Compare commits

..

30 Commits

Author SHA1 Message Date
Andrew Kane
2b125a1956 Moved bit functions to separate file 2024-03-27 13:42:24 -07:00
Andrew Kane
fcd655d2a3 Improved test [skip ci] 2024-03-26 19:26:52 -07:00
Andrew Kane
97fe28940d Updated readme [skip ci] 2024-03-25 23:33:59 -07:00
Andrew Kane
23c5bf6ef6 Fixed CI [skip ci] 2024-03-25 23:22:03 -07:00
Andrew Kane
9ed4303a5e Fixed test 2024-03-25 23:16:38 -07:00
Andrew Kane
acd066463a Added Jaccard distance to test 2024-03-25 23:15:14 -07:00
Andrew Kane
be936075eb Improved test [skip ci] 2024-03-25 23:13:30 -07:00
Andrew Kane
a9959fede2 Improved code [skip ci] 2024-03-25 23:03:12 -07:00
Andrew Kane
02c4f4884c Added support for indexing Jaccard distance 2024-03-25 22:44:51 -07:00
Andrew Kane
791fc2436f Added jaccard_distance function 2024-03-25 22:35:53 -07:00
Andrew Kane
e7a7936bb2 Improved code [skip ci] 2024-03-25 19:23:30 -07:00
Andrew Kane
ce2ba65906 Fixed CI 2024-03-25 19:07:09 -07:00
Andrew Kane
023633a274 Added test for build recall 2024-03-25 19:02:37 -07:00
Andrew Kane
9baa051b5b Fixed CI 2024-03-25 17:20:35 -07:00
Andrew Kane
ac94ac7cf1 Updated changelog [skip ci] 2024-03-25 17:18:09 -07:00
Andrew Kane
8b819dfdc2 Moved bit code to separate files 2024-03-25 17:10:01 -07:00
Andrew Kane
d9ca850faf Updated max dimensions [skip ci] 2024-03-25 16:56:32 -07:00
Andrew Kane
131782999b Updated changelog [skip ci] 2024-03-25 16:47:25 -07:00
Andrew Kane
d57ef873c2 Improved test [skip ci] 2024-03-25 16:43:15 -07:00
Andrew Kane
30c86fb05a Improved test [skip ci] 2024-03-25 16:41:40 -07:00
Andrew Kane
833f379ebe Added quantize_binary function 2024-03-25 16:39:55 -07:00
Andrew Kane
709fc75ce0 Updated changelog [skip ci] 2024-03-25 16:11:52 -07:00
Andrew Kane
2bc959b3eb Updated readme [skip ci] 2024-03-25 15:53:44 -07:00
Andrew Kane
ec9e13b5fb Improved test [skip ci] 2024-03-25 15:50:38 -07:00
Andrew Kane
95e476d570 Fixed CI 2024-03-25 15:49:18 -07:00
Andrew Kane
6bc0c47a0a Updated comments [skip ci] 2024-03-25 15:43:05 -07:00
Andrew Kane
58eeefeef4 Handle nulls 2024-03-25 15:40:15 -07:00
Andrew Kane
263e684824 Started indexing 2024-03-25 15:25:04 -07:00
Andrew Kane
21dfed5719 Fixed CI 2024-03-25 14:24:19 -07:00
Andrew Kane
f3aec9fd03 Added hamming_distance function 2024-03-25 14:22:23 -07:00
33 changed files with 611 additions and 115 deletions

View File

@@ -1,3 +1,10 @@
## 0.7.0 (unreleased)
- Added support for binary vectors to HNSW
- Added `hamming_distance` function
- Added `jaccard_distance` function
- Added `quantize_binary` function
## 0.6.2 (2024-03-18)
- Reduced lock contention with parallel HNSW index builds

View File

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

View File

@@ -1,7 +1,7 @@
EXTENSION = vector
EXTVERSION = 0.6.2
OBJS = src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\vector.obj
OBJS = src\bitvector.obj src\hnsw.obj src\hnswbuild.obj src\hnswinsert.obj src\hnswscan.obj src\hnswutils.obj src\hnswvacuum.obj src\ivfbuild.obj src\ivfflat.obj src\ivfinsert.obj src\ivfkmeans.obj src\ivfscan.obj src\ivfutils.obj src\ivfvacuum.obj src\vector.obj
HEADERS = src\vector.h
REGRESS = btree cast copy functions input ivfflat_cosine ivfflat_ip ivfflat_l2 ivfflat_options ivfflat_unlogged

View File

@@ -5,7 +5,7 @@ Open-source vector similarity search for Postgres
Store your vectors with the rest of your data. Supports:
- exact and approximate nearest neighbor search
- L2 distance, inner product, and cosine distance
- L2 distance, inner product, cosine distance, and more
- any [language](#languages) with a Postgres client
Plus [ACID](https://en.wikipedia.org/wiki/ACID) compliance, point-in-time recovery, JOINs, and all of the other [great features](https://www.postgresql.org/about/) of Postgres
@@ -221,7 +221,19 @@ Cosine distance
CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops);
```
Vectors with up to 2,000 dimensions can be indexed.
