@driftengine/core 4.8.3 → 4.8.5

package/dist/render/backend/webgpu/shaders/clothSolve.wgsl.d.ts

@@ -0,0 +1,33 @@
+/**
+ * The skinned cloth's solver on the device, in WGSL, hand-authored under the compute exception
+ * `AGENTS.md` records (2026-08-24): a compute shader has no WebGL2 twin to generate from, and
+ * `npm run wgsl` neither writes nor polices this file.
+ *
+ * **Written to be read beside `@driftengine/physics`' `skinnedClothSolve.ts` and `clothLimits.ts`**:
+ * each kernel there is an entry point here, one invocation a particle or a constraint, and the loop
+ * over a batch is the dispatch. `clothPose.ts`' `skinParticles` is `pose`. The order of a step —
+ * begin, then per substep predict, the batches, the limits, finish; then the blend — is
+ * `gpuSkinnedCloth.ts`', taken from the same `ClothControl` the CPU solver is driven by.
+ *
+ * **Where it differs, it says so.** Single precision throughout, where the CPU rounds to single
+ * only on a store. And the bending angle is `atan2(|n1 × n2|, n1 · n2)` rather than `acos(n1 · n2)`
+ * with its sine from `sqrt(1 − d²)`: the same function, but a garment spends its life near flat,
+ * where `d` is near −1 and `1 − d²` in single precision is rounding — a sine of a few ten-thousandths
+ * out of nothing, and a fold pushed by it. The cross product's length is exact there.
+ *
+ * **How far the two solvers part, measured** (`demo/dev/skinnedCloth.html?parity=1`, RX 9070 XT):
+ * a cape hanging in still air agrees to 0.016 mm after 150 frames. With tethers, backstops and a
+ * gusting wind it parts by up to 11 mm — and the CPU solver against itself, nudged a tenth of a
+ * micrometre a frame, parts by 9 mm at the same frame, and by as much as the device at every
+ * checkpoint from the sixtieth on. Contact, buckling and gusts carry a rounding difference that
+ * far; the device adds rounding and nothing else. Dropping one multiplier reset parts them by
+ * 67 mm, which is the instrument seeing a real difference.
+ *
+ * **Every particle's state in a few buffers of sections**, because WebGPU grants eight storage
+ * buffers to a stage: `state` is the positions, the substep's start, the velocities and the step's
+ * start; `pose` the two skinned poses a step interpolates between and the latest skinned normals;
+ * `statics` everything per particle that never changes. A section's offset is a multiple of the
+ * particle count, which the constants carry — so no offset is a uniform to keep in step.
+ */
+export declare const CLOTH_WORKGROUP = 64;
+export declare const CLOTH_SOLVE_WGSL = "\nstruct Constants {\n  count: u32,\n  distances: u32,\n  bendings: u32,\n  joints: u32,\n  colliders: u32,\n  /* 1 a max distance, 2 a backstop, 4 a frontstop: which limits the set-up has at all. */\n  limits: u32,\n  width: u32,\n  pad0: u32,\n  h: f32,\n  invH: f32,\n  invH2: f32,\n  keep: f32,\n  dragKeep: f32,\n  maxDistanceScale: f32,\n  margin: f32,\n  pad1: f32,\n  gravity: vec4<f32>,\n}\n\n/* One step's numbers, at a dynamic offset: every step of a frame has its own slot. */\nstruct Step {\n  fraction: f32,\n  blend: f32,\n  alpha: f32,\n  carry: u32,\n  /* Where the pose a step starts from and the one it moves toward sit in pose, in floats. */\n  fromAt: u32,\n  toAt: u32,\n  pad0: u32,\n  pad1: u32,\n  /* The character's turn since the last step, by columns; origin.w is the share carried. */\n  turn0: vec4<f32>,\n  turn1: vec4<f32>,\n  turn2: vec4<f32>,\n  origin: vec4<f32>,\n  shift: vec4<f32>,\n  /* The frame's wind: one wind, sampled once a frame, so a step's rather than the set-up's. */\n  wind: vec4<f32>,\n}\n\n/* One batch of one kind of constraint, at a dynamic offset; first resets the multipliers. */\nstruct Batch {\n  begin: u32,\n  end: u32,\n  first: u32,\n  pad: u32,\n}\n\n@group(0) @binding(0) var<uniform> c: Constants;\n@group(0) @binding(1) var<uniform> pace: Step;\n@group(0) @binding(2) var<uniform> batch: Batch;\n@group(0) @binding(3) var<storage, read_write> state: array<f32>;\n@group(0) @binding(4) var<storage, read_write> pose: array<f32>;\n@group(0) @binding(5) var<storage, read> statics: array<f32>;\n@group(0) @binding(6) var<storage, read> constraints: array<u32>;\n@group(0) @binding(7) var<storage, read_write> lambda: array<f32>;\n@group(0) @binding(8) var<storage, read> frame: array<f32>;\n\nfn stateAt(section: u32, i: u32) -> u32 {\n  return section * 3u * c.count + i * 3u;\n}\n\nfn load(at: u32) -> vec3<f32> {\n  return vec3<f32>(state[at], state[at + 1u], state[at + 2u]);\n}\n\nfn store(at: u32, v: vec3<f32>) {\n  state[at] = v.x;\n  state[at + 1u] = v.y;\n  state[at + 2u] = v.z;\n}\n\nfn poseOf(at: u32) -> vec3<f32> {\n  return vec3<f32>(pose[at], pose[at + 1u], pose[at + 2u]);\n}\n\n/* Sections of state. */\nconst POSITION = 0u;\nconst PREVIOUS = 1u;\nconst VELOCITY = 2u;\nconst START = 3u;\n\nfn inverseMass(i: u32) -> f32 {\n  return statics[i];\n}\n\n/* Where skinning puts particle i, fraction of the way from the last pose to the latest. */\nfn skinnedAt(i: u32) -> vec3<f32> {\n  let a = poseOf(pace.fromAt + i * 3u);\n  let b = poseOf(pace.toAt + i * 3u);\n  return a + (b - a) * pace.fraction;\n}\n\nfn skinMatrix(j: u32) -> mat4x4<f32> {\n  let at = j * 16u;\n  return mat4x4<f32>(\n    vec4<f32>(frame[at], frame[at + 1u], frame[at + 2u], frame[at + 3u]),\n    vec4<f32>(frame[at + 4u], frame[at + 5u], frame[at + 6u], frame[at + 7u]),\n    vec4<f32>(frame[at + 8u], frame[at + 9u], frame[at + 10u], frame[at + 11u]),\n    vec4<f32>(frame[at + 12u], frame[at + 13u], frame[at + 14u], frame[at + 15u]));\n}\n\n/* skinParticles: each particle's skinned position and normal by up to eight influences. */\n@compute @workgroup_size(64)\nfn clothPose(@builtin(global_invocation_id) id: vec3<u32>) {\n  let i = id.x;\n  if (i >= c.count) {\n    return;\n  }\n  let n = c.count;\n  let restAt = 7u * n + i * 3u;\n  let normalAt = 10u * n + i * 3u;\n  let x = vec4<f32>(statics[restAt], statics[restAt + 1u], statics[restAt + 2u], 1.0);\n  let q0 = vec3<f32>(statics[normalAt], statics[normalAt + 1u], statics[normalAt + 2u]);\n  var p = vec3<f32>(0.0);\n  var q = vec3<f32>(0.0);\n  for (var half = 0u; half < 2u; half++) {\n    for (var k = 0u; k < 4u; k++) {\n      let at = 13u * n + i * 16u + half * 8u;\n      let w = statics[at + 4u + k];\n      if (w == 0.0) {\n        continue;\n      }\n      let m = skinMatrix(u32(statics[at + k]));\n      p += w * (m * x).xyz;\n      q += w * (mat3x3<f32>(m[0].xyz, m[1].xyz, m[2].xyz) * q0);\n    }\n  }\n  let to = pace.toAt + i * 3u;\n  pose[to] = p.x;\n  pose[to + 1u] = p.y;\n  pose[to + 2u] = p.z;\n  let len = sqrt(dot(q, q));\n  let inv = select(0.0, 1.0 / len, len > 1e-12);\n  let normal = 6u * n + i * 3u;\n  pose[normal] = q.x * inv;\n  pose[normal + 1u] = q.y * inv;\n  pose[normal + 2u] = q.z * inv;\n}\n\n/* SkinnedCloth.rest: every particle at the latest skinned position, at rest. */\n@compute @workgroup_size(64)\nfn clothRest(@builtin(global_invocation_id) id: vec3<u32>) {\n  let i = id.x;\n  if (i >= c.count) {\n    return;\n  }\n  let p = poseOf(pace.toAt + i * 3u);\n  store(stateAt(POSITION, i), p);\n  store(stateAt(START, i), p);\n  store(stateAt(VELOCITY, i), vec3<f32>(0.0));\n}\n\n/* The step's start: the step start kept, a free particle carried, a kinematic one placed. */\n@compute @workgroup_size(64)\nfn clothBegin(@builtin(global_invocation_id) id: vec3<u32>) {\n  let i = id.x;\n  if (i >= c.count) {\n    return;\n  }\n  let at = stateAt(POSITION, i);\n  let p = load(at);\n  store(stateAt(START, i), p);\n  