Compare commits

..

42 Commits

Author SHA1 Message Date
Andrew Kane
0c0cb3e35e Improved code 2023-08-08 16:15:36 -07:00
Andrew Kane
19c7b4e85b Started docs [skip ci] 2023-08-08 01:36:35 -07:00
Andrew Kane
a4a6ca6780 Updated changelog and thanks [skip ci] 2023-08-08 01:13:02 -07:00
Andrew Kane
d027cb586e Improved element size estimate [skip ci] 2023-08-06 14:11:33 -07:00
Andrew Kane
0ffdefe138 Improved memory clean-up [skip ci] 2023-08-06 14:09:42 -07:00
Andrew Kane
72a8f68dc5 Improving naming [skip ci] 2023-08-06 14:07:48 -07:00
Andrew Kane
11e0b87abe Added basic cost estimation [skip ci] 2023-08-05 11:55:24 -07:00
Andrew Kane
405b7d98dc Improved message [skip ci] 2023-08-04 23:59:51 -07:00
Andrew Kane
7c217bad0d Removed todos [skip ci] 2023-08-04 23:49:56 -07:00
Andrew Kane
f82c90ce14 Skip deleted tuples [skip ci] 2023-08-04 23:38:37 -07:00
Andrew Kane
3390eb783b Added vacuum stats [skip ci] 2023-08-04 23:32:12 -07:00
Andrew Kane
247bc14ca1 Improved code [skip ci] 2023-08-04 23:31:58 -07:00
Andrew Kane
8f0e865137 Removed todo [skip ci] 2023-08-04 21:38:20 -07:00
Andrew Kane
7194c0281c Fixed repairing graph entry point 2023-08-04 21:02:45 -07:00
Andrew Kane
f55bf54588 No need to free [skip ci] 2023-08-04 20:02:01 -07:00
Andrew Kane
3a6cd544ff Use memory context for index scan [skip ci] 2023-08-04 19:59:09 -07:00
Andrew Kane
4f4f444396 Revert "Fixed repairing graph entry point"
This reverts commit f27ac1e338.
2023-08-04 16:30:29 -07:00
Andrew Kane
89835d394e Test counts 2023-08-04 16:15:25 -07:00
Andrew Kane
348bfb71ff Non-deterministic order for NULL and NaN distances 2023-08-04 16:08:19 -07:00
Andrew Kane
f27ac1e338 Fixed repairing graph entry point 2023-08-04 15:59:11 -07:00
Andrew Kane
6e0ddf26f0 Fixed comment [skip ci] 2023-08-04 14:01:09 -07:00
Andrew Kane
029c336c62 Fixed results for NULL and NaN distances [skip ci] 2023-08-04 13:51:16 -07:00
Andrew Kane
7f4acf9d43 Use neighbor page to be consistent with insert page [skip ci] 2023-08-04 09:51:32 -07:00
Andrew Kane
fe934c1465 Only update insert page if first free element 2023-08-04 09:47:32 -07:00
Andrew Kane
e8a6becff7 Added entry level to metapage [skip ci] 2023-08-02 22:06:52 -07:00
Andrew Kane
62067b298d Updated test recall for inner product [skip ci] 2023-08-02 18:29:31 -07:00
Andrew Kane
3424b49033 Improved test time for WAL test [skip ci] 2023-08-02 18:24:22 -07:00
Andrew Kane
c49557674c Removed level check 2023-08-02 18:21:07 -07:00
Andrew Kane
8a89f9deb6 Partially fixed repairing graph entry point [skip ci] 2023-08-02 09:48:01 -07:00
Andrew Kane
1a4fce10be Improved naming [skip ci] 2023-08-01 23:27:54 -07:00
Andrew Kane
18b315bb08 Improved naming [skip ci] 2023-08-01 23:25:42 -07:00
Andrew Kane
fae2b445d6 Removed eager loading of neighbors for better performance [skip ci] 2023-08-01 21:36:04 -07:00
Andrew Kane
f3bd7c30d4 Fixed comment [skip ci] 2023-08-01 20:38:10 -07:00
Andrew Kane
5802009c86 Improved variable names [skip ci] 2023-08-01 20:15:03 -07:00
Andrew Kane
3169a60e5c Restored interrupts [skip ci] 2023-08-01 19:53:45 -07:00
Andrew Kane
e2a8dd6594 Fixed CI 2023-08-01 19:33:38 -07:00
Andrew Kane
ae78f732ef Fixed CI 2023-08-01 19:25:52 -07:00
Andrew Kane
d3e08fdf99 Reduced index size 2023-08-01 19:17:39 -07:00
Andrew Kane
95eded091f DRY code [skip ci] 2023-07-30 21:00:00 -07:00
Andrew Kane
9e115c629f Improved UpdateMetaPage logic [skip ci] 2023-07-30 17:23:24 -07:00
Andrew Kane
d902e1daff Improved performance 2023-07-22 09:33:32 -07:00
Andrew Kane
f0760eee76 Added HNSW index type 2023-07-21 16:25:39 -07:00
113 changed files with 1935 additions and 8998 deletions

View File

@@ -8,8 +8,6 @@ jobs:
fail-fast: false
matrix:
include:
- postgres: 17
os: ubuntu-22.04
- postgres: 16
os: ubuntu-22.04
- postgres: 15
@@ -20,15 +18,17 @@ jobs:
os: ubuntu-20.04
- postgres: 12
os: ubuntu-20.04
- postgres: 11
os: ubuntu-20.04
steps:
- uses: actions/checkout@v4
- uses: actions/checkout@v3
- 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
PG_CFLAGS: -Wall -Wextra -Werror -Wno-unused-parameter -Wno-sign-compare
- run: |
export PG_CONFIG=`which pg_config`
sudo --preserve-env=PG_CONFIG make install
@@ -40,54 +40,37 @@ jobs:
sudo apt-get install libipc-run-perl
- run: make prove_installcheck
mac:
runs-on: ${{ matrix.os }}
runs-on: macos-latest
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: actions/checkout@v3
- uses: ankane/setup-postgres@v1
with:
postgres-version: ${{ matrix.postgres }}
postgres-version: 14
- run: make
env:
PG_CFLAGS: -DUSE_ASSERT_CHECKING -Wall -Wextra -Werror -Wno-unused-parameter
PG_CFLAGS: -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
wget -q https://github.com/postgres/postgres/archive/refs/tags/REL_14_5.tar.gz
tar xf REL_14_5.tar.gz
- run: make prove_installcheck PROVE_FLAGS="-I ./postgres-REL_14_5/src/test/perl" PERL5LIB="/Users/runner/perl5/lib/perl5"
- run: make clean && /usr/local/opt/llvm@15/bin/scan-build --status-bugs make
windows:
runs-on: windows-latest
if: ${{ !startsWith(github.ref_name, 'mac') }}
steps:
- uses: actions/checkout@v4
- uses: actions/checkout@v3
- 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 && ^
@@ -95,13 +78,12 @@ jobs:
nmake /NOLOGO /F Makefile.win uninstall
shell: cmd
i386:
if: ${{ !startsWith(github.ref_name, 'mac') && !startsWith(github.ref_name, 'windows') }}
runs-on: ubuntu-latest
container:
image: debian:12
image: debian:11
options: --platform linux/386
steps:
- run: apt-get update && apt-get install -y build-essential git libipc-run-perl postgresql-15 postgresql-server-dev-15 sudo
- run: apt-get update && apt-get install -y build-essential git libipc-run-perl postgresql-13 postgresql-server-dev-13 sudo
- run: service postgresql start
- run: |
git clone https://github.com/${{ github.repository }}.git pgvector
@@ -114,15 +96,4 @@ jobs:
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
valgrind:
if: ${{ !startsWith(github.ref_name, 'mac') && !startsWith(github.ref_name, 'windows') }}
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@v4
- uses: ankane/setup-postgres-valgrind@v1
with:
postgres-version: 16
- run: make OPTFLAGS=""
- run: sudo --preserve-env=PG_CONFIG make install
- run: make installcheck
PG_CFLAGS: -Wall -Wextra -Werror -Wno-unused-parameter -Wno-sign-compare

View File

@@ -1,55 +1,11 @@
## 0.7.0 (unreleased)
- Added `halfvec` type
- Added `sparsevec` type
- Added support for `bit` vectors to HNSW
- Added `hamming_distance` function
- Added `jaccard_distance` function
- Added `quantize_binary` function
- Added `subvector` function
- Updated comparison operators to support vectors with different dimensions
## 0.6.2 (2024-03-18)
- Reduced lock contention with parallel HNSW index builds
## 0.6.1 (2024-03-04)
- Fixed error with `ANALYZE` and vectors with different dimensions
- Fixed segmentation fault with `shared_preload_libraries`
- Fixed vector subtraction being marked as commutative
## 0.6.0 (2024-01-29)
If upgrading with Postgres 12 or Docker, see [these notes](https://github.com/pgvector/pgvector#060).
- Added support for parallel index builds for HNSW
- Added validation for GUC parameters
- Changed storage for vector from `extended` to `external`
- Improved performance of HNSW
- Reduced memory usage for HNSW index builds
- Reduced WAL generation for HNSW index builds
- Fixed error with logical replication
- Fixed `invalid memory alloc request size` error with HNSW index builds
- Moved Docker image to `pgvector` org
- Added Docker tags for each supported version of Postgres
- Dropped support for Postgres 11
## 0.5.1 (2023-10-10)
- Improved performance of HNSW index builds
- Added check for MVCC-compliant snapshot for index scans
## 0.5.0 (2023-08-28)
## 0.5.0 (unreleased)
- Added HNSW index type
- Added support for parallel index builds for IVFFlat
- Added support for parallel index builds
- Added `l1_distance` function
- Added element-wise multiplication for vectors
- Added `sum` aggregate
- Improved performance of distance functions
- Fixed out of range results for cosine distance
- Fixed results for NULL and NaN distances for IVFFlat
## 0.4.4 (2023-06-12)
@@ -81,7 +37,7 @@ If upgrading with Postgres 12 or Docker, see [these notes](https://github.com/pg
## 0.4.0 (2023-01-11)
If upgrading with Postgres < 13, see [this note](https://github.com/pgvector/pgvector/blob/v0.4.0/README.md#040).
If upgrading with Postgres < 13, see [this note](https://github.com/pgvector/pgvector#040).
- Changed text representation for vector elements to match `real`
- Changed storage for vector from `plain` to `extended`
@@ -98,7 +54,7 @@ If upgrading with Postgres < 13, see [this note](https://github.com/pgvector/pgv
## 0.3.1 (2022-11-02)
If upgrading from 0.2.7 or 0.3.0, [recreate](https://github.com/pgvector/pgvector/blob/v0.3.1/README.md#031) all `ivfflat` indexes after upgrading to ensure all data is indexed.
If upgrading from 0.2.7 or 0.3.0, [recreate](https://github.com/pgvector/pgvector#031) all `ivfflat` indexes after upgrading to ensure all data is indexed.
- Fixed issue with inserts silently corrupting `ivfflat` indexes (introduced in 0.2.7)
- Fixed segmentation fault with index creation when lists > 6500

View File

@@ -1,11 +1,10 @@
ARG PG_MAJOR=16
ARG PG_MAJOR=15
FROM postgres:$PG_MAJOR
ARG PG_MAJOR
COPY . /tmp/pgvector
RUN apt-get update && \
apt-mark hold locales && \
apt-get install -y --no-install-recommends build-essential postgresql-server-dev-$PG_MAJOR && \
cd /tmp/pgvector && \
make clean && \
@@ -16,5 +15,4 @@ RUN apt-get update && \
rm -r /tmp/pgvector && \
apt-get remove -y build-essential postgresql-server-dev-$PG_MAJOR && \
apt-get autoremove -y && \
apt-mark unhold locales && \
rm -rf /var/lib/apt/lists/*

View File

@@ -1,4 +1,4 @@
Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
Portions Copyright (c) 1996-2023, PostgreSQL Global Development Group
Portions Copyright (c) 1994, The Regents of the University of California

View File

@@ -2,7 +2,7 @@
"name": "vector",
"abstract": "Open-source vector similarity search for Postgres",
"description": "Supports L2 distance, inner product, and cosine distance",
"version": "0.6.2",
"version": "0.4.4",
"maintainer": [
"Andrew Kane <andrew@ankane.org>"
],
@@ -12,7 +12,7 @@
"prereqs": {
"runtime": {
"requires": {
"PostgreSQL": "12.0.0"
"PostgreSQL": "11.0.0"
}
}
},
@@ -20,7 +20,7 @@
"vector": {
"file": "sql/vector.sql",
"docfile": "README.md",
"version": "0.6.2",
"version": "0.4.4",
"abstract": "Open-source vector similarity search for Postgres"
}
},

View File

@@ -1,18 +1,17 @@
EXTENSION = vector
EXTVERSION = 0.6.2
EXTVERSION = 0.4.4
MODULE_big = vector
DATA = $(wildcard sql/*--*.sql)
OBJS = src/bitvector.o src/halfvec.o src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/sparsevec.o src/vector.o
HEADERS = src/halfvec.h src/sparsevec.h src/vector.h
OBJS = 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/vector.o
TESTS = $(wildcard test/sql/*.sql)
REGRESS = $(patsubst test/sql/%.sql,%,$(TESTS))
REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)
REGRESS_OPTS = --inputdir=test --load-extension=vector
OPTFLAGS = -march=native
# Mac ARM doesn't always support -march=native
# Mac ARM doesn't support -march=native
ifeq ($(shell uname -s), Darwin)
ifeq ($(shell uname -p), arm)
# no difference with -march=armv8.5-a
@@ -65,15 +64,13 @@ dist:
mkdir -p dist
git archive --format zip --prefix=$(EXTENSION)-$(EXTVERSION)/ --output dist/$(EXTENSION)-$(EXTVERSION).zip master
# for Docker
PG_MAJOR ?= 16
.PHONY: docker
docker:
docker build --pull --no-cache --build-arg PG_MAJOR=$(PG_MAJOR) -t pgvector/pgvector:pg$(PG_MAJOR) -t pgvector/pgvector:$(EXTVERSION)-pg$(PG_MAJOR) .
docker build --pull --no-cache --platform linux/amd64 -t ankane/pgvector:latest .
.PHONY: docker-release
docker-release:
docker buildx build --push --pull --no-cache --platform linux/amd64,linux/arm64 --build-arg PG_MAJOR=$(PG_MAJOR) -t pgvector/pgvector:pg$(PG_MAJOR) -t pgvector/pgvector:$(EXTVERSION)-pg$(PG_MAJOR) .
docker buildx build --push --pull --no-cache --platform linux/amd64,linux/arm64 -t ankane/pgvector:latest .
docker buildx build --push --platform linux/amd64,linux/arm64 -t ankane/pgvector:v$(EXTVERSION) .

View File

@@ -1,11 +1,10 @@
EXTENSION = vector
EXTVERSION = 0.6.2
EXTVERSION = 0.4.4
OBJS = src\bitvector.obj src\halfvec.obj src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\sparsevec.obj src\vector.obj
HEADERS = src\halfvec.h src\sparsevec.h src\vector.h
OBJS = 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\vector.obj
REGRESS = bit_functions btree cast copy halfvec_functions halfvec_input hnsw_bit_hamming hnsw_bit_jaccard hnsw_halfvec_cosine hnsw_halfvec_ip hnsw_halfvec_l2 hnsw_options hnsw_sparsevec_cosine hnsw_sparsevec_ip hnsw_sparsevec_l2 hnsw_unlogged hnsw_vector_cosine hnsw_vector_ip hnsw_vector_l2 ivfflat_options ivfflat_unlogged ivfflat_vector_cosine ivfflat_vector_ip ivfflat_vector_l2 sparsevec_functions sparsevec_input vector_functions vector_input
REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)
REGRESS = btree cast copy functions input ivfflat_cosine ivfflat_ip ivfflat_l2 ivfflat_options ivfflat_unlogged
REGRESS_OPTS = --inputdir=test --load-extension=vector
# For /arch flags
# https://learn.microsoft.com/en-us/cpp/build/reference/arch-minimum-cpu-architecture
@@ -55,8 +54,6 @@ install:
copy $(SHLIB) "$(PKGLIBDIR)"
copy $(EXTENSION).control "$(SHAREDIR)\extension"
copy sql\$(EXTENSION)--*.sql "$(SHAREDIR)\extension"
mkdir "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
for %f in ($(HEADERS)) do copy %f "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
installcheck:
"$(BINDIR)\pg_regress" --bindir="$(BINDIR)" $(REGRESS_OPTS) $(REGRESS)
@@ -64,9 +61,7 @@ installcheck:
uninstall:
del /f "$(PKGLIBDIR)\$(SHLIB)"
del /f "$(SHAREDIR)\extension\$(EXTENSION).control"
del /f "$(SHAREDIR)\extension\$(EXTENSION)--*.sql"
del /f "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)\*.h"
rmdir "$(INCLUDEDIR_SERVER)\extension\$(EXTENSION)"
del /f "$(SHAREDIR)\extension\vector--*.sql"
clean:
del /f $(SHLIB) $(EXTENSION).lib $(EXTENSION).exp

768
README.md

File diff suppressed because it is too large Load Diff

View File

@@ -1,2 +0,0 @@
-- complain if script is sourced in psql, rather than via CREATE EXTENSION
\echo Use "ALTER EXTENSION vector UPDATE TO '0.5.1'" to load this file. \quit

View File

@@ -1,5 +0,0 @@
-- complain if script is sourced in psql, rather than via CREATE EXTENSION
\echo Use "ALTER EXTENSION vector UPDATE TO '0.6.0'" to load this file. \quit
-- remove this single line for Postgres < 13
ALTER TYPE vector SET (STORAGE = external);

View File

@@ -1,16 +0,0 @@
-- complain if script is sourced in psql, rather than via CREATE EXTENSION
\echo Use "ALTER EXTENSION vector UPDATE TO '0.6.1'" to load this file. \quit
DROP OPERATOR - (vector, vector);
CREATE OPERATOR - (
LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_sub
);
ALTER OPERATOR <= (vector, vector) SET (
RESTRICT = scalarlesel, JOIN = scalarlejoinsel
);
ALTER OPERATOR >= (vector, vector) SET (
RESTRICT = scalargesel, JOIN = scalargejoinsel
);

View File

@@ -1,2 +0,0 @@
-- complain if script is sourced in psql, rather than via CREATE EXTENSION
\echo Use "ALTER EXTENSION vector UPDATE TO '0.6.2'" to load this file. \quit

View File

@@ -1,259 +0,0 @@
-- complain if script is sourced in psql, rather than via CREATE EXTENSION
\echo Use "ALTER EXTENSION vector UPDATE TO '0.7.0'" to load this file. \quit
CREATE FUNCTION quantize_binary(vector) RETURNS bit
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION subvector(vector, int, int) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION 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 OPERATOR <~> (
LEFTARG = bit, RIGHTARG = bit, PROCEDURE = hamming_distance,
COMMUTATOR = '<~>'
);
CREATE OPERATOR <%> (
LEFTARG = bit, RIGHTARG = bit, PROCEDURE = jaccard_distance,
COMMUTATOR = '<%>'
);
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);
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);
CREATE TYPE halfvec;
CREATE FUNCTION halfvec_in(cstring, oid, integer) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_out(halfvec) RETURNS cstring
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_typmod_in(cstring[]) RETURNS integer
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_recv(internal, oid, integer) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_send(halfvec) RETURNS bytea
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE TYPE halfvec (
INPUT = halfvec_in,
OUTPUT = halfvec_out,
TYPMOD_IN = halfvec_typmod_in,
RECEIVE = halfvec_recv,
SEND = halfvec_send,
STORAGE = external
);
CREATE FUNCTION l2_distance(halfvec, halfvec) RETURNS float8
AS 'MODULE_PATHNAME', 'halfvec_l2_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION inner_product(halfvec, halfvec) RETURNS float8
AS 'MODULE_PATHNAME', 'halfvec_inner_product' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION cosine_distance(halfvec, halfvec) RETURNS float8
AS 'MODULE_PATHNAME', 'halfvec_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION l1_distance(halfvec, halfvec) RETURNS float8
AS 'MODULE_PATHNAME', 'halfvec_l1_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_norm(halfvec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION quantize_binary(halfvec) RETURNS bit
AS 'MODULE_PATHNAME', 'halfvec_quantize_binary' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION subvector(halfvec, int, int) RETURNS halfvec
AS 'MODULE_PATHNAME', 'halfvec_subvector' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_l2_squared_distance(halfvec, halfvec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_negative_inner_product(halfvec, halfvec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec(halfvec, integer, boolean) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_halfvec(integer[], integer, boolean) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_halfvec(real[], integer, boolean) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_halfvec(double precision[], integer, boolean) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_halfvec(numeric[], integer, boolean) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_to_float4(halfvec, integer, boolean) RETURNS real[]
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE CAST (halfvec AS halfvec)
WITH FUNCTION halfvec(halfvec, integer, boolean) AS IMPLICIT;
CREATE CAST (halfvec AS real[])
WITH FUNCTION halfvec_to_float4(halfvec, integer, boolean) AS IMPLICIT;
CREATE CAST (integer[] AS halfvec)
WITH FUNCTION array_to_halfvec(integer[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (real[] AS halfvec)
WITH FUNCTION array_to_halfvec(real[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (double precision[] AS halfvec)
WITH FUNCTION array_to_halfvec(double precision[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (numeric[] AS halfvec)
WITH FUNCTION array_to_halfvec(numeric[], integer, boolean) AS ASSIGNMENT;
CREATE OPERATOR <-> (
LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = l2_distance,
COMMUTATOR = '<->'
);
CREATE OPERATOR <#> (
LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_negative_inner_product,
COMMUTATOR = '<#>'
);
CREATE OPERATOR <=> (
LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = cosine_distance,
COMMUTATOR = '<=>'
);
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);
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);
CREATE OPERATOR CLASS halfvec_cosine_ops
FOR TYPE halfvec USING hnsw AS
OPERATOR 1 <=> (halfvec, halfvec) FOR ORDER BY float_ops,
FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
FUNCTION 2 halfvec_norm(halfvec);
CREATE FUNCTION halfvec_to_vector(halfvec, integer, boolean) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION vector_to_halfvec(vector, integer, boolean) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE CAST (halfvec AS vector)
WITH FUNCTION halfvec_to_vector(halfvec, integer, boolean) AS IMPLICIT;
CREATE CAST (vector AS halfvec)
WITH FUNCTION vector_to_halfvec(vector, integer, boolean) AS IMPLICIT;
CREATE TYPE sparsevec;
CREATE FUNCTION sparsevec_in(cstring, oid, integer) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_out(sparsevec) RETURNS cstring
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_typmod_in(cstring[]) RETURNS integer
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_recv(internal, oid, integer) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_send(sparsevec) RETURNS bytea
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE TYPE sparsevec (
INPUT = sparsevec_in,
OUTPUT = sparsevec_out,
TYPMOD_IN = sparsevec_typmod_in,
RECEIVE = sparsevec_recv,
SEND = sparsevec_send,
STORAGE = external
);
CREATE FUNCTION l2_distance(sparsevec, sparsevec) RETURNS float8
AS 'MODULE_PATHNAME', 'sparsevec_l2_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION inner_product(sparsevec, sparsevec) RETURNS float8
AS 'MODULE_PATHNAME', 'sparsevec_inner_product' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION cosine_distance(sparsevec, sparsevec) RETURNS float8
AS 'MODULE_PATHNAME', 'sparsevec_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_norm(sparsevec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_l2_squared_distance(sparsevec, sparsevec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_negative_inner_product(sparsevec, sparsevec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec(sparsevec, integer, boolean) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION vector_to_sparsevec(vector, integer, boolean) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_to_vector(sparsevec, integer, boolean) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE CAST (sparsevec AS sparsevec)
WITH FUNCTION sparsevec(sparsevec, integer, boolean) AS IMPLICIT;
CREATE CAST (sparsevec AS vector)
WITH FUNCTION sparsevec_to_vector(sparsevec, integer, boolean) AS IMPLICIT;
CREATE CAST (vector AS sparsevec)
WITH FUNCTION vector_to_sparsevec(vector, integer, boolean) AS IMPLICIT;
CREATE OPERATOR <-> (
LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = l2_distance,
COMMUTATOR = '<->'
);
CREATE OPERATOR <#> (
LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_negative_inner_product,
COMMUTATOR = '<#>'
);
CREATE OPERATOR <=> (
LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = cosine_distance,
COMMUTATOR = '<=>'
);
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);
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);
CREATE OPERATOR CLASS sparsevec_cosine_ops
FOR TYPE sparsevec USING hnsw AS
OPERATOR 1 <=> (sparsevec, sparsevec) FOR ORDER BY float_ops,
FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec),
FUNCTION 2 sparsevec_norm(sparsevec);

View File

@@ -1,7 +1,7 @@
-- complain if script is sourced in psql, rather than via CREATE EXTENSION
\echo Use "CREATE EXTENSION vector" to load this file. \quit
-- vector type
-- type
CREATE TYPE vector;
@@ -26,10 +26,10 @@ CREATE TYPE vector (
TYPMOD_IN = vector_typmod_in,
RECEIVE = vector_recv,
SEND = vector_send,
STORAGE = external
STORAGE = extended
);
-- vector functions
-- functions
CREATE FUNCTION l2_distance(vector, vector) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
@@ -58,13 +58,7 @@ CREATE FUNCTION vector_sub(vector, vector) RETURNS vector
CREATE FUNCTION vector_mul(vector, vector) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION quantize_binary(vector) RETURNS bit
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION subvector(vector, int, int) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- vector private functions
-- private functions
CREATE FUNCTION vector_lt(vector, vector) RETURNS bool
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
@@ -105,7 +99,7 @@ CREATE FUNCTION vector_avg(double precision[]) RETURNS vector
CREATE FUNCTION vector_combine(double precision[], double precision[]) RETURNS double precision[]
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- vector aggregates
-- aggregates
CREATE AGGREGATE avg(vector) (
SFUNC = vector_accum,
@@ -123,7 +117,7 @@ CREATE AGGREGATE sum(vector) (
PARALLEL = SAFE
);
-- vector cast functions
-- cast functions
CREATE FUNCTION vector(vector, integer, boolean) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
@@ -143,7 +137,7 @@ CREATE FUNCTION array_to_vector(numeric[], integer, boolean) RETURNS vector
CREATE FUNCTION vector_to_float4(vector, integer, boolean) RETURNS real[]
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- vector casts
-- casts
CREATE CAST (vector AS vector)
WITH FUNCTION vector(vector, integer, boolean) AS IMPLICIT;
@@ -163,7 +157,7 @@ CREATE CAST (double precision[] AS vector)
CREATE CAST (numeric[] AS vector)
WITH FUNCTION array_to_vector(numeric[], integer, boolean) AS ASSIGNMENT;
-- vector operators
-- operators
CREATE OPERATOR <-> (
LEFTARG = vector, RIGHTARG = vector, PROCEDURE = l2_distance,
@@ -186,7 +180,8 @@ CREATE OPERATOR + (
);
CREATE OPERATOR - (
LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_sub
LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_sub,
COMMUTATOR = -
);
CREATE OPERATOR * (
@@ -200,10 +195,11 @@ CREATE OPERATOR < (
RESTRICT = scalarltsel, JOIN = scalarltjoinsel
);
-- should use scalarlesel and scalarlejoinsel, but not supported in Postgres < 11
CREATE OPERATOR <= (
LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_le,
COMMUTATOR = >= , NEGATOR = > ,
RESTRICT = scalarlesel, JOIN = scalarlejoinsel
RESTRICT = scalarltsel, JOIN = scalarltjoinsel
);
CREATE OPERATOR = (
@@ -218,10 +214,11 @@ CREATE OPERATOR <> (
RESTRICT = eqsel, JOIN = eqjoinsel
);
-- should use scalargesel and scalargejoinsel, but not supported in Postgres < 11
CREATE OPERATOR >= (
LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_ge,
COMMUTATOR = <= , NEGATOR = < ,
RESTRICT = scalargesel, JOIN = scalargejoinsel
RESTRICT = scalargtsel, JOIN = scalargtjoinsel
);
CREATE OPERATOR > (
@@ -246,7 +243,7 @@ CREATE ACCESS METHOD hnsw TYPE INDEX HANDLER hnswhandler;
COMMENT ON ACCESS METHOD hnsw IS 'hnsw index access method';
-- vector opclasses
-- opclasses
CREATE OPERATOR CLASS vector_ops
DEFAULT FOR TYPE vector USING btree AS
@@ -293,286 +290,3 @@ CREATE OPERATOR CLASS vector_cosine_ops
OPERATOR 1 <=> (vector, vector) FOR ORDER BY float_ops,
FUNCTION 1 vector_negative_inner_product(vector, vector),
FUNCTION 2 vector_norm(vector);
-- bit functions
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 OPERATOR <~> (
LEFTARG = bit, RIGHTARG = bit, PROCEDURE = hamming_distance,
COMMUTATOR = '<~>'
);
CREATE OPERATOR <%> (
LEFTARG = bit, RIGHTARG = bit, PROCEDURE = jaccard_distance,
COMMUTATOR = '<%>'
);
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);
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);
-- halfvec type
CREATE TYPE halfvec;
CREATE FUNCTION halfvec_in(cstring, oid, integer) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_out(halfvec) RETURNS cstring
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_typmod_in(cstring[]) RETURNS integer
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_recv(internal, oid, integer) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_send(halfvec) RETURNS bytea
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE TYPE halfvec (
INPUT = halfvec_in,
OUTPUT = halfvec_out,
TYPMOD_IN = halfvec_typmod_in,
RECEIVE = halfvec_recv,
SEND = halfvec_send,
STORAGE = external
);
-- halfvec functions
CREATE FUNCTION l2_distance(halfvec, halfvec) RETURNS float8
AS 'MODULE_PATHNAME', 'halfvec_l2_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION inner_product(halfvec, halfvec) RETURNS float8
AS 'MODULE_PATHNAME', 'halfvec_inner_product' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION cosine_distance(halfvec, halfvec) RETURNS float8
AS 'MODULE_PATHNAME', 'halfvec_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION l1_distance(halfvec, halfvec) RETURNS float8
AS 'MODULE_PATHNAME', 'halfvec_l1_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_norm(halfvec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION quantize_binary(halfvec) RETURNS bit
AS 'MODULE_PATHNAME', 'halfvec_quantize_binary' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION subvector(halfvec, int, int) RETURNS halfvec
AS 'MODULE_PATHNAME', 'halfvec_subvector' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- halfvec private functions
CREATE FUNCTION halfvec_l2_squared_distance(halfvec, halfvec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_negative_inner_product(halfvec, halfvec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- halfvec cast functions
CREATE FUNCTION halfvec(halfvec, integer, boolean) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_halfvec(integer[], integer, boolean) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_halfvec(real[], integer, boolean) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_halfvec(double precision[], integer, boolean) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_halfvec(numeric[], integer, boolean) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION halfvec_to_float4(halfvec, integer, boolean) RETURNS real[]
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- halfvec casts
CREATE CAST (halfvec AS halfvec)
WITH FUNCTION halfvec(halfvec, integer, boolean) AS IMPLICIT;
CREATE CAST (halfvec AS real[])
WITH FUNCTION halfvec_to_float4(halfvec, integer, boolean) AS IMPLICIT;
CREATE CAST (integer[] AS halfvec)
WITH FUNCTION array_to_halfvec(integer[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (real[] AS halfvec)
WITH FUNCTION array_to_halfvec(real[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (double precision[] AS halfvec)
WITH FUNCTION array_to_halfvec(double precision[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (numeric[] AS halfvec)
WITH FUNCTION array_to_halfvec(numeric[], integer, boolean) AS ASSIGNMENT;
-- halfvec operators
CREATE OPERATOR <-> (
LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = l2_distance,
COMMUTATOR = '<->'
);
CREATE OPERATOR <#> (
LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = halfvec_negative_inner_product,
COMMUTATOR = '<#>'
);
CREATE OPERATOR <=> (
LEFTARG = halfvec, RIGHTARG = halfvec, PROCEDURE = cosine_distance,
COMMUTATOR = '<=>'
);
-- halfvec opclasses
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);
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);
CREATE OPERATOR CLASS halfvec_cosine_ops
FOR TYPE halfvec USING hnsw AS
OPERATOR 1 <=> (halfvec, halfvec) FOR ORDER BY float_ops,
FUNCTION 1 halfvec_negative_inner_product(halfvec, halfvec),
FUNCTION 2 halfvec_norm(halfvec);
-- extension casts
CREATE FUNCTION halfvec_to_vector(halfvec, integer, boolean) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION vector_to_halfvec(vector, integer, boolean) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE CAST (halfvec AS vector)
WITH FUNCTION halfvec_to_vector(halfvec, integer, boolean) AS IMPLICIT;
CREATE CAST (vector AS halfvec)
WITH FUNCTION vector_to_halfvec(vector, integer, boolean) AS IMPLICIT;
--- sparsevec type
CREATE TYPE sparsevec;
CREATE FUNCTION sparsevec_in(cstring, oid, integer) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_out(sparsevec) RETURNS cstring
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_typmod_in(cstring[]) RETURNS integer
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_recv(internal, oid, integer) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_send(sparsevec) RETURNS bytea
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE TYPE sparsevec (
INPUT = sparsevec_in,
OUTPUT = sparsevec_out,
TYPMOD_IN = sparsevec_typmod_in,
RECEIVE = sparsevec_recv,
SEND = sparsevec_send,
STORAGE = external
);
-- sparsevec functions
CREATE FUNCTION l2_distance(sparsevec, sparsevec) RETURNS float8
AS 'MODULE_PATHNAME', 'sparsevec_l2_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION inner_product(sparsevec, sparsevec) RETURNS float8
AS 'MODULE_PATHNAME', 'sparsevec_inner_product' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION cosine_distance(sparsevec, sparsevec) RETURNS float8
AS 'MODULE_PATHNAME', 'sparsevec_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_norm(sparsevec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- sparsevec private functions
CREATE FUNCTION sparsevec_l2_squared_distance(sparsevec, sparsevec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_negative_inner_product(sparsevec, sparsevec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- sparsevec cast functions
CREATE FUNCTION sparsevec(sparsevec, integer, boolean) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION vector_to_sparsevec(vector, integer, boolean) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION sparsevec_to_vector(sparsevec, integer, boolean) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- sparsevec casts
CREATE CAST (sparsevec AS sparsevec)
WITH FUNCTION sparsevec(sparsevec, integer, boolean) AS IMPLICIT;
CREATE CAST (sparsevec AS vector)
WITH FUNCTION sparsevec_to_vector(sparsevec, integer, boolean) AS IMPLICIT;
CREATE CAST (vector AS sparsevec)
WITH FUNCTION vector_to_sparsevec(vector, integer, boolean) AS IMPLICIT;
-- sparsevec operators
CREATE OPERATOR <-> (
LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = l2_distance,
COMMUTATOR = '<->'
);
CREATE OPERATOR <#> (
LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = sparsevec_negative_inner_product,
COMMUTATOR = '<#>'
);
CREATE OPERATOR <=> (
LEFTARG = sparsevec, RIGHTARG = sparsevec, PROCEDURE = cosine_distance,
COMMUTATOR = '<=>'
);
-- sparsevec opclasses
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);
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);
CREATE OPERATOR CLASS sparsevec_cosine_ops
FOR TYPE sparsevec USING hnsw AS
OPERATOR 1 <=> (sparsevec, sparsevec) FOR ORDER BY float_ops,
FUNCTION 1 sparsevec_negative_inner_product(sparsevec, sparsevec),
FUNCTION 2 sparsevec_norm(sparsevec);