Hamming distance - added in 0.7.0
```sql
CREATE INDEX ON items USING hnsw (embedding bit_hamming_ops);
```
Jaccard distance - added in 0.7.0
```sql
CREATE INDEX ON items USING hnsw (embedding bit_jaccard_ops);
```
Vectors with up to 2,000 dimensions can be indexed, or bit vectors with up to 64,000 dimensions.
### Index Options
@@ -415,19 +427,7 @@ You can use [Reciprocal Rank Fusion](https://github.com/pgvector/pgvector-python
### Tuning
Use a tool like [PgTune](https://pgtune.leopard.in.ua/) to set initial values for Postgres server parameters. For instance, `shared_buffers` should typically be 25% of the servers memory. You can find the config file with:
```sql
SHOW config_file;
```
And check individual settings with:
```sql
SHOW shared_buffers;
```
Be sure to restart Postgres for changes to take effect.
Use a tool like [PgTune](https://pgtune.leopard.in.ua/) to set initial values for Postgres server parameters.
### Loading
@@ -683,8 +683,6 @@ ALTER TABLE items ALTER COLUMN embedding SET STORAGE PLAIN;
Results are limited by the size of the dynamic candidate list (`hnsw.ef_search`). There may be even less results due to dead tuples or filtering conditions in the query. We recommend setting `hnsw.ef_search` to at least twice the `LIMIT` of the query. If you need more than 500 results, use an IVFFlat index instead.
Also, note that `NULL` vectors are not indexed (as well as zero vectors for cosine distance).
#### Why are there less results for a query after adding an IVFFlat index?
The index was likely created with too little data for the number of lists. Drop the index until the table has more data.
@@ -695,13 +693,11 @@ DROP INDEX index_name;
Results can also be limited by the number of probes (`ivfflat.probes`).
Also, note that `NULL` vectors are not indexed (as well as zero vectors for cosine distance).
## Reference
### Vector Type
Each vector takes `4 * dimensions + 8` bytes of storage. Each element is a single-precision floating-point number (like the `real` type in Postgres), and all elements must be finite (no `NaN`, `Infinity` or `-Infinity`). Vectors can have up to 16,000 dimensions.
Each vector takes `4 * dimensions + 8` bytes of storage. Each element is a single precision floating-point number (like the `real` type in Postgres), and all elements must be finite (no `NaN`, `Infinity` or `-Infinity`). Vectors can have up to 16,000 dimensions.
### Vector Operators
@@ -722,16 +718,31 @@ cosine_distance(vector, vector) → double precision | cosine distance |
inner_product(vector, vector) → double precision | inner product |
l2_distance(vector, vector) → double precision | Euclidean distance |
l1_distance(vector, vector) → double precision | taxicab distance | 0.5.0
quantize_binary(vector) → bit | quantize | 0.7.0
vector_dims(vector) → integer | number of dimensions |
vector_norm(vector) → double precision | Euclidean norm |
### Vector Aggregate Functions
### Aggregate Functions
Function | Description | Added
--- | --- | ---
avg(vector) → vector | average |
sum(vector) → vector | sum | 0.5.0
### Bit Operators
Operator | Description | Added
--- | --- | ---
<~> | Hamming distance | 0.7.0
<%> | Jaccard distance | 0.7.0
### Bit Functions
Function | Description | Added
--- | --- | ---
hamming_distance(bit, bit) → double precision | Hamming distance | 0.7.0
jaccard_distance(bit, bit) → double precision | Jaccard distance | 0.7.0
## Installation Notes - Linux and Mac
### Postgres Location

View File

@@ -0,0 +1,31 @@
-- complain if script is sourced in psql, rather than via CREATE EXTENSION
\echo Use "ALTER EXTENSION vector UPDATE TO '0.7.0'" to load this file. \quit
CREATE FUNCTION quantize_binary(vector) RETURNS bit
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION hamming_distance(bit, bit) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION jaccard_distance(bit, bit) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE OPERATOR <~> (
