Compare commits

..

77 Commits

Author SHA1 Message Date
Andrew Kane
3c09df83f6 Updated readme [skip ci] 2024-03-31 22:52:51 -07:00
Andrew Kane
17e0058927 Improved recall test 2024-03-31 22:42:34 -07:00
Andrew Kane
3fc734fb36 Improved recall test [skip ci] 2024-03-31 22:38:17 -07:00
Andrew Kane
caca8990fc Use union to swap [skip ci] 2024-03-31 20:27:04 -07:00
Andrew Kane
b640d84381 Added todo [skip ci] 2024-03-31 20:13:15 -07:00
Andrew Kane
3efa99fe77 Added test for overflow [skip ci] 2024-03-31 20:11:28 -07:00
Andrew Kane
7087a45ea3 Improved halfvec_out code [skip ci] 2024-03-31 10:08:26 -07:00
Andrew Kane
5a229ba984 Check typmod when casting [skip ci] 2024-03-30 23:32:07 -07:00
Andrew Kane
56622e73fb Fixed underflow error 2024-03-30 15:04:33 -07:00
Andrew Kane
968173690e Changed casting to implicit 2024-03-30 10:28:38 -07:00
Andrew Kane
d82f6634fe Added dimension checks [skip ci] 2024-03-30 10:23:56 -07:00
Andrew Kane
4d041d9282 Fixed CI 2024-03-30 10:19:48 -07:00
Andrew Kane
e68db812aa Fixed underflow error 2024-03-30 10:06:21 -07:00
Andrew Kane
49bc34caf3 Make max dimensions consistent across types [skip ci] 2024-03-28 16:41:13 -07:00
Andrew Kane
4f1a379638 Added casting between vector and halfvec 2024-03-28 14:04:59 -07:00
Andrew Kane
45ef8f8a45 Improved code [skip ci] 2024-03-27 17:28:15 -07:00
Andrew Kane
6eedf78c25 Improved HnswNormValue [skip ci] 2024-03-27 17:19:58 -07:00
Andrew Kane
e2978d1365 Updated test [skip ci] 2024-03-27 17:09:38 -07:00
Andrew Kane
ccf44136dd Updated readme [skip ci] 2024-03-27 14:03:59 -07:00
Andrew Kane
daf27d5bd4 Fixed upgrade file [skip ci] 2024-03-27 13:51:12 -07:00
Andrew Kane
868498abec Fixed CI 2024-03-27 13:49:37 -07:00
Andrew Kane
264494b3e2 Added halfvec type 2024-03-27 13:47:19 -07:00
Andrew Kane
08abb63cbe Added notes about NULL vectors [skip ci] 2024-03-27 11:50:37 -07:00
Andrew Kane
06b8556a49 Revert "Updated readme [skip ci]"
This reverts commit 3f674c9994.
2024-03-25 23:33:46 -07:00
Andrew Kane
3f674c9994 Updated readme [skip ci] 2024-03-25 23:33:17 -07:00
Andrew Kane
31e41b3ba9 Added FAQ about binary vectors [skip ci] 2024-03-24 11:07:34 -07:00
Andrew Kane
903a925662 Improved type modifier tests 2024-03-21 17:31:08 -07:00
Andrew Kane
96ff19be44 Version bump to 0.6.2 [skip ci] 2024-03-18 10:21:04 -07:00
Andrew Kane
6c969bebad Updated changelog [skip ci] 2024-03-18 10:11:45 -07:00
Andrew Kane
b64a1482d9 Moved example [skip ci] 2024-03-16 15:20:26 -07:00
Andrew Kane
a5f2d70bc2 Use temp directory for installation instructions on Windows [skip ci] 2024-03-16 12:02:45 -07:00
Andrew Kane
f3fcb5e005 Moved installation notes for Windows [skip ci] 2024-03-16 11:47:18 -07:00
Andrew Kane
3a6e0afb9c Added installation notes for Windows [skip ci] 2024-03-16 11:35:55 -07:00
Andrew Kane
183d50bdbd Added note about creating indexes concurrently [skip ci] 2024-03-16 10:45:09 -07:00
Andrew Kane
bd776fee68 Updated readme [skip ci] 2024-03-16 10:44:45 -07:00
Andrew Kane
d30b113e4b Updated readme [skip ci] 2024-03-15 21:54:58 -07:00
Andrew Kane
fd3200f718 Updated readme [skip ci] 2024-03-15 21:47:57 -07:00
Andrew Kane
02c815d876 Added docs on tuning, monitoring, and scaling [skip ci] 2024-03-15 19:00:49 -07:00
Andrew Kane
4b2a7cc49d Improved performance section [skip ci] 2024-03-15 17:54:14 -07:00
Andrew Kane
da0ff998e9 Updated readme [skip ci] 2024-03-15 14:23:56 -07:00
Andrew Kane
cb36e24289 Improved portability section [skip ci] 2024-03-15 14:23:04 -07:00
Andrew Kane
b1d0d4c7a3 Improved troubleshooting docs [skip ci] 2024-03-15 14:01:24 -07:00
Andrew Kane
1dc6514b66 Updated comment [skip ci] 2024-03-15 12:38:14 -07:00
Andrew Kane
6c53f7ca02 Updated comment [skip ci] 2024-03-15 12:37:47 -07:00
Heikki Linnakangas
0d35a14198 Fix compiler warnings in strict C99 mode (#487)
Redefining a typedef is a C11 feature:

    In file included from src/hnsw.c:10:
    src/hnsw.h:147:5: warning: redefinition of typedef 'HnswElementData' is a C11 feature [-Wtypedef-redefinition]
    }                       HnswElementData;
                            ^
    src/hnsw.h:118:32: note: previous definition is here
    typedef struct HnswElementData HnswElementData;
                                   ^
    src/hnsw.h:163:5: warning: redefinition of typedef 'HnswNeighborArray' is a C11 feature [-Wtypedef-redefinition]
    }                       HnswNeighborArray;
                            ^
    src/hnsw.h:119:34: note: previous definition is here
    typedef struct HnswNeighborArray HnswNeighborArray;
                                     ^
    2 warnings generated.

I got these warnings when I built PostgreSQL with "CC=clang
CFLAGS=-std=gnu99"; other similar options would surely produce the
warnings too.
2024-03-12 02:02:33 -07:00
Andrew Kane
3ea2ce89be Reduced lock contention with parallel HNSW index builds 2024-03-11 20:16:55 -07:00
Andrew Kane
62350b1589 Added note about IVFFlat results [skip ci] 2024-03-06 00:29:55 -08:00
Andrew Kane
dd57309281 Added section about HNSW results - #480 [skip ci] 2024-03-06 00:27:50 -08:00
Andrew Kane
c6ddf62a29 Version bump to 0.6.1 [skip ci] 2024-03-04 10:27:59 -08:00
Andrew Kane
801be04d8b Updated CI [skip ci] 2024-03-02 17:06:11 -08:00
Andrew Kane
587e9ba97c Added link to bulk loading example [skip ci] 2024-02-29 21:25:11 -08:00
Andrew Kane
d57047a935 Moved COPY example and added FORMAT [skip ci] 2024-02-29 20:35:01 -08:00
oneturkmen
f1db1f17e3 Update README.md (#472) 2024-02-29 20:27:53 -08:00
Andrew Kane
2f48c0fac4 Updated changelog [skip ci] 2024-02-29 18:10:19 -08:00
Andrew Kane
84a8aa8176 Added note about HNSW build time [skip ci] 2024-02-29 16:25:58 -08:00
Andrew Kane
f64ebbef50 Added OPTFLAGS to readme [skip ci] 2024-02-29 16:15:22 -08:00
Andrew Kane
ac8156509b Updated vector operators for <= and >= 2024-02-29 15:43:29 -08:00
Andrew Kane
82bf69b479 Fixed vector subtraction being marked as commutative - fixes #475 2024-02-29 14:36:18 -08:00
Andrew Kane
4be2f57916 Updated license year [skip ci] 2024-02-29 13:44:58 -08:00
Andrew Kane
91e3d2905f Fixed sort function for Postgres 12 2024-02-28 16:26:41 -08:00
Andrew Kane
fe2406564f Replaced pairing heap with array in SelectNeighbors - closes #447
Co-authored-by: Heikki Linnakangas <heikki.linnakangas@iki.fi>
2024-02-28 15:47:26 -08:00
Andrew Kane
fa52511eaa Fixed closer caching for Postgres 12 2024-02-28 15:44:38 -08:00
Andrew Kane
447ef4d27a Specify Postgres version for Valgrind [skip ci] 2024-02-28 14:16:48 -08:00
Andrew Kane
b36cd22ccc Enable assertions on CI 2024-02-28 14:15:34 -08:00
Andrew Kane
b447ae4989 Avoid base address for relptr for Postgres < 14.5 2024-02-28 14:10:14 -08:00
Andrew Kane
efed873a3e Revert "Replaced pairing heap with array in SelectNeighbors - closes #447"
This reverts commit 14b278dec9.
2024-02-28 11:33:14 -08:00
Andrew Kane
14b278dec9 Replaced pairing heap with array in SelectNeighbors - closes #447
Co-authored-by: Heikki Linnakangas <heikki.linnakangas@iki.fi>
2024-02-28 11:25:12 -08:00
Andrew Kane
133a728e48 Updated changelog [skip ci] 2024-02-20 16:38:05 -08:00
Andrew Kane
ca10cbaa7d Revert "Remove offsethash"
This reverts commit 1cbd204f52.
2024-02-20 16:07:32 -08:00
Andrew Kane
eb29019a14 Revert "Eliminate a few HnswPtrAccess invocations"
This reverts commit 334c386a45.
2024-02-20 16:07:21 -08:00
Heikki Linnakangas
334c386a45 Eliminate a few HnswPtrAccess invocations
HnswPtrAccess() is pretty cheap, but it stills seems worthwhile to
avoid repeated calls in the hot paths when it can be easily avoided.
2024-02-19 14:13:18 +02:00
Heikki Linnakangas
1cbd204f52 Remove offsethash
The original motivation was to eliminate the superfluous HnswPtrAccess
call from AddToVisited. The caller has to call HnswPtrAccess() anyway,
so it makes sense to pass the HnswElement rather than HnswElementPtr
to AddToVisited(). But then I realized that we can use the
pointer-variant even with shared memory, because the visited-hash is
backend-private, and the addresses where the elements are mapped to in
shared memory are stable within the backend.
2024-02-19 14:12:26 +02:00
Andrew Kane
5ba62fca84 Fixed crash with shared_preload_libraries - fixes #460 2024-02-14 17:13:30 -08:00
Andrew Kane
22cb2a3fe7 Simplified Docker tasks [skip ci] 2024-02-07 12:03:33 -08:00
Andrew Kane
72b144906a Added Valgrind to CI 2024-02-07 09:43:01 -08:00
Andrew Kane
a618c1bc78 Updated port name [skip ci] 2024-02-06 12:43:20 -08:00
Andrew Kane
f43cd0ed98 Improved type [skip ci] 2024-02-05 19:38:52 -08:00
43 changed files with 2391 additions and 620 deletions

View File

@@ -28,7 +28,7 @@ jobs:
dev-files: true
- run: make
env:
PG_CFLAGS: -Wall -Wextra -Werror -Wno-unused-parameter -Wno-sign-compare
PG_CFLAGS: -DUSE_ASSERT_CHECKING -Wall -Wextra -Werror -Wno-unused-parameter -Wno-sign-compare
- run: |
export PG_CONFIG=`which pg_config`
sudo --preserve-env=PG_CONFIG make install
@@ -49,7 +49,7 @@ jobs:
postgres-version: 14
- run: make
env:
PG_CFLAGS: -Wall -Wextra -Werror -Wno-unused-parameter
PG_CFLAGS: -DUSE_ASSERT_CHECKING -Wall -Wextra -Werror -Wno-unused-parameter
- run: make install
- run: make installcheck
- if: ${{ failure() }}
@@ -60,7 +60,9 @@ jobs:
wget -q https://github.com/postgres/postgres/archive/refs/tags/REL_14_10.tar.gz
tar xf REL_14_10.tar.gz
- run: make prove_installcheck PROVE_FLAGS="-I ./postgres-REL_14_10/src/test/perl" PERL5LIB="/Users/runner/perl5/lib/perl5"
- run: make clean && /usr/local/opt/llvm@15/bin/scan-build --status-bugs make PG_CFLAGS="-DUSE_ASSERT_CHECKING"
- run: make clean && /usr/local/opt/llvm@15/bin/scan-build --status-bugs make
env:
PG_CFLAGS: -DUSE_ASSERT_CHECKING
windows:
runs-on: windows-latest
if: ${{ !startsWith(github.ref_name, 'mac') }}
@@ -71,6 +73,7 @@ jobs:
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 && ^
@@ -97,4 +100,15 @@ jobs:
sudo -u postgres make installcheck
sudo -u postgres make prove_installcheck
env:
PG_CFLAGS: -Wall -Wextra -Werror -Wno-unused-parameter -Wno-sign-compare
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
- run: sudo --preserve-env=PG_CONFIG make install
- run: make installcheck

View File

@@ -1,10 +1,16 @@
## 0.7.0 (unreleased)
- Added support for inline filtering with HNSW
- Added `halfvec` type
## 0.6.1 (unreleased)
## 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)

