Added category/preferred to types for new installations [skip ci]

.github/workflows/build.yml

@@ -8,8 +8,18 @@ jobs:
       fail-fast: false
       matrix:
         include:
+          - postgres: 17
+            os: ubuntu-22.04
+          - postgres: 16
+            os: ubuntu-22.04
+          - postgres: 15
+            os: ubuntu-22.04
           - postgres: 14
             os: ubuntu-22.04
+          - postgres: 13
+            os: ubuntu-20.04
+          - postgres: 12
+            os: ubuntu-20.04
     steps:
       - uses: actions/checkout@v4
       - uses: ankane/setup-postgres@v1
@@ -28,4 +38,93 @@ jobs:
       - run: |
           sudo apt-get update
           sudo apt-get install libipc-run-perl
-      - run: for run in {1..100}; do make prove_installcheck PROVE_TESTS=test/t/031_hnsw_sparsevec_vacuum_recall.pl || break; done
+      - run: make prove_installcheck
+  mac:
+    runs-on: ${{ matrix.os }}
+    if: ${{ !startsWith(github.ref_name, 'windows') }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - postgres: 16
+            os: macos-14
+          - postgres: 14
+            os: macos-12
+    steps:
+      - uses: actions/checkout@v4
+      - uses: ankane/setup-postgres@v1
+        with:
+          postgres-version: ${{ matrix.postgres }}
+      - run: make
+        env:
+          PG_CFLAGS: -DUSE_ASSERT_CHECKING -Wall -Wextra -Werror -Wno-unused-parameter
+      - run: make install
+      - run: make installcheck
+      - if: ${{ failure() }}
+        run: cat regression.diffs
+      # Homebrew Postgres does not enable TAP tests, so need to download
+      - run: |
+          brew install cpanm
+          cpanm --notest IPC::Run
+          wget -q https://github.com/postgres/postgres/archive/refs/tags/$TAG.tar.gz
+          tar xf $TAG.tar.gz
+          mv postgres-$TAG postgres
+        env:
+          TAG: ${{ matrix.postgres == 16 && 'REL_16_2' || 'REL_14_11' }}
+      - run: make prove_installcheck PROVE_FLAGS="-I ./postgres/src/test/perl -I ./test/perl"
+        env:
+          PERL5LIB: /Users/runner/perl5/lib/perl5
+      - run: make clean && $(brew --prefix llvm@15)/bin/scan-build --status-bugs make
+        env:
+          PG_CFLAGS: -DUSE_ASSERT_CHECKING
+  windows:
+    runs-on: windows-latest
+    if: ${{ !startsWith(github.ref_name, 'mac') }}
+    steps:
+      - uses: actions/checkout@v4
+      - uses: ankane/setup-postgres@v1
+        with:
+          postgres-version: 14
+      - run: |
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat" && ^
+          cd %TEMP% && ^
+          nmake /NOLOGO /F Makefile.win && ^
+          nmake /NOLOGO /F Makefile.win install && ^
+          nmake /NOLOGO /F Makefile.win installcheck && ^
+          nmake /NOLOGO /F Makefile.win clean && ^
+          nmake /NOLOGO /F Makefile.win uninstall
+        shell: cmd
+  i386:
+    if: ${{ !startsWith(github.ref_name, 'mac') && !startsWith(github.ref_name, 'windows') }}
+    runs-on: ubuntu-latest
+    container:
+      image: debian:12
+      options: --platform linux/386
+    steps:
+      - run: apt-get update && apt-get install -y build-essential git libipc-run-perl postgresql-15 postgresql-server-dev-15 sudo
+      - run: service postgresql start
+      - run: |
+          git clone https://github.com/${{ github.repository }}.git pgvector
+          cd pgvector
+          git fetch origin ${{ github.ref }}
+          git reset --hard FETCH_HEAD
+          make
+          make install
+          chown -R postgres .
+          sudo -u postgres make installcheck
+          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
+    steps:
+      - uses: actions/checkout@v4
+      - uses: ankane/setup-postgres-valgrind@v1
+        with:
+          postgres-version: 16
+      - run: make OPTFLAGS=""
+      - run: sudo --preserve-env=PG_CONFIG make install
+      - run: make installcheck

CHANGELOG.md

@@ -2,8 +2,7 @@
 
 - Added `halfvec` type
 - Added `sparsevec` type
-- Added support for indexing `bit` type
-- Added support for indexing L1 distance with HNSW
+- Added support for `bit` vectors to HNSW
 - Added `binary_quantize` function
 - Added `hamming_distance` function
 - Added `jaccard_distance` function

Makefile

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

Makefile.win

@@ -1,7 +1,7 @@
 EXTENSION = vector
 EXTVERSION = 0.6.2
 
-OBJS = src\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
+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
 HEADERS = src\halfvec.h src\sparsevec.h src\vector.h
 
 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

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 (`<#>`), cosine distance (`<=>`), and L1 distance (`<+>`, unreleased)
+Also supports inner product (`<#>`) and cosine distance (`<=>`)
 
 Note: `<#>` returns the negative inner product since Postgres only supports `ASC` order index scans on operators
 
@@ -143,7 +143,6 @@ Supported distance functions are:
 - `<->` - L2 distance
 - `<#>` - (negative) inner product
 - `<=>` - cosine distance
-- `<+>` - L1 distance (unreleased)
 
 Get the nearest neighbors to a row
 
@@ -228,12 +227,6 @@ 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
@@ -356,17 +349,10 @@ 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
 
@@ -548,7 +534,7 @@ 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.
 
-## Indexing Subvectors
+## Subvector Indexing
 
 *Unreleased*
 
@@ -868,7 +854,6 @@ Operator | Description | Added
 <-> | Euclidean distance |
 <#> | negative inner product |
 <=> | cosine distance |
-<+> | taxicab distance | unreleased
 
 ### Vector Functions
 
@@ -906,7 +891,6 @@ Operator | Description | Added
 <-> | Euclidean distance | unreleased
 <#> | negative inner product | unreleased
 <=> | cosine distance | unreleased
-<+> | taxicab distance | unreleased
 
 ### Halfvec Functions
 
@@ -958,7 +942,6 @@ Operator | Description | Added
 <-> | Euclidean distance | unreleased
 <#> | negative inner product | unreleased
 <=> | cosine distance | unreleased
-<+> | taxicab distance | unreleased
 
 ### Sparsevec Functions
 

sql/vector--0.6.2--0.7.0.sql

@@ -13,32 +13,16 @@ CREATE FUNCTION subvector(vector, int, int) RETURNS vector
 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 = '<~>'
@@ -49,24 +33,15 @@ 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 4 bit_hnsw_support(internal);
+	FUNCTION 1 hamming_distance(bit, bit);
 
 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 4 bit_hnsw_support(internal);
+	FUNCTION 1 jaccard_distance(bit, bit);
 
 CREATE TYPE halfvec;
 
@@ -169,12 +144,6 @@ CREATE FUNCTION halfvec_accum(double precision[], halfvec) RETURNS double precis
 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[],
@@ -254,11 +223,6 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
-CREATE OPERATOR <+> (
-	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = l1_distance,
-	COMMUTATOR = '<+>'
-);
-
 CREATE OPERATOR + (
 	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_add,
 	COMMUTATOR = +
@@ -326,17 +290,14 @@ 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 6 halfvec_ivfflat_support(internal);
+	FUNCTION 3 l2_distance(halfvec, halfvec);
 
 CREATE OPERATOR CLASS halfvec_ip_ops
 	FOR TYPE halfvec USING ivfflat AS
 	OPERATOR 1 <#> (halfvec, halfvec) FOR ORDER BY float_ops,
 	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
 	FUNCTION 3 halfvec_spherical_distance(halfvec, halfvec),
-	FUNCTION 4 l2_norm(halfvec),
-	FUNCTION 5 l2_normalize(halfvec),
-	FUNCTION 6 halfvec_ivfflat_support(internal);
+	FUNCTION 4 l2_norm(halfvec);
 
 CREATE OPERATOR CLASS halfvec_cosine_ops
 	FOR TYPE halfvec USING ivfflat AS
@@ -344,35 +305,23 @@ CREATE OPERATOR CLASS halfvec_cosine_ops
 	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
 	FUNCTION 2 l2_norm(halfvec),
 	FUNCTION 3 halfvec_spherical_distance(halfvec, halfvec),
-	FUNCTION 4 l2_norm(halfvec),
-	FUNCTION 5 l2_normalize(halfvec),
-	FUNCTION 6 halfvec_ivfflat_support(internal);
+	FUNCTION 4 l2_norm(halfvec);
 