LEFTARG = bit, RIGHTARG = bit, PROCEDURE = hamming_distance,
COMMUTATOR = '<~>'
);
CREATE OPERATOR <%> (
LEFTARG = bit, RIGHTARG = bit, PROCEDURE = jaccard_distance,
COMMUTATOR = '<%>'
);
CREATE OPERATOR CLASS bit_hamming_ops
FOR TYPE bit USING hnsw AS
OPERATOR 1 <~> (bit, bit) FOR ORDER BY float_ops,
FUNCTION 1 hamming_distance(bit, bit);
CREATE OPERATOR CLASS bit_jaccard_ops
FOR TYPE bit USING hnsw AS
OPERATOR 1 <%> (bit, bit) FOR ORDER BY float_ops,
FUNCTION 1 jaccard_distance(bit, bit);

View File

@@ -1,7 +1,7 @@
-- complain if script is sourced in psql, rather than via CREATE EXTENSION
\echo Use "CREATE EXTENSION vector" to load this file. \quit
-- vector type
-- type
CREATE TYPE vector;
@@ -29,7 +29,7 @@ CREATE TYPE vector (
STORAGE = external
);
-- vector functions
-- functions
CREATE FUNCTION l2_distance(vector, vector) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
@@ -58,7 +58,10 @@ CREATE FUNCTION vector_sub(vector, vector) RETURNS vector
CREATE FUNCTION vector_mul(vector, vector) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- vector private functions
CREATE FUNCTION quantize_binary(vector) RETURNS bit
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- private functions
CREATE FUNCTION vector_lt(vector, vector) RETURNS bool
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
@@ -99,7 +102,7 @@ CREATE FUNCTION vector_avg(double precision[]) RETURNS vector
CREATE FUNCTION vector_combine(double precision[], double precision[]) RETURNS double precision[]
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- vector aggregates
-- aggregates
CREATE AGGREGATE avg(vector) (
SFUNC = vector_accum,
@@ -117,7 +120,7 @@ CREATE AGGREGATE sum(vector) (
PARALLEL = SAFE
);
-- vector cast functions
-- cast functions
CREATE FUNCTION vector(vector, integer, boolean) RETURNS vector
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
@@ -137,7 +140,7 @@ CREATE FUNCTION array_to_vector(numeric[], integer, boolean) RETURNS vector
CREATE FUNCTION vector_to_float4(vector, integer, boolean) RETURNS real[]
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
-- vector casts
-- casts
CREATE CAST (vector AS vector)
WITH FUNCTION vector(vector, integer, boolean) AS IMPLICIT;
@@ -157,7 +160,7 @@ CREATE CAST (double precision[] AS vector)
CREATE CAST (numeric[] AS vector)
WITH FUNCTION array_to_vector(numeric[], integer, boolean) AS ASSIGNMENT;
-- vector operators
-- operators
CREATE OPERATOR <-> (
LEFTARG = vector, RIGHTARG = vector, PROCEDURE = l2_distance,
@@ -240,7 +243,7 @@ CREATE ACCESS METHOD hnsw TYPE INDEX HANDLER hnswhandler;
COMMENT ON ACCESS METHOD hnsw IS 'hnsw index access method';
-- vector opclasses
-- opclasses
CREATE OPERATOR CLASS vector_ops
DEFAULT FOR TYPE vector USING btree AS
@@ -287,3 +290,31 @@ CREATE OPERATOR CLASS vector_cosine_ops
OPERATOR 1 <=> (vector, vector) FOR ORDER BY float_ops,
FUNCTION 1 vector_negative_inner_product(vector, vector),
FUNCTION 2 vector_norm(vector);
-- bit functions
CREATE FUNCTION hamming_distance(bit, bit) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE FUNCTION jaccard_distance(bit, bit) RETURNS float8
AS 'MODULE_PATHNAME' LANGUAGE C IMMUTABLE STRICT PARALLEL SAFE;
CREATE OPERATOR <~> (
LEFTARG = bit, RIGHTARG = bit, PROCEDURE = hamming_distance,
COMMUTATOR = '<~>'
);
CREATE OPERATOR <%> (
LEFTARG = bit, RIGHTARG = bit, PROCEDURE = jaccard_distance,
COMMUTATOR = '<%>'
);
CREATE OPERATOR CLASS bit_hamming_ops
FOR TYPE bit USING hnsw AS
OPERATOR 1 <~> (bit, bit) FOR ORDER BY float_ops,
FUNCTION 1 hamming_distance(bit, bit);
CREATE OPERATOR CLASS bit_jaccard_ops
FOR TYPE bit USING hnsw AS
OPERATOR 1 <%> (bit, bit) FOR ORDER BY float_ops,
FUNCTION 1 jaccard_distance(bit, bit);

90
src/bitvector.c Normal file
View File

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

8
src/bitvector.h Normal file