View File

@@ -1,4 +1,4 @@
Portions Copyright (c) 1996-2023, PostgreSQL Global Development Group
Portions Copyright (c) 1996-2024, 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": "0.6.0",
"version": "0.6.2",
"maintainer": [
"Andrew Kane <andrew@ankane.org>"
],
@@ -20,7 +20,7 @@
"vector": {
"file": "sql/vector.sql",
"docfile": "README.md",
"version": "0.6.0",
"version": "0.6.2",
"abstract": "Open-source vector similarity search for Postgres"
}
},

View File

@@ -1,10 +1,10 @@
EXTENSION = vector
EXTVERSION = 0.6.0
EXTVERSION = 0.6.2
MODULE_big = vector
DATA = $(wildcard sql/*--*.sql)
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
HEADERS = src/vector.h
OBJS = 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/vector.o
HEADERS = src/halfvec.h src/vector.h
TESTS = $(wildcard test/sql/*.sql)
REGRESS = $(patsubst test/sql/%.sql,%,$(TESTS))
@@ -71,11 +71,9 @@ PG_MAJOR ?= 16
.PHONY: docker
docker:
docker build --pull --no-cache --build-arg PG_MAJOR=$(PG_MAJOR) -t pgvector/pgvector:pg$(PG_MAJOR) .
docker build --build-arg PG_MAJOR=$(PG_MAJOR) -t pgvector/pgvector:$(EXTVERSION)-pg$(PG_MAJOR) .
docker build --pull --no-cache --build-arg PG_MAJOR=$(PG_MAJOR) -t pgvector/pgvector:pg$(PG_MAJOR) -t pgvector/pgvector:$(EXTVERSION)-pg$(PG_MAJOR) .
.PHONY: docker-release
docker-release:
docker buildx build --push --pull --no-cache --platform linux/amd64,linux/arm64 --build-arg PG_MAJOR=$(PG_MAJOR) -t pgvector/pgvector:pg$(PG_MAJOR) .
docker buildx build --push --platform linux/amd64,linux/arm64 --build-arg PG_MAJOR=$(PG_MAJOR) -t pgvector/pgvector:$(EXTVERSION)-pg$(PG_MAJOR) .
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) .

View File

@@ -1,8 +1,8 @@
EXTENSION = vector
EXTVERSION = 0.6.0
EXTVERSION = 0.6.2
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
HEADERS = src\vector.h
OBJS = 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\vector.obj
HEADERS = src\halfvec.h src\vector.h
REGRESS = btree cast copy functions input ivfflat_cosine ivfflat_ip ivfflat_l2 ivfflat_options ivfflat_unlogged
REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)

