Compare commits

...

53 Commits

Author SHA1 Message Date
Andrew Kane
421116c8cd Improved name [skip ci] 2024-09-24 11:48:13 -07:00
Andrew Kane
3623c64d82 Added comment [skip ci] 2024-09-24 11:46:13 -07:00
Andrew Kane
18be1fda24 Fixed lookup table [skip ci] 2024-09-24 01:25:20 -07:00
Andrew Kane
e10a5c90df Fixed code [skip ci] 2024-09-24 00:36:23 -07:00
Andrew Kane
3b0976b725 Fixed rounding [skip ci] 2024-09-24 00:24:02 -07:00
Andrew Kane
724bde5020 Improved code [skip ci] 2024-09-23 22:07:30 -07:00
Andrew Kane
261ba4a0d9 Added test for IVFFlat [skip ci] 2024-09-23 18:57:13 -07:00
Andrew Kane
4293bc439e Exclude zero vectors for cosine distance to be consistent with other types [skip ci] 2024-09-23 18:55:10 -07:00
Andrew Kane
044f0c6441 Improved precision for cosine distance [skip ci] 2024-09-23 18:53:23 -07:00
Andrew Kane
22efa8ec22 Improved precision for cosine distance [skip ci] 2024-09-23 18:50:10 -07:00
Andrew Kane
0a2803130f Added support for IVFFlat [skip ci] 2024-09-23 18:40:49 -07:00
Andrew Kane
6105a01881 Added more tests [skip ci] 2024-09-23 18:33:08 -07:00
Andrew Kane
9925b92d73 Added more minivec tests [skip ci] 2024-09-23 18:29:17 -07:00
Andrew Kane
e49bae9dc3 Fixed CI 2024-09-23 18:24:21 -07:00
Andrew Kane
eea2c44fae Added minivec type 2024-09-23 18:12:26 -07:00
Andrew Kane
87ac108bf7 Removed code for Postgres 12 [skip ci] 2024-09-23 15:26:31 -07:00
Andrew Kane
97cf990e0f Free TupleDesc [skip ci] 2024-09-21 19:15:34 -07:00
Andrew Kane
55dc735e1a Moved allocations out of GetScanItems [skip ci] 2024-09-21 19:10:25 -07:00
Andrew Kane
be4e9a9df2 Added macros for IvfflatScanList [skip ci] 2024-09-21 18:10:37 -07:00
Andrew Kane
d5e8fc96a5 Changed HnswPairingHeapNode to HnswSearchCandidate to reduce allocations and improve code 2024-09-21 12:07:44 -07:00
Andrew Kane
6d2af6d3f9 Improved code [skip ci] 2024-09-20 15:21:57 -07:00
Andrew Kane
a6ab5d07c0 Fixed CI 2024-09-19 20:50:51 -07:00
Andrew Kane
aa77346103 Improved code [skip ci] 2024-09-19 19:57:16 -07:00
Andrew Kane
b0da2d95d9 Fixed array_to_sparsevec on Windows [skip ci] 2024-09-19 19:52:16 -07:00
Andrew Kane
3fb05eb847 Added casts for arrays to sparsevec - #604
Co-authored-by: Narek Galstyan <narekg@berkeley.edu>
Co-authored-by: Di Qi <di@lantern.dev>
2024-09-19 19:17:05 -07:00
Andrew Kane
b738ffecc1 Dropped support for Postgres 12 2024-09-19 18:13:54 -07:00
Heikki Linnakangas
7117513532 Add error codes to a few errors (#657)
With elog(), you get XX000 "internal_error", which sounds scary.

It's not self-evident what the right error codes for some of these
errors are, but I tried to use my best judgment.
2024-09-19 18:04:23 -07:00
Andrew Kane
85d877d540 Updated changelog [skip ci] 2024-09-19 18:03:20 -07:00
Jonathan S. Katz
05fb382031 Swap max costing values to align with upstream guidance (#658)
A feature targeted for PostgreSQL 18 (postgres/postgres@e2225346)
that makes optimizations around disabled path nodes impacted pgvector
such that PostgreSQL would choose to perform an index scan when it
should have used a different scan (e.g. `SELECT count(*) FROM table`).
Per upstream guidance[1], the recommendation is to switch to using
`get_float8_infinity()`, which achieves the same behavior in backbranches,
and can be adapated to work with the new behavior introduced in PostgreSQL 18.

[1] https://www.postgresql.org/message-id/2281822.1724441531%40sss.pgh.pa.us
2024-09-19 18:01:59 -07:00
Andrew Kane
8e1853fbf3 Improved variable name [skip ci] 2024-09-19 15:09:40 -07:00
Andrew Kane
f9d68a061a Simplified HnswLoadUnvisitedFromMemory [skip ci] 2024-09-19 04:39:46 -07:00
Andrew Kane
4f8ab574c9 Simplified CountElement [skip ci] 2024-09-19 04:32:38 -07:00
Andrew Kane
a15806196e Keep scan-build happy 2024-09-19 04:02:09 -07:00
Andrew Kane
5c9429a0f8 Reduced memory usage for HNSW index scans 2024-09-19 03:27:35 -07:00
Andrew Kane
4b44d6e745 Updated changelog [skip ci] 2024-09-19 02:42:33 -07:00
Andrew Kane
16ca608f42 Updated AddToVisited to use HnswElementPtr 2024-09-19 02:41:20 -07:00
Andrew Kane
8dde14a736 Reduced memory usage for HNSW index scans
Co-authored-by: Heikki Linnakangas <heikki.linnakangas@iki.fi>
2024-09-19 02:17:51 -07:00
Andrew Kane
d74d3065bc Reduced allocations for pairing heap 2024-09-19 01:59:46 -07:00
Andrew Kane
a1b80faa67 Updated readme 2024-09-05 23:13:12 -07:00
Andrew Kane
4af5a127e0 Revert "Improved cleanup for IVFFlat index scans [skip ci]"
This reverts commit da7d3959a3.
2024-09-02 01:52:28 -07:00
Andrew Kane
d02d71a398 Fixed CI 2024-09-02 01:44:03 -07:00
Andrew Kane
2aca04b8de Updated links [skip ci] 2024-08-28 13:39:03 -07:00
Andrew Kane
e47984e616 Reset tuple sort for Postgres 12 [skip ci] 2024-08-24 22:10:26 -07:00
Andrew Kane
da7d3959a3 Improved cleanup for IVFFlat index scans [skip ci] 2024-08-24 21:59:44 -07:00
Andrew Kane
dadbbc3758 Renamed InitSortState to InitScanSortState [skip ci] 2024-08-24 21:53:15 -07:00
Andrew Kane
6af0a43d62 Added InitBuildSortState function [skip ci] 2024-08-24 21:50:31 -07:00
Andrew Kane
ffcb90d094 Added InitSortState function [skip ci] 2024-08-24 21:42:18 -07:00
Andrew Kane
8a312c3c8e Added memory usage for IVFFlat index scans [skip ci] 2024-08-24 21:30:40 -07:00
Andrew Kane
5d86b177ab Fixed -DIVFFLAT_MEMORY [skip ci] 2024-08-24 20:56:33 -07:00
Andrew Kane
ea99957fae Added fields to IndexAmRoutine 2024-08-22 20:39:16 -07:00
Samuel Marks
4cede1a9c9 [src/hnswutils.c] Resolve 1 -Wmaybe-uninitialized (#654) 2024-08-22 19:51:16 -07:00
Andrew Kane
d0dbc8b4d1 Added Postgres 18 to CI [skip ci] 2024-08-13 02:24:42 -07:00
Andrew Kane
bb855e6cb4 Updated comment [skip ci] 2024-08-06 10:35:26 -07:00
43 changed files with 3809 additions and 368 deletions

View File

@@ -8,17 +8,17 @@ jobs:
fail-fast: false
matrix:
include:
# - postgres: 18
# os: ubuntu-24.04
- postgres: 17
os: ubuntu-24.04
- postgres: 16
os: ubuntu-24.04
os: ubuntu-22.04
- postgres: 15
os: ubuntu-22.04
- postgres: 14
os: ubuntu-22.04
- postgres: 13
os: ubuntu-20.04
- postgres: 12
- postgres: 13
os: ubuntu-20.04
steps:
- uses: actions/checkout@v4

View File

@@ -1,3 +1,9 @@
## 0.8.0 (unreleased)
- Added casts for arrays to `sparsevec`
- Reduced memory usage for HNSW index scans
- Dropped support for Postgres 12
## 0.7.4 (2024-08-05)
- Fixed locking for parallel HNSW index builds

View File

@@ -4,8 +4,8 @@ EXTVERSION = 0.7.4
MODULE_big = vector
DATA = $(wildcard sql/*--*--*.sql)
DATA_built = sql/$(EXTENSION)--$(EXTVERSION).sql
OBJS = src/bitutils.o src/bitvec.o src/halfutils.o src/halfvec.o src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/ivfvacuum.o src/sparsevec.o src/vector.o
HEADERS = src/halfvec.h src/sparsevec.h src/vector.h
OBJS = src/bitutils.o src/bitvec.o src/halfutils.o src/halfvec.o src/hnsw.o src/hnswbuild.o src/hnswinsert.o src/hnswscan.o src/hnswutils.o src/hnswvacuum.o src/ivfbuild.o src/ivfflat.o src/ivfinsert.o src/ivfkmeans.o src/ivfscan.o src/ivfutils.o src/minivec.o src/ivfvacuum.o src/sparsevec.o src/vector.o
HEADERS = src/halfvec.h src/minivec.h src/sparsevec.h src/vector.h
TESTS = $(wildcard test/sql/*.sql)
REGRESS = $(patsubst test/sql/%.sql,%,$(TESTS))

View File

@@ -2,10 +2,10 @@ EXTENSION = vector
EXTVERSION = 0.7.4
DATA_built = sql\$(EXTENSION)--$(EXTVERSION).sql
OBJS = src\bitutils.obj src\bitvec.obj src\halfutils.obj src\halfvec.obj src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\sparsevec.obj src\vector.obj
HEADERS = src\halfvec.h src\sparsevec.h src\vector.h
OBJS = src\bitutils.obj src\bitvec.obj src\halfutils.obj src\halfvec.obj src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\minivec.obj src\sparsevec.obj src\vector.obj
HEADERS = src\halfvec.h src\minivec.h src\sparsevec.h src\vector.h
REGRESS = bit btree cast copy halfvec hnsw_bit hnsw_halfvec hnsw_sparsevec hnsw_vector ivfflat_bit ivfflat_halfvec ivfflat_vector sparsevec vector_type
REGRESS = bit btree cast copy halfvec hnsw_bit hnsw_halfvec hnsw_sparsevec hnsw_vector ivfflat_bit ivfflat_halfvec ivfflat_vector minivec sparsevec vector_type
REGRESS_OPTS = --inputdir=test --load-extension=$(EXTENSION)
# For /arch flags

View File

@@ -106,7 +106,7 @@ 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))
Or load vectors in bulk using `COPY` ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/loading/example.py))
```sql
COPY items (embedding) FROM STDIN WITH (FORMAT BINARY);
@@ -202,7 +202,7 @@ You can add an index to use approximate nearest neighbor search, which trades so
Supported index types are:
- [HNSW](#hnsw) - added in 0.5.0
- [HNSW](#hnsw)
- [IVFFlat](#ivfflat)
## HNSW
@@ -473,7 +473,7 @@ SELECT * FROM items ORDER BY embedding::halfvec(3) <-> '[1,2,3]' LIMIT 5;
## Binary Vectors
Use the `bit` type to store binary vectors ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/hash_image_search.py))
Use the `bit` type to store binary vectors ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/imagehash/example.py))
```sql
CREATE TABLE items (id bigserial PRIMARY KEY, embedding bit(3));
@@ -551,7 +551,7 @@ SELECT id, content FROM items, plainto_tsquery('hello search') query
WHERE textsearch @@ query ORDER BY ts_rank_cd(textsearch, query) DESC LIMIT 5;
```
You can use [Reciprocal Rank Fusion](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search_rrf.py) or a [cross-encoder](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search.py) to combine results.
You can use [Reciprocal Rank Fusion](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search/rrf.py) or a [cross-encoder](https://github.com/pgvector/pgvector-python/blob/master/examples/hybrid_search/cross_encoder.py) to combine results.
## Indexing Subvectors
@@ -597,7 +597,7 @@ Be sure to restart Postgres for changes to take effect.
### Loading
Use `COPY` for bulk loading data ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/bulk_loading.py)).
Use `COPY` for bulk loading data ([example](https://github.com/pgvector/pgvector-python/blob/master/examples/loading/example.py)).
```sql
COPY items (embedding) FROM STDIN WITH (FORMAT BINARY);
@@ -687,7 +687,7 @@ 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)).
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/example.py)).
## Languages
@@ -934,6 +934,37 @@ Function | Description | Added
avg(halfvec) → halfvec | average | 0.7.0
sum(halfvec) → halfvec | sum | 0.7.0
### Minivec Type
Each mini vector takes `dimensions + 8` bytes of storage. Each element is a E4M3 8-bit floating-point number, and all elements must be finite (no `NaN`). Mini vectors can have up to 16,000 dimensions.
### Minivec Operators
Operator | Description | Added
--- | --- | ---
\+ | element-wise addition | 0.8.0
\- | element-wise subtraction | 0.8.0
\* | element-wise multiplication | 0.8.0
\|\| | concatenate | 0.8.0
<-> | Euclidean distance | 0.8.0
<#> | negative inner product | 0.8.0
<=> | cosine distance | 0.8.0
<+> | taxicab distance | 0.8.0
### Minivec Functions
Function | Description | Added
--- | --- | ---
binary_quantize(minivec) → bit | binary quantize | 0.8.0
cosine_distance(minivec, minivec) → double precision | cosine distance | 0.8.0
inner_product(minivec, minivec) → double precision | inner product | 0.8.0
l1_distance(minivec, minivec) → double precision | taxicab distance | 0.8.0
l2_distance(minivec, minivec) → double precision | Euclidean distance | 0.8.0
l2_norm(minivec) → double precision | Euclidean norm | 0.8.0
l2_normalize(minivec) → minivec | Normalize with Euclidean norm | 0.8.0
subvector(minivec, integer, integer) → minivec | subvector | 0.8.0
vector_dims(minivec) → integer | number of dimensions | 0.8.0
### Bit Type
Each bit vector takes `dimensions / 8 + 8` bytes of storage. See the [Postgres docs](https://www.postgresql.org/docs/current/datatype-bit.html) for more info.

View File

@@ -0,0 +1,28 @@
-- complain if script is sourced in psql, rather than via CREATE EXTENSION
\echo Use "ALTER EXTENSION vector UPDATE TO '0.8.0'" to load this file. \quit
-- TODO minivec functions
CREATE FUNCTION array_to_sparsevec(integer[], integer, boolean) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_sparsevec(real[], integer, boolean) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_sparsevec(double precision[], integer, boolean) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_sparsevec(numeric[], integer, boolean) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE CAST (integer[] AS sparsevec)
WITH FUNCTION array_to_sparsevec(integer[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (real[] AS sparsevec)
WITH FUNCTION array_to_sparsevec(real[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (double precision[] AS sparsevec)
WITH FUNCTION array_to_sparsevec(double precision[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (numeric[] AS sparsevec)
WITH FUNCTION array_to_sparsevec(numeric[], integer, boolean) AS ASSIGNMENT;