View File

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

View File

@@ -55,11 +55,6 @@
#define HNSW_UPDATE_ENTRY_GREATER 1
#define HNSW_UPDATE_ENTRY_ALWAYS 2
typedef enum HnswType
{
HNSW_TYPE_VECTOR
} HnswType;
/* Build phases */
/* PROGRESS_CREATEIDX_SUBPHASE_INITIALIZE is 1 */
#define PROGRESS_HNSW_PHASE_LOAD 2
@@ -247,7 +242,6 @@ typedef struct HnswBuildState
Relation index;
IndexInfo *indexInfo;
ForkNumber forkNum;
HnswType type;
/* Settings */
int dimensions;
@@ -268,6 +262,7 @@ typedef struct HnswBuildState
HnswGraph *graph;
double ml;
int maxLevel;
Vector *normvec;
/* Memory */
MemoryContext graphCtx;
@@ -372,8 +367,7 @@ typedef struct HnswVacuumState
int HnswGetM(Relation index);
int HnswGetEfConstruction(Relation index);
FmgrInfo *HnswOptionalProcInfo(Relation index, uint16 procnum);
HnswType HnswGetType(Relation index);
bool HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, HnswType type);
bool HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
Buffer HnswNewBuffer(Relation index, ForkNumber forkNum);
void HnswInitPage(Buffer buf, Page page);
void HnswInit(void);

View File

@@ -44,6 +44,7 @@
#include "access/xact.h"
#include "access/xloginsert.h"
#include "catalog/index.h"
#include "catalog/pg_type_d.h"
#include "commands/progress.h"
#include "hnsw.h"
#include "miscadmin.h"
@@ -489,7 +490,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heaptid, Hn
/* Normalize if needed */
if (buildstate->normprocinfo != NULL)
{
if (!HnswNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->type))
if (!HnswNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->normvec))
return false;
}
@@ -672,12 +673,15 @@ static void
InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, IndexInfo *indexInfo, ForkNumber forkNum)
{
int maxDimensions = HNSW_MAX_DIM;
Oid typid = TupleDescAttr(index->rd_att, 0)->atttypid;
if (typid == BITOID || typid == VARBITOID)
maxDimensions *= 32;
buildstate->heap = heap;
buildstate->index = index;
buildstate->indexInfo = indexInfo;
buildstate->forkNum = forkNum;
buildstate->type = HnswGetType(index);
buildstate->m = HnswGetM(index);
buildstate->efConstruction = HnswGetEfConstruction(index);
@@ -706,6 +710,9 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
buildstate->ml = HnswGetMl(buildstate->m);
buildstate->maxLevel = HnswGetMaxLevel(buildstate->m);
/* Reuse for each tuple */
buildstate->normvec = InitVector(buildstate->dimensions);
buildstate->graphCtx = GenerationContextCreate(CurrentMemoryContext,
"Hnsw build graph context",
#if PG_VERSION_NUM >= 150000
@@ -729,6 +736,7 @@ InitBuildState(HnswBuildState * buildstate, Relation heap, Relation index, Index
static void
FreeBuildState(HnswBuildState * buildstate)
{
pfree(buildstate->normvec);
MemoryContextDelete(buildstate->graphCtx);
MemoryContextDelete(buildstate->tmpCtx);
}

View File

@@ -622,7 +622,7 @@ HnswInsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_ti
normprocinfo = HnswOptionalProcInfo(index, HNSW_NORM_PROC);
if (normprocinfo != NULL)
{
if (!HnswNormValue(normprocinfo, collation, &value, HnswGetType(index)))
if (!HnswNormValue(normprocinfo, collation, &value, NULL))
return;
}

View File

@@ -1,6 +1,8 @@
#include "postgres.h"
#include "access/relscan.h"
#include "bitvector.h"
#include "catalog/pg_type_d.h"
#include "hnsw.h"
#include "pgstat.h"
#include "storage/bufmgr.h"
@@ -40,6 +42,29 @@ GetScanItems(IndexScanDesc scan, Datum q)
return HnswSearchLayer(base, q, ep, hnsw_ef_search, 0, index, procinfo, collation, m, false, NULL);
}
/*
* Get dimensions from metapage
*/
static int
GetDimensions(Relation index)
{
Buffer buf;
Page page;
HnswMetaPage metap;
int dimensions;
buf = ReadBuffer(index, HNSW_METAPAGE_BLKNO);
LockBuffer(buf, BUFFER_LOCK_SHARE);
page = BufferGetPage(buf);
metap = HnswPageGetMeta(page);
dimensions = metap->dimensions;
UnlockReleaseBuffer(buf);
return dimensions;
}
