Use 4-phase graph-coloring for constraints

Split cloth constraints into 4 graph-colored phases (horizontal even/odd, vertical even/odd) instead of one big constraints array. Create dynamic JS arrays (constraintsP0..P3) with an addConstraint helper, allocate four GPU constraint buffers and four corresponding solve compute shaders (csSolve0..csSolve3) via a createSolver helper, and dispatch them per substep to avoid write-write races. Update integrate/velocity shader bindings setup and dispatch logic; keep positions/prevPositions/velocities buffers as before. In WGSL, mark constraints as read-only and use arrayLength(&constraints) to bound-check the constraint index instead of relying on a CPU-side count. Also tweak sim parameter (compliance lowered) and minor refactors/cleanups for clarity and consistency.
This commit is contained in:
2026-02-23 11:19:28 +01:00
parent 746022c48d
commit 728dbc047f
2 changed files with 77 additions and 102 deletions

View File

@@ -47,27 +47,30 @@ export class ClothComponent {
// Calculate approximate constraints (horizontal + vertical edges) // Calculate approximate constraints (horizontal + vertical edges)
const numConstraints = (gridWidth - 1) * gridHeight + gridWidth * (gridHeight - 1); const numConstraints = (gridWidth - 1) * gridHeight + gridWidth * (gridHeight - 1);
// --- 2. INITIALIZE CPU ARRAYS (Strict vec4<f32> alignment) ---
const positionsData = new Float32Array(numVertices * 4); const positionsData = new Float32Array(numVertices * 4);
const prevPositionsData = new Float32Array(numVertices * 4); const prevPositionsData = new Float32Array(numVertices * 4);
const velocitiesData = new Float32Array(numVertices * 4); const velocitiesData = new Float32Array(numVertices * 4);
const constraintsData = new Float32Array(numConstraints * 4);
// Fill Initial Positions // Arrays für unsere 4 Phasen (dynamische Größe, da wir pushen)
const constraintsP0: number[] = [];
const constraintsP1: number[] = [];
const constraintsP2: number[] = [];
const constraintsP3: number[] = [];
// Hilfsfunktion zum sauberen Hinzufügen (vec4-Struktur)
const addConstraint = (arr: number[], a: number, b: number) => {
arr.push(a, b, spacing, 1.0);
};
// Positionen füllen (bleibt wie vorher)
for (let y = 0; y < gridHeight; y++) { for (let y = 0; y < gridHeight; y++) {
for (let x = 0; x < gridWidth; x++) { for (let x = 0; x < gridWidth; x++) {
const idx = (y * gridWidth + x) * 4; const idx = (y * gridWidth + x) * 4;
positionsData[idx + 0] = (x - gridWidth / 2) * spacing;
positionsData[idx + 1] = 5.0 - (y * spacing);
positionsData[idx + 2] = 0.0;
positionsData[idx + 3] = (y === 0) ? 0.0 : 1.0; // Oben festpinnen
// Center the cloth around X=0, let it hang down in Y
positionsData[idx + 0] = (x - gridWidth / 2) * spacing; // X
positionsData[idx + 1] = 5.0 - (y * spacing); // Y (Start at height 5)
positionsData[idx + 2] = 0.0; // Z
// Inverse Mass (w-component): Pin the top row!