View File

@@ -266,12 +266,18 @@ COMMENT ON ACCESS METHOD hnsw IS 'hnsw index access method';
CREATE FUNCTION ivfflat_halfvec_support(internal) RETURNS internal
AS 'MODULE_PATHNAME' LANGUAGE C;
CREATE FUNCTION ivfflat_minivec_support(internal) RETURNS internal
AS 'MODULE_PATHNAME' LANGUAGE C;
CREATE FUNCTION ivfflat_bit_support(internal) RETURNS internal
AS 'MODULE_PATHNAME' LANGUAGE C;
CREATE FUNCTION hnsw_halfvec_support(internal) RETURNS internal
AS 'MODULE_PATHNAME' LANGUAGE C;
CREATE FUNCTION hnsw_minivec_support(internal) RETURNS internal
AS 'MODULE_PATHNAME' LANGUAGE C;
CREATE FUNCTION hnsw_bit_support(internal) RETURNS internal
AS 'MODULE_PATHNAME' LANGUAGE C;
@@ -647,6 +653,295 @@ CREATE OPERATOR CLASS halfvec_l1_ops
FUNCTION 1 l1_distance(halfvec, halfvec),
FUNCTION 3 hnsw_halfvec_support(internal);
-- minivec type
CREATE TYPE minivec;
CREATE FUNCTION minivec_in(cstring, oid, integer) RETURNS minivec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_out(minivec) RETURNS cstring
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_typmod_in(cstring[]) RETURNS integer
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_recv(internal, oid, integer) RETURNS minivec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_send(minivec) RETURNS bytea
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE TYPE minivec (
INPUT = minivec_in,
OUTPUT = minivec_out,
TYPMOD_IN = minivec_typmod_in,
RECEIVE = minivec_recv,
SEND = minivec_send,
STORAGE = external
);
-- minivec functions
CREATE FUNCTION l2_distance(minivec, minivec) RETURNS float8
AS 'MODULE_PATHNAME', 'minivec_l2_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION inner_product(minivec, minivec) RETURNS float8
AS 'MODULE_PATHNAME', 'minivec_inner_product' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION cosine_distance(minivec, minivec) RETURNS float8
AS 'MODULE_PATHNAME', 'minivec_cosine_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION l1_distance(minivec, minivec) RETURNS float8
AS 'MODULE_PATHNAME', 'minivec_l1_distance' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION vector_dims(minivec) RETURNS integer
AS 'MODULE_PATHNAME', 'minivec_vector_dims' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION l2_norm(minivec) RETURNS float8
AS 'MODULE_PATHNAME', 'minivec_l2_norm' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION l2_normalize(minivec) RETURNS minivec
AS 'MODULE_PATHNAME', 'minivec_l2_normalize' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION binary_quantize(minivec) RETURNS bit
AS 'MODULE_PATHNAME', 'minivec_binary_quantize' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION subvector(minivec, int, int) RETURNS minivec
AS 'MODULE_PATHNAME', 'minivec_subvector' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- minivec private functions
CREATE FUNCTION minivec_add(minivec, minivec) RETURNS minivec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_sub(minivec, minivec) RETURNS minivec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_mul(minivec, minivec) RETURNS minivec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_concat(minivec, minivec) RETURNS minivec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_lt(minivec, minivec) RETURNS bool
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_le(minivec, minivec) RETURNS bool
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_eq(minivec, minivec) RETURNS bool
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_ne(minivec, minivec) RETURNS bool
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_ge(minivec, minivec) RETURNS bool
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_gt(minivec, minivec) RETURNS bool
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_cmp(minivec, minivec) RETURNS int4
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_l2_squared_distance(minivec, minivec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_negative_inner_product(minivec, minivec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_spherical_distance(minivec, minivec) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- minivec cast functions
CREATE FUNCTION minivec(minivec, integer, boolean) RETURNS minivec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_to_vector(minivec, integer, boolean) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION vector_to_minivec(vector, integer, boolean) RETURNS minivec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_minivec(integer[], integer, boolean) RETURNS minivec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_minivec(real[], integer, boolean) RETURNS minivec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_minivec(double precision[], integer, boolean) RETURNS minivec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_minivec(numeric[], integer, boolean) RETURNS minivec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION minivec_to_float4(minivec, integer, boolean) RETURNS real[]
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- minivec casts
CREATE CAST (minivec AS minivec)
WITH FUNCTION minivec(minivec, integer, boolean) AS IMPLICIT;
CREATE CAST (minivec AS vector)
WITH FUNCTION minivec_to_vector(minivec, integer, boolean) AS ASSIGNMENT;
CREATE CAST (vector AS minivec)
WITH FUNCTION vector_to_minivec(vector, integer, boolean) AS IMPLICIT;
CREATE CAST (minivec AS real[])
WITH FUNCTION minivec_to_float4(minivec, integer, boolean) AS ASSIGNMENT;
CREATE CAST (integer[] AS minivec)
WITH FUNCTION array_to_minivec(integer[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (real[] AS minivec)
WITH FUNCTION array_to_minivec(real[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (double precision[] AS minivec)
WITH FUNCTION array_to_minivec(double precision[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (numeric[] AS minivec)
WITH FUNCTION array_to_minivec(numeric[], integer, boolean) AS ASSIGNMENT;
-- minivec operators
CREATE OPERATOR <-> (
LEFTARG = minivec, RIGHTARG = minivec, PROCEDURE = l2_distance,
COMMUTATOR = '<->'
);
CREATE OPERATOR <#> (
LEFTARG = minivec, RIGHTARG = minivec, PROCEDURE = minivec_negative_inner_product,
COMMUTATOR = '<#>'
);
CREATE OPERATOR <=> (
LEFTARG = minivec, RIGHTARG = minivec, PROCEDURE = cosine_distance,
COMMUTATOR = '<=>'
);
CREATE OPERATOR <+> (
LEFTARG = minivec, RIGHTARG = minivec, PROCEDURE = l1_distance,
COMMUTATOR = '<+>'
);
CREATE OPERATOR + (
LEFTARG = minivec, RIGHTARG = minivec, PROCEDURE = minivec_add,
COMMUTATOR = +
);
CREATE OPERATOR - (
LEFTARG = minivec, RIGHTARG = minivec, PROCEDURE = minivec_sub
);
CREATE OPERATOR * (
LEFTARG = minivec, RIGHTARG = minivec, PROCEDURE = minivec_mul,
COMMUTATOR = *
);
CREATE OPERATOR || (
LEFTARG = minivec, RIGHTARG = minivec, PROCEDURE = minivec_concat
);
CREATE OPERATOR < (
LEFTARG = minivec, RIGHTARG = minivec, PROCEDURE = minivec_lt,
COMMUTATOR = > , NEGATOR = >= ,
RESTRICT = scalarltsel, JOIN = scalarltjoinsel
);
CREATE OPERATOR <= (
LEFTARG = minivec, RIGHTARG = minivec, PROCEDURE = minivec_le,
COMMUTATOR = >= , NEGATOR = > ,
RESTRICT = scalarlesel, JOIN = scalarlejoinsel
);
CREATE OPERATOR = (
LEFTARG = minivec, RIGHTARG = minivec, PROCEDURE = minivec_eq,
COMMUTATOR = = , NEGATOR = <> ,
RESTRICT = eqsel, JOIN = eqjoinsel
);
CREATE OPERATOR <> (
LEFTARG = minivec, RIGHTARG = minivec, PROCEDURE = minivec_ne,
COMMUTATOR = <> , NEGATOR = = ,
RESTRICT = eqsel, JOIN = eqjoinsel
);
CREATE OPERATOR >= (
LEFTARG = minivec, RIGHTARG = minivec, PROCEDURE = minivec_ge,
COMMUTATOR = <= , NEGATOR = < ,
RESTRICT = scalargesel, JOIN = scalargejoinsel
);
CREATE OPERATOR > (
LEFTARG = minivec, RIGHTARG = minivec, PROCEDURE = minivec_gt,
COMMUTATOR = < , NEGATOR = <= ,
RESTRICT = scalargtsel, JOIN = scalargtjoinsel
);
-- minivec op classes
CREATE OPERATOR CLASS minivec_ops
DEFAULT FOR TYPE minivec USING btree AS
OPERATOR 1 < ,
OPERATOR 2 <= ,
OPERATOR 3 = ,
OPERATOR 4 >= ,
OPERATOR 5 > ,
FUNCTION 1 minivec_cmp(minivec, minivec);
CREATE OPERATOR CLASS minivec_l2_ops
FOR TYPE minivec USING ivfflat AS
OPERATOR 1 <-> (minivec, minivec) FOR ORDER BY float_ops,
FUNCTION 1 minivec_l2_squared_distance(minivec, minivec),
FUNCTION 3 l2_distance(minivec, minivec),
FUNCTION 5 ivfflat_minivec_support(internal);
CREATE OPERATOR CLASS minivec_ip_ops
FOR TYPE minivec USING ivfflat AS
OPERATOR 1 <#> (minivec, minivec) FOR ORDER BY float_ops,
FUNCTION 1 minivec_negative_inner_product(minivec, minivec),
FUNCTION 3 minivec_spherical_distance(minivec, minivec),
FUNCTION 4 l2_norm(minivec),
FUNCTION 5 ivfflat_minivec_support(internal);
CREATE OPERATOR CLASS minivec_cosine_ops
FOR TYPE minivec USING ivfflat AS
OPERATOR 1 <=> (minivec, minivec) FOR ORDER BY float_ops,
FUNCTION 1 cosine_distance(minivec, minivec),
FUNCTION 2 l2_norm(minivec),
FUNCTION 3 minivec_spherical_distance(minivec, minivec),
FUNCTION 4 l2_norm(minivec),
FUNCTION 5 ivfflat_minivec_support(internal);
CREATE OPERATOR CLASS minivec_l2_ops
FOR TYPE minivec USING hnsw AS
OPERATOR 1 <-> (minivec, minivec) FOR ORDER BY float_ops,
FUNCTION 1 minivec_l2_squared_distance(minivec, minivec),
FUNCTION 3 hnsw_minivec_support(internal);
CREATE OPERATOR CLASS minivec_ip_ops
FOR TYPE minivec USING hnsw AS
OPERATOR 1 <#> (minivec, minivec) FOR ORDER BY float_ops,
FUNCTION 1 minivec_negative_inner_product(minivec, minivec),
FUNCTION 3 hnsw_minivec_support(internal);
CREATE OPERATOR CLASS minivec_cosine_ops
FOR TYPE minivec USING hnsw AS
OPERATOR 1 <=> (minivec, minivec) FOR ORDER BY float_ops,
FUNCTION 1 cosine_distance(minivec, minivec),
FUNCTION 2 l2_norm(minivec),
FUNCTION 3 hnsw_minivec_support(internal);
CREATE OPERATOR CLASS minivec_l1_ops
FOR TYPE minivec USING hnsw AS
OPERATOR 1 <+> (minivec, minivec) FOR ORDER BY float_ops,
FUNCTION 1 l1_distance(minivec, minivec),
FUNCTION 3 hnsw_minivec_support(internal);
-- bit functions
CREATE FUNCTION hamming_distance(bit, bit) RETURNS float8
@@ -782,6 +1077,18 @@ CREATE FUNCTION halfvec_to_sparsevec(halfvec, integer, boolean) RETURNS sparseve
CREATE FUNCTION sparsevec_to_halfvec(sparsevec, integer, boolean) RETURNS halfvec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_sparsevec(integer[], integer, boolean) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_sparsevec(real[], integer, boolean) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_sparsevec(double precision[], integer, boolean) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION array_to_sparsevec(numeric[], integer, boolean) RETURNS sparsevec
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- sparsevec casts
CREATE CAST (sparsevec AS sparsevec)
@@ -799,6 +1106,18 @@ CREATE CAST (sparsevec AS halfvec)
CREATE CAST (halfvec AS sparsevec)
WITH FUNCTION halfvec_to_sparsevec(halfvec, integer, boolean) AS IMPLICIT;
CREATE CAST (integer[] AS sparsevec)
WITH FUNCTION array_to_sparsevec(integer[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (real[] AS sparsevec)
WITH FUNCTION array_to_sparsevec(real[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (double precision[] AS sparsevec)
WITH FUNCTION array_to_sparsevec(double precision[], integer, boolean) AS ASSIGNMENT;
CREATE CAST (numeric[] AS sparsevec)
WITH FUNCTION array_to_sparsevec(numeric[], integer, boolean) AS ASSIGNMENT;
-- sparsevec operators
CREATE OPERATOR <-> (

View File

@@ -4,8 +4,8 @@
#include "postgres.h"
/* Check version in first header */
#if PG_VERSION_NUM < 120000
#error "Requires PostgreSQL 12+"
#if PG_VERSION_NUM < 130000
#error "Requires PostgreSQL 13+"
#endif
extern uint64 (*BitHammingDistance) (uint32 bytes, unsigned char *ax, unsigned char *bx, uint64 distance);

View File

@@ -19,11 +19,6 @@
#include "utils/numeric.h"
#include "vector.h"
#if PG_VERSION_NUM < 130000
#define TYPALIGN_DOUBLE 'd'
#define TYPALIGN_INT 'i'
#endif
#define STATE_DIMS(x) (ARR_DIMS(x)[0] - 1)
#define CreateStateDatums(dim) palloc(sizeof(Datum) * (dim + 1))
@@ -164,24 +159,6 @@ CheckStateArray(ArrayType *statearray, const char *caller)
return (float8 *) ARR_DATA_PTR(statearray);
}
#if PG_VERSION_NUM < 120003
static pg_noinline void
float_overflow_error(void)
{
ereport(ERROR,
(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
errmsg("value out of range: overflow")));
}
static pg_noinline void
float_underflow_error(void)
{
ereport(ERROR,
(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
errmsg("value out of range: underflow")));
}
#endif
/*
* Convert textual representation to internal representation
*/

View File

@@ -9,6 +9,7 @@
#include "commands/vacuum.h"
#include "hnsw.h"
#include "miscadmin.h"
#include "utils/float.h"
#include "utils/guc.h"
#include "utils/selfuncs.h"
@@ -59,17 +60,9 @@ HnswInit(void)
hnsw_relopt_kind = add_reloption_kind();
add_int_reloption(hnsw_relopt_kind, "m", "Max number of connections",
HNSW_DEFAULT_M, HNSW_MIN_M, HNSW_MAX_M
#if PG_VERSION_NUM >= 130000
,AccessExclusiveLock
#endif
);
HNSW_DEFAULT_M, HNSW_MIN_M, HNSW_MAX_M, AccessExclusiveLock);
add_int_reloption(hnsw_relopt_kind, "ef_construction", "Size of the dynamic candidate list for construction",
HNSW_DEFAULT_EF_CONSTRUCTION, HNSW_MIN_EF_CONSTRUCTION, HNSW_MAX_EF_CONSTRUCTION
#if PG_VERSION_NUM >= 130000
,AccessExclusiveLock
#endif
);
HNSW_DEFAULT_EF_CONSTRUCTION, HNSW_MIN_EF_CONSTRUCTION, HNSW_MAX_EF_CONSTRUCTION, AccessExclusiveLock);
DefineCustomIntVariable("hnsw.ef_search", "Sets the size of the dynamic candidate list for search",
"Valid range is 1..1000.", &hnsw_ef_search,
@@ -112,8 +105,8 @@ hnswcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
/* Never use index without order */
if (path->indexorderbys == NULL)
{
*indexStartupCost = DBL_MAX;
*indexTotalCost = DBL_MAX;
*indexStartupCost = get_float8_infinity();
*indexTotalCost = get_float8_infinity();
*indexSelectivity = 0;
*indexCorrelation = 0;
*indexPages = 0;
@@ -154,23 +147,10 @@ hnswoptions(Datum reloptions, bool validate)
{"ef_construction", RELOPT_TYPE_INT, offsetof(HnswOptions, efConstruction)},
};
#if PG_VERSION_NUM >= 130000
return (bytea *) build_reloptions(reloptions, validate,
hnsw_relopt_kind,
sizeof(HnswOptions),
tab, lengthof(tab));
#else
relopt_value *options;
int numoptions;
HnswOptions *rdopts;
options = parseRelOptions(reloptions, validate, hnsw_relopt_kind, &numoptions);
rdopts = allocateReloptStruct(sizeof(HnswOptions), options, numoptions);
fillRelOptions((void *) rdopts, sizeof(HnswOptions), options, numoptions,
validate, tab, lengthof(tab));
return (bytea *) rdopts;
#endif
}
/*
@@ -195,9 +175,7 @@ hnswhandler(PG_FUNCTION_ARGS)
amroutine->amstrategies = 0;
amroutine->amsupport = 3;
#if PG_VERSION_NUM >= 130000
amroutine->amoptsprocnum = 0;
#endif
amroutine->amcanorder = false;
amroutine->amcanorderbyop = true;
amroutine->amcanbackward = false; /* can change direction mid-scan */
@@ -210,17 +188,24 @@ hnswhandler(PG_FUNCTION_ARGS)
amroutine->amclusterable = false;
amroutine->ampredlocks = false;
amroutine->amcanparallel = false;
amroutine->amcaninclude = false;
#if PG_VERSION_NUM >= 130000
amroutine->amusemaintenanceworkmem = false; /* not used during VACUUM */
amroutine->amparallelvacuumoptions = VACUUM_OPTION_PARALLEL_BULKDEL;
#if PG_VERSION_NUM >= 170000
amroutine->amcanbuildparallel = true;
#endif
amroutine->amcaninclude = false;
amroutine->amusemaintenanceworkmem = false; /* not used during VACUUM */
#if PG_VERSION_NUM >= 160000
amroutine->amsummarizing = false;
#endif
amroutine->amparallelvacuumoptions = VACUUM_OPTION_PARALLEL_BULKDEL;
amroutine->amkeytype = InvalidOid;
/* Interface functions */
amroutine->ambuild = hnswbuild;
amroutine->ambuildempty = hnswbuildempty;
amroutine->aminsert = hnswinsert;
#if PG_VERSION_NUM >= 170000
amroutine->aminsertcleanup = NULL;
#endif
amroutine->ambulkdelete = hnswbulkdelete;
amroutine->amvacuumcleanup = hnswvacuumcleanup;
amroutine->amcanreturn = NULL;

View File

@@ -76,11 +76,6 @@
#define SeedRandom(seed) srandom(seed)
#endif
#if PG_VERSION_NUM < 130000
#define list_delete_last(list) list_truncate(list, list_length(list) - 1)
#define list_sort(list, cmp) ((list) = list_qsort(list, cmp))
#endif
#define HnswIsElementTuple(tup) ((tup)->type == HNSW_ELEMENT_TUPLE_TYPE)
#define HnswIsNeighborTuple(tup) ((tup)->type == HNSW_NEIGHBOR_TUPLE_TYPE)
@@ -160,11 +155,13 @@ struct HnswNeighborArray
HnswCandidate items[FLEXIBLE_ARRAY_MEMBER];
};
typedef struct HnswPairingHeapNode
typedef struct HnswSearchCandidate
{
pairingheap_node ph_node;
HnswCandidate *inner;
} HnswPairingHeapNode;
pairingheap_node c_node;
pairingheap_node w_node;
HnswElementPtr element;
float distance;
} HnswSearchCandidate;
/* HNSW index options */
typedef struct HnswOptions
@@ -385,7 +382,7 @@ void *HnswAlloc(HnswAllocator * allocator, Size size);
HnswElement HnswInitElement(char *base, ItemPointer tid, int m, double ml, int maxLevel, HnswAllocator * alloc);
HnswElement HnswInitElementFromBlock(BlockNumber blkno, OffsetNumber offno);
void HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint, Relation index, FmgrInfo *procinfo, Oid collation, int m, int efConstruction, bool existing);
HnswCandidate *HnswEntryCandidate(char *base, HnswElement em, Datum q, Relation rel, FmgrInfo *procinfo, Oid collation, bool loadVec);
HnswSearchCandidate *HnswEntryCandidate(char *base, HnswElement em, Datum q, Relation rel, FmgrInfo *procinfo, Oid collation, bool loadVec);
void HnswUpdateMetaPage(Relation index, int updateEntry, HnswElement entryPoint, BlockNumber insertPage, ForkNumber forkNum, bool building);
void HnswSetNeighborTuple(char *base, HnswNeighborTuple ntup, HnswElement e, int m);
void HnswAddHeapTid(HnswElement element, ItemPointer heaptid);