 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 4 halfvec_hnsw_support(internal);
+	FUNCTION 1 halfvec_l2_squared_distance(halfvec, halfvec);
 
 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 4 halfvec_hnsw_support(internal);
+	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec);
 
 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 l2_norm(halfvec),
-	FUNCTION 3 l2_normalize(halfvec),
-	FUNCTION 4 halfvec_hnsw_support(internal);
-
-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);
+	FUNCTION 2 l2_norm(halfvec);
 
 CREATE TYPE sparsevec;
 
@@ -454,15 +403,6 @@ 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;
 
@@ -472,12 +412,6 @@ CREATE CAST (sparsevec AS vector)
 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 = '<->'
@@ -493,11 +427,6 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
-CREATE OPERATOR <+> (
-	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = l1_distance,
-	COMMUTATOR = '<+>'
-);
-
 CREATE OPERATOR < (
 	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_lt,
 	COMMUTATOR = > , NEGATOR = >= ,
@@ -546,25 +475,15 @@ CREATE OPERATOR CLASS sparsevec_ops
 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 4 sparsevec_hnsw_support(internal);
+	FUNCTION 1 sparsevec_l2_squared_distance(sparsevec, sparsevec);
 
 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 4 sparsevec_hnsw_support(internal);
+	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec);
 
 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 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);
+	FUNCTION 2 l2_norm(sparsevec);

sql/vector.sql

@@ -26,7 +26,9 @@ CREATE TYPE vector (
 	TYPMOD_IN = vector_typmod_in,
 	RECEIVE   = vector_recv,
 	SEND      = vector_send,
-	STORAGE   = external
+	STORAGE   = external,
+	CATEGORY  = 'v',
+	PREFERRED = true
 );
 
 -- vector functions
@@ -186,11 +188,6 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
-CREATE OPERATOR <+> (
-	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = l1_distance,
-	COMMUTATOR = '<+>'
-);
-
 CREATE OPERATOR + (
 	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_add,
 	COMMUTATOR = +
@@ -309,11 +306,6 @@ 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
@@ -322,16 +314,6 @@ 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 = '<~>'
@@ -342,26 +324,15 @@ 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 4 bit_hnsw_support(internal);
+	FUNCTION 1 hamming_distance(bit, bit);
 
 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 4 bit_hnsw_support(internal);
+	FUNCTION 1 jaccard_distance(bit, bit);
 
 -- halfvec type
 
@@ -388,7 +359,8 @@ CREATE TYPE halfvec (
 	TYPMOD_IN = halfvec_typmod_in,
 	RECEIVE   = halfvec_recv,
 	SEND      = halfvec_send,
-	STORAGE   = external
+	STORAGE   = external,
+	CATEGORY  = 'v'
 );
 
 -- halfvec functions
@@ -470,12 +442,6 @@ CREATE FUNCTION halfvec_accum(double precision[], halfvec) RETURNS double precis
 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) (
@@ -563,11 +529,6 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
-CREATE OPERATOR <+> (
-	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = l1_distance,
-	COMMUTATOR = '<+>'
-);
-
 CREATE OPERATOR + (
 	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_add,
 	COMMUTATOR = +
@@ -637,17 +598,14 @@ 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 6 halfvec_ivfflat_support(internal);
+	FUNCTION 3 l2_distance(halfvec, halfvec);
 
 CREATE OPERATOR CLASS halfvec_ip_ops
 	FOR TYPE halfvec USING ivfflat AS
 	OPERATOR 1 <#> (halfvec, halfvec) FOR ORDER BY float_ops,
 	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
 	FUNCTION 3 halfvec_spherical_distance(halfvec, halfvec),
-	FUNCTION 4 l2_norm(halfvec),
-	FUNCTION 5 l2_normalize(halfvec),
-	FUNCTION 6 halfvec_ivfflat_support(internal);
+	FUNCTION 4 l2_norm(halfvec);
 
 CREATE OPERATOR CLASS halfvec_cosine_ops
 	FOR TYPE halfvec USING ivfflat AS
@@ -655,35 +613,23 @@ CREATE OPERATOR CLASS halfvec_cosine_ops
 	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
 	FUNCTION 2 l2_norm(halfvec),
 	FUNCTION 3 halfvec_spherical_distance(halfvec, halfvec),
-	FUNCTION 4 l2_norm(halfvec),
-	FUNCTION 5 l2_normalize(halfvec),
-	FUNCTION 6 halfvec_ivfflat_support(internal);
+	FUNCTION 4 l2_norm(halfvec);
 
 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 4 halfvec_hnsw_support(internal);
+	FUNCTION 1 halfvec_l2_squared_distance(halfvec, halfvec);
 
 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 4 halfvec_hnsw_support(internal);
+	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec);
 
 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 l2_norm(halfvec),
-	FUNCTION 3 l2_normalize(halfvec),
-	FUNCTION 4 halfvec_hnsw_support(internal);
-
-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);
+	FUNCTION 2 l2_norm(halfvec);
 
 --- sparsevec type
 
@@ -710,7 +656,8 @@ CREATE TYPE sparsevec (
 	TYPMOD_IN = sparsevec_typmod_in,
 	RECEIVE   = sparsevec_recv,
 	SEND      = sparsevec_send,
-	STORAGE   = external
+	STORAGE   = external,
+	CATEGORY  = 'v'
 );
 
 -- sparsevec functions
@@ -773,15 +720,6 @@ 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)
@@ -793,12 +731,6 @@ CREATE CAST (sparsevec AS vector)
 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 <-> (
@@ -816,11 +748,6 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
-CREATE OPERATOR <+> (
-	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = l1_distance,
-	COMMUTATOR = '<+>'
-);
-
 CREATE OPERATOR < (
 	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_lt,
 	COMMUTATOR = > , NEGATOR = >= ,
@@ -871,25 +798,15 @@ CREATE OPERATOR CLASS sparsevec_ops
 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 4 sparsevec_hnsw_support(internal);
+	FUNCTION 1 sparsevec_l2_squared_distance(sparsevec, sparsevec);
 
 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 4 sparsevec_hnsw_support(internal);
+	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec);
 
 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 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);
+	FUNCTION 2 l2_norm(sparsevec);

src/bitutils.c

@@ -1,217 +0,0 @@
-#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

@@ -1,31 +0,0 @@
-#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/bitvector.c

@@ -1,7 +1,7 @@
 #include "postgres.h"
 
-#include "bitutils.h"
-#include "bitvec.h"
+#include "bitvector.h"
+#include "port/pg_bitutils.h"
 #include "utils/varbit.h"
 
 #if PG_VERSION_NUM >= 160000
@@ -46,10 +46,17 @@ 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);
 
-	PG_RETURN_FLOAT8((double) BitHammingDistance(VARBITBYTES(a), VARBITS(a), VARBITS(b), 0));
+	/* 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);
 }
 
 /*
@@ -61,8 +68,23 @@ 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);
 
-	PG_RETURN_FLOAT8(BitJaccardDistance(VARBITBYTES(a), VARBITS(a), VARBITS(b), 0, 0, 0));
+	/* 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))));
 }

src/halfutils.c

@@ -8,21 +8,20 @@
 
 #if defined(HAVE__GET_CPUID)
 #include <cpuid.h>
-#else
+#elif defined(HAVE__CPUID)
 #include <intrin.h>
 #endif
 
 #ifdef _MSC_VER
-#define TARGET_F16C
+#define TARGET_F16C_FMA
 #else
-#define TARGET_F16C __attribute__((target("avx,f16c,fma")))
+#define TARGET_F16C_FMA __attribute__((target("f16c,fma")))
 #endif
 #endif
 
 float		(*HalfvecL2SquaredDistance) (int dim, half * ax, half * bx);
 float		(*HalfvecInnerProduct) (int dim, half * ax, half * bx);
 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)
@@ -41,8 +40,8 @@ HalfvecL2SquaredDistanceDefault(int dim, half * ax, half * bx)
 }
 
 #ifdef HALFVEC_DISPATCH
-TARGET_F16C static float
-HalfvecL2SquaredDistanceF16c(int dim, half * ax, half * bx)
+TARGET_F16C_FMA static float
+HalfvecL2SquaredDistanceF16cFma(int dim, half * ax, half * bx)
 {
 	float		distance;
 	int			i;
@@ -89,8 +88,8 @@ HalfvecInnerProductDefault(int dim, half * ax, half * bx)
 }
 