if (inverseMass(i) == 0.0) {\n    store(at, skinnedAt(i));\n    return;\n  }\n  if (pace.carry == 0u) {\n    return;\n  }\n  /* carryParticles: the share of the character's turn about its old origin, and of its move. */\n  let turn = mat3x3<f32>(pace.turn0.xyz, pace.turn1.xyz, pace.turn2.xyz);\n  let share = pace.origin.w;\n  let x = p - pace.origin.xyz;\n  store(at, p + (turn * x - x) * share + pace.shift.xyz);\n  let vAt = stateAt(VELOCITY, i);\n  let v = load(vAt);\n  store(vAt, v + (turn * v - v) * share);\n}\n\n/* predictParticles. */\n@compute @workgroup_size(64)\nfn clothPredict(@builtin(global_invocation_id) id: vec3<u32>) {\n  let i = id.x;\n  if (i >= c.count) {\n    return;\n  }\n  let at = stateAt(POSITION, i);\n  let p = load(at);\n  store(stateAt(PREVIOUS, i), p);\n  if (inverseMass(i) == 0.0) {\n    return;\n  }\n  let vAt = stateAt(VELOCITY, i);\n  let air = pace.wind.xyz;\n  let v = air + (load(vAt) * c.keep - air) * c.dragKeep + c.gravity.xyz * c.h;\n  store(vAt, v);\n  store(at, p + v * c.h);\n}\n\n/* solveDistanceBatch: one invocation a constraint of the batch, constraints sorted by batch. */\n@compute @workgroup_size(64)\nfn clothDistance(@builtin(global_invocation_id) id: vec3<u32>) {\n  let k = batch.begin + id.x;\n  if (k >= batch.end) {\n    return;\n  }\n  /* The first pass of a substep starts the multiplier at nothing, as the CPU's fill(0) does \u2014\n     written before any early return, so a skipped constraint does not carry the last one. */\n  if (batch.first == 1u) {\n    lambda[k] = 0.0;\n  }\n  let a = constraints[k * 4u];\n  let b = constraints[k * 4u + 1u];\n  let wa = inverseMass(a);\n  let wb = inverseMass(b);\n  let w = wa + wb;\n  if (w == 0.0) {\n    return;\n  }\n  let aAt = stateAt(POSITION, a);\n  let bAt = stateAt(POSITION, b);\n  let d = load(bAt) - load(aAt);\n  let len = sqrt(dot(d, d));\n  if (len < 1e-9) {\n    return;\n  }\n  let alpha = bitcast<f32>(constraints[k * 4u + 3u]) * c.invH2;\n  let old = lambda[k];\n  let delta = (-(len - bitcast<f32>(constraints[k * 4u + 2u])) - alpha * old) / (w + alpha);\n  lambda[k] = old + delta;\n  let s = delta / len;\n  /* A pinned particle is shared across a batch, so it is never written, even with itself. */\n  if (wa > 0.0) {\n    store(aAt, load(aAt) - d * s * wa);\n  }\n  if (wb > 0.0) {\n    store(bAt, load(bAt) + d * s * wb);\n  }\n}\n\n/* solveBendingBatch, with the angle from atan2 (see the header). */\n@compute @workgroup_size(64)\nfn clothBending(@builtin(global_invocation_id) id: vec3<u32>) {\n  let k = batch.begin + id.x;\n  if (k >= batch.end) {\n    return;\n  }\n  let at = c.distances * 4u + k * 6u;\n  let slot = c.distances + k;\n  if (batch.first == 1u) {\n    lambda[slot] = 0.0;\n  }\n  let ids = vec4<u32>(constraints[at], constraints[at + 1u], constraints[at + 2u], constraints[at + 3u]);\n  let w = vec4<f32>(inverseMass(ids.x), inverseMass(ids.y), inverseMass(ids.z), inverseMass(ids.w));\n  if (w.x + w.y + w.z + w.w == 0.0) {\n    return;\n  }\n  let origin = load(stateAt(POSITION, ids.x));\n  let e = load(stateAt(POSITION, ids.y)) - origin;\n  let p3 = load(stateAt(POSITION, ids.z)) - origin;\n  let p4 = load(stateAt(POSITION, ids.w)) - origin;\n  var n1 = cross(e, p3);\n  var n2 = cross(e, p4);\n  let l1 = sqrt(dot(n1, n1));\n  let l2 = sqrt(dot(n2, n2));\n  if (l1 < 1e-12 || l2 < 1e-12) {\n    return;\n  }\n  n1 = n1 / l1;\n  n2 = n2 / l2;\n  let d = clamp(dot(n1, n2), -1.0, 1.0);\n  let sine = sqrt(dot(cross(n1, n2), cross(n1, n2)));\n  if (sine < 1e-6) {\n    return;\n  }\n  /* The q vectors: the negative of the derivative of n1 \u00B7 n2. See solveBendingBatch. */\n  let q3 = (cross(e, n2) + cross(n1, e) * d) / l1;\n  let q4 = (cross(e, n1) + cross(n2, e) * d) / l2;\n  let q2 = -((cross(p3, n2) + cross(n1, p3) * d) / l1 + (cross(p4, n1) + cross(n2, p4) * d) / l2);\n  let q1 = -(q2 + q3 + q4);\n  var sum = w.x * dot(q1, q1) + w.y * dot(q2, q2) + w.z * dot(q3, q3) + w.w * dot(q4, q4);\n  sum = sum / (sine * sine);\n  if (sum < 1e-12) {\n    return;\n  }\n  let alpha = bitcast<f32>(constraints[at + 5u]) * c.invH2;\n  let old = lambda[slot];\n  let angle = atan2(sine, d);\n  let delta = (-(angle - bitcast<f32>(constraints[at + 4u])) - alpha * old) / (sum + alpha);\n  lambda[slot] = old + delta;\n  let scale = delta / sine;\n  if (w.x > 0.0) {\n    let p = stateAt(POSITION, ids.x);\n    store(p, load(p) + w.x * scale * q1);\n  }\n  if (w.y > 0.0) {\n    let p = stateAt(POSITION, ids.y);\n    store(p, load(p) + w.y * scale * q2);\n  }\n  if (w.z > 0.0) {\n    let p = stateAt(POSITION, ids.z);\n    store(p, load(p) + w.z * scale * q3);\n  }\n  if (w.w > 0.0) {\n    let p = stateAt(POSITION, ids.w);\n    store(p, load(p) + w.w * scale * q4);\n  }\n}\n\n/* solveTethers, one invocation a tether of the batch: a batch holds a particle once. */\n@compute @workgroup_size(64)\nfn clothTether(@builtin(global_invocation_id) id: vec3<u32>) {\n  let k = batch.begin + id.x;\n  if (k >= batch.end) {\n    return;\n  }\n  let at = c.distances * 4u + c.bendings * 6u + k * 3u;\n  let pAt = stateAt(POSITION, constraints[at]);\n  let anchor = load(stateAt(POSITION, constraints[at + 1u]));\n  let d = load(pAt) - anchor;\n  let dist = sqrt(dot(d, d));\n  let reach = bitcast<f32>(constraints[at + 2u]);\n  if (dist <= reach || dist < 1e-9) {\n    return;\n  }\n  store(pAt, anchor + d * (reach / dist));\n}\n\n/* pushOut: the point onto the sphere's surface if it is inside. A point at the centre stays. */\nfn pushOut(p: vec3<f32>, centre: vec3<f32>, radius: f32) -> vec3<f32> {\n  let d = p - centre;\n  let d2 = dot(d, d);\n  if (d2 >= radius * radius || d2 < 1e-18) {\n    return p;\n  }\n  return centre + d * (radius / sqrt(d2));\n}\n\n/* applyMaxDistance, applyStops (back, then front) and applyColliders, in that order. */\n@compute @workgroup_size(64)\nfn clothLimit(@builtin(global_invocation_id) id: vec3<u32>) {\n  let i = id.x;\n  if (i >= c.count || inverseMass(i) == 0.0) {\n    return;\n  }\n  let n = c.count;\n  let at = stateAt(POSITION, i);\n  var p = load(at);\n  let t = skinnedAt(i);\n  /* A max distance below zero is none: the CPU's Infinity, which a device may not be handed. */\n  let maxDistance = statics[n + i];\n  if ((c.limits & 1u) != 0u && maxDistance >= 0.0) {\n    let limit = maxDistance * c.maxDistanceScale;\n    let d = p - t;\n    let d2 = dot(d, d);\n    if (d2 > limit * limit) {\n      p = t + d * (limit / sqrt(d2));\n    }\n  }\n  let normal = poseOf(6u * n + i * 3u);\n  if ((c.limits & 2u) != 0u) {\n    let radius = statics[2u * n + i * 2u + 1u];\n    if (radius > 0.0) {\n      p = pushOut(p, t - normal * (statics[2u * n + i * 2u] + radius), radius);\n    }\n  }\n  if ((c.limits & 4u) != 0u) {\n    let radius = statics[4u * n + i * 2u + 1u];\n    if (radius > 0.0) {\n      p = pushOut(p, t + normal * (statics[4u * n + i * 2u] + radius), radius);\n    }\n  }\n  let pad = statics[6u * n + i] + c.margin;\n  let ends = c.joints * 16u;\n  let radii = ends + c.colliders * 6u;\n  for (var k = 0u; k < c.colliders; k++) {\n    let a = vec3<f32>(frame[ends + k * 6u], frame[ends + k * 6u + 1u], frame[ends + k * 6u + 2u]);\n    let e = vec3<f32>(frame[ends + k * 6u + 3u], frame[ends + k * 6u + 4u], frame[ends + k * 6u + 5u]) - a;\n    let e2 = dot(e, e);\n    var s = 0.0;\n    if (e2 > 1e-18) {\n      s = clamp(dot(p - a, e) / e2, 0.0, 1.0);\n    }\n    let r0 = frame[radii + k * 2u];\n    let r1 = frame[radii + k * 2u + 1u];\n    p = pushOut(p, a + e * s, r0 + (r1 - r0) * s + pad);\n  }\n  