Request for NEON/SVE Optimized Implementations for ARM in HNSWlib

[ThatikondV]

Hii,

I’m currently working on performance optimization of Milvus on ARM architectures, and since Milvus internally leverages Knowhere, which in turn uses HNSWlib for the HNSW index, I’ve been analyzing the end-to-end execution flow.

While going through the source code (particularly space_l2.h and space_ip.h), I noticed that:

• HNSWlib currently does not contain any architecture-specific vectorized kernels (e.g., NEON or SVE for ARM).

• The existing distance computation functions (for L2 and Inner Product) are implemented using scalar C++ loops, and rely purely on compiler auto-vectorization.

• This leads to suboptimal performance on ARM compared to x86 builds, where other ecosystems (e.g., FAISS) sometimes include architecture-tuned kernels (SSE/AVX).

My questions:

• Is there any plan to include NEON or SVE-optimized kernels for ARM platforms (especially for the distance functions inside space_l2.h and space_ip.h)?

• If not yet planned, would the maintainers be open to a contribution/PR that introduces such ARM-specific optimizations (behind compile-time flags like __ARM_NEON / __ARM_FEATURE_SVE)?

• Are there any ongoing discussions or design considerations around this that I can align with before starting the implementation?