// If y == 0, mass is 0.0 (pinned). Otherwise 1.0 (moves freely)
positionsData[idx + 3] = (y === 0) ? 0.0 : 1.0;
// PrevPositions start identical
prevPositionsData[idx + 0] = positionsData[idx + 0]; prevPositionsData[idx + 0] = positionsData[idx + 0];
prevPositionsData[idx + 1] = positionsData[idx + 1]; prevPositionsData[idx + 1] = positionsData[idx + 1];
prevPositionsData[idx + 2] = positionsData[idx + 2]; prevPositionsData[idx + 2] = positionsData[idx + 2];
@@ -75,33 +78,25 @@ export class ClothComponent {
} }
} }
// Fill Constraints (Simple Grid: connect right and connect down) // --- GRAPH COLORING: Constraints in 4 Phasen füllen ---
let cIdx = 0; // Phase 0: Horizontal Gerade
for (let y = 0; y < gridHeight; y++) { for (let y = 0; y < gridHeight; y++) {
for (let x = 0; x < gridWidth; x++) { for (let x = 0; x < gridWidth - 1; x += 2) addConstraint(constraintsP0, y * gridWidth + x, y * gridWidth + x + 1);
const indexA = y * gridWidth + x;
// Connect to right neighbor
if (x < gridWidth - 1) {
constraintsData[cIdx * 4 + 0] = indexA; // Vertex A
constraintsData[cIdx * 4 + 1] = indexA + 1; // Vertex B
constraintsData[cIdx * 4 + 2] = spacing; // Rest length
constraintsData[cIdx * 4 + 3] = 1.0; // Active flag
cIdx++;
} }
// Connect to bottom neighbor // Phase 1: Horizontal Ungerade
if (y < gridHeight - 1) { for (let y = 0; y < gridHeight; y++) {
constraintsData[cIdx * 4 + 0] = indexA; // Vertex A for (let x = 1; x < gridWidth - 1; x += 2) addConstraint(constraintsP1, y * gridWidth + x, y * gridWidth + x + 1);
constraintsData[cIdx * 4 + 1] = indexA + gridWidth; // Vertex B
constraintsData[cIdx * 4 + 2] = spacing; // Rest length
constraintsData[cIdx * 4 + 3] = 1.0; // Active flag
cIdx++;
} }
// Phase 2: Vertikal Gerade
for (let y = 0; y < gridHeight - 1; y += 2) {
for (let x = 0; x < gridWidth; x++) addConstraint(constraintsP2, y * gridWidth + x, (y + 1) * gridWidth + x);
} }
// Phase 3: Vertikal Ungerade
for (let y = 1; y < gridHeight - 1; y += 2) {
for (let x = 0; x < gridWidth; x++) addConstraint(constraintsP3, y * gridWidth + x, (y + 1) * gridWidth + x);
} }
// Parameters Data const paramsData = new Float32Array(8);
const paramsData = new Float32Array(8); // Matches the WGSL struct (dt, gravity, etc.)
// --- 3. CREATE GPU STORAGE BUFFERS --- // --- 3. CREATE GPU STORAGE BUFFERS ---
const positionsBuffer = new StorageBuffer(engine, positionsData.byteLength); const positionsBuffer = new StorageBuffer(engine, positionsData.byteLength);
@@ -111,52 +106,38 @@ export class ClothComponent {
prevPositionsBuffer.update(prevPositionsData); prevPositionsBuffer.update(prevPositionsData);
const velocitiesBuffer = new StorageBuffer(engine, velocitiesData.byteLength); const velocitiesBuffer = new StorageBuffer(engine, velocitiesData.byteLength);
// Automatically initialized to 0 by WebGPU, no update needed initially
const constraintsBuffer = new StorageBuffer(engine, constraintsData.byteLength);
constraintsBuffer.update(constraintsData);
const paramsBuffer = new StorageBuffer(engine, paramsData.byteLength); const paramsBuffer = new StorageBuffer(engine, paramsData.byteLength);
// Erstelle 4 separate Buffer für die 4 Phasen
const cBuffer0 = new StorageBuffer(engine, constraintsP0.length * 4); cBuffer0.update(new Float32Array(constraintsP0));
const cBuffer1 = new StorageBuffer(engine, constraintsP1.length * 4); cBuffer1.update(new Float32Array(constraintsP1));
const cBuffer2 = new StorageBuffer(engine, constraintsP2.length * 4); cBuffer2.update(new Float32Array(constraintsP2));
const cBuffer3 = new StorageBuffer(engine, constraintsP3.length * 4); cBuffer3.update(new Float32Array(constraintsP3));
// --- 4. SETUP COMPUTE SHADERS --- // --- 4. SETUP COMPUTE SHADERS ---
const csIntegrate = new ComputeShader("integrate", engine, { computeSource: CLOTH_INTEGRATE_COMPUTE_WGSL }, { const csIntegrate = new ComputeShader("integrate", engine, { computeSource: CLOTH_INTEGRATE_COMPUTE_WGSL }, {
bindingsMapping: { bindingsMapping: { "p": { group: 0, binding: 0 }, "positions": { group: 0, binding: 1 }, "prev_positions": { group: 0, binding: 2 }, "velocities": { group: 0, binding: 3 } }
"p": { group: 0, binding: 0 },
"positions": { group: 0, binding: 1 },