View File

@@ -60,12 +60,6 @@
#include "pgstat.h"
#endif
#if PG_VERSION_NUM >= 130000
#define CALLBACK_ITEM_POINTER ItemPointer tid
#else
#define CALLBACK_ITEM_POINTER HeapTuple hup
#endif
#if PG_VERSION_NUM >= 140000
#include "utils/backend_status.h"
#include "utils/wait_event.h"
@@ -75,10 +69,6 @@
#define PARALLEL_KEY_HNSW_AREA UINT64CONST(0xA000000000000002)
#define PARALLEL_KEY_QUERY_TEXT UINT64CONST(0xA000000000000003)
#if PG_VERSION_NUM < 130000
#define GENERATIONCHUNK_RAWSIZE (SIZEOF_SIZE_T + SIZEOF_VOID_P * 2)
#endif
/*
* Create the metapage
*/
@@ -192,7 +182,9 @@ CreateGraphPages(HnswBuildState * buildstate)
/* Initial size check */
if (etupSize > HNSW_TUPLE_ALLOC_SIZE)
elog(ERROR, "index tuple too large");
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
errmsg("index tuple too large")));
HnswSetElementTuple(base, etup, element);
@@ -583,17 +575,13 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
* Callback for table_index_build_scan
*/
static void
BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
BuildCallback(Relation index, ItemPointer tid, Datum *values,
bool *isnull, bool tupleIsAlive, void *state)
{
HnswBuildState *buildstate = (HnswBuildState *) state;
HnswGraph *graph = buildstate->graph;
MemoryContext oldCtx;
#if PG_VERSION_NUM < 130000
ItemPointer tid = &hup->t_self;
#endif
/* Skip nulls */
if (isnull[0])
return;
@@ -656,11 +644,7 @@ HnswMemoryContextAlloc(Size size, void *state)
HnswBuildState *buildstate = (HnswBuildState *) state;
void *chunk = MemoryContextAlloc(buildstate->graphCtx, size);
#if PG_VERSION_NUM >= 130000
buildstate->graphData.memoryUsed = MemoryContextMemAllocated(buildstate->graphCtx, false);
#else
buildstate->graphData.memoryUsed += MAXALIGN(size);
#endif
return chunk;
}
@@ -696,17 +680,25 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
/* Disallow varbit since require fixed dimensions */
if (TupleDescAttr(index->rd_att, 0)->atttypid == VARBITOID)
elog(ERROR, "type not supported for hnsw index");
ereport(ERROR,
(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
errmsg("type not supported for hnsw index")));
/* Require column to have dimensions to be indexed */
if (buildstate->dimensions < 0)
elog(ERROR, "column does not have dimensions");
ereport(ERROR,
(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("column does not have dimensions")));
if (buildstate->dimensions > buildstate->typeInfo->maxDimensions)
elog(ERROR, "column cannot have more than %d dimensions for hnsw index", buildstate->typeInfo->maxDimensions);
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
errmsg("column cannot have more than %d dimensions for hnsw index", buildstate->typeInfo->maxDimensions)));
if (buildstate->efConstruction < 2 * buildstate->m)
elog(ERROR, "ef_construction must be greater than or equal to 2 * m");
ereport(ERROR,
(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("ef_construction must be greater than or equal to 2 * m")));
buildstate->reltuples = 0;
buildstate->indtuples = 0;

View File

@@ -379,8 +379,12 @@ HnswUpdateNeighborsOnDisk(Relation index, FmgrInfo *procinfo, Oid collation, Hns
HnswElement neighborElement = HnswPtrAccess(base, hc->element);
OffsetNumber offno = neighborElement->neighborOffno;
/* Get latest neighbors since they may have changed */
/* Do not lock yet since selecting neighbors can take time */
/*
* Get latest neighbors since they may have changed. Do not lock
* yet since selecting neighbors can take time. Could use
* optimistic locking to retry if another update occurs before
* getting exclusive lock.
*/
HnswLoadNeighbors(neighborElement, index, m);
/*

View File

@@ -160,15 +160,15 @@ hnswgettuple(IndexScanDesc scan, ScanDirection dir)
so->first = false;
#if defined(HNSW_MEMORY) && PG_VERSION_NUM >= 130000
elog(INFO, "memory: %zu MB", MemoryContextMemAllocated(so->tmpCtx, false) / (1024 * 1024));
#if defined(HNSW_MEMORY)
elog(INFO, "memory: %zu KB", MemoryContextMemAllocated(so->tmpCtx, false) / 1024);
#endif
}
while (list_length(so->w) > 0)
{
char *base = NULL;
HnswCandidate *hc = llast(so->w);
HnswSearchCandidate *hc = llast(so->w);
HnswElement element = HnswPtrAccess(base, hc->element);
ItemPointer heaptid;

View File

@@ -5,6 +5,7 @@
#include "access/generic_xlog.h"
#include "catalog/pg_type.h"
#include "catalog/pg_type_d.h"
#include "common/hashfn.h"
#include "fmgr.h"
#include "hnsw.h"
#include "lib/pairingheap.h"
@@ -14,12 +15,6 @@
#include "utils/memdebug.h"
#include "utils/rel.h"
#if PG_VERSION_NUM >= 130000
#include "common/hashfn.h"
#else
#include "utils/hashutils.h"
#endif
#if PG_VERSION_NUM < 170000
static inline uint64
murmurhash64(uint64 data)
@@ -112,6 +107,12 @@ typedef union
tidhash_hash *tids;
} visited_hash;
typedef union
{
HnswElement element;
ItemPointerData indextid;
} HnswUnvisited;
/*
* Get the max number of connections in an upper layer for each element in the index
*/
@@ -158,6 +159,9 @@ HnswOptionalProcInfo(Relation index, uint16 procnum)
Datum
HnswNormValue(const HnswTypeInfo * typeInfo, Oid collation, Datum value)
{
if (!typeInfo->normalize)
return value;
return DirectFunctionCall1Coll(typeInfo->normalize, collation, value);
}
@@ -547,19 +551,19 @@ HnswLoadElementFromTuple(HnswElement element, HnswElementTuple etup, bool loadHe
/*
* 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, float *maxDistance)
static void
HnswLoadElementImpl(BlockNumber blkno, OffsetNumber offno, float *distance, Datum *q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec, float *maxDistance, HnswElement * element)
{
Buffer buf;
Page page;
HnswElementTuple etup;
/* Read vector */
buf = ReadBuffer(index, element->blkno);
buf = ReadBuffer(index, blkno);
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, element->offno));
etup = (HnswElementTuple) PageGetItem(page, PageGetItemId(page, offno));
Assert(HnswIsElementTuple(etup));
@@ -574,19 +578,32 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
/* Load element */
if (distance == NULL || maxDistance == NULL || *distance < *maxDistance)
HnswLoadElementFromTuple(element, etup, true, loadVec);
{
if (*element == NULL)
*element = HnswInitElementFromBlock(blkno, offno);
HnswLoadElementFromTuple(*element, etup, true, loadVec);
}
UnlockReleaseBuffer(buf);
}
/*
* Get the distance for a candidate
* 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, float *maxDistance)
{
HnswLoadElementImpl(element->blkno, element->offno, distance, q, index, procinfo, collation, loadVec, maxDistance, &element);
}
/*
* Get the distance for an element
*/
static float
GetCandidateDistance(char *base, HnswCandidate * hc, Datum q, FmgrInfo *procinfo, Oid collation)
GetElementDistance(char *base, HnswElement element, Datum q, FmgrInfo *procinfo, Oid collation)
{
HnswElement hce = HnswPtrAccess(base, hc->element);
Datum value = HnswGetValue(base, hce);
Datum value = HnswGetValue(base, element);
return DatumGetFloat8(FunctionCall2Coll(procinfo, collation, q, value));
}
@@ -594,29 +611,32 @@ GetCandidateDistance(char *base, HnswCandidate * hc, Datum q, FmgrInfo *procinfo
/*
* Create a candidate for the entry point
*/
HnswCandidate *
HnswSearchCandidate *
HnswEntryCandidate(char *base, HnswElement entryPoint, Datum q, Relation index, FmgrInfo *procinfo, Oid collation, bool loadVec)
{
HnswCandidate *hc = palloc(sizeof(HnswCandidate));
HnswSearchCandidate *hc = palloc(sizeof(HnswSearchCandidate));
HnswPtrStore(base, hc->element, entryPoint);
if (index == NULL)
hc->distance = GetCandidateDistance(base, hc, q, procinfo, collation);
hc->distance = GetElementDistance(base, entryPoint, q, procinfo, collation);
else
HnswLoadElement(entryPoint, &hc->distance, &q, index, procinfo, collation, loadVec, NULL);
return hc;
}
#define HnswGetSearchCandidate(membername, ptr) pairingheap_container(HnswSearchCandidate, membername, ptr)
#define HnswGetSearchCandidateConst(membername, ptr) pairingheap_const_container(HnswSearchCandidate, membername, ptr)
/*
* Compare candidate distances
*/
static int
CompareNearestCandidates(const pairingheap_node *a, const pairingheap_node *b, void *arg)
{
if (((const HnswPairingHeapNode *) a)->inner->distance < ((const HnswPairingHeapNode *) b)->inner->distance)
if (HnswGetSearchCandidateConst(c_node, a)->distance < HnswGetSearchCandidateConst(c_node, b)->distance)
return 1;
if (((const HnswPairingHeapNode *) a)->inner->distance > ((const HnswPairingHeapNode *) b)->inner->distance)
if (HnswGetSearchCandidateConst(c_node, a)->distance > HnswGetSearchCandidateConst(c_node, b)->distance)
return -1;
return 0;
@@ -628,27 +648,15 @@ CompareNearestCandidates(const pairingheap_node *a, const pairingheap_node *b, v
static int
CompareFurthestCandidates(const pairingheap_node *a, const pairingheap_node *b, void *arg)
{
if (((const HnswPairingHeapNode *) a)->inner->distance < ((const HnswPairingHeapNode *) b)->inner->distance)
if (HnswGetSearchCandidateConst(w_node, a)->distance < HnswGetSearchCandidateConst(w_node, b)->distance)
return -1;
if (((const HnswPairingHeapNode *) a)->inner->distance > ((const HnswPairingHeapNode *) b)->inner->distance)
if (HnswGetSearchCandidateConst(w_node, a)->distance > HnswGetSearchCandidateConst(w_node, b)->distance)
return 1;
return 0;
}
/*
* Create a pairing heap node for a candidate
*/
static HnswPairingHeapNode *
CreatePairingHeapNode(HnswCandidate * c)
{
HnswPairingHeapNode *node = palloc(sizeof(HnswPairingHeapNode));
node->inner = c;
return node;
}
/*
* Init visited
*/
@@ -667,11 +675,11 @@ InitVisited(char *base, visited_hash * v, Relation index, int ef, int m)
* Add to visited
*/
static inline void
AddToVisited(char *base, visited_hash * v, HnswCandidate * hc, Relation index, bool *found)
AddToVisited(char *base, visited_hash * v, HnswElementPtr elementPtr, Relation index, bool *found)
{
if (index != NULL)
{
HnswElement element = HnswPtrAccess(base, hc->element);
HnswElement element = HnswPtrAccess(base, elementPtr);
ItemPointerData indextid;
ItemPointerSet(&indextid, element->blkno, element->offno);
@@ -679,23 +687,15 @@ AddToVisited(char *base, visited_hash * v, HnswCandidate * hc, Relation index, b
}
else if (base != NULL)
{
#if PG_VERSION_NUM >= 130000
HnswElement element = HnswPtrAccess(base, hc->element);
HnswElement element = HnswPtrAccess(base, elementPtr);
offsethash_insert_hash(v->offsets, HnswPtrOffset(hc->element), element->hash, found);
#else
offsethash_insert(v->offsets, HnswPtrOffset(hc->element), found);
#endif
offsethash_insert_hash(v->offsets, HnswPtrOffset(elementPtr), element->hash, found);
}
else
{
#if PG_VERSION_NUM >= 130000
HnswElement element = HnswPtrAccess(base, hc->element);
HnswElement element = HnswPtrAccess(base, elementPtr);
pointerhash_insert_hash(v->pointers, (uintptr_t) HnswPtrPointer(hc->element), element->hash, found);
#else
pointerhash_insert(v->pointers, (uintptr_t) HnswPtrPointer(hc->element), found);
#endif
pointerhash_insert_hash(v->pointers, (uintptr_t) HnswPtrPointer(elementPtr), element->hash, found);
}
}
@@ -703,20 +703,89 @@ AddToVisited(char *base, visited_hash * v, HnswCandidate * hc, Relation index, b
* Count element towards ef
*/
static inline bool
CountElement(char *base, HnswElement skipElement, HnswCandidate * hc)
CountElement(HnswElement skipElement, HnswElement e)
{
HnswElement e;
if (skipElement == NULL)
return true;
/* Ensure does not access heaptidsLength during in-memory build */
pg_memory_barrier();
e = HnswPtrAccess(base, hc->element);
/* Keep scan-build happy on Mac x86-64 */
Assert(e);
return e->heaptidsLength != 0;
}
/*
* Load unvisited neighbors from memory
*/
static void
HnswLoadUnvisitedFromMemory(char *base, HnswElement element, HnswUnvisited * unvisited, int *unvisitedLength, visited_hash * v, int lc, HnswNeighborArray * localNeighborhood, Size neighborhoodSize)
{
/* Get the neighborhood at layer lc */
HnswNeighborArray *neighborhood = HnswGetNeighbors(base, element, lc);
/* Copy neighborhood to local memory */
LWLockAcquire(&element->lock, LW_SHARED);
memcpy(localNeighborhood, neighborhood, neighborhoodSize);
LWLockRelease(&element->lock);
*unvisitedLength = 0;
for (int i = 0; i < localNeighborhood->length; i++)
{
HnswCandidate *hc = &localNeighborhood->items[i];
bool found;
AddToVisited(base, v, hc->element, NULL, &found);
if (!found)
unvisited[(*unvisitedLength)++].element = HnswPtrAccess(base, hc->element);
}
}
/*
* Load unvisited neighbors from disk
*/
static void
HnswLoadUnvisitedFromDisk(HnswElement element, HnswUnvisited * unvisited, int *unvisitedLength, visited_hash * v, Relation index, int m, int lm, int lc)
{
Buffer buf;
Page page;
HnswNeighborTuple ntup;
int start;
ItemPointerData indextids[HNSW_MAX_M * 2];
buf = ReadBuffer(index, element->neighborPage);
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
ntup = (HnswNeighborTuple) PageGetItem(page, PageGetItemId(page, element->neighborOffno));
start = (element->level - lc) * m;
/* Copy to minimize lock time */
memcpy(&indextids, ntup->indextids + start, lm * sizeof(ItemPointerData));
UnlockReleaseBuffer(buf);
*unvisitedLength = 0;
for (int i = 0; i < lm; i++)
{
ItemPointer indextid = &indextids[i];
bool found;
if (!ItemPointerIsValid(indextid))
break;
tidhash_insert(v->tids, *indextid, &found);
if (!found)
unvisited[(*unvisitedLength)++].indextid = *indextid;
}
}
/*
* Algorithm 2 from paper
*/
@@ -729,43 +798,45 @@ HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, F
int wlen = 0;
visited_hash v;
ListCell *lc2;
HnswNeighborArray *neighborhoodData = NULL;
Size neighborhoodSize;
HnswNeighborArray *localNeighborhood = NULL;
Size neighborhoodSize = 0;
int lm = HnswGetLayerM(m, lc);
HnswUnvisited *unvisited = palloc(lm * sizeof(HnswUnvisited));
int unvisitedLength;
InitVisited(base, &v, index, ef, m);
/* Create local memory for neighborhood if needed */
if (index == NULL)
{
neighborhoodSize = HNSW_NEIGHBOR_ARRAY_SIZE(HnswGetLayerM(m, lc));
neighborhoodData = palloc(neighborhoodSize);
neighborhoodSize = HNSW_NEIGHBOR_ARRAY_SIZE(lm);
localNeighborhood = palloc(neighborhoodSize);
}
/* Add entry points to v, C, and W */
foreach(lc2, ep)
{
HnswCandidate *hc = (HnswCandidate *) lfirst(lc2);
HnswSearchCandidate *hc = (HnswSearchCandidate *) lfirst(lc2);
bool found;
AddToVisited(base, &v, hc, index, &found);
AddToVisited(base, &v, hc->element, index, &found);
pairingheap_add(C, &(CreatePairingHeapNode(hc)->ph_node));
pairingheap_add(W, &(CreatePairingHeapNode(hc)->ph_node));
pairingheap_add(C, &hc->c_node);
pairingheap_add(W, &hc->w_node);
/*
* Do not count elements being deleted towards ef when vacuuming. It
* would be ideal to do this for inserts as well, but this could
* affect insert performance.
*/
if (CountElement(base, skipElement, hc))
if (CountElement(skipElement, HnswPtrAccess(base, hc->element)))
wlen++;
}
while (!pairingheap_is_empty(C))
{
HnswNeighborArray *neighborhood;
HnswCandidate *c = ((HnswPairingHeapNode *) pairingheap_remove_first(C))->inner;
HnswCandidate *f = ((HnswPairingHeapNode *) pairingheap_first(W))->inner;
HnswSearchCandidate *c = HnswGetSearchCandidate(c_node, pairingheap_remove_first(C));
HnswSearchCandidate *f = HnswGetSearchCandidate(w_node, pairingheap_first(W));
HnswElement cElement;
if (c->distance > f->distance)
@@ -773,73 +844,67 @@ HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, F
cElement = HnswPtrAccess(base, c->element);
if (HnswPtrIsNull(base, cElement->neighbors))
HnswLoadNeighbors(cElement, index, m);
/* Get the neighborhood at layer lc */
neighborhood = HnswGetNeighbors(base, cElement, lc);
/* Copy neighborhood to local memory if needed */
if (index == NULL)
HnswLoadUnvisitedFromMemory(base, cElement, unvisited, &unvisitedLength, &v, lc, localNeighborhood, neighborhoodSize);
else
HnswLoadUnvisitedFromDisk(cElement, unvisited, &unvisitedLength, &v, index, m, lm, lc);
for (int i = 0; i < unvisitedLength; i++)
{
LWLockAcquire(&cElement->lock, LW_SHARED);
memcpy(neighborhoodData, neighborhood, neighborhoodSize);
LWLockRelease(&cElement->lock);
neighborhood = neighborhoodData;
}
HnswElement eElement;
HnswSearchCandidate *e;
float eDistance;
bool alwaysAdd = wlen < ef;
for (int i = 0; i < neighborhood->length; i++)
{
HnswCandidate *e = &neighborhood->items[i];
bool visited;
f = HnswGetSearchCandidate(w_node, pairingheap_first(W));
AddToVisited(base, &v, e, index, &visited);
if (!visited)
if (index == NULL)
{
float eDistance;
HnswElement eElement = HnswPtrAccess(base, e->element);
bool alwaysAdd = wlen < ef;
eElement = unvisited[i].element;
eDistance = GetElementDistance(base, eElement, q, procinfo, collation);
}
else
{
ItemPointer indextid = &unvisited[i].indextid;
BlockNumber blkno = ItemPointerGetBlockNumber(indextid);
OffsetNumber offno = ItemPointerGetOffsetNumber(indextid);
f = ((HnswPairingHeapNode *) pairingheap_first(W))->inner;
/* Avoid any allocations if not adding */
eElement = NULL;
HnswLoadElementImpl(blkno, offno, &eDistance, &q, index, procinfo, collation, inserting, alwaysAdd ? NULL : &f->distance, &eElement);
if (index == NULL)
eDistance = GetCandidateDistance(base, e, q, procinfo, collation);
else
HnswLoadElement(eElement, &eDistance, &q, index, procinfo, collation, inserting, alwaysAdd ? NULL : &f->distance);
if (eElement == NULL)
continue;
}
if (eDistance < f->distance || alwaysAdd)
{
HnswCandidate *ec;
if (!(eDistance < f->distance || alwaysAdd))
continue;
Assert(!eElement->deleted);
Assert(!eElement->deleted);
/* Make robust to issues */
if (eElement->level < lc)
continue;
/* Make robust to issues */
if (eElement->level < lc)
continue;
/* Copy e */
ec = palloc(sizeof(HnswCandidate));
HnswPtrStore(base, ec->element, eElement);
ec->distance = eDistance;
/* Create a new candidate */
e = palloc(sizeof(HnswSearchCandidate));
HnswPtrStore(base, e->element, eElement);
e->distance = eDistance;
pairingheap_add(C, &e->c_node);
pairingheap_add(W, &e->w_node);
pairingheap_add(C, &(CreatePairingHeapNode(ec)->ph_node));
pairingheap_add(W, &(CreatePairingHeapNode(ec)->ph_node));
/*
* Do not count elements being deleted towards ef when vacuuming.
* It would be ideal to do this for inserts as well, but this
* could affect insert performance.
*/
if (CountElement(skipElement, eElement))
{
wlen++;
/*
* Do not count elements being deleted towards ef when
* vacuuming. It would be ideal to do this for inserts as
* well, but this could affect insert performance.
*/
if (CountElement(base, skipElement, e))
{
wlen++;
/* No need to decrement wlen */
if (wlen > ef)
pairingheap_remove_first(W);
}
}
/* No need to decrement wlen */
if (wlen > ef)
pairingheap_remove_first(W);
}
}
}
@@ -847,7 +912,7 @@ HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, F
/* Add each element of W to w */
while (!pairingheap_is_empty(W))
{
HnswCandidate *hc = ((HnswPairingHeapNode *) pairingheap_remove_first(W))->inner;
HnswSearchCandidate *hc = HnswGetSearchCandidate(w_node, pairingheap_remove_first(W));
w = lappend(w, hc);
}
@@ -859,17 +924,10 @@ HnswSearchLayer(char *base, Datum q, List *ep, int ef, int lc, Relation index, F
* Compare candidate distances with pointer tie-breaker
*/
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)
{
HnswCandidate *hca = lfirst(*(ListCell **) a);
HnswCandidate *hcb = lfirst(*(ListCell **) b);
#endif
if (hca->distance < hcb->distance)
return 1;
@@ -890,17 +948,10 @@ CompareCandidateDistances(const void *a, const void *b)
* Compare candidate distances with offset tie-breaker
*/
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)
{
HnswCandidate *hca = lfirst(*(ListCell **) a);
HnswCandidate *hcb = lfirst(*(ListCell **) b);
#endif
if (hca->distance < hcb->distance)
return 1;
@@ -1110,7 +1161,7 @@ HnswUpdateConnection(char *base, HnswElement element, HnswCandidate * hc, int lm
if (HnswPtrIsNull(base, hc3Element->value))
HnswLoadElement(hc3Element, &hc3->distance, &q, index, procinfo, collation, true, NULL);
else
hc3->distance = GetCandidateDistance(base, hc3, q, procinfo, collation);
hc3->distance = GetElementDistance(base, hc3Element, q, procinfo, collation);
/* Prune element if being deleted */
if (hc3Element->heaptidsLength == 0)
@@ -1182,7 +1233,6 @@ RemoveElements(char *base, List *w, HnswElement skipElement)
return w2;
}
#if PG_VERSION_NUM >= 130000
/*
* Precompute hash
*/
@@ -1198,7 +1248,6 @@ PrecomputeHash(char *base, HnswElement element)
else
element->hash = hash_offset(HnswPtrOffset(ptr));
}
#endif
/*
* Algorithm 1 from paper
@@ -1213,11 +1262,9 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
Datum q = HnswGetValue(base, element);
HnswElement skipElement = existing ? element : NULL;
#if PG_VERSION_NUM >= 130000
/* Precompute hash */
if (index == NULL)
PrecomputeHash(base, element);
#endif
/* No neighbors if no entry point */
if (entryPoint == NULL)
@@ -1246,16 +1293,27 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
{
int lm = HnswGetLayerM(m, lc);
List *neighbors;
List *lw;
List *lw = NIL;
ListCell *lc2;
w = HnswSearchLayer(base, q, ep, efConstruction, lc, index, procinfo, collation, m, true, skipElement);
/* Convert search candidates to candidates */
foreach(lc2, w)
{
HnswSearchCandidate *sc = lfirst(lc2);
HnswCandidate *hc = palloc(sizeof(HnswCandidate));
hc->element = sc->element;
hc->distance = sc->distance;
lw = lappend(lw, hc);
}
/* Elements being deleted or skipped can help with search */
/* but should be removed before selecting neighbors */
if (index != NULL)
lw = RemoveElements(base, w, skipElement);
else
lw = w;
lw = RemoveElements(base, lw, skipElement);
/*
* Candidates are sorted, but not deterministically. Could set
@@ -1272,6 +1330,7 @@ HnswFindElementNeighbors(char *base, HnswElement element, HnswElement entryPoint
PGDLLEXPORT Datum l2_normalize(PG_FUNCTION_ARGS);
PGDLLEXPORT Datum halfvec_l2_normalize(PG_FUNCTION_ARGS);
PGDLLEXPORT Datum minivec_l2_normalize(PG_FUNCTION_ARGS);
PGDLLEXPORT Datum sparsevec_l2_normalize(PG_FUNCTION_ARGS);
static void
@@ -1280,7 +1339,9 @@ SparsevecCheckValue(Pointer v)
SparseVector *vec = (SparseVector *) v;
if (vec->nnz > HNSW_MAX_NNZ)
elog(ERROR, "sparsevec cannot have more than %d non-zero elements for hnsw index", HNSW_MAX_NNZ);
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
errmsg("sparsevec cannot have more than %d non-zero elements for hnsw index", HNSW_MAX_NNZ)));
}
/*
@@ -1318,6 +1379,20 @@ hnsw_halfvec_support(PG_FUNCTION_ARGS)
PG_RETURN_POINTER(&typeInfo);
};
FUNCTION_PREFIX PG_FUNCTION_INFO_V1(hnsw_minivec_support);
Datum
hnsw_minivec_support(PG_FUNCTION_ARGS)
{
static const HnswTypeInfo typeInfo = {
.maxDimensions = HNSW_MAX_DIM * 4,
/* Do not normalize to maximize precision */
.normalize = NULL,
.checkValue = NULL
};
PG_RETURN_POINTER(&typeInfo);
};
FUNCTION_PREFIX PG_FUNCTION_INFO_V1(hnsw_bit_support);
Datum
hnsw_bit_support(PG_FUNCTION_ARGS)