/*
* Get scan value
*/
@@ -50,7 +75,15 @@ GetScanValue(IndexScanDesc scan)
Datum value;
if (scan->orderByData->sk_flags & SK_ISNULL)
value = PointerGetDatum(NULL);
{
Oid typid = TupleDescAttr(scan->indexRelation->rd_att, 0)->atttypid;
int dimensions = GetDimensions(scan->indexRelation);
if (typid == BITOID || typid == VARBITOID)
value = PointerGetDatum(InitBitVector(dimensions));
else
value = PointerGetDatum(InitVector(dimensions));
}
else
{
value = scan->orderByData->sk_argument;
@@ -61,7 +94,7 @@ GetScanValue(IndexScanDesc scan)
/* Fine if normalization fails */
if (so->normprocinfo != NULL)
HnswNormValue(so->normprocinfo, so->collation, &value, HnswGetType(scan->indexRelation));
HnswNormValue(so->normprocinfo, so->collation, &value, NULL);
}
return value;

View File

@@ -149,15 +149,6 @@ HnswOptionalProcInfo(Relation index, uint16 procnum)
return index_getprocinfo(index, 1, procnum);
}
/*
* Get vector type
*/
HnswType
HnswGetType(Relation index)
{
return HNSW_TYPE_VECTOR;
}
/*
* Divide by the norm
*
@@ -167,25 +158,21 @@ HnswGetType(Relation index)
* if it's different than the original value
*/
bool
HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, HnswType type)
HnswNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result)
{
double norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
if (norm > 0)
{
/* TODO Remove vector-specific code */
if (type == HNSW_TYPE_VECTOR)
{
Vector *v = DatumGetVector(*value);
Vector *result = InitVector(v->dim);
Vector *v = DatumGetVector(*value);
for (int i = 0; i < v->dim; i++)
result->x[i] = v->x[i] / norm;
if (result == NULL)
result = InitVector(v->dim);
*value = PointerGetDatum(result);
}
else
elog(ERROR, "Unsupported type");
for (int i = 0; i < v->dim; i++)
result->x[i] = v->x[i] / norm;
*value = PointerGetDatum(result);
return true;
}
@@ -588,12 +575,7 @@ HnswLoadElement(HnswElement element, float *distance, Datum *q, Relation index,
/* Calculate distance */
if (distance != NULL)
{
if (DatumGetPointer(*q) == NULL)
*distance = 0;
else
*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->data)));
}
*distance = (float) DatumGetFloat8(FunctionCall2Coll(procinfo, collation, *q, PointerGetDatum(&etup->data)));
UnlockReleaseBuffer(buf);
}

View File

@@ -57,7 +57,7 @@ AddSample(Datum *values, IvfflatBuildState * buildstate)
*/
if (buildstate->kmeansnormprocinfo != NULL)
{
if (!IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->collation, &value))
if (!IvfflatNormValue(buildstate->kmeansnormprocinfo, buildstate->collation, &value, buildstate->normvec))
return;
}
@@ -105,7 +105,7 @@ SampleCallback(Relation index, CALLBACK_ITEM_POINTER, Datum *values,
oldCtx = MemoryContextSwitchTo(buildstate->tmpCtx);
/* Add sample */
AddSample(values, buildstate);
AddSample(values, state);
/* Reset memory context */
MemoryContextSwitchTo(oldCtx);
@@ -153,7 +153,7 @@ AddTupleToSort(Relation index, ItemPointer tid, Datum *values, IvfflatBuildState
/* Normalize if needed */
if (buildstate->normprocinfo != NULL)
{
if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value))
if (!IvfflatNormValue(buildstate->normprocinfo, buildstate->collation, &value, buildstate->normvec))
return;
}
@@ -356,6 +356,9 @@ InitBuildState(IvfflatBuildState * buildstate, Relation heap, Relation index, In
buildstate->centers = VectorArrayInit(buildstate->lists, buildstate->dimensions);
buildstate->listInfo = palloc(sizeof(ListInfo) * buildstate->lists);
/* Reuse for each tuple */
buildstate->normvec = InitVector(buildstate->dimensions);
buildstate->tmpCtx = AllocSetContextCreate(CurrentMemoryContext,
"Ivfflat build temporary context",
ALLOCSET_DEFAULT_SIZES);
@@ -377,6 +380,7 @@ FreeBuildState(IvfflatBuildState * buildstate)
{
VectorArrayFree(buildstate->centers);
pfree(buildstate->listInfo);
pfree(buildstate->normvec);
#ifdef IVFFLAT_KMEANS_DEBUG
pfree(buildstate->listSums);

View File