store(at, p);\n}\n\n/* finishParticles. */\n@compute @workgroup_size(64)\nfn clothFinish(@builtin(global_invocation_id) id: vec3<u32>) {\n  let i = id.x;\n  if (i >= c.count) {\n    return;\n  }\n  let vAt = stateAt(VELOCITY, i);\n  if (inverseMass(i) == 0.0) {\n    store(vAt, vec3<f32>(0.0));\n    return;\n  }\n  store(vAt, (load(stateAt(POSITION, i)) - load(stateAt(PREVIOUS, i))) * c.invH);\n}\n\n/* SkinnedCloth.blend: after a reset, the simulation's share grows over the blend's steps. */\n@compute @workgroup_size(64)\nfn clothBlend(@builtin(global_invocation_id) id: vec3<u32>) {\n  let i = id.x;\n  if (i >= c.count) {\n    return;\n  }\n  let t = skinnedAt(i);\n  let at = stateAt(POSITION, i);\n  store(at, t + (load(at) - t) * pace.blend);\n  let vAt = stateAt(VELOCITY, i);\n  store(vAt, load(vAt) * pace.blend);\n}\n\n@group(1) @binding(0) var output: texture_storage_2d<rgba32float, write>;\n\n/* SkinnedCloth.interpolate, into the particle texture the vertex stage reads. */\n@compute @workgroup_size(64)\nfn clothPublish(@builtin(global_invocation_id) id: vec3<u32>) {\n  let i = id.x;\n  if (i >= c.count) {\n    return;\n  }\n  let a = load(stateAt(START, i));\n  let p = a + (load(stateAt(POSITION, i)) - a) * pace.alpha;\n  textureStore(output, vec2<u32>(i % c.width, i / c.width), vec4<f32>(p, 0.0));\n}\n";

package/dist/render/backend/webgpu/shaders/clothSolve.wgsl.js

@@ -0,0 +1,469 @@
+/**
+ * The skinned cloth's solver on the device, in WGSL, hand-authored under the compute exception
+ * `AGENTS.md` records (2026-08-24): a compute shader has no WebGL2 twin to generate from, and
+ * `npm run wgsl` neither writes nor polices this file.
+ *
+ * **Written to be read beside `@driftengine/physics`' `skinnedClothSolve.ts` and `clothLimits.ts`**:
+ * each kernel there is an entry point here, one invocation a particle or a constraint, and the loop
+ * over a batch is the dispatch. `clothPose.ts`' `skinParticles` is `pose`. The order of a step —
+ * begin, then per substep predict, the batches, the limits, finish; then the blend — is
+ * `gpuSkinnedCloth.ts`', taken from the same `ClothControl` the CPU solver is driven by.
+ *
+ * **Where it differs, it says so.** Single precision throughout, where the CPU rounds to single
+ * only on a store. And the bending angle is `atan2(|n1 × n2|, n1 · n2)` rather than `acos(n1 · n2)`
+ * with its sine from `sqrt(1 − d²)`: the same function, but a garment spends its life near flat,
+ * where `d` is near −1 and `1 − d²` in single precision is rounding — a sine of a few ten-thousandths
+ * out of nothing, and a fold pushed by it. The cross product's length is exact there.
+ *
+ * **How far the two solvers part, measured** (`demo/dev/skinnedCloth.html?parity=1`, RX 9070 XT):
+ * a cape hanging in still air agrees to 0.016 mm after 150 frames. With tethers, backstops and a
+ * gusting wind it parts by up to 11 mm — and the CPU solver against itself, nudged a tenth of a
+ * micrometre a frame, parts by 9 mm at the same frame, and by as much as the device at every
+ * checkpoint from the sixtieth on. Contact, buckling and gusts carry a rounding difference that
+ * far; the device adds rounding and nothing else. Dropping one multiplier reset parts them by
+ * 67 mm, which is the instrument seeing a real difference.
+ *
+ * **Every particle's state in a few buffers of sections**, because WebGPU grants eight storage
+ * buffers to a stage: `state` is the positions, the substep's start, the velocities and the step's
+ * start; `pose` the two skinned poses a step interpolates between and the latest skinned normals;
+ * `statics` everything per particle that never changes. A section's offset is a multiple of the
+ * particle count, which the constants carry — so no offset is a uniform to keep in step.
