world_cache_query.wgsl

#define_import_path bevy_solari::world_cache

/// How responsive the world cache is to changes in lighting (higher is less responsive, lower is more responsive)
const WORLD_CACHE_MAX_TEMPORAL_SAMPLES: f32 = 10.0;
/// Maximum amount of frames a cell can live for without being queried
const WORLD_CACHE_CELL_LIFETIME: u32 = 30u;
/// Maximum amount of attempts to find a cache entry after a hash collision
const WORLD_CACHE_MAX_SEARCH_STEPS: u32 = 3u;

/// The size of a cache cell at the lowest LOD in meters
const WORLD_CACHE_POSITION_BASE_CELL_SIZE: f32 = 0.25;
/// How fast the world cache transitions between LODs as a function of distance to the camera
const WORLD_CACHE_POSITION_LOD_SCALE: f32 = 30.0;

/// Marker value for an empty cell
const WORLD_CACHE_EMPTY_CELL: u32 = 0u;

struct WorldCacheGeometryData {
    world_position: vec3<f32>,
    padding_a: u32,
    world_normal: vec3<f32>,
    padding_b: u32
}

@group(1) @binding(14) var<storage, read_write> world_cache_checksums: array<atomic<u32>, #{WORLD_CACHE_SIZE}>;
#ifdef WORLD_CACHE_NON_ATOMIC_LIFE_BUFFER
@group(1) @binding(15) var<storage, read_write> world_cache_life: array<u32, #{WORLD_CACHE_SIZE}>;
#else
@group(1) @binding(15) var<storage, read_write> world_cache_life: array<atomic<u32>, #{WORLD_CACHE_SIZE}>;
#endif
@group(1) @binding(16) var<storage, read_write> world_cache_radiance: array<vec4<f32>, #{WORLD_CACHE_SIZE}>;
@group(1) @binding(17) var<storage, read_write> world_cache_geometry_data: array<WorldCacheGeometryData, #{WORLD_CACHE_SIZE}>;
@group(1) @binding(18) var<storage, read_write> world_cache_active_cells_new_radiance: array<vec3<f32>, #{WORLD_CACHE_SIZE}>;
@group(1) @binding(19) var<storage, read_write> world_cache_a: array<u32, #{WORLD_CACHE_SIZE}>;
@group(1) @binding(20) var<storage, read_write> world_cache_b: array<u32, 1024u>;
@group(1) @binding(21) var<storage, read_write> world_cache_active_cell_indices: array<u32, #{WORLD_CACHE_SIZE}>;
@group(1) @binding(22) var<storage, read_write> world_cache_active_cells_count: u32;

#ifndef WORLD_CACHE_NON_ATOMIC_LIFE_BUFFER
fn query_world_cache(world_position: vec3<f32>, world_normal: vec3<f32>, view_position: vec3<f32>) -> vec3<f32> {
    let cell_size = get_cell_size(world_position, view_position);
    let world_position_quantized = bitcast<vec3<u32>>(quantize_position(world_position, cell_size));
    let world_normal_quantized = bitcast<vec3<u32>>(quantize_normal(world_normal));
    var key = compute_key(world_position_quantized, world_normal_quantized);
    let checksum = compute_checksum(world_position_quantized, world_normal_quantized);

    for (var i = 0u; i < WORLD_CACHE_MAX_SEARCH_STEPS; i++) {
        let existing_checksum = atomicCompareExchangeWeak(&world_cache_checksums[key], WORLD_CACHE_EMPTY_CELL, checksum).old_value;
        if existing_checksum == checksum {
            // Cache entry already exists - get radiance and reset cell lifetime
            atomicStore(&world_cache_life[key], WORLD_CACHE_CELL_LIFETIME);
            return world_cache_radiance[key].rgb;
        } else if existing_checksum == WORLD_CACHE_EMPTY_CELL {
            // Cell is empty - reset cell lifetime so that it starts getting updated next frame
            atomicStore(&world_cache_life[key], WORLD_CACHE_CELL_LIFETIME);
            world_cache_geometry_data[key].world_position = world_position;
            world_cache_geometry_data[key].world_normal = world_normal;
            return vec3(0.0);
        } else {
            // Collision - jump to another entry
            key = wrap_key(pcg_hash(key));
        }
    }

    return vec3(0.0);
}
#endif

fn get_cell_size(world_position: vec3<f32>, view_position: vec3<f32>) -> f32 {
    let camera_distance = distance(view_position, world_position) / WORLD_CACHE_POSITION_LOD_SCALE;
    let lod = exp2(floor(log2(1.0 + camera_distance)));
    return WORLD_CACHE_POSITION_BASE_CELL_SIZE * lod;
}

fn quantize_position(world_position: vec3<f32>, quantization_factor: f32) -> vec3<f32> {
    return floor(world_position / quantization_factor + 0.0001);
}

fn quantize_normal(world_normal: vec3<f32>) -> vec3<f32> {
    return floor(world_normal + 0.0001);
}

// TODO: Clustering
fn compute_key(world_position: vec3<u32>, world_normal: vec3<u32>) -> u32 {
    var key = pcg_hash(world_position.x);
    key = pcg_hash(key + world_position.y);
    key = pcg_hash(key + world_position.z);
    key = pcg_hash(key + world_normal.x);
    key = pcg_hash(key + world_normal.y);
    key = pcg_hash(key + world_normal.z);
    return wrap_key(key);
}

fn compute_checksum(world_position: vec3<u32>, world_normal: vec3<u32>) -> u32 {
    var key = iqint_hash(world_position.x);
    key = iqint_hash(key + world_position.y);
    key = iqint_hash(key + world_position.z);
    key = iqint_hash(key + world_normal.x);
    key = iqint_hash(key + world_normal.y);
    key = iqint_hash(key + world_normal.z);
    return key;
}

fn pcg_hash(input: u32) -> u32 {
    let state = input * 747796405u + 2891336453u;
    let word = ((state >> ((state >> 28u) + 4u)) ^ state) * 277803737u;
    return (word >> 22u) ^ word;
}

fn iqint_hash(input: u32) -> u32 {
    let n = (input << 13u) ^ input;
    return n * (n * n * 15731u + 789221u) + 1376312589u;
}

fn wrap_key(key: u32) -> u32 {
    return key & (#{WORLD_CACHE_SIZE} - 1u);
}