[pfoxARM]

Add an AArch64-specific implementation of get_lower_bound that uses a scalar linear search.

Benchmarks showed that scalar linear search gives the best overall QPS across different vector dimensions (dim128, dim256, dim768) and L sizes (10,20...,80).

We tested several alternatives:

• NEON SIMD (vcgeq_f32)
• Scalar unrolled search
• Rust partition_point
• Manual binary search

It was clear from testing that a scalar linear approach is best.

[Copilot]

🟡 Changes recommended

The architecture guards leave other targets without get_lower_bound, causing missing-method compilation errors.

1 open issue

• Restore scalar fallback for get_lower_bound on non-x86_64 targets · New

What changed in this PR

This PR specializes the neighbor priority queue’s lower-bound search for AArch64, using scalar linear search based on the reported benchmark results.

Changes:

• Adds an AArch64 scalar implementation of get_lower_bound.
• Restricts the existing SIMD implementation and its imports to x86-64.

File	Description
diskann/​src/​neighbor/​queue.rs	Adds architecture-specific lower-bound implementations and gates SIMD imports.

_{Review effort: Balanced}

[Mark Hildebrand (hildebrandmw)]

Thanks! One very tiny suggestion from me.

[Codecov Comments Bot (codecov-commenter)]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 92.07%. Comparing base (448e520) to head (dbb571a).

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1512      +/-   ##
==========================================
+ Coverage   91.02%   92.07%   +1.04%     
==========================================
  Files         602      602              
  Lines      124080   124080              
==========================================
+ Hits       112948   114245    +1297     
+ Misses      11132     9835    -1297     

Flag	Coverage Δ
miri	92.07% <ø> (+1.04%)	⬆️
unittests	92.01% <ø> (+1.28%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
diskann/src/neighbor/queue.rs	98.93% <ø> (ø)

... and 50 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Mark Hildebrand (hildebrandmw)]

Thanks - looks good to me! CI failure should be taken care of by merging/rebasing with the latest main. Apologies for the flakiness.

[pfoxARM]

No worries, hopefully CI passes now :)

[Jack Moffitt (metajack)]

Note that openai v3 vectors at 3072 dimensions, and naive binary search will win there I think. At least it did last I benchmarked this. In the future I think we will have some way to dynamically select an NPQ algorithm based on dimension, but for now matching the existing thing is probably the right call.

[Jack Moffitt (metajack)]

It seems quite surprising that a) our fallback is much less efficient and b) not using SIMD is as good as doing the work 8x as fast?

[Mark Hildebrand (hildebrandmw)]

Jack Moffitt (@metajack) - The big issue with the SIMD approach on aarch64 is that .first() is surprisingly expensive on aarch64, where-as it's much cheaper on x86. We could do some work on the masks types for aarch64 to maybe get it a bit faster, but it's still unfortunately not free.