View File

@@ -26,12 +26,6 @@
#include "pgstat.h"
#endif
#if PG_VERSION_NUM >= 130000
#define CALLBACK_ITEM_POINTER ItemPointer tid
#else
#define CALLBACK_ITEM_POINTER HeapTuple hup
#endif
#if PG_VERSION_NUM >= 140000
#include "utils/backend_status.h"
#include "utils/wait_event.h"
@@ -96,7 +90,7 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
* Callback for sampling
*/
static void
SampleCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
SampleCallback(Relation index, ItemPointer tid, Datum *values,
bool *isnull, bool tupleIsAlive, void *state)
{
IvfflatBuildState *buildstate = (IvfflatBuildState *) state;
@@ -207,16 +201,12 @@ AddTupleToSort(Relation index, ItemPointer tid, Datum *values, IvfflatBuildState
* Callback for table_index_build_scan
*/
static void
BuildCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
BuildCallback(Relation index, ItemPointer tid, Datum *values,
bool *isnull, bool tupleIsAlive, void *state)
{
IvfflatBuildState *buildstate = (IvfflatBuildState *) state;
MemoryContext oldCtx;
#if PG_VERSION_NUM < 130000
ItemPointer tid = &hup->t_self;
#endif
/* Skip nulls */
if (isnull[0])
return;
@@ -335,14 +325,20 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
/* Disallow varbit since require fixed dimensions */
if (TupleDescAttr(index->rd_att, 0)->atttypid == VARBITOID)
elog(ERROR, "type not supported for ivfflat index");
ereport(ERROR,
(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
errmsg("type not supported for ivfflat index")));
/* Require column to have dimensions to be indexed */
if (buildstate->dimensions < 0)
elog(ERROR, "column does not have dimensions");
ereport(ERROR,
(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("column does not have dimensions")));
if (buildstate->dimensions > buildstate->typeInfo->maxDimensions)
elog(ERROR, "column cannot have more than %d dimensions for ivfflat index", buildstate->typeInfo->maxDimensions);
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
errmsg("column cannot have more than %d dimensions for ivfflat index", buildstate->typeInfo->maxDimensions)));
buildstate->reltuples = 0;
buildstate->indtuples = 0;
@@ -355,7 +351,9 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
/* Require more than one dimension for spherical k-means */
if (buildstate->kmeansnormprocinfo != NULL && buildstate->dimensions == 1)
elog(ERROR, "dimensions must be greater than one for this opclass");
ereport(ERROR,
(errcode(ERRCODE_INVALID_PARAMETER_VALUE),
errmsg("dimensions must be greater than one for this opclass")));
/* Create tuple description for sorting */
buildstate->tupdesc = CreateTemplateTupleDesc(3);
@@ -562,6 +560,20 @@ PrintKmeansMetrics(IvfflatBuildState * buildstate)
}
#endif
/*
* Initialize build sort state
*/
static Tuplesortstate *
InitBuildSortState(TupleDesc tupdesc, int memory, SortCoordinate coordinate)
{
AttrNumber attNums[] = {1};
Oid sortOperators[] = {Int4LessOperator};
Oid sortCollations[] = {InvalidOid};
bool nullsFirstFlags[] = {false};
return tuplesort_begin_heap(tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, memory, coordinate, false);
}
/*
* Within leader, wait for end of heap scan
*/
@@ -609,12 +621,6 @@ IvfflatParallelScanAndSort(IvfflatSpool * ivfspool, IvfflatShared * ivfshared, S
double reltuples;
IndexInfo *indexInfo;
/* Sort options, which must match AssignTuples */
AttrNumber attNums[] = {1};
Oid sortOperators[] = {Int4LessOperator};
Oid sortCollations[] = {InvalidOid};
bool nullsFirstFlags[] = {false};
/* Initialize local tuplesort coordination state */
coordinate = palloc0(sizeof(SortCoordinateData));
coordinate->isWorker = true;
@@ -627,7 +633,7 @@ IvfflatParallelScanAndSort(IvfflatSpool * ivfspool, IvfflatShared * ivfshared, S
InitBuildState(&buildstate, ivfspool->heap, ivfspool->index, indexInfo);
memcpy(buildstate.centers->items, ivfcenters, buildstate.centers->itemsize * buildstate.centers->maxlen);
buildstate.centers->length = buildstate.centers->maxlen;
ivfspool->sortstate = tuplesort_begin_heap(buildstate.tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, sortmem, coordinate, false);
ivfspool->sortstate = InitBuildSortState(buildstate.tupdesc, sortmem, coordinate);
buildstate.sortstate = ivfspool->sortstate;
scan = table_beginscan_parallel(ivfspool->heap,
ParallelTableScanFromIvfflatShared(ivfshared));
@@ -924,12 +930,6 @@ AssignTuples(IvfflatBuildState * buildstate)
int parallel_workers = 0;
SortCoordinate coordinate = NULL;
/* Sort options, which must match IvfflatParallelScanAndSort */
AttrNumber attNums[] = {1};
Oid sortOperators[] = {Int4LessOperator};
Oid sortCollations[] = {InvalidOid};
bool nullsFirstFlags[] = {false};
pgstat_progress_update_param(PROGRESS_CREATEIDX_SUBPHASE, PROGRESS_IVFFLAT_PHASE_ASSIGN);
/* Calculate parallel workers */
@@ -950,7 +950,7 @@ AssignTuples(IvfflatBuildState * buildstate)
}
/* Begin serial/leader tuplesort */
buildstate->sortstate = tuplesort_begin_heap(buildstate->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, maintenance_work_mem, coordinate, false);
buildstate->sortstate = InitBuildSortState(buildstate->tupdesc, maintenance_work_mem, coordinate);
/* Add tuples to sort */
if (buildstate->heap != NULL)