206
README.md
View File

@@ -20,13 +20,13 @@ Compile and install the extension (supports Postgres 12+)
```sh
cd /tmp
git clone --branch v0.6.0 https://github.com/pgvector/pgvector.git
git clone --branch v0.6.2 https://github.com/pgvector/pgvector.git
cd pgvector
make
make install # may need sudo
```
See the [installation notes](#installation-notes) if you run into issues
See the [installation notes](#installation-notes---linux-and-mac) if you run into issues
You can also install it with [Docker](#docker), [Homebrew](#homebrew), [PGXN](#pgxn), [APT](#apt), [Yum](#yum), [pkg](#pkg), or [conda-forge](#conda-forge), and it comes preinstalled with [Postgres.app](#postgresapp) and many [hosted providers](#hosted-postgres). There are also instructions for [GitHub Actions](https://github.com/pgvector/setup-pgvector).
@@ -44,12 +44,15 @@ Then use `nmake` to build:
```cmd
set "PGROOT=C:\Program Files\PostgreSQL\16"
git clone --branch v0.6.0 https://github.com/pgvector/pgvector.git
cd %TEMP%
git clone --branch v0.6.2 https://github.com/pgvector/pgvector.git
cd pgvector
nmake /F Makefile.win
nmake /F Makefile.win install
```
See the [installation notes](#installation-notes---windows) if you run into issues
You can also install it with [Docker](#docker) or [conda-forge](#conda-forge).
## Getting Started
@@ -102,6 +105,12 @@ Insert vectors
INSERT INTO items (embedding) VALUES ('[1,2,3]'), ('[4,5,6]');
```
Or load vectors in bulk using `COPY` ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/bulk_loading.py))
```sql
COPY items (embedding) FROM STDIN WITH (FORMAT BINARY);
```
Upsert vectors
```sql
@@ -264,6 +273,8 @@ HINT: Increase maintenance_work_mem to speed up builds.
Note: Do not set `maintenance_work_mem` so high that it exhausts the memory on the server
Like other index types, its faster to create an index after loading your initial data
Starting with 0.6.0, you can also speed up index creation by increasing the number of parallel workers (2 by default)
```sql
@@ -377,12 +388,6 @@ Create an index on one [or more](https://www.postgresql.org/docs/current/indexes
CREATE INDEX ON items (category_id);
```
Or a composite HNSW index for approximate search (added in 0.7.0)
```sql
CREATE INDEX ON items USING hnsw (embedding vector_l2_ops, category_id);
```
Or a [partial index](https://www.postgresql.org/docs/current/indexes-partial.html) on the vector column for approximate search
```sql
@@ -395,6 +400,29 @@ Use [partitioning](https://www.postgresql.org/docs/current/ddl-partitioning.html
CREATE TABLE items (embedding vector(3), category_id int) PARTITION BY LIST(category_id);
```
## Types
- `vector`
- `halfvec`
## Half-Precision
```sql
CREATE TABLE items (id bigserial PRIMARY KEY, embedding halfvec(3));
INSERT INTO items (embedding) VALUES ('[1,2,3]'), ('[4,5,6]');
CREATE INDEX ON items USING hnsw ((embedding::halfvec(3)) halfvec_l2_ops);
```
```sql
SELECT id FROM items ORDER BY embedding::halfvec(3) <-> '[1,2,3]' LIMIT 5;
```
```sql
SELECT id FROM (
SELECT * FROM items ORDER BY embedding::halfvec(3) <-> '[1,2,3]' LIMIT 20
) ORDER BY embedding <-> '[1,2,3]' LIMIT 5;
```
## Hybrid Search
Use together with Postgres [full-text search](https://www.postgresql.org/docs/current/textsearch-intro.html) for hybrid search.
@@ -408,13 +436,39 @@ You can use [Reciprocal Rank Fusion](https://github.com/pgvector/pgvector-python
## Performance
### Tuning
Use a tool like [PgTune](https://pgtune.leopard.in.ua/) to set initial values for Postgres server parameters.
### Loading
Use `COPY` for bulk loading data ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/bulk_loading.py)).
```sql
COPY items (embedding) FROM STDIN WITH (FORMAT BINARY);
```
Add any indexes *after* loading the initial data for best performance.
### Indexing
See index build time for [HNSW](#index-build-time) and [IVFFlat](#index-build-time-1).
In production environments, create indexes concurrently to avoid blocking writes.
```sql
CREATE INDEX CONCURRENTLY ...
```
### Querying
Use `EXPLAIN ANALYZE` to debug performance.
```sql
EXPLAIN ANALYZE SELECT * FROM items ORDER BY embedding <-> '[3,1,2]' LIMIT 5;
```
### Exact Search
#### Exact Search
To speed up queries without an index, increase `max_parallel_workers_per_gather`.
@@ -428,7 +482,7 @@ If vectors are normalized to length 1 (like [OpenAI embeddings](https://platform
SELECT * FROM items ORDER BY embedding <#> '[3,1,2]' LIMIT 5;
```
### Approximate Search
#### Approximate Search
To speed up queries with an IVFFlat index, increase the number of inverted lists (at the expense of recall).
@@ -436,7 +490,7 @@ To speed up queries with an IVFFlat index, increase the number of inverted lists
CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops) WITH (lists = 1000);
```
## Vacuuming
### Vacuuming
Vacuuming can take a while for HNSW indexes. Speed it up by reindexing first.
@@ -445,6 +499,41 @@ REINDEX INDEX CONCURRENTLY index_name;
VACUUM table_name;
```
## Monitoring
Monitor performance with [pg_stat_statements](https://www.postgresql.org/docs/current/pgstatstatements.html) (be sure to add it to `shared_preload_libraries`).
```sql
CREATE EXTENSION pg_stat_statements;
```
Get the most time-consuming queries with:
```sql
SELECT query, calls, ROUND((total_plan_time + total_exec_time) / calls) AS avg_time_ms,
ROUND((total_plan_time + total_exec_time) / 60000) AS total_time_min
FROM pg_stat_statements ORDER BY total_plan_time + total_exec_time DESC LIMIT 20;
```
Note: Replace `total_plan_time + total_exec_time` with `total_time` for Postgres < 13
Monitor recall by comparing results from approximate search with exact search.
```sql
BEGIN;
SET LOCAL enable_indexscan = off; -- use exact search
SELECT ...
COMMIT;
```
## Scaling
Scale pgvector the same way you scale Postgres.
Scale vertically by increasing memory, CPU, and storage on a single instance. Use existing tools to [tune parameters](#tuning) and [monitor performance](#monitoring).
Scale horizontally with [replicas](https://www.postgresql.org/docs/current/hot-standby.html), or use [Citus](https://github.com/citusdata/citus) or another approach for sharding ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/citus.py)).
## Languages
Use pgvector from any language with a Postgres client. You can even generate and store vectors in one language and query them in another.
@@ -538,6 +627,18 @@ and query with:
SELECT * FROM items ORDER BY embedding::vector(3) <-> '[3,1,2]' LIMIT 5;
```
#### Are binary vectors supported?
You can store binary vectors and perform exact nearest neighbor search by Hamming distance in Postgres without an extension ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/hash_image_search.py)).
```tsql
CREATE TABLE items (id bigserial PRIMARY KEY, embedding bit(3));
INSERT INTO items (embedding) VALUES (B'000'), (B'111');
SELECT * FROM items ORDER BY bit_count(embedding # B'101') LIMIT 5;
```
Indexing is not currently supported.
#### Do indexes need to fit into memory?
No, but like other index types, youll likely see better performance if they do. You can get the size of an index with:
@@ -550,7 +651,17 @@ SELECT pg_size_pretty(pg_relation_size('index_name'));
#### Why isnt a query using an index?
The cost estimation in pgvector < 0.4.3 does not always work well with the planner. You can encourage the planner to use an index for a query with:
The query needs to have an `ORDER BY` and `LIMIT`, and the `ORDER BY` must be the result of a distance operator, not an expression.
```sql
-- index
ORDER BY embedding <=> '[3,1,2]' LIMIT 5;
-- no index
ORDER BY 1 - (embedding <=> '[3,1,2]') DESC LIMIT 5;
```
You can encourage the planner to use an index for a query with:
```sql
BEGIN;
@@ -579,6 +690,12 @@ or choose to store vectors inline:
ALTER TABLE items ALTER COLUMN embedding SET STORAGE PLAIN;
```
#### Why are there less results for a query after adding an HNSW index?
Results are limited by the size of the dynamic candidate list (`hnsw.ef_search`). There may be even less results due to dead tuples or filtering conditions in the query. We recommend setting `hnsw.ef_search` to at least twice the `LIMIT` of the query. If you need more than 500 results, use an IVFFlat index instead.
Also, note that `NULL` vectors are not indexed (as well as zero vectors for cosine distance).
#### Why are there less results for a query after adding an IVFFlat index?
The index was likely created with too little data for the number of lists. Drop the index until the table has more data.
@@ -587,6 +704,10 @@ The index was likely created with too little data for the number of lists. Drop
DROP INDEX index_name;
```
Results can also be limited by the number of probes (`ivfflat.probes`).
Also, note that `NULL` vectors are not indexed (as well as zero vectors for cosine distance).
## Reference
### Vector Type
@@ -622,7 +743,28 @@ Function | Description | Added
avg(vector) → vector | average |
sum(vector) → vector | sum | 0.5.0
## Installation Notes
### Halfvec Type
Each half vector takes `2 * dimensions + 8` bytes of storage. Each element is a half-precision floating-point number, and all elements must be finite (no `NaN`, `Infinity` or `-Infinity`). Half vectors can have up to 16,000 dimensions.
### Halfvec Operators
Operator | Description | Added
--- | --- | ---
<-> | Euclidean distance | 0.7.0
<#> | negative inner product | 0.7.0
<=> | cosine distance | 0.7.0
### Halfvec Functions
Function | Description | Added
--- | --- | ---
cosine_distance(halfvec, halfvec) → double precision | cosine distance | 0.7.0
inner_product(halfvec, halfvec) → double precision | inner product | 0.7.0
l2_distance(halfvec, halfvec) → double precision | Euclidean distance | 0.7.0
l1_distance(halfvec, halfvec) → double precision | taxicab distance | 0.7.0
## Installation Notes - Linux and Mac
### Postgres Location
@@ -662,6 +804,26 @@ Note: Replace `16` with your Postgres server version
If compilation fails and the output includes `warning: no such sysroot directory` on Mac, reinstall Xcode Command Line Tools.
### Portability
By default, pgvector compiles with `-march=native` on some platforms for best performance. However, this can lead to `Illegal instruction` errors if trying to run the compiled extension on a different machine.
To compile for portability, use:
```sh
make OPTFLAGS=""
```
## Installation Notes - Windows
### Missing Header
If compilation fails with `Cannot open include file: 'postgres.h': No such file or directory`, make sure `PGROOT` is correct.
### Permissions
If installation fails with `Access is denied`, re-run the installation instructions as an administrator.
## Additional Installation Methods
### Docker
@@ -677,7 +839,7 @@ This adds pgvector to the [Postgres image](https://hub.docker.com/_/postgres) (r
You can also build the image manually:
```sh
git clone --branch v0.6.0 https://github.com/pgvector/pgvector.git
git clone --branch v0.6.2 https://github.com/pgvector/pgvector.git
cd pgvector
docker build --build-arg PG_MAJOR=16 -t myuser/pgvector .
```
@@ -733,7 +895,7 @@ pkg install postgresql15-pg_vector
or the port with:
```sh
cd /usr/ports/databases/pg_vector
cd /usr/ports/databases/pgvector
make install
```
@@ -846,6 +1008,12 @@ make installcheck REGRESS=functions # regression test
make prove_installcheck PROVE_TESTS=test/t/001_ivfflat_wal.pl # TAP test
```
To enable assertions:
```sh
make clean && PG_CFLAGS="-DUSE_ASSERT_CHECKING" make && make install
```
To enable benchmarking:
```sh
@@ -858,12 +1026,6 @@ To show memory usage:
make clean && PG_CFLAGS="-DHNSW_MEMORY -DIVFFLAT_MEMORY" make && make install
```
To enable assertions:
```sh
make clean && PG_CFLAGS="-DUSE_ASSERT_CHECKING" make && make install
```
To get k-means metrics:
```sh

View File

@@ -0,0 +1,16 @@
-- 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

@@ -0,0 +1,2 @@
-- 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

@@ -1,34 +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 OPERATOR CLASS vector_bigint_ops
DEFAULT FOR TYPE bigint USING hnsw AS
OPERATOR 2 < ,
OPERATOR 3 <= ,
OPERATOR 4 = ,
OPERATOR 5 >= ,
OPERATOR 6 > ;
CREATE OPERATOR CLASS vector_integer_ops
DEFAULT FOR TYPE integer USING hnsw AS
OPERATOR 2 < ,
OPERATOR 3 <= ,
OPERATOR 4 = ,
OPERATOR 5 >= ,
OPERATOR 6 > ;
CREATE OPERATOR CLASS vector_text_ops
DEFAULT FOR TYPE text USING hnsw AS
OPERATOR 2 < ,
OPERATOR 3 <= ,
OPERATOR 4 = ,
OPERATOR 5 >= ,
OPERATOR 6 > ;
CREATE OPERATOR CLASS vector_uuid_ops
DEFAULT FOR TYPE uuid USING hnsw AS
OPERATOR 2 < ,
OPERATOR 3 <= ,
OPERATOR 4 = ,
OPERATOR 5 >= ,
OPERATOR 6 > ;

View File

@@ -0,0 +1,128 @@
-- 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 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 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 ASSIGNMENT;
CREATE CAST (vector AS halfvec)
WITH FUNCTION vector_to_halfvec(vector, integer, boolean) AS ASSIGNMENT;

View File

@@ -180,8 +180,7 @@ CREATE OPERATOR + (
);
CREATE OPERATOR - (
LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_sub,
COMMUTATOR = -
LEFTARG = vector, RIGHTARG = vector, PROCEDURE = vector_sub
);
CREATE OPERATOR * (
@@ -195,11 +194,10 @@ 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 = scalarltsel, JOIN = scalarltjoinsel
RESTRICT = scalarlesel, JOIN = scalarlejoinsel
);
CREATE OPERATOR = (
@@ -214,11 +212,10 @@ 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 = scalargtsel, JOIN = scalargtjoinsel
RESTRICT = scalargesel, JOIN = scalargejoinsel
);
CREATE OPERATOR > (
@@ -291,36 +288,144 @@ CREATE OPERATOR CLASS vector_cosine_ops
FUNCTION 1 vector_negative_inner_product(vector, vector),
FUNCTION 2 vector_norm(vector);
-- hnsw attributes
-- halfvec type
CREATE OPERATOR CLASS vector_bigint_ops
DEFAULT FOR TYPE bigint USING hnsw AS
OPERATOR 2 < ,
OPERATOR 3 <= ,
OPERATOR 4 = ,
OPERATOR 5 >= ,
OPERATOR 6 > ;
CREATE TYPE halfvec;
CREATE OPERATOR CLASS vector_integer_ops
DEFAULT FOR TYPE integer USING hnsw AS
OPERATOR 2 < ,
OPERATOR 3 <= ,
OPERATOR 4 = ,
OPERATOR 5 >= ,
OPERATOR 6 > ;
CREATE FUNCTION halfvec_in(cstring, oid, integer) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE OPERATOR CLASS vector_text_ops
DEFAULT FOR TYPE text USING hnsw AS
OPERATOR 2 < ,
OPERATOR 3 <= ,
OPERATOR 4 = ,
OPERATOR 5 >= ,
OPERATOR 6 > ;
CREATE FUNCTION halfvec_out(halfvec) RETURNS cstring
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE OPERATOR CLASS vector_uuid_ops
DEFAULT FOR TYPE uuid USING hnsw AS
OPERATOR 2 < ,
OPERATOR 3 <= ,
OPERATOR 4 = ,
OPERATOR 5 >= ,
OPERATOR 6 > ;
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;
-- 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;

969
src/halfvec.c Normal file
View File

@@ -0,0 +1,969 @@
#include "postgres.h"
#include <math.h>
#include "catalog/pg_type.h"
#include "common/shortest_dec.h"
#include "fmgr.h"
#include "halfvec.h"
#include "lib/stringinfo.h"
#include "libpq/pqformat.h"
#include "port.h" /* for strtof() */
#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 < 130000
#define TYPALIGN_DOUBLE 'd'
#define TYPALIGN_INT 'i'
#endif
/*
* Check if half is NaN
*/
static inline bool
HalfIsNan(half num)
{
#ifdef FLT16_SUPPORT
return isnan(num);
#else
return (num & 0x7C00) == 0x7C00 && (num & 0x7FFF) != 0x7C00;
#endif
}
/*
* Check if half is infinite
*/
static inline bool
HalfIsInf(half num)
{
#ifdef FLT16_SUPPORT
return isinf(num);
#else
return (num & 0x7FFF) == 0x7C00;
#endif
}
/*
* Check if half is zero
*/
static inline bool
HalfIsZero(half num)
{
#ifdef FLT16_SUPPORT
return num == 0;
#else
return (num & 0x7FFF) == 0x0000;
#endif
}
/*
* Get a half from a message buffer
*/
static half
pq_getmsghalf(StringInfo msg)
{
union
{
half h;
uint16 i;
} swap;
swap.i = pq_getmsgint(msg, 2);
return swap.h;
}
/*
* Append a half to a StringInfo buffer
*/
static void
pq_sendhalf(StringInfo buf, half h)
{
union
{
half h;
uint16 i;
} swap;
swap.h = h;
pq_sendint16(buf, swap.i);
}
/*
* Convert a half to a float4
*/
float
HalfToFloat4(half num)
{
#ifdef FLT16_SUPPORT
return (float) num;
#else
/* TODO Improve performance */
/* Assumes same endianness for floats and integers */
union
{
float f;
uint32 i;
} swapfloat;
union
{
half h;
uint16 i;
} swaphalf;
uint16 bin;
uint32 exponent;
uint32 mantissa;
uint32 result;
swaphalf.h = num;
bin = swaphalf.i;
exponent = (bin & 0x7C00) >> 10;
mantissa = bin & 0x03FF;
/* Sign */
result = (bin & 0x8000) << 16;
if (exponent == 31)
{
if (mantissa == 0)
{
/* Infinite */
result |= 0x7F800000;
}
else
{
/* NaN */
result |= 0x7FC00000;
result |= mantissa << 13;
}
}
else if (exponent == 0)
{
/* Subnormal */
if (mantissa != 0)
{
exponent = -14;
for (int i = 0; i < 10; i++)
{
mantissa <<= 1;
exponent -= 1;
if ((mantissa >> 10) % 2 == 1)
{
mantissa &= 0x03ff;
break;
}
}
result |= (exponent + 127) << 23;
result |= mantissa << 13;
}
}
else
{
/* Normal */
result |= (exponent - 15 + 127) << 23;
result |= mantissa << 13;
}
swapfloat.i = result;
return swapfloat.f;
#endif
}
/*
* Convert a float4 to a half
*/
half
Float4ToHalfUnchecked(float num)
{
#ifdef FLT16_SUPPORT
return (_Float16) num;
#else
/* TODO Improve performance */
/* Assumes same endianness for floats and integers */
union
{
float f;
uint32 i;
} swapfloat;
union
{
half h;
uint16 i;
} swaphalf;
uint32 bin;
int exponent;
int mantissa;
uint16 result;
swapfloat.f = num;
bin = swapfloat.i;
exponent = (bin & 0x7F800000) >> 23;
mantissa = bin & 0x007FFFFF;
/* Sign */
result = (bin & 0x80000000) >> 16;
if (isinf(num))
{
/* Infinite */
result |= 0x7C00;
}
else if (isnan(num))
{
/* NaN */
result |= 0x7E00;
result |= mantissa >> 13;
}
else if (exponent > 98)
{
int m;
int gr;
int s;
exponent -= 127;
s = mantissa & 0x00000FFF;
/* Subnormal */
if (exponent < -14)
{
int diff = -exponent - 14;
mantissa >>= diff;
mantissa += 1 << (23 - diff);
s |= mantissa & 0x00000FFF;
}
m = mantissa >> 13;
/* Round */
gr = (mantissa >> 12) % 4;
if (gr == 3 || (gr == 1 && s != 0))
m += 1;
if (m == 1024)
{
m = 0;
exponent += 1;
}
if (exponent > 15)
{
/* Infinite */
result |= 0x7C00;
}
else
{
if (exponent >= -14)
result |= (exponent + 15) << 10;
result |= m;
}
}
swaphalf.i = result;
return swaphalf.h;
#endif
}
/*
* Convert a float4 to a half
*/
half
Float4ToHalf(float num)
{
half result = Float4ToHalfUnchecked(num);
if (unlikely(HalfIsInf(result)) && !isinf(num))
float_overflow_error();
if (unlikely(HalfIsZero(result)) && num != 0.0)
float_underflow_error();
return result;
}
/*
* Ensure same dimensions
*/
static inline void
CheckDims(HalfVector * a, HalfVector * b)
{
if (a->dim != b->dim)
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("different halfvec 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("halfvec must have at least 1 dimension")));
if (dim > HALFVEC_MAX_DIM)
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
errmsg("halfvec cannot have more than %d dimensions", HALFVEC_MAX_DIM)));
}
/*
* Ensure finite element
*/
static inline void
CheckElement(half value)
{
if (HalfIsNan(value))
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("NaN not allowed in halfvec")));
if (HalfIsInf(value))
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("infinite value not allowed in halfvec")));
}
/*
* Allocate and initialize a new half vector
*/
HalfVector *
InitHalfVector(int dim)
{
HalfVector *result;
int size;
size = HALFVEC_SIZE(dim);
result = (HalfVector *) palloc0(size);
SET_VARSIZE(result, size);
result->dim = dim;
return result;
}
/*
* Check for whitespace, since array_isspace() is static
*/
static inline bool
halfvec_isspace(char ch)
{
if (ch == ' ' ||
ch == '\t' ||
ch == '\n' ||
ch == '\r' ||
ch == '\v' ||
ch == '\f')
return true;
return false;
}
#if PG_VERSION_NUM < 120003
static pg_noinline void
float_overflow_error(void)
{
ereport(ERROR,
(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
errmsg("value out of range: overflow")));
}
static pg_noinline void
float_underflow_error(void)
{
ereport(ERROR,
(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
errmsg("value out of range: underflow")));
}
#endif
/*
* Convert textual representation to internal representation
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_in);
Datum
halfvec_in(PG_FUNCTION_ARGS)
{
char *lit = PG_GETARG_CSTRING(0);
int32 typmod = PG_GETARG_INT32(2);
half x[HALFVEC_MAX_DIM];
int dim = 0;
char *pt;
char *stringEnd;
HalfVector *result;
char *litcopy = pstrdup(lit);
char *str = litcopy;
while (halfvec_isspace(*str))
str++;
if (*str != '[')
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("malformed halfvec literal: \"%s\"", lit),
errdetail("Vector contents must start with \"[\".")));
str++;
pt = strtok(str, ",");
stringEnd = pt;
while (pt != NULL && *stringEnd != ']')
{
if (dim == HALFVEC_MAX_DIM)
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
errmsg("halfvec cannot have more than %d dimensions", HALFVEC_MAX_DIM)));
while (halfvec_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 halfvec: \"%s\"", lit)));
/* Use strtof like float4in to avoid a double-rounding problem */
x[dim] = Float4ToHalf(strtof(pt, &stringEnd));
CheckElement(x[dim]);
dim++;
if (stringEnd == pt)
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("invalid input syntax for type halfvec: \"%s\"", lit)));
while (halfvec_isspace(*stringEnd))
stringEnd++;
if (*stringEnd != '\0' && *stringEnd != ']')
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("invalid input syntax for type halfvec: \"%s\"", lit)));
pt = strtok(NULL, ",");
}
if (stringEnd == NULL || *stringEnd != ']')
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("malformed halfvec literal: \"%s\"", lit),
errdetail("Unexpected end of input.")));
stringEnd++;
/* Only whitespace is allowed after the closing brace */
while (halfvec_isspace(*stringEnd))
stringEnd++;
if (*stringEnd != '\0')
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("malformed halfvec literal: \"%s\"", lit),
errdetail("Junk after closing right brace.")));
/* Ensure no consecutive delimiters since strtok skips */
for (pt = lit + 1; *pt != '\0'; pt++)
{
if (pt[-1] == ',' && *pt == ',')
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("malformed halfvec literal: \"%s\"", lit)));
}
if (dim < 1)
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("halfvec must have at least 1 dimension")));
pfree(litcopy);
CheckExpectedDim(typmod, dim);
result = InitHalfVector(dim);
for (int i = 0; i < dim; i++)
result->x[i] = x[i];
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
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_out);
Datum
halfvec_out(PG_FUNCTION_ARGS)
{
HalfVector *vector = PG_GETARG_HALFVEC_P(0);
int dim = vector->dim;
char *buf;
char *ptr;
/*
* Need:
*
* dim * (FLOAT_SHORTEST_DECIMAL_LEN - 1) bytes for
* float_to_shortest_decimal_bufn
*
* dim - 1 bytes for separator
*
* 3 bytes for [, ], and \0
*/
buf = (char *) palloc(FLOAT_SHORTEST_DECIMAL_LEN * dim + 2);
ptr = buf;
AppendChar(ptr, '[');
for (int i = 0; i < dim; i++)
{
if (i > 0)
AppendChar(ptr, ',');
AppendFloat(ptr, HalfToFloat4(vector->x[i]));
}
AppendChar(ptr, ']');
*ptr = '\0';
PG_FREE_IF_COPY(vector, 0);
PG_RETURN_CSTRING(buf);
}
/*
* Convert type modifier
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_typmod_in);
Datum
halfvec_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 halfvec must be at least 1")));
if (*tl > HALFVEC_MAX_DIM)
ereport(ERROR,
(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("dimensions for type halfvec cannot exceed %d", HALFVEC_MAX_DIM)));
PG_RETURN_INT32(*tl);
}
/*
* Convert external binary representation to internal representation
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_recv);
Datum
halfvec_recv(PG_FUNCTION_ARGS)
{
StringInfo buf = (StringInfo) PG_GETARG_POINTER(0);
int32 typmod = PG_GETARG_INT32(2);
HalfVector *result;
int16 dim;
int16 unused;
dim = pq_getmsgint(buf, sizeof(int16));
unused = pq_getmsgint(buf, sizeof(int16));
CheckDim(dim);
CheckExpectedDim(typmod, dim);
if (unused != 0)
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("expected unused to be 0, not %d", unused)));
result = InitHalfVector(dim);
for (int i = 0; i < dim; i++)
{
result->x[i] = pq_getmsghalf(buf);
CheckElement(result->x[i]);
}
PG_RETURN_POINTER(result);
}
/*
* Convert internal representation to the external binary representation
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_send);
Datum
halfvec_send(PG_FUNCTION_ARGS)
{
HalfVector *vec = PG_GETARG_HALFVEC_P(0);
StringInfoData buf;
pq_begintypsend(&buf);
pq_sendint(&buf, vec->dim, sizeof(int16));
pq_sendint(&buf, vec->unused, sizeof(int16));
for (int i = 0; i < vec->dim; i++)
pq_sendhalf(&buf, vec->x[i]);
PG_RETURN_BYTEA_P(pq_endtypsend(&buf));
}
/*
* Convert half vector to half vector
* This is needed to check the type modifier
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec);
Datum
halfvec(PG_FUNCTION_ARGS)
{
HalfVector *vec = PG_GETARG_HALFVEC_P(0);
int32 typmod = PG_GETARG_INT32(1);
CheckExpectedDim(typmod, vec->dim);
PG_RETURN_POINTER(vec);
}
/*
* Convert array to half vector
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(array_to_halfvec);
Datum
array_to_halfvec(PG_FUNCTION_ARGS)
{
ArrayType *array = PG_GETARG_ARRAYTYPE_P(0);
int32 typmod = PG_GETARG_INT32(1);
HalfVector *result;
int16 typlen;
bool typbyval;
char typalign;
Datum *elemsp;
int nelemsp;
if (ARR_NDIM(array) > 1)
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("array must be 1-D")));
if (ARR_HASNULL(array) && array_contains_nulls(array))
ereport(ERROR,
(errcode(ERRCODE_NULL_VALUE_NOT_ALLOWED),
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);
CheckDim(nelemsp);
CheckExpectedDim(typmod, nelemsp);
result = InitHalfVector(nelemsp);
if (ARR_ELEMTYPE(array) == INT4OID)
{
for (int i = 0; i < nelemsp; i++)
result->x[i] = Float4ToHalf(DatumGetInt32(elemsp[i]));
}
else if (ARR_ELEMTYPE(array) == FLOAT8OID)
{
for (int i = 0; i < nelemsp; i++)
result->x[i] = Float4ToHalf(DatumGetFloat8(elemsp[i]));
}
else if (ARR_ELEMTYPE(array) == FLOAT4OID)
{
for (int i = 0; i < nelemsp; i++)