+ */
+export const CLOTH_WORKGROUP = 64;
+export const CLOTH_SOLVE_WGSL = /* wgsl */ `
+struct Constants {
+  count: u32,
+  distances: u32,
+  bendings: u32,
+  joints: u32,
+  colliders: u32,
+  /* 1 a max distance, 2 a backstop, 4 a frontstop: which limits the set-up has at all. */
+  limits: u32,
+  width: u32,
+  pad0: u32,
+  h: f32,
+  invH: f32,
+  invH2: f32,
+  keep: f32,
+  dragKeep: f32,
+  maxDistanceScale: f32,
+  margin: f32,
+  pad1: f32,
+  gravity: vec4<f32>,
+}
+
+/* One step's numbers, at a dynamic offset: every step of a frame has its own slot. */
+struct Step {
+  fraction: f32,
+  blend: f32,
+  alpha: f32,
+  carry: u32,
+  /* Where the pose a step starts from and the one it moves toward sit in pose, in floats. */
+  fromAt: u32,
+  toAt: u32,
+  pad0: u32,
+  pad1: u32,
+  /* The character's turn since the last step, by columns; origin.w is the share carried. */
+  turn0: vec4<f32>,
+  turn1: vec4<f32>,
+  turn2: vec4<f32>,
+  origin: vec4<f32>,
+  shift: vec4<f32>,
+  /* The frame's wind: one wind, sampled once a frame, so a step's rather than the set-up's. */
+  wind: vec4<f32>,
+}
+
+/* One batch of one kind of constraint, at a dynamic offset; first resets the multipliers. */
+struct Batch {
+  begin: u32,
+  end: u32,
+  first: u32,
+  pad: u32,
+}
+
+@group(0) @binding(0) var<uniform> c: Constants;
+@group(0) @binding(1) var<uniform> pace: Step;
+@group(0) @binding(2) var<uniform> batch: Batch;
+@group(0) @binding(3) var<storage, read_write> state: array<f32>;
+@group(0) @binding(4) var<storage, read_write> pose: array<f32>;
+@group(0) @binding(5) var<storage, read> statics: array<f32>;
+@group(0) @binding(6) var<storage, read> constraints: array<u32>;
+@group(0) @binding(7) var<storage, read_write> lambda: array<f32>;
+@group(0) @binding(8) var<storage, read> frame: array<f32>;
+
+fn stateAt(section: u32, i: u32) -> u32 {
+  return section * 3u * c.count + i * 3u;
+}
+
+fn load(at: u32) -> vec3<f32> {
+  return vec3<f32>(state[at], state[at + 1u], state[at + 2u]);
+}
+
+fn store(at: u32, v: vec3<f32>) {
+  state[at] = v.x;
+  state[at + 1u] = v.y;
+  state[at + 2u] = v.z;
+}
+
+fn poseOf(at: u32) -> vec3<f32> {
+  return vec3<f32>(pose[at], pose[at + 1u], pose[at + 2u]);
+}
+
+/* Sections of state. */
+const POSITION = 0u;
+const PREVIOUS = 1u;
+const VELOCITY = 2u;
+const START = 3u;
+
+fn inverseMass(i: u32) -> f32 {
+  return statics[i];
+}
+
+/* Where skinning puts particle i, fraction of the way from the last pose to the latest. */
+fn skinnedAt(i: u32) -> vec3<f32> {
+  let a = poseOf(pace.fromAt + i * 3u);
+  let b = poseOf(pace.toAt + i * 3u);
+  return a + (b - a) * pace.fraction;
+}
+
+fn skinMatrix(j: u32) -> mat4x4<f32> {
+  let at = j * 16u;
+  return mat4x4<f32>(
+    vec4<f32>(frame[at], frame[at + 1u], frame[at + 2u], frame[at + 3u]),
+    vec4<f32>(frame[at + 4u], frame[at + 5u], frame[at + 6u], frame[at + 7u]),
+    vec4<f32>(frame[at + 8u], frame[at + 9u], frame[at + 10u], frame[at + 11u]),
+    vec4<f32>(frame[at + 12u], frame[at + 13u], frame[at + 14u], frame[at + 15u]));
+}
+
+/* skinParticles: each particle's skinned position and normal by up to eight influences. */
+@compute @workgroup_size(64)
+fn clothPose(@builtin(global_invocation_id) id: vec3<u32>) {
+  let i = id.x;
+  if (i >= c.count) {
+    return;
+  }
+  let n = c.count;
+  let restAt = 7u * n + i * 3u;
+  let normalAt = 10u * n + i * 3u;
+  let x = vec4<f32>(statics[restAt], statics[restAt + 1u], statics[restAt + 2u], 1.0);