View File

@@ -7,6 +7,7 @@
#include "commands/progress.h"
#include "commands/vacuum.h"
#include "ivfflat.h"
#include "utils/float.h"
#include "utils/guc.h"
#include "utils/selfuncs.h"
#include "utils/spccache.h"
@@ -26,11 +27,7 @@ IvfflatInit(void)
{
ivfflat_relopt_kind = add_reloption_kind();
add_int_reloption(ivfflat_relopt_kind, "lists", "Number of inverted lists",
IVFFLAT_DEFAULT_LISTS, IVFFLAT_MIN_LISTS, IVFFLAT_MAX_LISTS
#if PG_VERSION_NUM >= 130000
,AccessExclusiveLock
#endif
);
IVFFLAT_DEFAULT_LISTS, IVFFLAT_MIN_LISTS, IVFFLAT_MAX_LISTS, AccessExclusiveLock);
DefineCustomIntVariable("ivfflat.probes", "Sets the number of probes",
"Valid range is 1..lists.", &ivfflat_probes,
@@ -78,8 +75,8 @@ ivfflatcostestimate(PlannerInfo *root, IndexPath *path, double loop_count,
/* Never use index without order */
if (path->indexorderbys == NULL)
{
*indexStartupCost = DBL_MAX;
*indexTotalCost = DBL_MAX;
*indexStartupCost = get_float8_infinity();
*indexTotalCost = get_float8_infinity();
*indexSelectivity = 0;
*indexCorrelation = 0;
*indexPages = 0;
@@ -148,23 +145,10 @@ ivfflatoptions(Datum reloptions, bool validate)
{"lists", RELOPT_TYPE_INT, offsetof(IvfflatOptions, lists)},
};
#if PG_VERSION_NUM >= 130000
return (bytea *) build_reloptions(reloptions, validate,
ivfflat_relopt_kind,
sizeof(IvfflatOptions),
tab, lengthof(tab));
#else
relopt_value *options;
int numoptions;
IvfflatOptions *rdopts;
options = parseRelOptions(reloptions, validate, ivfflat_relopt_kind, &numoptions);
rdopts = allocateReloptStruct(sizeof(IvfflatOptions), options, numoptions);
fillRelOptions((void *) rdopts, sizeof(IvfflatOptions), options, numoptions,
validate, tab, lengthof(tab));
return (bytea *) rdopts;
#endif
}
/*
@@ -189,9 +173,7 @@ ivfflathandler(PG_FUNCTION_ARGS)
amroutine->amstrategies = 0;
amroutine->amsupport = 5;
#if PG_VERSION_NUM >= 130000
amroutine->amoptsprocnum = 0;
#endif
amroutine->amcanorder = false;
amroutine->amcanorderbyop = true;
amroutine->amcanbackward = false; /* can change direction mid-scan */
@@ -204,17 +186,24 @@ ivfflathandler(PG_FUNCTION_ARGS)
amroutine->amclusterable = false;
amroutine->ampredlocks = false;
amroutine->amcanparallel = false;
amroutine->amcaninclude = false;
#if PG_VERSION_NUM >= 130000
amroutine->amusemaintenanceworkmem = false; /* not used during VACUUM */
amroutine->amparallelvacuumoptions = VACUUM_OPTION_PARALLEL_BULKDEL;
#if PG_VERSION_NUM >= 170000
amroutine->amcanbuildparallel = true;
#endif
amroutine->amcaninclude = false;
amroutine->amusemaintenanceworkmem = false; /* not used during VACUUM */
#if PG_VERSION_NUM >= 160000
amroutine->amsummarizing = false;
#endif
amroutine->amparallelvacuumoptions = VACUUM_OPTION_PARALLEL_BULKDEL;
amroutine->amkeytype = InvalidOid;
/* Interface functions */
amroutine->ambuild = ivfflatbuild;
amroutine->ambuildempty = ivfflatbuildempty;
amroutine->aminsert = ivfflatinsert;
#if PG_VERSION_NUM >= 170000
amroutine->aminsertcleanup = NULL;
#endif
amroutine->ambulkdelete = ivfflatbulkdelete;
amroutine->amvacuumcleanup = ivfflatvacuumcleanup;
amroutine->amcanreturn = NULL; /* tuple not included in heapsort */

View File

@@ -253,8 +253,9 @@ typedef struct IvfflatScanOpaqueData
/* Sorting */
Tuplesortstate *sortstate;
TupleDesc tupdesc;
TupleTableSlot *slot;
bool isnull;
TupleTableSlot *vslot;
TupleTableSlot *mslot;
BufferAccessStrategy bas;
/* Support functions */
FmgrInfo *procinfo;

View File

@@ -151,12 +151,8 @@ RandomCenters(Relation index, VectorArray centers, const IvfflatTypeInfo * typeI
static void
ShowMemoryUsage(MemoryContext context, Size estimatedSize)
{
#if PG_VERSION_NUM >= 130000
elog(INFO, "total memory: %zu MB",
MemoryContextMemAllocated(context, true) / (1024 * 1024));
#else
MemoryContextStats(context);
#endif
elog(INFO, "estimated memory: %zu MB", estimatedSize / (1024 * 1024));
}
#endif
@@ -327,7 +323,7 @@ ElkanKmeans(Relation index, VectorArray samples, VectorArray centers, const Ivff
newCenters->length = numCenters;
#ifdef IVFFLAT_MEMORY
ShowMemoryUsage(MemoryContextGetParent(CurrentMemoryContext));
ShowMemoryUsage(MemoryContextGetParent(CurrentMemoryContext), totalSize);
#endif
/* Pick initial centers */

View File

@@ -11,16 +11,23 @@
#include "pgstat.h"
#include "storage/bufmgr.h"
#ifdef IVFFLAT_MEMORY
#include "utils/memutils.h"
#endif
#define GetScanList(ptr) pairingheap_container(IvfflatScanList, ph_node, ptr)
#define GetScanListConst(ptr) pairingheap_const_container(IvfflatScanList, ph_node, ptr)
/*
* Compare list distances
*/
static int
CompareLists(const pairingheap_node *a, const pairingheap_node *b, void *arg)
{
if (((const IvfflatScanList *) a)->distance > ((const IvfflatScanList *) b)->distance)
if (GetScanListConst(a)->distance > GetScanListConst(b)->distance)
return 1;
if (((const IvfflatScanList *) a)->distance < ((const IvfflatScanList *) b)->distance)
if (GetScanListConst(a)->distance < GetScanListConst(b)->distance)
return -1;
return 0;
@@ -72,14 +79,14 @@ GetScanLists(IndexScanDesc scan, Datum value)
/* Calculate max distance */
if (listCount == so->probes)
maxDistance = ((IvfflatScanList *) pairingheap_first(so->listQueue))->distance;
maxDistance = GetScanList(pairingheap_first(so->listQueue))->distance;
}
else if (distance < maxDistance)
{
IvfflatScanList *scanlist;
/* Remove */
scanlist = (IvfflatScanList *) pairingheap_remove_first(so->listQueue);
scanlist = GetScanList(pairingheap_remove_first(so->listQueue));
/* Reuse */
scanlist->startPage = list->startPage;
@@ -87,7 +94,7 @@ GetScanLists(IndexScanDesc scan, Datum value)
pairingheap_add(so->listQueue, &scanlist->ph_node);
/* Update max distance */
maxDistance = ((IvfflatScanList *) pairingheap_first(so->listQueue))->distance;
maxDistance = GetScanList(pairingheap_first(so->listQueue))->distance;
}
}
@@ -106,19 +113,12 @@ GetScanItems(IndexScanDesc scan, Datum value)
IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
TupleDesc tupdesc = RelationGetDescr(scan->indexRelation);
double tuples = 0;
TupleTableSlot *slot = MakeSingleTupleTableSlot(so->tupdesc, &TTSOpsVirtual);
/*
* Reuse same set of shared buffers for scan
*
* See postgres/src/backend/storage/buffer/README for description
*/
BufferAccessStrategy bas = GetAccessStrategy(BAS_BULKREAD);
TupleTableSlot *slot = so->vslot;
/* Search closest probes lists */
while (!pairingheap_is_empty(so->listQueue))
{
BlockNumber searchPage = ((IvfflatScanList *) pairingheap_remove_first(so->listQueue))->startPage;
BlockNumber searchPage = GetScanList(pairingheap_remove_first(so->listQueue))->startPage;
/* Search all entry pages for list */
while (BlockNumberIsValid(searchPage))
@@ -127,7 +127,7 @@ GetScanItems(IndexScanDesc scan, Datum value)
Page page;
OffsetNumber maxoffno;
buf = ReadBufferExtended(scan->indexRelation, MAIN_FORKNUM, searchPage, RBM_NORMAL, bas);
buf = ReadBufferExtended(scan->indexRelation, MAIN_FORKNUM, searchPage, RBM_NORMAL, so->bas);
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
maxoffno = PageGetMaxOffsetNumber(page);
@@ -166,8 +166,6 @@ GetScanItems(IndexScanDesc scan, Datum value)
}
}
FreeAccessStrategy(bas);
if (tuples < 100)
ereport(DEBUG1,
(errmsg("index scan found few tuples"),
@@ -217,6 +215,20 @@ GetScanValue(IndexScanDesc scan)
return value;
}
/*
* Initialize scan sort state
*/
static Tuplesortstate *
InitScanSortState(TupleDesc tupdesc)
{
AttrNumber attNums[] = {1};
Oid sortOperators[] = {Float8LessOperator};
Oid sortCollations[] = {InvalidOid};
bool nullsFirstFlags[] = {false};
return tuplesort_begin_heap(tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, work_mem, NULL, false);
}
/*
* Prepare for an index scan
*/
@@ -227,10 +239,6 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
IvfflatScanOpaque so;
int lists;
int dimensions;
AttrNumber attNums[] = {1};
Oid sortOperators[] = {Float8LessOperator};
Oid sortCollations[] = {InvalidOid};
bool nullsFirstFlags[] = {false};
int probes = ivfflat_probes;
scan = RelationGetIndexScan(index, nkeys, norderbys);
@@ -258,9 +266,18 @@ ivfflatbeginscan(Relation index, int nkeys, int norderbys)
TupleDescInitEntry(so->tupdesc, (AttrNumber) 2, "heaptid", TIDOID, -1, 0);
/* Prep sort */
so->sortstate = tuplesort_begin_heap(so->tupdesc, 1, attNums, sortOperators, sortCollations, nullsFirstFlags, work_mem, NULL, false);
so->sortstate = InitScanSortState(so->tupdesc);
so->slot = MakeSingleTupleTableSlot(so->tupdesc, &TTSOpsMinimalTuple);
/* Need separate slots for puttuple and gettuple */
so->vslot = MakeSingleTupleTableSlot(so->tupdesc, &TTSOpsVirtual);
so->mslot = MakeSingleTupleTableSlot(so->tupdesc, &TTSOpsMinimalTuple);
/*
* Reuse same set of shared buffers for scan
*
* See postgres/src/backend/storage/buffer/README for description
*/
so->bas = GetAccessStrategy(BAS_BULKREAD);
so->listQueue = pairingheap_allocate(CompareLists, scan);
@@ -277,10 +294,8 @@ ivfflatrescan(IndexScanDesc scan, ScanKey keys, int nkeys, ScanKey orderbys, int
{
IvfflatScanOpaque so = (IvfflatScanOpaque) scan->opaque;
#if PG_VERSION_NUM >= 130000
if (!so->first)
tuplesort_reset(so->sortstate);
#endif
so->first = true;
pairingheap_reset(so->listQueue);
@@ -327,14 +342,19 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
IvfflatBench("GetScanItems", GetScanItems(scan, value));
so->first = false;
#if defined(IVFFLAT_MEMORY)
elog(INFO, "memory: %zu MB", MemoryContextMemAllocated(CurrentMemoryContext, true) / (1024 * 1024));
#endif
/* Clean up if we allocated a new value */
if (value != scan->orderByData->sk_argument)
pfree(DatumGetPointer(value));
}
if (tuplesort_gettupleslot(so->sortstate, true, false, so->slot, NULL))
if (tuplesort_gettupleslot(so->sortstate, true, false, so->mslot, NULL))
{
ItemPointer heaptid = (ItemPointer) DatumGetPointer(slot_getattr(so->slot, 2, &so->isnull));
bool isnull;
ItemPointer heaptid = (ItemPointer) DatumGetPointer(slot_getattr(so->mslot, 2, &isnull));
scan->xs_heaptid = *heaptid;
scan->xs_recheck = false;
@@ -355,6 +375,10 @@ ivfflatendscan(IndexScanDesc scan)
pairingheap_free(so->listQueue);
tuplesort_end(so->sortstate);
FreeAccessStrategy(so->bas);
FreeTupleDesc(so->tupdesc);
/* TODO Free vslot and mslot without freeing TupleDesc */
pfree(so);
scan->opaque = NULL;

View File

@@ -7,6 +7,7 @@
#include "halfutils.h"
#include "halfvec.h"
#include "ivfflat.h"
#include "minivec.h"
#include "storage/bufmgr.h"
/*
@@ -70,6 +71,9 @@ IvfflatOptionalProcInfo(Relation index, uint16 procnum)
Datum
IvfflatNormValue(const IvfflatTypeInfo * typeInfo, Oid collation, Datum value)
{
if (!typeInfo->normalize)
return value;
return DirectFunctionCall1Coll(typeInfo->normalize, collation, value);
}
@@ -231,6 +235,7 @@ IvfflatUpdateList(Relation index, ListInfo listInfo,
PGDLLEXPORT Datum l2_normalize(PG_FUNCTION_ARGS);
PGDLLEXPORT Datum halfvec_l2_normalize(PG_FUNCTION_ARGS);
PGDLLEXPORT Datum minivec_l2_normalize(PG_FUNCTION_ARGS);
PGDLLEXPORT Datum sparsevec_l2_normalize(PG_FUNCTION_ARGS);
static Size
@@ -245,6 +250,12 @@ HalfvecItemSize(int dimensions)
return HALFVEC_SIZE(dimensions);
}
static Size
MinivecItemSize(int dimensions)
{
return MINIVEC_SIZE(dimensions);
}
static Size
BitItemSize(int dimensions)
{
@@ -275,6 +286,18 @@ HalfvecUpdateCenter(Pointer v, int dimensions, float *x)
vec->x[k] = Float4ToHalfUnchecked(x[k]);
}
static void
MinivecUpdateCenter(Pointer v, int dimensions, float *x)
{
MiniVector *vec = (MiniVector *) v;
SET_VARSIZE(vec, MINIVEC_SIZE(dimensions));
vec->dim = dimensions;
for (int k = 0; k < dimensions; k++)
vec->x[k] = Float4ToFp8Unchecked(x[k]);
}
static void
BitUpdateCenter(Pointer v, int dimensions, float *x)
{
@@ -309,6 +332,15 @@ HalfvecSumCenter(Pointer v, float *x)
x[k] += HalfToFloat4(vec->x[k]);
}
static void
MinivecSumCenter(Pointer v, float *x)
{
MiniVector *vec = (MiniVector *) v;
for (int k = 0; k < vec->dim; k++)
x[k] += Fp8ToFloat4(vec->x[k]);
}
static void
BitSumCenter(Pointer v, float *x)
{
@@ -357,6 +389,22 @@ ivfflat_halfvec_support(PG_FUNCTION_ARGS)
PG_RETURN_POINTER(&typeInfo);
};
FUNCTION_PREFIX PG_FUNCTION_INFO_V1(ivfflat_minivec_support);
Datum
ivfflat_minivec_support(PG_FUNCTION_ARGS)
{
static const IvfflatTypeInfo typeInfo = {
.maxDimensions = IVFFLAT_MAX_DIM * 4,
/* Do not normalize to maximize precision */
.normalize = NULL,
.itemSize = MinivecItemSize,
.updateCenter = MinivecUpdateCenter,
.sumCenter = MinivecSumCenter
};
PG_RETURN_POINTER(&typeInfo);
};
FUNCTION_PREFIX PG_FUNCTION_INFO_V1(ivfflat_bit_support);
Datum
ivfflat_bit_support(PG_FUNCTION_ARGS)

1051
src/minivec.c Normal file

File diff suppressed because it is too large Load Diff

163
src/minivec.h Normal file
View File

@@ -0,0 +1,163 @@
#ifndef MINIVEC_H
#define MINIVEC_H
#define MINIVEC_MAX_DIM 16000
#define fp8 uint8
#define MINIVEC_SIZE(_dim) (offsetof(MiniVector, x) + sizeof(fp8)*(_dim))
#define DatumGetMiniVector(x) ((MiniVector *) PG_DETOAST_DATUM(x))
#define PG_GETARG_MINIVEC_P(x) DatumGetMiniVector(PG_GETARG_DATUM(x))
#define PG_RETURN_MINIVEC_P(x) PG_RETURN_POINTER(x)
typedef struct MiniVector
{
int32 vl_len_; /* varlena header (do not touch directly!) */
int16 dim; /* number of dimensions */
int16 unused; /* reserved for future use, always zero */
fp8 x[FLEXIBLE_ARRAY_MEMBER];
} MiniVector;
MiniVector *InitMiniVector(int dim);
/*
* Check if fp8 is NaN
*/
static inline bool
Fp8IsNan(fp8 num)
{
return (num & 0x7F) == 0x7F;
}
/*
* Check if fp8 is zero
*/
static inline bool
Fp8IsZero(fp8 num)
{
return num == 0;
}
/*
* Convert a fp8 to a float4
*/
static inline float
Fp8ToFloat4(fp8 num)
{
/* Lookup table for non-sign bits */
/* Uses uint32 for correctness */
uint32 lookup[128] = {0, 989855744, 998244352, 1002438656, 1006632960, 1008730112, 1010827264, 1012924416, 1015021568, 1016070144, 1017118720, 1018167296, 1019215872, 1020264448, 1021313024, 1022361600, 1023410176, 1024458752, 1025507328, 1026555904, 1027604480, 1028653056, 1029701632, 1030750208, 1031798784, 1032847360, 1033895936, 1034944512, 1035993088, 1037041664, 1038090240, 1039138816, 1040187392, 1041235968, 1042284544, 1043333120, 1044381696, 1045430272, 1046478848, 1047527424, 1048576000, 1049624576, 1050673152, 1051721728, 1052770304, 1053818880, 1054867456, 1055916032, 1056964608, 1058013184, 1059061760, 1060110336, 1061158912, 1062207488, 1063256064, 1064304640, 1065353216, 1066401792, 1067450368, 1068498944, 1069547520, 1070596096, 1071644672, 1072693248, 1073741824, 1074790400, 1075838976, 1076887552, 1077936128, 1078984704, 1080033280, 1081081856, 1082130432, 1083179008, 1084227584, 1085276160, 1086324736, 1087373312, 1088421888, 1089470464, 1090519040, 1091567616, 1092616192, 1093664768, 1094713344, 1095761920, 1096810496, 1097859072, 1098907648, 1099956224, 1101004800, 1102053376, 1103101952, 1104150528, 1105199104, 1106247680, 1107296256, 1108344832, 1109393408, 1110441984, 1111490560, 1112539136, 1113587712, 1114636288, 1115684864, 1116733440, 1117782016, 1118830592, 1119879168, 1120927744, 1121976320, 1123024896, 1124073472, 1125122048, 1126170624, 1127219200, 1128267776, 1129316352, 1130364928, 1131413504, 1132462080, 1133510656, 1134559232, 1135607808, 1136656384, 1137704960, 1138753536, 2146435072};
union
{
float f;
uint32 i;
} swap;
swap.i = lookup[num & 0x7F];
return (num & 0x80) == 0x80 ? -swap.f : swap.f;
}
/*
* Convert a float4 to a fp8
*/
static inline fp8
Float4ToFp8Unchecked(float num)
{
union
{
float f;
uint32 i;
} swap;
uint32 bin;
int exponent;
int mantissa;
uint8 result;
swap.f = num;
bin = swap.i;
exponent = (bin & 0x7F800000) >> 23;
mantissa = bin & 0x007FFFFF;
/* Sign */
result = (bin & 0x80000000) >> 24;
if (isinf(num) || isnan(num))
{
/* NaN */
result |= 0x7F;
}
else if (exponent > 114)
{
int m;
int gr;
int s;
exponent -= 127;
s = mantissa & 0x0007FFFF;
/* Subnormal */
if (exponent < -6)
{
int diff = -exponent - 6;
mantissa >>= diff;
mantissa += 1 << (23 - diff);
s |= mantissa & 0x0007FFFF;
}
m = mantissa >> 20;
/* Round */
gr = (mantissa >> 19) % 4;
if (gr == 3 || (gr == 1 && s != 0))
m += 1;
if (m == 8)
{
m = 0;
exponent += 1;
}
if (exponent > 8)
{
/* Infinite, which is NaN */
result |= 0x7F;
}
else
{
if (exponent >= -6)
result |= (exponent + 7) << 3;
result |= m;
}
}
return result;
}
/*
* Convert a float4 to a fp8
*/
static inline fp8
Float4ToFp8(float num)
{
fp8 result = Float4ToFp8Unchecked(num);
if (unlikely(Fp8IsNan(result)) && !isinf(num))
{
char *buf = palloc(FLOAT_SHORTEST_DECIMAL_LEN);
float_to_shortest_decimal_buf(num, buf);
ereport(ERROR,
(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
errmsg("\"%s\" is out of range for type minivec", buf)));
}
return result;
}
#endif