"prev_positions": { group: 0, binding: 2 },
"velocities": { group: 0, binding: 3 }
}
}); });
csIntegrate.setStorageBuffer("p", paramsBuffer); csIntegrate.setStorageBuffer("p", paramsBuffer); csIntegrate.setStorageBuffer("positions", positionsBuffer); csIntegrate.setStorageBuffer("prev_positions", prevPositionsBuffer); csIntegrate.setStorageBuffer("velocities", velocitiesBuffer);
csIntegrate.setStorageBuffer("positions", positionsBuffer);
csIntegrate.setStorageBuffer("prev_positions", prevPositionsBuffer);
csIntegrate.setStorageBuffer("velocities", velocitiesBuffer);
// --- SETUP: csSolve (XPBD Constraints) --- // Hilfsfunktion, um die 4 Solve-Shader sauber zu erstellen
const csSolve = new ComputeShader("solve", engine, { computeSource: CLOTH_SOLVE_COMPUTE_WGSL }, { const createSolver = (name: string, cBuffer: StorageBuffer) => {
bindingsMapping: { const cs = new ComputeShader(name, engine, { computeSource: CLOTH_SOLVE_COMPUTE_WGSL }, {
"p": { group: 0, binding: 0 }, bindingsMapping: { "p": { group: 0, binding: 0 }, "positions": { group: 0, binding: 1 }, "constraints": { group: 0, binding: 2 } }
"positions": { group: 0, binding: 1 },
"constraints": { group: 0, binding: 2 }
}
}); });
csSolve.setStorageBuffer("p", paramsBuffer); cs.setStorageBuffer("p", paramsBuffer); cs.setStorageBuffer("positions", positionsBuffer); cs.setStorageBuffer("constraints", cBuffer);
csSolve.setStorageBuffer("positions", positionsBuffer); return cs;
csSolve.setStorageBuffer("constraints", constraintsBuffer); };
const csSolve0 = createSolver("solve0", cBuffer0);
const csSolve1 = createSolver("solve1", cBuffer1);
const csSolve2 = createSolver("solve2", cBuffer2);
const csSolve3 = createSolver("solve3", cBuffer3);
// --- SETUP: csVelocity (Update Velocities) ---
const csVelocity = new ComputeShader("velocity", engine, { computeSource: CLOTH_VELOCITY_COMPUTE_WGSL }, { const csVelocity = new ComputeShader("velocity", engine, { computeSource: CLOTH_VELOCITY_COMPUTE_WGSL }, {
bindingsMapping: { bindingsMapping: { "p": { group: 0, binding: 0 }, "positions": { group: 0, binding: 1 }, "prev_positions": { group: 0, binding: 2 }, "velocities": { group: 0, binding: 3 } }
"p": { group: 0, binding: 0 },
"positions": { group: 0, binding: 1 },
"prev_positions": { group: 0, binding: 2 },
"velocities": { group: 0, binding: 3 }
}
}); });
csVelocity.setStorageBuffer("p", paramsBuffer); csVelocity.setStorageBuffer("p", paramsBuffer); csVelocity.setStorageBuffer("positions", positionsBuffer); csVelocity.setStorageBuffer("prev_positions", prevPositionsBuffer); csVelocity.setStorageBuffer("velocities", velocitiesBuffer);
csVelocity.setStorageBuffer("positions", positionsBuffer);
csVelocity.setStorageBuffer("prev_positions", prevPositionsBuffer);
csVelocity.setStorageBuffer("velocities", velocitiesBuffer);
// --- 5. SETUP RENDER MESH --- // --- 5. SETUP RENDER MESH ---
// We create a ground mesh that matches our grid size, but we will OVERWRITE its vertices in the shader. // We create a ground mesh that matches our grid size, but we will OVERWRITE its vertices in the shader.
@@ -187,26 +168,27 @@ export class ClothComponent {
scene.onBeforeRenderObservable.clear(); scene.onBeforeRenderObservable.clear();
scene.onBeforeRenderObservable.add(() => { scene.onBeforeRenderObservable.add(() => {
// 1. Update Parameters (just an example, bind your simParams here) paramsData[0] = 0.016;
paramsData[0] = 0.016; // dt paramsData[1] = -9.81;
paramsData[1] = -9.81; // gravity paramsData[2] = 0.0001; // Compliance (sehr klein = steifer Stoff)
paramsData[2] = 0.001; // compliance (stiffness)
paramsData[3] = numVertices; paramsData[3] = numVertices;
paramsData[4] = numConstraints;
paramsBuffer.update(paramsData); paramsBuffer.update(paramsData);
// 2. Dispatch Compute Shaders in sequence!
const dispatchXVertices = Math.ceil(numVertices / 64); const dispatchXVertices = Math.ceil(numVertices / 64);
const dispatchXConstraints = Math.ceil(numConstraints / 64);
/*csIntegrate.dispatch(dispatchXVertices, 1, 1); // 1. Positionen vorhersehen
csIntegrate.dispatch(dispatchXVertices, 1, 1);
// For XPBD stability, you often run the solver multiple times (substeps) // 2. XPBD Solver (Substeps) - Jede Farbe einzeln lösen!
