mirror of
https://github.com/hanzoai/datastore.git
synced 2026-09-26 13:55:07 +00:00
Alexey MilovidovandClaude Opus 4.8
33ec1d50cd
`QBit` already supports the transposed (bit-plane) distance functions `L2DistanceTransposed` and `cosineDistanceTransposed`, but had no inner product counterpart. This adds `dotProductTransposed`, the transposed analogue of `dotProduct`. Unlike the distance functions, the dot product is a similarity measure: larger values mean more similar vectors. It returns `Float64` and accepts the same `(qbit, reference, p)` arguments, where `p` selects how many bit planes to read for the precision-speed trade-off. Implementation reuses the generic `FunctionArrayDistance` kernel machinery: the new `DotProductTransposed` kernel resolves the SimSIMD `dot` metric and provides a scalar fallback. The function is also recognized by `DistanceTransposedPartialReadsPass` so the partial-reads optimization applies to it as well. `scalarProductTransposed` is registered as an alias, mirroring the `dotProduct`/`scalarProduct` pair. Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>
70 lines
2.5 KiB
XML
70 lines
2.5 KiB
XML
<test>
|
|
<!--
|
|
Microbenchmark isolating the per-row cost of the QBit distance functions.
|
|
|
|
Unlike qbit_l2_distance_*_memory.xml (dim=100, ORDER BY ... LIMIT, multi-threaded), this test is deliberately shaped so that almost
|
|
all of the wall-clock time is spent inside the per-row distance kernel call, so changes to that call
|
|
-->
|
|
<settings>
|
|
<max_threads>1</max_threads>
|
|
</settings>
|
|
|
|
<substitutions>
|
|
<substitution>
|
|
<name>vector_size</name>
|
|
<values>
|
|
<value>8</value>
|
|
</values>
|
|
</substitution>
|
|
|
|
<substitution>
|
|
<name>elements</name>
|
|
<values>
|
|
<value>3000000</value>
|
|
</values>
|
|
</substitution>
|
|
|
|
<substitution>
|
|
<name>type</name>
|
|
<values>
|
|
<value>Float64</value>
|
|
</values>
|
|
</substitution>
|
|
|
|
<substitution>
|
|
<name>func</name>
|
|
<values>
|
|
<value>L2DistanceTransposed</value>
|
|
<value>cosineDistanceTransposed</value>
|
|
<value>dotProductTransposed</value>
|
|
</values>
|
|
</substitution>
|
|
|
|
<substitution>
|
|
<name>p</name>
|
|
<values>
|
|
<value>1</value>
|
|
</values>
|
|
</substitution>
|
|
</substitutions>
|
|
|
|
<create_query>DROP TABLE IF EXISTS tab_{type}</create_query>
|
|
<create_query>CREATE TABLE tab_{type} (id UInt32, vec QBit({type}, {vector_size})) ENGINE = Memory</create_query>
|
|
<create_query>INSERT INTO tab_{type} SELECT number + 1 AS id, arrayMap(i -> id + i, range({vector_size})) AS vec FROM numbers({elements})</create_query>
|
|
|
|
<query>
|
|
SELECT
|
|
sum({func}(vec, arrayMap(i -> to{type}(i + 1), range({vector_size})), {p}))
|
|
+ sum({func}(vec, arrayMap(i -> to{type}(i + 2), range({vector_size})), {p}))
|
|
+ sum({func}(vec, arrayMap(i -> to{type}(i + 3), range({vector_size})), {p}))
|
|
+ sum({func}(vec, arrayMap(i -> to{type}(i + 4), range({vector_size})), {p}))
|
|
+ sum({func}(vec, arrayMap(i -> to{type}(i + 5), range({vector_size})), {p}))
|
|
+ sum({func}(vec, arrayMap(i -> to{type}(i + 6), range({vector_size})), {p}))
|
|
+ sum({func}(vec, arrayMap(i -> to{type}(i + 7), range({vector_size})), {p}))
|
|
+ sum({func}(vec, arrayMap(i -> to{type}(i + 8), range({vector_size})), {p}))
|
|
FROM tab_{type}
|
|
</query>
|
|
|
|
<drop_query>DROP TABLE tab_{type}</drop_query>
|
|
</test>
|