+  let q0 = vec3<f32>(statics[normalAt], statics[normalAt + 1u], statics[normalAt + 2u]);
+  var p = vec3<f32>(0.0);
+  var q = vec3<f32>(0.0);
+  for (var half = 0u; half < 2u; half++) {
+    for (var k = 0u; k < 4u; k++) {
+      let at = 13u * n + i * 16u + half * 8u;
+      let w = statics[at + 4u + k];
+      if (w == 0.0) {
+        continue;
+      }
+      let m = skinMatrix(u32(statics[at + k]));
+      p += w * (m * x).xyz;
+      q += w * (mat3x3<f32>(m[0].xyz, m[1].xyz, m[2].xyz) * q0);
+    }
+  }
+  let to = pace.toAt + i * 3u;
+  pose[to] = p.x;
+  pose[to + 1u] = p.y;
+  pose[to + 2u] = p.z;
+  let len = sqrt(dot(q, q));
+  let inv = select(0.0, 1.0 / len, len > 1e-12);
+  let normal = 6u * n + i * 3u;
+  pose[normal] = q.x * inv;
+  pose[normal + 1u] = q.y * inv;
+  pose[normal + 2u] = q.z * inv;
+}
+
+/* SkinnedCloth.rest: every particle at the latest skinned position, at rest. */
+@compute @workgroup_size(64)
+fn clothRest(@builtin(global_invocation_id) id: vec3<u32>) {
+  let i = id.x;
+  if (i >= c.count) {
+    return;
+  }
+  let p = poseOf(pace.toAt + i * 3u);
+  store(stateAt(POSITION, i), p);
+  store(stateAt(START, i), p);
+  store(stateAt(VELOCITY, i), vec3<f32>(0.0));
+}
+
+/* The step's start: the step start kept, a free particle carried, a kinematic one placed. */
+@compute @workgroup_size(64)
+fn clothBegin(@builtin(global_invocation_id) id: vec3<u32>) {
+  let i = id.x;
+  if (i >= c.count) {
+    return;
+  }
+  let at = stateAt(POSITION, i);
+  let p = load(at);
+  store(stateAt(START, i), p);
+  if (inverseMass(i) == 0.0) {
+    store(at, skinnedAt(i));
+    return;
+  }
+  if (pace.carry == 0u) {
+    return;
+  }
+  /* carryParticles: the share of the character's turn about its old origin, and of its move. */
+  let turn = mat3x3<f32>(pace.turn0.xyz, pace.turn1.xyz, pace.turn2.xyz);
+  let share = pace.origin.w;
+  let x = p - pace.origin.xyz;
+  store(at, p + (turn * x - x) * share + pace.shift.xyz);
+  let vAt = stateAt(VELOCITY, i);
+  let v = load(vAt);
+  store(vAt, v + (turn * v - v) * share);
+}
+
+/* predictParticles. */
+@compute @workgroup_size(64)
+fn clothPredict(@builtin(global_invocation_id) id: vec3<u32>) {
+  let i = id.x;
+  if (i >= c.count) {
+    return;
+  }
+  let at = stateAt(POSITION, i);
+  let p = load(at);
+  store(stateAt(PREVIOUS, i), p);
+  if (inverseMass(i) == 0.0) {
+    return;
+  }
+  let vAt = stateAt(VELOCITY, i);
+  let air = pace.wind.xyz;
+  let v = air + (load(vAt) * c.keep - air) * c.dragKeep + c.gravity.xyz * c.h;
+  store(vAt, v);
+  store(at, p + v * c.h);
+}
+
+/* solveDistanceBatch: one invocation a constraint of the batch, constraints sorted by batch. */
+@compute @workgroup_size(64)
+fn clothDistance(@builtin(global_invocation_id) id: vec3<u32>) {
+  let k = batch.begin + id.x;
+  if (k >= batch.end) {
+    return;
+  }
+  /* The first pass of a substep starts the multiplier at nothing, as the CPU's fill(0) does —
+     written before any early return, so a skipped constraint does not carry the last one. */
+  if (batch.first == 1u) {
+    lambda[k] = 0.0;
+  }
+  let a = constraints[k * 4u];
+  let b = constraints[k * 4u + 1u];
+  let wa = inverseMass(a);
+  let wb = inverseMass(b);
+  let w = wa + wb;
+  if (w == 0.0) {
+    return;
+  }
+  let aAt = stateAt(POSITION, a);
+  let bAt = stateAt(POSITION, b);
+  let d = load(bAt) - load(aAt);
+  let len = sqrt(dot(d, d));
+  if (len < 1e-9) {
+    return;
+  }
+  let alpha = bitcast<f32>(constraints[k * 4u + 3u]) * c.invH2;
+  let old = lambda[k];
