Test sparsevec vacuum recall

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

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

.github/workflows/build.yml

@@ -1,122 +1,6 @@
 name: build
 on: [push, pull_request]
 jobs:
-  ubuntu:
-    runs-on: ${{ matrix.os }}
-    if: ${{ !startsWith(github.ref_name, 'mac') && !startsWith(github.ref_name, 'windows') }}
-    strategy:
-      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
-        with:
-          postgres-version: ${{ matrix.postgres }}
-          dev-files: true
-      - run: make
-        env:
-          PG_CFLAGS: -DUSE_ASSERT_CHECKING -Wall -Wextra -Werror -Wno-unused-parameter -Wno-sign-compare
-      - run: |
-          export PG_CONFIG=`which pg_config`
-          sudo --preserve-env=PG_CONFIG make install
-      - run: make installcheck
-      - if: ${{ failure() }}
-        run: cat regression.diffs
-      - run: |
-          sudo apt-get update
-          sudo apt-get install libipc-run-perl
-      - 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
@@ -128,4 +12,4 @@ jobs:
           check-ub: yes
       - run: make OPTFLAGS=""
       - run: sudo --preserve-env=PG_CONFIG make install
-      - run: make installcheck
+      - run: make prove_installcheck PROVE_TESTS=test/t/031_hnsw_sparsevec_vacuum_recall.pl

CHANGELOG.md

@@ -3,6 +3,7 @@
 - Added `halfvec` type
 - Added `sparsevec` type
 - Added support for indexing `bit` type
+- Added support for indexing L1 distance with HNSW
 - Added `binary_quantize` function
 - Added `hamming_distance` function
 - Added `jaccard_distance` function

Makefile

@@ -3,7 +3,7 @@ EXTVERSION = 0.6.2
 
 MODULE_big = vector
 DATA = $(wildcard sql/*--*.sql)
-OBJS = src/bitvector.o src/halfutils.o src/halfvec.o src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/sparsevec.o src/vector.o
+OBJS = src/bitutils.o src/bitvec.o src/halfutils.o src/halfvec.o src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/sparsevec.o src/vector.o
 HEADERS = src/halfvec.h src/sparsevec.h src/vector.h
 
 TESTS = $(wildcard test/sql/*.sql)

Makefile.win

@@ -1,7 +1,7 @@
 EXTENSION = vector
 EXTVERSION = 0.6.2
 
-OBJS = src\bitvector.obj src\halfutils.obj src\halfvec.obj src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\sparsevec.obj src\vector.obj
+OBJS = src\bitutils.obj src\bitvec.obj src\halfutils.obj src\halfvec.obj src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\sparsevec.obj src\vector.obj
 HEADERS = src\halfvec.h src\sparsevec.h src\vector.h
 
 REGRESS = bit_functions btree_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 (`<#>`) and cosine distance (`<=>`)
+Also supports inner product (`<#>`), cosine distance (`<=>`), and L1 distance (`<+>`, unreleased)
 
 Note: `<#>` returns the negative inner product since Postgres only supports `ASC` order index scans on operators
 
@@ -143,6 +143,7 @@ Supported distance functions are:
 - `<->` - L2 distance
 - `<#>` - (negative) inner product
 - `<=>` - cosine distance
+- `<+>` - L1 distance (unreleased)
 
 Get the nearest neighbors to a row
 
@@ -227,6 +228,12 @@ Cosine distance
 CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops);
 ```
 
+L1 distance - unreleased
+
+```sql
+CREATE INDEX ON items USING hnsw (embedding vector_l1_ops);
+```
+
 Hamming distance - unreleased
 
 ```sql
@@ -349,6 +356,12 @@ 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
@@ -535,7 +548,7 @@ SELECT id, content FROM items, plainto_tsquery('hello search') query
 
 You can use [Reciprocal Rank Fusion](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search_rrf.py) or a [cross-encoder](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search.py) to combine results.
 
-## Subvector Indexing
+## Indexing Subvectors
 
 *Unreleased*
 
@@ -855,6 +868,7 @@ Operator | Description | Added
 <-> | Euclidean distance |
 <#> | negative inner product |
 <=> | cosine distance |
+<+> | taxicab distance | unreleased
 
 ### Vector Functions
 
@@ -892,6 +906,7 @@ Operator | Description | Added
 <-> | Euclidean distance | unreleased
 <#> | negative inner product | unreleased
 <=> | cosine distance | unreleased
+<+> | taxicab distance | unreleased
 
 ### Halfvec Functions
 
@@ -943,6 +958,7 @@ 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

@@ -13,16 +13,32 @@ 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 = '<~>'
@@ -37,23 +53,20 @@ 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 3 hamming_distance(bit, bit),
-
+	FUNCTION 6 bit_ivfflat_support(internal);
-CREATE OPERATOR CLASS bit_jaccard_ops
-	FOR TYPE bit USING ivfflat AS
-	OPERATOR 1 <%> (bit, bit) FOR ORDER BY float_ops,
-	FUNCTION 1 jaccard_distance(bit, bit),
-	FUNCTION 3 jaccard_distance(bit, bit);
 
 CREATE OPERATOR CLASS bit_hamming_ops
 	FOR TYPE bit USING hnsw AS
 	OPERATOR 1 <~> (bit, bit) FOR ORDER BY float_ops,
-	FUNCTION 1 hamming_distance(bit, bit);
+	FUNCTION 1 hamming_distance(bit, bit),
+	FUNCTION 4 bit_hnsw_support(internal);
 
 CREATE OPERATOR CLASS bit_jaccard_ops
 	FOR TYPE bit USING hnsw AS
 	OPERATOR 1 <%> (bit, bit) FOR ORDER BY float_ops,
-	FUNCTION 1 jaccard_distance(bit, bit);
+	FUNCTION 1 jaccard_distance(bit, bit),
+	FUNCTION 4 bit_hnsw_support(internal);
 
 CREATE TYPE halfvec;
 
@@ -156,6 +169,12 @@ 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[],
@@ -235,6 +254,11 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
+CREATE OPERATOR <+> (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = l1_distance,
+	COMMUTATOR = '<+>'
+);
+
 CREATE OPERATOR + (
 	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_add,
 	COMMUTATOR = +
@@ -302,14 +326,17 @@ CREATE OPERATOR CLASS halfvec_l2_ops
 	FOR TYPE halfvec USING ivfflat AS
 	OPERATOR 1 <-> (halfvec, halfvec) FOR ORDER BY float_ops,
 	FUNCTION 1 halfvec_l2_squared_distance(halfvec, halfvec),
-	FUNCTION 3 l2_distance(halfvec, halfvec);
+	FUNCTION 3 l2_distance(halfvec, halfvec),
+	FUNCTION 6 halfvec_ivfflat_support(internal);
 
 CREATE OPERATOR CLASS halfvec_ip_ops
 	FOR TYPE halfvec USING ivfflat AS
 	OPERATOR 1 <#> (halfvec, halfvec) FOR ORDER BY float_ops,
 	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
 	FUNCTION 3 halfvec_spherical_distance(halfvec, halfvec),
-	FUNCTION 4 l2_norm(halfvec);
+	FUNCTION 4 l2_norm(halfvec),
+	FUNCTION 5 l2_normalize(halfvec),
+	FUNCTION 6 halfvec_ivfflat_support(internal);
 
 CREATE OPERATOR CLASS halfvec_cosine_ops
 	FOR TYPE halfvec USING ivfflat AS
@@ -317,23 +344,35 @@ 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 4 l2_norm(halfvec),
+	FUNCTION 5 l2_normalize(halfvec),
+	FUNCTION 6 halfvec_ivfflat_support(internal);
 