result->x[i] = Float4ToHalf(DatumGetFloat4(elemsp[i]));
}
else if (ARR_ELEMTYPE(array) == NUMERICOID)
{
for (int i = 0; i < nelemsp; i++)
result->x[i] = Float4ToHalf(DatumGetFloat4(DirectFunctionCall1(numeric_float4, elemsp[i])));
}
else
{
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
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]);
PG_RETURN_POINTER(result);
}
/*
* Convert half vector to float4[]
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_to_float4);
Datum
halfvec_to_float4(PG_FUNCTION_ARGS)
{
HalfVector *vec = PG_GETARG_HALFVEC_P(0);
Datum *datums;
ArrayType *result;
datums = (Datum *) palloc(sizeof(Datum) * vec->dim);
for (int i = 0; i < vec->dim; i++)
datums[i] = Float4GetDatum(HalfToFloat4(vec->x[i]));
/* Use TYPALIGN_INT for float4 */
result = construct_array(datums, vec->dim, FLOAT4OID, sizeof(float4), true, TYPALIGN_INT);
pfree(datums);
PG_RETURN_POINTER(result);
}
/*
* Convert vector to half vec
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(vector_to_halfvec);
Datum
vector_to_halfvec(PG_FUNCTION_ARGS)
{
Vector *vec = PG_GETARG_VECTOR_P(0);
int32 typmod = PG_GETARG_INT32(1);
HalfVector *result;
CheckDim(vec->dim);
CheckExpectedDim(typmod, vec->dim);
result = InitHalfVector(vec->dim);
for (int i = 0; i < vec->dim; i++)
{
result->x[i] = Float4ToHalfUnchecked(vec->x[i]);
/* TODO Better error for overflow */
CheckElement(result->x[i]);
}
PG_RETURN_POINTER(result);
}
/*
* Get the L2 distance between half vectors
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_l2_distance);
Datum
halfvec_l2_distance(PG_FUNCTION_ARGS)
{
HalfVector *a = PG_GETARG_HALFVEC_P(0);
HalfVector *b = PG_GETARG_HALFVEC_P(1);
half *ax = a->x;
half *bx = b->x;
float distance = 0.0;
CheckDims(a, b);
/* Auto-vectorized */
for (int i = 0; i < a->dim; i++)
{
float diff = HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]);
distance += diff * diff;
}
PG_RETURN_FLOAT8(sqrt((double) distance));
}
/*
* Get the L2 squared distance between half vectors
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_l2_squared_distance);
Datum
halfvec_l2_squared_distance(PG_FUNCTION_ARGS)
{
HalfVector *a = PG_GETARG_HALFVEC_P(0);
HalfVector *b = PG_GETARG_HALFVEC_P(1);
half *ax = a->x;
half *bx = b->x;
float distance = 0.0;
CheckDims(a, b);
/* Auto-vectorized */
for (int i = 0; i < a->dim; i++)
{
float diff = HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]);
distance += diff * diff;
}
PG_RETURN_FLOAT8((double) distance);
}
/*
* Get the inner product of two half vectors
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_inner_product);
Datum
halfvec_inner_product(PG_FUNCTION_ARGS)
{
HalfVector *a = PG_GETARG_HALFVEC_P(0);
HalfVector *b = PG_GETARG_HALFVEC_P(1);
half *ax = a->x;
half *bx = b->x;
float distance = 0.0;
CheckDims(a, b);
/* Auto-vectorized */
for (int i = 0; i < a->dim; i++)
distance += HalfToFloat4(ax[i]) * HalfToFloat4(bx[i]);
PG_RETURN_FLOAT8((double) distance);
}
/*
* Get the negative inner product of two half vectors
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_negative_inner_product);
Datum
halfvec_negative_inner_product(PG_FUNCTION_ARGS)
{
HalfVector *a = PG_GETARG_HALFVEC_P(0);
HalfVector *b = PG_GETARG_HALFVEC_P(1);
half *ax = a->x;
half *bx = b->x;
float distance = 0.0;
CheckDims(a, b);
/* Auto-vectorized */
for (int i = 0; i < a->dim; i++)
distance += HalfToFloat4(ax[i]) * HalfToFloat4(bx[i]);
PG_RETURN_FLOAT8((double) distance * -1);
}
/*
* Get the cosine distance between two half vectors
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_cosine_distance);
Datum
halfvec_cosine_distance(PG_FUNCTION_ARGS)
{
HalfVector *a = PG_GETARG_HALFVEC_P(0);
HalfVector *b = PG_GETARG_HALFVEC_P(1);
half *ax = a->x;
half *bx = b->x;
float distance = 0.0;
float norma = 0.0;
float normb = 0.0;
double similarity;
CheckDims(a, b);
/* Auto-vectorized */
for (int i = 0; i < a->dim; i++)
{
float axi = HalfToFloat4(ax[i]);
float bxi = HalfToFloat4(bx[i]);
distance += axi * bxi;
norma += axi * axi;
normb += bxi * bxi;
}
/* Use sqrt(a * b) over sqrt(a) * sqrt(b) */
similarity = (double) distance / sqrt((double) norma * (double) normb);
#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;
else if (similarity < -1)
similarity = -1;
PG_RETURN_FLOAT8(1 - similarity);
}
/*
* Get the L1 distance between two half vectors
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_l1_distance);
Datum
halfvec_l1_distance(PG_FUNCTION_ARGS)
{
HalfVector *a = PG_GETARG_HALFVEC_P(0);
HalfVector *b = PG_GETARG_HALFVEC_P(1);
half *ax = a->x;
half *bx = b->x;
float distance = 0.0;
CheckDims(a, b);
/* Auto-vectorized */
for (int i = 0; i < a->dim; i++)
distance += fabsf(HalfToFloat4(ax[i]) - HalfToFloat4(bx[i]));
PG_RETURN_FLOAT8((double) distance);
}
/*
* Get the L2 norm of a half vector
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(halfvec_norm);
Datum
halfvec_norm(PG_FUNCTION_ARGS)
{
HalfVector *a = PG_GETARG_HALFVEC_P(0);
half *ax = a->x;
double norm = 0.0;
/* Auto-vectorized */
for (int i = 0; i < a->dim; i++)
{
double axi = (double) HalfToFloat4(ax[i]);
norm += axi * axi;
}
PG_RETURN_FLOAT8(sqrt(norm));
}

41
src/halfvec.h Normal file
View File

@@ -0,0 +1,41 @@
#ifndef HALFVEC_H
#define HALFVEC_H
#define __STDC_WANT_IEC_60559_TYPES_EXT__
#include <float.h>
#ifdef __FLT16_MAX__
#define FLT16_SUPPORT
#endif
#ifdef FLT16_SUPPORT
#define half _Float16
#define HALF_MAX FLT16_MAX
#else
/* TODO #pragma message("")? */
#define half uint16
#define HALF_MAX 65504
#endif
#define HALFVEC_MAX_DIM 16000
#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

@@ -8,6 +8,7 @@
#include "commands/progress.h"
#include "commands/vacuum.h"
#include "hnsw.h"
#include "miscadmin.h"
#include "utils/guc.h"
#include "utils/selfuncs.h"
@@ -28,7 +29,7 @@ static relopt_kind hnsw_relopt_kind;
* this grows bigger, we should use a shmem_request_hook and
* RequestAddinShmemSpace() to pre-reserve space for this.
*/
static void
void
HnswInitLockTranche(void)
{
int *tranche_ids;
@@ -53,7 +54,8 @@ HnswInitLockTranche(void)
void
HnswInit(void)
{
HnswInitLockTranche();
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",
@@ -200,7 +202,7 @@ hnswhandler(PG_FUNCTION_ARGS)
amroutine->amcanorderbyop = true;
amroutine->amcanbackward = false; /* can change direction mid-scan */
amroutine->amcanunique = false;
amroutine->amcanmulticol = true;
amroutine->amcanmulticol = false;
amroutine->amoptionalkey = true;
amroutine->amsearcharray = false;
amroutine->amsearchnulls = false;

View File

@@ -55,6 +55,10 @@
#define HNSW_UPDATE_ENTRY_GREATER 1
#define HNSW_UPDATE_ENTRY_ALWAYS 2
/* Data types */
#define HNSW_TYPE_VECTOR 1
#define HNSW_TYPE_HALFVEC 2
/* Build phases */
/* PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE is 1 */
#define PROGRESS_HNSW_PHASE_LOAD 2
@@ -80,7 +84,7 @@
#if PG_VERSION_NUM < 130000
#define list_delete_last(list) list_truncate(list, list_length(list) - 1)
#define list_sort(list, cmp) list_qsort(list, cmp)
#define list_sort(list, cmp) ((list) = list_qsort(list, cmp))
#endif
#define HnswIsElementTuple(tup) ((tup)->type == HNSW_ELEMENT_TUPLE_TYPE)
@@ -128,9 +132,8 @@ HnswPtrDeclare(HnswElementData, HnswElementRelptr, HnswElementPtr);
HnswPtrDeclare(HnswNeighborArray, HnswNeighborArrayRelptr, HnswNeighborArrayPtr);
HnswPtrDeclare(HnswNeighborArrayPtr, HnswNeighborsRelptr, HnswNeighborsPtr);
HnswPtrDeclare(char, DatumRelptr, DatumPtr);
HnswPtrDeclare(IndexTupleData, IndexTupleRelptr, IndexTuplePtr);
typedef struct HnswElementData
struct HnswElementData
{
HnswElementPtr next;
ItemPointerData heaptids[HNSW_HEAPTIDS];
@@ -144,9 +147,8 @@ typedef struct HnswElementData
OffsetNumber neighborOffno;
BlockNumber neighborPage;
DatumPtr value;
IndexTuplePtr itup;
LWLock lock;
} HnswElementData;
};
typedef HnswElementData * HnswElement;
@@ -157,12 +159,12 @@ typedef struct HnswCandidate
bool closer;
} HnswCandidate;
typedef struct HnswNeighborArray
struct HnswNeighborArray
{
int length;
bool closerSet;
HnswCandidate items[FLEXIBLE_ARRAY_MEMBER];
} HnswNeighborArray;
};
typedef struct HnswPairingHeapNode
{
@@ -187,6 +189,7 @@ typedef struct HnswGraph
/* Entry state */
LWLock entryLock;
LWLock entryWaitLock;
HnswElementPtr entryPoint;
/* Allocations state */
@@ -243,6 +246,7 @@ typedef struct HnswBuildState
Relation index;
IndexInfo *indexInfo;
ForkNumber forkNum;
int type;
/* Settings */
int dimensions;
@@ -263,7 +267,6 @@ typedef struct HnswBuildState
HnswGraph *graph;
double ml;
int maxLevel;
Vector *normvec;
/* Memory */
MemoryContext graphCtx;
@@ -368,11 +371,12 @@ typedef struct HnswVacuumState
int HnswGetM(Relation index);
int HnswGetEfConstruction(Relation index);
FmgrInfo *HnswOptionalProcInfo(Relation index, uint16 procnum);
bool HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
int HnswGetType(Relation index);
bool HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, int type);
Buffer HnswNewBuffer(Relation index, ForkNumber forkNum);
void HnswInitPage(Buffer buf, Page page);
void HnswInit(void);
List *HnswSearchLayer(char *base, Datum q, IndexScanDesc scan, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement);
List *HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement);
HnswElement HnswGetEntryPoint(Relation index);
void HnswGetMetaPageInfo(Relation index, int *m, HnswElement * entryPoint);
void *HnswAlloc(HnswAllocator * allocator, Size size);
@@ -386,14 +390,12 @@ 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, Relation index);
void HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec, IndexScanDesc scan, bool *matches);
void HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element, bool useIndexTuple);
void HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec);
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);
TupleDesc HnswTupleDesc(Relation index);
IndexTuple HnswFormIndexTuple(Relation index, TupleDesc tupdesc, Datum value, Datum *values, bool *isnull);
bool HnswElementIsDuplicate(char *base, HnswElement a, HnswElement b, Relation index);
void HnswInitLockTranche(void);
PGDLLEXPORT void HnswParallelBuildMain(dsm_segment *seg, shm_toc *toc);
/* Index access methods */

View File

@@ -157,7 +157,6 @@ CreateGraphPages(HnswBuildState * buildstate)
Page page;
HnswElementPtr iter = buildstate->graph->head;
char *base = buildstate->hnswarea;
bool useIndexTuple = IndexRelationGetNumberOfAttributes(index) > 1;
/* Calculate sizes */
maxSize = HNSW_MAX_SIZE;
@@ -178,7 +177,6 @@ CreateGraphPages(HnswBuildState * buildstate)
Size ntupSize;
Size combinedSize;
Pointer valuePtr = HnswPtrAccess(base, element->value);
IndexTuple itup = HnswPtrAccess(base, element->itup);
/* Update iterator */
iter = element->next;
@@ -187,7 +185,7 @@ CreateGraphPages(HnswBuildState * buildstate)
MemSet(etup, 0, HNSW_TUPLE_ALLOC_SIZE);
/* Calculate sizes */
etupSize = HNSW_ELEMENT_TUPLE_SIZE(useIndexTuple ? IndexTupleSize(itup) : VARSIZE_ANY(valuePtr));
etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(valuePtr));
ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(element->level, buildstate->m);
combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
@@ -195,7 +193,7 @@ CreateGraphPages(HnswBuildState * buildstate)
if (etupSize > HNSW_TUPLE_ALLOC_SIZE)
elog(ERROR, "index tuple too large");
HnswSetElementTuple(base, etup, element, useIndexTuple);
HnswSetElementTuple(base, etup, element);
/* Keep element and neighbors on the same page if possible */
if (PageGetFreeSpace(page) < etupSize || (combinedSize <= maxSize && PageGetFreeSpace(page) < combinedSize))
@@ -336,17 +334,19 @@ AddDuplicateInMemory(HnswElement element, HnswElement dup)
* Find duplicate element
*/
static bool
FindDuplicateInMemory(char *base, HnswElement element, Relation index)
FindDuplicateInMemory(char *base, HnswElement element)
{
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 (!HnswElementIsDuplicate(base, element, neighborElement, index))
if (!datumIsEqual(value, neighborValue, false, -1))
return false;
/* Check for space */
@@ -406,7 +406,7 @@ UpdateGraphInMemory(FmgrInfo *procinfo, Oid collation, HnswElement element, int
char *base = buildstate->hnswarea;
/* Look for duplicate */
if (FindDuplicateInMemory(base, element, buildstate->index))
if (FindDuplicateInMemory(base, element))
return;
/* Add element */
@@ -431,10 +431,15 @@ InsertTupleInMemory(HnswBuildState * buildstate, HnswElement element)
HnswGraph *graph = buildstate->graph;
HnswElement entryPoint;
LWLock *entryLock = &graph->entryLock;
LWLock *entryWaitLock = &graph->entryWaitLock;
int efConstruction = buildstate->efConstruction;
int m = buildstate->m;
char *base = buildstate->hnswarea;
/* Wait if another process needs exclusive lock on entry lock */
LWLockAcquire(entryWaitLock, LW_EXCLUSIVE);
LWLockRelease(entryWaitLock);
/* Get entry point */
LWLockAcquire(entryLock, LW_SHARED);
entryPoint = HnswPtrAccess(base, graph->entryPoint);
@@ -445,8 +450,10 @@ InsertTupleInMemory(HnswBuildState * buildstate, HnswElement element)
/* Release shared lock */