 #ifdef HALFVEC_DISPATCH
-TARGET_F16C static float
-HalfvecInnerProductF16c(int dim, half * ax, half * bx)
+TARGET_F16C_FMA static float
+HalfvecInnerProductF16cFma(int dim, half * ax, half * bx)
 {
 	float		distance;
 	int			i;
@@ -142,8 +141,8 @@ HalfvecCosineSimilarityDefault(int dim, half * ax, half * bx)
 }
 
 #ifdef HALFVEC_DISPATCH
-TARGET_F16C static double
-HalfvecCosineSimilarityF16c(int dim, half * ax, half * bx)
+TARGET_F16C_FMA static double
+HalfvecCosineSimilarityF16cFma(int dim, half * ax, half * bx)
 {
 	float		similarity;
 	float		norma;
@@ -192,83 +191,21 @@ HalfvecCosineSimilarityF16c(int dim, half * ax, half * bx)
 }
 #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);
+#define CPU_FEATURE_FMA  (1 << 12)
+#define CPU_FEATURE_F16C (1 << 29)
 
-	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
+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]);
-#else
+#elif defined(HAVE__CPUID)
 	__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
@@ -283,16 +220,13 @@ HalfvecInit(void)
 	HalfvecL2SquaredDistance = HalfvecL2SquaredDistanceDefault;
 	HalfvecInnerProduct = HalfvecInnerProductDefault;
 	HalfvecCosineSimilarity = HalfvecCosineSimilarityDefault;
-	HalfvecL1Distance = HalfvecL1DistanceDefault;
 
 #ifdef HALFVEC_DISPATCH
-	if (SupportsCpuFeature(CPU_FEATURE_AVX | CPU_FEATURE_F16C | CPU_FEATURE_FMA))
+	if (SupportsCpuFeature(CPU_FEATURE_FMA | CPU_FEATURE_F16C))
 	{
-		HalfvecL2SquaredDistance = HalfvecL2SquaredDistanceF16c;
-		HalfvecInnerProduct = HalfvecInnerProductF16c;
-		HalfvecCosineSimilarity = HalfvecCosineSimilarityF16c;
-		/* Does not require FMA, but keep logic simple */
-		HalfvecL1Distance = HalfvecL1DistanceF16c;
+		HalfvecL2SquaredDistance = HalfvecL2SquaredDistanceF16cFma;
+		HalfvecInnerProduct = HalfvecInnerProductF16cFma;
+		HalfvecCosineSimilarity = HalfvecCosineSimilarityF16cFma;
 	}
 #endif
 }

src/halfutils.h

@@ -13,7 +13,6 @@
 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);
 

src/halfvec.c

@@ -2,7 +2,7 @@
 
 #include <math.h>
 
-#include "bitvec.h"
+#include "bitvector.h"
 #include "catalog/pg_type.h"
 #include "common/shortest_dec.h"
 #include "fmgr.h"
@@ -11,7 +11,6 @@
 #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"
@@ -683,10 +682,17 @@ 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);
 
-	PG_RETURN_FLOAT8((double) HalfvecL1Distance(a->dim, a->x, b->x));
+	/* Auto-vectorized */
+	for (int i = 0; i < a->dim; i++)
+		distance += fabsf(HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]));
+
+	PG_RETURN_FLOAT8((double) distance);
 }
 
 /*
@@ -1168,26 +1174,3 @@ halfvec_avg(PG_FUNCTION_ARGS)
 
 	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

@@ -5,11 +5,10 @@
 
 #include <float.h>
 
-#include "bitutils.h"
 #include "fmgr.h"
 #include "vector.h"
 
-#if defined(USE_DISPATCH)
+#if defined(__x86_64__) || defined(_M_AMD64)
 #define HALFVEC_DISPATCH
 #endif
 
@@ -45,5 +44,6 @@ typedef struct HalfVector
 
 HalfVector *InitHalfVector(int dim);
 int			halfvec_cmp_internal(HalfVector * a, HalfVector * b);
+PGDLLEXPORT Datum halfvec_l2_normalize(PG_FUNCTION_ARGS);
 
 #endif

src/hnsw.c

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

src/hnsw.h

@@ -22,8 +22,6 @@
 /* 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
@@ -63,8 +61,7 @@ typedef enum HnswType
 	HNSW_TYPE_VECTOR,
 	HNSW_TYPE_HALFVEC,
 	HNSW_TYPE_BIT,
-	HNSW_TYPE_SPARSEVEC,
-	HNSW_TYPE_UNSUPPORTED
+	HNSW_TYPE_SPARSEVEC
 }			HnswType;
 
 /* Build phases */
@@ -268,7 +265,6 @@ typedef struct HnswBuildState
 	/* Support functions */
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
-	FmgrInfo   *normalizeprocinfo;
 	Oid			collation;
 
 	/* Variables */
@@ -345,7 +341,6 @@ typedef struct HnswScanOpaqueData
 	/* Support functions */
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
-	FmgrInfo   *normalizeprocinfo;
 	Oid			collation;
 }			HnswScanOpaqueData;
 
@@ -382,7 +377,7 @@ int			HnswGetM(Relation index);
 int			HnswGetEfConstruction(Relation index);
 FmgrInfo   *HnswOptionalProcInfo(Relation index, uint16 procnum);
 HnswType	HnswGetType(Relation index);
-Datum		HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum value);
+Datum		HnswNormValue(Datum value, HnswType type);
 bool		HnswCheckNorm(FmgrInfo *procinfo, Oid collation, Datum value);
 void		HnswCheckValue(Datum value, HnswType type);
 Buffer		HnswNewBuffer(Relation index, ForkNumber forkNum);

src/hnswbuild.c

@@ -496,7 +496,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
 		if (!HnswCheckNorm(buildstate->normprocinfo, buildstate->collation, value))
 			return false;
 
-		value = HnswNormValue(buildstate->normalizeprocinfo, buildstate->collation, value);
+		value = HnswNormValue(value, buildstate->type);
 	}
 
 	/* Get datum size */
@@ -725,7 +725,6 @@ 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

@@ -614,12 +614,13 @@ 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, HnswGetType(index));
+	HnswCheckValue(value, type);
 
 	/* Normalize if needed */
 	normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
@@ -628,7 +629,7 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 		if (!HnswCheckNorm(normprocinfo, collation, value))
 			return;
 
-		value = HnswNormValue(HnswOptionalProcInfo(index, HNSW_NORMALIZE_PROC), collation, value);
+		value = HnswNormValue(value, type);
 	}
 
 	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)
-			value = HnswNormValue(so->normalizeprocinfo, so->collation, value);
+			value = HnswNormValue(value, HnswGetType(scan->indexRelation));
 	}
 
 	return value;
@@ -87,7 +87,6 @@ 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

@@ -6,6 +6,8 @@
 #include "catalog/pg_type.h"
 #include "catalog/pg_type_d.h"
 #include "fmgr.h"
+#include "halfutils.h"
+#include "halfvec.h"
 #include "hnsw.h"
 #include "lib/pairingheap.h"
 #include "sparsevec.h"
@@ -13,6 +15,8 @@
 #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"
@@ -158,17 +162,32 @@ 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 (procinfo == NULL)
-		return HNSW_TYPE_VECTOR;
+	if (typid == BITOID)
+		return HNSW_TYPE_BIT;
 
-	result = (HnswType) DatumGetInt32(FunctionCall1(procinfo, ObjectIdGetDatum(typid)));
+	tuple = SearchSysCache1(TYPEOID, ObjectIdGetDatum(typid));
+	if (!HeapTupleIsValid(tuple))
+		elog(ERROR, "cache lookup failed for type %u", typid);
 
-	if (result == HNSW_TYPE_UNSUPPORTED)
+	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);
 		elog(ERROR, "type not supported for hnsw index");
+	}
+
+	ReleaseSysCache(tuple);
 
 	return result;
 }
@@ -177,12 +196,17 @@ HnswGetType(Relation index)
  * Normalize value
  */
 Datum
-HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum value)
+HnswNormValue(Datum value, HnswType type)
 {
-	if (procinfo == NULL)
+	/* TODO Remove type-specific code */
+	if (type == HNSW_TYPE_VECTOR)
 		return DirectFunctionCall1(l2_normalize, value);
-
-	return FunctionCall1Coll(procinfo, collation, value);
+	else if (type == HNSW_TYPE_HALFVEC)
+		return DirectFunctionCall1(halfvec_l2_normalize, value);
+	else if (type == HNSW_TYPE_SPARSEVEC)
+		return DirectFunctionCall1(sparsevec_l2_normalize, value);
+	else
+		elog(ERROR, "Unsupported type");
 }
 
