Module simd 

SIMD-accelerated vector operations for embedding similarity.

Provides optimized implementations with automatic fallback:

• x86_64 (float32): AVX-512F > AVX2 + FMA > scalar
• x86_64 (int8): AVX-512 VNNI > AVX2 > scalar
• aarch64: ARM NEON with multiple accumulators and loop unrolling
• Other: Scalar fallback

Optimizations

• Multiple accumulators: 4 parallel accumulators to break dependency chains
• Loop unrolling: Process 16/32/64 elements per iteration depending on ISA
• AVX-512F: Wide float32 kernels for dot, cosine, normalize, and distance
• AVX-512 VNNI: Integer VNNI instructions when available (quantized int8 path)

Structs

BinaryVector

Unstable: binary quantization format and struct layout are under active design.

Int4Params

Unstable: INT4 quantization internals; scale/bias scheme may change.

Int4Vector

Unstable: INT4 quantization format is under active design; struct layout may change.

PreparedQueryWithMeta

A prepared query annotated with a normalization hint for fast-path dispatch.

QuantizationParams

Unstable: INT8 quantization parameters; scale/bias scheme may change.

QuantizedVector

Unstable: INT8 quantized vector; struct layout and invariants may change.

SimdConfig

Unstable: SIMD dispatch internals; fields may be added as new ISAs are supported.

Enums

NormalizationHint

Caller assertion that a vector is L2-unit-normalized (norm ≈ 1).

PreparedQuery

Unstable: pre-quantized query for repeated distance computation.

QuantizationTier

Unstable: tier design is under active iteration; tier boundaries may change.

QuantizedData

Unstable: unified quantized data container; variants may change with tier redesign.

Functions

approximate_cosine_distance

Unstable: tiered distance dispatch; tier mix and formula may change.

approximate_cosine_distance_prepared

Unstable: prepared cosine distance; query tier must match stored data tier.

approximate_cosine_distance_prepared_with_meta

Cosine distance with unit-norm fast path.

approximate_dot_product

Unstable: tiered dot product dispatch; tier mix and formula may change.

approximate_dot_product_prepared

Unstable: prepared dot product dispatch; query tier must match stored data tier.

batch_approximate_cosine_distance_prepared

Compute cosine distances from one prepared query to a slice of stored vectors.

batch_approximate_cosine_distance_prepared_into

Like batch_approximate_cosine_distance_prepared but writes into a caller-supplied buffer.

batch_cosine_similarity

Unstable: SIMD batch dispatch; use lattice_embed::utils::batch_cosine_similarity for stable wrapper.

batch_dot_product

Unstable: SIMD batch dispatch; use lattice_embed::utils::batch_dot_product for stable wrapper.

cosine_similarity

Unstable: SIMD dispatch layer; use lattice_embed::utils::cosine_similarity for the stable wrapper.

cosine_similarity_i8

Unstable: SIMD INT8 cosine similarity; norm storage approach may change.

dot_product

Unstable: SIMD dispatch layer; use lattice_embed::utils::dot_product for the stable wrapper.

dot_product_batch4

Compute dot product of one query against 4 candidates simultaneously.

dot_product_i8

Unstable: SIMD INT8 dot product; VNNI/AVX2/NEON dispatch may change.

dot_product_i8_raw

Unstable: raw SIMD INT8 hot path; signature and scaling semantics may change.

euclidean_distance

Unstable: SIMD dispatch layer; use lattice_embed::utils::euclidean_distance for the stable wrapper.

is_unit_norm

Returns true when the squared norm of v is within 1e-4 of 1.0.

normalize

Unstable: SIMD dispatch layer; use lattice_embed::utils::normalize for the stable wrapper.

prepare_query

Prepare a query vector for repeated distance computation against a homogeneous tier.

prepare_query_with_norm

Prepare a query annotated with the given normalization hint.

resolved_dot_product_batch4_kernel

Resolve the best available batch-4 f32 dot-product kernel once and return it.

resolved_dot_product_kernel

Resolve the best available f32 dot-product kernel once and return it.

resolved_i8_dot_kernel

Return the cached INT8 dot-product kernel.

simd_config

Unstable: SIMD dispatch internal; shape may change as new backends are added.

squared_euclidean_distance

Unstable: squared Euclidean distance — skips the final sqrt.

try_approximate_cosine_distance_prepared

Non-panicking variant of approximate_cosine_distance_prepared.

try_approximate_dot_product_prepared

Non-panicking variant of approximate_dot_product_prepared.

Type Aliases

DotBatch4Kernel

SIMD kernel type for batch-4 f32 dot product.

DotKernel

SIMD kernel function pointer type for f32 dot product.

I8DotKernel

INT8 dot-product kernel function pointer type.