 CREATE OPERATOR CLASS halfvec_l2_ops
 	FOR TYPE halfvec USING hnsw AS
 	OPERATOR 1 <-> (halfvec, halfvec) FOR ORDER BY float_ops,
-	FUNCTION 1 halfvec_l2_squared_distance(halfvec, halfvec);
+	FUNCTION 1 halfvec_l2_squared_distance(halfvec, halfvec),
+	FUNCTION 4 halfvec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS halfvec_ip_ops
 	FOR TYPE halfvec USING hnsw AS
 	OPERATOR 1 <#> (halfvec, halfvec) FOR ORDER BY float_ops,
-	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec);
+	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
+	FUNCTION 4 halfvec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS halfvec_cosine_ops
 	FOR TYPE halfvec USING hnsw AS
 	OPERATOR 1 <=> (halfvec, halfvec) FOR ORDER BY float_ops,
 	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
-	FUNCTION 2 l2_norm(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);
 
 CREATE TYPE sparsevec;
 
@@ -415,6 +454,15 @@ 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;
 
@@ -424,6 +472,12 @@ 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 = '<->'
@@ -439,6 +493,11 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
+CREATE OPERATOR <+> (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = l1_distance,
+	COMMUTATOR = '<+>'
+);
+
 CREATE OPERATOR < (
 	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_lt,
 	COMMUTATOR = > , NEGATOR = >= ,
@@ -487,15 +546,25 @@ 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 1 sparsevec_l2_squared_distance(sparsevec, sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS sparsevec_ip_ops
 	FOR TYPE sparsevec USING hnsw AS
 	OPERATOR 1 <#> (sparsevec, sparsevec) FOR ORDER BY float_ops,
-	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec);
+	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS sparsevec_cosine_ops
 	FOR TYPE sparsevec USING hnsw AS
 	OPERATOR 1 <=> (sparsevec, sparsevec) FOR ORDER BY float_ops,
 	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec),
-	FUNCTION 2 l2_norm(sparsevec);
+	FUNCTION 2 l2_norm(sparsevec),
+	FUNCTION 3 l2_normalize(sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);
+
+CREATE OPERATOR CLASS sparsevec_l1_ops
+	FOR TYPE sparsevec USING hnsw AS
+	OPERATOR 1 <+> (sparsevec, sparsevec) FOR ORDER BY float_ops,
+	FUNCTION 1 l1_distance(sparsevec, sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);

sql/vector.sql

@@ -186,6 +186,11 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
+CREATE OPERATOR <+> (
+	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = l1_distance,
+	COMMUTATOR = '<+>'
+);
+
 CREATE OPERATOR + (
 	LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_add,
 	COMMUTATOR = +
@@ -304,6 +309,11 @@ CREATE OPERATOR CLASS vector_cosine_ops
 	FUNCTION 1 vector_negative_inner_product(vector, vector),
 	FUNCTION 2 vector_norm(vector);
 
+CREATE OPERATOR CLASS vector_l1_ops
+	FOR TYPE vector USING hnsw AS
+	OPERATOR 1 <+> (vector, vector) FOR ORDER BY float_ops,
+	FUNCTION 1 l1_distance(vector, vector);
+
 -- bit functions
 
 CREATE FUNCTION hamming_distance(bit, bit) RETURNS float8
@@ -312,6 +322,14 @@ 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 <~> (
@@ -330,23 +348,20 @@ 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 3 hamming_distance(bit, bit),
-
+	FUNCTION 6 bit_ivfflat_support(internal);
-CREATE OPERATOR CLASS bit_jaccard_ops
-	FOR TYPE bit USING ivfflat AS
-	OPERATOR 1 <%> (bit, bit) FOR ORDER BY float_ops,
-	FUNCTION 1 jaccard_distance(bit, bit),
-	FUNCTION 3 jaccard_distance(bit, bit);
 
 CREATE OPERATOR CLASS bit_hamming_ops
 	FOR TYPE bit USING hnsw AS
 	OPERATOR 1 <~> (bit, bit) FOR ORDER BY float_ops,
-	FUNCTION 1 hamming_distance(bit, bit);
+	FUNCTION 1 hamming_distance(bit, bit),
+	FUNCTION 4 bit_hnsw_support(internal);
 
 CREATE OPERATOR CLASS bit_jaccard_ops
 	FOR TYPE bit USING hnsw AS
 	OPERATOR 1 <%> (bit, bit) FOR ORDER BY float_ops,
-	FUNCTION 1 jaccard_distance(bit, bit);
+	FUNCTION 1 jaccard_distance(bit, bit),
+	FUNCTION 4 bit_hnsw_support(internal);
 
 -- halfvec type
 
@@ -455,6 +470,12 @@ 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) (
@@ -542,6 +563,11 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
+CREATE OPERATOR <+> (
+	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = l1_distance,
+	COMMUTATOR = '<+>'
+);
+
 CREATE OPERATOR + (
 	LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_add,
 	COMMUTATOR = +
@@ -611,14 +637,17 @@ CREATE OPERATOR CLASS halfvec_l2_ops
 	FOR TYPE halfvec USING ivfflat AS
 	OPERATOR 1 <-> (halfvec, halfvec) FOR ORDER BY float_ops,
 	FUNCTION 1 halfvec_l2_squared_distance(halfvec, halfvec),
-	FUNCTION 3 l2_distance(halfvec, halfvec);
+	FUNCTION 3 l2_distance(halfvec, halfvec),
+	FUNCTION 6 halfvec_ivfflat_support(internal);
 
 CREATE OPERATOR CLASS halfvec_ip_ops
 	FOR TYPE halfvec USING ivfflat AS
 	OPERATOR 1 <#> (halfvec, halfvec) FOR ORDER BY float_ops,
 	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
 	FUNCTION 3 halfvec_spherical_distance(halfvec, halfvec),
-	FUNCTION 4 l2_norm(halfvec);
+	FUNCTION 4 l2_norm(halfvec),
+	FUNCTION 5 l2_normalize(halfvec),
+	FUNCTION 6 halfvec_ivfflat_support(internal);
 
 CREATE OPERATOR CLASS halfvec_cosine_ops
 	FOR TYPE halfvec USING ivfflat AS
@@ -626,23 +655,35 @@ 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 4 l2_norm(halfvec),
+	FUNCTION 5 l2_normalize(halfvec),
+	FUNCTION 6 halfvec_ivfflat_support(internal);
 
 CREATE OPERATOR CLASS halfvec_l2_ops
 	FOR TYPE halfvec USING hnsw AS
 	OPERATOR 1 <-> (halfvec, halfvec) FOR ORDER BY float_ops,
-	FUNCTION 1 halfvec_l2_squared_distance(halfvec, halfvec);
+	FUNCTION 1 halfvec_l2_squared_distance(halfvec, halfvec),
+	FUNCTION 4 halfvec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS halfvec_ip_ops
 	FOR TYPE halfvec USING hnsw AS
 	OPERATOR 1 <#> (halfvec, halfvec) FOR ORDER BY float_ops,
-	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec);
+	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
+	FUNCTION 4 halfvec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS halfvec_cosine_ops
 	FOR TYPE halfvec USING hnsw AS
 	OPERATOR 1 <=> (halfvec, halfvec) FOR ORDER BY float_ops,
 	FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
-	FUNCTION 2 l2_norm(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);
 
 --- sparsevec type
 
@@ -732,6 +773,15 @@ 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)
@@ -743,6 +793,12 @@ 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 <-> (
@@ -760,6 +816,11 @@ CREATE OPERATOR <=> (
 	COMMUTATOR = '<=>'
 );
 
