Struct DecodeTable 

pub struct DecodeTable { /* private fields */ }

Precomputed 256-entry lookup table mapping every possible packed byte to the sequence of bucket_weights values it decodes to.

PLAID §4.5 notes that naive decompression pays a chain of shift-mask-weight-lookup operations per residual dimension; a one-off table that already composes the shift/mask with the weight lookup reduces decoding to a single load per code position. For nbits=2 the whole table is 256 × 4 f32 = 4 KiB and easily stays in L1.

Implementations

impl DecodeTable

pub fn new(codec: &ResidualCodec) -> Self

Build the table for codec. Call once per search/decode batch and reuse across every encoded vector.

pub fn weights_for(&self, byte: u8) -> &[f32]

Weights for the codes_per_byte positions inside packed byte byte. Length always equals codes_per_byte.

pub fn weights_flat(&self) -> &[f32]

Raw row-major [256, codes_per_byte] weights buffer.

Exposed so the search path can upload the table once per query and decode residuals via batched index_select on the device, matching the GPU decompression kernel described in PLAID §4.5 (one thread per packed byte).

pub fn codes_per_byte(&self) -> usize

Number of codes packed into one byte at this table’s nbits.

pub fn nbits(&self) -> u32

Bit-width the table was built for.