 /*
@@ -1302,29 +1326,3 @@ 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

@@ -6,7 +6,6 @@
 #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"
@@ -60,10 +59,8 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 	 */
 	if (buildstate->kmeansnormprocinfo != NULL)
 	{
-		if (!IvfflatCheckNorm(buildstate->kmeansnormprocinfo, buildstate->collation, value))
+		if (!IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->collation, &value, buildstate->type))
 			return;
-
-		value = IvfflatNormValue(buildstate->normalizeprocinfo, buildstate->collation, value);
 	}
 
 	if (samples->length < targsamples)
@@ -158,10 +155,8 @@ AddTupleToSort(Relation index, ItemPointer tid, Datum *values, IvfflatBuildState
 	/* Normalize if needed */
 	if (buildstate->normprocinfo != NULL)
 	{
-		if (!IvfflatCheckNorm(buildstate->normprocinfo, buildstate->collation, value))
+		if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->type))
 			return;
-
-		value = IvfflatNormValue(buildstate->normalizeprocinfo, buildstate->collation, value);
 	}
 
 	/* Find the list that minimizes the distance */
@@ -329,8 +324,6 @@ GetMaxDimensions(IvfflatType type)
 
 	if (type == IVFFLAT_TYPE_HALFVEC)
 		maxDimensions *= 2;
-	else if (type == IVFFLAT_TYPE_BIT)
-		maxDimensions *= 32;
 
 	return maxDimensions;
 }
@@ -345,8 +338,6 @@ 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");
 }
@@ -383,7 +374,6 @@ 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 = 6;
+	amroutine->amsupport = 4;
 #if PG_VERSION_NUM >= 130000
 	amroutine->amoptsprocnum = 0;
 #endif

src/ivfflat.h

@@ -28,8 +28,6 @@
 #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
@@ -48,9 +46,7 @@
 typedef enum IvfflatType
 {
 	IVFFLAT_TYPE_VECTOR,
-	IVFFLAT_TYPE_HALFVEC,
-	IVFFLAT_TYPE_BIT,
-	IVFFLAT_TYPE_UNSUPPORTED
+	IVFFLAT_TYPE_HALFVEC
 }			IvfflatType;
 
 /* Build phases */
@@ -178,7 +174,6 @@ typedef struct IvfflatBuildState
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
 	FmgrInfo   *kmeansnormprocinfo;
-	FmgrInfo   *normalizeprocinfo;
 	Oid			collation;
 
 	/* Variables */
@@ -259,9 +254,7 @@ 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;
@@ -270,7 +263,7 @@ typedef struct IvfflatScanOpaqueData
 
 typedef IvfflatScanOpaqueData * IvfflatScanOpaque;
 
-#define VECTOR_ARRAY_SIZE(_length, _size) (sizeof(VectorArrayData) + (_length) * MAXALIGN(_size))
+#define VECTOR_ARRAY_SIZE(_length, _size) (sizeof(VectorArrayData) + (_length) * _size)
 #define VECTOR_ARRAY_OFFSET(_arr, _offset) ((char*) (_arr)->items + (_offset) * (_arr)->itemsize)
 #define VectorArrayGet(_arr, _offset) VECTOR_ARRAY_OFFSET(_arr, _offset)
 #define VectorArraySet(_arr, _offset, _val) memcpy(VECTOR_ARRAY_OFFSET(_arr, _offset), _val, (_arr)->itemsize)
@@ -281,8 +274,7 @@ void		VectorArrayFree(VectorArray arr);
 void		IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatType type);
 FmgrInfo   *IvfflatOptionalProcInfo(Relation index, uint16 procnum);
 IvfflatType IvfflatGetType(Relation index);
-Datum		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum value);
-bool		IvfflatCheckNorm(FmgrInfo *procinfo, Oid collation, Datum value);
+bool		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, IvfflatType type);
 int			IvfflatGetLists(Relation index);
 void		IvfflatGetMetaPageInfo(Relation index, int *lists, int *dimensions);
 void		IvfflatUpdateList(Relation index, ListInfo listInfo, BlockNumber insertPage, BlockNumber originalInsertPage, BlockNumber startPage, ForkNumber forkNum);

src/ivfinsert.c

@@ -85,12 +85,8 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
 	normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
 	if (normprocinfo != NULL)
 	{
-		Oid			collation = index->rd_indcollation[0];
-
-		if (!IvfflatCheckNorm(normprocinfo, collation, value))
+		if (!IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value, IvfflatGetType(index)))
 			return;
-
-		value = IvfflatNormValue(IvfflatOptionalProcInfo(index, IVFFLAT_NORMALIZE_PROC), collation, value);
 	}
 
 	/* Find the insert page - sets the page and list info */

src/ivfkmeans.c

@@ -3,12 +3,10 @@
 #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"
@@ -89,31 +87,33 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
 }
 
 /*
- * Norm centers
+ * Apply norm to vector
  */
-static void
-NormCenters(FmgrInfo *normalizeprocinfo, Oid collation, VectorArray centers)
+static inline void
+ApplyNorm(FmgrInfo *normprocinfo, Oid collation, Datum value, IvfflatType type)
 {
-	MemoryContext normCtx = AllocSetContextCreate(CurrentMemoryContext,
-												  "Ivfflat norm temporary context",
-												  ALLOCSET_DEFAULT_SIZES);
-	MemoryContext oldCtx = MemoryContextSwitchTo(normCtx);
+	double		norm = DatumGetFloat8(FunctionCall1Coll(normprocinfo, collation, value));
 
-	for (int j = 0; j < centers->length; j++)
+	/* TODO Handle zero norm */
+	if (norm > 0)
 	{
-		Datum		center = PointerGetDatum(VectorArrayGet(centers, j));
-		Datum		newCenter = IvfflatNormValue(normalizeprocinfo, collation, center);
-		Size		size = VARSIZE_ANY(DatumGetPointer(newCenter));
+		if (type == IVFFLAT_TYPE_VECTOR)
+		{
+			Vector	   *vec = DatumGetVector(value);
 
-		if (size > centers->itemsize)
-			elog(ERROR, "safety check failed");
+			for (int i = 0; i < vec->dim; i++)
+				vec->x[i] /= norm;
+		}
+		else if (type == IVFFLAT_TYPE_HALFVEC)
+		{
+			HalfVector *vec = DatumGetHalfVector(value);
 
-		memcpy(DatumGetPointer(center), DatumGetPointer(newCenter), size);
-		MemoryContextReset(normCtx);
+			for (int i = 0; i < vec->dim; i++)
+				vec->x[i] = Float4ToHalfUnchecked(HalfToFloat4(vec->x[i]) / norm);
+		}
+		else
+			elog(ERROR, "Unsupported type");
 	}
-
-	MemoryContextSwitchTo(oldCtx);
-	MemoryContextDelete(normCtx);
 }
 
 /*
@@ -134,31 +134,6 @@ 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));
-}
-
-/*
- * Sort vector array
- */
-static void
-SortVectorArray(VectorArray arr, IvfflatType type)
-{
-	if (type == IVFFLAT_TYPE_VECTOR)
-		qsort(arr->items, arr->length, arr->itemsize, CompareVectors);
-	else if (type == IVFFLAT_TYPE_HALFVEC)
-		qsort(arr->items, arr->length, arr->itemsize, CompareHalfVectors);
-	else if (type == IVFFLAT_TYPE_BIT)
-		qsort(arr->items, arr->length, arr->itemsize, CompareBitVectors);
-	else
-		elog(ERROR, "Unsupported type");
-}
-
 /*
  * Quick approach if we have little data
  */
@@ -168,12 +143,16 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers, IvfflatTy
 	int			dimensions = centers->dim;
 	Oid			collation = index->rd_indcollation[0];
 	FmgrInfo   *normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
-	FmgrInfo   *normalizeprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORMALIZE_PROC);
 