@@ -172,6 +172,7 @@ typedef struct IvfflatBuildState
VectorArray samples;
VectorArray centers;
ListInfo *listInfo;
Vector *normvec;
#ifdef IVFFLAT_KMEANS_DEBUG
double inertia;
@@ -266,7 +267,7 @@ void VectorArrayFree(VectorArray arr);
void PrintVectorArray(char *msg, VectorArray arr);
void IvfflatKmeans(Relation index, VectorArray samples, VectorArray centers);
FmgrInfo *IvfflatOptionalProcInfo(Relation index, uint16 procnum);
bool IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value);
bool IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result);
int IvfflatGetLists(Relation index);
void IvfflatGetMetaPageInfo(Relation index, int *lists, int *dimensions);
void IvfflatUpdateList(Relation index, ListInfo listInfo, BlockNumber insertPage, BlockNumber originalInsertPage, BlockNumber startPage, ForkNumber forkNum);

View File

@@ -85,7 +85,7 @@ InsertTuple(Relation index, Datum *values, bool *isnull, ItemPointer heap_tid, R
normprocinfo = IvfflatOptionalProcInfo(index, IVFFLAT_NORM_PROC);
if (normprocinfo != NULL)
{
if (!IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value))
if (!IvfflatNormValue(normprocinfo, index->rd_indcollation[0], &value, NULL))
return;
}

View File

@@ -293,7 +293,7 @@ ivfflatgettuple(IndexScanDesc scan, ScanDirection dir)
/* Fine if normalization fails */
if (so->normprocinfo != NULL)
IvfflatNormValue(so->normprocinfo, so->collation, &value);
IvfflatNormValue(so->normprocinfo, so->collation, &value, NULL);
}
IvfflatBench("GetScanLists", GetScanLists(scan, value));

View File

@@ -75,14 +75,16 @@ IvfflatOptionalProcInfo(Relation index, uint16 procnum)
* if it's different than the original value
*/
bool
IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value)
IvfflatNormValue(FmgrInfo *procinfo, Oid collation, Datum *value, Vector * result)
{
double norm = DatumGetFloat8(FunctionCall1Coll(procinfo, collation, *value));
if (norm > 0)
{
Vector *v = DatumGetVector(*value);
Vector *result = InitVector(v->dim);
if (result == NULL)
result = InitVector(v->dim);
for (int i = 0; i < v->dim; i++)
result->x[i] = v->x[i] / norm;

View File

@@ -2,6 +2,7 @@
#include <math.h>
#include "bitvector.h"
#include "catalog/pg_type.h"
#include "common/shortest_dec.h"
#include "fmgr.h"
@@ -197,8 +198,6 @@ vector_in(PG_FUNCTION_ARGS)
while (pt != NULL && *stringEnd != ']')
{
float val;
if (dim == VECTOR_MAX_DIM)
ereport(ERROR,
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED),
@@ -214,23 +213,15 @@ vector_in(PG_FUNCTION_ARGS)
errmsg("invalid input syntax for type vector: \"%s\"", lit)));
/* Use strtof like float4in to avoid a double-rounding problem */
errno = 0;
val = strtof(pt, &stringEnd);
x[dim] = strtof(pt, &stringEnd);
CheckElement(x[dim]);
dim++;
if (stringEnd == pt)
ereport(ERROR,
(errcode(ERRCODE_INVALID_TEXT_REPRESENTATION),
errmsg("invalid input syntax for type vector: \"%s\"", lit)));
/* Check for range error like float4in */
if (errno == ERANGE && (val == 0 || isinf(val)))
ereport(ERROR,
(errcode(ERRCODE_NUMERIC_VALUE_OUT_OF_RANGE),
errmsg("\"%s\" is out of range for type vector", pt)));
CheckElement(val);
x[dim++] = val;
while (vector_isspace(*stringEnd))
stringEnd++;
@@ -870,6 +861,26 @@ vector_mul(PG_FUNCTION_ARGS)
PG_RETURN_POINTER(result);
}
/*
* Quantize a vector
*/
PGDLLEXPORT PG_FUNCTION_INFO_V1(quantize_binary);
Datum
quantize_binary(PG_FUNCTION_ARGS)
{
Vector *a = PG_GETARG_VECTOR_P(0);
float *ax = a->x;
VarBit *result = InitBitVector(a->dim);
unsigned char *rx = VARBITS(result);
/* TODO Improve */
for (int i = 0; i < a->dim; i++)
rx[i / 8] |= (ax[i] > 0) << (7 - (i % 8));
PG_RETURN_VARBIT_P(result);
}
/*
* Internal helper to compare vectors
*/

View File

@@ -0,0 +1,64 @@
SELECT hamming_distance(B'111', B'111');
hamming_distance
------------------
0
(1 row)
SELECT hamming_distance(B'111', B'110');
hamming_distance
------------------
1
(1 row)