View File

@@ -1,90 +0,0 @@
#include "postgres.h"
#include "bitvector.h"
#include "port/pg_bitutils.h"
#include "utils/varbit.h"
#if PG_VERSION_NUM >= 160000
#include "varatt.h"
#endif
/*
* Allocate and initialize a new bit vector
*/
VarBit *
InitBitVector(int dim)
{
VarBit *result;
int size;
size = VARBITTOTALLEN(dim);
result = (VarBit *) palloc0(size);
SET_VARSIZE(result, size);
VARBITLEN(result) = dim;
return result;
}
/*
* Ensure same dimensions
*/
static inline void
CheckDims(VarBit *a, VarBit *b)
{
if (VARBITLEN(a) != VARBITLEN(b))
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("different bit lengths %u and %u", VARBITLEN(a), VARBITLEN(b))));
}
/*
* Get the Hamming distance between two bit vectors
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(hamming_distance);
Datum
hamming_distance(PG_FUNCTION_ARGS)
{
VarBit *a = PG_GETARG_VARBIT_P(0);
VarBit *b = PG_GETARG_VARBIT_P(1);
unsigned char *ax = VARBITS(a);
unsigned char *bx = VARBITS(b);
uint64 distance = 0;
CheckDims(a, b);
/* TODO Improve performance */
for (uint32 i = 0; i < VARBITBYTES(a); i++)
distance += pg_number_of_ones[ax[i] ^ bx[i]];
PG_RETURN_FLOAT8((double) distance);
}
/*
* Get the Jaccard distance between two bit vectors
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(jaccard_distance);
Datum
jaccard_distance(PG_FUNCTION_ARGS)
{
VarBit *a = PG_GETARG_VARBIT_P(0);
VarBit *b = PG_GETARG_VARBIT_P(1);
unsigned char *ax = VARBITS(a);
unsigned char *bx = VARBITS(b);
uint64 ab = 0;
uint64 aa;
uint64 bb;
CheckDims(a, b);
/* TODO Improve performance */
for (uint32 i = 0; i < VARBITBYTES(a); i++)
ab += pg_number_of_ones[ax[i] & bx[i]];
if (ab == 0)
PG_RETURN_FLOAT8(1);
aa = pg_popcount((char *) ax, VARBITBYTES(a));
bb = pg_popcount((char *) bx, VARBITBYTES(b));
PG_RETURN_FLOAT8(1 - (ab / ((double) (aa + bb - ab))));
}

View File

@@ -1,8 +0,0 @@
#ifndef BITVECTOR_H
#define BITVECTOR_H
#include "utils/varbit.h"
VarBit *InitBitVector(int dim);
#endif

File diff suppressed because it is too large Load Diff

View File

@@ -1,45 +0,0 @@
#ifndef HALFVEC_H
#define HALFVEC_H
#define __STDC_WANT_IEC_60559_TYPES_EXT__
#include <float.h>
#include "vector.h"
/* F16C has better performance than _Float16 (on x86-64) */
#if defined(__F16C__) || defined(_MSC_VER)
#define F16C_SUPPORT
#elif defined(__FLT16_MAX__)
#define FLT16_SUPPORT
#endif
#ifdef FLT16_SUPPORT
#define half _Float16
#define HALF_MAX FLT16_MAX
#else
#define half uint16
#define HALF_MAX 65504
#endif
#define HALFVEC_MAX_DIM VECTOR_MAX_DIM
#define HALFVEC_SIZE(_dim) (offsetof(HalfVector, x) + sizeof(half)*(_dim))
#define DatumGetHalfVector(x) ((HalfVector *) PG_DETOAST_DATUM(x))
#define PG_GETARG_HALFVEC_P(x) DatumGetHalfVector(PG_GETARG_DATUM(x))
#define PG_RETURN_HALFVEC_P(x) PG_RETURN_POINTER(x)
typedef struct HalfVector
{
int32 vl_len_; /* varlena header (do not touch directly!) */
int16 dim; /* number of dimensions */
int16 unused;
half x[FLEXIBLE_ARRAY_MEMBER];
} HalfVector;
HalfVector *InitHalfVector(int dim);
float HalfToFloat4(half num);
half Float4ToHalf(float num);
half Float4ToHalfUnchecked(float num);
#endif

View File

@@ -4,59 +4,24 @@
#include <math.h>
#include "access/amapi.h"
#include "access/reloptions.h"
#include "commands/progress.h"
#include "commands/vacuum.h"
#include "hnsw.h"
#include "miscadmin.h"
#include "utils/guc.h"
#include "utils/selfuncs.h"
#if PG_VERSION_NUM < 150000
#define MarkGUCPrefixReserved(x) EmitWarningsOnPlaceholders(x)
#if PG_VERSION_NUM >= 120000
#include "commands/progress.h"
#endif
int hnsw_ef_search;
int hnsw_lock_tranche_id;
static relopt_kind hnsw_relopt_kind;
/*
* Assign a tranche ID for our LWLocks. This only needs to be done by one
* backend, as the tranche ID is remembered in shared memory.
*
* This shared memory area is very small, so we just allocate it from the
* "slop" that PostgreSQL reserves for small allocations like this. If
* this grows bigger, we should use a shmem_request_hook and
* RequestAddinShmemSpace() to pre-reserve space for this.
*/
void
HnswInitLockTranche(void)
{
int *tranche_ids;
bool found;
LWLockAcquire(AddinShmemInitLock, LW_EXCLUSIVE);
tranche_ids = ShmemInitStruct("hnsw LWLock ids",
sizeof(int) * 1,
&found);
if (!found)
tranche_ids[0] = LWLockNewTrancheId();
hnsw_lock_tranche_id = tranche_ids[0];
LWLockRelease(AddinShmemInitLock);
/* Per-backend registration of the tranche ID */
LWLockRegisterTranche(hnsw_lock_tranche_id, "HnswBuild");
}
/*
* Initialize index options and variables
*/
void
HnswInit(void)
{
if (!process_shared_preload_libraries_in_progress)
HnswInitLockTranche();
hnsw_relopt_kind = add_reloption_kind();
add_int_reloption(hnsw_relopt_kind, "m", "Max number of connections",
HNSW_DEFAULT_M, HNSW_MIN_M, HNSW_MAX_M
@@ -72,15 +37,14 @@ HnswInit(void)
);
DefineCustomIntVariable("hnsw.ef_search", "Sets the size of the dynamic candidate list for search",
"Valid range is 1..1000.", &hnsw_ef_search,
"Valid range is 10..1000.", &hnsw_ef_search,
HNSW_DEFAULT_EF_SEARCH, HNSW_MIN_EF_SEARCH, HNSW_MAX_EF_SEARCH, PGC_USERSET, 0, NULL, NULL, NULL);
MarkGUCPrefixReserved("hnsw");
}
/*
* Get the name of index build phase
*/
#if PG_VERSION_NUM >= 120000
static char *
hnswbuildphasename(int64 phasenum)
{
@@ -94,6 +58,7 @@ hnswbuildphasename(int64 phasenum)
return NULL;
}
}
#endif
/*
* Estimate the cost of an index scan
@@ -108,6 +73,9 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
int m;
int entryLevel;
Relation index;
#if PG_VERSION_NUM < 120000
List *qinfos;
#endif
/* Never use index without order */
if (path->indexorderbys == NULL)
@@ -123,7 +91,7 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
MemSet(&costs, 0, sizeof(costs));
index = index_open(path->indexinfo->indexoid, NoLock);
HnswGetMetaPageInfo(index, &m, NULL);
m = HnswGetM(index);
index_close(index, NoLock);
/* Approximate entry level */
@@ -133,7 +101,12 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
/* Account for number of tuples (or entry level), m, and ef_search */
costs.numIndexTuples = (entryLevel + 2) * m;
#if PG_VERSION_NUM >= 120000
genericcostestimate(root, path, loop_count, &costs);
#else
qinfos = deconstruct_indexquals(path);
genericcostestimate(root, path, loop_count, qinfos, &costs);
#endif
/* Use total cost since most work happens before first tuple is returned */
*indexStartupCost = costs.indexTotalCost;
@@ -223,11 +196,13 @@ hnswhandler(PG_FUNCTION_ARGS)
amroutine->aminsert = hnswinsert;
amroutine->ambulkdelete = hnswbulkdelete;
amroutine->amvacuumcleanup = hnswvacuumcleanup;
amroutine->amcanreturn = NULL;
amroutine->amcanreturn = NULL; /* tuple not included in heapsort */
amroutine->amcostestimate = hnswcostestimate;
amroutine->amoptions = hnswoptions;
amroutine->amproperty = NULL; /* TODO AMPROP_DISTANCE_ORDERABLE */
#if PG_VERSION_NUM >= 120000
amroutine->ambuildphasename = hnswbuildphasename;
#endif
amroutine->amvalidate = hnswvalidate;
#if PG_VERSION_NUM >= 140000
amroutine->amadjustmembers = NULL;

View File

@@ -3,21 +3,18 @@
#include "postgres.h"
#include "access/genam.h"
#include "access/parallel.h"
#include "lib/pairingheap.h"
#include "access/generic_xlog.h"
#include "access/reloptions.h"
#include "nodes/execnodes.h"
#include "port.h" /* for random() */
#include "utils/relptr.h"
#include "utils/sampling.h"
#include "vector.h"
#if PG_VERSION_NUM < 120000
#error "Requires PostgreSQL 12+"
#if PG_VERSION_NUM < 110000
#error "Requires PostgreSQL 11+"
#endif
#define HNSW_MAX_DIM 2000
#define HNSW_MAX_NNZ 1000
/* Support functions */
#define HNSW_DISTANCE_PROC 1
@@ -25,151 +22,93 @@
#define HNSW_VERSION 1
#define HNSW_MAGIC_NUMBER 0xA953A953
#define HNSW_PAGE_ID 0xFF90
#define HNSW_PAGE_ID 0xFF85
/* Preserved page numbers */
#define HNSW_METAPAGE_BLKNO 0
#define HNSW_HEAD_BLKNO 1 /* first element page */
/* Must correspond to page numbers since page lock is used */
#define HNSW_UPDATE_LOCK 0
#define HNSW_SCAN_LOCK 1
/* HNSW parameters */
#define HNSW_DEFAULT_M 16
#define HNSW_MIN_M 2
#define HNSW_MIN_M 4
#define HNSW_MAX_M 100
#define HNSW_DEFAULT_EF_CONSTRUCTION 64
#define HNSW_MIN_EF_CONSTRUCTION 4
#define HNSW_DEFAULT_EF_CONSTRUCTION 40
#define HNSW_MIN_EF_CONSTRUCTION 10
#define HNSW_MAX_EF_CONSTRUCTION 1000
#define HNSW_DEFAULT_EF_SEARCH 40
#define HNSW_MIN_EF_SEARCH 1
#define HNSW_MIN_EF_SEARCH 10
#define HNSW_MAX_EF_SEARCH 1000
/* Tuple types */
#define HNSW_ELEMENT_TUPLE_TYPE 1
#define HNSW_NEIGHBOR_TUPLE_TYPE 2
/* Make graph robust against non-HOT updates */
#define HNSW_HEAPTIDS 10
#define HNSW_UPDATE_ENTRY_GREATER 1
#define HNSW_UPDATE_ENTRY_ALWAYS 2
typedef enum HnswType
{
HNSW_TYPE_VECTOR,
HNSW_TYPE_HALFVEC,
HNSW_TYPE_BIT,
HNSW_TYPE_SPARSEVEC
} HnswType;
/* Build phases */
/* PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE is 1 */
#define PROGRESS_HNSW_PHASE_LOAD 2
#define HNSW_MAX_SIZE (BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - sizeof(ItemIdData))
#define HNSW_TUPLE_ALLOC_SIZE BLCKSZ
#define HNSW_ELEMENT_TUPLE_SIZE(size) MAXALIGN(offsetof(HnswElementTupleData, data) + (size))
#define HNSW_NEIGHBOR_TUPLE_SIZE(level, m) MAXALIGN(offsetof(HnswNeighborTupleData, indextids) + ((level) + 2) * (m) * sizeof(ItemPointerData))
#define HNSW_NEIGHBOR_ARRAY_SIZE(lm) (offsetof(HnswNeighborArray, items) + sizeof(HnswCandidate) * (lm))
#define HNSW_ELEMENT_TUPLE_SIZE(_dim) MAXALIGN(offsetof(HnswElementTupleData, vec) + VECTOR_SIZE(_dim))
#define HNSW_NEIGHBOR_TUPLE_SIZE(level, m) MAXALIGN(offsetof(HnswNeighborTupleData, neighbors) + ((level) + 2) * (m) * sizeof(HnswNeighborTupleItem))
#define HnswPageGetOpaque(page) ((HnswPageOpaque) PageGetSpecialPointer(page))
#define HnswPageGetMeta(page) ((HnswMetaPageData *) PageGetContents(page))
#if PG_VERSION_NUM >= 150000
#define RandomDouble() pg_prng_double(&pg_global_prng_state)
#define SeedRandom(seed) pg_prng_seed(&pg_global_prng_state, seed)
#else
#define RandomDouble() (((double) random()) / MAX_RANDOM_VALUE)
#define SeedRandom(seed) srandom(seed)
#endif
#if PG_VERSION_NUM < 130000
#define list_delete_last(list) list_truncate(list, list_length(list) - 1)
#define list_sort(list, cmp) ((list) = list_qsort(list, cmp))
#define list_sort(list, cmp) list_qsort(list, cmp)
#endif
#define HnswIsElementTuple(tup) ((tup)->type == HNSW_ELEMENT_TUPLE_TYPE)
#define HnswIsNeighborTuple(tup) ((tup)->type == HNSW_NEIGHBOR_TUPLE_TYPE)
/* 2 * M connections for ground layer */
#define HnswGetLayerM(m, layer) (layer == 0 ? (m) * 2 : (m))
/* Optimal ML from paper */
#define HnswGetLayerM(m, layer) (layer == 0 ? m * 2 : m)
#define HnswGetMl(m) (1 / log(m))
/* Ensure fits on page and in uint8 */
#define HnswGetMaxLevel(m) Min(((BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - offsetof(HnswNeighborTupleData, indextids) - sizeof(ItemIdData)) / (sizeof(ItemPointerData)) / (m)) - 2, 255)
#define HnswGetValue(base, element) PointerGetDatum(HnswPtrAccess(base, (element)->value))
#if PG_VERSION_NUM < 140005
#define relptr_offset(rp) ((rp).relptr_off - 1)
#endif
/* Pointer macros */
#define HnswPtrAccess(base, hp) ((base) == NULL ? (hp).ptr : relptr_access(base, (hp).relptr))
#define HnswPtrStore(base, hp, value) ((base) == NULL ? (void) ((hp).ptr = (value)) : (void) relptr_store(base, (hp).relptr, value))
#define HnswPtrIsNull(base, hp) ((base) == NULL ? (hp).ptr == NULL : relptr_is_null((hp).relptr))
#define HnswPtrEqual(base, hp1, hp2) ((base) == NULL ? (hp1).ptr == (hp2).ptr : relptr_offset((hp1).relptr) == relptr_offset((hp2).relptr))
/* For code paths dedicated to each type */
#define HnswPtrPointer(hp) (hp).ptr
#define HnswPtrOffset(hp) relptr_offset((hp).relptr)
/* Variables */
extern int hnsw_ef_search;
extern int hnsw_lock_tranche_id;
typedef struct HnswElementData HnswElementData;
typedef struct HnswNeighborArray HnswNeighborArray;
#define HnswPtrDeclare(type, relptrtype, ptrtype) \
relptr_declare(type, relptrtype); \
typedef union { type *ptr; relptrtype relptr; } ptrtype;
/* Pointers that can be absolute or relative */
/* Use char for DatumPtr so works with Pointer */
HnswPtrDeclare(HnswElementData, HnswElementRelptr, HnswElementPtr);
HnswPtrDeclare(HnswNeighborArray, HnswNeighborArrayRelptr, HnswNeighborArrayPtr);
HnswPtrDeclare(HnswNeighborArrayPtr, HnswNeighborsRelptr, HnswNeighborsPtr);
HnswPtrDeclare(char, DatumRelptr, DatumPtr);
struct HnswElementData
typedef struct HnswElementData
{
HnswElementPtr next;
ItemPointerData heaptids[HNSW_HEAPTIDS];
uint8 heaptidsLength;
List *heaptids;
uint8 level;
uint8 deleted;
uint32 hash;
HnswNeighborsPtr neighbors;
HnswNeighborArray *neighbors;
BlockNumber blkno;
OffsetNumber offno;
OffsetNumber neighborOffno;
BlockNumber neighborPage;
DatumPtr value;
LWLock lock;
};
Vector *vec;
} HnswElementData;
typedef HnswElementData * HnswElement;
typedef struct HnswCandidate
{
HnswElementPtr element;
HnswElement element;
float distance;
bool closer;
} HnswCandidate;
struct HnswNeighborArray
typedef struct HnswNeighborArray
{
int length;
bool closerSet;
HnswCandidate items[FLEXIBLE_ARRAY_MEMBER];
};
HnswCandidate *items;
} HnswNeighborArray;
typedef struct HnswUpdate
{
HnswCandidate hc;
int level;
int index;
} HnswUpdate;
typedef struct HnswPairingHeapNode
{
@@ -185,65 +124,6 @@ typedef struct HnswOptions
int efConstruction; /* size of dynamic candidate list */
} HnswOptions;
typedef struct HnswGraph
{
/* Graph state */
slock_t lock;
HnswElementPtr head;
double indtuples;
/* Entry state */
LWLock entryLock;
LWLock entryWaitLock;
HnswElementPtr entryPoint;
/* Allocations state */
LWLock allocatorLock;
long memoryUsed;
long memoryTotal;
/* Flushed state */
LWLock flushLock;
bool flushed;
} HnswGraph;
typedef struct HnswShared
{
/* Immutable state */
Oid heaprelid;
Oid indexrelid;
bool isconcurrent;
/* Worker progress */
ConditionVariable workersdonecv;
/* Mutex for mutable state */
slock_t mutex;
/* Mutable state */
int nparticipantsdone;
double reltuples;
HnswGraph graphData;
} HnswShared;
#define ParallelTableScanFromHnswShared(shared) \
(ParallelTableScanDesc) ((char *) (shared) + BUFFERALIGN(sizeof(HnswShared)))
typedef struct HnswLeader
{
ParallelContext *pcxt;
int nparticipanttuplesorts;
HnswShared *hnswshared;
Snapshot snapshot;
char *hnswarea;
} HnswLeader;
typedef struct HnswAllocator
{
void *(*alloc) (Size size, void *state);
void *state;
} HnswAllocator;
typedef struct HnswBuildState
{
/* Info */
@@ -251,7 +131,6 @@ typedef struct HnswBuildState
Relation index;
IndexInfo *indexInfo;
ForkNumber forkNum;
HnswType type;
/* Settings */
int dimensions;
@@ -268,20 +147,16 @@ typedef struct HnswBuildState
Oid collation;
/* Variables */
HnswGraph graphData;
HnswGraph *graph;
List *elements;
HnswElement entryPoint;
double ml;
int maxLevel;
double maxInMemoryElements;
bool flushed;
Vector *normvec;
/* Memory */
MemoryContext graphCtx;
MemoryContext tmpCtx;
HnswAllocator allocator;
/* Parallel builds */
HnswLeader *hnswleader;
HnswShared *hnswshared;
char *hnswarea;
} HnswBuildState;
typedef struct HnswMetaPageData
@@ -317,17 +192,24 @@ typedef struct HnswElementTupleData
ItemPointerData heaptids[HNSW_HEAPTIDS];
ItemPointerData neighbortid;
uint16 unused2;
Vector data;
Vector vec;
} HnswElementTupleData;
typedef HnswElementTupleData * HnswElementTuple;
typedef struct HnswNeighborTupleItem
{
ItemPointerData indextid;
uint16 unused;
float distance; /* improves performance of inserts */
} HnswNeighborTupleItem;
typedef struct HnswNeighborTupleData
{
uint8 type;
uint8 unused;
uint16 count;
ItemPointerData indextids[FLEXIBLE_ARRAY_MEMBER];
HnswNeighborTupleItem neighbors[FLEXIBLE_ARRAY_MEMBER];
} HnswNeighborTupleData;
typedef HnswNeighborTupleData * HnswNeighborTuple;
@@ -335,6 +217,7 @@ typedef HnswNeighborTupleData * HnswNeighborTuple;
typedef struct HnswScanOpaqueData
{
bool first;
Buffer buf;
List *w;
MemoryContext tmpCtx;
@@ -363,7 +246,7 @@ typedef struct HnswVacuumState
Oid collation;
/* Variables */
struct tidhash_hash *deleted;
HTAB *deleted;
BufferAccessStrategy bas;
HnswNeighborTuple ntup;
HnswElementData highestPoint;
@@ -375,34 +258,26 @@ typedef struct HnswVacuumState
/* Methods */
int HnswGetM(Relation index);
int HnswGetEfConstruction(Relation index);
FmgrInfo *HnswOptionalProcInfo(Relation index, uint16 procnum);
HnswType HnswGetType(Relation index);
bool HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, HnswType type);
void HnswCheckValue(Datum value, HnswType type);
FmgrInfo *HnswOptionalProcInfo(Relation rel, uint16 procnum);
bool HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
void HnswCommitBuffer(Buffer buf, GenericXLogState *state);
Buffer HnswNewBuffer(Relation index, ForkNumber forkNum);
void HnswInitPage(Buffer buf, Page page);
void HnswInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state);
void HnswInit(void);
List *HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement);
List *HnswSearchLayer(Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, bool inserting, BlockNumber *skipPage, OffsetNumber *skipOffno);
HnswElement HnswGetEntryPoint(Relation index);
void HnswGetMetaPageInfo(Relation index, int *m, HnswElement * entryPoint);
void *HnswAlloc(HnswAllocator * allocator, Size size);
HnswElement HnswInitElement(char *base, ItemPointer tid, int m, double ml, int maxLevel, HnswAllocator * alloc);
HnswElement HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno);
void HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, bool existing);
HnswCandidate *HnswEntryCandidate(char *base, HnswElement em, Datum q, Relation rel, FmgrInfo *procinfo, Oid collation, bool loadVec);
void HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum, bool building);
void HnswSetNeighborTuple(char *base, HnswNeighborTuple ntup, HnswElement e, int m);
HnswElement HnswInitElement(ItemPointer tid, int m, double ml, int maxLevel);
void HnswFreeElement(HnswElement element);
HnswElement HnswInsertElement(HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, List **updates, bool vacuuming);
HnswCandidate *HnswEntryCandidate(HnswElement em, Datum q, Relation rel, FmgrInfo *procinfo, Oid collation, bool loadvec);
void HnswUpdateMetaPage(Relation index, bool updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum);
void HnswSetNeighborTuple(HnswNeighborTuple ntup, HnswElement e, int m);
void HnswAddHeapTid(HnswElement element, ItemPointer heaptid);
void HnswInitNeighbors(char *base, HnswElement element, int m, HnswAllocator * alloc);
bool HnswInsertTupleOnDisk(Relation index, Datum value, Datum *values, bool *isnull, ItemPointer heap_tid, bool building);
void HnswUpdateNeighborsOnDisk(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting, bool building);
void HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec);
void HnswInitNeighbors(HnswElement element, int m);
bool HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel);
void HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec);
void HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element);
void HnswUpdateConnection(char *base, HnswElement element, HnswCandidate * hc, int lm, int lc, int *updateIdx, Relation index, FmgrInfo *procinfo, Oid collation);
void HnswLoadNeighbors(HnswElement element, Relation index, int m);
void HnswInitLockTranche(void);
PGDLLEXPORT void HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc);
void HnswSetElementTuple(HnswElementTuple etup, HnswElement element);
/* Index access methods */
IndexBuildResult *hnswbuild(Relation heap, Relation index, IndexInfo *indexInfo);
@@ -420,54 +295,7 @@ void hnswrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys,
bool hnswgettuple(IndexScanDesc scan, ScanDirection dir);
void hnswendscan(IndexScanDesc scan);
static inline HnswNeighborArray *
HnswGetNeighbors(char *base, HnswElement element, int lc)
{
HnswNeighborArrayPtr *neighborList = HnswPtrAccess(base, element->neighbors);
Assert(element->level >= lc);
return HnswPtrAccess(base, neighborList[lc]);
}
/* Hash tables */
typedef struct TidHashEntry
{
ItemPointerData tid;
char status;
} TidHashEntry;
#define SH_PREFIX tidhash
#define SH_ELEMENT_TYPE TidHashEntry
#define SH_KEY_TYPE ItemPointerData
#define SH_SCOPE extern
#define SH_DECLARE
#include "lib/simplehash.h"
typedef struct PointerHashEntry
{
uintptr_t ptr;
char status;
} PointerHashEntry;
#define SH_PREFIX pointerhash
#define SH_ELEMENT_TYPE PointerHashEntry
#define SH_KEY_TYPE uintptr_t
#define SH_SCOPE extern
#define SH_DECLARE
#include "lib/simplehash.h"
typedef struct OffsetHashEntry
{
Size offset;
char status;
} OffsetHashEntry;
#define SH_PREFIX offsethash
#define SH_ELEMENT_TYPE OffsetHashEntry
#define SH_KEY_TYPE Size
#define SH_SCOPE extern
#define SH_DECLARE
#include "lib/simplehash.h"
/* Ensure fits in uint8 */
#define HnswGetMaxLevel(m) Min(((BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(HnswPageOpaqueData)) - offsetof(HnswNeighborTupleData, neighbors) - sizeof(ItemIdData)) / (sizeof(HnswNeighborTupleItem)) / m) - 2, 255)
#endif