 	/* Copy existing vectors while avoiding duplicates */
 	if (samples->length > 0)
 	{
-		SortVectorArray(samples, type);
+		if (type == IVFFLAT_TYPE_VECTOR)
+			qsort(samples->items, samples->length, samples->itemsize, CompareVectors);
+		else if (type == IVFFLAT_TYPE_HALFVEC)
+			qsort(samples->items, samples->length, samples->itemsize, CompareHalfVectors);
+		else
+			elog(ERROR, "Unsupported type");
 
 		for (int i = 0; i < samples->length; i++)
 		{
@@ -212,25 +191,15 @@ 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");
 
+		/* Normalize if needed (only needed for random centers) */
+		if (normprocinfo != NULL)
+			ApplyNorm(normprocinfo, collation, center, type);
+
 		centers->length++;
 	}
-
-	/* Fine if existing vectors are normalized twice */
-	if (normprocinfo != NULL)
-		NormCenters(normalizeprocinfo, collation, centers);
 }
 
 #ifdef IVFFLAT_MEMORY
@@ -251,14 +220,27 @@ ShowMemoryUsage(MemoryContext context, Size estimatedSize)
 #endif
 
 /*
- * Sum centers
+ * Compute new centers
  */
 static void
-SumCenters(VectorArray samples, VectorArray aggCenters, int *closestCenters, IvfflatType type)
+ComputeNewCenters(VectorArray samples, VectorArray aggCenters, VectorArray newCenters, int *centerCounts, int *closestCenters, FmgrInfo *normprocinfo, Oid collation, IvfflatType type)
 {
 	int			dimensions = aggCenters->dim;
+	int			numCenters = aggCenters->maxlen;
 	int			numSamples = samples->length;
 
+	/* Reset sum and count */
+	for (int j = 0; j < numCenters; j++)
+	{
+		Vector	   *vec = (Vector *) VectorArrayGet(aggCenters, j);
+
+		for (int k = 0; k < dimensions; k++)
+			vec->x[k] = 0.0;
+
+		centerCounts[j] = 0;
+	}
+
+	/* Increment sum of closest center */
 	if (type == IVFFLAT_TYPE_VECTOR)
 	{
 		for (int j = 0; j < numSamples; j++)
@@ -281,81 +263,8 @@ SumCenters(VectorArray samples, VectorArray aggCenters, int *closestCenters, Ivf
 				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");
-}
-
-/*
- * Set new centers
- */
-static void
-SetNewCenters(VectorArray aggCenters, VectorArray newCenters, IvfflatType type)
-{
-	int			dimensions = aggCenters->dim;
-	int			numCenters = aggCenters->maxlen;
-
-	if (type == IVFFLAT_TYPE_HALFVEC)
-	{
-		for (int j = 0; j < numCenters; j++)
-		{
-			Vector	   *aggCenter = (Vector *) VectorArrayGet(aggCenters, j);
-			HalfVector *newCenter = (HalfVector *) VectorArrayGet(newCenters, j);
-
-			for (int k = 0; k < dimensions; k++)
-				newCenter->x[k] = Float4ToHalfUnchecked(aggCenter->x[k]);
-		}
-	}
-	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));
-		}
-	}
-}
-
-/*
- * Compute new centers
- */
-static void
-ComputeNewCenters(VectorArray samples, VectorArray aggCenters, VectorArray newCenters, int *centerCounts, int *closestCenters, FmgrInfo *normprocinfo, FmgrInfo *normalizeprocinfo, Oid collation, IvfflatType type)
-{
-	int			dimensions = aggCenters->dim;
-	int			numCenters = aggCenters->maxlen;
-	int			numSamples = samples->length;
-
-	/* Reset sum and count */
-	for (int j = 0; j < numCenters; j++)
-	{
-		Vector	   *vec = (Vector *) VectorArrayGet(aggCenters, j);
-
-		for (int k = 0; k < dimensions; k++)
-			vec->x[k] = 0.0;
-
-		centerCounts[j] = 0;
-	}
-
-	/* Increment sum of closest center */
-	SumCenters(samples, aggCenters, closestCenters, type);
 
 	/* Increment count of closest center */
 	for (int j = 0; j < numSamples; j++)
@@ -388,11 +297,28 @@ ComputeNewCenters(VectorArray samples, VectorArray aggCenters, VectorArray newCe
 	}
 
 	/* Set new centers if different from agg centers */
-	SetNewCenters(aggCenters, newCenters, type);
+	if (type == IVFFLAT_TYPE_HALFVEC)
+	{
+		for (int j = 0; j < numCenters; j++)
+		{
+			Vector	   *aggCenter = (Vector *) VectorArrayGet(aggCenters, j);
+			HalfVector *newCenter = (HalfVector *) VectorArrayGet(newCenters, j);
+
+			for (int k = 0; k < dimensions; k++)
+				newCenter->x[k] = Float4ToHalfUnchecked(aggCenter->x[k]);
+		}
+	}
 
 	/* Normalize if needed */
 	if (normprocinfo != NULL)
-		NormCenters(normalizeprocinfo, collation, newCenters);
+	{
+		for (int j = 0; j < numCenters; j++)
+		{
+			Datum		newCenter = PointerGetDatum(VectorArrayGet(newCenters, j));
+
+			ApplyNorm(normprocinfo, collation, newCenter, type);
+		}
+	}
 }
 
 /*
@@ -408,7 +334,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 {
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
-	FmgrInfo   *normalizeprocinfo;
 	Oid			collation;
 	int			dimensions = centers->dim;
 	int			numCenters = centers->maxlen;
@@ -456,7 +381,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 	/* Set support functions */
 	procinfo = index_getprocinfo(index, 1, IVFFLAT_KMEANS_DISTANCE_PROC);
 	normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
-	normalizeprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORMALIZE_PROC);
 	collation = index->rd_indcollation[0];
 
 	/* Use memory context */
@@ -476,8 +400,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 	newcdist = palloc(newcdistSize);
 
 	aggCenters = VectorArrayInit(numCenters, dimensions, VECTOR_SIZE(dimensions));
-	aggCenters->length = numCenters;
-
 	for (int j = 0; j < numCenters; j++)
 	{
 		Vector	   *vec = (Vector *) VectorArrayGet(aggCenters, j);
@@ -494,7 +416,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 	else if (type == IVFFLAT_TYPE_HALFVEC)
 	{
 		newCenters = VectorArrayInit(numCenters, dimensions, centers->itemsize);
-		newCenters->length = numCenters;
 
 		for (int j = 0; j < numCenters; j++)
 		{
@@ -504,19 +425,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 			vec->dim = dimensions;
 		}
 	}
-	else if (type == IVFFLAT_TYPE_BIT)
-	{
-		newCenters = VectorArrayInit(numCenters, dimensions, centers->itemsize);
-		newCenters->length = numCenters;
-
-		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");
 
@@ -658,7 +566,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 		}
 
 		/* Step 4: For each center c, let m(c) be mean of all points assigned */
-		ComputeNewCenters(samples, aggCenters, newCenters, centerCounts, closestCenters, normprocinfo, normalizeprocinfo, collation, type);
+		ComputeNewCenters(samples, aggCenters, newCenters, centerCounts, closestCenters, normprocinfo, collation, type);
 
 		/* Step 5 */
 		for (int j = 0; j < numCenters; j++)
@@ -734,20 +642,23 @@ CheckCenters(Relation index, VectorArray centers, IvfflatType type)
 					elog(ERROR, "Infinite value detected. Please report a bug.");
 			}
 		}
-		else if (type != IVFFLAT_TYPE_BIT)
+		else
 			elog(ERROR, "Unsupported type");
 	}
 
-	if (type != IVFFLAT_TYPE_BIT)
-	{
-		/* Ensure no duplicate centers */
-		SortVectorArray(centers, 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 (datumIsEqual(PointerGetDatum(VectorArrayGet(centers, i)), PointerGetDatum(VectorArrayGet(centers, i - 1)), false, -1))
-				elog(ERROR, "Duplicate centers detected. Please report a bug.");
-		}
+	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

@@ -5,6 +5,7 @@
 #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"
@@ -56,7 +57,7 @@ GetScanLists(IndexScanDesc scan, Datum value)
 			double		distance;
 
