Struct HashEncoder 

pub struct HashEncoder {
    pub hash_tables: Vec<Vec<Vec<f64>>>,
    pub n_levels: usize,
    pub n_features_per_level: usize,
    pub base_resolution: usize,
    pub resolution_growth: f64,
    pub table_size: usize,
}

Multi-resolution hash encoder from Instant-NGP (Müller et al. 2022).

Fields

• hash_tables – [L] levels, each a [table_size × F] array of feature scalars.
• n_levels – number of resolution levels L.
• n_features_per_level – number of features F per hash entry.
• base_resolution – grid resolution at level 0.
• resolution_growth – per-level resolution multiplier b.
• table_size – capacity of every hash table (= 2^{log2_hashmap_size}).

Fields

hash_tables: Vec<Vec<Vec<f64>>>

[L][table_size][F] – learnable feature vectors.

n_levels: usize

Number of resolution levels L.

n_features_per_level: usize

Number of feature scalars F stored per hash-table entry.

base_resolution: usize

Grid resolution at the coarsest level (level 0).

resolution_growth: f64

Per-level resolution growth factor b = (N_max/N_min)^{1/(L-1)}.

table_size: usize

Capacity of each hash table (= 2^{log2_hashmap_size}).

Implementations

impl HashEncoder

pub fn new( n_levels: usize, n_features_per_level: usize, log2_hashmap_size: usize, base_resolution: usize, finest_resolution: usize, seed: u64, ) -> Self

Construct and randomly initialise a new HashEncoder.

Features are initialised uniformly in [-0.0001, 0.0001].

Arguments

• n_levels – number of resolution levels.
• n_features_per_level – features per hash entry.
• log2_hashmap_size – log₂ of hash-table capacity.
• base_resolution – coarsest grid resolution.
• finest_resolution – finest grid resolution.
• seed – PRNG seed.

pub fn n_output_features(&self) -> usize

Total output feature dimension = n_levels × n_features_per_level.

pub fn lookup_features(&self, pos: &[f64; 3], level: usize) -> Vec<f64>

Trilinearly interpolate the 8 corner features of the voxel containing pos at resolution level l.

Returns a Vec<f64> of length n_features_per_level.

pub fn encode(&self, pos: &[f64; 3]) -> Vec<f64>

Encode a 3-D position by looking up and concatenating features from all n_levels resolution levels.

Output length: n_levels × n_features_per_level.