tweaks

src/utils/bloom1_sparse_index_params.h

@@ -5,6 +5,16 @@
  */
 #pragma once
 
-#define BLOOM1_HASHES 3
+#include <common/hashfn.h>
+
+#define BLOOM1_HASHES 4
 #define BLOOM1_SEED_1 0x71d924af
 #define BLOOM1_SEED_2 0xba48b314
+
+static inline uint32
+bloom1_get_one_hash(uint32 value_hash, uint32 index)
+{
+	const uint32 h1 = hash_combine(value_hash, BLOOM1_SEED_1);
+	const uint32 h2 = hash_combine(value_hash, BLOOM1_SEED_2);
+	return h1 + index * h2 + index * index;
+}

src/utils/ts_bloom1_matches.c

@@ -20,34 +20,28 @@ TS_FUNCTION_INFO_V1(ts_bloom1_matches);
 Datum
 ts_bloom1_matches(PG_FUNCTION_ARGS)
 {
-	bytea *bloom = PG_GETARG_VARLENA_PP(0);
-	Datum val = PG_GETARG_DATUM(1);
-
-	const int nbits = VARSIZE_ANY_EXHDR(bloom) * 8;
-
 	Oid val_type = get_fn_expr_argtype(fcinfo->flinfo, 1);
 	Ensure(OidIsValid(val_type), "cannot determine argument type");
 	TypeCacheEntry *val_entry = lookup_type_cache(val_type, TYPECACHE_HASH_PROC);
 	Ensure(OidIsValid(val_entry->hash_proc), "cannot determine hash function");
 	const Oid hash_proc_oid = val_entry->hash_proc;
 
 	/* compute the hashes, used for the bloom filter */
-	uint32 datum_hash = DatumGetUInt32(OidFunctionCall1Coll(hash_proc_oid, C_COLLATION_OID, val));
-	uint32 h1 = hash_bytes_uint32_extended(datum_hash, BLOOM1_SEED_1) % nbits;
-	uint32 h2 = hash_bytes_uint32_extended(datum_hash, BLOOM1_SEED_2) % nbits;
+	Datum val = PG_GETARG_DATUM(1);
+	const uint32 datum_hash =
+		DatumGetUInt32(OidFunctionCall1Coll(hash_proc_oid, C_COLLATION_OID, val));
 
 	/* compute the requested number of hashes */
-	const char *words = VARDATA_ANY(bloom);
+	bytea *bloom = PG_GETARG_VARLENA_PP(0);
+	const int nbits = VARSIZE_ANY_EXHDR(bloom) * 8;
+	const uint64 *words = (const uint64 *) VARDATA_ANY(bloom);
 	const int word_bits = sizeof(*words) * 8;
 	bool match = true;
 	for (int i = 0; i < BLOOM1_HASHES; i++)
 	{
-		/* h1 + h2 + f(i) */
-		uint32 h = (h1 + i * h2) % nbits;
-		uint32 word_index = (h / word_bits);
-		uint32 bit = (h % word_bits);
-
-		/* if the bit is not set, set it and remember we did that */
+		const uint32 h = bloom1_get_one_hash(datum_hash, i) % nbits;
+		const uint32 word_index = (h / word_bits);
+		const uint32 bit = (h % word_bits);
 		match = (words[word_index] & (0x01 << bit)) && match;
 	}
 

tsl/src/compression/batch_metadata_builder_bloom1.c

@@ -76,7 +76,7 @@ batch_metadata_builder_bloom1_create(Oid type_oid, int bloom_attr_offset)
 		.nbits_set = 0,
 	};
 
-	Assert(builder->nbits % 8 == 0);
+	Assert(builder->nbits % 64 == 0);
 	const int bytea_size = VARHDRSZ + builder->nbits / 8;
 	builder->bloom_bytea = palloc0(bytea_size);
 	SET_VARSIZE(builder->bloom_bytea, bytea_size);
@@ -89,30 +89,22 @@ bloom1_update_val(void *builder_, Datum val)
 {
 	Bloom1MetadataBuilder *builder = (Bloom1MetadataBuilder *) builder_;
 
-	const int nbits = builder->nbits;
 	const Oid hash_proc_oid = builder->hash_proc_oid;
 
 	/* compute the hashes, used for the bloom filter */
-	uint32 datum_hash =
-		DatumGetUInt32(OidFunctionCall1Coll(hash_proc_oid, /* collation = */ C_COLLATION_OID, val));
-	uint32 h1 = hash_bytes_uint32_extended(datum_hash, BLOOM1_SEED_1) % nbits;
-	uint32 h2 = hash_bytes_uint32_extended(datum_hash, BLOOM1_SEED_2) % nbits;
+	const uint32 datum_hash =
+		DatumGetUInt32(OidFunctionCall1Coll(hash_proc_oid, C_COLLATION_OID, val));
 
 	/* compute the requested number of hashes */
-	char *restrict words = VARDATA(builder->bloom_bytea);
+	const int nbits = builder->nbits;
+	uint64 *restrict words = (uint64 *restrict) VARDATA(builder->bloom_bytea);
+	const int word_bits = sizeof(*words) * 8;
 	for (int i = 0; i < BLOOM1_HASHES; i++)
 	{
-		/* h1 + h2 + f(i) */
-		uint32 h = (h1 + i * h2) % builder->nbits;
-		uint32 byte = (h / 8);
-		uint32 bit = (h % 8);
-
-		/* if the bit is not set, set it and remember we did that */
-		if (!(words[byte] & (0x01 << bit)))
-		{
-			words[byte] |= (0x01 << bit);
-			builder->nbits_set++;
-		}
+		const uint32 h = bloom1_get_one_hash(datum_hash, i) % nbits;
+		const uint32 byte = (h / word_bits);
+		const uint32 bit = (h % word_bits);
+		words[byte] |= (0x01 << bit);
 	}
 }
 

tsl/test/expected/compress_bloom_sparse.out

@@ -53,13 +53,13 @@ select count(*) from bloom where value = md5(7248::text);
    Output: count(*)
    ->  Custom Scan (DecompressChunk) on _timescaledb_internal._hyper_1_1_chunk (actual rows=1 loops=1)
          Vectorized Filter: (_hyper_1_1_chunk.value = '1f4183315762e30ea441d3caef5e64ad'::text)
-         Rows Removed by Filter: 1999
-         Batches Removed by Filter: 1
+         Rows Removed by Filter: 2999
+         Batches Removed by Filter: 2
          Bulk Decompression: true
-         ->  Seq Scan on _timescaledb_internal.compress_hyper_2_2_chunk (actual rows=2 loops=1)
+         ->  Seq Scan on _timescaledb_internal.compress_hyper_2_2_chunk (actual rows=3 loops=1)
                Output: compress_hyper_2_2_chunk._ts_meta_count, compress_hyper_2_2_chunk._ts_meta_min_1, compress_hyper_2_2_chunk._ts_meta_max_1, compress_hyper_2_2_chunk.ts, compress_hyper_2_2_chunk._ts_meta_v2_bloom1_value, compress_hyper_2_2_chunk.value
                Filter: _timescaledb_functions.ts_bloom1_matches(compress_hyper_2_2_chunk._ts_meta_v2_bloom1_value, '1f4183315762e30ea441d3caef5e64ad'::text)
-               Rows Removed by Filter: 8
+               Rows Removed by Filter: 7
 (11 rows)
 
 select count(*) from bloom where value = md5(7248::text);