 			/* Use procinfo from the index instead of scan key for performance */
-			distance = DatumGetFloat8(so->distfunc(so->procinfo, so->collation, PointerGetDatum(&list->center), value));
+			distance = DatumGetFloat8(FunctionCall2Coll(so->procinfo, so->collation, PointerGetDatum(&list->center), value));
 
 			if (listCount < so->probes)
 			{
@@ -149,7 +150,7 @@ GetScanItems(IndexScanDesc scan, Datum value)
 				 * performance
 				 */
 				ExecClearTuple(slot);
-				slot->tts_values[0] = so->distfunc(so->procinfo, so->collation, datum, value);
+				slot->tts_values[0] = FunctionCall2Coll(so->procinfo, so->collation, datum, value);
 				slot->tts_isnull[0] = false;
 				slot->tts_values[1] = PointerGetDatum(&itup->t_tid);
 				slot->tts_isnull[1] = false;
@@ -177,15 +178,6 @@ 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
  */
@@ -197,21 +189,26 @@ GetScanValue(IndexScanDesc scan)
 
 	if (scan->orderByData->sk_flags & SK_ISNULL)
 	{
-		value = PointerGetDatum(NULL);
-		so->distfunc = ZeroDistance;
+		IvfflatType type = IvfflatGetType(scan->indexRelation);
+
+		if (type == IVFFLAT_TYPE_VECTOR)
+			value = PointerGetDatum(InitVector(so->dimensions));
+		else if (type == IVFFLAT_TYPE_HALFVEC)
+			value = PointerGetDatum(InitHalfVector(so->dimensions));
+		else
+			elog(ERROR, "Unsupported type");
 	}
 	else
 	{
 		value = scan->orderByData->sk_argument;
-		so->distfunc = FunctionCall2Coll;
 
 		/* Value should not be compressed or toasted */
 		Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
 		Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
 
-		/* Check normprocinfo since normalizeprocinfo not set for vector */
+		/* Fine if normalization fails */
 		if (so->normprocinfo != NULL)
-			value = IvfflatNormValue(so->normalizeprocinfo, so->collation, value);
+			IvfflatNormValue(so->normprocinfo, so->collation, &value, IvfflatGetType(scan->indexRelation));
 	}
 
 	return value;
@@ -249,7 +246,6 @@ 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

@@ -3,8 +3,12 @@
 #include "access/generic_xlog.h"
 #include "catalog/pg_type.h"
 #include "fmgr.h"
+#include "halfutils.h"
+#include "halfvec.h"
 #include "ivfflat.h"
 #include "storage/bufmgr.h"
+#include "vector.h"
+#include "utils/syscache.h"
 
 /*
  * Allocate a vector array
@@ -14,9 +18,6 @@ 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;
@@ -67,40 +68,57 @@ 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;
 
-	if (procinfo == NULL)
-		return IVFFLAT_TYPE_VECTOR;
+	tuple = SearchSysCache1(TYPEOID, ObjectIdGetDatum(typid));
+	if (!HeapTupleIsValid(tuple))
+		elog(ERROR, "cache lookup failed for type %u", typid);
 
-	result = (IvfflatType) DatumGetInt32(FunctionCall1(procinfo, ObjectIdGetDatum(typid)));
-
-	if (result == IVFFLAT_TYPE_UNSUPPORTED)
+	type = (Form_pg_type) GETSTRUCT(tuple);
+	if (strcmp(NameStr(type->typname), "vector") == 0)
+		result = IVFFLAT_TYPE_VECTOR;
+	else if (strcmp(NameStr(type->typname), "halfvec") == 0)
+		result = IVFFLAT_TYPE_HALFVEC;
+	else
+	{
+		ReleaseSysCache(tuple);
 		elog(ERROR, "type not supported for ivfflat index");
+	}
+
+	ReleaseSysCache(tuple);
 
 	return result;
 }
 
 /*
- * 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
+ * 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
  */
 bool
-IvfflatCheckNorm(FmgrInfo *procinfo, Oid collation, Datum value)
+IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, IvfflatType type)
 {
-	return DatumGetFloat8(FunctionCall1Coll(procinfo, collation, value)) > 0;
+	double		norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
+
+	if (norm > 0)
+	{
+		if (type == IVFFLAT_TYPE_VECTOR)
+			*value = DirectFunctionCall1(l2_normalize, *value);
+		else if (type == IVFFLAT_TYPE_HALFVEC)
+			*value = DirectFunctionCall1(halfvec_l2_normalize, *value);
+		else
+			elog(ERROR, "Unsupported type");
+
+		return true;
+	}
+
+	return false;
 }
 
 /*
@@ -245,22 +263,3 @@ 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

@@ -4,8 +4,6 @@
 #include <math.h>
 
 #include "fmgr.h"
-#include "halfutils.h"
-#include "halfvec.h"
 #include "libpq/pqformat.h"
 #include "sparsevec.h"
 #include "utils/array.h"
@@ -20,12 +18,6 @@
 #include "utils/builtins.h"
 #endif
 
-typedef struct SparseInputElement
-{
-	int32		index;
-	float		value;
-}			SparseInputElement;
-
 /*
  * Ensure same dimensions
  */
@@ -172,21 +164,6 @@ 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
  */
@@ -201,7 +178,8 @@ sparsevec_in(PG_FUNCTION_ARGS)
 	char	   *stringEnd;
 	SparseVector *result;
 	float	   *rvalues;
-	SparseInputElement *elements;
+	int32	   *indices;
+	float	   *values;
 	int			maxNnz;
 	int			nnz = 0;
 
@@ -219,7 +197,8 @@ sparsevec_in(PG_FUNCTION_ARGS)
 				(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
 				 errmsg("sparsevec cannot have more than %d non-zero elements", SPARSEVEC_MAX_NNZ)));
 
-	elements = palloc(maxNnz * sizeof(SparseInputElement));
+	indices = palloc(maxNnz * sizeof(int32));
+	values = palloc(maxNnz * sizeof(float));
 
 	pt = lit;
 
@@ -312,8 +291,8 @@ sparsevec_in(PG_FUNCTION_ARGS)
 			/* Do not store zero values */
 			if (value != 0)
 			{
-				elements[nnz].index = index;
-				elements[nnz].value = value;
+				indices[nnz] = index;
+				values[nnz] = value;
 				nnz++;
 			}
 
@@ -374,14 +353,12 @@ 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] = elements[i].index;
-		rvalues[i] = elements[i].value;
+		result->indices[i] = indices[i];
+		rvalues[i] = values[i];
 
 		CheckIndex(result->indices, i, dim);
 	}
@@ -527,10 +504,6 @@ 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);
@@ -618,49 +591,6 @@ 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
  */
@@ -683,7 +613,7 @@ SparsevecL2SquaredDistance(SparseVector * a, SparseVector * b)
 
 			if (ai == bi)
 			{
-				float		diff = ax[i] - bx[j];
+				double		diff = ax[i] - bx[j];
 
 				distance += diff * diff;
 			}
@@ -721,7 +651,7 @@ sparsevec_l2_distance(PG_FUNCTION_ARGS)
 
 	CheckDims(a, b);
 
-	PG_RETURN_FLOAT8(sqrt((double) SparsevecL2SquaredDistance(a, b)));
+	PG_RETURN_FLOAT8(sqrt(SparsevecL2SquaredDistance(a, b)));
 }
 
 /*
@@ -737,7 +667,7 @@ sparsevec_l2_squared_distance(PG_FUNCTION_ARGS)
 
 	CheckDims(a, b);
 
-	PG_RETURN_FLOAT8((double) SparsevecL2SquaredDistance(a, b));
+	PG_RETURN_FLOAT8(SparsevecL2SquaredDistance(a, b));
 }
 
 /*
@@ -788,7 +718,7 @@ sparsevec_inner_product(PG_FUNCTION_ARGS)
 
 	CheckDims(a, b);
 
-	PG_RETURN_FLOAT8((double) SparsevecInnerProduct(a, b));
+	PG_RETURN_FLOAT8(SparsevecInnerProduct(a, b));
 }
 
 /*
@@ -803,7 +733,7 @@ sparsevec_negative_inner_product(PG_FUNCTION_ARGS)
 
 	CheckDims(a, b);
 
-	PG_RETURN_FLOAT8((double) -SparsevecInnerProduct(a, b));
+	PG_RETURN_FLOAT8(-SparsevecInnerProduct(a, b));
 }
 
 /*
@@ -862,7 +792,7 @@ sparsevec_l1_distance(PG_FUNCTION_ARGS)
 	SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
 	float	   *ax = SPARSEVEC_VALUES(a);
 	float	   *bx = SPARSEVEC_VALUES(b);
-	float		distance = 0.0;
+	double		distance = 0.0;
 	int			bpos = 0;
 