+CREATE OPERATOR <+> (
+	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = l1_distance,
+	COMMUTATOR = '<+>'
+);
+
 CREATE OPERATOR < (
 	LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_lt,
 	COMMUTATOR = > , NEGATOR = >= ,
@@ -810,15 +871,25 @@ 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 1 sparsevec_l2_squared_distance(sparsevec, sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS sparsevec_ip_ops
 	FOR TYPE sparsevec USING hnsw AS
 	OPERATOR 1 <#> (sparsevec, sparsevec) FOR ORDER BY float_ops,
-	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec);
+	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);
 
 CREATE OPERATOR CLASS sparsevec_cosine_ops
 	FOR TYPE sparsevec USING hnsw AS
 	OPERATOR 1 <=> (sparsevec, sparsevec) FOR ORDER BY float_ops,
 	FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec),
-	FUNCTION 2 l2_norm(sparsevec);
+	FUNCTION 2 l2_norm(sparsevec),
+	FUNCTION 3 l2_normalize(sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);
+
+CREATE OPERATOR CLASS sparsevec_l1_ops
+	FOR TYPE sparsevec USING hnsw AS
+	OPERATOR 1 <+> (sparsevec, sparsevec) FOR ORDER BY float_ops,
+	FUNCTION 1 l1_distance(sparsevec, sparsevec),
+	FUNCTION 4 sparsevec_hnsw_support(internal);

src/bitutils.c

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

src/bitutils.h

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

src/bitvec.c

@@ -1,7 +1,7 @@
 #include "postgres.h"
 
-#include "bitvector.h"
+#include "bitutils.h"
-#include "port/pg_bitutils.h"
+#include "bitvec.h"
 #include "utils/varbit.h"
 
 #if PG_VERSION_NUM >= 160000
@@ -46,17 +46,10 @@ hamming_distance(PG_FUNCTION_ARGS)
 {
 	VarBit	   *a = PG_GETARG_VARBIT_P(0);
 	VarBit	   *b = PG_GETARG_VARBIT_P(1);
-	unsigned char *ax = VARBITS(a);
-	unsigned char *bx = VARBITS(b);
-	uint64		distance = 0;
 
 	CheckDims(a, b);
 
-	/* TODO Improve performance */
+	PG_RETURN_FLOAT8((double) BitHammingDistance(VARBITBYTES(a), VARBITS(a), VARBITS(b), 0));
-	for (uint32 i = 0; i < VARBITBYTES(a); i++)
-		distance += pg_number_of_ones[ax[i] ^ bx[i]];
-
-	PG_RETURN_FLOAT8((double) distance);
 }
 
 /*
@@ -68,23 +61,8 @@ jaccard_distance(PG_FUNCTION_ARGS)
 {
 	VarBit	   *a = PG_GETARG_VARBIT_P(0);
 	VarBit	   *b = PG_GETARG_VARBIT_P(1);
-	unsigned char *ax = VARBITS(a);
-	unsigned char *bx = VARBITS(b);
-	uint64		ab = 0;
-	uint64		aa;
-	uint64		bb;
 
 	CheckDims(a, b);
 
-	/* TODO Improve performance */
+	PG_RETURN_FLOAT8(BitJaccardDistance(VARBITBYTES(a), VARBITS(a), VARBITS(b), 0, 0, 0));
-	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 +8,21 @@
 
 #if defined(HAVE__GET_CPUID)
 #include <cpuid.h>
-#elif defined(HAVE__CPUID)
+#else
 #include <intrin.h>
 #endif
 
 #ifdef _MSC_VER
-#define TARGET_F16C_FMA
+#define TARGET_F16C
 #else
-#define TARGET_F16C_FMA __attribute__((target("f16c,fma")))
+#define TARGET_F16C __attribute__((target("avx,f16c,fma")))
 #endif
 #endif
 
-
 float		(*HalfvecL2SquaredDistance) (int dim, half * ax, half * bx);
 float		(*HalfvecInnerProduct) (int dim, half * ax, half * bx);
 double		(*HalfvecCosineSimilarity) (int dim, half * ax, half * bx);
+float		(*HalfvecL1Distance) (int dim, half * ax, half * bx);
 
 static float
 HalfvecL2SquaredDistanceDefault(int dim, half * ax, half * bx)
@@ -41,8 +41,8 @@ HalfvecL2SquaredDistanceDefault(int dim, half * ax, half * bx)
 }
 
 #ifdef HALFVEC_DISPATCH
-TARGET_F16C_FMA static float
+TARGET_F16C static float
-HalfvecL2SquaredDistanceF16cFma(int dim, half * ax, half * bx)
+HalfvecL2SquaredDistanceF16c(int dim, half * ax, half * bx)
 {
 	float		distance;
 	int			i;
@@ -89,8 +89,8 @@ HalfvecInnerProductDefault(int dim, half * ax, half * bx)
 }
 
 #ifdef HALFVEC_DISPATCH
-TARGET_F16C_FMA static float
+TARGET_F16C static float
-HalfvecInnerProductF16cFma(int dim, half * ax, half * bx)
+HalfvecInnerProductF16c(int dim, half * ax, half * bx)
 {
 	float		distance;
 	int			i;
@@ -142,8 +142,8 @@ HalfvecCosineSimilarityDefault(int dim, half * ax, half * bx)
 }
 
 #ifdef HALFVEC_DISPATCH
-TARGET_F16C_FMA static double
+TARGET_F16C static double
-HalfvecCosineSimilarityF16cFma(int dim, half * ax, half * bx)
+HalfvecCosineSimilarityF16c(int dim, half * ax, half * bx)
 {
 	float		similarity;
 	float		norma;
@@ -192,6 +192,51 @@ HalfvecCosineSimilarityF16cFma(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);
+
+	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)
@@ -211,16 +256,19 @@ SupportsCpuFeature(unsigned int feature)
 
 #if defined(HAVE__GET_CPUID)
 	__get_cpuid(1, &exx[0], &exx[1], &exx[2], &exx[3]);
-#elif defined(HAVE__CPUID)
+#else
 	__cpuid(exx, 1);
 #endif
 
+	/* Check OS supports XSAVE */
 	if ((exx[2] & CPU_FEATURE_OSXSAVE) != CPU_FEATURE_OSXSAVE)
 		return false;
 
+	/* Check XMM and YMM registers are enabled */
 	if ((_xgetbv(0) & 6) != 6)
 		return false;
 
+	/* Now check features */
 	return (exx[2] & feature) == feature;
 }
 #endif
@@ -235,13 +283,16 @@ HalfvecInit(void)
 	HalfvecL2SquaredDistance = HalfvecL2SquaredDistanceDefault;
 	HalfvecInnerProduct = HalfvecInnerProductDefault;
 	HalfvecCosineSimilarity = HalfvecCosineSimilarityDefault;
+	HalfvecL1Distance = HalfvecL1DistanceDefault;
 
 #ifdef HALFVEC_DISPATCH
-	if (SupportsCpuFeature(CPU_FEATURE_AVX | CPU_FEATURE_FMA | CPU_FEATURE_F16C))
+	if (SupportsCpuFeature(CPU_FEATURE_AVX | CPU_FEATURE_F16C | CPU_FEATURE_FMA))
 	{
-		HalfvecL2SquaredDistance = HalfvecL2SquaredDistanceF16cFma;
+		HalfvecL2SquaredDistance = HalfvecL2SquaredDistanceF16c;
-		HalfvecInnerProduct = HalfvecInnerProductF16cFma;
+		HalfvecInnerProduct = HalfvecInnerProductF16c;
-		HalfvecCosineSimilarity = HalfvecCosineSimilarityF16cFma;
+		HalfvecCosineSimilarity = HalfvecCosineSimilarityF16c;
+		/* Does not require FMA, but keep logic simple */
+		HalfvecL1Distance = HalfvecL1DistanceF16c;
 	}
 #endif
 }

src/halfutils.h

@@ -13,6 +13,7 @@
 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 "bitvector.h"
+#include "bitvec.h"
 #include "catalog/pg_type.h"
 #include "common/shortest_dec.h"
 #include "fmgr.h"
@@ -11,6 +11,7 @@
 #include "lib/stringinfo.h"
 #include "libpq/pqformat.h"
 #include "port.h"				/* for strtof() */
+#include "sparsevec.h"
 #include "utils/array.h"
 #include "utils/builtins.h"
 #include "utils/float.h"
@@ -682,17 +683,10 @@ halfvec_l1_distance(PG_FUNCTION_ARGS)
 {
 	HalfVector *a = PG_GETARG_HALFVEC_P(0);
 	HalfVector *b = PG_GETARG_HALFVEC_P(1);
-	half	   *ax = a->x;
-	half	   *bx = b->x;
-	float		distance = 0.0;
 