LWLockRelease(entryLock);
/* Get exclusive lock */
/* Tell other processes to wait and get exclusive lock */
LWLockAcquire(entryWaitLock, LW_EXCLUSIVE);
LWLockAcquire(entryLock, LW_EXCLUSIVE);
LWLockRelease(entryWaitLock);
/* Get latest entry point after lock is acquired */
entryPoint = HnswPtrAccess(base, graph->entryPoint);
@@ -471,14 +478,10 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
HnswGraph *graph = buildstate->graph;
HnswElement element;
HnswAllocator *allocator = &buildstate->allocator;
IndexTuple itup;
Size itupSize;
IndexTuple itupPtr;
Size valueSize;
Pointer valuePtr;
LWLock *flushLock = &graph->flushLock;
char *base = buildstate->hnswarea;
TupleDesc tupdesc = HnswTupleDesc(index);
bool unused;
/* Detoast once for all calls */
Datum value = PointerGetDatum(PG_DETOAST_DATUM(values[0]));
@@ -486,13 +489,12 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
/* Normalize if needed */
if (buildstate->normprocinfo != NULL)
{
if (!HnswNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->normvec))
if (!HnswNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->type))
return false;
}
/* Get datum size */
itup = HnswFormIndexTuple(index, tupdesc, value, values, isnull);
itupSize = IndexTupleSize(itup);
valueSize = VARSIZE_ANY(DatumGetPointer(value));
/* Ensure graph not flushed when inserting */
LWLockAcquire(flushLock, LW_SHARED);
@@ -539,7 +541,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
/* Ok, we can proceed to allocate the element */
element = HnswInitElement(base, heaptid, buildstate->m, buildstate->ml, buildstate->maxLevel, allocator);
itupPtr = HnswAlloc(allocator, itupSize);
valuePtr = HnswAlloc(allocator, valueSize);
/*
* We have now allocated the space needed for the element, so we don't
@@ -548,10 +550,8 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
*/
LWLockRelease(&graph->allocatorLock);
/* Copy the index tuple */
memcpy(itupPtr, itup, itupSize);
HnswPtrStore(base, element->itup, itupPtr);
valuePtr = DatumGetPointer(index_getattr(itupPtr, 1, tupdesc, &unused));
/* Copy the datum */
memcpy(valuePtr, DatumGetPointer(value), valueSize);
HnswPtrStore(base, element->value, valuePtr);
/* Create a lock for the element */
@@ -608,6 +608,9 @@ BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
static void
InitGraph(HnswGraph * graph, char *base, long memoryTotal)
{
/* Initialize the lock tranche if needed */
HnswInitLockTranche();
HnswPtrStore(base, graph->head, (HnswElement) NULL);
HnswPtrStore(base, graph->entryPoint, (HnswElement) NULL);
graph->memoryUsed = 0;
@@ -616,6 +619,7 @@ InitGraph(HnswGraph * graph, char *base, long memoryTotal)
graph->indtuples = 0;
SpinLockInit(&graph->lock);
LWLockInitialize(&graph->entryLock, hnsw_lock_tranche_id);
LWLockInitialize(&graph->entryWaitLock, hnsw_lock_tranche_id);
LWLockInitialize(&graph->allocatorLock, hnsw_lock_tranche_id);
LWLockInitialize(&graph->flushLock, hnsw_lock_tranche_id);
}
@@ -667,34 +671,27 @@ HnswSharedMemoryAlloc(Size size, void *state)
static void
InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, IndexInfo *indexInfo, ForkNumber forkNum)
{
int maxDimensions = HNSW_MAX_DIM;
buildstate->heap = heap;
buildstate->index = index;
buildstate->indexInfo = indexInfo;
buildstate->forkNum = forkNum;
buildstate->type = HnswGetType(index);
buildstate->m = HnswGetM(index);
buildstate->efConstruction = HnswGetEfConstruction(index);
buildstate->dimensions = TupleDescAttr(index->rd_att, 0)->atttypmod;
/* For now */
if (IndexRelationGetNumberOfKeyAttributes(index) > 3)
elog(ERROR, "index cannot have more than three columns");
if (!OidIsValid(index_getprocid(index, 1, HNSW_DISTANCE_PROC)))
elog(ERROR, "first column must be a vector");
for (int i = 1; i < IndexRelationGetNumberOfKeyAttributes(index); i++)
{
if (OidIsValid(index_getprocid(index, i + 1, HNSW_DISTANCE_PROC)))
elog(ERROR, "column %d cannot be a vector", i + 1);
}
if (buildstate->type == HNSW_TYPE_HALFVEC)
maxDimensions *= 2;
/* Require column to have dimensions to be indexed */
if (buildstate->dimensions < 0)
elog(ERROR, "column does not have dimensions");
if (buildstate->dimensions > HNSW_MAX_DIM)
elog(ERROR, "column cannot have more than %d dimensions for hnsw index", HNSW_MAX_DIM);
if (buildstate->dimensions > maxDimensions)
elog(ERROR, "column cannot have more than %d dimensions for hnsw index", maxDimensions);
if (buildstate->efConstruction < 2 * buildstate->m)
elog(ERROR, "ef_construction must be greater than or equal to 2 * m");
@@ -712,9 +709,6 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
buildstate->ml = HnswGetMl(buildstate->m);
buildstate->maxLevel = HnswGetMaxLevel(buildstate->m);
/* Reuse for each tuple */
buildstate->normvec = InitVector(buildstate->dimensions);
buildstate->graphCtx = GenerationContextCreate(CurrentMemoryContext,
"Hnsw build graph context",
#if PG_VERSION_NUM >= 150000
@@ -738,7 +732,6 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
static void
FreeBuildState(HnswBuildState * buildstate)
{
pfree(buildstate->normvec);
MemoryContextDelete(buildstate->graphCtx);
MemoryContextDelete(buildstate->tmpCtx);
}
@@ -994,6 +987,14 @@ HnswBeginParallel(HnswBuildState * buildstate, bool isconcurrent, int request)
/* Report less than allocated so never fails */
InitGraph(&hnswshared->graphData, hnswarea, esthnswarea - 1024 * 1024);
/*
* Avoid base address for relptr for Postgres < 14.5
* https://github.com/postgres/postgres/commit/7201cd18627afc64850537806da7f22150d1a83b
*/
#if PG_VERSION_NUM < 140005
hnswshared->graphData.memoryUsed += MAXALIGN(1);
#endif
shm_toc_insert(pcxt->toc, PARALLEL_KEY_HNSW_SHARED, hnswshared);
shm_toc_insert(pcxt->toc, PARALLEL_KEY_HNSW_AREA, hnswarea);

View File

@@ -136,12 +136,9 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
OffsetNumber freeNeighborOffno = InvalidOffsetNumber;
BlockNumber newInsertPage = InvalidBlockNumber;
char *base = NULL;
bool useIndexTuple = IndexRelationGetNumberOfAttributes(index) > 1;
Pointer valuePtr = HnswPtrAccess(base, e->value);
IndexTuple itup = HnswPtrAccess(base, e->itup);
/* Calculate sizes */
etupSize = HNSW_ELEMENT_TUPLE_SIZE(useIndexTuple ? IndexTupleSize(itup) : VARSIZE_ANY(valuePtr));
etupSize = HNSW_ELEMENT_TUPLE_SIZE(VARSIZE_ANY(HnswPtrAccess(base, e->value)));
ntupSize = HNSW_NEIGHBOR_TUPLE_SIZE(e->level, m);
combinedSize = etupSize + ntupSize + sizeof(ItemIdData);
maxSize = HNSW_MAX_SIZE;
@@ -149,7 +146,7 @@ AddElementOnDisk(Relation index, HnswElement e, int m, BlockNumber insertPage, B
/* Prepare element tuple */
etup = palloc0(etupSize);
HnswSetElementTuple(base, etup, e, useIndexTuple);
HnswSetElementTuple(base, etup, e);
/* Prepare neighbor tuple */
ntup = palloc0(ntupSize);
@@ -507,14 +504,16 @@ 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 (!HnswElementIsDuplicate(base, element, neighborElement, index))
if (!datumIsEqual(value, neighborValue, false, -1))
return false;
if (AddDuplicateOnDisk(index, element, neighborElement, building))
@@ -565,10 +564,6 @@ HnswInsertTupleOnDisk(Relation index, Datum value, Datum *values, bool *isnull,
Oid collation = index->rd_indcollation[0];
LOCKMODE lockmode = ShareLock;
char *base = NULL;
TupleDesc tupdesc = HnswTupleDesc(index);
bool unused;
IndexTuple itup;
Pointer valuePtr;
/*
* Get a shared lock. This allows vacuum to ensure no in-flight inserts
@@ -582,10 +577,7 @@ HnswInsertTupleOnDisk(Relation index, Datum value, Datum *values, bool *isnull,
/* Create an element */
element = HnswInitElement(base, heap_tid, m, HnswGetMl(m), HnswGetMaxLevel(m), NULL);
itup = HnswFormIndexTuple(index, tupdesc, value, values, isnull);
HnswPtrStore(base, element->itup, itup);
valuePtr = DatumGetPointer(index_getattr(itup, 1, tupdesc, &unused));
HnswPtrStore(base, element->value, valuePtr);
HnswPtrStore(base, element->value, DatumGetPointer(value));
/* Prevent concurrent inserts when likely updating entry point */
if (entryPoint == NULL || element->level > entryPoint->level)
@@ -630,7 +622,7 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
if (normprocinfo != NULL)
{
if (!HnswNormValue(normprocinfo, collation, &value, NULL))
if (!HnswNormValue(normprocinfo, collation, &value, HnswGetType(index)))
return;
}

View File

@@ -1,6 +1,7 @@
#include "postgres.h"
#include "access/relscan.h"
#include "halfvec.h"
#include "hnsw.h"
#include "pgstat.h"
#include "storage/bufmgr.h"
@@ -33,12 +34,11 @@ GetScanItems(IndexScanDesc scan, Datum q)
for (int lc = entryPoint->level; lc >= 1; lc--)
{
w = HnswSearchLayer(base, q, NULL, ep, 1, lc, index, procinfo, collation, m, false, NULL);
w = HnswSearchLayer(base, q, ep, 1, lc, index, procinfo, collation, m, false, NULL);
ep = w;
}
/* Only pass scan to check matches for layer 0 */
return HnswSearchLayer(base, q, scan, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL);
return HnswSearchLayer(base, q, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL);
}
/*
@@ -74,7 +74,14 @@ GetScanValue(IndexScanDesc scan)
Datum value;
if (scan->orderByData->sk_flags & SK_ISNULL)
value = PointerGetDatum(InitVector(GetDimensions(scan->indexRelation)));
{
int dimensions = GetDimensions(scan->indexRelation);
if (HnswGetType(scan->indexRelation) == HNSW_TYPE_HALFVEC)
value = PointerGetDatum(InitHalfVector(dimensions));
else
value = PointerGetDatum(InitVector(dimensions));
}
else
{
value = scan->orderByData->sk_argument;
@@ -85,7 +92,7 @@ GetScanValue(IndexScanDesc scan)
/* Fine if normalization fails */
if (so->normprocinfo != NULL)
HnswNormValue(so->normprocinfo, so->collation, &value, NULL);
HnswNormValue(so->normprocinfo, so->collation, &value, HnswGetType(scan->indexRelation));
}
return value;
@@ -203,7 +210,7 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
MemoryContextSwitchTo(oldCtx);
scan->xs_heaptid = *heaptid;
scan->xs_recheck = scan->numberOfKeys > 0;
scan->xs_recheck = false;
scan->xs_recheckorderby = false;
return true;
}

View File

@@ -3,19 +3,17 @@
#include <math.h>
#include "access/generic_xlog.h"
#include "access/relscan.h"
#include "catalog/pg_type.h"
#include "halfvec.h"
#include "hnsw.h"
#include "lib/pairingheap.h"
#include "storage/bufmgr.h"
#include "utils/datum.h"
#include "utils/memdebug.h"
#include "utils/rel.h"
#include "utils/syscache.h"
#include "vector.h"
#if PG_VERSION_NUM < 130000
#define TYPSTORAGE_PLAIN 'p'
#endif
#if PG_VERSION_NUM >= 130000
#include "common/hashfn.h"
#else
@@ -154,6 +152,32 @@ HnswOptionalProcInfo(Relation index, uint16 procnum)
return index_getprocinfo(index, 1, procnum);
}
/*
* Get type
*/
int
HnswGetType(Relation index)
{
Oid typeOid = TupleDescAttr(index->rd_att, 0)->atttypid;
HeapTuple tuple;
Form_pg_type type;
int result;
tuple = SearchSysCache1(TYPEOID, ObjectIdGetDatum(typeOid));
if (!HeapTupleIsValid(tuple))
elog(ERROR, "cache lookup failed for type %u", typeOid);
type = (Form_pg_type) GETSTRUCT(tuple);
if (strcmp(NameStr(type->typname), "halfvec") == 0)
result = HNSW_TYPE_HALFVEC;
else
result = HNSW_TYPE_VECTOR;
ReleaseSysCache(tuple);
return result;
}
/*
* Divide by the norm
*
@@ -163,21 +187,37 @@ HnswOptionalProcInfo(Relation index, uint16 procnum)
* if it's different than the original value
*/
bool
HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result)
HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, int type)
{
double norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
if (norm > 0)
{
Vector *v = DatumGetVector(*value);
if (type == HNSW_TYPE_HALFVEC)
{
HalfVector *v = DatumGetHalfVector(*value);
HalfVector *result = InitHalfVector(v->dim);
if (result == NULL)
result = InitVector(v->dim);
for (int i = 0; i < v->dim; i++)
{
/* TODO Fix */
result->x[i] = Float4ToHalfUnchecked(HalfToFloat4(v->x[i]) / norm);
}
for (int i = 0; i < v->dim; i++)
result->x[i] = v->x[i] / norm;
*value = PointerGetDatum(result);
}
else if (type == HNSW_TYPE_VECTOR)
{
Vector *v = DatumGetVector(*value);
Vector *result = InitVector(v->dim);
*value = PointerGetDatum(result);
for (int i = 0; i < v->dim; i++)
result->x[i] = v->x[i] / norm;
*value = PointerGetDatum(result);
}
else
elog(ERROR, "Unsupported type");
return true;
}
@@ -300,82 +340,6 @@ HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno)
return element;
}
/*
* Get the tuple descriptor
*/
TupleDesc
HnswTupleDesc(Relation index)
{
TupleDesc tupdesc = CreateTupleDescCopyConstr(RelationGetDescr(index));
/* Prevent compression */
TupleDescAttr(tupdesc, 0)->attstorage = TYPSTORAGE_PLAIN;
return tupdesc;
}
/*
* Form an index tuple
*/
IndexTuple
HnswFormIndexTuple(Relation index, TupleDesc tupdesc, Datum value, Datum *values, bool *isnull)
{
Size size = sizeof(Datum) * IndexRelationGetNumberOfAttributes(index);
Datum *newValues = palloc(size);
memcpy(newValues, values, size);
newValues[0] = value;
return index_form_tuple(tupdesc, newValues, isnull);
}
/*
* Check if elements are duplicates
*/
bool
HnswElementIsDuplicate(char *base, HnswElement a, HnswElement b, Relation index)
{
if (IndexRelationGetNumberOfAttributes(index) == 1)