File diff suppressed because it is too large Load Diff

View File

@@ -2,11 +2,9 @@
#include <math.h>
#include "access/generic_xlog.h"
#include "hnsw.h"
#include "storage/bufmgr.h"
#include "storage/lmgr.h"
#include "utils/datum.h"
#include "utils/memutils.h"
/*
@@ -36,7 +34,7 @@ GetInsertPage(Relation index)
* Check for a free offset
*/
static bool
HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size ntupSize, Buffer *nbuf, Page *npage, OffsetNumber *freeOffno, OffsetNumber *freeNeighborOffno, BlockNumber *newInsertPage)
HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size ntupSize, Buffer *nbuf, Page *npage, OffsetNumber *freeOffno, OffsetNumber *freeNeighborOffno, BlockNumber *firstFreePage)
{
OffsetNumber offno;
OffsetNumber maxoffno = PageGetMaxOffsetNumber(page);
@@ -51,15 +49,14 @@ HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size
if (etup->deleted)
{
BlockNumber elementPage = BufferGetBlockNumber(buf);
BlockNumber neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
OffsetNumber neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
ItemId itemid;
if (!BlockNumberIsValid(*newInsertPage))
*newInsertPage = elementPage;
if (!BlockNumberIsValid(*firstFreePage))
*firstFreePage = neighborPage;
if (neighborPage == elementPage)
if (neighborPage == BufferGetBlockNumber(buf))
{
*nbuf = buf;
*npage = page;
@@ -94,7 +91,7 @@ HnswFreeOffset(Relation index, Buffer buf, Page page, HnswElement element, Size
* Add a new page
*/
static void
HnswInsertAppendPage(Relation index, Buffer *nbuf, Page *npage, GenericXLogState *state, Page page, bool building)
HnswInsertAppendPage(Relation index, Buffer *nbuf, Page *npage, GenericXLogState *state, Page page)
{
/* Add a new page */
LockRelationForExtension(index, ExclusiveLock);
@@ -102,11 +99,7 @@ HnswInsertAppendPage(Relation index, Buffer *nbuf, Page *npage, GenericXLogState
UnlockRelationForExtension(index, ExclusiveLock);
/* Init new page */
if (building)
*npage = BufferGetPage(*nbuf);
else
*npage = GenericXLogRegisterBuffer(state, *nbuf, GENERIC_XLOG_FULL_IMAGE);
*npage = GenericXLogRegisterBuffer(state, *nbuf, GENERIC_XLOG_FULL_IMAGE);
HnswInitPage(*nbuf, *npage);
/* Update previous buffer */
@@ -117,7 +110,7 @@ HnswInsertAppendPage(Relation index, Buffer *nbuf, Page *npage, GenericXLogState
* Add to element and neighbor pages
*/
static void
AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, BlockNumber *updatedInsertPage, bool building)
WriteNewElementPages(Relation index, HnswElement e, int m)
{
Buffer buf;
Page page;
@@ -125,57 +118,40 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
Size etupSize;
Size ntupSize;
Size combinedSize;
Size maxSize;
Size minCombinedSize;
HnswElementTuple etup;
BlockNumber currentPage = insertPage;
BlockNumber insertPage = GetInsertPage(index);
BlockNumber originalInsertPage = insertPage;
int dimensions = e->vec->dim;
HnswNeighborTuple ntup;
Buffer nbuf;
Page npage;
OffsetNumber freeOffno = InvalidOffsetNumber;
OffsetNumber freeNeighborOffno = InvalidOffsetNumber;
BlockNumber newInsertPage = InvalidBlockNumber;
char *base = NULL;
BlockNumber firstFreePage = InvalidBlockNumber;
/* Calculate sizes */
etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(HnswPtrAccess(base, e->value)));
etupSize = HNSW_ELEMENT_TUPLE_SIZE(dimensions);
ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
maxSize = HNSW_MAX_SIZE;
minCombinedSize = etupSize + HNSW_NEIGHBOR_TUPLE_SIZE(0, m) + sizeof(ItemIdData);
/* Prepare element tuple */
etup = palloc0(etupSize);
HnswSetElementTuple(base, etup, e);
HnswSetElementTuple(etup, e);
/* Prepare neighbor tuple */
ntup = palloc0(ntupSize);
HnswSetNeighborTuple(base, ntup, e, m);
HnswSetNeighborTuple(ntup, e, m);
/* Find a page (or two if needed) to insert the tuples */
/* Find a page to insert the item */
for (;;)
{
buf = ReadBuffer(index, currentPage);
buf = ReadBuffer(index, insertPage);
LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
if (building)
{
state = NULL;
page = BufferGetPage(buf);
}
else
{
state = GenericXLogStart(index);
page = GenericXLogRegisterBuffer(state, buf, 0);
}
state = GenericXLogStart(index);
page = GenericXLogRegisterBuffer(state, buf, 0);
/* Keep track of first page where element at level 0 can fit */
if (!BlockNumberIsValid(newInsertPage) && PageGetFreeSpace(page) >= minCombinedSize)
newInsertPage = currentPage;
/* First, try the fastest path */
/* Space for both tuples on the current page */
/* This can split existing tuples in rare cases */
/* Space for both */
if (PageGetFreeSpace(page) >= combinedSize)
{
nbuf = buf;
@@ -183,35 +159,28 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
break;
}
/* Next, try space from a deleted element */
if (HnswFreeOffset(index, buf, page, e, ntupSize, &nbuf, &npage, &freeOffno, &freeNeighborOffno, &newInsertPage))
/* Space for element but not neighbors and last page */
if (PageGetFreeSpace(page) >= etupSize && !BlockNumberIsValid(HnswPageGetOpaque(page)->nextblkno))
{
HnswInsertAppendPage(index, &nbuf, &npage, state, page);
break;
}
/* Space from deleted item */
if (HnswFreeOffset(index, buf, page, e, ntupSize, &nbuf, &npage, &freeOffno, &freeNeighborOffno, &firstFreePage))
{
if (nbuf != buf)
{
if (building)
npage = BufferGetPage(nbuf);
else
npage = GenericXLogRegisterBuffer(state, nbuf, 0);
}
npage = GenericXLogRegisterBuffer(state, nbuf, 0);
break;
}
/* Finally, try space for element only if last page */
/* Skip if both tuples can fit on the same page */
if (combinedSize > maxSize && PageGetFreeSpace(page) >= etupSize && !BlockNumberIsValid(HnswPageGetOpaque(page)->nextblkno))
{
HnswInsertAppendPage(index, &nbuf, &npage, state, page, building);
break;
}
insertPage = HnswPageGetOpaque(page)->nextblkno;
currentPage = HnswPageGetOpaque(page)->nextblkno;
if (BlockNumberIsValid(currentPage))
if (BlockNumberIsValid(insertPage))
{
/* Move to next page */
if (!building)
GenericXLogAbort(state);
GenericXLogAbort(state);
UnlockReleaseBuffer(buf);
}
else
@@ -219,33 +188,24 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
Buffer newbuf;
Page newpage;
HnswInsertAppendPage(index, &newbuf, &newpage, state, page, building);
HnswInsertAppendPage(index, &newbuf, &newpage, state, page);
/* Commit */
if (building)
MarkBufferDirty(buf);
else
GenericXLogFinish(state);
MarkBufferDirty(newbuf);
MarkBufferDirty(buf);
GenericXLogFinish(state);
/* Unlock previous buffer */
UnlockReleaseBuffer(buf);
/* Prepare new buffer */
state = GenericXLogStart(index);
buf = newbuf;
if (building)
{
state = NULL;
page = BufferGetPage(buf);
}
else
{
state = GenericXLogStart(index);
page = GenericXLogRegisterBuffer(state, buf, 0);
}
page = GenericXLogRegisterBuffer(state, buf, 0);
/* Create new page for neighbors if needed */
if (PageGetFreeSpace(page) < combinedSize)
HnswInsertAppendPage(index, &nbuf, &npage, state, page, building);
HnswInsertAppendPage(index, &nbuf, &npage, state, page);
else
{
nbuf = buf;
@@ -259,10 +219,7 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
e->blkno = BufferGetBlockNumber(buf);
e->neighborPage = BufferGetBlockNumber(nbuf);
/* Added tuple to new page if newInsertPage is not set */
/* So can set to neighbor page instead of element page */
if (!BlockNumberIsValid(newInsertPage))
newInsertPage = e->neighborPage;
insertPage = e->neighborPage;
if (OffsetNumberIsValid(freeOffno))
{
@@ -299,144 +256,85 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
}
/* Commit */
if (building)
{
MarkBufferDirty(buf);
if (nbuf != buf)
MarkBufferDirty(nbuf);
}
else
GenericXLogFinish(state);
MarkBufferDirty(buf);
if (nbuf != buf)
MarkBufferDirty(nbuf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
if (nbuf != buf)
UnlockReleaseBuffer(nbuf);
/* Update the insert page */
if (BlockNumberIsValid(newInsertPage) && newInsertPage != insertPage)
*updatedInsertPage = newInsertPage;
if (insertPage != originalInsertPage && (!OffsetNumberIsValid(freeOffno) || firstFreePage == insertPage))
HnswUpdateMetaPage(index, false, NULL, insertPage, MAIN_FORKNUM);
}
/*
* Check if connection already exists
* Calculate index for update
*/
static bool
ConnectionExists(HnswElement e, HnswNeighborTuple ntup, int startIdx, int lm)
static int
HnswGetIndex(HnswUpdate * update, int m)
{
for (int i = 0; i < lm; i++)
{
ItemPointer indextid = &ntup->indextids[startIdx + i];
if (!ItemPointerIsValid(indextid))
break;
if (ItemPointerGetBlockNumber(indextid) == e->blkno && ItemPointerGetOffsetNumber(indextid) == e->offno)
return true;
}
return false;
return (update->hc.element->level - update->level) * m + update->index;
}
/*
* Update neighbors
*/
void
HnswUpdateNeighborsOnDisk(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement e, int m, bool checkExisting, bool building)
static void
UpdateNeighborPages(Relation index, HnswElement e, int m, List *updates)
{
char *base = NULL;
ListCell *lc;
for (int lc = e->level; lc >= 0; lc--)
/* Could update multiple at once for same element */
/* but should only happen a low percent of time, so keep simple for now */
foreach(lc, updates)
{
int lm = HnswGetLayerM(m, lc);
HnswNeighborArray *neighbors = HnswGetNeighbors(base, e, lc);
Buffer buf;
Page page;
GenericXLogState *state;
HnswUpdate *update = lfirst(lc);
ItemId itemid;
HnswNeighborTuple ntup;
Size ntupSize;
int idx;
OffsetNumber offno = update->hc.element->neighborOffno;
for (int i = 0; i < neighbors->length; i++)
/* Register page */
buf = ReadBuffer(index, update->hc.element->neighborPage);
LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
state = GenericXLogStart(index);
page = GenericXLogRegisterBuffer(state, buf, 0);
/* Get tuple */
itemid = PageGetItemId(page, offno);
ntup = (HnswNeighborTuple) PageGetItem(page, itemid);
ntupSize = ItemIdGetLength(itemid);
/* Calculate index */
idx = HnswGetIndex(update, m);
/* Make robust to issues */
if (idx < ntup->count)
{
HnswCandidate *hc = &neighbors->items[i];
Buffer buf;
Page page;
GenericXLogState *state;
HnswNeighborTuple ntup;
int idx = -1;
int startIdx;
HnswElement neighborElement = HnswPtrAccess(base, hc->element);
OffsetNumber offno = neighborElement->neighborOffno;
HnswNeighborTupleItem *neighbor = &ntup->neighbors[idx];
/* Get latest neighbors since they may have changed */
/* Do not lock yet since selecting neighbors can take time */
HnswLoadNeighbors(neighborElement, index, m);
/* Update neighbor */
ItemPointerSet(&neighbor->indextid, e->blkno, e->offno);
neighbor->distance = update->hc.distance;
/*
* Could improve performance for vacuuming by checking neighbors
* against list of elements being deleted to find index. It's
* important to exclude already deleted elements for this since
* they can be replaced at any time.
*/
/* Overwrite tuple */
if (!PageIndexTupleOverwrite(page, offno, (Item) ntup, ntupSize))
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Select neighbors */
HnswUpdateConnection(NULL, e, hc, lm, lc, &idx, index, procinfo, collation);
/* New element was not selected as a neighbor */
if (idx == -1)
continue;
/* Register page */
buf = ReadBuffer(index, neighborElement->neighborPage);
LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
if (building)
{
state = NULL;
page = BufferGetPage(buf);
}
else
{
state = GenericXLogStart(index);
page = GenericXLogRegisterBuffer(state, buf, 0);
}
/* Get tuple */
ntup = (HnswNeighborTuple) PageGetItem(page, PageGetItemId(page, offno));
/* Calculate index for update */
startIdx = (neighborElement->level - lc) * m;
/* Check for existing connection */
if (checkExisting && ConnectionExists(e, ntup, startIdx, lm))
idx = -1;
else if (idx == -2)
{
/* Find free offset if still exists */
/* TODO Retry updating connections if not */
for (int j = 0; j < lm; j++)
{
if (!ItemPointerIsValid(&ntup->indextids[startIdx + j]))
{
idx = startIdx + j;
break;
}
}
}
else
idx += startIdx;
/* Make robust to issues */
if (idx >= 0 && idx < ntup->count)
{
ItemPointer indextid = &ntup->indextids[idx];
/* Update neighbor on the buffer */
ItemPointerSet(indextid, e->blkno, e->offno);
/* Commit */
if (building)
MarkBufferDirty(buf);
else
GenericXLogFinish(state);
}
else if (!building)
GenericXLogAbort(state);
UnlockReleaseBuffer(buf);
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
}
else
GenericXLogAbort(state);
UnlockReleaseBuffer(buf);
}
}
@@ -444,27 +342,20 @@ HnswUpdateNeighborsOnDisk(Relation index, FmgrInfo *procinfo, Oid collation, Hns
* Add a heap TID to an existing element
*/
static bool
AddDuplicateOnDisk(Relation index, HnswElement element, HnswElement dup, bool building)
HnswAddDuplicate(Relation index, HnswElement element, HnswElement dup)
{
Buffer buf;
Page page;
GenericXLogState *state;
Size etupSize = HNSW_ELEMENT_TUPLE_SIZE(dup->vec->dim);
HnswElementTuple etup;
int i;
/* Read page */
buf = ReadBuffer(index, dup->blkno);
LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
if (building)
{
state = NULL;
page = BufferGetPage(buf);
}
else
{
state = GenericXLogStart(index);
page = GenericXLogRegisterBuffer(state, buf, 0);
}
state = GenericXLogStart(index);
page = GenericXLogRegisterBuffer(state, buf, 0);
/* Find space */
etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, dup->offno));
@@ -477,160 +368,91 @@ AddDuplicateOnDisk(Relation index, HnswElement element, HnswElement dup, bool bu
/* Either being deleted or we lost our chance to another backend */
if (i == 0 || i == HNSW_HEAPTIDS)
{
if (!building)
GenericXLogAbort(state);
GenericXLogAbort(state);
UnlockReleaseBuffer(buf);
return false;
}
/* Add heap TID, modifying the tuple on the page directly */
etup->heaptids[i] = element->heaptids[0];
/* Add heap TID */
etup->heaptids[i] = *((ItemPointer) linitial(element->heaptids));
/* Overwrite tuple */
if (!PageIndexTupleOverwrite(page, dup->offno, (Item) etup, etupSize))
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
if (building)
MarkBufferDirty(buf);
else
GenericXLogFinish(state);
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
return true;
}
/*
* Find duplicate element
*/
static bool
FindDuplicateOnDisk(Relation index, HnswElement element, bool building)
{
char *base = NULL;
HnswNeighborArray *neighbors = HnswGetNeighbors(base, element, 0);
Datum value = HnswGetValue(base, element);
for (int i = 0; i < neighbors->length; i++)
{
HnswCandidate *neighbor = &neighbors->items[i];
HnswElement neighborElement = HnswPtrAccess(base, neighbor->element);
Datum neighborValue = HnswGetValue(base, neighborElement);
/* Exit early since ordered by distance */
if (!datumIsEqual(value, neighborValue, false, -1))
return false;
if (AddDuplicateOnDisk(index, element, neighborElement, building))
return true;
}
return false;
}
/*
* Update graph on disk
* Write changes to disk
*/
static void
UpdateGraphOnDisk(Relation index, FmgrInfo *procinfo, Oid collation, HnswElement element, int m, int efConstruction, HnswElement entryPoint, bool building)
WriteElement(Relation index, HnswElement element, int m, List *updates, HnswElement dup, HnswElement entryPoint)
{
BlockNumber newInsertPage = InvalidBlockNumber;
/* Try to add to existing page */
if (dup != NULL)
{
if (HnswAddDuplicate(index, element, dup))
return;
}
/* Look for duplicate */
if (FindDuplicateOnDisk(index, element, building))
return;
/* If fails, take this path */
WriteNewElementPages(index, element, m);
UpdateNeighborPages(index, element, m, updates);
/* Add element */
AddElementOnDisk(index, element, m, GetInsertPage(index), &newInsertPage, building);
/* Update insert page if needed */
if (BlockNumberIsValid(newInsertPage))
HnswUpdateMetaPage(index, 0, NULL, newInsertPage, MAIN_FORKNUM, building);
/* Update neighbors */
HnswUpdateNeighborsOnDisk(index, procinfo, collation, element, m, false, building);
/* Update entry point if needed */
/* Update metapage if needed */
if (entryPoint == NULL || element->level > entryPoint->level)
HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_GREATER, element, InvalidBlockNumber, MAIN_FORKNUM, building);
HnswUpdateMetaPage(index, true, element, InvalidBlockNumber, MAIN_FORKNUM);
}
/*
* Insert a tuple into the index
*/
bool
HnswInsertTupleOnDisk(Relation index, Datum value, Datum *values, bool *isnull, ItemPointer heap_tid, bool building)
{
HnswElement entryPoint;
HnswElement element;
int m;
int efConstruction = HnswGetEfConstruction(index);
FmgrInfo *procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
Oid collation = index->rd_indcollation[0];
LOCKMODE lockmode = ShareLock;
char *base = NULL;
/*
* Get a shared lock. This allows vacuum to ensure no in-flight inserts
* before repairing graph. Use a page lock so it does not interfere with
* buffer lock (or reads when vacuuming).
*/
LockPage(index, HNSW_UPDATE_LOCK, lockmode);
/* Get m and entry point */
HnswGetMetaPageInfo(index, &m, &entryPoint);
/* Create an element */
element = HnswInitElement(base, heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m), NULL);
HnswPtrStore(base, element->value, DatumGetPointer(value));
/* Prevent concurrent inserts when likely updating entry point */
if (entryPoint == NULL || element->level > entryPoint->level)
{
/* Release shared lock */
UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
/* Get exclusive lock */
lockmode = ExclusiveLock;
LockPage(index, HNSW_UPDATE_LOCK, lockmode);
/* Get latest entry point after lock is acquired */
entryPoint = HnswGetEntryPoint(index);
}
/* Find neighbors for element */
HnswFindElementNeighbors(base, element, entryPoint, index, procinfo, collation, m, efConstruction, false);
/* Update graph on disk */
UpdateGraphOnDisk(index, procinfo, collation, element, m, efConstruction, entryPoint, building);
/* Release lock */
UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
return true;
}
/*
* Insert a tuple into the index
*/
static void
HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid)
HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel)
{
Datum value;
FmgrInfo *normprocinfo;
HnswElement entryPoint;
HnswElement element;
int m = HnswGetM(index);
int efConstruction = HnswGetEfConstruction(index);
double ml = HnswGetMl(m);
FmgrInfo *procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
Oid collation = index->rd_indcollation[0];
HnswType type = HnswGetType(index);
List *updates = NIL;
HnswElement dup;
/* Detoast once for all calls */
value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
/* Check value */
HnswCheckValue(value, type);
/* Normalize if needed */
normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
if (normprocinfo != NULL)
{
if (!HnswNormValue(normprocinfo, collation, &value, type))
return;
if (!HnswNormValue(normprocinfo, collation, &value, NULL))
return false;
}
HnswInsertTupleOnDisk(index, value, values, isnull, heap_tid, false);
/* Create an element */
element = HnswInitElement(heap_tid, m, ml, HnswGetMaxLevel(m));
element->vec = DatumGetVector(value);
/* Get entry point */
entryPoint = HnswGetEntryPoint(index);
/* Insert element in graph */
dup = HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, &updates, false);
/* Write to disk */
WriteElement(index, element, m, updates, dup, entryPoint);
return true;
}
/*
@@ -659,7 +481,7 @@ hnswinsert(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid,
oldCtx = MemoryContextSwitchTo(insertCtx);
/* Insert tuple */
HnswInsertTuple(index, values, isnull, heap_tid);
HnswInsertTuple(index, values, isnull, heap_tid, heap);
/* Delete memory context */
MemoryContextSwitchTo(oldCtx);

View File

@@ -4,67 +4,57 @@
#include "hnsw.h"
#include "pgstat.h"
#include "storage/bufmgr.h"
#include "storage/lmgr.h"
#include "utils/memutils.h"
/*
* Algorithm 5 from paper
*/
static List *
static void
GetScanItems(IndexScanDesc scan, Datum q)
{
HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
Relation index = scan->indexRelation;
FmgrInfo *procinfo = so->procinfo;
Oid collation = so->collation;
List *ep;
List *ep = NIL;
List *w;
int m;
HnswElement entryPoint;
char *base = NULL;
/* Get m and entry point */
HnswGetMetaPageInfo(index, &m, &entryPoint);
HnswElement entryPoint = HnswGetEntryPoint(index);
if (entryPoint == NULL)
return NIL;
return;
ep = list_make1(HnswEntryCandidate(base, entryPoint, q, index, procinfo, collation, false));
ep = lappend(ep, HnswEntryCandidate(entryPoint, q, index, procinfo, collation, false));
for (int lc = entryPoint->level; lc >= 1; lc--)
{
w = HnswSearchLayer(base, q, ep, 1, lc, index, procinfo, collation, m, false, NULL);
w = HnswSearchLayer(q, ep, 1, lc, index, procinfo, collation, false, NULL, NULL);
ep = w;
}
return HnswSearchLayer(base, q, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL);
so->w = HnswSearchLayer(q, ep, hnsw_ef_search, 0, index, procinfo, collation, false, NULL, NULL);
}
/*
* Get scan value
* Get dimensions from metapage
*/
static Datum
GetScanValue(IndexScanDesc scan)
static int
GetDimensions(Relation index)
{
HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
Datum value;
Buffer buf;
Page page;
HnswMetaPage metap;
int dimensions;
if (scan->orderByData->sk_flags & SK_ISNULL)
value = PointerGetDatum(NULL);
else
{
value = scan->orderByData->sk_argument;
buf = ReadBuffer(index, HNSW_METAPAGE_BLKNO);
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
metap = HnswPageGetMeta(page);
/* Value should not be compressed or toasted */
Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
dimensions = metap->dimensions;
/* Fine if normalization fails */
if (so->normprocinfo != NULL)
HnswNormValue(so->normprocinfo, so->collation, &value, HnswGetType(scan->indexRelation));
}
UnlockReleaseBuffer(buf);
return value;
return dimensions;
}
/*
@@ -79,6 +69,7 @@ hnswbeginscan(Relation index, int nkeys, int norderbys)
scan = RelationGetIndexScan(index, nkeys, norderbys);
so = (HnswScanOpaque) palloc(sizeof(HnswScanOpaqueData));
so->buf = InvalidBuffer;
so->first = true;
so->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
"Hnsw scan temporary context",
@@ -138,52 +129,62 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
if (scan->orderByData == NULL)
elog(ERROR, "cannot scan hnsw index without order");
/* Requires MVCC-compliant snapshot as not able to maintain a pin */
/* https://www.postgresql.org/docs/current/index-locking.html */
if (!IsMVCCSnapshot(scan->xs_snapshot))
elog(ERROR, "non-MVCC snapshots are not supported with hnsw");
if (scan->orderByData->sk_flags & SK_ISNULL)
value = PointerGetDatum(InitVector(GetDimensions(scan->indexRelation)));
else
{
value = scan->orderByData->sk_argument;
/* Get scan value */
value = GetScanValue(scan);
/* Value should not be compressed or toasted */
Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
/*
* Get a shared lock. This allows vacuum to ensure no in-flight scans
* before marking tuples as deleted.
*/
LockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
so->w = GetScanItems(scan, value);
/* Release shared lock */
UnlockPage(scan->indexRelation, HNSW_SCAN_LOCK, ShareLock);
/* Fine if normalization fails */
if (so->normprocinfo != NULL)
HnswNormValue(so->normprocinfo, so->collation, &value, NULL);
}
GetScanItems(scan, value);
so->first = false;
#if defined(HNSW_MEMORY) && PG_VERSION_NUM >= 130000
elog(INFO, "memory: %zu MB", MemoryContextMemAllocated(so->tmpCtx, false) / (1024 * 1024));
#endif
}
while (list_length(so->w) > 0)
{
char *base = NULL;
HnswCandidate *hc = llast(so->w);
HnswElement element = HnswPtrAccess(base, hc->element);
ItemPointer heaptid;
ItemPointer tid;
BlockNumber indexblkno;
/* Move to next element if no valid heap TIDs */
if (element->heaptidsLength == 0)
/* Move to next element if no valid heap tids */
if (list_length(hc->element->heaptids) == 0)
{
so->w = list_delete_last(so->w);
continue;
}
heaptid = &element->heaptids[--element->heaptidsLength];
tid = llast(hc->element->heaptids);
indexblkno = hc->element->blkno;
hc->element->heaptids = list_delete_last(hc->element->heaptids);
MemoryContextSwitchTo(oldCtx);
scan->xs_heaptid = *heaptid;
scan->xs_recheck = false;
#if PG_VERSION_NUM >= 120000
scan->xs_heaptid = *tid;
#else
scan->xs_ctup.t_self = *tid;
#endif
if (BufferIsValid(so->buf))
ReleaseBuffer(so->buf);
/*
* An index scan must maintain a pin on the index page holding the
* item last returned by amgettuple
*
* https://www.postgresql.org/docs/current/index-locking.html
*/
so->buf = ReadBuffer(scan->indexRelation, indexblkno);
scan->xs_recheckorderby = false;
return true;
}
@@ -200,6 +201,10 @@ hnswendscan(IndexScanDesc scan)
{
HnswScanOpaque so = (HnswScanOpaque) scan->opaque;
/* Release pin */
if (BufferIsValid(so->buf))
ReleaseBuffer(so->buf);
MemoryContextDelete(so->tmpCtx);
pfree(so);