 	CheckDims(a, b);
 
-	/* Auto-vectorized */
+	PG_RETURN_FLOAT8((double) HalfvecL1Distance(a->dim, a->x, b->x));
-	for (int i = 0; i < a->dim; i++)
-		distance += fabsf(HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]));
-
-	PG_RETURN_FLOAT8((double) distance);
 }
 
 /*
@@ -1174,3 +1168,26 @@ 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

@@ -5,10 +5,11 @@
 
 #include <float.h>
 
+#include "bitutils.h"
 #include "fmgr.h"
 #include "vector.h"
 
-#if defined(__x86_64__) || defined(_M_AMD64)
+#if defined(USE_DISPATCH)
 #define HALFVEC_DISPATCH
 #endif
 
@@ -44,6 +45,5 @@ 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 = 2;
+	amroutine->amsupport = 4;
 #if PG_VERSION_NUM >= 130000
 	amroutine->amoptsprocnum = 0;
 #endif

src/hnsw.h

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

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(value, buildstate->type);
+		value = HnswNormValue(buildstate->normalizeprocinfo, buildstate->collation, value);
 	}
 
 	/* Get datum size */
@@ -725,6 +725,7 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
 	/* Get support functions */
 	buildstate->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
 	buildstate->normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
+	buildstate->normalizeprocinfo = HnswOptionalProcInfo(index, HNSW_NORMALIZE_PROC);
 	buildstate->collation = index->rd_indcollation[0];
 
 	InitGraph(&buildstate->graphData, NULL, maintenance_work_mem * 1024L);

src/hnswinsert.c

@@ -614,13 +614,12 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 	Datum		value;
 	FmgrInfo   *normprocinfo;
 	Oid			collation = index->rd_indcollation[0];
-	HnswType	type = HnswGetType(index);
 
 	/* Detoast once for all calls */
 	value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
 
 	/* Check value */
-	HnswCheckValue(value, type);
+	HnswCheckValue(value, HnswGetType(index));
 
 	/* Normalize if needed */
 	normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
@@ -629,7 +628,7 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
 		if (!HnswCheckNorm(normprocinfo, collation, value))
 			return;
 
-		value = HnswNormValue(value, type);
+		value = HnswNormValue(HnswOptionalProcInfo(index, HNSW_NORMALIZE_PROC), collation, value);
 	}
 
 	HnswInsertTupleOnDisk(index, value, values, isnull, heap_tid, false);

src/hnswscan.c

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

src/hnswutils.c

@@ -6,8 +6,6 @@
 #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"
@@ -15,8 +13,6 @@
 #include "utils/datum.h"
 #include "utils/memdebug.h"
 #include "utils/rel.h"
-#include "utils/syscache.h"
-#include "vector.h"
 
 #if PG_VERSION_NUM >= 130000
 #include "common/hashfn.h"
@@ -162,32 +158,17 @@ HnswOptionalProcInfo(Relation index, uint16 procnum)
 HnswType
 HnswGetType(Relation index)
 {
+	FmgrInfo   *procinfo = HnswOptionalProcInfo(index, HNSW_TYPE_SUPPORT_PROC);
 	Oid			typid = TupleDescAttr(index->rd_att, 0)->atttypid;
-	HeapTuple	tuple;
-	Form_pg_type type;
 	HnswType	result;
 
-	if (typid == BITOID)
+	if (procinfo == NULL)
-		return HNSW_TYPE_BIT;
+		return HNSW_TYPE_VECTOR;
 
-	tuple = SearchSysCache1(TYPEOID, ObjectIdGetDatum(typid));
+	result = (HnswType) DatumGetInt32(FunctionCall1(procinfo, ObjectIdGetDatum(typid)));
-	if (!HeapTupleIsValid(tuple))
-		elog(ERROR, "cache lookup failed for type %u", typid);
 
-	type = (Form_pg_type) GETSTRUCT(tuple);
+	if (result == HNSW_TYPE_UNSUPPORTED)
-	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;
 }
@@ -196,17 +177,12 @@ HnswGetType(Relation index)
  * Normalize value
  */
 Datum
-HnswNormValue(Datum value, HnswType type)
+HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum value)
 {
-	/* TODO Remove type-specific code */
+	if (procinfo == NULL)
-	if (type == HNSW_TYPE_VECTOR)
 		return DirectFunctionCall1(l2_normalize, value);
-	else if (type == HNSW_TYPE_HALFVEC)
+
-		return DirectFunctionCall1(halfvec_l2_normalize, value);
+	return FunctionCall1Coll(procinfo, collation, value);
-	else if (type == HNSW_TYPE_SPARSEVEC)
-		return DirectFunctionCall1(sparsevec_l2_normalize, value);
-	else
-		elog(ERROR, "Unsupported type");
 }
 
 /*
@@ -1326,3 +1302,29 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
 		ep = w;
 	}
 }
+
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_hnsw_support);
+Datum
+halfvec_hnsw_support(PG_FUNCTION_ARGS)
+{
+	PG_RETURN_INT32(HNSW_TYPE_HALFVEC);
+};
+
+PGDLLEXPORT PG_FUNCTION_INFO_V1(bit_hnsw_support);
+Datum
+bit_hnsw_support(PG_FUNCTION_ARGS)
+{
+	Oid			typid = PG_GETARG_OID(0);
+
+	if (typid == BITOID)
+		PG_RETURN_INT32(HNSW_TYPE_BIT);
+	else
+		PG_RETURN_INT32(HNSW_TYPE_UNSUPPORTED);
+};
+
+PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_hnsw_support);
+Datum
+sparsevec_hnsw_support(PG_FUNCTION_ARGS)
+{
+	PG_RETURN_INT32(HNSW_TYPE_SPARSEVEC);
+};

src/ivfbuild.c

@@ -6,7 +6,7 @@
 #include "access/tableam.h"
 #include "access/parallel.h"
 #include "access/xact.h"
-#include "bitvector.h"
+#include "bitvec.h"
 #include "catalog/index.h"
 #include "catalog/pg_operator_d.h"
 #include "catalog/pg_type_d.h"
@@ -60,8 +60,10 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
 	 */
 	if (buildstate->kmeansnormprocinfo != NULL)
 	{
-		if (!IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->collation, &value, buildstate->type))
+		if (!IvfflatCheckNorm(buildstate->kmeansnormprocinfo, buildstate->collation, value))
 			return;
+
+		value = IvfflatNormValue(buildstate->normalizeprocinfo, buildstate->collation, value);
 	}
 
 	if (samples->length < targsamples)
@@ -156,8 +158,10 @@ AddTupleToSort(Relation index, ItemPointer tid, Datum *values, IvfflatBuildState
 	/* Normalize if needed */
 	if (buildstate->normprocinfo != NULL)
 	{
-		if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->type))
+		if (!IvfflatCheckNorm(buildstate->normprocinfo, buildstate->collation, value))
 			return;
+
+		value = IvfflatNormValue(buildstate->normalizeprocinfo, buildstate->collation, value);
 	}
 