View File

@@ -3,6 +3,7 @@
#include <limits.h>
#include <math.h>
#include "catalog/pg_type.h"
#include "common/string.h"
#include "fmgr.h"
#include "halfutils.h"
@@ -11,6 +12,7 @@
#include "sparsevec.h"
#include "utils/array.h"
#include "utils/builtins.h"
#include "utils/lsyscache.h"
#include "vector.h"
#if PG_VERSION_NUM >= 120000
@@ -670,6 +672,137 @@ halfvec_to_sparsevec(PG_FUNCTION_ARGS)
PG_RETURN_POINTER(result);
}
/*
* Convert array to sparse vector
*/
FUNCTION_PREFIX PG_FUNCTION_INFO_V1(array_to_sparsevec);
Datum
array_to_sparsevec(PG_FUNCTION_ARGS)
{
ArrayType *array = PG_GETARG_ARRAYTYPE_P(0);
int32 typmod = PG_GETARG_INT32(1);
SparseVector *result;
int16 typlen;
bool typbyval;
char typalign;
Datum *elemsp;
int nelemsp;
int nnz = 0;
float *values;
int j = 0;
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);
#ifdef _MSC_VER
/* /fp:fast may not propagate +/-Infinity or NaN */
#define IS_NOT_ZERO(v) (isnan((float) (v)) || isinf((float) (v)) || ((float) (v)) != 0)
#else
#define IS_NOT_ZERO(v) (((float) (v)) != 0)
#endif
if (ARR_ELEMTYPE(array) == INT4OID)
{
for (int i = 0; i < nelemsp; i++)
nnz += IS_NOT_ZERO(DatumGetInt32(elemsp[i]));
}
else if (ARR_ELEMTYPE(array) == FLOAT8OID)
{
for (int i = 0; i < nelemsp; i++)
nnz += IS_NOT_ZERO(DatumGetFloat8(elemsp[i]));
}
else if (ARR_ELEMTYPE(array) == FLOAT4OID)
{
for (int i = 0; i < nelemsp; i++)
nnz += IS_NOT_ZERO(DatumGetFloat4(elemsp[i]));
}
else if (ARR_ELEMTYPE(array) == NUMERICOID)
{
for (int i = 0; i < nelemsp; i++)
nnz += IS_NOT_ZERO(DirectFunctionCall1(numeric_float4, elemsp[i]));
}
else
{
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("unsupported array type")));
}
result = InitSparseVector(nelemsp, nnz);
values = SPARSEVEC_VALUES(result);
#define PROCESS_ARRAY_ELEM(elem) \
do { \
float v = (float) (elem); \
if (IS_NOT_ZERO(v)) { \
/* Safety check */ \
if (j >= result->nnz) \
elog(ERROR, "safety check failed"); \
result->indices[j] = i; \
values[j] = v; \
j++; \
} \
} while (0)
if (ARR_ELEMTYPE(array) == INT4OID)
{
for (int i = 0; i < nelemsp; i++)
PROCESS_ARRAY_ELEM(DatumGetInt32(elemsp[i]));
}
else if (ARR_ELEMTYPE(array) == FLOAT8OID)
{
for (int i = 0; i < nelemsp; i++)
PROCESS_ARRAY_ELEM(DatumGetFloat8(elemsp[i]));
}
else if (ARR_ELEMTYPE(array) == FLOAT4OID)
{
for (int i = 0; i < nelemsp; i++)
PROCESS_ARRAY_ELEM(DatumGetFloat4(elemsp[i]));
}
else if (ARR_ELEMTYPE(array) == NUMERICOID)
{
for (int i = 0; i < nelemsp; i++)
PROCESS_ARRAY_ELEM(DatumGetFloat4(DirectFunctionCall1(numeric_float4, elemsp[i])));
}
else
{
ereport(ERROR,
(errcode(ERRCODE_DATA_EXCEPTION),
errmsg("unsupported array type")));
}
#undef PROCESS_ARRAY_ELEM
#undef IS_NOT_ZERO
/*
* Free allocation from deconstruct_array. Do not free individual elements
* when pass-by-reference since they point to original array.
*/
pfree(elemsp);
if (j != result->nnz)
elog(ERROR, "correctness check failed");
/* Check elements */
for (int i = 0; i < result->nnz; i++)
CheckElement(values[i]);
PG_RETURN_POINTER(result);
}
/*
* Get the L2 squared distance between sparse vectors
*/

View File

@@ -13,6 +13,7 @@
#include "ivfflat.h"
#include "lib/stringinfo.h"
#include "libpq/pqformat.h"
#include "minivec.h"
#include "port.h" /* for strtof() */
#include "sparsevec.h"
#include "utils/array.h"
@@ -26,11 +27,6 @@
#include "varatt.h"
#endif
#if PG_VERSION_NUM < 130000
#define TYPALIGN_DOUBLE 'd'
#define TYPALIGN_INT 'i'
#endif
#define STATE_DIMS(x) (ARR_DIMS(x)[0] - 1)
#define CreateStateDatums(dim) palloc(sizeof(Datum) * (dim + 1))
@@ -160,24 +156,6 @@ CheckStateArray(ArrayType *statearray, const char *caller)
return (float8 *) ARR_DATA_PTR(statearray);
}
#if PG_VERSION_NUM < 120003
static pg_noinline void
float_overflow_error(void)
{
ereport(ERROR,
(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
errmsg("value out of range: overflow")));
}
static pg_noinline void
float_underflow_error(void)
{
ereport(ERROR,
(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
errmsg("value out of range: underflow")));
}
#endif
/*
* Convert textual representation to internal representation
*/
@@ -565,6 +543,28 @@ halfvec_to_vector(PG_FUNCTION_ARGS)
PG_RETURN_POINTER(result);
}
/*
* Convert fp8 vector to vector
*/
FUNCTION_PREFIX PG_FUNCTION_INFO_V1(minivec_to_vector);
Datum
minivec_to_vector(PG_FUNCTION_ARGS)
{
MiniVector *vec = PG_GETARG_MINIVEC_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] = Fp8ToFloat4(vec->x[i]);
PG_RETURN_POINTER(result);
}
VECTOR_TARGET_CLONES static float
VectorL2SquaredDistance(int dim, float *ax, float *bx)
{

View File

@@ -38,6 +38,26 @@ SELECT * FROM t ORDER BY val;
(4 rows)
DROP TABLE t;
-- minivec
CREATE TABLE t (val minivec(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE INDEX ON t (val);
SELECT * FROM t WHERE val = '[1,2,3]';
val
---------
[1,2,3]
(1 row)
SELECT * FROM t ORDER BY val;
val
---------
[0,0,0]
[1,1,1]
[1,2,3]
(4 rows)
DROP TABLE t;
-- sparsevec
CREATE TABLE t (val sparsevec(3));

View File

@@ -140,6 +140,64 @@ SELECT '{1e-8,-1e-8}'::real[]::halfvec;
[0,-0]
(1 row)
SELECT '[1,2,3]'::vector::minivec;
minivec
---------
[1,2,3]
(1 row)
SELECT '[1,2,3]'::vector::minivec(3);
minivec
---------
[1,2,3]
(1 row)
SELECT '[1,2,3]'::vector::minivec(2);
ERROR: expected 2 dimensions, not 3
SELECT '[465]'::vector::minivec;
ERROR: "465" is out of range for type minivec
SELECT '[1e-8]'::vector::minivec;
minivec
---------
[0]
(1 row)
SELECT '[1,2,3]'::minivec::vector;
vector
---------
[1,2,3]
(1 row)
SELECT '[1,2,3]'::minivec::vector(3);
vector
---------
[1,2,3]
(1 row)
SELECT '[1,2,3]'::minivec::vector(2);
ERROR: expected 2 dimensions, not 3
SELECT '{1,2,3}'::real[]::minivec;
minivec
---------
[1,2,3]
(1 row)
SELECT '{1,2,3}'::real[]::minivec(3);
minivec
---------
[1,2,3]
(1 row)
SELECT '{1,2,3}'::real[]::minivec(2);
ERROR: expected 2 dimensions, not 3
SELECT '{465,-465}'::real[]::minivec;
ERROR: "465" is out of range for type minivec
SELECT '{1e-8,-1e-8}'::real[]::minivec;
minivec
---------
[0,-0]
(1 row)
SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec;
sparsevec
-----------------
@@ -208,6 +266,62 @@ SELECT '{1:1e-8}/1'::sparsevec::halfvec;
[0]
(1 row)
SELECT ARRAY[1,0,2,0,3,0]::sparsevec;
array
-----------------
{1:1,3:2,5:3}/6
(1 row)
SELECT ARRAY[1.0,0.0,2.0,0.0,3.0,0.0]::sparsevec;
array
-----------------
{1:1,3:2,5:3}/6
(1 row)
SELECT ARRAY[1,0,2,0,3,0]::float4[]::sparsevec;
array
-----------------
{1:1,3:2,5:3}/6
(1 row)
SELECT ARRAY[1,0,2,0,3,0]::float8[]::sparsevec;
array
-----------------
{1:1,3:2,5:3}/6
(1 row)
SELECT ARRAY[1,0,2,0,3,0]::numeric[]::sparsevec;
array
-----------------
{1:1,3:2,5:3}/6
(1 row)
SELECT '{1,0,2,0,3,0}'::real[]::sparsevec;
sparsevec
-----------------
{1:1,3:2,5:3}/6
(1 row)
SELECT '{1,0,2,0,3,0}'::real[]::sparsevec(6);
sparsevec
-----------------
{1:1,3:2,5:3}/6
(1 row)
SELECT '{1,0,2,0,3,0}'::real[]::sparsevec(5);
ERROR: expected 5 dimensions, not 6
SELECT '{NULL}'::real[]::sparsevec;
ERROR: array must not contain nulls
SELECT '{NaN}'::real[]::sparsevec;
ERROR: NaN not allowed in sparsevec
SELECT '{Infinity}'::real[]::sparsevec;
ERROR: infinite value not allowed in sparsevec
SELECT '{-Infinity}'::real[]::sparsevec;
ERROR: infinite value not allowed in sparsevec
SELECT '{}'::real[]::sparsevec;
ERROR: sparsevec must have at least 1 dimension
SELECT '{{1}}'::real[]::sparsevec;
ERROR: array must be 1-D
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

@@ -30,6 +30,23 @@ SELECT * FROM t2 ORDER BY val;
(4 rows)
DROP TABLE t;
DROP TABLE t2;
-- minivec
CREATE TABLE t (val minivec(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE TABLE t2 (val minivec(3));
\copy t TO 'results/minivec.bin' WITH (FORMAT binary)
\copy t2 FROM 'results/minivec.bin' WITH (FORMAT binary)
SELECT * FROM t2 ORDER BY val;
val
---------
[0,0,0]
[1,1,1]
[1,2,3]
(4 rows)
DROP TABLE t;
DROP TABLE t2;
-- sparsevec

View File

@@ -0,0 +1,84 @@
SET enable_seqscan = off;
-- L2
CREATE TABLE t (val minivec(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE INDEX ON t USING ivfflat (val minivec_l2_ops) WITH (lists = 1);
INSERT INTO t (val) VALUES ('[1,2,4]');
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
val
---------
[1,2,3]
[1,2,4]
[1,1,1]
[0,0,0]
(4 rows)
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::minivec)) t2;
count
-------
4
(1 row)
SELECT COUNT(*) FROM t;
count
-------
5
(1 row)
TRUNCATE t;
NOTICE: ivfflat index created with little data
DETAIL: This will cause low recall.
HINT: Drop the index until the table has more data.
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
val
-----
(0 rows)
DROP TABLE t;
-- inner product
CREATE TABLE t (val minivec(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE INDEX ON t USING ivfflat (val minivec_ip_ops) WITH (lists = 1);
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::minivec)) t2;
count
-------
4
(1 row)
DROP TABLE t;
-- cosine
CREATE TABLE t (val minivec(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE INDEX ON t USING ivfflat (val minivec_cosine_ops) WITH (lists = 1);
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::minivec)) t2;
count
-------
3
(1 row)
DROP TABLE t;