File diff suppressed because it is too large Load Diff

View File

@@ -2,20 +2,21 @@
#include <math.h>
#include "access/generic_xlog.h"
#include "commands/vacuum.h"
#include "hnsw.h"
#include "storage/bufmgr.h"
#include "storage/lmgr.h"
#include "utils/memutils.h"
/*
* Check if deleted list contains an index TID
* Check if deleted list contains an index tid
*/
static bool
DeletedContains(tidhash_hash * deleted, ItemPointer indextid)
DeletedContains(HTAB *deleted, ItemPointer indextid)
{
return tidhash_lookup(deleted, *indextid) != NULL;
bool found;
hash_search(deleted, indextid, HASH_FIND, &found);
return found;
}
/*
@@ -31,7 +32,6 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
Relation index = vacuumstate->index;
BufferAccessStrategy bas = vacuumstate->bas;
HnswElement entryPoint = HnswGetEntryPoint(vacuumstate->index);
IndexBulkDeleteResult *stats = vacuumstate->stats;
/* Store separately since highestPoint.level is uint8 */
int highestLevel = -1;
@@ -77,24 +77,25 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
break;
if (vacuumstate->callback(&etup->heaptids[i], vacuumstate->callback_state))
{
itemUpdated = true;
stats->tuples_removed++;
}
else
{
/* Move to front of list */
etup->heaptids[idx++] = etup->heaptids[i];
stats->num_index_tuples++;
}
}
if (itemUpdated)
{
Size etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
/* Mark rest as invalid */
for (int i = idx; i < HNSW_HEAPTIDS; i++)
ItemPointerSetInvalid(&etup->heaptids[i]);
if (!PageIndexTupleOverwrite(page, offno, (Item) etup, etupSize))
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
updated = true;
}
}
@@ -102,15 +103,13 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
if (!ItemPointerIsValid(&etup->heaptids[0]))
{
ItemPointerData ip;
bool found;
/* Add to deleted list */
ItemPointerSet(&ip, blkno, offno);
tidhash_insert(vacuumstate->deleted, ip, &found);
Assert(!found);
(void) hash_search(vacuumstate->deleted, &ip, HASH_ENTER, NULL);
}
else if (etup->level > highestLevel && !(entryPoint != NULL && blkno == entryPoint->blkno && offno == entryPoint->offno))
else if (etup->level > highestLevel && !(blkno == entryPoint->blkno && offno == entryPoint->offno))
{
/* Keep track of highest non-entry point */
highestPoint->blkno = blkno;
@@ -123,7 +122,10 @@ RemoveHeapTids(HnswVacuumState * vacuumstate)
blkno = HnswPageGetOpaque(page)->nextblkno;
if (updated)
{
MarkBufferDirty(buf);
GenericXLogFinish(state);
}
else
GenericXLogAbort(state);
@@ -154,24 +156,19 @@ NeedsUpdated(HnswVacuumState * vacuumstate, HnswElement element)
/* Check neighbors */
for (int i = 0; i < ntup->count; i++)
{
ItemPointer indextid = &ntup->indextids[i];
HnswNeighborTupleItem *neighbor = &ntup->neighbors[i];
if (!ItemPointerIsValid(indextid))
if (!ItemPointerIsValid(&neighbor->indextid))
continue;
/* Check if in deleted list */
if (DeletedContains(vacuumstate->deleted, indextid))
if (DeletedContains(vacuumstate->deleted, &neighbor->indextid))
{
needsUpdated = true;
break;
}
}
/* Also update if layer 0 is not full */
/* This could indicate too many candidates being deleted during insert */
if (!needsUpdated)
needsUpdated = !ItemPointerIsValid(&ntup->indextids[ntup->count - 1]);
UnlockReleaseBuffer(buf);
return needsUpdated;
@@ -181,7 +178,7 @@ NeedsUpdated(HnswVacuumState * vacuumstate, HnswElement element)
* Repair graph for a single element
*/
static void
RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswElement entryPoint)
RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element)
{
Relation index = vacuumstate->index;
Buffer buf;
@@ -191,28 +188,46 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
int efConstruction = vacuumstate->efConstruction;
FmgrInfo *procinfo = vacuumstate->procinfo;
Oid collation = vacuumstate->collation;
HnswElement entryPoint;
BufferAccessStrategy bas = vacuumstate->bas;
HnswNeighborTuple ntup = vacuumstate->ntup;
Size ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, m);
char *base = NULL;
/* Skip if element is entry point */
if (entryPoint != NULL && element->blkno == entryPoint->blkno && element->offno == entryPoint->offno)
/* Check if any neighbors point to deleted values */
if (!NeedsUpdated(vacuumstate, element))
return;
/* Refresh entry point for each element */
entryPoint = HnswGetEntryPoint(index);
/* Special case for entry point */
if (element->blkno == entryPoint->blkno && element->offno == entryPoint->offno)
{
if (BlockNumberIsValid(vacuumstate->highestPoint.blkno))
{
/* Already updated */
if (vacuumstate->highestPoint.blkno == element->blkno && vacuumstate->highestPoint.offno == element->offno)
return;
entryPoint = &vacuumstate->highestPoint;
/* Reset neighbors from previous update */
entryPoint->neighbors = NULL;
}
else
entryPoint = NULL;
}
/* Init fields */
HnswInitNeighbors(base, element, m, NULL);
element->heaptidsLength = 0;
HnswInitNeighbors(element, m);
element->heaptids = NIL;
/* Find neighbors for element, skipping itself */
HnswFindElementNeighbors(base, element, entryPoint, index, procinfo, collation, m, efConstruction, true);
/* Zero memory for each element */
MemSet(ntup, 0, HNSW_TUPLE_ALLOC_SIZE);
/* Add element to graph, skipping itself */
HnswInsertElement(element, entryPoint, index, procinfo, collation, m, efConstruction, NULL, true);
/* Update neighbor tuple */
/* Do this before getting page to minimize locking */
HnswSetNeighborTuple(base, ntup, element, m);
HnswSetNeighborTuple(ntup, element, m);
/* Get neighbor page */
buf = ReadBufferExtended(index, MAIN_FORKNUM, element->neighborPage, RBM_NORMAL, bas);
@@ -225,11 +240,9 @@ RepairGraphElement(HnswVacuumState * vacuumstate, HnswElement element, HnswEleme
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
/* Update neighbors */
HnswUpdateNeighborsOnDisk(index, procinfo, collation, element, m, true, false);
}
/*
@@ -243,35 +256,16 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
HnswElement entryPoint;
MemoryContext oldCtx = MemoryContextSwitchTo(vacuumstate->tmpCtx);
if (!BlockNumberIsValid(highestPoint->blkno))
highestPoint = NULL;
/*
* Repair graph for highest non-entry point. Highest point may be outdated
* due to inserts that happen during and after RemoveHeapTids.
*/
if (highestPoint != NULL)
/* Repair graph for highest non-entry point */
/* This may not be the highest with new inserts, but should be fine */
if (BlockNumberIsValid(highestPoint->blkno))
{
/* Get a shared lock */
LockPage(index, HNSW_UPDATE_LOCK, ShareLock);
/* Load element */
HnswLoadElement(highestPoint, NULL, NULL, index, vacuumstate->procinfo, vacuumstate->collation, true);
/* Repair if needed */
if (NeedsUpdated(vacuumstate, highestPoint))
RepairGraphElement(vacuumstate, highestPoint, HnswGetEntryPoint(index));
/* Release lock */
UnlockPage(index, HNSW_UPDATE_LOCK, ShareLock);
RepairGraphElement(vacuumstate, highestPoint);
}
/* Prevent concurrent inserts when possibly updating entry point */
LockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
/* Get latest entry point */
/* See if entry point needs updated */
entryPoint = HnswGetEntryPoint(index);
if (entryPoint != NULL)
{
ItemPointerData epData;
@@ -279,37 +273,15 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
ItemPointerSet(&epData, entryPoint->blkno, entryPoint->offno);
if (DeletedContains(vacuumstate->deleted, &epData))
{
/*
* Replace the entry point with the highest point. If highest
* point is outdated and empty, the entry point will be empty
* until an element is repaired.
*/
HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_ALWAYS, highestPoint, InvalidBlockNumber, MAIN_FORKNUM, false);
}
HnswUpdateMetaPage(index, true, highestPoint, InvalidBlockNumber, MAIN_FORKNUM);
else
{
/*
* Repair the entry point with the highest point. If highest point
* is outdated, this can remove connections at higher levels in
* the graph until they are repaired, but this should be fine.
*/
/* Highest point will be used to repair */
HnswLoadElement(entryPoint, NULL, NULL, index, vacuumstate->procinfo, vacuumstate->collation, true);
if (NeedsUpdated(vacuumstate, entryPoint))
{
/* Reset neighbors from previous update */
if (highestPoint != NULL)
HnswPtrStore((char *) NULL, highestPoint->neighbors, (HnswNeighborArrayPtr *) NULL);
RepairGraphElement(vacuumstate, entryPoint, highestPoint);
}
RepairGraphElement(vacuumstate, entryPoint);
}
}
/* Release lock */
UnlockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
/* Reset memory context */
MemoryContextSwitchTo(oldCtx);
MemoryContextReset(vacuumstate->tmpCtx);
@@ -325,14 +297,6 @@ RepairGraph(HnswVacuumState * vacuumstate)
BufferAccessStrategy bas = vacuumstate->bas;
BlockNumber blkno = HNSW_HEAD_BLKNO;
/*
* Wait for inserts to complete. Inserts before this point may have
* neighbors about to be deleted. Inserts after this point will not.
*/
LockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
UnlockPage(index, HNSW_UPDATE_LOCK, ExclusiveLock);
/* Repair entry point first */
RepairGraphEntryPoint(vacuumstate);
while (BlockNumberIsValid(blkno))
@@ -369,8 +333,14 @@ RepairGraph(HnswVacuumState * vacuumstate)
continue;
/* Create an element */
element = HnswInitElementFromBlock(blkno, offno);
HnswLoadElementFromTuple(element, etup, false, true);
element = palloc(sizeof(HnswElementData));
element->neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
element->neighborOffno = ItemPointerGetOffsetNumber(&etup->neighbortid);
element->level = etup->level;
element->blkno = blkno;
element->offno = offno;
element->vec = palloc(VECTOR_SIZE(etup->vec.dim));
memcpy(element->vec, &etup->vec, VECTOR_SIZE(etup->vec.dim));
elements = lappend(elements, element);
}
@@ -381,48 +351,7 @@ RepairGraph(HnswVacuumState * vacuumstate)
/* Update neighbor pages */
foreach(lc2, elements)
{
HnswElement element = (HnswElement) lfirst(lc2);
HnswElement entryPoint;
LOCKMODE lockmode = ShareLock;
/* Check if any neighbors point to deleted values */
if (!NeedsUpdated(vacuumstate, element))
continue;
/* Get a shared lock */
LockPage(index, HNSW_UPDATE_LOCK, lockmode);
/* Refresh entry point for each element */
entryPoint = HnswGetEntryPoint(index);
/* Prevent concurrent inserts when likely updating entry point */
if (entryPoint == NULL || element->level > entryPoint->level)
{
/* Release shared lock */
UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
/* Get exclusive lock */
lockmode = ExclusiveLock;
LockPage(index, HNSW_UPDATE_LOCK, lockmode);
/* Get latest entry point after lock is acquired */
entryPoint = HnswGetEntryPoint(index);
}
/* Repair connections */
RepairGraphElement(vacuumstate, element, entryPoint);
/*
* Update metapage if needed. Should only happen if entry point
* was replaced and highest point was outdated.
*/
if (entryPoint == NULL || element->level > entryPoint->level)
HnswUpdateMetaPage(index, HNSW_UPDATE_ENTRY_GREATER, element, InvalidBlockNumber, MAIN_FORKNUM, false);
/* Release lock */
UnlockPage(index, HNSW_UPDATE_LOCK, lockmode);
}
RepairGraphElement(vacuumstate, (HnswElement) lfirst(lc2));
/* Reset memory context */
MemoryContextSwitchTo(oldCtx);
@@ -440,14 +369,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
BlockNumber insertPage = InvalidBlockNumber;
Relation index = vacuumstate->index;
BufferAccessStrategy bas = vacuumstate->bas;
/*
* Wait for index scans to complete. Scans before this point may contain
* tuples about to be deleted. Scans after this point will not, since the
* graph has been repaired.
*/
LockPage(index, HNSW_SCAN_LOCK, ExclusiveLock);
UnlockPage(index, HNSW_SCAN_LOCK, ExclusiveLock);
IndexBulkDeleteResult *stats = vacuumstate->stats;
while (BlockNumberIsValid(blkno))
{
@@ -478,6 +400,8 @@ MarkDeleted(HnswVacuumState * vacuumstate)
{
HnswElementTuple etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
HnswNeighborTuple ntup;
Size etupSize;
Size ntupSize;
Buffer nbuf;
Page npage;
BlockNumber neighborPage;
@@ -489,17 +413,21 @@ MarkDeleted(HnswVacuumState * vacuumstate)
/* Skip deleted tuples */
if (etup->deleted)
{
/* Set to first free page */
if (!BlockNumberIsValid(insertPage))
insertPage = blkno;
continue;
}
/* Skip live tuples */
if (ItemPointerIsValid(&etup->heaptids[0]))
{
stats->num_index_tuples++;
continue;
}
/* Update stats */
stats->tuples_removed++;
/* Calculate sizes */
etupSize = HNSW_ELEMENT_TUPLE_SIZE(etup->vec.dim);
ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(etup->level, vacuumstate->m);
/* Get neighbor page */
neighborPage = ItemPointerGetBlockNumber(&etup->neighbortid);
@@ -521,18 +449,27 @@ MarkDeleted(HnswVacuumState * vacuumstate)
/* Overwrite element */
etup->deleted = 1;
MemSet(&etup->data, 0, VARSIZE_ANY(&etup->data));
MemSet(&etup->vec.x, 0, etup->vec.dim * sizeof(float));
/* Overwrite neighbors */
for (int i = 0; i < ntup->count; i++)
ItemPointerSetInvalid(&ntup->indextids[i]);
{
ItemPointerSetInvalid(&ntup->neighbors[i].indextid);
ntup->neighbors[i].distance = NAN;
}
/*
* We modified the tuples in place, no need to call
* PageIndexTupleOverwrite
*/
/* Overwrite element tuple */
if (!PageIndexTupleOverwrite(page, offno, (Item) etup, etupSize))
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Overwrite neighbor tuple */
if (!PageIndexTupleOverwrite(npage, neighborOffno, (Item) ntup, ntupSize))
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
/* Commit */
MarkBufferDirty(buf);
if (nbuf != buf)
MarkBufferDirty(nbuf);
GenericXLogFinish(state);
if (nbuf != buf)
UnlockReleaseBuffer(nbuf);
@@ -552,8 +489,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
UnlockReleaseBuffer(buf);
}
/* Update insert page last, after everything has been marked as deleted */
HnswUpdateMetaPage(index, 0, NULL, insertPage, MAIN_FORKNUM, false);
HnswUpdateMetaPage(index, false, NULL, insertPage, MAIN_FORKNUM);
}
/*
@@ -563,6 +499,7 @@ static void
InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkDeleteResult *stats, IndexBulkDeleteCallback callback, void *callback_state)
{
Relation index = info->index;
HASHCTL hash_ctl;
if (stats == NULL)
stats = (IndexBulkDeleteResult *) palloc0(sizeof(IndexBulkDeleteResult));
@@ -571,20 +508,21 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
vacuumstate->stats = stats;
vacuumstate->callback = callback;
vacuumstate->callback_state = callback_state;
vacuumstate->m = HnswGetM(index);
vacuumstate->efConstruction = HnswGetEfConstruction(index);
vacuumstate->bas = GetAccessStrategy(BAS_BULKREAD);
vacuumstate->procinfo = index_getprocinfo(index, 1, HNSW_DISTANCE_PROC);
vacuumstate->collation = index->rd_indcollation[0];
vacuumstate->ntup = palloc0(HNSW_TUPLE_ALLOC_SIZE);
vacuumstate->ntup = palloc0(BLCKSZ);
vacuumstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
"Hnsw vacuum temporary context",
ALLOCSET_DEFAULT_SIZES);
/* Get m from metapage */
HnswGetMetaPageInfo(index, &vacuumstate->m, NULL);
/* Create hash table */
vacuumstate->deleted = tidhash_create(CurrentMemoryContext, 256, NULL);
hash_ctl.keysize = sizeof(ItemPointerData);
hash_ctl.entrysize = sizeof(ItemPointerData);
hash_ctl.hcxt = CurrentMemoryContext;
vacuumstate->deleted = hash_create("hnswbulkdelete indextids", 256, &hash_ctl, HASH_ELEM | HASH_BLOBS | HASH_CONTEXT);
}
/*
@@ -593,7 +531,7 @@ InitVacuumState(HnswVacuumState * vacuumstate, IndexVacuumInfo *info, IndexBulkD
static void
FreeVacuumState(HnswVacuumState * vacuumstate)
{
tidhash_destroy(vacuumstate->deleted);
hash_destroy(vacuumstate->deleted);
FreeAccessStrategy(vacuumstate->bas);
pfree(vacuumstate->ntup);
MemoryContextDelete(vacuumstate->tmpCtx);

View File

@@ -2,38 +2,58 @@
#include <float.h>
#include "access/table.h"
#include "access/tableam.h"
#include "access/parallel.h"
#include "access/xact.h"
#include "catalog/index.h"
#include "catalog/pg_operator_d.h"
#include "catalog/pg_type_d.h"
#include "commands/progress.h"
#include "ivfflat.h"
#include "miscadmin.h"
#include "optimizer/optimizer.h"
#include "storage/bufmgr.h"
#include "tcop/tcopprot.h"
#include "utils/memutils.h"
#include "tcop/tcopprot.h"
#if PG_VERSION_NUM >= 140000
#include "utils/backend_progress.h"
#else
#elif PG_VERSION_NUM >= 120000
#include "pgstat.h"
#endif
#if PG_VERSION_NUM >= 120000
#include "access/tableam.h"
#include "commands/progress.h"
#else
#define PROGRESS_CREATEIDX_SUBPHASE 0
#define PROGRESS_CREATEIDX_TUPLES_TOTAL 0
#define PROGRESS_CREATEIDX_TUPLES_DONE 0
#endif
#if PG_VERSION_NUM >= 130000
#define CALLBACK_ITEM_POINTER ItemPointer tid
#else
#define CALLBACK_ITEM_POINTER HeapTuple hup
#endif
#if PG_VERSION_NUM >= 120000
#define UpdateProgress(index, val) pgstat_progress_update_param(index, val)
#else
#define UpdateProgress(index, val) ((void)val)
#endif
#if PG_VERSION_NUM >= 140000
#include "utils/backend_status.h"
#include "utils/wait_event.h"
#endif
#if PG_VERSION_NUM >= 120000
#include "access/table.h"
#include "optimizer/optimizer.h"
#else
#include "access/heapam.h"
#include "optimizer/planner.h"
#include "pgstat.h"
#endif
#define PARALLEL_KEY_IVFFLAT_SHARED UINT64CONST(0xA000000000000001)
#define PARALLEL_KEY_TUPLESORT UINT64CONST(0xA000000000000002)
#define PARALLEL_KEY_IVFFLAT_CENTERS UINT64CONST(0xA000000000000003)
@@ -57,7 +77,7 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
*/
if (buildstate->kmeansnormprocinfo != NULL)
{
if (!IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->collation, &value, buildstate->type))
if (!IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->collation, &value, buildstate->normvec))
return;
}
@@ -105,7 +125,7 @@ SampleCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
/* Add sample */
AddSample(values, buildstate);
AddSample(values, state);
/* Reset memory context */
MemoryContextSwitchTo(oldCtx);
@@ -130,8 +150,13 @@ SampleRows(IvfflatBuildState * buildstate)
{
BlockNumber targblock = BlockSampler_Next(&buildstate->bs);
#if PG_VERSION_NUM >= 120000
table_index_build_range_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
false, true, false, targblock, 1, SampleCallback, (void *) buildstate, NULL);
#else
IndexBuildHeapRangeScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
false, true, targblock, 1, SampleCallback, (void *) buildstate, NULL);
#endif
}
}
@@ -153,7 +178,7 @@ AddTupleToSort(Relation index, ItemPointer tid, Datum *values, IvfflatBuildState
/* Normalize if needed */
if (buildstate->normprocinfo != NULL)
{
if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->type))
if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->normvec))
return;
}
@@ -257,12 +282,16 @@ InsertTuples(Relation index, IvfflatBuildState * buildstate, ForkNumber forkNum)
IndexTuple itup = NULL; /* silence compiler warning */
int64 inserted = 0;
#if PG_VERSION_NUM >= 120000
TupleTableSlot *slot = MakeSingleTupleTableSlot(buildstate->tupdesc, &TTSOpsMinimalTuple);
#else
TupleTableSlot *slot = MakeSingleTupleTableSlot(buildstate->tupdesc);
#endif
TupleDesc tupdesc = RelationGetDescr(index);
pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_LOAD);
UpdateProgress(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_LOAD);
pgstat_progress_update_param(PROGRESS_CREATEIDX_TUPLES_TOTAL, buildstate->indtuples);
UpdateProgress(PROGRESS_CREATEIDX_TUPLES_TOTAL, buildstate->indtuples);
GetNextTuple(buildstate->sortstate, tupdesc, slot, &itup, &list);
@@ -298,7 +327,7 @@ InsertTuples(Relation index, IvfflatBuildState * buildstate, ForkNumber forkNum)
pfree(itup);
pgstat_progress_update_param(PROGRESS_CREATEIDX_TUPLES_DONE, ++inserted);
UpdateProgress(PROGRESS_CREATEIDX_TUPLES_DONE, ++inserted);
GetNextTuple(buildstate->sortstate, tupdesc, slot, &itup, &list);
}
@@ -312,39 +341,25 @@ InsertTuples(Relation index, IvfflatBuildState * buildstate, ForkNumber forkNum)
}
}
/*
* Get max dimensions
*/
static int
GetMaxDimensions(IvfflatType type)
{
return IVFFLAT_MAX_DIM;
}
/*
* Initialize the build state
*/
static void
InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, IndexInfo *indexInfo)
{
int maxDimensions;
buildstate->heap = heap;
buildstate->index = index;
buildstate->indexInfo = indexInfo;
buildstate->type = IvfflatGetType(index);
buildstate->lists = IvfflatGetLists(index);
buildstate->dimensions = TupleDescAttr(index->rd_att, 0)->atttypmod;
maxDimensions = GetMaxDimensions(buildstate->type);
/* Require column to have dimensions to be indexed */
if (buildstate->dimensions < 0)
elog(ERROR, "column does not have dimensions");
if (buildstate->dimensions > maxDimensions)
elog(ERROR, "column cannot have more than %d dimensions for ivfflat index", maxDimensions);
if (buildstate->dimensions > IVFFLAT_MAX_DIM)
elog(ERROR, "column cannot have more than %d dimensions for ivfflat index", IVFFLAT_MAX_DIM);
buildstate->reltuples = 0;
buildstate->indtuples = 0;
@@ -360,16 +375,27 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
elog(ERROR, "dimensions must be greater than one for this opclass");
/* Create tuple description for sorting */
#if PG_VERSION_NUM >= 120000
buildstate->tupdesc = CreateTemplateTupleDesc(3);
#else
buildstate->tupdesc = CreateTemplateTupleDesc(3, false);
#endif
TupleDescInitEntry(buildstate->tupdesc, (AttrNumber) 1, "list", INT4OID, -1, 0);
TupleDescInitEntry(buildstate->tupdesc, (AttrNumber) 2, "tid", TIDOID, -1, 0);
TupleDescInitEntry(buildstate->tupdesc, (AttrNumber) 3, "vector", RelationGetDescr(index)->attrs[0].atttypid, -1, 0);
#if PG_VERSION_NUM >= 120000
buildstate->slot = MakeSingleTupleTableSlot(buildstate->tupdesc, &TTSOpsVirtual);
#else
buildstate->slot = MakeSingleTupleTableSlot(buildstate->tupdesc);
#endif
buildstate->centers = VectorArrayInit(buildstate->lists, buildstate->dimensions);
buildstate->listInfo = palloc(sizeof(ListInfo) * buildstate->lists);
/* Reuse for each tuple */
buildstate->normvec = InitVector(buildstate->dimensions);
buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
"Ivfflat build temporary context",
ALLOCSET_DEFAULT_SIZES);
@@ -391,6 +417,7 @@ FreeBuildState(IvfflatBuildState * buildstate)
{
VectorArrayFree(buildstate->centers);
pfree(buildstate->listInfo);
pfree(buildstate->normvec);
#ifdef IVFFLAT_KMEANS_DEBUG
pfree(buildstate->listSums);
@@ -408,7 +435,7 @@ ComputeCenters(IvfflatBuildState * buildstate)
{
int numSamples;
pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_KMEANS);
UpdateProgress(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_KMEANS);
/* Target 50 samples per list, with at least 10000 samples */
/* The number of samples has a large effect on index build time */
@@ -479,30 +506,29 @@ CreateListPages(Relation index, VectorArray centers, int dimensions,
Buffer buf;
Page page;
GenericXLogState *state;
Size listSize;
OffsetNumber offno;
Size itemsz;
IvfflatList list;
listSize = MAXALIGN(IVFFLAT_LIST_SIZE(dimensions));
list = palloc0(listSize);
itemsz = MAXALIGN(IVFFLAT_LIST_SIZE(dimensions));
list = palloc(itemsz);
buf = IvfflatNewBuffer(index, forkNum);
IvfflatInitRegisterPage(index, &buf, &page, &state);
for (int i = 0; i < lists; i++)
{
OffsetNumber offno;
/* Load list */
list->startPage = InvalidBlockNumber;
list->insertPage = InvalidBlockNumber;
memcpy(&list->center, VectorArrayGet(centers, i), VECTOR_SIZE(dimensions));
/* Ensure free space */
if (PageGetFreeSpace(page) < listSize)
if (PageGetFreeSpace(page) < itemsz)
IvfflatAppendPage(index, &buf, &page, &state, forkNum);
/* Add the item */
offno = PageAddItem(page, (Item) list, listSize, InvalidOffsetNumber, false, false);
offno = PageAddItem(page, (Item) list, itemsz, InvalidOffsetNumber, false, false);
if (offno == InvalidOffsetNumber)
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
@@ -516,10 +542,10 @@ CreateListPages(Relation index, VectorArray centers, int dimensions,
pfree(list);
}
#ifdef IVFFLAT_KMEANS_DEBUG
/*
* Print k-means metrics
*/
#ifdef IVFFLAT_KMEANS_DEBUG
static void
PrintKmeansMetrics(IvfflatBuildState * buildstate)
{
@@ -604,7 +630,11 @@ IvfflatParallelScanAndSort(IvfflatSpool * ivfspool, IvfflatShared * ivfshared, S
{
SortCoordinate coordinate;
IvfflatBuildState buildstate;
#if PG_VERSION_NUM >= 120000
TableScanDesc scan;
#else
HeapScanDesc scan;
#endif
double reltuples;
IndexInfo *indexInfo;
@@ -628,11 +658,18 @@ IvfflatParallelScanAndSort(IvfflatSpool * ivfspool, IvfflatShared * ivfshared, S
buildstate.centers->length = buildstate.centers->maxlen;
ivfspool->sortstate = tuplesort_begin_heap(buildstate.tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, sortmem, coordinate, false);
buildstate.sortstate = ivfspool->sortstate;
#if PG_VERSION_NUM >= 120000
scan = table_beginscan_parallel(ivfspool->heap,
ParallelTableScanFromIvfflatShared(ivfshared));
reltuples = table_index_build_scan(ivfspool->heap, ivfspool->index, indexInfo,
true, progress, BuildCallback,
(void *) &buildstate, scan);
#else
scan = heap_beginscan_parallel(ivfspool->heap, &ivfshared->heapdesc);
reltuples = IndexBuildHeapScan(ivfspool->heap, ivfspool->index, indexInfo,
true, BuildCallback,
(void *) &buildstate, scan);
#endif
/* Execute this worker's part of the sort */
tuplesort_performsort(ivfspool->sortstate);
@@ -702,7 +739,11 @@ IvfflatParallelBuildMain(dsm_segment *seg, shm_toc *toc)
}
/* Open relations within worker */
#if PG_VERSION_NUM >= 120000
heapRel = table_open(ivfshared->heaprelid, heapLockmode);
#else
heapRel = heap_open(ivfshared->heaprelid, heapLockmode);
#endif
indexRel = index_open(ivfshared->indexrelid, indexLockmode);
/* Initialize worker's own spool */
@@ -722,7 +763,11 @@ IvfflatParallelBuildMain(dsm_segment *seg, shm_toc *toc)
/* Close relations within worker */
index_close(indexRel, indexLockmode);
#if PG_VERSION_NUM >= 120000
table_close(heapRel, heapLockmode);
#else
heap_close(heapRel, heapLockmode);
#endif
}
/*
@@ -747,7 +792,19 @@ IvfflatEndParallel(IvfflatLeader * ivfleader)
static Size
ParallelEstimateShared(Relation heap, Snapshot snapshot)
{
#if PG_VERSION_NUM >= 120000
return add_size(BUFFERALIGN(sizeof(IvfflatShared)), table_parallelscan_estimate(heap, snapshot));
#else
if (!IsMVCCSnapshot(snapshot))
{
Assert(snapshot == SnapshotAny);
return sizeof(IvfflatShared);
}
return add_size(offsetof(IvfflatShared, heapdesc) +
offsetof(ParallelHeapScanDescData, phs_snapshot_data),
EstimateSnapshotSpace(snapshot));
#endif
}
/*
@@ -798,7 +855,11 @@ IvfflatBeginParallel(IvfflatBuildState * buildstate, bool isconcurrent, int requ
/* Enter parallel mode and create context */
EnterParallelMode();
Assert(request > 0);
#if PG_VERSION_NUM >= 120000
pcxt = CreateParallelContext("vector", "IvfflatParallelBuildMain", request);
#else
pcxt = CreateParallelContext("vector", "IvfflatParallelBuildMain", request, true);
#endif
scantuplesortstates = leaderparticipates ? request + 1 : request;
@@ -856,9 +917,13 @@ IvfflatBeginParallel(IvfflatBuildState * buildstate, bool isconcurrent, int requ
#ifdef IVFFLAT_KMEANS_DEBUG
ivfshared->inertia = 0;
#endif
#if PG_VERSION_NUM >= 120000
table_parallelscan_initialize(buildstate->heap,
ParallelTableScanFromIvfflatShared(ivfshared),
snapshot);
#else
heap_parallelscan_initialize(&ivfshared->heapdesc, buildstate->heap, snapshot);
#endif
/* Store shared tuplesort-private state, for which we reserved space */
sharedsort = (Sharedsort *) shm_toc_allocate(pcxt->toc, estsort);
@@ -929,7 +994,7 @@ AssignTuples(IvfflatBuildState * buildstate)
Oid sortCollations[] = {InvalidOid};
bool nullsFirstFlags[] = {false};
pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_ASSIGN);
UpdateProgress(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_ASSIGN);
/* Calculate parallel workers */
if (buildstate->heap != NULL)
@@ -957,8 +1022,15 @@ AssignTuples(IvfflatBuildState * buildstate)
if (buildstate->ivfleader)
buildstate->reltuples = ParallelHeapScan(buildstate);
else
{
#if PG_VERSION_NUM >= 120000
buildstate->reltuples = table_index_build_scan(buildstate->heap, buildstate->index, buildstate->indexInfo,
true, true, BuildCallback, (void *) buildstate, NULL);
#else
buildstate->reltuples = IndexBuildHeapScan(buildstate->heap, buildstate->index, buildstate->indexInfo,
true, BuildCallback, (void *) buildstate, NULL);
#endif
}
#ifdef IVFFLAT_KMEANS_DEBUG
PrintKmeansMetrics(buildstate);

View File

@@ -3,16 +3,14 @@
#include <float.h>
#include "access/amapi.h"
#include "access/reloptions.h"
#include "commands/progress.h"
#include "commands/vacuum.h"
#include "ivfflat.h"
#include "utils/guc.h"
#include "utils/selfuncs.h"
#include "utils/spccache.h"
#if PG_VERSION_NUM < 150000
#define MarkGUCPrefixReserved(x) EmitWarningsOnPlaceholders(x)
#if PG_VERSION_NUM >= 120000
#include "commands/progress.h"
#endif
int ivfflat_probes;
@@ -26,7 +24,7 @@ IvfflatInit(void)
{
ivfflat_relopt_kind = add_reloption_kind();
add_int_reloption(ivfflat_relopt_kind, "lists", "Number of inverted lists",
IVFFLAT_DEFAULT_LISTS, IVFFLAT_MIN_LISTS, IVFFLAT_MAX_LISTS
IVFFLAT_DEFAULT_LISTS, 1, IVFFLAT_MAX_LISTS
#if PG_VERSION_NUM >= 130000
,AccessExclusiveLock
#endif
@@ -34,14 +32,13 @@ IvfflatInit(void)
DefineCustomIntVariable("ivfflat.probes", "Sets the number of probes",
"Valid range is 1..lists.", &ivfflat_probes,
IVFFLAT_DEFAULT_PROBES, IVFFLAT_MIN_LISTS, IVFFLAT_MAX_LISTS, PGC_USERSET, 0, NULL, NULL, NULL);
MarkGUCPrefixReserved("ivfflat");
1, 1, IVFFLAT_MAX_LISTS, PGC_USERSET, 0, NULL, NULL, NULL);
}
/*
* Get the name of index build phase
*/
#if PG_VERSION_NUM >= 120000
static char *
ivfflatbuildphasename(int64 phasenum)
{
@@ -59,6 +56,7 @@ ivfflatbuildphasename(int64 phasenum)
return NULL;
}
}
#endif
/*
* Estimate the cost of an index scan
@@ -73,7 +71,10 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
int lists;
double ratio;
double spc_seq_page_cost;
Relation index;
Relation indexRel;
#if PG_VERSION_NUM < 120000
List *qinfos;
#endif
/* Never use index without order */
if (path->indexorderbys == NULL)
@@ -88,9 +89,9 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
MemSet(&costs, 0, sizeof(costs));
index = index_open(path->indexinfo->indexoid, NoLock);
IvfflatGetMetaPageInfo(index, &lists, NULL);
index_close(index, NoLock);
indexRel = index_open(path->indexinfo->indexoid, NoLock);
lists = IvfflatGetLists(indexRel);
index_close(indexRel, NoLock);
/* Get the ratio of lists that we need to visit */
ratio = ((double) ivfflat_probes) / lists;
@@ -104,7 +105,12 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
*/
costs.numIndexTuples = path->indexinfo->tuples * ratio;
#if PG_VERSION_NUM >= 120000
genericcostestimate(root, path, loop_count, &costs);
#else
qinfos = deconstruct_indexquals(path);
genericcostestimate(root, path, loop_count, qinfos, &costs);
#endif
get_tablespace_page_costs(path->indexinfo->reltablespace, NULL, &spc_seq_page_cost);
@@ -221,7 +227,9 @@ ivfflathandler(PG_FUNCTION_ARGS)
amroutine->amcostestimate = ivfflatcostestimate;
amroutine->amoptions = ivfflatoptions;
amroutine->amproperty = NULL; /* TODO AMPROP_DISTANCE_ORDERABLE */
#if PG_VERSION_NUM >= 120000
amroutine->ambuildphasename = ivfflatbuildphasename;
#endif
amroutine->amvalidate = ivfflatvalidate;
#if PG_VERSION_NUM >= 140000
amroutine->amadjustmembers = NULL;

View File

@@ -3,10 +3,9 @@
#include "postgres.h"
#include "access/genam.h"
#include "access/generic_xlog.h"
#include "access/parallel.h"
#include "lib/pairingheap.h"
#include "access/reloptions.h"
#include "nodes/execnodes.h"
#include "port.h" /* for random() */
#include "utils/sampling.h"
@@ -17,6 +16,10 @@
#include "common/pg_prng.h"
#endif
#if PG_VERSION_NUM < 120000
#include "access/relscan.h"
#endif
#ifdef IVFFLAT_BENCH
#include "portability/instr_time.h"
#endif
@@ -37,16 +40,8 @@
#define IVFFLAT_METAPAGE_BLKNO 0
#define IVFFLAT_HEAD_BLKNO 1 /* first list page */
/* IVFFlat parameters */
#define IVFFLAT_DEFAULT_LISTS 100
#define IVFFLAT_MIN_LISTS 1
#define IVFFLAT_MAX_LISTS 32768
#define IVFFLAT_DEFAULT_PROBES 1
typedef enum IvfflatType
{
IVFFLAT_TYPE_VECTOR
} IvfflatType;
/* Build phases */
/* PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE is 1 */
@@ -137,10 +132,16 @@ typedef struct IvfflatShared
#ifdef IVFFLAT_KMEANS_DEBUG
double inertia;
#endif
#if PG_VERSION_NUM < 120000
ParallelHeapScanDescData heapdesc; /* must come last */
#endif
} IvfflatShared;
#if PG_VERSION_NUM >= 120000
#define ParallelTableScanFromIvfflatShared(shared) \
(ParallelTableScanDesc) ((char *) (shared) + BUFFERALIGN(sizeof(IvfflatShared)))
#endif
typedef struct IvfflatLeader
{
@@ -158,7 +159,6 @@ typedef struct IvfflatBuildState
Relation heap;
Relation index;
IndexInfo *indexInfo;
IvfflatType type;
/* Settings */
int dimensions;
@@ -178,6 +178,7 @@ typedef struct IvfflatBuildState
VectorArray samples;
VectorArray centers;
ListInfo *listInfo;
Vector *normvec;
#ifdef IVFFLAT_KMEANS_DEBUG
double inertia;
@@ -240,8 +241,8 @@ typedef struct IvfflatScanList
typedef struct IvfflatScanOpaqueData
{
int probes;
int dimensions;
bool first;
Buffer buf;
/* Sorting */
Tuplesortstate *sortstate;
@@ -271,11 +272,9 @@ VectorArray VectorArrayInit(int maxlen, int dimensions);
void VectorArrayFree(VectorArray arr);
void PrintVectorArray(char *msg, VectorArray arr);
void IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers);
FmgrInfo *IvfflatOptionalProcInfo(Relation index, uint16 procnum);
IvfflatType IvfflatGetType(Relation index);
bool IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, IvfflatType type);
FmgrInfo *IvfflatOptionalProcInfo(Relation rel, uint16 procnum);
bool IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
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);
void IvfflatCommitBuffer(Buffer buf, GenericXLogState *state);
void IvfflatAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **state, ForkNumber forkNum);