 	/* Find the list that minimizes the distance */
@@ -379,6 +383,7 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
 	buildstate->procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
 	buildstate->normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
 	buildstate->kmeansnormprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
+	buildstate->normalizeprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORMALIZE_PROC);
 	buildstate->collation = index->rd_indcollation[0];
 
 	/* Require more than one dimension for spherical k-means */

src/ivfflat.c

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

src/ivfflat.h

@@ -28,6 +28,8 @@
 #define IVFFLAT_NORM_PROC 2
 #define IVFFLAT_KMEANS_DISTANCE_PROC 3
 #define IVFFLAT_KMEANS_NORM_PROC 4
+#define IVFFLAT_NORMALIZE_PROC 5
+#define IVFFLAT_TYPE_SUPPORT_PROC 6
 
 #define IVFFLAT_VERSION	1
 #define IVFFLAT_MAGIC_NUMBER 0x14FF1A7
@@ -47,7 +49,8 @@ typedef enum IvfflatType
 {
 	IVFFLAT_TYPE_VECTOR,
 	IVFFLAT_TYPE_HALFVEC,
-	IVFFLAT_TYPE_BIT
+	IVFFLAT_TYPE_BIT,
+	IVFFLAT_TYPE_UNSUPPORTED
 }			IvfflatType;
 
 /* Build phases */
@@ -175,6 +178,7 @@ typedef struct IvfflatBuildState
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
 	FmgrInfo   *kmeansnormprocinfo;
+	FmgrInfo   *normalizeprocinfo;
 	Oid			collation;
 
 	/* Variables */
@@ -255,7 +259,9 @@ typedef struct IvfflatScanOpaqueData
 	/* Support functions */
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
+	FmgrInfo   *normalizeprocinfo;
 	Oid			collation;
+	Datum		(*distfunc) (FmgrInfo *flinfo, Oid collation, Datum arg1, Datum arg2);
 
 	/* Lists */
 	pairingheap *listQueue;
@@ -275,7 +281,8 @@ void		VectorArrayFree(VectorArray arr);
 void		IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatType type);
 FmgrInfo   *IvfflatOptionalProcInfo(Relation index, uint16 procnum);
 IvfflatType IvfflatGetType(Relation index);
-bool		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, IvfflatType type);
+Datum		IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum value);
+bool		IvfflatCheckNorm(FmgrInfo *procinfo, Oid collation, Datum value);
 int			IvfflatGetLists(Relation index);
 void		IvfflatGetMetaPageInfo(Relation index, int *lists, int *dimensions);
 void		IvfflatUpdateList(Relation index, ListInfo listInfo, BlockNumber insertPage, BlockNumber originalInsertPage, BlockNumber startPage, ForkNumber forkNum);

src/ivfinsert.c

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

src/ivfkmeans.c

@@ -3,7 +3,7 @@
 #include <float.h>
 #include <math.h>
 
-#include "bitvector.h"
+#include "bitvec.h"
 #include "halfutils.h"
 #include "halfvec.h"
 #include "ivfflat.h"
@@ -89,33 +89,31 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
 }
 
 /*
- * Apply norm to vector
+ * Norm centers
  */
-static inline void
+static void
-ApplyNorm(FmgrInfo *normprocinfo, Oid collation, Datum value, IvfflatType type)
+NormCenters(FmgrInfo *normalizeprocinfo, Oid collation, VectorArray centers)
 {
-	double		norm = DatumGetFloat8(FunctionCall1Coll(normprocinfo, collation, value));
+	MemoryContext normCtx = AllocSetContextCreate(CurrentMemoryContext,
+												  "Ivfflat norm temporary context",
+												  ALLOCSET_DEFAULT_SIZES);
+	MemoryContext oldCtx = MemoryContextSwitchTo(normCtx);
 
-	/* TODO Handle zero norm */
+	for (int j = 0; j < centers->length; j++)
-	if (norm > 0)
 	{
-		if (type == IVFFLAT_TYPE_VECTOR)
+		Datum		center = PointerGetDatum(VectorArrayGet(centers, j));
-		{
+		Datum		newCenter = IvfflatNormValue(normalizeprocinfo, collation, center);
-			Vector	   *vec = DatumGetVector(value);
+		Size		size = VARSIZE_ANY(DatumGetPointer(newCenter));
 
-			for (int i = 0; i < vec->dim; i++)
+		if (size > centers->itemsize)
-				vec->x[i] /= norm;
+			elog(ERROR, "safety check failed");
-		}
-		else if (type == IVFFLAT_TYPE_HALFVEC)
-		{
-			HalfVector *vec = DatumGetHalfVector(value);
 
-			for (int i = 0; i < vec->dim; i++)
+		memcpy(DatumGetPointer(center), DatumGetPointer(newCenter), size);
-				vec->x[i] = Float4ToHalfUnchecked(HalfToFloat4(vec->x[i]) / norm);
+		MemoryContextReset(normCtx);
-		}
-		else
-			elog(ERROR, "Unsupported type");
 	}
+
+	MemoryContextSwitchTo(oldCtx);
+	MemoryContextDelete(normCtx);
 }
 
 /*
@@ -145,6 +143,22 @@ 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
  */
@@ -154,18 +168,12 @@ 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)
 	{
-		if (type == IVFFLAT_TYPE_VECTOR)
+		SortVectorArray(samples, type);
-			qsort(samples->items, samples->length, samples->itemsize, CompareVectors);
-		else if (type == IVFFLAT_TYPE_HALFVEC)
-			qsort(samples->items, samples->length, samples->itemsize, CompareHalfVectors);
-		else if (type == IVFFLAT_TYPE_BIT)
-			qsort(samples->items, samples->length, samples->itemsize, CompareBitVectors);
-		else
-			elog(ERROR, "Unsupported type");
 
 		for (int i = 0; i < samples->length; i++)
 		{
@@ -217,12 +225,12 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers, IvfflatTy
 		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
@@ -243,27 +251,14 @@ ShowMemoryUsage(MemoryContext context, Size estimatedSize)
 #endif
 
 /*
- * Compute new centers
+ * Sum centers
  */
 static void
-ComputeNewCenters(VectorArray samples, VectorArray aggCenters, VectorArray newCenters, int *centerCounts, int *closestCenters, FmgrInfo *normprocinfo, Oid collation, IvfflatType type)
+SumCenters(VectorArray samples, VectorArray aggCenters, int *closestCenters, 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++)
@@ -299,6 +294,68 @@ ComputeNewCenters(VectorArray samples, VectorArray aggCenters, VectorArray newCe
 	}
 	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++)
@@ -331,43 +388,11 @@ ComputeNewCenters(VectorArray samples, VectorArray aggCenters, VectorArray newCe
 	}
 
 	/* Set new centers if different from agg centers */
-	if (type == IVFFLAT_TYPE_HALFVEC)
+	SetNewCenters(aggCenters, newCenters, type);
-	{
-		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));
-		}
-	}
 
