Module simd 

Shared SIMD-accelerated functions for posting list compression

This module provides platform-optimized implementations for common operations:

• Unpacking: Convert packed 8/16/32-bit values to u32 arrays
• Delta decoding: Prefix sum for converting deltas to absolute values
• Add one: Increment all values in an array (for TF decoding)

Supports:

• NEON on aarch64 (Apple Silicon, ARM servers)
• SSE/SSE4.1 on x86_64 (Intel/AMD)
• Scalar fallback for other architectures

Enums

RoundedBitWidth

Rounded bit width type for SIMD-friendly encoding

Functions

add_one

Add 1 to all values with SIMD acceleration

batch_cosine_scores

Batch cosine similarity: query vs N contiguous vectors.

batch_cosine_scores_f16

Batch cosine similarity: f32 query vs N contiguous f16 vectors.

batch_cosine_scores_u8

Batch cosine similarity: f32 query vs N contiguous u8 vectors.

batch_f32_to_f16

Batch convert f32 slice to f16 (stored as u16)

batch_f32_to_u8

Batch convert f32 slice to u8 with [-1,1] → [0,255] mapping

batch_squared_euclidean_distances

Batch compute squared Euclidean distances from one query to multiple vectors

bits_needed

Compute the number of bits needed to represent a value

cosine_similarity

Compute cosine similarity between two f32 vectors with SIMD acceleration

delta_decode

Delta decode with SIMD acceleration

dequantize_uint8

Dequantize UInt8 weights to f32 with SIMD acceleration

dot_product_f32

Compute dot product of two f32 arrays with SIMD acceleration

f16_to_f32

Convert f16 (stored as u16) to f32

f32_to_f16

Convert f32 to f16 (IEEE 754 half-precision), stored as u16

f32_to_u8_saturating

Quantize f32 in [-1, 1] to u8 [0, 255]

max_f32

Find maximum value in f32 array with SIMD acceleration

pack_rounded

Pack values using rounded bit width (SIMD-friendly)

round_bit_width

Round a bit width to the nearest SIMD-friendly width (0, 8, 16, or 32)

squared_euclidean_distance

Compute squared Euclidean distance between two f32 vectors with SIMD acceleration

u8_to_f32

Dequantize u8 [0, 255] to f32 in [-1, 1]

unpack_8bit

Unpack 8-bit packed values to u32 with SIMD acceleration

unpack_8bit_delta_decode

Fused unpack 8-bit + delta decode in a single pass

unpack_16bit

Unpack 16-bit packed values to u32 with SIMD acceleration

unpack_16bit_delta_decode

Fused unpack 16-bit + delta decode in a single pass

unpack_32bit

Unpack 32-bit packed values to u32 with SIMD acceleration

unpack_delta_decode

Fused unpack + delta decode for arbitrary bit widths

unpack_rounded

Unpack values using rounded bit width with SIMD acceleration

unpack_rounded_delta_decode

Fused unpack + delta decode using rounded bit width