588
test/expected/minivec.out Normal file
View File

@@ -0,0 +1,588 @@
SELECT '[1,2,3]'::minivec;
minivec
---------
[1,2,3]
(1 row)
SELECT '[-1,-2,-3]'::minivec;
minivec
------------
[-1,-2,-3]
(1 row)
SELECT '[1.,2.,3.]'::minivec;
minivec
---------
[1,2,3]
(1 row)
SELECT ' [ 1, 2 , 3 ] '::minivec;
minivec
---------
[1,2,3]
(1 row)
SELECT '[1.23456]'::minivec;
minivec
---------
[1.25]
(1 row)
SELECT '[hello,1]'::minivec;
ERROR: invalid input syntax for type minivec: "[hello,1]"
LINE 1: SELECT '[hello,1]'::minivec;
^
SELECT '[NaN,1]'::minivec;
ERROR: NaN not allowed in minivec
LINE 1: SELECT '[NaN,1]'::minivec;
^
SELECT '[Infinity,1]'::minivec;
ERROR: "Infinity" is out of range for type minivec
LINE 1: SELECT '[Infinity,1]'::minivec;
^
SELECT '[-Infinity,1]'::minivec;
ERROR: "-Infinity" is out of range for type minivec
LINE 1: SELECT '[-Infinity,1]'::minivec;
^
SELECT '[65519,-65519]'::minivec;
ERROR: "65519" is out of range for type minivec
LINE 1: SELECT '[65519,-65519]'::minivec;
^
SELECT '[65520,-65520]'::minivec;
ERROR: "65520" is out of range for type minivec
LINE 1: SELECT '[65520,-65520]'::minivec;
^
SELECT '[1e-8,-1e-8]'::minivec;
minivec
---------
[0,-0]
(1 row)
SELECT '[4e38,1]'::minivec;
ERROR: "4e38" is out of range for type minivec
LINE 1: SELECT '[4e38,1]'::minivec;
^
SELECT '[1e-46,1]'::minivec;
minivec
---------
[0,1]
(1 row)
SELECT '[1,2,3'::minivec;
ERROR: invalid input syntax for type minivec: "[1,2,3"
LINE 1: SELECT '[1,2,3'::minivec;
^
SELECT '[1,2,3]9'::minivec;
ERROR: invalid input syntax for type minivec: "[1,2,3]9"
LINE 1: SELECT '[1,2,3]9'::minivec;
^
DETAIL: Junk after closing right brace.
SELECT '1,2,3'::minivec;
ERROR: invalid input syntax for type minivec: "1,2,3"
LINE 1: SELECT '1,2,3'::minivec;
^
DETAIL: Vector contents must start with "[".
SELECT ''::minivec;
ERROR: invalid input syntax for type minivec: ""
LINE 1: SELECT ''::minivec;
^
DETAIL: Vector contents must start with "[".
SELECT '['::minivec;
ERROR: invalid input syntax for type minivec: "["
LINE 1: SELECT '['::minivec;
^
SELECT '[ '::minivec;
ERROR: invalid input syntax for type minivec: "[ "
LINE 1: SELECT '[ '::minivec;
^
SELECT '[,'::minivec;
ERROR: invalid input syntax for type minivec: "[,"
LINE 1: SELECT '[,'::minivec;
^
SELECT '[]'::minivec;
ERROR: minivec must have at least 1 dimension
LINE 1: SELECT '[]'::minivec;
^
SELECT '[ ]'::minivec;
ERROR: minivec must have at least 1 dimension
LINE 1: SELECT '[ ]'::minivec;
^
SELECT '[,]'::minivec;
ERROR: invalid input syntax for type minivec: "[,]"
LINE 1: SELECT '[,]'::minivec;
^
SELECT '[1,]'::minivec;
ERROR: invalid input syntax for type minivec: "[1,]"
LINE 1: SELECT '[1,]'::minivec;
^
SELECT '[1a]'::minivec;
ERROR: invalid input syntax for type minivec: "[1a]"
LINE 1: SELECT '[1a]'::minivec;
^
SELECT '[1,,3]'::minivec;
ERROR: invalid input syntax for type minivec: "[1,,3]"
LINE 1: SELECT '[1,,3]'::minivec;
^
SELECT '[1, ,3]'::minivec;
ERROR: invalid input syntax for type minivec: "[1, ,3]"
LINE 1: SELECT '[1, ,3]'::minivec;
^
SELECT '[1,2,3]'::minivec(3);
minivec
---------
[1,2,3]
(1 row)
SELECT '[1,2,3]'::minivec(2);
ERROR: expected 2 dimensions, not 3
SELECT '[1,2,3]'::minivec(3, 2);
ERROR: invalid type modifier
LINE 1: SELECT '[1,2,3]'::minivec(3, 2);
^
SELECT '[1,2,3]'::minivec('a');
ERROR: invalid input syntax for type integer: "a"
LINE 1: SELECT '[1,2,3]'::minivec('a');
^
SELECT '[1,2,3]'::minivec(0);
ERROR: dimensions for type minivec must be at least 1
LINE 1: SELECT '[1,2,3]'::minivec(0);
^
SELECT '[1,2,3]'::minivec(16001);
ERROR: dimensions for type minivec cannot exceed 16000
LINE 1: SELECT '[1,2,3]'::minivec(16001);
^
SELECT unnest('{"[1,2,3]", "[4,5,6]"}'::minivec[]);
unnest
---------
[1,2,3]
[4,5,6]
(2 rows)
SELECT '{"[1,2,3]"}'::minivec(2)[];
ERROR: expected 2 dimensions, not 3
SELECT '[1,2,3]'::minivec + '[4,5,6]';
?column?
----------
[5,7,9]
(1 row)
SELECT '[448]'::minivec + '[448]';
ERROR: value out of range: overflow
SELECT '[1,2]'::minivec + '[3]';
ERROR: different minivec dimensions 2 and 1
SELECT '[1,2,3]'::minivec - '[4,5,6]';
?column?
------------
[-3,-3,-3]
(1 row)
SELECT '[-448]'::minivec - '[448]';
ERROR: value out of range: overflow
SELECT '[1,2]'::minivec - '[3]';
ERROR: different minivec dimensions 2 and 1
SELECT '[1,2,3]'::minivec * '[4,5,6]';
?column?
-----------
[4,10,18]
(1 row)
SELECT '[448]'::minivec * '[448]';
ERROR: value out of range: overflow
SELECT '[1e-7]'::minivec * '[1e-7]';
?column?
----------
[0]
(1 row)
SELECT '[1,2]'::minivec * '[3]';
ERROR: different minivec dimensions 2 and 1
SELECT '[1,2,3]'::minivec || '[4,5]';
?column?
-------------
[1,2,3,4,5]
(1 row)
SELECT array_fill(0, ARRAY[16000])::minivec || '[1]';
ERROR: minivec cannot have more than 16000 dimensions
SELECT '[1,2,3]'::minivec < '[1,2,3]';
?column?
----------
f
(1 row)
SELECT '[1,2,3]'::minivec < '[1,2]';
?column?
----------
f
(1 row)
SELECT '[1,2,3]'::minivec <= '[1,2,3]';
?column?
----------
t
(1 row)
SELECT '[1,2,3]'::minivec <= '[1,2]';
?column?
----------
f
(1 row)
SELECT '[1,2,3]'::minivec = '[1,2,3]';
?column?
----------
t
(1 row)
SELECT '[1,2,3]'::minivec = '[1,2]';
?column?
----------
f
(1 row)
SELECT '[1,2,3]'::minivec != '[1,2,3]';
?column?
----------
f
(1 row)
SELECT '[1,2,3]'::minivec != '[1,2]';
?column?
----------
t
(1 row)
SELECT '[1,2,3]'::minivec >= '[1,2,3]';
?column?
----------
t
(1 row)
SELECT '[1,2,3]'::minivec >= '[1,2]';
?column?
----------
t
(1 row)
SELECT '[1,2,3]'::minivec > '[1,2,3]';
?column?
----------
f
(1 row)
SELECT '[1,2,3]'::minivec > '[1,2]';
?column?
----------
t
(1 row)
SELECT minivec_cmp('[1,2,3]', '[1,2,3]');
minivec_cmp
-------------
0
(1 row)
SELECT minivec_cmp('[1,2,3]', '[0,0,0]');
minivec_cmp
-------------
1
(1 row)
SELECT minivec_cmp('[0,0,0]', '[1,2,3]');
minivec_cmp
-------------
-1
(1 row)
SELECT minivec_cmp('[1,2]', '[1,2,3]');
minivec_cmp
-------------
-1
(1 row)
SELECT minivec_cmp('[1,2,3]', '[1,2]');
minivec_cmp
-------------
1
(1 row)
SELECT minivec_cmp('[1,2]', '[2,3,4]');
minivec_cmp
-------------
-1
(1 row)
SELECT minivec_cmp('[2,3]', '[1,2,3]');
minivec_cmp
-------------
1
(1 row)
SELECT vector_dims('[1,2,3]'::minivec);
vector_dims
-------------
3
(1 row)
SELECT round(l2_norm('[1,1]'::minivec)::numeric, 5);
round
---------
1.41421
(1 row)
SELECT l2_norm('[3,4]'::minivec);
l2_norm
---------
5
(1 row)
SELECT l2_norm('[0,1]'::minivec);
l2_norm
---------
1
(1 row)
SELECT l2_norm('[0,0]'::minivec);
l2_norm
---------
0
(1 row)
SELECT l2_norm('[2]'::minivec);
l2_norm
---------
2
(1 row)
SELECT l2_distance('[0,0]'::minivec, '[3,4]');
l2_distance
-------------
5
(1 row)
SELECT l2_distance('[0,0]'::minivec, '[0,1]');
l2_distance
-------------
1
(1 row)
SELECT l2_distance('[1,2]'::minivec, '[3]');
ERROR: different minivec dimensions 2 and 1
SELECT l2_distance('[1,1,1,1,1,1,1,1,1]'::minivec, '[1,1,1,1,1,1,1,4,5]');
l2_distance
-------------
5
(1 row)
SELECT '[0,0]'::minivec <-> '[3,4]';
?column?
----------
5
(1 row)
SELECT inner_product('[1,2]'::minivec, '[3,4]');
inner_product
---------------
11
(1 row)
SELECT inner_product('[1,2]'::minivec, '[3]');
ERROR: different minivec dimensions 2 and 1
SELECT inner_product('[448]'::minivec, '[448]');
inner_product
---------------
200704
(1 row)
SELECT inner_product('[1,1,1,1,1,1,1,1,1]'::minivec, '[1,2,3,4,5,6,7,8,9]');
inner_product
---------------
45
(1 row)
SELECT '[1,2]'::minivec <#> '[3,4]';
?column?
----------
-11
(1 row)
SELECT cosine_distance('[1,2]'::minivec, '[2,4]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,2]'::minivec, '[0,0]');
cosine_distance
-----------------
NaN
(1 row)
SELECT cosine_distance('[1,1]'::minivec, '[1,1]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,0]'::minivec, '[0,2]');
cosine_distance
-----------------
1
(1 row)
SELECT cosine_distance('[1,1]'::minivec, '[-1,-1]');
cosine_distance
-----------------
2
(1 row)
SELECT cosine_distance('[1,2]'::minivec, '[3]');
ERROR: different minivec dimensions 2 and 1
SELECT cosine_distance('[1,1]'::minivec, '[1.1,1.1]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,1]'::minivec, '[-1.1,-1.1]');
cosine_distance
-----------------
2
(1 row)
SELECT cosine_distance('[1,2,3,4,5,6,7,8,9]'::minivec, '[1,2,3,4,5,6,7,8,9]');
cosine_distance
-----------------
0
(1 row)
SELECT cosine_distance('[1,2,3,4,5,6,7,8,9]'::minivec, '[-1,-2,-3,-4,-5,-6,-7,-8,-9]');
cosine_distance
-----------------
2
(1 row)
SELECT '[1,2]'::minivec <=> '[2,4]';
?column?
----------
0
(1 row)
SELECT l1_distance('[0,0]'::minivec, '[3,4]');
l1_distance
-------------
7
(1 row)
SELECT l1_distance('[0,0]'::minivec, '[0,1]');
l1_distance
-------------
1
(1 row)
SELECT l1_distance('[1,2]'::minivec, '[3]');
ERROR: different minivec dimensions 2 and 1
SELECT l1_distance('[1,2,3,4,5,6,7,8,9]'::minivec, '[1,2,3,4,5,6,7,8,9]');
l1_distance
-------------
0
(1 row)
SELECT l1_distance('[1,2,3,4,5,6,7,8,9]'::minivec, '[0,3,2,5,4,7,6,9,8]');
l1_distance
-------------
9
(1 row)
SELECT '[0,0]'::minivec <+> '[3,4]';
?column?
----------
7
(1 row)
SELECT l2_normalize('[3,4]'::minivec);
l2_normalize
----------------
[0.625,0.8125]
(1 row)
SELECT l2_normalize('[3,0]'::minivec);
l2_normalize
--------------
[1,0]
(1 row)
SELECT l2_normalize('[0,0.1]'::minivec);
l2_normalize
--------------
[0,1]
(1 row)
SELECT l2_normalize('[0,0]'::minivec);
l2_normalize
--------------
[0,0]
(1 row)
SELECT l2_normalize('[448]'::minivec);
l2_normalize
--------------
[1]
(1 row)
SELECT binary_quantize('[1,0,-1]'::minivec);
binary_quantize
-----------------
100
(1 row)
SELECT binary_quantize('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]'::minivec);
binary_quantize
-----------------
01001110101
(1 row)
SELECT subvector('[1,2,3,4,5]'::minivec, 1, 3);
subvector
-----------
[1,2,3]
(1 row)
SELECT subvector('[1,2,3,4,5]'::minivec, 3, 2);
subvector
-----------
[3,4]
(1 row)
SELECT subvector('[1,2,3,4,5]'::minivec, -1, 3);
subvector
-----------
[1]
(1 row)
SELECT subvector('[1,2,3,4,5]'::minivec, 3, 9);
subvector
-----------
[3,4,5]
(1 row)
SELECT subvector('[1,2,3,4,5]'::minivec, 1, 0);
ERROR: minivec must have at least 1 dimension
SELECT subvector('[1,2,3,4,5]'::minivec, 3, -1);
ERROR: minivec must have at least 1 dimension
SELECT subvector('[1,2,3,4,5]'::minivec, -1, 2);
ERROR: minivec must have at least 1 dimension
SELECT subvector('[1,2,3,4,5]'::minivec, 2147483647, 10);
ERROR: minivec must have at least 1 dimension
SELECT subvector('[1,2,3,4,5]'::minivec, 3, 2147483647);
subvector
-----------
[3,4,5]
(1 row)
SELECT subvector('[1,2,3,4,5]'::minivec, -2147483644, 2147483647);
subvector
-----------
[1,2]
(1 row)

View File

@@ -22,6 +22,17 @@ SELECT * FROM t ORDER BY val;
DROP TABLE t;
-- minivec
CREATE TABLE t (val minivec(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE INDEX ON t (val);
SELECT * FROM t WHERE val = '[1,2,3]';
SELECT * FROM t ORDER BY val;
DROP TABLE t;
-- sparsevec
CREATE TABLE t (val sparsevec(3));

View File

@@ -38,6 +38,22 @@ SELECT '{1,2,3}'::real[]::halfvec(2);
SELECT '{65520,-65520}'::real[]::halfvec;
SELECT '{1e-8,-1e-8}'::real[]::halfvec;
SELECT '[1,2,3]'::vector::minivec;
SELECT '[1,2,3]'::vector::minivec(3);
SELECT '[1,2,3]'::vector::minivec(2);
SELECT '[465]'::vector::minivec;
SELECT '[1e-8]'::vector::minivec;
SELECT '[1,2,3]'::minivec::vector;
SELECT '[1,2,3]'::minivec::vector(3);
SELECT '[1,2,3]'::minivec::vector(2);
SELECT '{1,2,3}'::real[]::minivec;
SELECT '{1,2,3}'::real[]::minivec(3);
SELECT '{1,2,3}'::real[]::minivec(2);
SELECT '{465,-465}'::real[]::minivec;
SELECT '{1e-8,-1e-8}'::real[]::minivec;
SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec;
SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec(5);
SELECT '[0,1.5,0,3.5,0]'::vector::sparsevec(4);
@@ -58,6 +74,22 @@ SELECT '{}/16001'::sparsevec::halfvec;
SELECT '{1:65520}/1'::sparsevec::halfvec;
SELECT '{1:1e-8}/1'::sparsevec::halfvec;
SELECT ARRAY[1,0,2,0,3,0]::sparsevec;
SELECT ARRAY[1.0,0.0,2.0,0.0,3.0,0.0]::sparsevec;
SELECT ARRAY[1,0,2,0,3,0]::float4[]::sparsevec;
SELECT ARRAY[1,0,2,0,3,0]::float8[]::sparsevec;
SELECT ARRAY[1,0,2,0,3,0]::numeric[]::sparsevec;
SELECT '{1,0,2,0,3,0}'::real[]::sparsevec;
SELECT '{1,0,2,0,3,0}'::real[]::sparsevec(6);
SELECT '{1,0,2,0,3,0}'::real[]::sparsevec(5);
SELECT '{NULL}'::real[]::sparsevec;
SELECT '{NaN}'::real[]::sparsevec;
SELECT '{Infinity}'::real[]::sparsevec;
SELECT '{-Infinity}'::real[]::sparsevec;
SELECT '{}'::real[]::sparsevec;
SELECT '{{1}}'::real[]::sparsevec;
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

@@ -28,6 +28,21 @@ SELECT * FROM t2 ORDER BY val;
DROP TABLE t;
DROP TABLE t2;
-- minivec
CREATE TABLE t (val minivec(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE TABLE t2 (val minivec(3));
\copy t TO 'results/minivec.bin' WITH (FORMAT binary)
\copy t2 FROM 'results/minivec.bin' WITH (FORMAT binary)
SELECT * FROM t2 ORDER BY val;
DROP TABLE t;
DROP TABLE t2;
-- sparsevec
CREATE TABLE t (val sparsevec(3));

View File

@@ -0,0 +1,45 @@
SET enable_seqscan = off;
-- L2
CREATE TABLE t (val minivec(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE INDEX ON t USING ivfflat (val minivec_l2_ops) WITH (lists = 1);
INSERT INTO t (val) VALUES ('[1,2,4]');
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::minivec)) t2;
SELECT COUNT(*) FROM t;
TRUNCATE t;
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
DROP TABLE t;
-- inner product
CREATE TABLE t (val minivec(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE INDEX ON t USING ivfflat (val minivec_ip_ops) WITH (lists = 1);
INSERT INTO t (val) VALUES ('[1,2,4]');
SELECT * FROM t ORDER BY val <#> '[3,3,3]';
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <#> (SELECT NULL::minivec)) t2;
DROP TABLE t;
-- cosine
CREATE TABLE t (val minivec(3));
INSERT INTO t (val) VALUES ('[0,0,0]'), ('[1,2,3]'), ('[1,1,1]'), (NULL);
CREATE INDEX ON t USING ivfflat (val minivec_cosine_ops) WITH (lists = 1);
INSERT INTO t (val) VALUES ('[1,2,4]');
SELECT * FROM t ORDER BY val <=> '[3,3,3]';
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> '[0,0,0]') t2;
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <=> (SELECT NULL::minivec)) t2;
DROP TABLE t;