 	/* 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);
-		}
-	}
 }
 
 /*
@@ -383,6 +408,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 {
 	FmgrInfo   *procinfo;
 	FmgrInfo   *normprocinfo;
+	FmgrInfo   *normalizeprocinfo;
 	Oid			collation;
 	int			dimensions = centers->dim;
 	int			numCenters = centers->maxlen;
@@ -430,6 +456,7 @@ 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 */
@@ -449,6 +476,8 @@ 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);
@@ -465,6 +494,7 @@ 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++)
 		{
@@ -477,6 +507,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, IvfflatTyp
 	else if (type == IVFFLAT_TYPE_BIT)
 	{
 		newCenters = VectorArrayInit(numCenters, dimensions, centers->itemsize);
+		newCenters->length = numCenters;
 
 		for (int j = 0; j < numCenters; j++)
 		{
@@ -627,7 +658,7 @@ 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, collation, type);
+		ComputeNewCenters(samples, aggCenters, newCenters, centerCounts, closestCenters, normprocinfo, normalizeprocinfo, collation, type);
 
 		/* Step 5 */
 		for (int j = 0; j < numCenters; j++)
@@ -710,13 +741,7 @@ CheckCenters(Relation index, VectorArray centers, IvfflatType type)
 	if (type != IVFFLAT_TYPE_BIT)
 	{
 		/* Ensure no duplicate centers */
-		/* Fine to sort in-place */
+		SortVectorArray(centers, type);
-		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++)
 		{

src/ivfscan.c

@@ -3,10 +3,8 @@
 #include <float.h>
 
 #include "access/relscan.h"
-#include "bitvector.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"
@@ -58,7 +56,7 @@ GetScanLists(IndexScanDesc scan, Datum value)
 			double		distance;
 
 			/* Use procinfo from the index instead of scan key for performance */
-			distance = DatumGetFloat8(FunctionCall2Coll(so->procinfo, so->collation, PointerGetDatum(&list->center), value));
+			distance = DatumGetFloat8(so->distfunc(so->procinfo, so->collation, PointerGetDatum(&list->center), value));
 
 			if (listCount < so->probes)
 			{
@@ -151,7 +149,7 @@ GetScanItems(IndexScanDesc scan, Datum value)
 				 * performance
 				 */
 				ExecClearTuple(slot);
-				slot->tts_values[0] = FunctionCall2Coll(so->procinfo, so->collation, datum, value);
+				slot->tts_values[0] = so->distfunc(so->procinfo, so->collation, datum, value);
 				slot->tts_isnull[0] = false;
 				slot->tts_values[1] = PointerGetDatum(&itup->t_tid);
 				slot->tts_isnull[1] = false;
@@ -179,6 +177,15 @@ GetScanItems(IndexScanDesc scan, Datum value)
 	tuplesort_performsort(so->sortstate);
 }
 
+/*
+ * Zero distance
+ */
+static Datum
+ZeroDistance(FmgrInfo *flinfo, Oid collation, Datum arg1, Datum arg2)
+{
+	return Float8GetDatum(0.0);
+}
+
 /*
  * Get scan value
  */
@@ -190,28 +197,21 @@ GetScanValue(IndexScanDesc scan)
 
 	if (scan->orderByData->sk_flags & SK_ISNULL)
 	{
-		IvfflatType type = IvfflatGetType(scan->indexRelation);
+		value = PointerGetDatum(NULL);
-
+		so->distfunc = ZeroDistance;
-		if (type == IVFFLAT_TYPE_VECTOR)
-			value = PointerGetDatum(InitVector(so->dimensions));
-		else if (type == IVFFLAT_TYPE_HALFVEC)
-			value = PointerGetDatum(InitHalfVector(so->dimensions));
-		else if (type == IVFFLAT_TYPE_BIT)
-			value = PointerGetDatum(InitBitVector(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)));
 
-		/* Fine if normalization fails */
+		/* Check normprocinfo since normalizeprocinfo not set for vector */
 		if (so->normprocinfo != NULL)
-			IvfflatNormValue(so->normprocinfo, so->collation, &value, IvfflatGetType(scan->indexRelation));
+			value = IvfflatNormValue(so->normalizeprocinfo, so->collation, value);
 	}
 
 	return value;
@@ -249,6 +249,7 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
 	/* Set support functions */
 	so->procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
 	so->normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
+	so->normalizeprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORMALIZE_PROC);
 	so->collation = index->rd_indcollation[0];
 
 	/* Create tuple description for sorting */

src/ivfutils.c

@@ -3,12 +3,8 @@
 #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
@@ -71,60 +67,40 @@ IvfflatOptionalProcInfo(Relation index, uint16 procnum)
 IvfflatType
 IvfflatGetType(Relation index)
 {
+	FmgrInfo   *procinfo = IvfflatOptionalProcInfo(index, IVFFLAT_TYPE_SUPPORT_PROC);
 	Oid			typid = TupleDescAttr(index->rd_att, 0)->atttypid;
-	HeapTuple	tuple;
-	Form_pg_type type;
 	IvfflatType result;
 
-	if (typid == BITOID)
+	if (procinfo == NULL)
-		return IVFFLAT_TYPE_BIT;
+		return IVFFLAT_TYPE_VECTOR;
 
-	tuple = SearchSysCache1(TYPEOID, ObjectIdGetDatum(typid));
+	result = (IvfflatType) DatumGetInt32(FunctionCall1(procinfo, ObjectIdGetDatum(typid)));
-	if (!HeapTupleIsValid(tuple))
-		elog(ERROR, "cache lookup failed for type %u", typid);
 
-	type = (Form_pg_type) GETSTRUCT(tuple);
+	if (result == IVFFLAT_TYPE_UNSUPPORTED)
-	if (strcmp(NameStr(type->typname), "vector") == 0)
-		result = IVFFLAT_TYPE_VECTOR;
-	else if (strcmp(NameStr(type->typname), "halfvec") == 0)
-		result = IVFFLAT_TYPE_HALFVEC;
-	else
-	{
-		ReleaseSysCache(tuple);
 		elog(ERROR, "type not supported for ivfflat index");
-	}
-
-	ReleaseSysCache(tuple);
 
 	return result;
 }
 
 /*
- * Divide by the norm
+ * Normalize value
- *
+ */
- * Returns false if value should not be indexed
+Datum
- *
+IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum value)
- * The caller needs to free the pointer stored in value
+{
- * if it's different than the original value
+	if (procinfo == NULL)
+		return DirectFunctionCall1(l2_normalize, value);
+
+	return FunctionCall1Coll(procinfo, collation, value);
+}
+
+/*
+ * Check if non-zero norm
  */
 bool
-IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, IvfflatType type)
+IvfflatCheckNorm(FmgrInfo *procinfo, Oid collation, Datum value)
 {
-	double		norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
+	return DatumGetFloat8(FunctionCall1Coll(procinfo, collation, value)) > 0;
-
-	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;
 }
 
 /*
@@ -269,3 +245,22 @@ IvfflatUpdateList(Relation index, ListInfo listInfo,
 		UnlockReleaseBuffer(buf);
 	}
 }
+
+PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_ivfflat_support);
+Datum
+halfvec_ivfflat_support(PG_FUNCTION_ARGS)
+{
+	PG_RETURN_INT32(IVFFLAT_TYPE_HALFVEC);
+};
+
+PGDLLEXPORT PG_FUNCTION_INFO_V1(bit_ivfflat_support);
+Datum
+bit_ivfflat_support(PG_FUNCTION_ARGS)
+{
+	Oid			typid = PG_GETARG_OID(0);
+
+	if (typid == BITOID)
+		PG_RETURN_INT32(IVFFLAT_TYPE_BIT);
+	else
+		PG_RETURN_INT32(IVFFLAT_TYPE_UNSUPPORTED);
+};

src/sparsevec.c

@@ -4,6 +4,8 @@
 #include <math.h>
 
 #include "fmgr.h"
+#include "halfutils.h"
+#include "halfvec.h"
 #include "libpq/pqformat.h"
 #include "sparsevec.h"
 #include "utils/array.h"
@@ -18,6 +20,12 @@
 #include "utils/builtins.h"
 #endif
 
