Add Neon Squared L2 kernels for USlice2 and USlice4

[pfoxARM]

What does this implement/fix? Briefly explain your changes.

• Adds U4 Squared L2 U2U2 kernel and U4U4 Squared L2 kernel
• Adds SIMDAbsDiff trait in diskann-wide with emulated paths and testing

Any other comments?

Tested for performance on C100 and C200 using graph-index-build-sq

[Copilot]

🟡 Changes recommended

A critical codegen arity error prevents compilation, with documentation nits also unresolved.

Get a fresh assessment by requesting another Copilot review.

Pull request overview

Adds AArch64 Neon squared-L2 kernels for packed 2-bit and 4-bit vectors, plus SIMD absolute-difference support and benchmark coverage.

Changes:

• Adds SIMDAbsDiff implementations, tests, and exports.
• Adds Neon U2/U4 squared-L2 kernels and dispatch updates.
• Registers the 2-bit scalar-quantization benchmark.

File summaries

File	Summary	Findings
diskann-wide/src/traits.rs	Defines SIMDAbsDiff.	—
diskann-wide/src/test_utils/ops.rs	Adds absolute-difference test support.	—
diskann-wide/src/reference.rs	Adds scalar reference implementations.	—
diskann-wide/src/lib.rs	Re-exports the trait.	—
diskann-wide/src/emulated.rs	Implements and tests emulated absolute differences.	—
diskann-wide/src/doubled.rs	Propagates the operation across doubled vectors.	—
diskann-wide/src/arch/aarch64/u8x16_.rs	Adds Neon U8 absolute-difference support and tests.	—
diskann-wide/src/arch/aarch64/macros.rs	Adds the Neon intrinsic wrapper macro.	—
diskann-quantization/src/spherical/__codegen/aarch64.rs	Adds Neon spherical L2 instantiations.	Critical (3 votes): Incorrect CompensatedSquaredL2::new arity causes codegen compilation failure in four entries. Nit (3 votes): Rename SquardL2 to SquaredL2.
diskann-quantization/src/bits/distances.rs	Adds Neon U2/U4 squared-L2 kernels and dispatch.	—
diskann-benchmark/src/index/inmem/scalar.rs	Registers the 2-bit benchmark.	Nit (2 votes): Update the policy comment to include the 2-bit variant.

Review details

Suppressed comments (3)

diskann-quantization/src/spherical/__codegen/aarch64.rs:117

• This L2 codegen entry point passes shift to CompensatedSquaredL2::new, but that constructor takes only dim (vectors.rs:543), so enabling the codegen feature cannot compile. Remove the unused shift parameter and call new(dim) here.

    shift: &[f32],
    dim: usize,
) -> Result<DistanceComputer, AllocatorError> {
    let reify = Reify::<_, _, AsQuery<2>, AsData<2>>::new(
        vectors::CompensatedSquaredL2::new(shift, dim),

diskann-quantization/src/spherical/__codegen/aarch64.rs:134

• This L2 codegen entry point passes shift to CompensatedSquaredL2::new, but that constructor takes only dim (vectors.rs:543), so enabling the codegen feature cannot compile. Remove the unused shift parameter and call new(dim) here.

    shift: &[f32],
    dim: usize,
) -> Result<DistanceComputer, AllocatorError> {
    let reify = Reify::<_, _, AsData<4>, AsData<4>>::new(
        vectors::CompensatedSquaredL2::new(shift, dim),

diskann-quantization/src/spherical/__codegen/aarch64.rs:153

• This L2 codegen entry point passes shift to CompensatedSquaredL2::new, but that constructor takes only dim (vectors.rs:543), so enabling the codegen feature cannot compile. Remove the unused shift parameter and call new(dim) here.

    shift: &[f32],
    dim: usize,
) -> Result<DistanceComputer, AllocatorError> {
    let reify = Reify::<_, _, AsQuery<4>, AsData<4>>::new(
        vectors::CompensatedSquaredL2::new(shift, dim),

• Files reviewed: 11/11 changed files
• Comments generated: 3
• Review effort level: Lite

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.