jpmschweitzer/settled-reach
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

perf(simulation): expose flow accumulation + optimize basin merge in D8 drainage (#953)

Browse Source 
- DrainageResult now exposes flow_accumulation + max_accumulation for
  D-209 attractor-strength normalization.
- Rewrite merge_small_basins from an O(merges x n) loop (rescanned the
  whole grid per merge) to an adjacency-graph + union-find pass: one grid
  scan, lazy merges. Cuts D8 drainage at 512x256 from ~299ms to ~45ms,
  meeting the D-208 ~50ms target (the module had never been run at
  canonical resolution before — it was orphaned). Determinism preserved
  (smallest by (size,id), largest neighbor by (size, lowest id)).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
jpmschweitzer
2026-05-23 00:35:16 +02:00
co-authored by Claude Opus 4.7
parent a409360802
commit 7cf7614342
1 changed files with 113 additions and 76 deletions
Download Patch File Download Diff File Expand all files Collapse all files
server/src/atlas/drainage.rs
+113 -76
View File
@@ -45,6 +45,13 @@ const D8: [(i32, i32); 8] = [
pub struct DrainageResult {
pub river_network: RiverNetwork,
pub drainage_basins: Vec<DrainageBasin>,
/// Per-cell flow accumulation (row-major, `w × h`): the upstream cell count
/// draining through each cell. Exposed for D-209 attractor-strength
/// normalization (`flow_accumulation[cell] / max_accumulation`).
pub flow_accumulation: Vec<i32>,
/// Maximum flow accumulation across the grid — the denominator for
/// normalized attractor strength (D-209). Always ≥ 1.
pub max_accumulation: i32,
}
// ---------------------------------------------------------------------------
@@ -87,9 +94,15 @@ pub fn analyze(elevation: &[f32], width: u32, height: u32, sea_level: f32) -> Dr
// 8. Build DrainageBasin structs.
let drainage_basins = build_basins(&labels, w, h);
// Max accumulation for D-209 strength normalization (clamped ≥ 1 so the
// division is always well-defined, even on a flat/empty world).
let max_accumulation = accum.iter().copied().max().unwrap_or(1).max(1);
DrainageResult {
river_network,
drainage_basins,
flow_accumulation: accum,
max_accumulation,
}
}
@@ -378,6 +391,30 @@ fn label_basins(fdir: &[i8], accum: &[i32], w: usize, h: usize) -> Vec<i32> {
// Step 7: Merge small basins
// ---------------------------------------------------------------------------
/// Union-find root with path compression.
fn uf_find(parent: &mut [i32], x: i32) -> i32 {
let mut root = x;
while parent[root as usize] != root {
root = parent[root as usize];
}
let mut cur = x;
while parent[cur as usize] != root {
let next = parent[cur as usize];
parent[cur as usize] = root;
cur = next;
}
root
}
/// Merge small basins into their largest neighbor until the count is in
/// `[min_count, max_count]` and every basin holds ≥ 2% of the surface.
///
/// Builds a basin adjacency graph + sizes in a single grid pass, then performs
/// all merges as union-find operations on that graph — the grid is rewritten
/// exactly once at the end. This replaces the former O(merges × n) loop (which
/// rescanned the whole grid per merge: ~250ms at 512×256) with O(n + merges).
/// Determinism: smallest basin chosen by `(size, id)`, largest neighbor by
/// `(size, then lowest id)` — both fixed orders.
fn merge_small_basins(
mut labels: Vec<i32>,
w: usize,
@@ -385,52 +422,91 @@ fn merge_small_basins(
min_count: usize,
max_count: usize,
) -> Vec<i32> {
use std::collections::BTreeSet;
let n = w * h;
let min_frac = 0.02f64; // 2% minimum basin area
for _ in 0..200 {
// Count basin sizes.
let mut sizes: std::collections::BTreeMap<i32, usize> = std::collections::BTreeMap::new();
for &l in &labels {
*sizes.entry(l).or_insert(0) += 1;
}
let n_basins = sizes.len();
let max_label = labels.iter().copied().max().unwrap_or(0);
let nb = (max_label + 1) as usize;
if nb <= 1 {
return labels; // single basin — nothing to merge
}
// Stop if within target range and all basins are large enough.
if n_basins <= max_count && sizes.values().all(|&s| s as f64 / n as f64 >= min_frac) {
break;
}
if n_basins <= min_count {
break;
}
// Find the smallest basin.