+typedef struct SparseInputElement
+{
+	int32		index;
+	float		value;
+}			SparseInputElement;
+
 /*
  * Ensure same dimensions
  */
@@ -164,6 +172,21 @@ sparsevec_isspace(char ch)
 	return false;
 }
 
+/*
+ * Compare indices
+ */
+static int
+CompareIndices(const void *a, const void *b)
+{
+	if (((SparseInputElement *) a)->index < ((SparseInputElement *) b)->index)
+		return -1;
+
+	if (((SparseInputElement *) a)->index > ((SparseInputElement *) b)->index)
+		return 1;
+
+	return 0;
+}
+
 /*
  * Convert textual representation to internal representation
  */
@@ -178,8 +201,7 @@ sparsevec_in(PG_FUNCTION_ARGS)
 	char	   *stringEnd;
 	SparseVector *result;
 	float	   *rvalues;
-	int32	   *indices;
+	SparseInputElement *elements;
-	float	   *values;
 	int			maxNnz;
 	int			nnz = 0;
 
@@ -197,8 +219,7 @@ sparsevec_in(PG_FUNCTION_ARGS)
 				(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
 				 errmsg("sparsevec cannot have more than %d non-zero elements", SPARSEVEC_MAX_NNZ)));
 
-	indices = palloc(maxNnz * sizeof(int32));
+	elements = palloc(maxNnz * sizeof(SparseInputElement));
-	values = palloc(maxNnz * sizeof(float));
 
 	pt = lit;
 
@@ -291,8 +312,8 @@ sparsevec_in(PG_FUNCTION_ARGS)
 			/* Do not store zero values */
 			if (value != 0)
 			{
-				indices[nnz] = index;
+				elements[nnz].index = index;
-				values[nnz] = value;
+				elements[nnz].value = value;
 				nnz++;
 			}
 
@@ -353,12 +374,14 @@ sparsevec_in(PG_FUNCTION_ARGS)
 	CheckDim(dim);
 	CheckExpectedDim(typmod, dim);
 
+	qsort(elements, nnz, sizeof(SparseInputElement), CompareIndices);
+
 	result = InitSparseVector(dim, nnz);
 	rvalues = SPARSEVEC_VALUES(result);
 	for (int i = 0; i < nnz; i++)
 	{
-		result->indices[i] = indices[i];
+		result->indices[i] = elements[i].index;
-		rvalues[i] = values[i];
+		rvalues[i] = elements[i].value;
 
 		CheckIndex(result->indices, i, dim);
 	}
@@ -504,6 +527,10 @@ sparsevec_recv(PG_FUNCTION_ARGS)
 	{
 		values[i] = pq_getmsgfloat4(buf);
 		CheckElement(values[i]);
+		if (values[i] == 0)
+			ereport(ERROR,
+					(errcode(ERRCODE_DATA_EXCEPTION),
+					 errmsg("binary representation of sparsevec cannot contain zero values")));
 	}
 
 	PG_RETURN_POINTER(result);
@@ -591,6 +618,49 @@ 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
  */

src/sparsevec.h

@@ -23,6 +23,5 @@ typedef struct SparseVector
 }			SparseVector;
 
 SparseVector *InitSparseVector(int dim, int nnz);
-PGDLLEXPORT Datum sparsevec_l2_normalize(PG_FUNCTION_ARGS);
 
 #endif

src/vector.c

@@ -2,7 +2,8 @@
 
 #include <math.h>
 
-#include "bitvector.h"
+#include "bitutils.h"
+#include "bitvec.h"
 #include "catalog/pg_type.h"
 #include "common/shortest_dec.h"
 #include "fmgr.h"
@@ -33,18 +34,10 @@
 #define STATE_DIMS(x) (ARR_DIMS(x)[0] - 1)
 #define CreateStateDatums(dim) palloc(sizeof(Datum) * (dim + 1))
 
-/* target_clones requires glibc */
+#if defined(USE_TARGET_CLONES) && !defined(__FMA__)
-#if defined(__gnu_linux__) && defined(__has_attribute)
+#define VECTOR_TARGET_CLONES __attribute__((target_clones("default", "fma")))
-/* Use separate line for portability */
-#if __has_attribute(target_clones)
-#define HAVE_TARGET_CLONES
-#endif
-#endif
-
-#if defined(__x86_64__) && defined(HAVE_TARGET_CLONES) && !defined(__FMA__)
-#define VECTOR_DISPATCH __attribute__((target_clones("default", "fma")))
 #else
-#define VECTOR_DISPATCH
+#define VECTOR_TARGET_CLONES
 #endif
 
 PG_MODULE_MAGIC;
@@ -56,6 +49,7 @@ PGDLLEXPORT void _PG_init(void);
 void
 _PG_init(void)
 {
+	BitvecInit();
 	HalfvecInit();
 	HnswInit();
 	IvfflatInit();
@@ -571,7 +565,7 @@ halfvec_to_vector(PG_FUNCTION_ARGS)
 	PG_RETURN_POINTER(result);
 }
 
-VECTOR_DISPATCH static float
+VECTOR_TARGET_CLONES static float
 VectorL2SquaredDistance(int dim, float *ax, float *bx)
 {
 	float		distance = 0.0;
@@ -618,7 +612,7 @@ vector_l2_squared_distance(PG_FUNCTION_ARGS)
 	PG_RETURN_FLOAT8((double) VectorL2SquaredDistance(a->dim, a->x, b->x));
 }
 
-VECTOR_DISPATCH static float
+VECTOR_TARGET_CLONES static float
 VectorInnerProduct(int dim, float *ax, float *bx)
 {
 	float		distance = 0.0;
@@ -660,7 +654,7 @@ vector_negative_inner_product(PG_FUNCTION_ARGS)
 	PG_RETURN_FLOAT8((double) -VectorInnerProduct(a->dim, a->x, b->x));
 }
 
-VECTOR_DISPATCH static double
+VECTOR_TARGET_CLONES static double
 VectorCosineSimilarity(int dim, float *ax, float *bx)
 {
 	float		similarity = 0.0;
@@ -735,7 +729,8 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
 	PG_RETURN_FLOAT8(acos(distance) / M_PI);
 }
 
-static float
+/* Does not require FMA, but keep logic simple */
+VECTOR_TARGET_CLONES static float
 VectorL1Distance(int dim, float *ax, float *bx)
 {
 	float		distance = 0.0;

test/expected/bit_functions.out

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

test/expected/cast.out

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

test/expected/halfvec_functions.out

@@ -304,6 +304,24 @@ 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

@@ -0,0 +1,21 @@
+SET enable_seqscan = off;
+CREATE TABLE t (val halfvec(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING hnsw (val halfvec_l1_ops);
+INSERT INTO t (val) VALUES ('[1,2,4]');
+SELECT * FROM t ORDER BY val <+> '[3,3,3]';
+   val   
+---------
+ [1,2,3]
+ [1,2,4]
+ [1,1,1]
+ [0,0,0]
+(4 rows)
+
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <+> (SELECT NULL::halfvec)) t2;
+ count 
+-------
+     4
+(1 row)
+
+DROP TABLE t;

test/expected/hnsw_sparsevec_l1.out

@@ -0,0 +1,21 @@
+SET enable_seqscan = off;
+CREATE TABLE t (val sparsevec(3));
+INSERT INTO t (val) VALUES ('{}/3'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
+CREATE INDEX ON t USING hnsw (val sparsevec_l1_ops);
+INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
+SELECT * FROM t ORDER BY val <+> '{1:3,2:3,3:3}/3';
+       val       
+-----------------
+ {1:1,2:2,3:3}/3
+ {1:1,2:2,3:4}/3
+ {1:1,2:1,3:1}/3
+ {}/3
+(4 rows)
+
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <+> (SELECT NULL::sparsevec)) t2;
+ count 
+-------
+     4
+(1 row)
+
+DROP TABLE t;