 	CheckDims(a, b);
@@ -897,7 +827,7 @@ sparsevec_l1_distance(PG_FUNCTION_ARGS)
 	for (int j = bpos; j < b->nnz; j++)
 		distance += fabsf(bx[j]);
 
-	PG_RETURN_FLOAT8((double) distance);
+	PG_RETURN_FLOAT8(distance);
 }
 
 /*

src/sparsevec.h

@@ -3,38 +3,26 @@
 
 #include "fmgr.h"
 
-#define SPARSEVEC_MAX_DIM 1000000
+#define SPARSEVEC_MAX_DIM 100000
 #define SPARSEVEC_MAX_NNZ 16000
 
+/* Ensure values are aligned */
+#define SPARSEVEC_SIZE(_nnz)		(offsetof(SparseVector, indices) + MAXALIGN((_nnz) * sizeof(int32)) + (_nnz * sizeof(float)))
+#define SPARSEVEC_VALUES(x)		((float *) (((char *) (x)) + offsetof(SparseVector, indices) + MAXALIGN((x)->nnz * sizeof(int32))))
 #define DatumGetSparseVector(x)		((SparseVector *) PG_DETOAST_DATUM(x))
 #define PG_GETARG_SPARSEVEC_P(x)	DatumGetSparseVector(PG_GETARG_DATUM(x))
 #define PG_RETURN_SPARSEVEC_P(x)	PG_RETURN_POINTER(x)
 
-/* Indices are always sorted */
-/* Values come after indices */
 typedef struct SparseVector
 {
 	int32		vl_len_;		/* varlena header (do not touch directly!) */
 	int32		dim;			/* number of dimensions */
-	int32		nnz;			/* number of non-zero elements */
+	int32		nnz;
 	int32		unused;
 	int32		indices[FLEXIBLE_ARRAY_MEMBER];
 }			SparseVector;
 
-/* Use functions instead of macros to avoid double evaluation */
-
-static inline Size
-SPARSEVEC_SIZE(int nnz)
-{
-	return offsetof(SparseVector, indices) + (nnz * sizeof(int32)) + (nnz * sizeof(float));
-}
-
-static inline float *
-SPARSEVEC_VALUES(SparseVector * x)
-{
-	return (float *) (((char *) x) + offsetof(SparseVector, indices) + (x->nnz * sizeof(int32)));
-}
-
 SparseVector *InitSparseVector(int dim, int nnz);
+PGDLLEXPORT Datum sparsevec_l2_normalize(PG_FUNCTION_ARGS);
 
 #endif

src/vector.c

@@ -2,8 +2,7 @@
 
 #include <math.h>
 
-#include "bitutils.h"
-#include "bitvec.h"
+#include "bitvector.h"
 #include "catalog/pg_type.h"
 #include "common/shortest_dec.h"
 #include "fmgr.h"
@@ -34,10 +33,11 @@
 #define STATE_DIMS(x) (ARR_DIMS(x)[0] - 1)
 #define CreateStateDatums(dim) palloc(sizeof(Datum) * (dim + 1))
 
-#if defined(USE_TARGET_CLONES) && !defined(__FMA__)
-#define VECTOR_TARGET_CLONES __attribute__((target_clones("default", "fma")))
+/* target_clones requires glibc */
+#if defined(__x86_64__) && defined(__gnu_linux__) && defined(__has_attribute) && __has_attribute(target_clones) && !defined(__FMA__)
+#define VECTOR_DISPATCH __attribute__((target_clones("default", "fma")))
 #else
-#define VECTOR_TARGET_CLONES
+#define VECTOR_DISPATCH
 #endif
 
 PG_MODULE_MAGIC;
@@ -49,7 +49,6 @@ PGDLLEXPORT void _PG_init(void);
 void
 _PG_init(void)
 {
-	BitvecInit();
 	HalfvecInit();
 	HnswInit();
 	IvfflatInit();
@@ -565,7 +564,7 @@ halfvec_to_vector(PG_FUNCTION_ARGS)
 	PG_RETURN_POINTER(result);
 }
 
-VECTOR_TARGET_CLONES static float
+VECTOR_DISPATCH static float
 VectorL2SquaredDistance(int dim, float *ax, float *bx)
 {
 	float		distance = 0.0;
@@ -612,7 +611,7 @@ vector_l2_squared_distance(PG_FUNCTION_ARGS)
 	PG_RETURN_FLOAT8((double) VectorL2SquaredDistance(a->dim, a->x, b->x));
 }
 
-VECTOR_TARGET_CLONES static float
+VECTOR_DISPATCH static float
 VectorInnerProduct(int dim, float *ax, float *bx)
 {
 	float		distance = 0.0;
@@ -654,7 +653,7 @@ vector_negative_inner_product(PG_FUNCTION_ARGS)
 	PG_RETURN_FLOAT8((double) -VectorInnerProduct(a->dim, a->x, b->x));
 }
 
-VECTOR_TARGET_CLONES static double
+VECTOR_DISPATCH static double
 VectorCosineSimilarity(int dim, float *ax, float *bx)
 {
 	float		similarity = 0.0;
@@ -729,8 +728,7 @@ 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
+static float
 VectorL1Distance(int dim, float *ax, float *bx)
 {
 	float		distance = 0.0;

test/expected/bit_functions.out

@@ -28,24 +28,6 @@ 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 
 ------------------
@@ -104,24 +86,6 @@ 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/cast.out

@@ -170,44 +170,6 @@ 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

@@ -176,18 +176,6 @@ SELECT l2_norm('[0,1]'::halfvec);
        1
 (1 row)
 
-SELECT l2_norm('[0,0]'::halfvec);
- l2_norm 
----------
-       0
-(1 row)
-
-SELECT l2_norm('[2]'::halfvec);
- l2_norm 
----------
-       2
-(1 row)
-
 SELECT l2_distance('[0,0]'::halfvec, '[3,4]');
  l2_distance 
 -------------
@@ -316,24 +304,6 @@ SELECT l1_distance('[0,0]'::halfvec, '[0,1]');
 
 SELECT l1_distance('[1,2]'::halfvec, '[3]');
 ERROR:  different halfvec dimensions 2 and 1
-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      
 ------------------------

test/expected/hnsw_halfvec_l1.out

@@ -1,21 +0,0 @@
-SET enable_seqscan = off;
-CREATE TABLE t (val halfvec(3));
-INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
-CREATE INDEX ON t USING 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

@@ -1,21 +0,0 @@
-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

@@ -1,21 +0,0 @@
-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

@@ -1,36 +0,0 @@
-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

@@ -136,18 +136,6 @@ SELECT l2_norm('{1:3e37,2:4e37}/2'::sparsevec)::real;
    5e+37
 (1 row)
 
-SELECT l2_norm('{}/2'::sparsevec);
- l2_norm 
----------
-       0
-(1 row)
-
-SELECT l2_norm('{1:2}/1'::sparsevec);
- l2_norm 
----------
-       2
-(1 row)
-
 SELECT l2_distance('{}/2'::sparsevec, '{1:3,2:4}/2');
  l2_distance 
 -------------
@@ -166,12 +154,6 @@ SELECT l2_distance('{2:4}/2'::sparsevec, '{1:3}/2');
            5
 (1 row)
 
-SELECT l2_distance('{1:3,2:4}/2'::sparsevec, '{}/2');
- l2_distance 
--------------
-           5
-(1 row)
-
 SELECT l2_distance('{}/2'::sparsevec, '{2:1}/2');
  l2_distance 
 -------------
@@ -192,36 +174,6 @@ SELECT inner_product('{1:1,2:2}/2'::sparsevec, '{1:2,2:4}/2');
 
 SELECT inner_product('{1:1,2:2}/2'::sparsevec, '{1:3}/1');
 ERROR:  different sparsevec dimensions 2 and 1
-SELECT inner_product('{1:1,3:3}/4'::sparsevec, '{2:2,4:4}/4');
- inner_product 
----------------
-             0
-(1 row)
-
-SELECT inner_product('{2:2,4:4}/4'::sparsevec, '{1:1,3:3}/4');
- inner_product 
----------------
-             0
-(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 inner_product('{1:1}/2'::sparsevec, '{}/2');
- inner_product 
----------------
-             0
-(1 row)
-
-SELECT inner_product('{}/2'::sparsevec, '{1:1}/2');
- inner_product 
----------------
-             0
-(1 row)
-
 SELECT inner_product('{1:3e38}/1'::sparsevec, '{1:3e38}/1');
  inner_product 
 ---------------
@@ -276,12 +228,6 @@ SELECT cosine_distance('{1:2}/2'::sparsevec, '{2:2}/2');
                1
 (1 row)
 