View File

@@ -2,7 +2,6 @@
#include <float.h>
#include "access/generic_xlog.h"
#include "ivfflat.h"
#include "storage/bufmgr.h"
#include "storage/lmgr.h"
@@ -12,37 +11,36 @@
* Find the list that minimizes the distance function
*/
static void
FindInsertPage(Relation index, Datum *values, BlockNumber *insertPage, ListInfo * listInfo)
FindInsertPage(Relation rel, Datum *values, BlockNumber *insertPage, ListInfo * listInfo)
{
Buffer cbuf;
Page cpage;
IvfflatList list;
double distance;
double minDistance = DBL_MAX;
BlockNumber nextblkno = IVFFLAT_HEAD_BLKNO;
FmgrInfo *procinfo;
Oid collation;
OffsetNumber offno;
OffsetNumber maxoffno;
/* Avoid compiler warning */
listInfo->blkno = nextblkno;
listInfo->offno = FirstOffsetNumber;
procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
collation = index->rd_indcollation[0];
procinfo = index_getprocinfo(rel, 1, IVFFLAT_DISTANCE_PROC);
collation = rel->rd_indcollation[0];
/* Search all list pages */
while (BlockNumberIsValid(nextblkno))
{
Buffer cbuf;
Page cpage;
OffsetNumber maxoffno;
cbuf = ReadBuffer(index, nextblkno);
cbuf = ReadBuffer(rel, nextblkno);
LockBuffer(cbuf, BUFFER_LOCK_SHARE);
cpage = BufferGetPage(cbuf);
maxoffno = PageGetMaxOffsetNumber(cpage);
for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
{
IvfflatList list;
double distance;
list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, offno));
distance = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, values[0], PointerGetDatum(&list->center)));
@@ -65,7 +63,7 @@ FindInsertPage(Relation index, Datum *values, BlockNumber *insertPage, ListInfo
* Insert a tuple into the index
*/
static void
InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel)
InsertTuple(Relation rel, Datum *values, bool *isnull, ItemPointer heap_tid, Relation heapRel)
{
IndexTuple itup;
Datum value;
@@ -82,33 +80,33 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
/* Normalize if needed */
normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
normprocinfo = IvfflatOptionalProcInfo(rel, IVFFLAT_NORM_PROC);
if (normprocinfo != NULL)
{
if (!IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value, IvfflatGetType(index)))
if (!IvfflatNormValue(normprocinfo, rel->rd_indcollation[0], &value, NULL))
return;
}
/* Find the insert page - sets the page and list info */
FindInsertPage(index, values, &insertPage, &listInfo);
FindInsertPage(rel, values, &insertPage, &listInfo);
Assert(BlockNumberIsValid(insertPage));
originalInsertPage = insertPage;
/* Form tuple */
itup = index_form_tuple(RelationGetDescr(index), &value, isnull);
itup = index_form_tuple(RelationGetDescr(rel), &value, isnull);
itup->t_tid = *heap_tid;
/* Get tuple size */
itemsz = MAXALIGN(IndexTupleSize(itup));
Assert(itemsz <= BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(IvfflatPageOpaqueData)) - sizeof(ItemIdData));
Assert(itemsz <= BLCKSZ - MAXALIGN(SizeOfPageHeaderData) - MAXALIGN(sizeof(IvfflatPageOpaqueData)));
/* Find a page to insert the item */
for (;;)
{
buf = ReadBuffer(index, insertPage);
buf = ReadBuffer(rel, insertPage);
LockBuffer(buf, BUFFER_LOCK_EXCLUSIVE);
state = GenericXLogStart(index);
state = GenericXLogStart(rel);
page = GenericXLogRegisterBuffer(state, buf, 0);
if (PageGetFreeSpace(page) >= itemsz)
@@ -128,9 +126,9 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
Page newpage;
/* Add a new page */
LockRelationForExtension(index, ExclusiveLock);
newbuf = IvfflatNewBuffer(index, MAIN_FORKNUM);
UnlockRelationForExtension(index, ExclusiveLock);
LockRelationForExtension(rel, ExclusiveLock);
newbuf = IvfflatNewBuffer(rel, MAIN_FORKNUM);
UnlockRelationForExtension(rel, ExclusiveLock);
/* Init new page */
newpage = GenericXLogRegisterBuffer(state, newbuf, GENERIC_XLOG_FULL_IMAGE);
@@ -143,13 +141,15 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
IvfflatPageGetOpaque(page)->nextblkno = insertPage;
/* Commit */
MarkBufferDirty(newbuf);
MarkBufferDirty(buf);
GenericXLogFinish(state);
/* Unlock previous buffer */
UnlockReleaseBuffer(buf);
/* Prepare new buffer */
state = GenericXLogStart(index);
state = GenericXLogStart(rel);
buf = newbuf;
page = GenericXLogRegisterBuffer(state, buf, 0);
break;
@@ -158,13 +158,13 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
/* Add to next offset */
if (PageAddItem(page, (Item) itup, itemsz, InvalidOffsetNumber, false, false) == InvalidOffsetNumber)
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(index));
elog(ERROR, "failed to add index item to \"%s\"", RelationGetRelationName(rel));
IvfflatCommitBuffer(buf, state);
/* Update the insert page */
if (insertPage != originalInsertPage)
IvfflatUpdateList(index, listInfo, insertPage, originalInsertPage, InvalidBlockNumber, MAIN_FORKNUM);
IvfflatUpdateList(rel, listInfo, insertPage, originalInsertPage, InvalidBlockNumber, MAIN_FORKNUM);
}
/*

View File

@@ -6,10 +6,6 @@
#include "ivfflat.h"
#include "miscadmin.h"
#ifdef IVFFLAT_MEMORY
#include "utils/memutils.h"
#endif
/*
* Initialize with kmeans++
*
@@ -21,6 +17,10 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
FmgrInfo *procinfo;
Oid collation;
int64 j;
double distance;
double sum;
double choice;
Vector *vec;
float *weight = palloc(samples->length * sizeof(float));
int numCenters = centers->maxlen;
int numSamples = samples->length;
@@ -33,21 +33,17 @@ InitCenters(Relation index, VectorArray samples, VectorArray centers, float *low
centers->length++;
for (j = 0; j < numSamples; j++)
weight[j] = FLT_MAX;
weight[j] = DBL_MAX;
for (int i = 0; i < numCenters; i++)
{
double sum;
double choice;
CHECK_FOR_INTERRUPTS();
sum = 0.0;
for (j = 0; j < numSamples; j++)
{
Vector *vec = VectorArrayGet(samples, j);
double distance;
vec = VectorArrayGet(samples, j);
/* Only need to compute distance for new center */
/* TODO Use triangle inequality to reduce distance calculations */
@@ -116,6 +112,7 @@ CompareVectors(const void *a, const void *b)
static void
QuickCenters(Relation index, VectorArray samples, VectorArray centers)
{
Vector *vec;
int dimensions = centers->dim;
Oid collation = index->rd_indcollation[0];
FmgrInfo *normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_KMEANS_NORM_PROC);
@@ -126,7 +123,7 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers)
qsort(samples->items, samples->length, VECTOR_SIZE(samples->dim), CompareVectors);
for (int i = 0; i < samples->length; i++)
{
Vector *vec = VectorArrayGet(samples, i);
vec = VectorArrayGet(samples, i);
if (i == 0 || CompareVectors(vec, VectorArrayGet(samples, i - 1)) != 0)
{
@@ -139,7 +136,7 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers)
/* Fill remaining with random data */
while (centers->length < centers->maxlen)
{
Vector *vec = VectorArrayGet(centers, centers->length);
vec = VectorArrayGet(centers, centers->length);
SET_VARSIZE(vec, VECTOR_SIZE(dimensions));
vec->dim = dimensions;
@@ -155,23 +152,6 @@ QuickCenters(Relation index, VectorArray samples, VectorArray centers)
}
}
#ifdef IVFFLAT_MEMORY
/*
* Show memory usage
*/
static void
ShowMemoryUsage(Size estimatedSize)
{
#if PG_VERSION_NUM >= 130000
elog(INFO, "total memory: %zu MB",
MemoryContextMemAllocated(CurrentMemoryContext, true) / (1024 * 1024));
#else
MemoryContextStats(CurrentMemoryContext);
#endif
elog(INFO, "estimated memory: %zu MB", estimatedSize / (1024 * 1024));
}
#endif
/*
* Use Elkan for performance. This requires distance function to satisfy triangle inequality.
*
@@ -188,6 +168,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
Oid collation;
Vector *vec;
Vector *newCenter;
int iteration;
int64 j;
int64 k;
int dimensions = centers->dim;
@@ -201,6 +182,14 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
float *s;
float *halfcdist;
float *newcdist;
int changes;
double minDistance;
int closestCenter;
double distance;
bool rj;
bool rjreset;
double dxcx;
double dxc;
/* Calculate allocation sizes */
Size samplesSize = VECTOR_ARRAY_SIZE(samples->maxlen, samples->dim);
@@ -252,24 +241,20 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
vec->dim = dimensions;
}
#ifdef IVFFLAT_MEMORY
ShowMemoryUsage(totalSize);
#endif
/* Pick initial centers */
InitCenters(index, samples, centers, lowerBound);
/* Assign each x to its closest initial center c(x) = argmin d(x,c) */
for (j = 0; j < numSamples; j++)
{
float minDistance = FLT_MAX;
int closestCenter = 0;
minDistance = DBL_MAX;
closestCenter = 0;
/* Find closest center */
for (k = 0; k < numCenters; k++)
{
/* TODO Use Lemma 1 in k-means++ initialization */
float distance = lowerBound[j * numCenters + k];
distance = lowerBound[j * numCenters + k];
if (distance < minDistance)
{
@@ -283,14 +268,13 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
}
/* Give 500 iterations to converge */
for (int iteration = 0; iteration < 500; iteration++)
for (iteration = 0; iteration < 500; iteration++)
{
int changes = 0;
bool rjreset;
/* Can take a while, so ensure we can interrupt */
CHECK_FOR_INTERRUPTS();
changes = 0;
/* Step 1: For all centers, compute distance */
for (j = 0; j < numCenters; j++)
{
@@ -298,8 +282,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
for (k = j + 1; k < numCenters; k++)
{
float distance = 0.5 * DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, k))));
distance = 0.5 * DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, k))));
halfcdist[j * numCenters + k] = distance;
halfcdist[k * numCenters + j] = distance;
}
@@ -308,12 +291,10 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
/* For all centers c, compute s(c) */
for (j = 0; j < numCenters; j++)
{
float minDistance = FLT_MAX;
minDistance = DBL_MAX;
for (k = 0; k < numCenters; k++)
{
float distance;
if (j == k)
continue;
@@ -329,8 +310,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
for (j = 0; j < numSamples; j++)
{
bool rj;
/* Step 2: Identify all points x such that u(x) <= s(c(x)) */
if (upperBound[j] <= s[closestCenters[j]])
continue;
@@ -339,8 +318,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
for (k = 0; k < numCenters; k++)
{
float dxcx;
/* Step 3: For all remaining points x and centers c */
if (k == closestCenters[j])
continue;
@@ -370,7 +347,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
/* Step 3b */
if (dxcx > lowerBound[j * numCenters + k] || dxcx > halfcdist[closestCenters[j] * numCenters + k])
{
float dxc = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, k))));
dxc = DatumGetFloat8(FunctionCall2Coll(procinfo, collation, PointerGetDatum(vec), PointerGetDatum(VectorArrayGet(centers, k))));
/* d(x,c) calculated */
lowerBound[j * numCenters + k] = dxc;
@@ -384,6 +361,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
changes++;
}
}
}
}
@@ -400,8 +378,6 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
for (j = 0; j < numSamples; j++)
{
int closestCenter;
vec = VectorArrayGet(samples, j);
closestCenter = closestCenters[j];
@@ -450,7 +426,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
{
for (k = 0; k < numCenters; k++)
{
float distance = lowerBound[j * numCenters + k] - newcdist[k];
distance = lowerBound[j * numCenters + k] - newcdist[k];
if (distance < 0)
distance = 0;
@@ -466,7 +442,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers)
/* Step 7 */
for (j = 0; j < numCenters; j++)
VectorArraySet(centers, j, VectorArrayGet(newCenters, j));
memcpy(VectorArrayGet(centers, j), VectorArrayGet(newCenters, j), VECTOR_SIZE(dimensions));
if (changes == 0 && iteration != 0)
break;
@@ -489,6 +465,9 @@ static void
CheckCenters(Relation index, VectorArray centers)
{
FmgrInfo *normprocinfo;
Oid collation;
Vector *vec;
double norm;
if (centers->length != centers->maxlen)
elog(ERROR, "Not enough centers. Please report a bug.");
@@ -496,7 +475,7 @@ CheckCenters(Relation index, VectorArray centers)
/* Ensure no NaN or infinite values */
for (int i = 0; i < centers->length; i++)
{
Vector *vec = VectorArrayGet(centers, i);
vec = VectorArrayGet(centers, i);
for (int j = 0; j < vec->dim; j++)
{
@@ -522,12 +501,11 @@ CheckCenters(Relation index, VectorArray centers)
normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
if (normprocinfo != NULL)
{
Oid collation = index->rd_indcollation[0];
collation = index->rd_indcollation[0];
for (int i = 0; i < centers->length; i++)
{
double norm = DatumGetFloat8(FunctionCall1Coll(normprocinfo, collation, PointerGetDatum(VectorArrayGet(centers, i))));
norm = DatumGetFloat8(FunctionCall1Coll(normprocinfo, collation, PointerGetDatum(VectorArrayGet(centers, i))));
if (norm == 0)
elog(ERROR, "Zero norm detected. Please report a bug.");
}

View File

@@ -5,7 +5,6 @@
#include "access/relscan.h"
#include "catalog/pg_operator_d.h"
#include "catalog/pg_type_d.h"
#include "lib/pairingheap.h"
#include "ivfflat.h"
#include "miscadmin.h"
#include "pgstat.h"
@@ -32,36 +31,36 @@ CompareLists(const pairingheap_node *a, const pairingheap_node *b, void *arg)
static void
GetScanLists(IndexScanDesc scan, Datum value)
{
IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
Buffer cbuf;
Page cpage;
IvfflatList list;
OffsetNumber offno;
OffsetNumber maxoffno;
BlockNumber nextblkno = IVFFLAT_HEAD_BLKNO;
int listCount = 0;
IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
double distance;
IvfflatScanList *scanlist;
double maxDistance = DBL_MAX;
/* Search all list pages */
while (BlockNumberIsValid(nextblkno))
{
Buffer cbuf;
Page cpage;
OffsetNumber maxoffno;
cbuf = ReadBuffer(scan->indexRelation, nextblkno);
LockBuffer(cbuf, BUFFER_LOCK_SHARE);
cpage = BufferGetPage(cbuf);
maxoffno = PageGetMaxOffsetNumber(cpage);
for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
{
IvfflatList list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, offno));
double distance;
list = (IvfflatList) PageGetItem(cpage, PageGetItemId(cpage, offno));
/* Use procinfo from the index instead of scan key for performance */
distance = DatumGetFloat8(FunctionCall2Coll(so->procinfo, so->collation, PointerGetDatum(&list->center), value));
if (listCount < so->probes)
{
IvfflatScanList *scanlist;
scanlist = &so->lists[listCount];
scanlist->startPage = list->startPage;
scanlist->distance = distance;
@@ -76,8 +75,6 @@ GetScanLists(IndexScanDesc scan, Datum value)
}
else if (distance < maxDistance)
{
IvfflatScanList *scanlist;
/* Remove */
scanlist = (IvfflatScanList *) pairingheap_remove_first(so->listQueue);
@@ -104,9 +101,22 @@ static void
GetScanItems(IndexScanDesc scan, Datum value)
{
IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
Buffer buf;
Page page;
IndexTuple itup;
BlockNumber searchPage;
OffsetNumber offno;
OffsetNumber maxoffno;
Datum datum;
bool isnull;
TupleDesc tupdesc = RelationGetDescr(scan->indexRelation);
double tuples = 0;
#if PG_VERSION_NUM >= 120000
TupleTableSlot *slot = MakeSingleTupleTableSlot(so->tupdesc, &TTSOpsVirtual);
#else
TupleTableSlot *slot = MakeSingleTupleTableSlot(so->tupdesc);
#endif
/*
* Reuse same set of shared buffers for scan
@@ -118,28 +128,19 @@ GetScanItems(IndexScanDesc scan, Datum value)
/* Search closest probes lists */
while (!pairingheap_is_empty(so->listQueue))
{
BlockNumber searchPage = ((IvfflatScanList *) pairingheap_remove_first(so->listQueue))->startPage;
searchPage = ((IvfflatScanList *) pairingheap_remove_first(so->listQueue))->startPage;
/* Search all entry pages for list */
while (BlockNumberIsValid(searchPage))
{
Buffer buf;
Page page;
OffsetNumber maxoffno;
buf = ReadBufferExtended(scan->indexRelation, MAIN_FORKNUM, searchPage, RBM_NORMAL, bas);
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
maxoffno = PageGetMaxOffsetNumber(page);
for (OffsetNumber offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
for (offno = FirstOffsetNumber; offno <= maxoffno; offno = OffsetNumberNext(offno))
{
IndexTuple itup;
Datum datum;
bool isnull;
ItemId itemid = PageGetItemId(page, offno);
itup = (IndexTuple) PageGetItem(page, itemid);
itup = (IndexTuple) PageGetItem(page, PageGetItemId(page, offno));
datum = index_getattr(itup, 1, tupdesc, &isnull);
/*
@@ -153,6 +154,8 @@ GetScanItems(IndexScanDesc scan, Datum value)
slot->tts_isnull[0] = false;
slot->tts_values[1] = PointerGetDatum(&itup->t_tid);
slot->tts_isnull[1] = false;
slot->tts_values[2] = Int32GetDatum((int) searchPage);
slot->tts_isnull[2] = false;
ExecStoreVirtualTuple(slot);
tuplesort_puttupleslot(so->sortstate, slot);
@@ -186,7 +189,6 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
IndexScanDesc scan;
IvfflatScanOpaque so;
int lists;
int dimensions;
AttrNumber attNums[] = {1};
Oid sortOperators[] = {Float8LessOperator};
Oid sortCollations[] = {InvalidOid};
@@ -194,17 +196,15 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
int probes = ivfflat_probes;
scan = RelationGetIndexScan(index, nkeys, norderbys);
/* Get lists and dimensions from metapage */
IvfflatGetMetaPageInfo(index, &lists, &dimensions);
lists = IvfflatGetLists(scan->indexRelation);
if (probes > lists)
probes = lists;
so = (IvfflatScanOpaque) palloc(offsetof(IvfflatScanOpaqueData, lists) + probes * sizeof(IvfflatScanList));
so->buf = InvalidBuffer;
so->first = true;
so->probes = probes;
so->dimensions = dimensions;
/* Set support functions */
so->procinfo = index_getprocinfo(index, 1, IVFFLAT_DISTANCE_PROC);
@@ -212,14 +212,23 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
so->collation = index->rd_indcollation[0];
/* Create tuple description for sorting */
so->tupdesc = CreateTemplateTupleDesc(2);
#if PG_VERSION_NUM >= 120000
so->tupdesc = CreateTemplateTupleDesc(3);
#else
so->tupdesc = CreateTemplateTupleDesc(3, false);
#endif
TupleDescInitEntry(so->tupdesc, (AttrNumber) 1, "distance", FLOAT8OID, -1, 0);
TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "heaptid", TIDOID, -1, 0);
TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "tid", TIDOID, -1, 0);
TupleDescInitEntry(so->tupdesc, (AttrNumber) 3, "indexblkno", INT4OID, -1, 0);
/* Prep sort */
so->sortstate = tuplesort_begin_heap(so->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, work_mem, NULL, false);
#if PG_VERSION_NUM >= 120000
so->slot = MakeSingleTupleTableSlot(so->tupdesc, &TTSOpsMinimalTuple);
#else
so->slot = MakeSingleTupleTableSlot(so->tupdesc);
#endif
so->listQueue = pairingheap_allocate(CompareLists, scan);
@@ -268,7 +277,6 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
if (so->first)
{
Datum value;
IvfflatType type = IvfflatGetType(scan->indexRelation);
/* Count index scan for stats */
pgstat_count_index_scan(scan->indexRelation);
@@ -277,29 +285,21 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
if (scan->orderByData == NULL)
elog(ERROR, "cannot scan ivfflat index without order");
/* Requires MVCC-compliant snapshot as not able to pin during sorting */
/* https://www.postgresql.org/docs/current/index-locking.html */
if (!IsMVCCSnapshot(scan->xs_snapshot))
elog(ERROR, "non-MVCC snapshots are not supported with ivfflat");
/* No items will match if null */
if (scan->orderByData->sk_flags & SK_ISNULL)
{
if (type == IVFFLAT_TYPE_VECTOR)
value = PointerGetDatum(InitVector(so->dimensions));
else
elog(ERROR, "Unsupported type");
}
else
{
value = scan->orderByData->sk_argument;
return false;
/* Value should not be compressed or toasted */
Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
value = scan->orderByData->sk_argument;
/* Fine if normalization fails */
if (so->normprocinfo != NULL)
IvfflatNormValue(so->normprocinfo, so->collation, &value, type);
/* Value should not be compressed or toasted */
Assert(!VARATT_IS_COMPRESSED(DatumGetPointer(value)));
Assert(!VARATT_IS_EXTENDED(DatumGetPointer(value)));
if (so->normprocinfo != NULL)
{
/* No items will match if normalization fails */
if (!IvfflatNormValue(so->normprocinfo, so->collation, &value, NULL))
return false;
}
IvfflatBench("GetScanLists", GetScanLists(scan, value));
@@ -313,10 +313,26 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
if (tuplesort_gettupleslot(so->sortstate, true, false, so->slot, NULL))
{
ItemPointer heaptid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
ItemPointer tid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
BlockNumber indexblkno = DatumGetInt32(slot_getattr(so->slot, 3, &so->isnull));
#if PG_VERSION_NUM >= 120000
scan->xs_heaptid = *tid;
#else
scan->xs_ctup.t_self = *tid;
#endif
if (BufferIsValid(so->buf))
ReleaseBuffer(so->buf);
/*
* An index scan must maintain a pin on the index page holding the
* item last returned by amgettuple
*
* https://www.postgresql.org/docs/current/index-locking.html
*/
so->buf = ReadBuffer(scan->indexRelation, indexblkno);
scan->xs_heaptid = *heaptid;
scan->xs_recheck = false;
scan->xs_recheckorderby = false;
return true;
}
@@ -332,6 +348,10 @@ ivfflatendscan(IndexScanDesc scan)
{
IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
/* Release pin */
if (BufferIsValid(so->buf))
ReleaseBuffer(so->buf);
pairingheap_free(so->listQueue);
tuplesort_end(so->sortstate);

View File

@@ -1,6 +1,5 @@
#include "postgres.h"
#include "access/generic_xlog.h"
#include "ivfflat.h"
#include "storage/bufmgr.h"
#include "vector.h"
@@ -58,21 +57,12 @@ IvfflatGetLists(Relation index)
* Get proc
*/
FmgrInfo *
IvfflatOptionalProcInfo(Relation index, uint16 procnum)
IvfflatOptionalProcInfo(Relation rel, uint16 procnum)
{
if (!OidIsValid(index_getprocid(index, 1, procnum)))
if (!OidIsValid(index_getprocid(rel, 1, procnum)))
return NULL;
return index_getprocinfo(index, 1, procnum);
}
/*
* Get type
*/
IvfflatType
IvfflatGetType(Relation index)
{
return IVFFLAT_TYPE_VECTOR;
return index_getprocinfo(rel, 1, procnum);
}
/*
@@ -84,24 +74,21 @@ IvfflatGetType(Relation index)
* if it's different than the original value
*/
bool
IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, IvfflatType type)
IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result)
{
double norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
if (norm > 0)
{
if (type == IVFFLAT_TYPE_VECTOR)
{
Vector *v = DatumGetVector(*value);
Vector *result = InitVector(v->dim);
Vector *v = DatumGetVector(*value);
for (int i = 0; i < v->dim; i++)
result->x[i] = v->x[i] / norm;
if (result == NULL)
result = InitVector(v->dim);
*value = PointerGetDatum(result);
}
else
elog(ERROR, "Unsupported type");
for (int i = 0; i < v->dim; i++)
result->x[i] = v->x[i] / norm;
*value = PointerGetDatum(result);
return true;
}
@@ -149,6 +136,7 @@ IvfflatInitRegisterPage(Relation index, Buffer *buf, Page *page, GenericXLogStat
void
IvfflatCommitBuffer(Buffer buf, GenericXLogState *state)
{
MarkBufferDirty(buf);
GenericXLogFinish(state);
UnlockReleaseBuffer(buf);
}
@@ -172,6 +160,8 @@ IvfflatAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **st
IvfflatInitPage(newbuf, newpage);
/* Commit */
MarkBufferDirty(*buf);
MarkBufferDirty(newbuf);
GenericXLogFinish(*state);
/* Unlock */
@@ -182,29 +172,6 @@ IvfflatAppendPage(Relation index, Buffer *buf, Page *page, GenericXLogState **st
*buf = newbuf;
}
/*
* Get the metapage info
*/
void
IvfflatGetMetaPageInfo(Relation index, int *lists, int *dimensions)
{
Buffer buf;
Page page;
IvfflatMetaPage metap;
buf = ReadBuffer(index, IVFFLAT_METAPAGE_BLKNO);
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
metap = IvfflatPageGetMeta(page);
*lists = metap->lists;
if (dimensions != NULL)
*dimensions = metap->dimensions;
UnlockReleaseBuffer(buf);
}
/*
* Update the start or insert page of a list
*/

View File

@@ -1,6 +1,5 @@
#include "postgres.h"
#include "access/generic_xlog.h"
#include "commands/vacuum.h"
#include "ivfflat.h"
#include "storage/bufmgr.h"
@@ -108,6 +107,7 @@ ivfflatbulkdelete(IndexVacuumInfo *info, IndexBulkDeleteResult *stats,
{
/* Delete tuples */
PageIndexMultiDelete(page, deletable, ndeletable);
MarkBufferDirty(buf);
GenericXLogFinish(state);
}
else