SELECT hamming_distance(B'111', B'100');
hamming_distance
------------------
2
(1 row)
SELECT hamming_distance(B'111', B'000');
hamming_distance
------------------
3
(1 row)
SELECT hamming_distance(B'111', B'00');
ERROR: different bit lengths 3 and 2
SELECT jaccard_distance(B'1111', B'1111');
jaccard_distance
------------------
0
(1 row)
SELECT jaccard_distance(B'1111', B'1110');
jaccard_distance
------------------
0.25
(1 row)
SELECT jaccard_distance(B'1111', B'1100');
jaccard_distance
------------------
0.5
(1 row)
SELECT jaccard_distance(B'1111', B'1000');
jaccard_distance
------------------
0.75
(1 row)
SELECT jaccard_distance(B'1111', B'0000');
jaccard_distance
------------------
1
(1 row)
SELECT jaccard_distance(B'1100', B'1000');
jaccard_distance
------------------
0.5
(1 row)
SELECT jaccard_distance(B'1111', B'000');
ERROR: different bit lengths 4 and 3

View File

@@ -208,6 +208,18 @@ SELECT l1_distance('[3e38]', '[-3e38]');
Infinity
(1 row)
SELECT quantize_binary('[1,0,-1]');
quantize_binary
-----------------
100
(1 row)
SELECT quantize_binary('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]');
quantize_binary
-----------------
01001110101
(1 row)
SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
avg
-----------

View File

@@ -0,0 +1,21 @@
SET enable_seqscan = off;
CREATE TABLE t (val bit(3));
INSERT INTO t (val) VALUES (B'000'), (B'100'), (B'111'), (NULL);
CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
INSERT INTO t (val) VALUES (B'110');
SELECT * FROM t ORDER BY val <~> B'111';
val
-----
111
110
100
000
(4 rows)
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <~> (SELECT NULL::bit)) t2;
count
-------
4
(1 row)
DROP TABLE t;

View File

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

View File

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

View File

@@ -63,21 +63,9 @@ SELECT '[1.5e-38,-1.5e-38]'::vector;
(1 row)
SELECT '[4e38,1]'::vector;
ERROR: "4e38" is out of range for type vector
ERROR: infinite value not allowed in vector
LINE 1: SELECT '[4e38,1]'::vector;
^
SELECT '[-4e38,1]'::vector;
ERROR: "-4e38" is out of range for type vector
LINE 1: SELECT '[-4e38,1]'::vector;
^
SELECT '[1e-46,1]'::vector;
ERROR: "1e-46" is out of range for type vector
LINE 1: SELECT '[1e-46,1]'::vector;
^
SELECT '[-1e-46,1]'::vector;
ERROR: "-1e-46" is out of range for type vector
LINE 1: SELECT '[-1e-46,1]'::vector;
^
SELECT '[1,2,3'::vector;
ERROR: malformed vector literal: "[1,2,3"
LINE 1: SELECT '[1,2,3'::vector;

View File

@@ -0,0 +1,13 @@
SELECT hamming_distance(B'111', B'111');
SELECT hamming_distance(B'111', B'110');
SELECT hamming_distance(B'111', B'100');
SELECT hamming_distance(B'111', B'000');
SELECT hamming_distance(B'111', B'00');
SELECT jaccard_distance(B'1111', B'1111');
SELECT jaccard_distance(B'1111', B'1110');
SELECT jaccard_distance(B'1111', B'1100');
SELECT jaccard_distance(B'1111', B'1000');
SELECT jaccard_distance(B'1111', B'0000');
SELECT jaccard_distance(B'1100', B'1000');
SELECT jaccard_distance(B'1111', B'000');

View File

@@ -48,6 +48,9 @@ SELECT l1_distance('[0,0]', '[0,1]');
SELECT l1_distance('[1,2]', '[3]');
SELECT l1_distance('[3e38]', '[-3e38]');
SELECT quantize_binary('[1,0,-1]');
SELECT quantize_binary('[0,0.1,-0.2,-0.3,0.4,0.5,0.6,-0.7,0.8,-0.9,1]');
SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]']) v;
SELECT avg(v) FROM unnest(ARRAY['[1,2,3]'::vector, '[3,5,7]', NULL]) v;
SELECT avg(v) FROM unnest(ARRAY[]::vector[]) v;

12
test/sql/hnsw_hamming.sql Normal file
View File

@@ -0,0 +1,12 @@
SET enable_seqscan = off;
CREATE TABLE t (val bit(3));
INSERT INTO t (val) VALUES (B'000'), (B'100'), (B'111'), (NULL);
CREATE INDEX ON t USING hnsw (val bit_hamming_ops);
INSERT INTO t (val) VALUES (B'110');
SELECT * FROM t ORDER BY val <~> B'111';
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <~> (SELECT NULL::bit)) t2;
DROP TABLE t;

12