{
Datum value = HnswGetValue(base, a);
Datum value2 = HnswGetValue(base, b);
return datumIsEqual(value, value2, false, -1);
}
else
{
TupleDesc tupdesc = RelationGetDescr(index);
IndexTuple itup = HnswPtrAccess(base, a->itup);
IndexTuple itup2 = HnswPtrAccess(base, b->itup);
for (int i = 0; i < tupdesc->natts; i++)
{
Datum value;
Datum value2;
bool isnull;
bool isnull2;
value = index_getattr(itup, i + 1, tupdesc, &isnull);
value2 = index_getattr(itup2, i + 1, tupdesc, &isnull2);
if (isnull || isnull2)
{
if (isnull != isnull2)
return false;
}
else
{
Form_pg_attribute att = TupleDescAttr(tupdesc, i);
if (!datumIsEqual(value, value2, att->attbyval, att->attlen))
return false;
}
}
return true;
}
}
/*
* Get the metapage info
*/
@@ -485,8 +449,10 @@ HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, Bloc
* Set element tuple, except for neighbor info
*/
void
HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element, bool useIndexTuple)
HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element)
{
Pointer valuePtr = HnswPtrAccess(base, element->value);
etup->type = HNSW_ELEMENT_TUPLE_TYPE;
etup->level = element->level;
etup->deleted = 0;
@@ -497,19 +463,7 @@ HnswSetElementTuple(char *base, HnswElementTuple etup, HnswElement element, bool
else
ItemPointerSetInvalid(&etup->heaptids[i]);
}
if (useIndexTuple)
{
IndexTuple itup = HnswPtrAccess(base, element->itup);
memcpy(&etup->data, itup, IndexTupleSize(itup));
}
else
{
Pointer valuePtr = HnswPtrAccess(base, element->value);
memcpy(&etup->data, valuePtr, VARSIZE_ANY(valuePtr));
}
memcpy(&etup->data, valuePtr, VARSIZE_ANY(valuePtr));
}
/*
@@ -613,7 +567,7 @@ HnswLoadNeighbors(HnswElement element, Relation index, int m)
* Load an element from a tuple
*/
void
HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec, Relation index)
HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHeaptids, bool loadVec)
{
element->level = etup->level;
element->deleted = etup->deleted;
@@ -636,64 +590,17 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
if (loadVec)
{
char *base = NULL;
Datum value = datumCopy(PointerGetDatum(&etup->data), false, -1);
if (IndexRelationGetNumberOfAttributes(index) > 1)
{
TupleDesc tupdesc = RelationGetDescr(index);
bool unused;
IndexTuple itup = CopyIndexTuple((IndexTuple) &etup->data);
Datum value = index_getattr(itup, 1, tupdesc, &unused);
HnswPtrStore(base, element->itup, itup);
HnswPtrStore(base, element->value, DatumGetPointer(value));
}
else
{
Datum value = datumCopy(PointerGetDatum(&etup->data), false, -1);
HnswPtrStore(base, element->value, DatumGetPointer(value));
}
HnswPtrStore(base, element->value, DatumGetPointer(value));
}
}
/*
* Check if an element matches
*/
static bool
HnswCheckMatches(Relation index, HnswElementTuple etup, IndexScanDesc scan)
{
if (scan == NULL)
return true;
for (int i = 0; i < scan->numberOfKeys; i++)
{
IndexTuple itup = (IndexTuple) &etup->data;
TupleDesc tupdesc = RelationGetDescr(index);
ScanKey key = &scan->keyData[i];
bool isnull;
Datum value = index_getattr(itup, key->sk_attno, tupdesc, &isnull);
bool attnull = key->sk_flags & SK_ISNULL;
if (isnull || attnull)
{
if (isnull != attnull)
return false;
}
else
{
if (!DatumGetBool(FunctionCall2Coll(&key->sk_func, key->sk_collation, value, key->sk_argument)))
return false;
}
}
return true;
}
/*
* Load an element and optionally get its distance from q
*/
void
HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec, IndexScanDesc scan, bool *matches)
HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec)
{
Buffer buf;
Page page;
@@ -709,29 +616,11 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
Assert(HnswIsElementTuple(etup));
/* Load element */
HnswLoadElementFromTuple(element, etup, true, loadVec, index);
HnswLoadElementFromTuple(element, etup, true, loadVec);
/* Calculate distance */
if (distance != NULL)
{
Datum value;
if (IndexRelationGetNumberOfAttributes(index) > 1)
{
IndexTuple itup = (IndexTuple) &etup->data;
TupleDesc tupdesc = RelationGetDescr(index);
bool unused;
value = index_getattr(itup, 1, tupdesc, &unused);
}
else
value = PointerGetDatum(&etup->data);
*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, value));
}
if (matches != NULL)
*matches = HnswCheckMatches(index, etup, scan);
*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->data)));
UnlockReleaseBuffer(buf);
}
@@ -760,7 +649,7 @@ HnswEntryCandidate(char *base, HnswElement entryPoint, Datum q, Relation index,
if (index == NULL)
hc->distance = GetCandidateDistance(base, hc, q, procinfo, collation);
else
HnswLoadElement(entryPoint, &hc->distance, &q, index, procinfo, collation, loadVec, NULL, NULL);
HnswLoadElement(entryPoint, &hc->distance, &q, index, procinfo, collation, loadVec);
return hc;
}
@@ -878,14 +767,12 @@ CountElement(char *base, HnswElement skipElement, HnswCandidate * hc)
* Algorithm 2 from paper
*/
List *
HnswSearchLayer(char *base, Datum q, IndexScanDesc scan, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement)
HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, FmgrInfo *procinfo, Oid collation, int m, bool inserting, HnswElement skipElement)
{
List *w = NIL;
pairingheap *C = pairingheap_allocate(CompareNearestCandidates, NULL);
pairingheap *W = pairingheap_allocate(CompareFurthestCandidates, NULL);
int wlen = 0;
uint64 additional = 0;
uint64 maxAdditional = (scan != NULL && scan->numberOfKeys > 0) ? 100 * ef : 0;
visited_hash v;
ListCell *lc2;
HnswNeighborArray *neighborhoodData = NULL;
@@ -957,18 +844,14 @@ HnswSearchLayer(char *base, Datum q, IndexScanDesc scan, List *ep, int ef, int l
if (!visited)
{
float eDistance;
bool eMatches;
HnswElement eElement = HnswPtrAccess(base, e->element);
f = ((HnswPairingHeapNode *) pairingheap_first(W))->inner;
if (index == NULL)
{
eDistance = GetCandidateDistance(base, e, q, procinfo, collation);
eMatches = true;
}
else
HnswLoadElement(eElement, &eDistance, &q, index, procinfo, collation, inserting, scan, &eMatches);
HnswLoadElement(eElement, &eDistance, &q, index, procinfo, collation, inserting);
Assert(!eElement->deleted);
@@ -987,16 +870,6 @@ HnswSearchLayer(char *base, Datum q, IndexScanDesc scan, List *ep, int ef, int l
pairingheap_add(C, &(CreatePairingHeapNode(ec)->ph_node));
pairingheap_add(W, &(CreatePairingHeapNode(ec)->ph_node));
/*
* Do not count elements that do not match filter towards
* ef
*/
if (!eMatches)
{
if ((++additional) <= maxAdditional)
continue;
}
/*
* Do not count elements being deleted towards ef when
* vacuuming. It would be ideal to do this for inserts as
@@ -1032,12 +905,15 @@ HnswSearchLayer(char *base, Datum q, IndexScanDesc scan, List *ep, int ef, int l
static int
#if PG_VERSION_NUM >= 130000
CompareCandidateDistances(const ListCell *a, const ListCell *b)
{
HnswCandidate *hca = lfirst(a);
HnswCandidate *hcb = lfirst(b);
#else
CompareCandidateDistances(const void *a, const void *b)
#endif
{
HnswCandidate *hca = lfirst((ListCell *) a);
HnswCandidate *hcb = lfirst((ListCell *) b);
HnswCandidate *hca = lfirst(*(ListCell **) a);
HnswCandidate *hcb = lfirst(*(ListCell **) b);
#endif
if (hca->distance < hcb->distance)
return 1;
@@ -1060,12 +936,15 @@ CompareCandidateDistances(const void *a, const void *b)
static int
#if PG_VERSION_NUM >= 130000
CompareCandidateDistancesOffset(const ListCell *a, const ListCell *b)
{
HnswCandidate *hca = lfirst(a);
HnswCandidate *hcb = lfirst(b);
#else
CompareCandidateDistancesOffset(const void *a, const void *b)
#endif
{
HnswCandidate *hca = lfirst((ListCell *) a);
HnswCandidate *hcb = lfirst((ListCell *) b);
HnswCandidate *hca = lfirst(*(ListCell **) a);
HnswCandidate *hcb = lfirst(*(ListCell **) b);
#endif
if (hca->distance < hcb->distance)
return 1;
@@ -1124,7 +1003,9 @@ SelectNeighbors(char *base, List *c, int lm, int lc, FmgrInfo *procinfo, Oid col
{
List *r = NIL;
List *w = list_copy(c);
pairingheap *wd;
HnswCandidate **wd;
int wdlen = 0;
int wdoff = 0;
HnswNeighborArray *neighbors = HnswGetNeighbors(base, e2, lc);
bool mustCalculate = !neighbors->closerSet;
List *added = NIL;
@@ -1133,7 +1014,7 @@ SelectNeighbors(char *base, List *c, int lm, int lc, FmgrInfo *procinfo, Oid col
if (list_length(w) <= lm)
return w;
wd = pairingheap_allocate(CompareNearestCandidates, NULL);
wd = palloc(sizeof(HnswCandidate *) * list_length(w));
/* Ensure order of candidates is deterministic for closer caching */
if (sortCandidates)
@@ -1199,21 +1080,21 @@ SelectNeighbors(char *base, List *c, int lm, int lc, FmgrInfo *procinfo, Oid col
if (e->closer)
r = lappend(r, e);
else
pairingheap_add(wd, &(CreatePairingHeapNode(e)->ph_node));
wd[wdlen++] = e;
}
/* Cached value can only be used in future if sorted deterministically */
neighbors->closerSet = sortCandidates;
/* Keep pruned connections */
while (!pairingheap_is_empty(wd) && list_length(r) < lm)
r = lappend(r, ((HnswPairingHeapNode *) pairingheap_remove_first(wd))->inner);
while (wdoff < wdlen && list_length(r) < lm)
r = lappend(r, wd[wdoff++]);
/* Return pruned for update connections */
if (pruned != NULL)
{
if (!pairingheap_is_empty(wd))
*pruned = ((HnswPairingHeapNode *) pairingheap_first(wd))->inner;
if (wdoff < wdlen)
*pruned = wd[wdoff];
else
*pruned = linitial(w);
}
@@ -1271,7 +1152,7 @@ HnswUpdateConnection(char *base, HnswElement element, HnswCandidate * hc, int lm
HnswElement hc3Element = HnswPtrAccess(base, hc3->element);
if (HnswPtrIsNull(base, hc3Element->value))
HnswLoadElement(hc3Element, &hc3->distance, &q, index, procinfo, collation, true, NULL, NULL);
HnswLoadElement(hc3Element, &hc3->distance, &q, index, procinfo, collation, true);
else
hc3->distance = GetCandidateDistance(base, hc3, q, procinfo, collation);
@@ -1393,7 +1274,7 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
/* 1st phase: greedy search to insert level */
for (int lc = entryLevel; lc >= level + 1; lc--)
{
w = HnswSearchLayer(base, q, NULL, ep, 1, lc, index, procinfo, collation, m, true, skipElement);
w = HnswSearchLayer(base, q, ep, 1, lc, index, procinfo, collation, m, true, skipElement);
ep = w;
}
@@ -1411,7 +1292,7 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
List *neighbors;
List *lw;
w = HnswSearchLayer(base, q, NULL, ep, efConstruction, lc, index, procinfo, collation, m, true, skipElement);
w = HnswSearchLayer(base, q, ep, efConstruction, lc, index, procinfo, collation, m, true, skipElement);
/* Elements being deleted or skipped can help with search */
/* but should be removed before selecting neighbors */

View File

@@ -256,7 +256,7 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
LockPage(index, HNSW_UPDATE_LOCK, ShareLock);
/* Load element */
HnswLoadElement(highestPoint, NULL, NULL, index, vacuumstate->procinfo, vacuumstate->collation, true, NULL, NULL);
HnswLoadElement(highestPoint, NULL, NULL, index, vacuumstate->procinfo, vacuumstate->collation, true);
/* Repair if needed */
if (NeedsUpdated(vacuumstate, highestPoint))
@@ -294,7 +294,7 @@ RepairGraphEntryPoint(HnswVacuumState * vacuumstate)
* is outdated, this can remove connections at higher levels in
* the graph until they are repaired, but this should be fine.
*/
HnswLoadElement(entryPoint, NULL, NULL, index, vacuumstate->procinfo, vacuumstate->collation, true, NULL, NULL);
HnswLoadElement(entryPoint, NULL, NULL, index, vacuumstate->procinfo, vacuumstate->collation, true);
if (NeedsUpdated(vacuumstate, entryPoint))
{
@@ -370,7 +370,7 @@ RepairGraph(HnswVacuumState * vacuumstate)
/* Create an element */
element = HnswInitElementFromBlock(blkno, offno);
HnswLoadElementFromTuple(element, etup, false, true, index);
HnswLoadElementFromTuple(element, etup, false, true);
elements = lappend(elements, element);
}
@@ -440,7 +440,6 @@ MarkDeleted(HnswVacuumState * vacuumstate)
BlockNumber insertPage = InvalidBlockNumber;
Relation index = vacuumstate->index;
BufferAccessStrategy bas = vacuumstate->bas;
bool useIndexTuple = IndexRelationGetNumberOfAttributes(index);
/*
* Wait for index scans to complete. Scans before this point may contain
@@ -522,18 +521,7 @@ MarkDeleted(HnswVacuumState * vacuumstate)
/* Overwrite element */
etup->deleted = 1;
if (useIndexTuple)
{
IndexTuple itup = (IndexTuple) &etup->data;
MemSet(itup, 0, IndexTupleSize(itup));
}
else
{
Vector *vec = (Vector *) (&etup->data);
MemSet(vec, 0, VARSIZE_ANY(vec));
}
MemSet(&etup->data, 0, VARSIZE_ANY(&etup->data));
/* Overwrite neighbors */
for (int i = 0; i < ntup->count; i++)

View File

@@ -5,6 +5,7 @@
#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"
@@ -532,6 +533,28 @@ 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
*/

View File

@@ -46,6 +46,30 @@ 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));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE TABLE t2 (val vector(3));
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));
\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
---------
[0,0,0]
[1,1,1]
[1,2,3]
val | val2
---------+---------
[0,0,0] | [0,0,0]
[1,1,1] | [1,1,1]
[1,2,3] | [1,2,3]
|
(4 rows)
DROP TABLE t;