134
test/sql/minivec.sql Normal file
View File

@@ -0,0 +1,134 @@
SELECT '[1,2,3]'::minivec;
SELECT '[-1,-2,-3]'::minivec;
SELECT '[1.,2.,3.]'::minivec;
SELECT ' [ 1, 2 , 3 ] '::minivec;
SELECT '[1.23456]'::minivec;
SELECT '[hello,1]'::minivec;
SELECT '[NaN,1]'::minivec;
SELECT '[Infinity,1]'::minivec;
SELECT '[-Infinity,1]'::minivec;
SELECT '[65519,-65519]'::minivec;
SELECT '[65520,-65520]'::minivec;
SELECT '[1e-8,-1e-8]'::minivec;
SELECT '[4e38,1]'::minivec;
SELECT '[1e-46,1]'::minivec;
SELECT '[1,2,3'::minivec;
SELECT '[1,2,3]9'::minivec;
SELECT '1,2,3'::minivec;
SELECT ''::minivec;
SELECT '['::minivec;
SELECT '[ '::minivec;
SELECT '[,'::minivec;
SELECT '[]'::minivec;
SELECT '[ ]'::minivec;
SELECT '[,]'::minivec;
SELECT '[1,]'::minivec;
SELECT '[1a]'::minivec;
SELECT '[1,,3]'::minivec;
SELECT '[1, ,3]'::minivec;
SELECT '[1,2,3]'::minivec(3);
SELECT '[1,2,3]'::minivec(2);
SELECT '[1,2,3]'::minivec(3, 2);
SELECT '[1,2,3]'::minivec('a');
SELECT '[1,2,3]'::minivec(0);
SELECT '[1,2,3]'::minivec(16001);
SELECT unnest('{"[1,2,3]", "[4,5,6]"}'::minivec[]);
SELECT '{"[1,2,3]"}'::minivec(2)[];
SELECT '[1,2,3]'::minivec + '[4,5,6]';
SELECT '[448]'::minivec + '[448]';
SELECT '[1,2]'::minivec + '[3]';
SELECT '[1,2,3]'::minivec - '[4,5,6]';
SELECT '[-448]'::minivec - '[448]';
SELECT '[1,2]'::minivec - '[3]';
SELECT '[1,2,3]'::minivec * '[4,5,6]';
SELECT '[448]'::minivec * '[448]';
SELECT '[1e-7]'::minivec * '[1e-7]';
SELECT '[1,2]'::minivec * '[3]';
SELECT '[1,2,3]'::minivec || '[4,5]';
SELECT array_fill(0, ARRAY[16000])::minivec || '[1]';
SELECT '[1,2,3]'::minivec < '[1,2,3]';
SELECT '[1,2,3]'::minivec < '[1,2]';
SELECT '[1,2,3]'::minivec <= '[1,2,3]';
SELECT '[1,2,3]'::minivec <= '[1,2]';
SELECT '[1,2,3]'::minivec = '[1,2,3]';
SELECT '[1,2,3]'::minivec = '[1,2]';
SELECT '[1,2,3]'::minivec != '[1,2,3]';
SELECT '[1,2,3]'::minivec != '[1,2]';
SELECT '[1,2,3]'::minivec >= '[1,2,3]';
SELECT '[1,2,3]'::minivec >= '[1,2]';
SELECT '[1,2,3]'::minivec > '[1,2,3]';
SELECT '[1,2,3]'::minivec > '[1,2]';
SELECT minivec_cmp('[1,2,3]', '[1,2,3]');
SELECT minivec_cmp('[1,2,3]', '[0,0,0]');
SELECT minivec_cmp('[0,0,0]', '[1,2,3]');
SELECT minivec_cmp('[1,2]', '[1,2,3]');
SELECT minivec_cmp('[1,2,3]', '[1,2]');
SELECT minivec_cmp('[1,2]', '[2,3,4]');
SELECT minivec_cmp('[2,3]', '[1,2,3]');
SELECT vector_dims('[1,2,3]'::minivec);
SELECT round(l2_norm('[1,1]'::minivec)::numeric, 5);
SELECT l2_norm('[3,4]'::minivec);
SELECT l2_norm('[0,1]'::minivec);
SELECT l2_norm('[0,0]'::minivec);
SELECT l2_norm('[2]'::minivec);
SELECT l2_distance('[0,0]'::minivec, '[3,4]');
SELECT l2_distance('[0,0]'::minivec, '[0,1]');
SELECT l2_distance('[1,2]'::minivec, '[3]');
SELECT l2_distance('[1,1,1,1,1,1,1,1,1]'::minivec, '[1,1,1,1,1,1,1,4,5]');
SELECT '[0,0]'::minivec <-> '[3,4]';
SELECT inner_product('[1,2]'::minivec, '[3,4]');
SELECT inner_product('[1,2]'::minivec, '[3]');
SELECT inner_product('[448]'::minivec, '[448]');
SELECT inner_product('[1,1,1,1,1,1,1,1,1]'::minivec, '[1,2,3,4,5,6,7,8,9]');
SELECT '[1,2]'::minivec <#> '[3,4]';
SELECT cosine_distance('[1,2]'::minivec, '[2,4]');
SELECT cosine_distance('[1,2]'::minivec, '[0,0]');
SELECT cosine_distance('[1,1]'::minivec, '[1,1]');
SELECT cosine_distance('[1,0]'::minivec, '[0,2]');
SELECT cosine_distance('[1,1]'::minivec, '[-1,-1]');
SELECT cosine_distance('[1,2]'::minivec, '[3]');
SELECT cosine_distance('[1,1]'::minivec, '[1.1,1.1]');
SELECT cosine_distance('[1,1]'::minivec, '[-1.1,-1.1]');
SELECT cosine_distance('[1,2,3,4,5,6,7,8,9]'::minivec, '[1,2,3,4,5,6,7,8,9]');
SELECT cosine_distance('[1,2,3,4,5,6,7,8,9]'::minivec, '[-1,-2,-3,-4,-5,-6,-7,-8,-9]');
SELECT '[1,2]'::minivec <=> '[2,4]';
SELECT l1_distance('[0,0]'::minivec, '[3,4]');
SELECT l1_distance('[0,0]'::minivec, '[0,1]');
SELECT l1_distance('[1,2]'::minivec, '[3]');
SELECT l1_distance('[1,2,3,4,5,6,7,8,9]'::minivec, '[1,2,3,4,5,6,7,8,9]');
SELECT l1_distance('[1,2,3,4,5,6,7,8,9]'::minivec, '[0,3,2,5,4,7,6,9,8]');
SELECT '[0,0]'::minivec <+> '[3,4]';
SELECT l2_normalize('[3,4]'::minivec);
SELECT l2_normalize('[3,0]'::minivec);
SELECT l2_normalize('[0,0.1]'::minivec);
SELECT l2_normalize('[0,0]'::minivec);
SELECT l2_normalize('[448]'::minivec);
SELECT binary_quantize('[1,0,-1]'::minivec);
SELECT binary_quantize('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]'::minivec);
SELECT subvector('[1,2,3,4,5]'::minivec, 1, 3);
SELECT subvector('[1,2,3,4,5]'::minivec, 3, 2);
SELECT subvector('[1,2,3,4,5]'::minivec, -1, 3);
SELECT subvector('[1,2,3,4,5]'::minivec, 3, 9);
SELECT subvector('[1,2,3,4,5]'::minivec, 1, 0);
SELECT subvector('[1,2,3,4,5]'::minivec, 3, -1);
SELECT subvector('[1,2,3,4,5]'::minivec, -1, 2);
SELECT subvector('[1,2,3,4,5]'::minivec, 2147483647, 10);
SELECT subvector('[1,2,3,4,5]'::minivec, 3, 2147483647);
SELECT subvector('[1,2,3,4,5]'::minivec, -2147483644, 2147483647);

View File

@@ -40,6 +40,10 @@ for (1 .. 50)
$actual = $node->safe_psql("postgres", "SELECT halfvec_cmp(v::halfvec, '$query'::real[]::halfvec) FROM tst");
is($expected, $actual);
# Test minivec
$actual = $node->safe_psql("postgres", "SELECT minivec_cmp(v::minivec, '$query'::real[]::minivec) FROM tst");
is($expected, $actual);
# Test sparsevec
$actual = $node->safe_psql("postgres", "SELECT sparsevec_cmp(v::vector::sparsevec, '$query'::real[]::vector::sparsevec) FROM tst");
is($expected, $actual);

View File

@@ -45,6 +45,10 @@ for my $function (@functions)
my $actual = $node->safe_psql("postgres", "SELECT $function(v::halfvec, '$query'::vector::halfvec) FROM tst");
is($expected, $actual, "halfvec $function");
# Test minivec
$actual = $node->safe_psql("postgres", "SELECT $function(v::minivec, '$query'::vector::minivec) FROM tst");
is($expected, $actual, "minivec $function");
# Test sparsevec
$actual = $node->safe_psql("postgres", "SELECT $function(v::sparsevec, '$query'::vector::sparsevec) FROM tst");
is($expected, $actual, "sparsevec $function");

View File

@@ -12,8 +12,8 @@ $node->start;
# Create extension
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
my @types = ("vector", "halfvec", "sparsevec");
my @inputs = ("[1.23,4.56,7.89]", "[1.23,4.56,7.89]", "{1:1.23,2:4.56,3:7.89}/3");
my @types = ("vector", "halfvec", "minivec", "sparsevec");
my @inputs = ("[1.23,4.56,7.89]", "[1.23,4.56,7.89]", "[1.23,4.56,7.89]", "{1:1.23,2:4.56,3:7.89}/3");
my @subs = (" ", " ", ",", ":", "-", "1", "9", "\0", "2147483648", "-2147483649");
for my $i (0 .. $#types)

View File

@@ -0,0 +1,132 @@
use strict;
use warnings FATAL => 'all';
use PostgreSQL::Test::Cluster;
use PostgreSQL::Test::Utils;
use Test::More;
my $node;
my @queries = ();
my @expected;
my $limit = 20;
my $dim = 10;
my $array_sql = join(",", ('2 * random() * 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 = PostgreSQL::Test::Cluster->new('node');
$node->init;
$node->start;
# Create table
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v minivec($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 = ("minivec_l2_ops", "minivec_ip_ops", "minivec_cosine_ops", "minivec_l1_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 $min = 0.98;
test_recall($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

@@ -0,0 +1,113 @@
use strict;
use warnings FATAL => 'all';
use PostgreSQL::Test::Cluster;
use PostgreSQL::Test::Utils;
use Test::More;
my $node;
my @queries = ();
my @expected;
my $limit = 20;
my $dim = 10;
my $array_sql = join(",", ('2 * random() * 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 = PostgreSQL::Test::Cluster->new('node');
$node->init;
$node->start;
# Create table
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
$node->safe_psql("postgres", "CREATE TABLE tst (i serial, v minivec($dim));");
# 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 = ("minivec_l2_ops", "minivec_ip_ops", "minivec_cosine_ops", "minivec_l1_ops");
for my $i (0 .. $#operators)
{
my $operator = $operators[$i];
my $opclass = $opclasses[$i];
# Add index
$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING hnsw (v $opclass);");
# Use concurrent inserts
$node->pgbench(
"--no-vacuum --client=10 --transactions=1000",
0,
[qr{actually processed}],
[qr{^$}],
"concurrent INSERTs",
{
"040_hnsw_minivec_insert_recall_$opclass" => "INSERT INTO tst (v) VALUES (ARRAY[$array_sql]);"
}
);
# Get exact results
@expected = ();
foreach (@queries)
{
my $res = $node->safe_psql("postgres", qq(
SET enable_indexscan = off;
SELECT i FROM tst ORDER BY v $operator '$_' LIMIT $limit;
));
push(@expected, $res);
}
# Test approximate results
my $min = 0.98;
test_recall($min, $operator);
$node->safe_psql("postgres", "DROP INDEX idx;");
$node->safe_psql("postgres", "TRUNCATE tst;");
}
done_testing();

View File

@@ -0,0 +1,97 @@
use strict;
use warnings FATAL => 'all';
use PostgreSQL::Test::Cluster;
use PostgreSQL::Test::Utils;
use Test::More;
my $node;
my @queries = ();
my @expected;
my $limit = 20;
sub test_recall
{
my ($min, $ef_search, $test_name) = @_;
my $correct = 0;
my $total = 0;
my $explain = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SET hnsw.ef_search = $ef_search;
EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v <-> '$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;
SET hnsw.ef_search = $ef_search;
SELECT i FROM tst ORDER BY v <-> '$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, $test_name);
}
# Initialize node
$node = PostgreSQL::Test::Cluster->new('node');
$node->init;
$node->start;
# Create table
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v minivec(3));");
$node->safe_psql("postgres", "ALTER TABLE tst SET (autovacuum_enabled = false);");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT i, ARRAY[random(), random(), random()] FROM generate_series(1, 10000) i;"
);
# Add index
$node->safe_psql("postgres", "CREATE INDEX ON tst USING hnsw (v minivec_l2_ops) WITH (m = 4, ef_construction = 8);");
# Delete data
$node->safe_psql("postgres", "DELETE FROM tst WHERE i > 2500;");
# Generate queries
for (1 .. 20)
{
my $r1 = rand();
my $r2 = rand();
my $r3 = rand();
push(@queries, "[$r1,$r2,$r3]");
}
# Get exact results
@expected = ();
foreach (@queries)
{
my $res = $node->safe_psql("postgres", qq(
SET enable_indexscan = off;
SELECT i FROM tst ORDER BY v <-> '$_' LIMIT $limit;
));
push(@expected, $res);
}
test_recall(0.18, $limit, "before vacuum");
test_recall(0.93, 100, "before vacuum");
# TODO Test concurrent inserts with vacuum
$node->safe_psql("postgres", "VACUUM tst;");
test_recall(0.95, $limit, "after vacuum");
done_testing();

View File

@@ -0,0 +1,58 @@
use strict;
use warnings FATAL => 'all';
use PostgreSQL::Test::Cluster;
use PostgreSQL::Test::Utils;
use Test::More;
# Initialize node
my $node = PostgreSQL::Test::Cluster->new('node');
$node->init;
$node->start;
# Create table
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
$node->safe_psql("postgres", "CREATE TABLE tst (v minivec(3));");
sub insert_vectors
{
for my $i (1 .. 20)
{
$node->safe_psql("postgres", "INSERT INTO tst VALUES ('[1,1,1]');");
}
}
sub test_duplicates
{
my $res = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SET hnsw.ef_search = 1;
SELECT COUNT(*) FROM (SELECT * FROM tst ORDER BY v <-> '[1,1,1]') t;
));
is($res, 10);
}
# Test duplicates with build
insert_vectors();
$node->safe_psql("postgres", "CREATE INDEX idx ON tst USING hnsw (v minivec_l2_ops);");
test_duplicates();
# Reset
$node->safe_psql("postgres", "TRUNCATE tst;");
# Test duplicates with inserts
insert_vectors();
test_duplicates();
# Test fallback path for inserts
$node->pgbench(
"--no-vacuum --client=5 --transactions=100",
0,
[qr{actually processed}],
[qr{^$}],
"concurrent INSERTs",
{
"042_hnsw_minivec_duplicates" => "INSERT INTO tst VALUES ('[1,1,1]');"
}
);
done_testing();

View File

@@ -0,0 +1,154 @@
use strict;
use warnings FATAL => 'all';
use PostgreSQL::Test::Cluster;
use PostgreSQL::Test::Utils;
use Test::More;
my $node;
my @queries = ();
my @expected;
my $limit = 20;
my $dim = 10;
my $array_sql = join(",", ('random()') x $dim);
sub test_recall
{
my ($probes, $min, $operator) = @_;
my $correct = 0;
my $total = 0;
my $explain = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SET ivfflat.probes = $probes;
EXPLAIN ANALYZE SELECT i FROM tst ORDER BY v $operator '$queries[0]' LIMIT $limit;
));
like($explain, qr/Index Scan using idx on tst/);
for my $i (0 .. $#queries)
{
my $actual = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SET ivfflat.probes = $probes;
SELECT i FROM tst ORDER BY v $operator '$queries[$i]' LIMIT $limit;
));
my @actual_ids = split("\n", $actual);
my @expected_ids = split("\n", $expected[$i]);
my %expected_set = map { $_ => 1 } @expected_ids;
foreach (@actual_ids)
{
if (exists($expected_set{$_}))
{
$correct++;
}
}
$total += $limit;
}
cmp_ok($correct / $total, ">=", $min, $operator);
}
# Initialize node
$node = PostgreSQL::Test::Cluster->new('node');
$node->init;
$node->start;
# Create table
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v minivec($dim));");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT i, ARRAY[$array_sql] FROM generate_series(1, 100000) 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 = ("minivec_l2_ops", "minivec_ip_ops", "minivec_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", qq(
WITH top AS (
SELECT v $operator '$_' AS distance FROM tst ORDER BY distance LIMIT $limit
)
SELECT i FROM tst WHERE (v $operator '$_') <= (SELECT MAX(distance) FROM top)
));
push(@expected, $res);
}
# Build index serially
$node->safe_psql("postgres", qq(
SET max_parallel_maintenance_workers = 0;
CREATE INDEX idx ON tst USING ivfflat (v $opclass);
));
# Test approximate results
if ($operator ne "<#>")
{
# TODO Fix test (uniform random vectors all have similar inner product)
test_recall(1, 0.33, $operator);
test_recall(10, 0.93, $operator);
}
# Test probes equals lists
if ($operator eq "<=>")
{
test_recall(100, 0.98, $operator);
}
else
{
test_recall(100, 1.00, $operator);
}
$node->safe_psql("postgres", "DROP INDEX idx;");
# Build index in parallel
my ($ret, $stdout, $stderr) = $node->psql("postgres", qq(
SET client_min_messages = DEBUG;
SET min_parallel_table_scan_size = 1;
CREATE INDEX idx ON tst USING ivfflat (v $opclass);
));
is($ret, 0, $stderr);
like($stderr, qr/using \d+ parallel workers/);
# Test approximate results
if ($operator ne "<#>")
{
# TODO Fix test (uniform random vectors all have similar inner product)
test_recall(1, 0.33, $operator);
test_recall(10, 0.93, $operator);
}
# Test probes equals lists
if ($operator eq "<=>")
{
test_recall(100, 0.98, $operator);
}
else
{
test_recall(100, 1.00, $operator);
}
$node->safe_psql("postgres", "DROP INDEX idx;");
}
done_testing();