test/sql/hnsw_jaccard.sql Normal file
View File

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

View File

@@ -7,7 +7,7 @@ CREATE INDEX ON t USING hnsw (val vector_l2_ops);
INSERT INTO t (val) VALUES ('[1,2,4]');
SELECT * FROM t ORDER BY val <-> '[3,3,3]';
SELECT COUNT(*) FROM (SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector)) t2;
SELECT * FROM t ORDER BY val <-> (SELECT NULL::vector);
SELECT COUNT(*) FROM t;
TRUNCATE t;

View File

@@ -11,9 +11,6 @@ SELECT '[1.5e38,-1.5e38]'::vector;
SELECT '[1.5e+38,-1.5e+38]'::vector;
SELECT '[1.5e-38,-1.5e-38]'::vector;
SELECT '[4e38,1]'::vector;
SELECT '[-4e38,1]'::vector;
SELECT '[1e-46,1]'::vector;
SELECT '[-1e-46,1]'::vector;
SELECT '[1,2,3'::vector;
SELECT '[1,2,3]9'::vector;
SELECT '1,2,3'::vector;

View File

@@ -86,7 +86,7 @@ foreach (@queries)
push(@expected, $res);
}
test_recall(0.19, $limit, "before vacuum");
test_recall(0.20, $limit, "before vacuum");
test_recall(0.95, 100, "before vacuum");
# TODO Test concurrent inserts with vacuum

View File

@@ -0,0 +1,137 @@
use strict;
use warnings;
use PostgresNode;
use TestLib;
use Test::More;
my $node;
my @queries = ();
my @expected;
my $limit = 20;
my $dim = 52;
my $max = 2**$dim;
sub test_recall
{
my ($min, $operator) = @_;
my $correct = 0;
my $total = 0;
my $explain = $node->safe_psql("postgres", qq(
SET enable_seqscan = off;
SET hnsw.ef_search = 100;
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;
SET hnsw.ef_search = 100;
SELECT i FROM tst ORDER BY v $operator $queries[$i] LIMIT $limit;
));
my @actual_ids = split("\n", $actual);
my @expected_ids = split("\n", $expected[$i]);
my %expected_set = map { $_ => 1 } @expected_ids;
foreach (@actual_ids)
{
if (exists($expected_set{$_}))
{
$correct++;
}
}
$total += $limit;
}
cmp_ok($correct / $total, ">=", $min, $operator);
}
# Initialize node
$node = get_new_node('node');
$node->init;
$node->start;
# Create table
$node->safe_psql("postgres", "CREATE EXTENSION vector;");
$node->safe_psql("postgres", "CREATE TABLE tst (i int4, v bit($dim));");
$node->safe_psql("postgres",
"INSERT INTO tst SELECT i, (random() * $max)::bigint::bit($dim) FROM generate_series(1, 10000) i;"
);
# Generate queries
for (1 .. 20)
{
my $r = int(rand() * $max);
push(@queries, "${r}::bigint::bit($dim)");
}
# Check each index type
my @operators = ("<~>", "<\%>");
my @opclasses = ("bit_hamming_ops", "bit_jaccard_ops");
for my $i (0 .. $#operators)
{
my $operator = $operators[$i];
my $opclass = $opclasses[$i];
# Get exact results
@expected = ();
foreach (@queries)
{
# Handle ties
my $res = $node->safe_psql("postgres", qq(
WITH top AS (
SELECT v $operator $_ AS distance FROM tst ORDER BY v $operator $_ 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 hnsw (v $opclass);
));
# Test approximate results
my $min = $operator eq "<\%>" ? 0.96 : 0.99;
test_recall($min, $operator);
$node->safe_psql("postgres", "DROP INDEX idx;");
# Build index in parallel in memory
my ($ret, $stdout, $stderr) = $node->psql("postgres", qq(
SET client_min_messages = DEBUG;
SET min_parallel_table_scan_size = 1;
CREATE INDEX idx ON tst USING hnsw (v $opclass);
));
is($ret, 0, $stderr);
like($stderr, qr/using \d+ parallel workers/);
# Test approximate results
test_recall($min, $operator);
$node->safe_psql("postgres", "DROP INDEX idx;");
# Build index in parallel on disk
# Set parallel_workers on table to use workers with low maintenance_work_mem
($ret, $stdout, $stderr) = $node->psql("postgres", qq(
ALTER TABLE tst SET (parallel_workers = 2);
SET client_min_messages = DEBUG;
SET maintenance_work_mem = '4MB';
CREATE INDEX idx ON tst USING hnsw (v $opclass);
ALTER TABLE tst RESET (parallel_workers);
));
is($ret, 0, $stderr);
like($stderr, qr/using \d+ parallel workers/);
like($stderr, qr/hnsw graph no longer fits into maintenance_work_mem/);
$node->safe_psql("postgres", "DROP INDEX idx;");
}
done_testing();