View File

@@ -104,105 +104,105 @@ SELECT vector_norm('[3e37,4e37]')::real;
5e+37
(1 row)
SELECT l2_distance('[0,0]', '[3,4]');
SELECT l2_distance('[0,0]'::vector, '[3,4]');
l2_distance
-------------
5
(1 row)
SELECT l2_distance('[0,0]', '[0,1]');
SELECT l2_distance('[0,0]'::vector, '[0,1]');
l2_distance
-------------
1
(1 row)
SELECT l2_distance('[1,2]', '[3]');
SELECT l2_distance('[1,2]'::vector, '[3]');
ERROR: different vector dimensions 2 and 1
SELECT l2_distance('[3e38]', '[-3e38]');
SELECT l2_distance('[3e38]'::vector, '[-3e38]');
l2_distance
-------------
Infinity
(1 row)
SELECT inner_product('[1,2]', '[3,4]');
SELECT inner_product('[1,2]'::vector, '[3,4]');
inner_product
---------------
11
(1 row)
SELECT inner_product('[1,2]', '[3]');
SELECT inner_product('[1,2]'::vector, '[3]');
ERROR: different vector dimensions 2 and 1
SELECT inner_product('[3e38]', '[3e38]');
SELECT inner_product('[3e38]'::vector, '[3e38]');
inner_product
---------------
Infinity
(1 row)
SELECT cosine_distance('[1,2]', '[2,4]');
SELECT cosine_distance('[1,2]'::vector, '[2,4]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,2]', '[0,0]');
SELECT cosine_distance('[1,2]'::vector, '[0,0]');
cosine_distance
-----------------
NaN
(1 row)
SELECT cosine_distance('[1,1]', '[1,1]');
SELECT cosine_distance('[1,1]'::vector, '[1,1]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,0]', '[0,2]');
SELECT cosine_distance('[1,0]'::vector, '[0,2]');
cosine_distance
-----------------
1
(1 row)
SELECT cosine_distance('[1,1]', '[-1,-1]');
SELECT cosine_distance('[1,1]'::vector, '[-1,-1]');
cosine_distance
-----------------
2
(1 row)
SELECT cosine_distance('[1,2]', '[3]');
SELECT cosine_distance('[1,2]'::vector, '[3]');
ERROR: different vector dimensions 2 and 1
SELECT cosine_distance('[1,1]', '[1.1,1.1]');
SELECT cosine_distance('[1,1]'::vector, '[1.1,1.1]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,1]', '[-1.1,-1.1]');
SELECT cosine_distance('[1,1]'::vector, '[-1.1,-1.1]');
cosine_distance
-----------------
2
(1 row)
SELECT cosine_distance('[3e38]', '[3e38]');
SELECT cosine_distance('[3e38]'::vector, '[3e38]');
cosine_distance
-----------------
NaN
(1 row)
SELECT l1_distance('[0,0]', '[3,4]');
SELECT l1_distance('[0,0]'::vector, '[3,4]');
l1_distance
-------------
7
(1 row)
SELECT l1_distance('[0,0]', '[0,1]');
SELECT l1_distance('[0,0]'::vector, '[0,1]');
l1_distance
-------------
1
(1 row)
SELECT l1_distance('[1,2]', '[3]');
SELECT l1_distance('[1,2]'::vector, '[3]');
ERROR: different vector dimensions 2 and 1
SELECT l1_distance('[3e38]', '[-3e38]');
SELECT l1_distance('[3e38]'::vector, '[-3e38]');
l1_distance
-------------
Infinity

View File

@@ -0,0 +1,104 @@
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

View File

@@ -0,0 +1,147 @@
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: value out of range: overflow
LINE 1: SELECT '[65520,-65520]'::halfvec;
^
SELECT '[1e-8,-1e-8]'::halfvec;
ERROR: value out of range: underflow
LINE 1: SELECT '[1e-8,-1e-8]'::halfvec;
^
SELECT '[4e38,1]'::halfvec;
ERROR: infinite value not allowed in halfvec
LINE 1: SELECT '[4e38,1]'::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

@@ -0,0 +1,26 @@
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

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

@@ -0,0 +1,36 @@
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]';
val
---------
[1,2,3]
[1,2,4]
[1,1,1]
[0,0,0]
(4 rows)
SELECT * FROM t ORDER BY val <-> (SELECT NULL::halfvec);
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;
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
val
-----
(0 rows)
DROP TABLE t;

View File

@@ -116,8 +116,30 @@ 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

@@ -10,6 +10,12 @@ 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));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
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));
CREATE TABLE t2 (val vector(3), val2 halfvec(3));
\copy t TO 'results/data.bin' WITH (FORMAT binary)
\copy t2 FROM 'results/data.bin' WITH (FORMAT binary)

View File

@@ -24,29 +24,29 @@ SELECT vector_norm('[3,4]');
SELECT vector_norm('[0,1]');
SELECT vector_norm('[3e37,4e37]')::real;
SELECT l2_distance('[0,0]', '[3,4]');
SELECT l2_distance('[0,0]', '[0,1]');
SELECT l2_distance('[1,2]', '[3]');
SELECT l2_distance('[3e38]', '[-3e38]');
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]', '[3,4]');
SELECT inner_product('[1,2]', '[3]');
SELECT inner_product('[3e38]', '[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]', '[2,4]');
SELECT cosine_distance('[1,2]', '[0,0]');
SELECT cosine_distance('[1,1]', '[1,1]');
SELECT cosine_distance('[1,0]', '[0,2]');
SELECT cosine_distance('[1,1]', '[-1,-1]');
SELECT cosine_distance('[1,2]', '[3]');
SELECT cosine_distance('[1,1]', '[1.1,1.1]');
SELECT cosine_distance('[1,1]', '[-1.1,-1.1]');
SELECT cosine_distance('[3e38]', '[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]', '[3,4]');
SELECT l1_distance('[0,0]', '[0,1]');
SELECT l1_distance('[1,2]', '[3]');
SELECT l1_distance('[3e38]', '[-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 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;

View File

@@ -0,0 +1,23 @@
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]');

View File

@@ -0,0 +1,34 @@
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 '[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

@@ -0,0 +1,13 @@
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

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

@@ -0,0 +1,16 @@
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 * FROM t ORDER BY val <-> (SELECT NULL::halfvec);
SELECT COUNT(*) FROM t;
TRUNCATE t;
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
DROP TABLE t;

View File

@@ -22,7 +22,13 @@ 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

@@ -1,154 +0,0 @@
use strict;
use warnings;
use PostgresNode;
use TestLib;
use Test::More;
my $node;
my @queries = ();
my @cs = ();
my @expected;
my $limit = 20;
my $dim = 3;
my $array_sql = join(",", ('random()') x $dim);
my $nc = 50;
my @types = ("int4", "int8", "text", "uuid", "varchar");
my $type = $types[rand(@types)];
sub cast_c
{
my ($c) = @_;
if ($type eq "uuid")
{
return "('00000000-0000-0000-0000-0000000000' || LPAD(($c)::text, 2, '0'))::uuid";
}
else
{
return "($c)::$type";
}
}
sub test_recall
{
my ($min, $operator) = @_;
my $correct = 0;
my $total = 0;
my $explain = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
EXPLAIN ANALYZE SELECT i FROM tst WHERE c = @{[cast_c($cs[0])]} ORDER BY v $operator '$queries[0]' LIMIT $limit;
));
like($explain, qr/Index Cond/);
for my $i (0 .. $#queries)
{
my $actual = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SELECT i FROM tst WHERE c = @{[cast_c($cs[$i])]} ORDER BY v $operator '$queries[$i]' LIMIT $limit;
));
my @actual_ids = split("\n", $actual);
my %actual_set = map { $_ => 1 } @actual_ids;
is(scalar(@actual_ids), $limit);
my @expected_ids = split("\n", $expected[$i]);
foreach (@expected_ids)
{
if (exists($actual_set{$_}))
{
$correct++;
}
$total++;
}
}
cmp_ok($correct / $total, ">=", $min, $operator);
}
# Initialize node
$node = get_new_node('node');
$node->init;
$node->start;
# Create table
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v vector($dim), c $type, c2 $type);");
$node->safe_psql("postgres", qq(
INSERT INTO tst SELECT i, ARRAY[$array_sql], @{[cast_c("i % $nc")]}, @{[cast_c("i % $nc")]} FROM generate_series(1, 10000) i;
));
$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops, c, c2);");
$node->safe_psql("postgres", qq(
INSERT INTO tst SELECT i, ARRAY[$array_sql], @{[cast_c("i % $nc")]}, @{[cast_c("i % $nc")]} FROM generate_series(1, 10000) i;
));
$node->safe_psql("postgres", "ANALYZE tst;");
# Generate queries
for (1 .. 20)
{
my @r = ();
for (1 .. $dim)
{
push(@r, rand());
}
push(@queries, "[" . join(",", @r) . "]");
push(@cs, int(rand() * $nc));
}
# Get exact results
@expected = ();
for my $i (0 .. $#queries)
{
my $res = $node->safe_psql("postgres", qq(
SET enable_indexscan = off;
SELECT i FROM tst WHERE c = @{[cast_c($cs[$i])]} ORDER BY v <-> '$queries[$i]' LIMIT $limit;
));
push(@expected, $res);
}
# Test recall
test_recall(0.99, '<->');
# Test no conditions
my $explain = $node->safe_psql("postgres", qq(
EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v <-> '$queries[0]' LIMIT $limit;
));
like($explain, qr/Index Scan/);
# Test range
$explain = $node->safe_psql("postgres", qq(
EXPLAIN ANALYZE SELECT i FROM tst WHERE c >= @{[cast_c(1)]} AND c <= @{[cast_c(3)]} ORDER BY v <-> '$queries[0]' LIMIT $limit;
));
like($explain, qr/Index Cond: \(\(\S+ >= \S+\) AND \(\S+ <= \S+\)\)/);
# Test multiple conditions
$explain = $node->safe_psql("postgres", qq(
EXPLAIN ANALYZE SELECT i FROM tst WHERE c = @{[cast_c($cs[0])]} AND c2 = @{[cast_c($cs[0])]} ORDER BY v <-> '$queries[0]' LIMIT $limit;
));
like($explain, qr/Index Cond: \(\(\S+ = \S+\) AND \(\S+ = \S+\)\)/);
# Test no order
$explain = $node->safe_psql("postgres", qq(
EXPLAIN ANALYZE SELECT i FROM tst WHERE c = @{[cast_c($cs[0])]} LIMIT $limit;
));
like($explain, qr/Seq Scan/);
# Test vacuum
$node->safe_psql("postgres", "DELETE FROM tst WHERE c > @{[cast_c(5)]};");
$node->safe_psql("postgres", "VACUUM tst;");
# Test columns
my ($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (c);");
like($stderr, qr/first column must be a vector/);
($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (c, v vector_l2_ops);");
like($stderr, qr/first column must be a vector/);
($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops, c, c, c);");
like($stderr, qr/index cannot have more than three columns/);
($ret, $stdout, $stderr) = $node->psql("postgres", "CREATE INDEX ON tst USING hnsw (v vector_l2_ops, v vector_l2_ops);");
like($stderr, qr/column 2 cannot be a vector/);
done_testing();

View File

@@ -0,0 +1,132 @@
use strict;
use warnings;
use PostgresNode;
use TestLib;
use Test::More;
my $node;
my @queries = ();
my @expected;
my $limit = 20;
my $dim = 10;
my $array_sql = join(",", ('random()') x $dim);
sub test_recall
{
my ($min, $operator) = @_;
my $correct = 0;
my $total = 0;
my $explain = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v $operator '$queries[0]' LIMIT $limit;
));
like($explain, qr/Index Scan/);
for my $i (0 .. $#queries)
{
my $actual = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SELECT i FROM tst ORDER BY v $operator '$queries[$i]' LIMIT $limit;
));
my @actual_ids = split("\n", $actual);
my %actual_set = map { $_ => 1 } @actual_ids;
my @expected_ids = split("\n", $expected[$i]);
foreach (@expected_ids)
{
if (exists($actual_set{$_}))
{
$correct++;
}
$total++;
}
}
cmp_ok($correct / $total, ">=", $min, $operator);
}
# Initialize node
$node = get_new_node('node');
$node->init;
$node->start;
# Create table
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v halfvec($dim));");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, 10000) i;"
);
# Generate queries
for (1 .. 20)
{
my @r = ();
for (1 .. $dim)
{
push(@r, rand());
}
push(@queries, "[" . join(",", @r) . "]");
}
# Check each index type
my @operators = ("<->", "<#>", "<=>");
my @opclasses = ("halfvec_l2_ops", "halfvec_ip_ops", "halfvec_cosine_ops");
for my $i (0 .. $#operators)
{
my $operator = $operators[$i];
my $opclass = $opclasses[$i];
# Get exact results
@expected = ();
foreach (@queries)
{
my $res = $node->safe_psql("postgres", "SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;");
push(@expected, $res);
}
# Build index serially
$node->safe_psql("postgres", qq(
SET max_parallel_maintenance_workers = 0;
CREATE INDEX idx ON tst USING hnsw (v $opclass);
));
# Test approximate results
my $min = $operator eq "<#>" ? 0.95 : 0.99;
test_recall($min, $operator);
$node->safe_psql("postgres", "DROP INDEX idx;");
# Build index in parallel in memory
my ($ret, $stdout, $stderr) = $node->psql("postgres", qq(
SET client_min_messages = DEBUG;
SET min_parallel_table_scan_size = 1;
CREATE INDEX idx ON tst USING hnsw (v $opclass);
));
is($ret, 0, $stderr);
like($stderr, qr/using \d+ parallel workers/);
# Test approximate results
test_recall($min, $operator);
$node->safe_psql("postgres", "DROP INDEX idx;");
# Build index in parallel on disk
# Set parallel_workers on table to use workers with low maintenance_work_mem
($ret, $stdout, $stderr) = $node->psql("postgres", qq(
ALTER TABLE tst SET (parallel_workers = 2);
SET client_min_messages = DEBUG;
SET maintenance_work_mem = '4MB';
CREATE INDEX idx ON tst USING hnsw (v $opclass);
ALTER TABLE tst RESET (parallel_workers);
));
is($ret, 0, $stderr);
like($stderr, qr/using \d+ parallel workers/);
like($stderr, qr/hnsw graph no longer fits into maintenance_work_mem/);
$node->safe_psql("postgres", "DROP INDEX idx;");
}
done_testing();

View File

@@ -1,4 +1,4 @@
comment = 'vector data type and ivfflat and hnsw access methods'
default_version = '0.6.0'
default_version = '0.6.2'
module_pathname = '$libdir/vector'
relocatable = true