View File

@@ -1,788 +0,0 @@
#include "postgres.h"
#include <limits.h>
#include <math.h>
#include "fmgr.h"
#include "libpq/pqformat.h"
#include "sparsevec.h"
#include "utils/array.h"
#include "utils/builtins.h"
#include "vector.h"
#if PG_VERSION_NUM >= 120000
#include "common/shortest_dec.h"
#include "utils/float.h"
#else
#include <float.h>
#include "utils/builtins.h"
#endif
/*
* Ensure same dimensions
*/
static inline void
CheckDims(SparseVector * a, SparseVector * b)
{
if (a->dim != b->dim)
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("different sparsevec dimensions %d and %d", a->dim, b->dim)));
}
/*
* Ensure expected dimensions
*/
static inline void
CheckExpectedDim(int32 typmod, int dim)
{
if (typmod != -1 && typmod != dim)
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("expected %d dimensions, not %d", typmod, dim)));
}
/*
* Ensure valid dimensions
*/
static inline void
CheckDim(int dim)
{
if (dim < 1)
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("sparsevec must have at least 1 dimension")));
if (dim > SPARSEVEC_MAX_DIM)
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
errmsg("sparsevec cannot have more than %d dimensions", SPARSEVEC_MAX_DIM)));
}
/*
* Ensure valid nnz
*/
static inline void
CheckNnz(int nnz, int dim)
{
if (nnz < 0)
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("sparsevec cannot have negative number of elements")));
if (nnz > SPARSEVEC_MAX_NNZ)
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
errmsg("sparsevec cannot have more than %d non-zero elements", SPARSEVEC_MAX_NNZ)));
if (nnz > dim)
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
errmsg("sparsevec cannot have more elements than dimensions")));
}
/*
* Ensure valid index
*/
static inline void
CheckIndex(int32 *indices, int i, int dim)
{
int32 index = indices[i];
if (index < 1)
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("index must be greater than zero")));
if (index > dim)
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("index must be less than or equal to dimensions")));
if (i > 0)
{
if (index < indices[i - 1])
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("indexes must be in ascending order")));
if (index == indices[i - 1])
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("indexes must not contain duplicates")));
}
}
/*
* Ensure finite element
*/
static inline void
CheckElement(float value)
{
if (isnan(value))
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("NaN not allowed in sparsevec")));
if (isinf(value))
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("infinite value not allowed in sparsevec")));
}
/*
* Allocate and initialize a new sparse vector
*/
SparseVector *
InitSparseVector(int dim, int nnz)
{
SparseVector *result;
int size;
size = SPARSEVEC_SIZE(nnz);
result = (SparseVector *) palloc0(size);
SET_VARSIZE(result, size);
result->dim = dim;
result->nnz = nnz;
return result;
}
/*
* Check for whitespace, since array_isspace() is static
*/
static inline bool
sparsevec_isspace(char ch)
{
if (ch == ' ' ||
ch == '\t' ||
ch == '\n' ||
ch == '\r' ||
ch == '\v' ||
ch == '\f')
return true;
return false;
}
/*
* Convert textual representation to internal representation
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_in);
Datum
sparsevec_in(PG_FUNCTION_ARGS)
{
char *lit = PG_GETARG_CSTRING(0);
int32 typmod = PG_GETARG_INT32(2);
int dim;
char *pt;
char *stringEnd;
SparseVector *result;
float *rvalues;
char *litcopy = pstrdup(lit);
char *str = litcopy;
int32 *indices;
float *values;
int maxNnz;
int nnz = 0;
maxNnz = 1;
pt = str;
while (*pt != '\0')
{
if (*pt == ',')
maxNnz++;
pt++;
}
if (maxNnz > SPARSEVEC_MAX_NNZ)
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
errmsg("sparsevec cannot have more than %d non-zero elements", SPARSEVEC_MAX_NNZ)));
indices = palloc(maxNnz * sizeof(int32));
values = palloc(maxNnz * sizeof(float));
while (sparsevec_isspace(*str))
str++;
if (*str != '{')
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("malformed sparsevec literal: \"%s\"", lit),
errdetail("Vector contents must start with \"{\".")));
str++;
pt = strtok(str, ",");
stringEnd = pt;
while (pt != NULL && *stringEnd != '}')
{
long index;
float value;
/* TODO Better error */
if (nnz == maxNnz)
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("ran out of buffer: \"%s\"", lit)));
while (sparsevec_isspace(*pt))
pt++;
/* Check for empty string like float4in */
if (*pt == '\0')
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
/* Use similar logic as int2vectorin */
errno = 0;
index = strtol(pt, &stringEnd, 10);
if (stringEnd == pt)
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
if (errno == ERANGE || index < 1 || index > INT_MAX)
ereport(ERROR,
(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
errmsg("index \"%ld\" is out of range for type sparsevec", index)));
if (stringEnd == pt)
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
while (sparsevec_isspace(*stringEnd))
stringEnd++;
if (*stringEnd != ':')
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
stringEnd++;
while (sparsevec_isspace(*stringEnd))
stringEnd++;
errno = 0;
pt = stringEnd;
value = strtof(pt, &stringEnd);
if (stringEnd == pt)
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
/* Check for range error like float4in */
if (errno == ERANGE && (value == 0 || isinf(value)))
ereport(ERROR,
(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
errmsg("\"%s\" is out of range for type sparsevec", pt)));
/* TODO Decide whether to store zero values */
if (value != 0)
{
indices[nnz] = index;
values[nnz] = value;
nnz++;
}
if (*stringEnd != '\0' && *stringEnd != '}')
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
pt = strtok(NULL, ",");
}
if (stringEnd == NULL || *stringEnd != '}')
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("malformed sparsevec literal: \"%s\"", lit),
errdetail("Unexpected end of input.")));
stringEnd++;
if (*stringEnd != '/')
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("malformed sparsevec literal: \"%s\"", lit),
errdetail("Unexpected end of input.")));
stringEnd++;
/* Use similar logic as int2vectorin */
errno = 0;
pt = stringEnd;
dim = strtol(pt, &stringEnd, 10);
if (stringEnd == pt)
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("invalid input syntax for type sparsevec: \"%s\"", lit)));
/* Only whitespace is allowed after the closing brace */
while (sparsevec_isspace(*stringEnd))
stringEnd++;
if (*stringEnd != '\0')
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("malformed sparsevec literal: \"%s\"", lit),
errdetail("Junk after closing.")));
pfree(litcopy);
CheckDim(dim);
CheckExpectedDim(typmod, dim);
result = InitSparseVector(dim, nnz);
rvalues = SPARSEVEC_VALUES(result);
for (int i = 0; i < nnz; i++)
{
result->indices[i] = indices[i];
rvalues[i] = values[i];
CheckIndex(result->indices, i, dim);
CheckElement(rvalues[i]);
}
PG_RETURN_POINTER(result);
}
#define AppendChar(ptr, c) (*(ptr)++ = (c))
#define AppendFloat(ptr, f) ((ptr) += float_to_shortest_decimal_bufn((f), (ptr)))
#if PG_VERSION_NUM >= 140000
#define AppendInt(ptr, i) ((ptr) += pg_ltoa((i), (ptr)))
#else
#define AppendInt(ptr, i) \
do { \
pg_ltoa(i, ptr); \
while (*ptr != '\0') \
ptr++; \
} while (0)
#endif
/*
* Convert internal representation to textual representation
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_out);
Datum
sparsevec_out(PG_FUNCTION_ARGS)
{
SparseVector *sparsevec = PG_GETARG_SPARSEVEC_P(0);
float *values = SPARSEVEC_VALUES(sparsevec);
char *buf;
char *ptr;
/*
* Need:
*
* nnz * 10 bytes for index (positive integer)
*
* nnz bytes for :
*
* nnz * (FLOAT_SHORTEST_DECIMAL_LEN - 1) bytes for
* float_to_shortest_decimal_bufn
*
* nnz - 1 bytes for ,
*
* 10 bytes for dimensions
*
* 4 bytes for {, }, /, and \0
*/
buf = (char *) palloc((11 + FLOAT_SHORTEST_DECIMAL_LEN) * sparsevec->nnz + 13);
ptr = buf;
AppendChar(ptr, '{');
for (int i = 0; i < sparsevec->nnz; i++)
{
if (i > 0)
AppendChar(ptr, ',');
AppendInt(ptr, sparsevec->indices[i]);
AppendChar(ptr, ':');
AppendFloat(ptr, values[i]);
}
AppendChar(ptr, '}');
AppendChar(ptr, '/');
AppendInt(ptr, sparsevec->dim);
*ptr = '\0';
PG_FREE_IF_COPY(sparsevec, 0);
PG_RETURN_CSTRING(buf);
}
/*
* Convert type modifier
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_typmod_in);
Datum
sparsevec_typmod_in(PG_FUNCTION_ARGS)
{
ArrayType *ta = PG_GETARG_ARRAYTYPE_P(0);
int32 *tl;
int n;
tl = ArrayGetIntegerTypmods(ta, &n);
if (n != 1)
ereport(ERROR,
(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("invalid type modifier")));
if (*tl < 1)
ereport(ERROR,
(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("dimensions for type sparsevec must be at least 1")));
if (*tl > SPARSEVEC_MAX_DIM)
ereport(ERROR,
(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("dimensions for type sparsevec cannot exceed %d", SPARSEVEC_MAX_DIM)));
PG_RETURN_INT32(*tl);
}
/*
* Convert external binary representation to internal representation
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_recv);
Datum
sparsevec_recv(PG_FUNCTION_ARGS)
{
StringInfo buf = (StringInfo) PG_GETARG_POINTER(0);
int32 typmod = PG_GETARG_INT32(2);
SparseVector *result;
int32 dim;
int32 nnz;
int32 unused;
float *values;
dim = pq_getmsgint(buf, sizeof(int32));
nnz = pq_getmsgint(buf, sizeof(int32));
unused = pq_getmsgint(buf, sizeof(int32));
CheckDim(dim);
CheckNnz(nnz, dim);
CheckExpectedDim(typmod, dim);
if (unused != 0)
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("expected unused to be 0, not %d", unused)));
result = InitSparseVector(dim, nnz);
values = SPARSEVEC_VALUES(result);
for (int i = 0; i < nnz; i++)
{
result->indices[i] = pq_getmsgint(buf, sizeof(int32));
CheckIndex(result->indices, i, dim);
}
for (int i = 0; i < nnz; i++)
{
values[i] = pq_getmsgfloat4(buf);
CheckElement(values[i]);
}
PG_RETURN_POINTER(result);
}
/*
* Convert internal representation to the external binary representation
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_send);
Datum
sparsevec_send(PG_FUNCTION_ARGS)
{
SparseVector *svec = PG_GETARG_SPARSEVEC_P(0);
float *values = SPARSEVEC_VALUES(svec);
StringInfoData buf;
pq_begintypsend(&buf);
pq_sendint(&buf, svec->dim, sizeof(int32));
pq_sendint(&buf, svec->nnz, sizeof(int32));
pq_sendint(&buf, svec->unused, sizeof(int32));
for (int i = 0; i < svec->nnz; i++)
pq_sendint(&buf, svec->indices[i], sizeof(int32));
for (int i = 0; i < svec->nnz; i++)
pq_sendfloat4(&buf, values[i]);
PG_RETURN_BYTEA_P(pq_endtypsend(&buf));
}
/*
* Convert sparse vector to sparse vector
* This is needed to check the type modifier
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec);
Datum
sparsevec(PG_FUNCTION_ARGS)
{
SparseVector *svec = PG_GETARG_SPARSEVEC_P(0);
int32 typmod = PG_GETARG_INT32(1);
CheckExpectedDim(typmod, svec->dim);
PG_RETURN_POINTER(svec);
}
/*
* Convert dense vector to sparse vector
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_to_sparsevec);
Datum
vector_to_sparsevec(PG_FUNCTION_ARGS)
{
Vector *vec = PG_GETARG_VECTOR_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 (vec->x[i] != 0)
nnz++;
}
result = InitSparseVector(dim, nnz);
values = SPARSEVEC_VALUES(result);
for (int i = 0; i < dim; i++)
{
if (vec->x[i] != 0)
{
/* Safety check */
if (j == nnz)
elog(ERROR, "safety check failed");
result->indices[j] = i + 1;
values[j] = vec->x[i];
j++;
}
}
PG_RETURN_POINTER(result);
}
/*
* Get the L2 squared distance between sparse vectors
*/
static double
l2_distance_squared_internal(SparseVector * a, SparseVector * b)
{
float *ax = SPARSEVEC_VALUES(a);
float *bx = SPARSEVEC_VALUES(b);
double distance = 0.0;
int bpos = 0;
for (int i = 0; i < a->nnz; i++)
{
int ai = a->indices[i];
int bi = -1;
for (int j = bpos; j < b->nnz; j++)
{
bi = b->indices[j];
if (ai == bi)
{
double diff = ax[i] - bx[j];
distance += diff * diff;
}
else if (ai > bi)
distance += bx[j] * bx[j];
/* Update start for next iteration */
if (ai >= bi)
bpos = j + 1;
/* Found or passed it */
if (bi >= ai)
break;
}
if (ai != bi)
distance += ax[i] * ax[i];
}
for (int j = bpos; j < b->nnz; j++)
distance += bx[j] * bx[j];
return distance;
}
/*
* Get the L2 distance between sparse vectors
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_l2_distance);
Datum
sparsevec_l2_distance(PG_FUNCTION_ARGS)
{
SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
CheckDims(a, b);
PG_RETURN_FLOAT8(sqrt(l2_distance_squared_internal(a, b)));
}
/*
* Get the L2 squared distance between sparse vectors
* This saves a sqrt calculation
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_l2_squared_distance);
Datum
sparsevec_l2_squared_distance(PG_FUNCTION_ARGS)
{
SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
CheckDims(a, b);
PG_RETURN_FLOAT8(l2_distance_squared_internal(a, b));
}
/*
* Get the inner product of two sparse vectors
*/
static double
inner_product_internal(SparseVector * a, SparseVector * b)
{
float *ax = SPARSEVEC_VALUES(a);
float *bx = SPARSEVEC_VALUES(b);
double distance = 0.0;
int bpos = 0;
for (int i = 0; i < a->nnz; i++)
{
int ai = a->indices[i];
for (int j = bpos; j < b->nnz; j++)
{
int bi = b->indices[j];
/* Only update when the same index */
if (ai == bi)
distance += ax[i] * bx[j];
/* Update start for next iteration */
if (ai >= bi)
bpos = j + 1;
/* Found or passed it */
if (bi >= ai)
break;
}
}
return distance;
}
/*
* Get the inner product of two sparse vectors
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_inner_product);
Datum
sparsevec_inner_product(PG_FUNCTION_ARGS)
{
SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
CheckDims(a, b);
PG_RETURN_FLOAT8(inner_product_internal(a, b));
}
/*
* Get the negative inner product of two sparse vectors
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_negative_inner_product);
Datum
sparsevec_negative_inner_product(PG_FUNCTION_ARGS)
{
SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
CheckDims(a, b);
PG_RETURN_FLOAT8(-inner_product_internal(a, b));
}
/*
* Get the cosine distance between two sparse vectors
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_cosine_distance);
Datum
sparsevec_cosine_distance(PG_FUNCTION_ARGS)
{
SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
SparseVector *b = PG_GETARG_SPARSEVEC_P(1);
float *ax = SPARSEVEC_VALUES(a);
float *bx = SPARSEVEC_VALUES(b);
float norma = 0.0;
float normb = 0.0;
double similarity;
CheckDims(a, b);
similarity = inner_product_internal(a, b);
/* Auto-vectorized */
for (int i = 0; i < a->nnz; i++)
norma += ax[i] * ax[i];
/* Auto-vectorized */
for (int i = 0; i < b->nnz; i++)
normb += bx[i] * bx[i];
/* Use sqrt(a * b) over sqrt(a) * sqrt(b) */
similarity /= sqrt((double) norma * (double) normb);
#ifdef _MSC_VER
/* /fp:fast may not propagate NaN */
if (isnan(similarity))
PG_RETURN_FLOAT8(NAN);
#endif
/* Keep in range */
if (similarity > 1)
similarity = 1.0;
else if (similarity < -1)
similarity = -1.0;
PG_RETURN_FLOAT8(1.0 - similarity);
}
/*
* Get the L2 norm of a sparse vector
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_norm);
Datum
sparsevec_norm(PG_FUNCTION_ARGS)
{
SparseVector *a = PG_GETARG_SPARSEVEC_P(0);
float *ax = SPARSEVEC_VALUES(a);
double norm = 0.0;
/* Auto-vectorized */
for (int i = 0; i < a->nnz; i++)
norm += (double) ax[i] * (double) ax[i];
PG_RETURN_FLOAT8(sqrt(norm));
}

View File

@@ -1,25 +0,0 @@
#ifndef SPARSEVEC_H
#define SPARSEVEC_H
#define SPARSEVEC_MAX_DIM 100000
#define SPARSEVEC_MAX_NNZ 16000
/* Ensure values are aligned */
#define SPARSEVEC_SIZE(_nnz) (offsetof(SparseVector, indices) + MAXALIGN((_nnz) * sizeof(int32)) + (_nnz * sizeof(float)))
#define SPARSEVEC_VALUES(x) ((float *) (((char *) (x)) + offsetof(SparseVector, indices) + MAXALIGN((x)->nnz * sizeof(int32))))
#define DatumGetSparseVector(x) ((SparseVector *) PG_DETOAST_DATUM(x))
#define PG_GETARG_SPARSEVEC_P(x) DatumGetSparseVector(PG_GETARG_DATUM(x))
#define PG_RETURN_SPARSEVEC_P(x) PG_RETURN_POINTER(x)
typedef struct SparseVector
{
int32 vl_len_; /* varlena header (do not touch directly!) */
int32 dim; /* number of dimensions */
int32 nnz;
int32 unused;
int32 indices[FLEXIBLE_ARRAY_MEMBER];
} SparseVector;
SparseVector *InitSparseVector(int dim, int nnz);
#endif

View File

@@ -2,26 +2,24 @@
#include <math.h>
#include "bitvector.h"
#include "catalog/pg_type.h"
#include "common/shortest_dec.h"
#include "fmgr.h"
#include "halfvec.h"
#include "hnsw.h"
#include "ivfflat.h"
#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"
#include "utils/lsyscache.h"
#include "utils/numeric.h"
#include "vector.h"
#if PG_VERSION_NUM >= 160000
#include "varatt.h"
#if PG_VERSION_NUM >= 120000
#include "common/shortest_dec.h"
#include "utils/float.h"
#else
#include <float.h>
#endif
#if PG_VERSION_NUM < 130000
@@ -37,7 +35,6 @@ PG_MODULE_MAGIC;
/*
* Initialize index options and variables
*/
PGDLLEXPORT void _PG_init(void);
void
_PG_init(void)
{
@@ -87,7 +84,7 @@ CheckDim(int dim)
}
/*
* Ensure finite element
* Ensure finite elements
*/
static inline void
CheckElement(float value)
@@ -103,23 +100,6 @@ CheckElement(float value)
errmsg("infinite value not allowed in vector")));
}
/*
* Allocate and initialize a new vector
*/
Vector *
InitVector(int dim)
{
Vector *result;
int size;
size = VECTOR_SIZE(dim);
result = (Vector *) palloc0(size);
SET_VARSIZE(result, size);
result->dim = dim;
return result;
}
/*
* Check for whitespace, since array_isspace() is static
*/
@@ -175,15 +155,14 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_in);
Datum
vector_in(PG_FUNCTION_ARGS)
{
char *lit = PG_GETARG_CSTRING(0);
char *str = PG_GETARG_CSTRING(0);
int32 typmod = PG_GETARG_INT32(2);
float x[VECTOR_MAX_DIM];
int dim = 0;
char *pt;
char *stringEnd;
Vector *result;
char *litcopy = pstrdup(lit);
char *str = litcopy;
char *lit = pstrdup(str);
while (vector_isspace(*str))
str++;
@@ -200,8 +179,6 @@ vector_in(PG_FUNCTION_ARGS)
while (pt != NULL && *stringEnd != ']')
{
float val;
if (dim == VECTOR_MAX_DIM)
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
@@ -217,23 +194,15 @@ vector_in(PG_FUNCTION_ARGS)
errmsg("invalid input syntax for type vector: \"%s\"", lit)));
/* Use strtof like float4in to avoid a double-rounding problem */
errno = 0;
val = strtof(pt, &stringEnd);
x[dim] = strtof(pt, &stringEnd);
CheckElement(x[dim]);
dim++;
if (stringEnd == pt)
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("invalid input syntax for type vector: \"%s\"", lit)));
/* Check for range error like float4in */
if (errno == ERANGE && isinf(val))
ereport(ERROR,
(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
errmsg("\"%s\" is out of range for type vector", pt)));
CheckElement(val);
x[dim++] = val;
while (vector_isspace(*stringEnd))
stringEnd++;
@@ -277,7 +246,7 @@ vector_in(PG_FUNCTION_ARGS)
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("vector must have at least 1 dimension")));
pfree(litcopy);
pfree(lit);
CheckExpectedDim(typmod, dim);
@@ -288,9 +257,6 @@ vector_in(PG_FUNCTION_ARGS)
PG_RETURN_POINTER(result);
}
#define AppendChar(ptr, c) (*(ptr)++ = (c))
#define AppendFloat(ptr, f) ((ptr) += float_to_shortest_decimal_bufn((f), (ptr)))
/*
* Convert internal representation to textual representation
*/
@@ -302,6 +268,16 @@ vector_out(PG_FUNCTION_ARGS)
int dim = vector->dim;
char *buf;
char *ptr;
int n;
#if PG_VERSION_NUM < 120000
int ndig = FLT_DIG + extra_float_digits;
if (ndig < 1)
ndig = 1;
#define FLOAT_SHORTEST_DECIMAL_LEN (ndig + 10)
#endif
/*
* Need:
@@ -316,17 +292,25 @@ vector_out(PG_FUNCTION_ARGS)
buf = (char *) palloc(FLOAT_SHORTEST_DECIMAL_LEN * dim + 2);
ptr = buf;
AppendChar(ptr, '[');
*ptr = '[';
ptr++;
for (int i = 0; i < dim; i++)
{
if (i > 0)
AppendChar(ptr, ',');
{
*ptr = ',';
ptr++;
}
AppendFloat(ptr, vector->x[i]);
#if PG_VERSION_NUM >= 120000
n = float_to_shortest_decimal_bufn(vector->x[i], ptr);
#else
n = sprintf(ptr, "%.*g", ndig, vector->x[i]);
#endif
ptr += n;
}
AppendChar(ptr, ']');
*ptr = ']';
ptr++;
*ptr = '\0';
PG_FREE_IF_COPY(vector, 0);
@@ -431,18 +415,17 @@ vector_send(PG_FUNCTION_ARGS)
/*
* Convert vector to vector
* This is needed to check the type modifier
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(vector);
Datum
vector(PG_FUNCTION_ARGS)
{
Vector *vec = PG_GETARG_VECTOR_P(0);
Vector *arg = PG_GETARG_VECTOR_P(0);
int32 typmod = PG_GETARG_INT32(1);
CheckExpectedDim(typmod, vec->dim);
CheckExpectedDim(typmod, arg->dim);
PG_RETURN_POINTER(vec);
PG_RETURN_POINTER(arg);
}
/*
@@ -459,6 +442,7 @@ array_to_vector(PG_FUNCTION_ARGS)
bool typbyval;
char typalign;
Datum *elemsp;
bool *nullsp;
int nelemsp;
if (ARR_NDIM(array) > 1)
@@ -472,7 +456,7 @@ array_to_vector(PG_FUNCTION_ARGS)
errmsg("array must not contain nulls")));
get_typlenbyvalalign(ARR_ELEMTYPE(array), &typlen, &typbyval, &typalign);
deconstruct_array(array, ARR_ELEMTYPE(array), typlen, typbyval, typalign, &elemsp, NULL, &nelemsp);
deconstruct_array(array, ARR_ELEMTYPE(array), typlen, typbyval, typalign, &elemsp, &nullsp, &nelemsp);
CheckDim(nelemsp);
CheckExpectedDim(typmod, nelemsp);
@@ -506,12 +490,6 @@ array_to_vector(PG_FUNCTION_ARGS)
errmsg("unsupported array type")));
}
/*
* Free allocation from deconstruct_array. Do not free individual elements
* when pass-by-reference since they point to original array.
*/
pfree(elemsp);
/* Check elements */
for (int i = 0; i < result->dim; i++)
CheckElement(result->x[i]);
@@ -543,28 +521,6 @@ vector_to_float4(PG_FUNCTION_ARGS)
PG_RETURN_POINTER(result);
}
/*
* Convert half vector to vector
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_to_vector);
Datum
halfvec_to_vector(PG_FUNCTION_ARGS)
{
HalfVector *vec = PG_GETARG_HALFVEC_P(0);
int32 typmod = PG_GETARG_INT32(1);
Vector *result;
CheckDim(vec->dim);
CheckExpectedDim(typmod, vec->dim);
result = InitVector(vec->dim);
for (int i = 0; i < vec->dim; i++)
result->x[i] = HalfToFloat4(vec->x[i]);
PG_RETURN_POINTER(result);
}
/*
* Get the L2 distance between vectors
*/
@@ -676,7 +632,6 @@ cosine_distance(PG_FUNCTION_ARGS)
float distance = 0.0;
float norma = 0.0;
float normb = 0.0;
double similarity;
CheckDims(a, b);
@@ -689,21 +644,7 @@ cosine_distance(PG_FUNCTION_ARGS)
}
/* Use sqrt(a * b) over sqrt(a) * sqrt(b) */
similarity = (double) distance / sqrt((double) norma * (double) normb);
#ifdef _MSC_VER
/* /fp:fast may not propagate NaN */
if (isnan(similarity))
PG_RETURN_FLOAT8(NAN);
#endif
/* Keep in range */
if (similarity > 1)
similarity = 1.0;
else if (similarity < -1)
similarity = -1.0;
PG_RETURN_FLOAT8(1.0 - similarity);
PG_RETURN_FLOAT8(1.0 - ((double) distance / sqrt((double) norma * (double) normb)));
}
/*
@@ -717,8 +658,6 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
Vector *b = PG_GETARG_VECTOR_P(1);
float *ax = a->x;
float *bx = b->x;
float dp = 0.0;
double distance;
@@ -726,10 +665,9 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
/* Auto-vectorized */
for (int i = 0; i < a->dim; i++)
dp += ax[i] * bx[i];
dp += a->x[i] * b->x[i];
distance = (double) dp;
/* Prevent NaN with acos with loss of precision */
if (distance > 1)
distance = 1;
@@ -740,7 +678,7 @@ vector_spherical_distance(PG_FUNCTION_ARGS)
}
/*
* Get the L1 distance between two vectors
* Get the L1 distance between vectors
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(l1_distance);
Datum
@@ -782,13 +720,13 @@ vector_norm(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
float *ax = a->x;
double norm = 0.0;
float norm = 0.0;
/* Auto-vectorized */
for (int i = 0; i < a->dim; i++)
norm += (double) ax[i] * (double) ax[i];
norm += ax[i] * ax[i];
PG_RETURN_FLOAT8(sqrt(norm));
PG_RETURN_FLOAT8(sqrt((double) norm));
}
/*
@@ -893,66 +831,15 @@ vector_mul(PG_FUNCTION_ARGS)
PG_RETURN_POINTER(result);
}
/*
* Quantize a vector
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(quantize_binary);
Datum
quantize_binary(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
float *ax = a->x;
VarBit *result = InitBitVector(a->dim);
unsigned char *rx = VARBITS(result);
for (int i = 0; i < a->dim; i++)
rx[i / 8] |= (ax[i] > 0) << (7 - (i % 8));
PG_RETURN_VARBIT_P(result);
}
/*
* Get a subvector
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(subvector);
Datum
subvector(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
int32 start = PG_GETARG_INT32(1);
int32 count = PG_GETARG_INT32(2);
int32 end = start + count;
float *ax = a->x;
Vector *result;
int dim;
/* Indexing starts at 1, like substring */
if (start < 1)
start = 1;
if (end > a->dim)
end = a->dim + 1;
dim = end - start;
CheckDim(dim);
result = InitVector(dim);
for (int i = 0; i < dim; i++)
result->x[i] = ax[start - 1 + i];
PG_RETURN_POINTER(result);
}
/*
* Internal helper to compare vectors
*/
int
vector_cmp_internal(Vector * a, Vector * b)
{
int dim = Min(a->dim, b->dim);
CheckDims(a, b);
/* Check values before dimensions to be consistent with Postgres arrays */
for (int i = 0; i < dim; i++)
for (int i = 0; i < a->dim; i++)
{
if (a->x[i] < b->x[i])
return -1;
@@ -960,13 +847,6 @@ vector_cmp_internal(Vector * a, Vector * b)
if (a->x[i] > b->x[i])
return 1;
}
if (a->dim < b->dim)
return -1;
if (a->dim > b->dim)
return 1;
return 0;
}
@@ -977,8 +857,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_lt);
Datum
vector_lt(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
Vector *b = PG_GETARG_VECTOR_P(1);
Vector *a = (Vector *) PG_GETARG_VECTOR_P(0);
Vector *b = (Vector *) PG_GETARG_VECTOR_P(1);
PG_RETURN_BOOL(vector_cmp_internal(a, b) < 0);
}
@@ -990,8 +870,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_le);
Datum
vector_le(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
Vector *b = PG_GETARG_VECTOR_P(1);
Vector *a = (Vector *) PG_GETARG_VECTOR_P(0);
Vector *b = (Vector *) PG_GETARG_VECTOR_P(1);
PG_RETURN_BOOL(vector_cmp_internal(a, b) <= 0);
}
@@ -1003,8 +883,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_eq);
Datum
vector_eq(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
Vector *b = PG_GETARG_VECTOR_P(1);
Vector *a = (Vector *) PG_GETARG_VECTOR_P(0);
Vector *b = (Vector *) PG_GETARG_VECTOR_P(1);
PG_RETURN_BOOL(vector_cmp_internal(a, b) == 0);
}
@@ -1016,8 +896,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_ne);
Datum
vector_ne(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
Vector *b = PG_GETARG_VECTOR_P(1);
Vector *a = (Vector *) PG_GETARG_VECTOR_P(0);
Vector *b = (Vector *) PG_GETARG_VECTOR_P(1);
PG_RETURN_BOOL(vector_cmp_internal(a, b) != 0);
}
@@ -1029,8 +909,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_ge);
Datum
vector_ge(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
Vector *b = PG_GETARG_VECTOR_P(1);
Vector *a = (Vector *) PG_GETARG_VECTOR_P(0);
Vector *b = (Vector *) PG_GETARG_VECTOR_P(1);
PG_RETURN_BOOL(vector_cmp_internal(a, b) >= 0);
}
@@ -1042,8 +922,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_gt);
Datum
vector_gt(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
Vector *b = PG_GETARG_VECTOR_P(1);
Vector *a = (Vector *) PG_GETARG_VECTOR_P(0);
Vector *b = (Vector *) PG_GETARG_VECTOR_P(1);
PG_RETURN_BOOL(vector_cmp_internal(a, b) > 0);
}
@@ -1055,8 +935,8 @@ PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_cmp);
Datum
vector_cmp(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
Vector *b = PG_GETARG_VECTOR_P(1);
Vector *a = (Vector *) PG_GETARG_VECTOR_P(0);
Vector *b = (Vector *) PG_GETARG_VECTOR_P(1);
PG_RETURN_INT32(vector_cmp_internal(a, b));
}
@@ -1225,26 +1105,3 @@ vector_avg(PG_FUNCTION_ARGS)
PG_RETURN_POINTER(result);
}
/*
* Convert sparse vector to dense vector
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(sparsevec_to_vector);
Datum
sparsevec_to_vector(PG_FUNCTION_ARGS)
{
SparseVector *svec = PG_GETARG_SPARSEVEC_P(0);
int32 typmod = PG_GETARG_INT32(1);
Vector *result;
int dim = svec->dim;
float *values = SPARSEVEC_VALUES(svec);
CheckDim(dim);
CheckExpectedDim(typmod, dim);
result = InitVector(dim);
for (int i = 0; i < svec->nnz; i++)
result->x[svec->indices[i] - 1] = values[i];
PG_RETURN_POINTER(result);
}

View File

@@ -1,6 +1,12 @@
#ifndef VECTOR_H
#define VECTOR_H
#include "postgres.h"
#if PG_VERSION_NUM >= 160000
#include "varatt.h"
#endif
#define VECTOR_MAX_DIM 16000
#define VECTOR_SIZE(_dim) (offsetof(Vector, x) + sizeof(float)*(_dim))
@@ -8,6 +14,9 @@
#define PG_GETARG_VECTOR_P(x) DatumGetVector(PG_GETARG_DATUM(x))
#define PG_RETURN_VECTOR_P(x) PG_RETURN_POINTER(x)
/* Exported functions */
PGDLLEXPORT void _PG_init(void);
typedef struct Vector
{
int32 vl_len_; /* varlena header (do not touch directly!) */
@@ -16,8 +25,24 @@ typedef struct Vector
float x[FLEXIBLE_ARRAY_MEMBER];
} Vector;
Vector *InitVector(int dim);
void PrintVector(char *msg, Vector * vector);
int vector_cmp_internal(Vector * a, Vector * b);
/*
* Allocate and initialize a new vector
*/
static inline Vector *
InitVector(int dim)
{
Vector *result;
int size;
size = VECTOR_SIZE(dim);
result = (Vector *) palloc0(size);
SET_VARSIZE(result, size);
result->dim = dim;
return result;
}
#endif

View File

@@ -1,104 +0,0 @@
SELECT hamming_distance('111', '111');
hamming_distance
------------------
0
(1 row)
SELECT hamming_distance('111', '110');
hamming_distance
------------------
1
(1 row)
SELECT hamming_distance('111', '100');
hamming_distance
------------------
2
(1 row)
SELECT hamming_distance('111', '000');
hamming_distance
------------------
3
(1 row)
SELECT hamming_distance('10101010101010101010', '01010101010101010101');
hamming_distance
------------------
20
(1 row)
SELECT hamming_distance('', '');
hamming_distance
------------------
0
(1 row)
SELECT hamming_distance('111', '00');
ERROR: different bit lengths 3 and 2
SELECT hamming_distance('111', '000'::varbit(4));
hamming_distance
------------------
3
(1 row)
SELECT hamming_distance('111', '0000'::varbit(4));
ERROR: different bit lengths 3 and 4
SELECT jaccard_distance('1111', '1111');
jaccard_distance
------------------
0
(1 row)
SELECT jaccard_distance('1111', '1110');
jaccard_distance
------------------
0.25
(1 row)
SELECT jaccard_distance('1111', '1100');
jaccard_distance
------------------
0.5
(1 row)
SELECT jaccard_distance('1111', '1000');
jaccard_distance
------------------
0.75
(1 row)
SELECT jaccard_distance('1111', '0000');
jaccard_distance
------------------
1
(1 row)
SELECT jaccard_distance('1100', '1000');
jaccard_distance
------------------
0.5
(1 row)
SELECT jaccard_distance('10101010101010101010', '01010101010101010101');
jaccard_distance
------------------
1
(1 row)
SELECT jaccard_distance('', '');
jaccard_distance
------------------
1
(1 row)
SELECT jaccard_distance('1111', '000');
ERROR: different bit lengths 4 and 3
SELECT jaccard_distance('1111', '0000'::varbit(5));
jaccard_distance
------------------
1
(1 row)
SELECT jaccard_distance('1111', '00000'::varbit(5));
ERROR: different bit lengths 4 and 5

