datastore/tests/performance/array_distance_simd_speedup.xml

<test>
    <!--
        Speedup-focused companion to array_distance.xml. It isolates the cases
        where PR #101310 has a clear net win over master: the non-const x
        non-const and mixed-type paths (which had NO SIMD on master at all),
        with the FMA-friendly L2 / L2Squared / cosine kernels, at the
        mid-to-long vector lengths typical of vector search.

        L1 / Linf and the const-needle hand-written paths are deliberately left
        to array_distance.xml and norm_distance: their kernels are either less
        SIMD-friendly (abs-sum / fmax reductions) or already optimal on master,
        so they belong in the broad regression suite rather than here.
    -->
    <substitutions>
        <substitution>
            <name>dim</name>
            <values>
                <value>512</value>
                <value>2048</value>
            </values>
        </substitution>
    </substitutions>

    <settings>
        <max_threads>1</max_threads>
    </settings>

    <!-- Two distinct Float32 vectors per row (so cosine is well-defined). -->
    <create_query>
        CREATE TABLE adss_f32_{dim} (a Array(Float32), b Array(Float32)) ENGINE = Memory
    </create_query>

    <!-- Quantised UInt8 storage vector against a Float64 query vector: the
         canonical mixed-type ANN layout that has no SIMD path on master. -->
    <create_query>
        CREATE TABLE adss_mixed_{dim} (u8 Array(UInt8), f64 Array(Float64)) ENGINE = Memory
    </create_query>

    <fill_query>
        INSERT INTO adss_f32_{dim}
        SELECT
            arrayMap(i -> toFloat32((number * 9 + i) % 1000) * 0.001, range({dim})),
            arrayMap(i -> toFloat32((number * 7 + i + 13) % 1000) * 0.001, range({dim}))
        FROM numbers(100000)
    </fill_query>
    <fill_query>
        INSERT INTO adss_mixed_{dim}
        SELECT
            arrayMap(i -> toUInt8((number + i) % 256), range({dim})),
            arrayMap(i -> toFloat64((number * 3 + i) % 997) * 0.5, range({dim}))
        FROM numbers(100000)
    </fill_query>

    <!-- ===== Native Float32, non-const x non-const: no SIMD on master ===== -->
    <query>SELECT sum(cosineDistance(a, b))    FROM adss_f32_{dim}</query>
    <query>SELECT sum(L2Distance(a, b))        FROM adss_f32_{dim}</query>
    <query>SELECT sum(L2SquaredDistance(a, b)) FROM adss_f32_{dim}</query>

    <!-- ===== Mixed UInt8 x Float64: no SIMD on master, non-const and const ===== -->
    <query>SELECT sum(L2Distance(u8, f64))     FROM adss_mixed_{dim}</query>
    <query>SELECT sum(cosineDistance(u8, f64)) FROM adss_mixed_{dim}</query>
    <query>
        WITH (SELECT u8 FROM adss_mixed_{dim} LIMIT 1) AS needle
        SELECT sum(L2Distance(needle, f64)) FROM adss_mixed_{dim}
    </query>

    <drop_query>DROP TABLE adss_f32_{dim}</drop_query>
    <drop_query>DROP TABLE adss_mixed_{dim}</drop_query>
</test>