test/expected/hnsw_vector_l1.out

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

test/expected/ivfflat_bit_jaccard.out

@@ -1,21 +0,0 @@
-SET enable_seqscan = off;
-CREATE TABLE t (val bit(4));
-INSERT INTO t (val) VALUES (B'0000'), (B'1100'), (B'1111'), (NULL);
-CREATE INDEX ON t USING ivfflat (val bit_jaccard_ops) WITH (lists = 1);
-INSERT INTO t (val) VALUES (B'1110');
-SELECT * FROM t ORDER BY val <%> B'1111';
- val  
-------
- 1111
- 1110
- 1100
- 0000
-(4 rows)
-
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <%> (SELECT NULL::bit)) t2;
- count 
--------
-     4
-(1 row)
-
-DROP TABLE t;

test/expected/sparsevec_input.out

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

test/expected/vector_functions.out

@@ -202,6 +202,12 @@ 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? 
 ----------
@@ -222,6 +228,12 @@ 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? 
 ----------
@@ -278,6 +290,18 @@ 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? 
 ----------
@@ -304,6 +328,24 @@ 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

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

test/sql/cast.sql

@@ -47,6 +47,17 @@ SELECT '{2:1.5,4:3.5}/5'::sparsevec::vector(5);
 SELECT '{2:1.5,4:3.5}/5'::sparsevec::vector(4);
 SELECT '{}/16001'::sparsevec::vector;
 
+SELECT '[0,1.5,0,3.5,0]'::halfvec::sparsevec;
+SELECT '[0,1.5,0,3.5,0]'::halfvec::sparsevec(5);
+SELECT '[0,1.5,0,3.5,0]'::halfvec::sparsevec(4);
+
+SELECT '{2:1.5,4:3.5}/5'::sparsevec::halfvec;
+SELECT '{2:1.5,4:3.5}/5'::sparsevec::halfvec(5);
+SELECT '{2:1.5,4:3.5}/5'::sparsevec::halfvec(4);
+SELECT '{}/16001'::sparsevec::halfvec;
+SELECT '{1:65520}/1'::sparsevec::halfvec;
+SELECT '{1:1e-8}/1'::sparsevec::halfvec;
+
 SELECT array_agg(n)::vector FROM generate_series(1, 16001) n;
 SELECT array_to_vector(array_agg(n), 16001, false) FROM generate_series(1, 16001) n;
 

test/sql/halfvec_functions.sql

@@ -68,6 +68,9 @@ SELECT '[1,2]'::halfvec <=> '[2,4]';
 SELECT l1_distance('[0,0]'::halfvec, '[3,4]');
 SELECT l1_distance('[0,0]'::halfvec, '[0,1]');
 SELECT l1_distance('[1,2]'::halfvec, '[3]');
+SELECT l1_distance('[1,2,3,4,5,6,7,8,9]'::halfvec, '[1,2,3,4,5,6,7,8,9]');
+SELECT l1_distance('[1,2,3,4,5,6,7,8,9]'::halfvec, '[0,3,2,5,4,7,6,9,8]');
+SELECT '[0,0]'::halfvec <+> '[3,4]';
 
 SELECT l2_normalize('[3,4]'::halfvec);
 SELECT l2_normalize('[3,0]'::halfvec);

test/sql/hnsw_halfvec_l1.sql

@@ -0,0 +1,12 @@
+SET enable_seqscan = off;
+
+CREATE TABLE t (val halfvec(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING hnsw (val halfvec_l1_ops);
+
+INSERT INTO t (val) VALUES ('[1,2,4]');
+
+SELECT * FROM t ORDER BY val <+> '[3,3,3]';
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <+> (SELECT NULL::halfvec)) t2;
+
+DROP TABLE t;

test/sql/hnsw_sparsevec_l1.sql

@@ -0,0 +1,12 @@
+SET enable_seqscan = off;
+
+CREATE TABLE t (val sparsevec(3));
+INSERT INTO t (val) VALUES ('{}/3'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
+CREATE INDEX ON t USING hnsw (val sparsevec_l1_ops);
+
+INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
+
+SELECT * FROM t ORDER BY val <+> '{1:3,2:3,3:3}/3';
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <+> (SELECT NULL::sparsevec)) t2;
+
+DROP TABLE t;

test/sql/hnsw_vector_l1.sql

@@ -0,0 +1,12 @@
+SET enable_seqscan = off;
+
+CREATE TABLE t (val vector(3));
+INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
+CREATE INDEX ON t USING hnsw (val vector_l1_ops);
+
+INSERT INTO t (val) VALUES ('[1,2,4]');
+
+SELECT * FROM t ORDER BY val <+> '[3,3,3]';
+SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <+> (SELECT NULL::vector)) t2;
+
+DROP TABLE t;

test/sql/ivfflat_bit_jaccard.sql

@@ -1,12 +0,0 @@
-SET enable_seqscan = off;
-
-CREATE TABLE t (val bit(4));
-INSERT INTO t (val) VALUES (B'0000'), (B'1100'), (B'1111'), (NULL);
-CREATE INDEX ON t USING ivfflat (val bit_jaccard_ops) WITH (lists = 1);
-
-INSERT INTO t (val) VALUES (B'1110');
-
-SELECT * FROM t ORDER BY val <%> B'1111';
-SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <%> (SELECT NULL::bit)) t2;
-
-DROP TABLE t;

test/sql/sparsevec_input.sql

@@ -34,6 +34,8 @@ SELECT '{1:1a}/1'::sparsevec;
 SELECT '{1:1,}/1'::sparsevec;
 SELECT '{1:0,2:1,3:0}/3'::sparsevec;
 SELECT '{2:1,1:1}/2'::sparsevec;
+SELECT '{1:1,1:1}/2'::sparsevec;
+SELECT '{1:1,2:1,1:1}/2'::sparsevec;
 SELECT '{}/5'::sparsevec;
 SELECT '{}/-1'::sparsevec;
 SELECT '{}/100001'::sparsevec;

test/sql/vector_functions.sql

@@ -46,11 +46,13 @@ 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]');
@@ -62,12 +64,17 @@ 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 @operators = ("<->", "<#>", "<=>", "<+>");
-my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
+my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops", "vector_l1_ops");
 
 for my $i (0 .. $#operators)
 {

test/t/013_hnsw_insert_recall.pl

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

test/t/021_hnsw_halfvec_build_recall.pl

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

test/t/022_hnsw_sparsevec_build_recall.pl

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

test/t/024_hnsw_halfvec_insert_recall.pl

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

test/t/025_hnsw_sparsevec_insert_recall.pl

@@ -64,8 +64,8 @@ for (1 .. 20)
 }
 
 # Check each index type
-my @operators = ("<->", "<#>", "<=>");
+my @operators = ("<->", "<#>", "<=>", "<+>");
-my @opclasses = ("sparsevec_l2_ops", "sparsevec_ip_ops", "sparsevec_cosine_ops");
+my @opclasses = ("sparsevec_l2_ops", "sparsevec_ip_ops", "sparsevec_cosine_ops", "sparsevec_l1_ops");
 
 for my $i (0 .. $#operators)
 {

test/t/035_ivfflat_bit_build_recall.pl

@@ -70,8 +70,8 @@ for (1 .. 20)
 }
 
 # Check each index type
-my @operators = ("<~>", "<\%>");
+my @operators = ("<~>");
-my @opclasses = ("bit_hamming_ops", "bit_jaccard_ops");
+my @opclasses = ("bit_hamming_ops");
 
 for my $i (0 .. $#operators)
 {