View File

@@ -46,30 +46,6 @@ SELECT '[1,2,3]'::vector::real[];
{1,2,3}
(1 row)
SELECT '[1,2,3]'::vector::halfvec;
halfvec
---------
[1,2,3]
(1 row)
SELECT '[1,2,3]'::halfvec::vector;
vector
---------
[1,2,3]
(1 row)
SELECT '[1,2,3]'::vector::halfvec(2);
ERROR: expected 2 dimensions, not 3
SELECT '[1,2,3]'::halfvec::vector(2);
ERROR: expected 2 dimensions, not 3
SELECT '[65520]'::vector::halfvec;
ERROR: infinite value not allowed in halfvec
SELECT '[1e-8]'::vector::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;

View File

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

162
test/expected/functions.out Normal file
View File

@@ -0,0 +1,162 @@
SELECT '[1,2,3]'::vector + '[4,5,6]';
?column?
----------
[5,7,9]
(1 row)
SELECT '[3e38]'::vector + '[3e38]';
ERROR: value out of range: overflow
SELECT '[1,2,3]'::vector - '[4,5,6]';
?column?
------------
[-3,-3,-3]
(1 row)
SELECT '[-3e38]'::vector - '[3e38]';
ERROR: value out of range: overflow
SELECT '[1,2,3]'::vector * '[4,5,6]';
?column?
-----------
[4,10,18]
(1 row)
SELECT '[1e37]'::vector * '[1e37]';
ERROR: value out of range: overflow
SELECT '[1e-37]'::vector * '[1e-37]';
ERROR: value out of range: underflow
SELECT vector_dims('[1,2,3]');
vector_dims
-------------
3
(1 row)
SELECT round(vector_norm('[1,1]')::numeric, 5);
round
---------
1.41421
(1 row)
SELECT vector_norm('[3,4]');
vector_norm
-------------
5
(1 row)
SELECT vector_norm('[0,1]');
vector_norm
-------------
1
(1 row)
SELECT l2_distance('[0,0]', '[3,4]');
l2_distance
-------------
5
(1 row)
SELECT l2_distance('[0,0]', '[0,1]');
l2_distance
-------------
1
(1 row)
SELECT l2_distance('[1,2]', '[3]');
ERROR: different vector dimensions 2 and 1
SELECT inner_product('[1,2]', '[3,4]');
inner_product
---------------
11
(1 row)
SELECT inner_product('[1,2]', '[3]');
ERROR: different vector dimensions 2 and 1
SELECT cosine_distance('[1,2]', '[2,4]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,2]', '[0,0]');
cosine_distance
-----------------
NaN
(1 row)
SELECT cosine_distance('[1,1]', '[1,1]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,1]', '[-1,-1]');
cosine_distance
-----------------
2
(1 row)
SELECT cosine_distance('[1,2]', '[3]');
ERROR: different vector dimensions 2 and 1
SELECT l1_distance('[0,0]', '[3,4]');
l1_distance
-------------
7
(1 row)
SELECT l1_distance('[0,0]', '[0,1]');
l1_distance
-------------
1
(1 row)
SELECT l1_distance('[1,2]', '[3]');
ERROR: different vector dimensions 2 and 1
SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
avg
-----------
[2,3.5,5]
(1 row)
SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;
avg
-----------
[2,3.5,5]
(1 row)
SELECT avg(v) FROM unnest(ARRAY[]::vector[]) v;
avg
-----
(1 row)
SELECT avg(v) FROM unnest(ARRAY['[1,2]'::vector, '[3]']) v;
ERROR: expected 2 dimensions, not 1
SELECT avg(v) FROM unnest(ARRAY['[3e38]'::vector, '[3e38]']) v;
avg
---------
[3e+38]
(1 row)
SELECT vector_avg(array_agg(n)) FROM generate_series(1, 16002) n;
ERROR: vector cannot have more than 16000 dimensions
SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
sum
----------
[4,7,10]
(1 row)
SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;
sum
----------
[4,7,10]
(1 row)
SELECT sum(v) FROM unnest(ARRAY[]::vector[]) v;
sum
-----
(1 row)
SELECT sum(v) FROM unnest(ARRAY['[1,2]'::vector, '[3]']) v;
ERROR: different vector dimensions 2 and 1
SELECT sum(v) FROM unnest(ARRAY['[3e38]'::vector, '[3e38]']) v;
ERROR: value out of range: overflow

View File

@@ -1,146 +0,0 @@
SELECT l2_distance('[0,0]'::halfvec, '[3,4]');
l2_distance
-------------
5
(1 row)
SELECT l2_distance('[0,0]'::halfvec, '[0,1]');
l2_distance
-------------
1
(1 row)
SELECT l2_distance('[1,2]'::halfvec, '[3]');
ERROR: different halfvec dimensions 2 and 1
SELECT '[0,0]'::halfvec <-> '[3,4]';
?column?
----------
5
(1 row)
SELECT inner_product('[1,2]'::halfvec, '[3,4]');
inner_product
---------------
11
(1 row)
SELECT inner_product('[1,2]'::halfvec, '[3]');
ERROR: different halfvec dimensions 2 and 1
SELECT inner_product('[65504]'::halfvec, '[65504]');
inner_product
---------------
4290774016
(1 row)
SELECT '[1,2]'::halfvec <#> '[3,4]';
?column?
----------
-11
(1 row)
SELECT cosine_distance('[1,2]'::halfvec, '[2,4]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,2]'::halfvec, '[0,0]');
cosine_distance
-----------------
NaN
(1 row)
SELECT cosine_distance('[1,1]'::halfvec, '[1,1]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,0]'::halfvec, '[0,2]');
cosine_distance
-----------------
1
(1 row)
SELECT cosine_distance('[1,1]'::halfvec, '[-1,-1]');
cosine_distance
-----------------
2
(1 row)
SELECT cosine_distance('[1,2]'::halfvec, '[3]');
ERROR: different halfvec dimensions 2 and 1
SELECT cosine_distance('[1,1]'::halfvec, '[1.1,1.1]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,1]'::halfvec, '[-1.1,-1.1]');
cosine_distance
-----------------
2
(1 row)
SELECT '[1,2]'::halfvec <=> '[2,4]';
?column?
----------
0
(1 row)
SELECT l1_distance('[0,0]'::halfvec, '[3,4]');
l1_distance
-------------
7
(1 row)
SELECT l1_distance('[0,0]'::halfvec, '[0,1]');
l1_distance
-------------
1
(1 row)
SELECT l1_distance('[1,2]'::halfvec, '[3]');
ERROR: different halfvec dimensions 2 and 1
SELECT quantize_binary('[1,0,-1]'::halfvec);
quantize_binary
-----------------
100
(1 row)
SELECT quantize_binary('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]'::halfvec);
quantize_binary
-----------------
01001110101
(1 row)
SELECT subvector('[1,2,3,4,5]'::halfvec, 1, 3);
subvector
-----------
[1,2,3]
(1 row)
SELECT subvector('[1,2,3,4,5]'::halfvec, 3, 2);
subvector
-----------
[3,4]
(1 row)
SELECT subvector('[1,2,3,4,5]'::halfvec, -1, 3);
subvector
-----------
[1]
(1 row)
SELECT subvector('[1,2,3,4,5]'::halfvec, 3, 9);
subvector
-----------
[3,4,5]
(1 row)
SELECT subvector('[1,2,3,4,5]'::halfvec, 1, 0);
ERROR: halfvec must have at least 1 dimension
SELECT subvector('[1,2,3,4,5]'::halfvec, 3, -1);
ERROR: halfvec must have at least 1 dimension
SELECT subvector('[1,2,3,4,5]'::halfvec, -1, 2);
ERROR: halfvec must have at least 1 dimension

View File

@@ -1,151 +0,0 @@
SELECT '[1,2,3]'::halfvec;
halfvec
---------
[1,2,3]
(1 row)
SELECT '[-1,-2,-3]'::halfvec;
halfvec
------------
[-1,-2,-3]
(1 row)
SELECT '[1.,2.,3.]'::halfvec;
halfvec
---------
[1,2,3]
(1 row)
SELECT ' [ 1, 2 , 3 ] '::halfvec;
halfvec
---------
[1,2,3]
(1 row)
SELECT '[1.23456]'::halfvec;
halfvec
------------
[1.234375]
(1 row)
SELECT '[hello,1]'::halfvec;
ERROR: invalid input syntax for type halfvec: "[hello,1]"
LINE 1: SELECT '[hello,1]'::halfvec;
^
SELECT '[NaN,1]'::halfvec;
ERROR: NaN not allowed in halfvec
LINE 1: SELECT '[NaN,1]'::halfvec;
^
SELECT '[Infinity,1]'::halfvec;
ERROR: infinite value not allowed in halfvec
LINE 1: SELECT '[Infinity,1]'::halfvec;
^
SELECT '[-Infinity,1]'::halfvec;
ERROR: infinite value not allowed in halfvec
LINE 1: SELECT '[-Infinity,1]'::halfvec;
^
SELECT '[65519,-65519]'::halfvec;
halfvec
----------------
[65504,-65504]
(1 row)
SELECT '[65520,-65520]'::halfvec;
ERROR: "65520" is out of range for type halfvec
LINE 1: SELECT '[65520,-65520]'::halfvec;
^
SELECT '[1e-8,-1e-8]'::halfvec;
ERROR: "1e-8" is out of range for type halfvec
LINE 1: SELECT '[1e-8,-1e-8]'::halfvec;
^
SELECT '[4e38,1]'::halfvec;
ERROR: "4e38" is out of range for type halfvec
LINE 1: SELECT '[4e38,1]'::halfvec;
^
SELECT '[1e-46,1]'::halfvec;
ERROR: "1e-46" is out of range for type halfvec
LINE 1: SELECT '[1e-46,1]'::halfvec;
^
SELECT '[1,2,3'::halfvec;
ERROR: malformed halfvec literal: "[1,2,3"
LINE 1: SELECT '[1,2,3'::halfvec;
^
DETAIL: Unexpected end of input.
SELECT '[1,2,3]9'::halfvec;
ERROR: malformed halfvec literal: "[1,2,3]9"
LINE 1: SELECT '[1,2,3]9'::halfvec;
^
DETAIL: Junk after closing right brace.
SELECT '1,2,3'::halfvec;
ERROR: malformed halfvec literal: "1,2,3"
LINE 1: SELECT '1,2,3'::halfvec;
^
DETAIL: Vector contents must start with "[".
SELECT ''::halfvec;
ERROR: malformed halfvec literal: ""
LINE 1: SELECT ''::halfvec;
^
DETAIL: Vector contents must start with "[".
SELECT '['::halfvec;
ERROR: malformed halfvec literal: "["
LINE 1: SELECT '['::halfvec;
^
DETAIL: Unexpected end of input.
SELECT '[,'::halfvec;
ERROR: malformed halfvec literal: "[,"
LINE 1: SELECT '[,'::halfvec;
^
DETAIL: Unexpected end of input.
SELECT '[]'::halfvec;
ERROR: halfvec must have at least 1 dimension
LINE 1: SELECT '[]'::halfvec;
^
SELECT '[1,]'::halfvec;
ERROR: invalid input syntax for type halfvec: "[1,]"
LINE 1: SELECT '[1,]'::halfvec;
^
SELECT '[1a]'::halfvec;
ERROR: invalid input syntax for type halfvec: "[1a]"
LINE 1: SELECT '[1a]'::halfvec;
^
SELECT '[1,,3]'::halfvec;
ERROR: malformed halfvec literal: "[1,,3]"
LINE 1: SELECT '[1,,3]'::halfvec;
^
SELECT '[1, ,3]'::halfvec;
ERROR: invalid input syntax for type halfvec: "[1, ,3]"
LINE 1: SELECT '[1, ,3]'::halfvec;
^
SELECT '[1,2,3]'::halfvec(3);
halfvec
---------
[1,2,3]
(1 row)
SELECT '[1,2,3]'::halfvec(2);
ERROR: expected 2 dimensions, not 3
SELECT '[1,2,3]'::halfvec(3, 2);
ERROR: invalid type modifier
LINE 1: SELECT '[1,2,3]'::halfvec(3, 2);
^
SELECT '[1,2,3]'::halfvec('a');
ERROR: invalid input syntax for type integer: "a"
LINE 1: SELECT '[1,2,3]'::halfvec('a');
^
SELECT '[1,2,3]'::halfvec(0);
ERROR: dimensions for type halfvec must be at least 1
LINE 1: SELECT '[1,2,3]'::halfvec(0);
^
SELECT '[1,2,3]'::halfvec(16001);
ERROR: dimensions for type halfvec cannot exceed 16000
LINE 1: SELECT '[1,2,3]'::halfvec(16001);
^
SELECT unnest('{"[1,2,3]", "[4,5,6]"}'::halfvec[]);
unnest
---------
[1,2,3]
[4,5,6]
(2 rows)
SELECT '{"[1,2,3]"}'::halfvec(2)[];
ERROR: expected 2 dimensions, not 3

View File

@@ -1,27 +0,0 @@
SET enable_seqscan = off;
CREATE TABLE t (val bit(3));
INSERT INTO t (val) VALUES (B'000'), (B'100'), (B'111'), (NULL);
CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
INSERT INTO t (val) VALUES (B'110');
SELECT * FROM t ORDER BY val <~> B'111';
val
-----
111
110
100
000
(4 rows)
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <~> (SELECT NULL::bit)) t2;
count
-------
4
(1 row)
DROP TABLE t;
-- TODO move
CREATE TABLE t (val varbit(3));
CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
ERROR: type not supported for hnsw index
CREATE INDEX ON t USING hnsw ((val::bit(3)) bit_hamming_ops);
DROP TABLE t;

View File

@@ -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 hnsw (val bit_jaccard_ops);
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;

View File

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

View File

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

View File

@@ -12,11 +12,14 @@ SELECT * FROM t ORDER BY val <-> '[3,3,3]';
[0,0,0]
(4 rows)
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector)) t2;
count
-------
4
(1 row)
SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
val
---------
[0,0,0]
[1,1,1]
[1,2,3]
[1,2,4]
(4 rows)
SELECT COUNT(*) FROM t;
count
@@ -24,10 +27,4 @@ SELECT COUNT(*) FROM t;
5
(1 row)
TRUNCATE t;
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
val
-----
(0 rows)
DROP TABLE t;

View File

@@ -1,26 +1,25 @@
SET enable_seqscan = off;
CREATE TABLE t (val vector(3));
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 1);
ERROR: value 1 out of bounds for option "m"
DETAIL: Valid values are between "2" and "100".
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 3);
ERROR: value 3 out of bounds for option "m"
DETAIL: Valid values are between "4" and "100".
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 101);
ERROR: value 101 out of bounds for option "m"
DETAIL: Valid values are between "2" and "100".
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 3);
ERROR: value 3 out of bounds for option "ef_construction"
DETAIL: Valid values are between "4" and "1000".
DETAIL: Valid values are between "4" and "100".
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 9);
ERROR: value 9 out of bounds for option "ef_construction"
DETAIL: Valid values are between "10" and "1000".
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 1001);
ERROR: value 1001 out of bounds for option "ef_construction"
DETAIL: Valid values are between "4" and "1000".
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 16, ef_construction = 31);
ERROR: ef_construction must be greater than or equal to 2 * m
DETAIL: Valid values are between "10" and "1000".
SHOW hnsw.ef_search;
hnsw.ef_search
----------------
40
(1 row)
SET hnsw.ef_search = 0;
ERROR: 0 is outside the valid range for parameter "hnsw.ef_search" (1 .. 1000)
SET hnsw.ef_search = 9;
ERROR: 9 is outside the valid range for parameter "hnsw.ef_search" (10 .. 1000)
SET hnsw.ef_search = 1001;
ERROR: 1001 is outside the valid range for parameter "hnsw.ef_search" (1 .. 1000)
ERROR: 1001 is outside the valid range for parameter "hnsw.ef_search" (10 .. 1000)
DROP TABLE t;

View File

@@ -1,26 +0,0 @@
SET enable_seqscan = off;
CREATE TABLE t (val sparsevec(3));
INSERT INTO t (val) VALUES ('{}/3'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
CREATE INDEX ON t USING hnsw (val sparsevec_cosine_ops);
INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
SELECT * FROM t ORDER BY val <=> '{1:3,2:3,3:3}/3';
val
-----------------
{1:1,2:1,3:1}/3
{1:1,2:2,3:3}/3
{1:1,2:2,3:4}/3
(3 rows)
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '{}/3') t2;
count
-------
3
(1 row)
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::sparsevec)) t2;
count
-------
3
(1 row)
DROP TABLE t;

View File

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

View File

@@ -1,43 +0,0 @@
SET enable_seqscan = off;
CREATE TABLE t (val sparsevec(3));
INSERT INTO t (val) VALUES ('{}/3'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
SELECT * FROM t ORDER BY val <-> '{1:3,2:3,3:3}/3';
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)
SELECT COUNT(*) FROM t;
count
-------
5
(1 row)
TRUNCATE t;
SELECT * FROM t ORDER BY val <-> '{1:3,2:3,3:3}/3';
val
-----
(0 rows)
DROP TABLE t;
-- TODO move
CREATE TABLE t (val sparsevec(1001));
INSERT INTO t (val) VALUES (array_fill(1, ARRAY[1001])::vector::sparsevec);
CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
ERROR: sparsevec cannot have more than 1000 non-zero elements for hnsw index
TRUNCATE t;
CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
INSERT INTO t (val) VALUES (array_fill(1, ARRAY[1001])::vector::sparsevec);
ERROR: sparsevec cannot have more than 1000 non-zero elements for hnsw index
DROP TABLE t;

View File

@@ -63,25 +63,9 @@ SELECT '[1.5e-38,-1.5e-38]'::vector;
(1 row)
SELECT '[4e38,1]'::vector;
ERROR: "4e38" is out of range for type vector
ERROR: infinite value not allowed in vector
LINE 1: SELECT '[4e38,1]'::vector;
^
SELECT '[-4e38,1]'::vector;
ERROR: "-4e38" is out of range for type vector
LINE 1: SELECT '[-4e38,1]'::vector;
^
SELECT '[1e-46,1]'::vector;
vector
--------
[0,1]
(1 row)
SELECT '[-1e-46,1]'::vector;
vector
--------
[-0,1]
(1 row)
SELECT '[1,2,3'::vector;
ERROR: malformed vector literal: "[1,2,3"
LINE 1: SELECT '[1,2,3'::vector;
@@ -132,30 +116,8 @@ SELECT '[1, ,3]'::vector;
ERROR: invalid input syntax for type vector: "[1, ,3]"
LINE 1: SELECT '[1, ,3]'::vector;
^
SELECT '[1,2,3]'::vector(3);
vector
---------
[1,2,3]
(1 row)
SELECT '[1,2,3]'::vector(2);
ERROR: expected 2 dimensions, not 3
SELECT '[1,2,3]'::vector(3, 2);
ERROR: invalid type modifier
LINE 1: SELECT '[1,2,3]'::vector(3, 2);
^
SELECT '[1,2,3]'::vector('a');
ERROR: invalid input syntax for type integer: "a"
LINE 1: SELECT '[1,2,3]'::vector('a');
^
SELECT '[1,2,3]'::vector(0);
ERROR: dimensions for type vector must be at least 1
LINE 1: SELECT '[1,2,3]'::vector(0);
^
SELECT '[1,2,3]'::vector(16001);
ERROR: dimensions for type vector cannot exceed 16000
LINE 1: SELECT '[1,2,3]'::vector(16001);
^
SELECT unnest('{"[1,2,3]", "[4,5,6]"}'::vector[]);
unnest
---------

View File

@@ -11,16 +11,9 @@ SELECT * FROM t ORDER BY val <=> '[3,3,3]';
[1,2,4]
(3 rows)
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
count
-------
3
(1 row)
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
count
-------
3
(1 row)
SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector);
val
-----
(0 rows)
DROP TABLE t;

View File

@@ -12,10 +12,9 @@ SELECT * FROM t ORDER BY val <#> '[3,3,3]';
[0,0,0]
(4 rows)
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector)) t2;
count
-------
4
(1 row)
SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector);
val
-----
(0 rows)
DROP TABLE t;

View File

@@ -1,7 +1,7 @@
SET enable_seqscan = off;
CREATE TABLE t (val halfvec(3));
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 halfvec_l2_ops);
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
INSERT INTO t (val) VALUES ('[1,2,4]');
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
val
@@ -12,11 +12,10 @@ SELECT * FROM t ORDER BY val <-> '[3,3,3]';
[0,0,0]
(4 rows)
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::halfvec)) t2;
count
-------
4
(1 row)
SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
val
-----
(0 rows)
SELECT COUNT(*) FROM t;
count
@@ -24,10 +23,4 @@ SELECT COUNT(*) FROM t;
5
(1 row)
TRUNCATE t;
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
val
-----
(0 rows)
DROP TABLE t;

View File

@@ -1,8 +1,9 @@
SET enable_seqscan = off;
CREATE TABLE t (val vector(3));
CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 0);
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 0);
ERROR: value 0 out of bounds for option "lists"
DETAIL: Valid values are between "1" and "32768".
CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 32769);
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 32769);
ERROR: value 32769 out of bounds for option "lists"
DETAIL: Valid values are between "1" and "32768".
SHOW ivfflat.probes;

View File

@@ -1,7 +1,7 @@
SET enable_seqscan = off;
CREATE UNLOGGED 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 ivfflat (val vector_l2_ops) WITH (lists = 1);
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
val
---------

View File

@@ -1,39 +0,0 @@
SET enable_seqscan = off;
CREATE TABLE t (val vector(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 1);
INSERT INTO t (val) VALUES ('[1,2,4]');
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
val
---------
[1,2,3]
[1,2,4]
[1,1,1]
[0,0,0]
(4 rows)
SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
val
---------
[0,0,0]
[1,1,1]
[1,2,3]
[1,2,4]
(4 rows)
SELECT COUNT(*) FROM t;
count
-------
5
(1 row)
TRUNCATE t;
NOTICE: ivfflat index created with little data
DETAIL: This will cause low recall.
HINT: Drop the index until the table has more data.
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
val
-----
(0 rows)
DROP TABLE t;

View File

@@ -1,62 +0,0 @@
SELECT l2_distance('{}/2'::sparsevec, '{1:3,2:4}/2');
l2_distance
-------------
5
(1 row)
SELECT l2_distance('{}/2'::sparsevec, '{2:1}/2');
l2_distance
-------------
1
(1 row)
SELECT '{}/2'::sparsevec <-> '{1:3,2:4}/2';
?column?
----------
5
(1 row)
SELECT inner_product('{1:1,2:2}/2'::sparsevec, '{1:2,2:4}/2');
inner_product
---------------
10
(1 row)
SELECT sparsevec_negative_inner_product('{1:1,2:2}/2', '{1:2,2:4}/2');
sparsevec_negative_inner_product
----------------------------------
-10
(1 row)
SELECT cosine_distance('{1:1,2:2}/2'::sparsevec, '{1:2,2:4}/2');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('{1:1,2:2}/2'::sparsevec, '{}/2');
cosine_distance
-----------------
NaN
(1 row)
SELECT cosine_distance('{1:1,2:1}/2'::sparsevec, '{1:-1,2:-1}/2');
cosine_distance
-----------------
2
(1 row)
SELECT cosine_distance('{1:2}/2'::sparsevec, '{2:2}/2');
cosine_distance
-----------------
1
(1 row)
SELECT cosine_distance('{}/1'::sparsevec, '{}/1');
cosine_distance
-----------------
NaN
(1 row)
SELECT cosine_distance('{1:2}/2'::sparsevec, '{1:1}/3');
ERROR: different sparsevec dimensions 2 and 3

View File

@@ -1,62 +0,0 @@
SELECT '{1:1.5,3:3.5}/5'::sparsevec;
sparsevec
-----------------
{1:1.5,3:3.5}/5
(1 row)
SELECT '{1:1.5,3:3.5}/5'::sparsevec::vector;
vector
-----------------
[1.5,0,3.5,0,0]
(1 row)
SELECT '{1:1.5,3:3.5}/5'::sparsevec::vector(5);
vector
-----------------
[1.5,0,3.5,0,0]
(1 row)
SELECT '{1:1.5,3:3.5}/5'::sparsevec::vector(4);
ERROR: expected 4 dimensions, not 5
SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec;
sparsevec
-----------------
{2:1.5,4:3.5}/5
(1 row)
SELECT '{1:0,2:1,3:0}/3'::sparsevec;
sparsevec
-----------
{2:1}/3
(1 row)
SELECT '{2:1,1:1}/2'::sparsevec;
ERROR: indexes must be in ascending order
LINE 1: SELECT '{2:1,1:1}/2'::sparsevec;
^
SELECT '{}/5'::sparsevec;
sparsevec
-----------
{}/5
(1 row)
SELECT '{}/-1'::sparsevec;
ERROR: sparsevec must have at least 1 dimension
LINE 1: SELECT '{}/-1'::sparsevec;
^
SELECT '{}/100001'::sparsevec;
ERROR: sparsevec cannot have more than 100000 dimensions
LINE 1: SELECT '{}/100001'::sparsevec;
^
SELECT '{}/16001'::sparsevec::vector;
ERROR: vector cannot have more than 16000 dimensions
SELECT '{0:1}/1'::sparsevec;
ERROR: index "0" is out of range for type sparsevec
LINE 1: SELECT '{0:1}/1'::sparsevec;
^
SELECT '{2:1}/1'::sparsevec;
ERROR: index must be less than or equal to dimensions
LINE 1: SELECT '{2:1}/1'::sparsevec;
^
SELECT '{}/1'::sparsevec(2);
ERROR: expected 2 dimensions, not 1

View File

@@ -1,366 +0,0 @@
SELECT '[1,2,3]'::vector + '[4,5,6]';
?column?
----------
[5,7,9]
(1 row)
SELECT '[3e38]'::vector + '[3e38]';
ERROR: value out of range: overflow
SELECT '[1,2,3]'::vector - '[4,5,6]';
?column?
------------
[-3,-3,-3]
(1 row)
SELECT '[-3e38]'::vector - '[3e38]';
ERROR: value out of range: overflow
SELECT '[1,2,3]'::vector * '[4,5,6]';
?column?
-----------
[4,10,18]
(1 row)
SELECT '[1e37]'::vector * '[1e37]';
ERROR: value out of range: overflow
SELECT '[1e-37]'::vector * '[1e-37]';
ERROR: value out of range: underflow
SELECT '[1,2,3]'::vector < '[1,2,3]';
?column?
----------
f
(1 row)
SELECT '[1,2,3]'::vector < '[1,2]';
?column?
----------
f
(1 row)
SELECT '[1,2,3]'::vector <= '[1,2,3]';
?column?
----------
t
(1 row)
SELECT '[1,2,3]'::vector <= '[1,2]';
?column?
----------
f
(1 row)
SELECT '[1,2,3]'::vector = '[1,2,3]';
?column?
----------
t
(1 row)
SELECT '[1,2,3]'::vector = '[1,2]';
?column?
----------
f
(1 row)
SELECT '[1,2,3]'::vector != '[1,2,3]';
?column?
----------
f
(1 row)
SELECT '[1,2,3]'::vector != '[1,2]';
?column?
----------
t
(1 row)
SELECT '[1,2,3]'::vector >= '[1,2,3]';
?column?
----------
t
(1 row)
SELECT '[1,2,3]'::vector >= '[1,2]';
?column?
----------
t
(1 row)
SELECT '[1,2,3]'::vector > '[1,2,3]';
?column?
----------
f
(1 row)
SELECT '[1,2,3]'::vector > '[1,2]';
?column?
----------
t
(1 row)
SELECT vector_cmp('[1,2,3]', '[1,2,3]');
vector_cmp
------------
0
(1 row)
SELECT vector_cmp('[1,2,3]', '[0,0,0]');
vector_cmp
------------
1
(1 row)
SELECT vector_cmp('[0,0,0]', '[1,2,3]');
vector_cmp
------------
-1
(1 row)
SELECT vector_cmp('[1,2]', '[1,2,3]');
vector_cmp
------------
-1
(1 row)
SELECT vector_cmp('[1,2,3]', '[1,2]');
vector_cmp
------------
1
(1 row)
SELECT vector_cmp('[1,2]', '[2,3,4]');
vector_cmp
------------
-1
(1 row)
SELECT vector_cmp('[2,3]', '[1,2,3]');
vector_cmp
------------
1
(1 row)
SELECT vector_dims('[1,2,3]');
vector_dims
-------------
3
(1 row)
SELECT round(vector_norm('[1,1]')::numeric, 5);
round
---------
1.41421
(1 row)
SELECT vector_norm('[3,4]');
vector_norm
-------------
5
(1 row)
SELECT vector_norm('[0,1]');
vector_norm
-------------
1
(1 row)
SELECT vector_norm('[3e37,4e37]')::real;
vector_norm
-------------
5e+37
(1 row)
SELECT l2_distance('[0,0]'::vector, '[3,4]');
l2_distance
-------------
5
(1 row)
SELECT l2_distance('[0,0]'::vector, '[0,1]');
l2_distance
-------------
1
(1 row)
SELECT l2_distance('[1,2]'::vector, '[3]');
ERROR: different vector dimensions 2 and 1
SELECT l2_distance('[3e38]'::vector, '[-3e38]');
l2_distance
-------------
Infinity
(1 row)
SELECT inner_product('[1,2]'::vector, '[3,4]');
inner_product
---------------
11
(1 row)
SELECT inner_product('[1,2]'::vector, '[3]');
ERROR: different vector dimensions 2 and 1
SELECT inner_product('[3e38]'::vector, '[3e38]');
inner_product
---------------
Infinity
(1 row)
SELECT cosine_distance('[1,2]'::vector, '[2,4]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,2]'::vector, '[0,0]');
cosine_distance
-----------------
NaN
(1 row)
SELECT cosine_distance('[1,1]'::vector, '[1,1]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,0]'::vector, '[0,2]');
cosine_distance
-----------------
1
(1 row)
SELECT cosine_distance('[1,1]'::vector, '[-1,-1]');
cosine_distance
-----------------
2
(1 row)
SELECT cosine_distance('[1,2]'::vector, '[3]');
ERROR: different vector dimensions 2 and 1
SELECT cosine_distance('[1,1]'::vector, '[1.1,1.1]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,1]'::vector, '[-1.1,-1.1]');
cosine_distance
-----------------
2
(1 row)
SELECT cosine_distance('[3e38]'::vector, '[3e38]');
cosine_distance
-----------------
NaN
(1 row)
SELECT l1_distance('[0,0]'::vector, '[3,4]');
l1_distance
-------------
7
(1 row)
SELECT l1_distance('[0,0]'::vector, '[0,1]');
l1_distance
-------------
1
(1 row)
SELECT l1_distance('[1,2]'::vector, '[3]');
ERROR: different vector dimensions 2 and 1
SELECT l1_distance('[3e38]'::vector, '[-3e38]');
l1_distance
-------------
Infinity
(1 row)
SELECT quantize_binary('[1,0,-1]'::vector);
quantize_binary
-----------------
100
(1 row)
SELECT quantize_binary('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]'::vector);
quantize_binary
-----------------
01001110101
(1 row)
SELECT subvector('[1,2,3,4,5]'::vector, 1, 3);
subvector
-----------
[1,2,3]
(1 row)
SELECT subvector('[1,2,3,4,5]'::vector, 3, 2);
subvector
-----------
[3,4]
(1 row)
SELECT subvector('[1,2,3,4,5]'::vector, -1, 3);
subvector
-----------
[1]
(1 row)
SELECT subvector('[1,2,3,4,5]'::vector, 3, 9);
subvector
-----------
[3,4,5]
(1 row)
SELECT subvector('[1,2,3,4,5]'::vector, 1, 0);
ERROR: vector must have at least 1 dimension
SELECT subvector('[1,2,3,4,5]'::vector, 3, -1);
ERROR: vector must have at least 1 dimension
SELECT subvector('[1,2,3,4,5]'::vector, -1, 2);
ERROR: vector must have at least 1 dimension
SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
avg
-----------
[2,3.5,5]
(1 row)
SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;
avg
-----------
[2,3.5,5]
(1 row)
SELECT avg(v) FROM unnest(ARRAY[]::vector[]) v;
avg
-----
(1 row)
SELECT avg(v) FROM unnest(ARRAY['[1,2]'::vector, '[3]']) v;
ERROR: expected 2 dimensions, not 1
SELECT avg(v) FROM unnest(ARRAY['[3e38]'::vector, '[3e38]']) v;
avg
---------
[3e+38]
(1 row)
SELECT vector_avg(array_agg(n)) FROM generate_series(1, 16002) n;
ERROR: vector cannot have more than 16000 dimensions
SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
sum
----------
[4,7,10]
(1 row)
SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;
sum
----------
[4,7,10]
(1 row)
SELECT sum(v) FROM unnest(ARRAY[]::vector[]) v;
sum
-----
(1 row)
SELECT sum(v) FROM unnest(ARRAY['[1,2]'::vector, '[3]']) v;
ERROR: different vector dimensions 2 and 1
SELECT sum(v) FROM unnest(ARRAY['[3e38]'::vector, '[3e38]']) v;
ERROR: value out of range: overflow

