AVX-512 SIMD not fully implemented - falls back to AVX2

[ruvnet]

Priority: High

Current Behavior

AVX-512 distance calculations fall back to AVX2 implementation. The AVX-512 code paths exist but are not fully optimized.

Expected Behavior

Full AVX-512 implementation for maximum performance on supported hardware (2-3x speedup over AVX2).

Technical Details

• Location: crates/ruvector-postgres/src/distance/simd.rs
• Current state: AVX-512 feature flags exist but implementation uses AVX2 intrinsics
• Affected operations: L2, cosine, inner product, Manhattan distance

Acceptance Criteria

• Implement AVX-512 intrinsics for all distance functions
• Add runtime detection for AVX-512 support
• Benchmark AVX-512 vs AVX2 performance
• Add CI tests on AVX-512 capable hardware

Impact

~2-3x potential performance improvement for distance calculations on modern Intel/AMD CPUs with AVX-512 support.

[ruvnet]

Status: AVX-512 Already Implemented ✅

Upon investigation, AVX-512 SIMD is already fully implemented in crates/ruvector-postgres/src/distance/simd.rs.

What's Implemented (lines 122-298):

• l2_distance_ptr_avx512 - Euclidean distance (16 floats per iteration)
• cosine_distance_ptr_avx512 - Cosine distance
• inner_product_ptr_avx512 - Inner product
• manhattan_distance_ptr_avx512 - Manhattan distance
• cosine_distance_normalized_avx512 - Pre-normalized cosine

Runtime Detection (lines 30-39):

pub fn is_avx512_available() -> bool {
    #[cfg(feature = "simd-avx512")]
    { is_x86_feature_detected!("avx512f") }
    #[cfg(not(feature = "simd-avx512"))]
    { false }
}

How to Enable:

AVX-512 requires the simd-avx512 feature flag (needs nightly Rust for intrinsics):

cargo build --release --features simd-avx512

Why It Falls Back:

Without the feature flag, is_avx512_available() always returns false, causing the code to fall back to AVX2.

Recommendation: Consider auto-enabling AVX-512 on stable Rust 1.72+ where AVX-512 intrinsics are stable, or document the feature flag requirement in the README.

The implementation uses:

• _mm512_loadu_ps for loading 16 floats
• _mm512_fmadd_ps for fused multiply-add
• _mm512_reduce_add_ps for horizontal reduction
• _mm512_abs_ps for absolute values

No code changes needed - just documentation update for enabling the feature.