Struct PolarCodec 

pub struct PolarCodec { /* private fields */ }

PolarQuant codec: batch encode, query preparation, and high-throughput scan.

Implementations

impl PolarCodec

pub fn new(dim: usize, bits: u8, seed: u64) -> Self

Create a new codec.

Panics

Panics if dim is 0, odd, or bits is 0 or > 8.

pub fn pairs(&self) -> usize

Number of pairs per vector.

pub fn encode(&self, vector: &[f32]) -> CompressedCode

Encode a single vector (convenience, allocates).

pub fn encode_batch(&self, vectors: &Array2<f32>) -> CompressedCorpus

Batch-encode into a flat CompressedCorpus (SoA layout).

Uses BLAS for the rotation: rotated = vectors × Πᵀ (one GEMM). Quantization is scalar but cache-friendly (sequential writes).

pub fn encode_batch_codes(&self, vectors: &Array2<f32>) -> Vec<CompressedCode>

Also produce the old per-vector codes (for backward compat with index.rs).

pub fn prepare_query(&self, query: &[f32]) -> QueryState

Prepare query-dependent centroid lookup table.

Cost: one 768×768 matvec + 384×16 multiply-adds = ~0.08ms. The returned QueryState is reused for ALL vectors in the scan.

pub fn scan_corpus( &self, corpus: &CompressedCorpus, qs: &QueryState, ) -> Vec<f32>

High-throughput scan of a CompressedCorpus against a prepared query.

Returns approximate inner product scores for all vectors. Memory access: sequential streaming through radii + indices (cache-optimal). Centroid table: 24 KB, stays in L1 throughout.

At 100K vectors, d=768: ~3.3ms on CPU, ~0.1ms on GPU (future Metal kernel).

pub fn batch_scan(&self, codes: &[CompressedCode], qs: &QueryState) -> Vec<f32>

Scan per-vector codes (old API, for backward compat).