
+		my $actual = $node->safe_psql("postgres", "SELECT $function(v::halfvec, '$query'::vector::halfvec) FROM tst");
+		is($expected, $actual, "halfvec $function");
+
+		# Test sparsevec
+		$actual = $node->safe_psql("postgres", "SELECT $function(v::sparsevec, '$query'::vector::sparsevec) FROM tst");
+		is($expected, $actual, "sparsevec $function");
+	}
+}
+
+done_testing();

test/t/035_ivfflat_bit_build_recall.pl

@@ -0,0 +1,128 @@
+use strict;
+use warnings;
+use PostgresNode;
+use TestLib;
+use Test::More;
+
+my $node;
+my @queries = ();
+my @expected;
+my $limit = 20;
+my $dim = 52;
+my $max = 2**$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 bit($dim));");
+$node->safe_psql("postgres",
+	"INSERT INTO tst SELECT i, (random() * $max)::bigint::bit($dim) FROM generate_series(1, 100000) i;"
+);
+
+# 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");
+
+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
+	test_recall(1, 0.08, $operator);
+	test_recall(10, 0.50, $operator);
+
+	# Test probes equals lists
+	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
+	test_recall(1, 0.08, $operator);
+	test_recall(10, 0.50, $operator);
+
+	# Test probes equals lists
+	test_recall(100, 1.00, $operator);
+
+	$node->safe_psql("postgres", "DROP INDEX idx;");
+}
+
+done_testing();