for (let i = 0; i < 5; i++) { for (let i = 0; i < 5; i++) {
csSolve.dispatch(dispatchXConstraints, 1, 1); csSolve0.dispatch(Math.ceil((constraintsP0.length / 4) / 64), 1, 1);
csSolve1.dispatch(Math.ceil((constraintsP1.length / 4) / 64), 1, 1);
csSolve2.dispatch(Math.ceil((constraintsP2.length / 4) / 64), 1, 1);
csSolve3.dispatch(Math.ceil((constraintsP3.length / 4) / 64), 1, 1);
} }
csVelocity.dispatch(dispatchXVertices, 1, 1);*/ // 3. Geschwindigkeiten aktualisieren
csVelocity.dispatch(dispatchXVertices, 1, 1);
}); });
} }
} }

View File

@@ -98,17 +98,18 @@ export const CLOTH_INTEGRATE_COMPUTE_WGSL = CLOTH_SHARED_STRUCTS + `
export const CLOTH_SOLVE_COMPUTE_WGSL = CLOTH_SHARED_STRUCTS + ` export const CLOTH_SOLVE_COMPUTE_WGSL = CLOTH_SHARED_STRUCTS + `
@group(0) @binding(0) var<storage, read> p : Params; @group(0) @binding(0) var<storage, read> p : Params;
@group(0) @binding(1) var<storage, read_write> positions : array<vec4<f32>>; @group(0) @binding(1) var<storage, read_write> positions : array<vec4<f32>>;
@group(0) @binding(2) var<storage, read_write> constraints : array<vec4<f32>>; @group(0) @binding(2) var<storage, read> constraints : array<vec4<f32>>; // <--- Nur "read", da wir sie hier nicht verändern
@compute @workgroup_size(64) @compute @workgroup_size(64)
fn main(@builtin(global_invocation_id) global_id : vec3<u32>) { fn main(@builtin(global_invocation_id) global_id : vec3<u32>) {
let idx = global_id.x; let idx = global_id.x;
if (f32(idx) >= p.numConstraints) { return; }
// HIER: Wir fragen die GPU direkt, wie groß das übergebene Array ist!
if (idx >= arrayLength(&constraints)) { return; }
let constraint = constraints[idx]; let constraint = constraints[idx];
let isActive = constraint.w; // 1.0 = Active, 0.0 = Cut/Broken let isActive = constraint.w;
// If the cloth is cut here, skip this constraint!
if (isActive < 0.5) { return; } if (isActive < 0.5) { return; }
let idA = u32(constraint.x); let idA = u32(constraint.x);
@@ -118,35 +119,27 @@ export const CLOTH_SOLVE_COMPUTE_WGSL = CLOTH_SHARED_STRUCTS + `
var pA = positions[idA]; var pA = positions[idA];
var pB = positions[idB]; var pB = positions[idB];
let wA = pA.w; // Inverse mass A let wA = pA.w;
let wB = pB.w; // Inverse mass B let wB = pB.w;
let wSum = wA + wB; let wSum = wA + wB;
// If both points are pinned, do nothing
if (wSum <= 0.0) { return; } if (wSum <= 0.0) { return; }
let dir = pA.xyz - pB.xyz; let dir = pA.xyz - pB.xyz;
let dist = length(dir); let dist = length(dir);
// Prevent division by zero
if (dist < 0.0001) { return; } if (dist < 0.0001) { return; }
// XPBD Calculation (Extended Position-Based Dynamics)
let n = dir / dist; let n = dir / dist;
let C = dist - restLength; // Constraint violation (how much it stretched) let C = dist - restLength;
// Calculate the correction factor (alpha represents the XPBD compliance)
let alpha = p.compliance / (p.dt * p.dt); let alpha = p.compliance / (p.dt * p.dt);
let lambda = -C / (wSum + alpha); let lambda = -C / (wSum + alpha);
// Apply position corrections directly to the points
let corrA = n * (lambda * wA); let corrA = n * (lambda * wA);
let corrB = n * (-lambda * wB); let corrB = n * (-lambda * wB);
// NOTE: In a multi-threaded GPU environment without "Graph Coloring", // This is because we are using graph coloring to be thread safe
// writing directly to positions like this can cause minor race conditions
// (flickering). We will handle Graph Coloring in the TypeScript setup!
if (wA > 0.0) { if (wA > 0.0) {
positions[idA].x = positions[idA].x + corrA.x; positions[idA].x = positions[idA].x + corrA.x;
positions[idA].y = positions[idA].y + corrA.y; positions[idA].y = positions[idA].y + corrA.y;