View File

@@ -1,21 +0,0 @@
SELECT hamming_distance('111', '111');
SELECT hamming_distance('111', '110');
SELECT hamming_distance('111', '100');
SELECT hamming_distance('111', '000');
SELECT hamming_distance('10101010101010101010', '01010101010101010101');
SELECT hamming_distance('', '');
SELECT hamming_distance('111', '00');
SELECT hamming_distance('111', '000'::varbit(4));
SELECT hamming_distance('111', '0000'::varbit(4));
SELECT jaccard_distance('1111', '1111');
SELECT jaccard_distance('1111', '1110');
SELECT jaccard_distance('1111', '1100');
SELECT jaccard_distance('1111', '1000');
SELECT jaccard_distance('1111', '0000');
SELECT jaccard_distance('1100', '1000');
SELECT jaccard_distance('10101010101010101010', '01010101010101010101');
SELECT jaccard_distance('', '');
SELECT jaccard_distance('1111', '000');
SELECT jaccard_distance('1111', '0000'::varbit(5));
SELECT jaccard_distance('1111', '00000'::varbit(5));

View File

@@ -10,12 +10,6 @@ SELECT '{-Infinity}'::real[]::vector;
SELECT '{}'::real[]::vector;
SELECT '{{1}}'::real[]::vector;
SELECT '[1,2,3]'::vector::real[];
SELECT '[1,2,3]'::vector::halfvec;
SELECT '[1,2,3]'::halfvec::vector;
SELECT '[1,2,3]'::vector::halfvec(2);
SELECT '[1,2,3]'::halfvec::vector(2);
SELECT '[65520]'::vector::halfvec;
SELECT '[1e-8]'::vector::halfvec;
SELECT 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;

View File

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

43
test/sql/functions.sql Normal file
View File

@@ -0,0 +1,43 @@
SELECT '[1,2,3]'::vector + '[4,5,6]';
SELECT '[3e38]'::vector + '[3e38]';
SELECT '[1,2,3]'::vector - '[4,5,6]';
SELECT '[-3e38]'::vector - '[3e38]';
SELECT '[1,2,3]'::vector * '[4,5,6]';
SELECT '[1e37]'::vector * '[1e37]';
SELECT '[1e-37]'::vector * '[1e-37]';
SELECT vector_dims('[1,2,3]');
SELECT round(vector_norm('[1,1]')::numeric, 5);
SELECT vector_norm('[3,4]');
SELECT vector_norm('[0,1]');
SELECT l2_distance('[0,0]', '[3,4]');
SELECT l2_distance('[0,0]', '[0,1]');
SELECT l2_distance('[1,2]', '[3]');
SELECT inner_product('[1,2]', '[3,4]');
SELECT inner_product('[1,2]', '[3]');
SELECT cosine_distance('[1,2]', '[2,4]');
SELECT cosine_distance('[1,2]', '[0,0]');
SELECT cosine_distance('[1,1]', '[1,1]');
SELECT cosine_distance('[1,1]', '[-1,-1]');
SELECT cosine_distance('[1,2]', '[3]');
SELECT l1_distance('[0,0]', '[3,4]');
SELECT l1_distance('[0,0]', '[0,1]');
SELECT l1_distance('[1,2]', '[3]');
SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;
SELECT avg(v) FROM unnest(ARRAY[]::vector[]) v;
SELECT avg(v) FROM unnest(ARRAY['[1,2]'::vector, '[3]']) v;
SELECT avg(v) FROM unnest(ARRAY['[3e38]'::vector, '[3e38]']) v;
SELECT vector_avg(array_agg(n)) FROM generate_series(1, 16002) n;
SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;
SELECT sum(v) FROM unnest(ARRAY[]::vector[]) v;
SELECT sum(v) FROM unnest(ARRAY['[1,2]'::vector, '[3]']) v;
SELECT sum(v) FROM unnest(ARRAY['[3e38]'::vector, '[3e38]']) v;

View File

@@ -1,34 +0,0 @@
SELECT l2_distance('[0,0]'::halfvec, '[3,4]');
SELECT l2_distance('[0,0]'::halfvec, '[0,1]');
SELECT l2_distance('[1,2]'::halfvec, '[3]');
SELECT '[0,0]'::halfvec <-> '[3,4]';
SELECT inner_product('[1,2]'::halfvec, '[3,4]');
SELECT inner_product('[1,2]'::halfvec, '[3]');
SELECT inner_product('[65504]'::halfvec, '[65504]');
SELECT '[1,2]'::halfvec <#> '[3,4]';
SELECT cosine_distance('[1,2]'::halfvec, '[2,4]');
SELECT cosine_distance('[1,2]'::halfvec, '[0,0]');
SELECT cosine_distance('[1,1]'::halfvec, '[1,1]');
SELECT cosine_distance('[1,0]'::halfvec, '[0,2]');
SELECT cosine_distance('[1,1]'::halfvec, '[-1,-1]');
SELECT cosine_distance('[1,2]'::halfvec, '[3]');
SELECT cosine_distance('[1,1]'::halfvec, '[1.1,1.1]');
SELECT cosine_distance('[1,1]'::halfvec, '[-1.1,-1.1]');
SELECT '[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 quantize_binary('[1,0,-1]'::halfvec);
SELECT quantize_binary('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]'::halfvec);
SELECT subvector('[1,2,3,4,5]'::halfvec, 1, 3);
SELECT subvector('[1,2,3,4,5]'::halfvec, 3, 2);
SELECT subvector('[1,2,3,4,5]'::halfvec, -1, 3);
SELECT subvector('[1,2,3,4,5]'::halfvec, 3, 9);
SELECT subvector('[1,2,3,4,5]'::halfvec, 1, 0);
SELECT subvector('[1,2,3,4,5]'::halfvec, 3, -1);
SELECT subvector('[1,2,3,4,5]'::halfvec, -1, 2);

View File

@@ -1,35 +0,0 @@
SELECT '[1,2,3]'::halfvec;
SELECT '[-1,-2,-3]'::halfvec;
SELECT '[1.,2.,3.]'::halfvec;
SELECT ' [ 1, 2 , 3 ] '::halfvec;
SELECT '[1.23456]'::halfvec;
SELECT '[hello,1]'::halfvec;
SELECT '[NaN,1]'::halfvec;
SELECT '[Infinity,1]'::halfvec;
SELECT '[-Infinity,1]'::halfvec;
SELECT '[65519,-65519]'::halfvec;
SELECT '[65520,-65520]'::halfvec;
SELECT '[1e-8,-1e-8]'::halfvec;
SELECT '[4e38,1]'::halfvec;
SELECT '[1e-46,1]'::halfvec;
SELECT '[1,2,3'::halfvec;
SELECT '[1,2,3]9'::halfvec;
SELECT '1,2,3'::halfvec;
SELECT ''::halfvec;
SELECT '['::halfvec;
SELECT '[,'::halfvec;
SELECT '[]'::halfvec;
SELECT '[1,]'::halfvec;
SELECT '[1a]'::halfvec;
SELECT '[1,,3]'::halfvec;
SELECT '[1, ,3]'::halfvec;
SELECT '[1,2,3]'::halfvec(3);
SELECT '[1,2,3]'::halfvec(2);
SELECT '[1,2,3]'::halfvec(3, 2);
SELECT '[1,2,3]'::halfvec('a');
SELECT '[1,2,3]'::halfvec(0);
SELECT '[1,2,3]'::halfvec(16001);
SELECT unnest('{"[1,2,3]", "[4,5,6]"}'::halfvec[]);
SELECT '{"[1,2,3]"}'::halfvec(2)[];

View File

@@ -1,18 +0,0 @@
SET enable_seqscan = off;
CREATE TABLE t (val bit(3));
INSERT INTO t (val) VALUES (B'000'), (B'100'), (B'111'), (NULL);
CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
INSERT INTO t (val) VALUES (B'110');
SELECT * FROM t ORDER BY val <~> B'111';
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <~> (SELECT NULL::bit)) t2;
DROP TABLE t;
-- TODO move
CREATE TABLE t (val varbit(3));
CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
CREATE INDEX ON t USING hnsw ((val::bit(3)) bit_hamming_ops);
DROP TABLE t;

View File

@@ -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 hnsw (val bit_jaccard_ops);
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;

View File

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

View File

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

View File

@@ -1,16 +0,0 @@
SET enable_seqscan = off;
CREATE TABLE t (val halfvec(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE INDEX ON t USING hnsw (val halfvec_l2_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;
SELECT COUNT(*) FROM t;
TRUNCATE t;
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
DROP TABLE t;

View File

@@ -2,7 +2,7 @@ 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 ivfflat (val vector_l2_ops) WITH (lists = 1);
CREATE INDEX ON t USING hnsw (val vector_l2_ops);
INSERT INTO t (val) VALUES ('[1,2,4]');
@@ -10,7 +10,4 @@ SELECT * FROM t ORDER BY val <-> '[3,3,3]';
SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
SELECT COUNT(*) FROM t;
TRUNCATE t;
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
DROP TABLE t;

View File

@@ -1,13 +1,14 @@
SET enable_seqscan = off;
CREATE TABLE t (val vector(3));
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 1);
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 3);
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 101);
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 3);
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 9);
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (ef_construction = 1001);
CREATE INDEX ON t USING hnsw (val vector_l2_ops) WITH (m = 16, ef_construction = 31);
SHOW hnsw.ef_search;
SET hnsw.ef_search = 0;
SET hnsw.ef_search = 9;
SET hnsw.ef_search = 1001;
DROP TABLE t;

View File

@@ -1,13 +0,0 @@
SET enable_seqscan = off;
CREATE TABLE t (val sparsevec(3));
INSERT INTO t (val) VALUES ('{}/3'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
CREATE INDEX ON t USING hnsw (val sparsevec_cosine_ops);
INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
SELECT * FROM t ORDER BY val <=> '{1:3,2:3,3:3}/3';
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '{}/3') t2;
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::sparsevec)) t2;
DROP TABLE t;

View File

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

View File

@@ -1,25 +0,0 @@
SET enable_seqscan = off;
CREATE TABLE t (val sparsevec(3));
INSERT INTO t (val) VALUES ('{}/3'), ('{1:1,2:2,3:3}/3'), ('{1:1,2:1,3:1}/3'), (NULL);
CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
INSERT INTO t (val) VALUES ('{1:1,2:2,3:4}/3');
SELECT * FROM t ORDER BY val <-> '{1:3,2:3,3:3}/3';
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::sparsevec)) t2;
SELECT COUNT(*) FROM t;
TRUNCATE t;
SELECT * FROM t ORDER BY val <-> '{1:3,2:3,3:3}/3';
DROP TABLE t;
-- TODO move
CREATE TABLE t (val sparsevec(1001));
INSERT INTO t (val) VALUES (array_fill(1, ARRAY[1001])::vector::sparsevec);
CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
TRUNCATE t;
CREATE INDEX ON t USING hnsw (val sparsevec_l2_ops);
INSERT INTO t (val) VALUES (array_fill(1, ARRAY[1001])::vector::sparsevec);
DROP TABLE t;

View File

@@ -11,9 +11,6 @@ SELECT '[1.5e38,-1.5e38]'::vector;
SELECT '[1.5e+38,-1.5e+38]'::vector;
SELECT '[1.5e-38,-1.5e-38]'::vector;
SELECT '[4e38,1]'::vector;
SELECT '[-4e38,1]'::vector;
SELECT '[1e-46,1]'::vector;
SELECT '[-1e-46,1]'::vector;
SELECT '[1,2,3'::vector;
SELECT '[1,2,3]9'::vector;
SELECT '1,2,3'::vector;
@@ -25,13 +22,7 @@ SELECT '[1,]'::vector;
SELECT '[1a]'::vector;
SELECT '[1,,3]'::vector;
SELECT '[1, ,3]'::vector;
SELECT '[1,2,3]'::vector(3);
SELECT '[1,2,3]'::vector(2);
SELECT '[1,2,3]'::vector(3, 2);
SELECT '[1,2,3]'::vector('a');
SELECT '[1,2,3]'::vector(0);
SELECT '[1,2,3]'::vector(16001);
SELECT unnest('{"[1,2,3]", "[4,5,6]"}'::vector[]);
SELECT '{"[1,2,3]"}'::vector(2)[];

View File

@@ -7,7 +7,6 @@ CREATE INDEX ON t USING ivfflat (val vector_cosine_ops) WITH (lists = 1);
INSERT INTO t (val) VALUES ('[1,2,4]');
SELECT * FROM t ORDER BY val <=> '[3,3,3]';
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector)) t2;
SELECT * FROM t ORDER BY val <=> (SELECT NULL::vector);
DROP TABLE t;

View File

@@ -7,6 +7,6 @@ CREATE INDEX ON t USING ivfflat (val vector_ip_ops) WITH (lists = 1);
INSERT INTO t (val) VALUES ('[1,2,4]');
SELECT * FROM t ORDER BY val <#> '[3,3,3]';
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector)) t2;
SELECT * FROM t ORDER BY val <#> (SELECT NULL::vector);
DROP TABLE t;

View File

@@ -2,15 +2,12 @@ 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_l2_ops);
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
INSERT INTO t (val) VALUES ('[1,2,4]');
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector)) t2;
SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
SELECT COUNT(*) FROM t;
TRUNCATE t;
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
DROP TABLE t;

View File

@@ -1,6 +1,8 @@
SET enable_seqscan = off;
CREATE TABLE t (val vector(3));
CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 0);
CREATE INDEX ON t USING ivfflat (val vector_l2_ops) WITH (lists = 32769);
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 0);
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 32769);
SHOW ivfflat.probes;

View File

@@ -2,7 +2,7 @@ SET enable_seqscan = off;
CREATE UNLOGGED 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 ivfflat (val vector_l2_ops) WITH (lists = 1);
CREATE INDEX ON t USING ivfflat (val) WITH (lists = 1);
SELECT * FROM t ORDER BY val <-> '[3,3,3]';

View File

@@ -1,13 +0,0 @@
SELECT l2_distance('{}/2'::sparsevec, '{1:3,2:4}/2');
SELECT l2_distance('{}/2'::sparsevec, '{2:1}/2');
SELECT '{}/2'::sparsevec <-> '{1:3,2:4}/2';
SELECT inner_product('{1:1,2:2}/2'::sparsevec, '{1:2,2:4}/2');
SELECT sparsevec_negative_inner_product('{1:1,2:2}/2', '{1:2,2:4}/2');
SELECT cosine_distance('{1:1,2:2}/2'::sparsevec, '{1:2,2:4}/2');
SELECT cosine_distance('{1:1,2:2}/2'::sparsevec, '{}/2');
SELECT cosine_distance('{1:1,2:1}/2'::sparsevec, '{1:-1,2:-1}/2');
SELECT cosine_distance('{1:2}/2'::sparsevec, '{2:2}/2');
SELECT cosine_distance('{}/1'::sparsevec, '{}/1');
SELECT cosine_distance('{1:2}/2'::sparsevec, '{1:1}/3');

View File

@@ -1,19 +0,0 @@
SELECT '{1:1.5,3:3.5}/5'::sparsevec;
SELECT '{1:1.5,3:3.5}/5'::sparsevec::vector;
SELECT '{1:1.5,3:3.5}/5'::sparsevec::vector(5);
SELECT '{1:1.5,3:3.5}/5'::sparsevec::vector(4);
SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec;
SELECT '{1:0,2:1,3:0}/3'::sparsevec;
SELECT '{2:1,1:1}/2'::sparsevec;
SELECT '{}/5'::sparsevec;
SELECT '{}/-1'::sparsevec;
SELECT '{}/100001'::sparsevec;
SELECT '{}/16001'::sparsevec::vector;
SELECT '{0:1}/1'::sparsevec;
SELECT '{2:1}/1'::sparsevec;
SELECT '{}/1'::sparsevec(2);

View File

@@ -1,83 +0,0 @@
SELECT '[1,2,3]'::vector + '[4,5,6]';
SELECT '[3e38]'::vector + '[3e38]';
SELECT '[1,2,3]'::vector - '[4,5,6]';
SELECT '[-3e38]'::vector - '[3e38]';
SELECT '[1,2,3]'::vector * '[4,5,6]';
SELECT '[1e37]'::vector * '[1e37]';
SELECT '[1e-37]'::vector * '[1e-37]';
SELECT '[1,2,3]'::vector < '[1,2,3]';
SELECT '[1,2,3]'::vector < '[1,2]';
SELECT '[1,2,3]'::vector <= '[1,2,3]';
SELECT '[1,2,3]'::vector <= '[1,2]';
SELECT '[1,2,3]'::vector = '[1,2,3]';
SELECT '[1,2,3]'::vector = '[1,2]';
SELECT '[1,2,3]'::vector != '[1,2,3]';
SELECT '[1,2,3]'::vector != '[1,2]';
SELECT '[1,2,3]'::vector >= '[1,2,3]';
SELECT '[1,2,3]'::vector >= '[1,2]';
SELECT '[1,2,3]'::vector > '[1,2,3]';
SELECT '[1,2,3]'::vector > '[1,2]';
SELECT vector_cmp('[1,2,3]', '[1,2,3]');
SELECT vector_cmp('[1,2,3]', '[0,0,0]');
SELECT vector_cmp('[0,0,0]', '[1,2,3]');
SELECT vector_cmp('[1,2]', '[1,2,3]');
SELECT vector_cmp('[1,2,3]', '[1,2]');
SELECT vector_cmp('[1,2]', '[2,3,4]');
SELECT vector_cmp('[2,3]', '[1,2,3]');
SELECT vector_dims('[1,2,3]');
SELECT round(vector_norm('[1,1]')::numeric, 5);
SELECT vector_norm('[3,4]');
SELECT vector_norm('[0,1]');
SELECT vector_norm('[3e37,4e37]')::real;
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 inner_product('[1,2]'::vector, '[3,4]');
SELECT inner_product('[1,2]'::vector, '[3]');
SELECT inner_product('[3e38]'::vector, '[3e38]');
SELECT cosine_distance('[1,2]'::vector, '[2,4]');
SELECT cosine_distance('[1,2]'::vector, '[0,0]');
SELECT cosine_distance('[1,1]'::vector, '[1,1]');
SELECT cosine_distance('[1,0]'::vector, '[0,2]');
SELECT cosine_distance('[1,1]'::vector, '[-1,-1]');
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 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 quantize_binary('[1,0,-1]'::vector);
SELECT quantize_binary('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]'::vector);
SELECT subvector('[1,2,3,4,5]'::vector, 1, 3);
SELECT subvector('[1,2,3,4,5]'::vector, 3, 2);
SELECT subvector('[1,2,3,4,5]'::vector, -1, 3);
SELECT subvector('[1,2,3,4,5]'::vector, 3, 9);
SELECT subvector('[1,2,3,4,5]'::vector, 1, 0);
SELECT subvector('[1,2,3,4,5]'::vector, 3, -1);
SELECT subvector('[1,2,3,4,5]'::vector, -1, 2);
SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;
SELECT avg(v) FROM unnest(ARRAY[]::vector[]) v;
SELECT avg(v) FROM unnest(ARRAY['[1,2]'::vector, '[3]']) v;
SELECT avg(v) FROM unnest(ARRAY['[3e38]'::vector, '[3e38]']) v;
SELECT vector_avg(array_agg(n)) FROM generate_series(1, 16002) n;
SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
SELECT sum(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;
SELECT sum(v) FROM unnest(ARRAY[]::vector[]) v;
SELECT sum(v) FROM unnest(ARRAY['[1,2]'::vector, '[3]']) v;
SELECT sum(v) FROM unnest(ARRAY['[3e38]'::vector, '[3e38]']) v;

View File

@@ -19,13 +19,14 @@ sub test_index_replay
# Wait for replica to catch up
my $applname = $node_replica->name;
my $server_version_num = $node_primary->safe_psql("postgres", "SHOW server_version_num");
my $caughtup_query = "SELECT pg_current_wal_lsn() <= replay_lsn FROM pg_stat_replication WHERE application_name = '$applname';";
$node_primary->poll_query_until('postgres', $caughtup_query)
or die "Timed out while waiting for replica 1 to catch up";
my @r = ();
for (1 .. $dim)
{
for (1 .. $dim) {
push(@r, rand());
}
my $sql = join(",", @r);
@@ -51,13 +52,11 @@ my $array_sql = join(",", ('random()') x $dim);
# Initialize primary node
$node_primary = get_new_node('primary');
$node_primary->init(allows_streaming => 1);
if ($dim > 32)
{
if ($dim > 32) {
# TODO use wal_keep_segments for Postgres < 13
$node_primary->append_conf('postgresql.conf', qq(wal_keep_size = 1GB));
}
if ($dim > 1500)
{
if ($dim > 1500) {
$node_primary->append_conf('postgresql.conf', qq(maintenance_work_mem = 128MB));
}
$node_primary->start;
@@ -68,7 +67,8 @@ $node_primary->backup($backup_name);
# Create streaming replica linking to primary
$node_replica = get_new_node('replica');
$node_replica->init_from_backup($node_primary, $backup_name, has_streaming => 1);
$node_replica->init_from_backup($node_primary, $backup_name,
has_streaming => 1);
$node_replica->start;
# Create ivfflat index on primary
@@ -77,7 +77,7 @@ $node_primary->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));
$node_primary->safe_psql("postgres",
"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
);
$node_primary->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
$node_primary->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v);");
# Test that queries give same result
test_index_replay('initial');

View File

@@ -7,8 +7,7 @@ use Test::More;
my $dim = 3;
my @r = ();
for (1 .. $dim)
{
for (1 .. $dim) {
my $v = int(rand(1000)) + 1;
push(@r, "i % $v");
}
@@ -25,7 +24,7 @@ $node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim));");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT i % 10, ARRAY[$array_sql] FROM generate_series(1, 100000) i;"
);
$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v);");
# Get size
my $size = $node->safe_psql("postgres", "SELECT pg_total_relation_size('tst_v_idx');");

View File

@@ -22,8 +22,7 @@ sub test_recall
));
like($explain, qr/Index Scan using idx on tst/);
for my $i (0 .. $#queries)
{
for my $i (0 .. $#queries) {
my $actual = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SET ivfflat.probes = $probes;
@@ -34,10 +33,8 @@ sub test_recall
my @expected_ids = split("\n", $expected[$i]);
foreach (@expected_ids)
{
if (exists($actual_set{$_}))
{
foreach (@expected_ids) {
if (exists($actual_set{$_})) {
$correct++;
}
$total++;
@@ -60,8 +57,7 @@ $node->safe_psql("postgres",
);
# Generate queries
for (1 .. 20)
{
for (1..20) {
my $r1 = rand();
my $r2 = rand();
my $r3 = rand();
@@ -70,21 +66,26 @@ for (1 .. 20)
# Check each index type
my @operators = ("<->", "<#>", "<=>");
my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
for my $i (0 .. $#operators)
{
my $operator = $operators[$i];
my $opclass = $opclasses[$i];
foreach (@operators) {
my $operator = $_;
# Get exact results
@expected = ();
foreach (@queries)
{
foreach (@queries) {
my $res = $node->safe_psql("postgres", "SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;");
push(@expected, $res);
}
my $opclass;
if ($operator eq "<->") {
$opclass = "vector_l2_ops";
} elsif ($operator eq "<#>") {
$opclass = "vector_ip_ops";
} else {
$opclass = "vector_cosine_ops";
}
# Build index serially
$node->safe_psql("postgres", qq(
SET max_parallel_maintenance_workers = 0;
@@ -92,14 +93,13 @@ for my $i (0 .. $#operators)
));
# Test approximate results
if ($operator ne "<#>")
{
# TODO Fix test (uniform random vectors all have similar inner product)
test_recall(1, 0.71, $operator);
if ($operator ne "<#>") {
# TODO fix test
test_recall(1, 0.75, $operator);
test_recall(10, 0.95, $operator);
}
# Account for equal distances
test_recall(100, 0.9925, $operator);
test_recall(100, 0.995, $operator);
$node->safe_psql("postgres", "DROP INDEX idx;");
@@ -113,14 +113,13 @@ for my $i (0 .. $#operators)
like($stderr, qr/using \d+ parallel workers/);
# Test approximate results
if ($operator ne "<#>")
{
# TODO Fix test (uniform random vectors all have similar inner product)
test_recall(1, 0.71, $operator);
if ($operator ne "<#>") {
# TODO fix test
test_recall(1, 0.75, $operator);
test_recall(10, 0.95, $operator);
}
# Account for equal distances
test_recall(100, 0.9925, $operator);
test_recall(100, 0.995, $operator);
$node->safe_psql("postgres", "DROP INDEX idx;");
}

View File

@@ -20,7 +20,7 @@ sub test_centers
{
my ($lists, $min) = @_;
my ($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops) WITH (lists = $lists);");
my ($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING ivfflat (v) WITH (lists = $lists);");
is($ret, 0, $stderr);
}

View File

@@ -18,21 +18,24 @@ $node->safe_psql("postgres",
# Check each index type
my @operators = ("<->", "<#>", "<=>");
my @opclasses = ("vector_l2_ops", "vector_ip_ops", "vector_cosine_ops");
for my $i (0 .. $#operators)
{
my $operator = $operators[$i];
my $opclass = $opclasses[$i];
foreach (@operators) {
my $operator = $_;
# Add index
my $opclass;
if ($operator eq "<->") {
$opclass = "vector_l2_ops";
} elsif ($operator eq "<#>") {
$opclass = "vector_ip_ops";
} else {
$opclass = "vector_cosine_ops";
}
$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v $opclass);");
# Test 100% recall
for (1 .. 20)
{
my $id = int(rand() * 100000);
my $query = $node->safe_psql("postgres", "SELECT v FROM tst WHERE i = $id;");
for (1..20) {
my $i = int(rand() * 100000);
my $query = $node->safe_psql("postgres", "SELECT v FROM tst WHERE i = $i;");
my $res = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SELECT v FROM tst ORDER BY v <-> '$query' LIMIT 1;

View File

@@ -16,8 +16,8 @@ $node->safe_psql("postgres",
"INSERT INTO tst SELECT ARRAY[random(), random(), random()] FROM generate_series(1, 100000) i;"
);
$node->safe_psql("postgres", "CREATE INDEX lists50 ON tst USING ivfflat (v vector_l2_ops) WITH (lists = 50);");
$node->safe_psql("postgres", "CREATE INDEX lists100 ON tst USING ivfflat (v vector_l2_ops) WITH (lists = 100);");
$node->safe_psql("postgres", "CREATE INDEX lists50 ON tst USING ivfflat (v) WITH (lists = 50);");
$node->safe_psql("postgres", "CREATE INDEX lists100 ON tst USING ivfflat (v) WITH (lists = 100);");
# Test prefers more lists
my $res = $node->safe_psql("postgres", "EXPLAIN SELECT v FROM tst ORDER BY v <-> '[0.5,0.5,0.5]' LIMIT 10;");
@@ -26,7 +26,7 @@ unlike($res, qr/lists50/);
# Test errors with too much memory
my ($ret, $stdout, $stderr) = $node->psql("postgres",
"CREATE INDEX lists10000 ON tst USING ivfflat (v vector_l2_ops) WITH (lists = 10000);"
"CREATE INDEX lists10000 ON tst USING ivfflat (v) WITH (lists = 10000);"
);
like($stderr, qr/memory required is/);

View File

@@ -19,7 +19,7 @@ $node->safe_psql("postgres", "CREATE TABLE tst (v vector($dim));");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
);
$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v vector_l2_ops);");
$node->safe_psql("postgres", "CREATE INDEX ON tst USING ivfflat (v);");
$node->pgbench(
"--no-vacuum --client=5 --transactions=100",
@@ -28,7 +28,7 @@ $node->pgbench(
[qr{^$}],
"concurrent INSERTs",
{
"007_ivfflat_inserts" => "INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10) i;"
"007_inserts" => "INSERT INTO tst SELECT ARRAY[$array_sql] FROM generate_series(1, 10) i;"
}
);

View File

@@ -30,8 +30,7 @@ sub test_aggregate
# Test matches real for avg
# Cannot test sum since sum(real) varies between calls
if ($agg eq 'avg')
{
if ($agg eq 'avg') {
my $r1 = $node->safe_psql("postgres", "SELECT $agg(r1)::float4 FROM tst;");
my $r2 = $node->safe_psql("postgres", "SELECT $agg(r2)::float4 FROM tst;");
my $r3 = $node->safe_psql("postgres", "SELECT $agg(r3)::float4 FROM tst;");

View File

@@ -19,13 +19,14 @@ sub test_index_replay
# Wait for replica to catch up
my $applname = $node_replica->name;
my $server_version_num = $node_primary->safe_psql("postgres", "SHOW server_version_num");
my $caughtup_query = "SELECT pg_current_wal_lsn() <= replay_lsn FROM pg_stat_replication WHERE application_name = '$applname';";
$node_primary->poll_query_until('postgres', $caughtup_query)
or die "Timed out while waiting for replica 1 to catch up";
my @r = ();
for (1 .. $dim)
{
for (1 .. $dim) {
push(@r, rand());
}
my $sql = join(",", @r);
@@ -51,13 +52,11 @@ my $array_sql = join(",", ('random()') x $dim);
# Initialize primary node
$node_primary = get_new_node('primary');
$node_primary->init(allows_streaming => 1);
if ($dim > 32)
{
if ($dim > 32) {
# TODO use wal_keep_segments for Postgres < 13
$node_primary->append_conf('postgresql.conf', qq(wal_keep_size = 1GB));
}
if ($dim > 1500)
{
if ($dim > 1500) {
$node_primary->append_conf('postgresql.conf', qq(maintenance_work_mem = 128MB));
}
$node_primary->start;
@@ -68,7 +67,8 @@ $node_primary->backup($backup_name);
# Create streaming replica linking to primary
$node_replica = get_new_node('replica');
$node_replica->init_from_backup($node_primary, $backup_name, has_streaming => 1);
$node_replica->init_from_backup($node_primary, $backup_name,
has_streaming => 1);
$node_replica->start;
# Create hnsw index on primary

Some files were not shown because too many files have changed in this diff Show More