let (&smallest_id, &smallest_size) = sizes.iter().min_by_key(|(_, &s)| s).unwrap();
if n_basins <= max_count && smallest_size as f64 / n as f64 >= min_frac {
break;
}
// Find its largest adjacent basin.
let nbr_id = find_largest_neighbor(&labels, smallest_id, &sizes, w, h);
let merge_into = nbr_id.unwrap_or(0);
// Merge.
for l in labels.iter_mut() {
if *l == smallest_id {
*l = merge_into;
// One pass: basin sizes + adjacency (neighbor labels per basin).
let mut size = vec![0usize; nb];
let mut adj: Vec<BTreeSet<i32>> = vec![BTreeSet::new(); nb];
for r in 0..h {
for c in 0..w {
let l = labels[r * w + c];
size[l as usize] += 1;
for &(dr, dc) in &D8 {
let nr = r as i32 + dr;
let nc = (c as i32 + dc).rem_euclid(w as i32) as usize;
if nr >= 0 && nr < h as i32 {
let nl = labels[nr as usize * w + nc];
if nl != l {
adj[l as usize].insert(nl);
}
}
}
}
}
// Renumber contiguously from 0.
let unique: Vec<i32> = {
let mut set: std::collections::BTreeSet<i32> = std::collections::BTreeSet::new();
for &l in &labels {
set.insert(l);
let mut parent: Vec<i32> = (0..nb as i32).collect();
let mut active: BTreeSet<i32> = (0..nb as i32).collect();
while active.len() > min_count {
// Smallest active basin (tie → lowest id; BTreeSet iterates ascending).
let smallest = *active.iter().min_by_key(|&&b| (size[b as usize], b)).unwrap();
let smallest_size = size[smallest as usize];
if active.len() <= max_count && smallest_size as f64 / n as f64 >= min_frac {
break;
}
set.into_iter().collect()
};
// Largest active neighbor (tie → lowest id).
let mut best: i32 = -1;
let mut best_size = 0usize;
for &nb_lbl in &adj[smallest as usize] {
let rep = uf_find(&mut parent, nb_lbl);
if rep == smallest {
continue;
}
let s = size[rep as usize];
if s > best_size || (s == best_size && (best < 0 || rep < best)) {
best_size = s;
best = rep;
}
}
// No neighbor (isolated basin) → merge into the next smallest active.
let merge_into = if best >= 0 {
best
} else {
match active.iter().find(|&&b| b != smallest) {
Some(&other) => other,
None => break,
}
};
// Union smallest → merge_into; fold size and adjacency.
parent[smallest as usize] = merge_into;
size[merge_into as usize] += smallest_size;
let small_adj = std::mem::take(&mut adj[smallest as usize]);
for nb_lbl in small_adj {
let rep = uf_find(&mut parent, nb_lbl);
if rep != merge_into {
adj[merge_into as usize].insert(rep);
}
}
active.remove(&smallest);
}
// Resolve every cell to its basin representative (single pass).
for l in labels.iter_mut() {
*l = uf_find(&mut parent, *l);
}
// Renumber contiguously from 0.
let unique: BTreeSet<i32> = labels.iter().copied().collect();
let remap: std::collections::BTreeMap<i32, i32> = unique
.iter()
.enumerate()
@@ -443,45 +519,6 @@ fn merge_small_basins(
labels
}
fn find_largest_neighbor(
labels: &[i32],
target_id: i32,
sizes: &std::collections::BTreeMap<i32, usize>,
w: usize,
h: usize,
) -> Option<i32> {
let n = w * h;
let mut neighbor_sizes: std::collections::BTreeMap<i32, usize> =
std::collections::BTreeMap::new();
for i in 0..n {
if labels[i] != target_id {
continue;
}
let r = i / w;
let c = i % w;
for &(dr, dc) in &D8 {
let nr = r as i32 + dr;
let nc = (c as i32 + dc).rem_euclid(w as i32) as usize;
if nr >= 0 && nr < h as i32 {
let nbr_id = labels[nr as usize * w + nc];
if nbr_id != target_id {
let size = sizes.get(&nbr_id).copied().unwrap_or(0);
let e = neighbor_sizes.entry(nbr_id).or_insert(0);
if size > *e {
*e = size;
}
}
}
}
}
neighbor_sizes
.into_iter()
.max_by_key(|(_, s)| *s)
.map(|(id, _)| id)
}
// ---------------------------------------------------------------------------
// Step 8: Build DrainageBasin structs
// ---------------------------------------------------------------------------