-SELECT cosine_distance('{2:2}/2'::sparsevec, '{1:2}/2');
- cosine_distance 
------------------
-               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');
@@ -346,12 +292,6 @@ SELECT l1_distance('{1:1,3:3,5:5,7:7,9:9}/9'::sparsevec, '{2:2,4:4,6:6,8:8}/9');
           45
 (1 row)
 
-SELECT '{}/2'::sparsevec <+> '{1:3,2:4}/2';
- ?column? 
-----------
-        7
-(1 row)
-
 SELECT l2_normalize('{1:3,2:4}/2'::sparsevec);
   l2_normalize   
 -----------------

test/expected/sparsevec_input.out

@@ -164,18 +164,8 @@ SELECT '{1:0,2:1,3:0}/3'::sparsevec;
 (1 row)
 
 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;
+ERROR:  indexes must be in ascending order
+LINE 1: SELECT '{2:1,1:1}/2'::sparsevec;
                ^
 SELECT '{}/5'::sparsevec;
  sparsevec 
@@ -187,9 +177,9 @@ SELECT '{}/-1'::sparsevec;
 ERROR:  sparsevec must have at least 1 dimension
 LINE 1: SELECT '{}/-1'::sparsevec;
                ^
-SELECT '{}/1000001'::sparsevec;
-ERROR:  sparsevec cannot have more than 1000000 dimensions
-LINE 1: SELECT '{}/1000001'::sparsevec;
+SELECT '{}/100001'::sparsevec;
+ERROR:  sparsevec cannot have more than 100000 dimensions
+LINE 1: SELECT '{}/100001'::sparsevec;
                ^
 SELECT '{0:1}/1'::sparsevec;
 ERROR:  index "0" is out of range for type sparsevec
@@ -219,7 +209,7 @@ SELECT '{}/3'::sparsevec(0);
 ERROR:  dimensions for type sparsevec must be at least 1
 LINE 1: SELECT '{}/3'::sparsevec(0);
                        ^
-SELECT '{}/3'::sparsevec(1000001);
-ERROR:  dimensions for type sparsevec cannot exceed 1000000
-LINE 1: SELECT '{}/3'::sparsevec(1000001);
+SELECT '{}/3'::sparsevec(100001);
+ERROR:  dimensions for type sparsevec cannot exceed 100000
+LINE 1: SELECT '{}/3'::sparsevec(100001);
                        ^

test/expected/vector_functions.out

@@ -182,18 +182,6 @@ SELECT vector_norm('[3e37,4e37]')::real;
        5e+37
 (1 row)
 
-SELECT vector_norm('[0,0]');
- vector_norm 
--------------
-           0
-(1 row)
-
-SELECT vector_norm('[2]');
- vector_norm 
--------------
-           2
-(1 row)
-
 SELECT l2_distance('[0,0]'::vector, '[3,4]');
  l2_distance 
 -------------
@@ -214,12 +202,6 @@ 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? 
 ----------
@@ -240,12 +222,6 @@ 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? 
 ----------
@@ -302,18 +278,6 @@ 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? 
 ----------
@@ -340,24 +304,6 @@ SELECT l1_distance('[3e38]'::vector, '[-3e38]');
     Infinity
 (1 row)
 
-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 
 --------------

test/sql/bit_functions.sql

@@ -3,9 +3,6 @@ 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));
@@ -18,9 +15,6 @@ 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/cast.sql

@@ -47,17 +47,6 @@ 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

@@ -40,8 +40,6 @@ 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_norm('[0,0]'::halfvec);
-SELECT l2_norm('[2]'::halfvec);
 
 SELECT l2_distance('[0,0]'::halfvec, '[3,4]');
 SELECT l2_distance('[0,0]'::halfvec, '[0,1]');
@@ -70,9 +68,6 @@ 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 l2_normalize('[3,4]'::halfvec);
 SELECT l2_normalize('[3,0]'::halfvec);

test/sql/hnsw_halfvec_l1.sql

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

@@ -1,12 +0,0 @@
-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

@@ -1,12 +0,0 @@
-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

@@ -1,20 +0,0 @@
-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

@@ -23,23 +23,15 @@ 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_norm('{}/2'::sparsevec);
-SELECT l2_norm('{1:2}/1'::sparsevec);
 
 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('{1:3,2:4}/2'::sparsevec, '{}/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 inner_product('{1:1,2:2}/2'::sparsevec, '{1:3}/1');
-SELECT inner_product('{1:1,3:3}/4'::sparsevec, '{2:2,4:4}/4');
-SELECT inner_product('{2:2,4:4}/4'::sparsevec, '{1:1,3:3}/4');
-SELECT inner_product('{1:1,3:3,5:5}/5'::sparsevec, '{2:4,3:6,4:8}/5');
-SELECT inner_product('{1:1}/2'::sparsevec, '{}/2');
-SELECT inner_product('{}/2'::sparsevec, '{1:1}/2');
 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';
@@ -50,7 +42,6 @@ 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('{2:2}/2'::sparsevec, '{1:2}/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');
@@ -64,7 +55,6 @@ 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 '{}/2'::sparsevec <+> '{1:3,2:4}/2';
 
 SELECT l2_normalize('{1:3,2:4}/2'::sparsevec);
 SELECT l2_normalize('{1:3}/2'::sparsevec);

test/sql/sparsevec_input.sql

@@ -34,11 +34,9 @@ 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 '{}/1000001'::sparsevec;
+SELECT '{}/100001'::sparsevec;
 SELECT '{0:1}/1'::sparsevec;
 SELECT '{2:1}/1'::sparsevec;
 
@@ -47,4 +45,4 @@ SELECT '{}/3'::sparsevec(2);
 SELECT '{}/3'::sparsevec(3, 2);
 SELECT '{}/3'::sparsevec('a');
 SELECT '{}/3'::sparsevec(0);
-SELECT '{}/3'::sparsevec(1000001);
+SELECT '{}/3'::sparsevec(100001);

test/sql/vector_functions.sql

@@ -41,20 +41,16 @@ SELECT round(vector_norm('[1,1]')::numeric, 5);
 SELECT vector_norm('[3,4]');
 SELECT vector_norm('[0,1]');
 SELECT vector_norm('[3e37,4e37]')::real;
-SELECT vector_norm('[0,0]');
-SELECT vector_norm('[2]');
 
 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]');
@@ -66,17 +62,12 @@ 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 l2_normalize('[3,4]'::vector);
 SELECT l2_normalize('[3,0]'::vector);

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", "vector_l1_ops");
+my @operators = ("<->", "<#>", "<=>");
+my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_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", "vector_l1_ops");
+my @operators = ("<->", "<#>", "<=>");
+my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_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", "halfvec_l1_ops");
+my @operators = ("<->", "<#>", "<=>");
+my @opclasses = ("halfvec_l2_ops", "halfvec_ip_ops", "halfvec_cosine_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", "sparsevec_l1_ops");
+my @operators = ("<->", "<#>", "<=>");
+my @opclasses = ("sparsevec_l2_ops", "sparsevec_ip_ops", "sparsevec_cosine_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", "halfvec_l1_ops");
+my @operators = ("<->", "<#>", "<=>");
+my @opclasses = ("halfvec_l2_ops", "halfvec_ip_ops", "halfvec_cosine_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", "sparsevec_l1_ops");
+my @operators = ("<->", "<#>", "<=>");
+my @opclasses = ("sparsevec_l2_ops", "sparsevec_ip_ops", "sparsevec_cosine_ops");
 
 for my $i (0 .. $#operators)
 {

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, 0.34, $operator);
-		test_recall(10, 0.93, $operator);
+		test_recall(1, 0.35, $operator);
+		test_recall(10, 0.95, $operator);
 	}
 
 	# Test probes equals lists

test/t/035_ivfflat_bit_build_recall.pl

@@ -1,128 +0,0 @@
-use strict;
-use warnings;
-use PostgresNode;
-use TestLib;
-use Test::More;
-
-my $node;
-my @queries = ();
-my @expected;
-my $limit = 20;
-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();