+  let delta = (-(len - bitcast<f32>(constraints[k * 4u + 2u])) - alpha * old) / (w + alpha);
+  lambda[k] = old + delta;
+  let s = delta / len;
+  /* A pinned particle is shared across a batch, so it is never written, even with itself. */
+  if (wa > 0.0) {
+    store(aAt, load(aAt) - d * s * wa);
+  }
+  if (wb > 0.0) {
+    store(bAt, load(bAt) + d * s * wb);
+  }
+}
+
+/* solveBendingBatch, with the angle from atan2 (see the header). */
+@compute @workgroup_size(64)
+fn clothBending(@builtin(global_invocation_id) id: vec3<u32>) {
+  let k = batch.begin + id.x;
+  if (k >= batch.end) {
+    return;
+  }
+  let at = c.distances * 4u + k * 6u;
+  let slot = c.distances + k;
+  if (batch.first == 1u) {
+    lambda[slot] = 0.0;
+  }
+  let ids = vec4<u32>(constraints[at], constraints[at + 1u], constraints[at + 2u], constraints[at + 3u]);
+  let w = vec4<f32>(inverseMass(ids.x), inverseMass(ids.y), inverseMass(ids.z), inverseMass(ids.w));
+  if (w.x + w.y + w.z + w.w == 0.0) {
+    return;
+  }
+  let origin = load(stateAt(POSITION, ids.x));
+  let e = load(stateAt(POSITION, ids.y)) - origin;
+  let p3 = load(stateAt(POSITION, ids.z)) - origin;
+  let p4 = load(stateAt(POSITION, ids.w)) - origin;
+  var n1 = cross(e, p3);
+  var n2 = cross(e, p4);
+  let l1 = sqrt(dot(n1, n1));
+  let l2 = sqrt(dot(n2, n2));
+  if (l1 < 1e-12 || l2 < 1e-12) {
+    return;
+  }
+  n1 = n1 / l1;
+  n2 = n2 / l2;
+  let d = clamp(dot(n1, n2), -1.0, 1.0);
+  let sine = sqrt(dot(cross(n1, n2), cross(n1, n2)));
+  if (sine < 1e-6) {
+    return;
+  }
+  /* The q vectors: the negative of the derivative of n1 · n2. See solveBendingBatch. */
+  let q3 = (cross(e, n2) + cross(n1, e) * d) / l1;
+  let q4 = (cross(e, n1) + cross(n2, e) * d) / l2;
+  let q2 = -((cross(p3, n2) + cross(n1, p3) * d) / l1 + (cross(p4, n1) + cross(n2, p4) * d) / l2);
+  let q1 = -(q2 + q3 + q4);
+  var sum = w.x * dot(q1, q1) + w.y * dot(q2, q2) + w.z * dot(q3, q3) + w.w * dot(q4, q4);
+  sum = sum / (sine * sine);
+  if (sum < 1e-12) {
+    return;
+  }
+  let alpha = bitcast<f32>(constraints[at + 5u]) * c.invH2;
+  let old = lambda[slot];
+  let angle = atan2(sine, d);
+  let delta = (-(angle - bitcast<f32>(constraints[at + 4u])) - alpha * old) / (sum + alpha);
+  lambda[slot] = old + delta;
+  let scale = delta / sine;
+  if (w.x > 0.0) {
+    let p = stateAt(POSITION, ids.x);
+    store(p, load(p) + w.x * scale * q1);
+  }
+  if (w.y > 0.0) {
+    let p = stateAt(POSITION, ids.y);
+    store(p, load(p) + w.y * scale * q2);
+  }
+  if (w.z > 0.0) {
+    let p = stateAt(POSITION, ids.z);
+    store(p, load(p) + w.z * scale * q3);
+  }
+  if (w.w > 0.0) {
+    let p = stateAt(POSITION, ids.w);
+    store(p, load(p) + w.w * scale * q4);
+  }
+}
+
+/* solveTethers, one invocation a tether of the batch: a batch holds a particle once. */
+@compute @workgroup_size(64)
+fn clothTether(@builtin(global_invocation_id) id: vec3<u32>) {
+  let k = batch.begin + id.x;
+  if (k >= batch.end) {
+    return;
+  }
+  let at = c.distances * 4u + c.bendings * 6u + k * 3u;
+  let pAt = stateAt(POSITION, constraints[at]);
+  let anchor = load(stateAt(POSITION, constraints[at + 1u]));
+  let d = load(pAt) - anchor;
+  let dist = sqrt(dot(d, d));
+  let reach = bitcast<f32>(constraints[at + 2u]);
+  if (dist <= reach || dist < 1e-9) {
+    return;
+  }
+  store(pAt, anchor + d * (reach / dist));
+}
+
+/* pushOut: the point onto the sphere's surface if it is inside. A point at the centre stays. */
+fn pushOut(p: vec3<f32>, centre: vec3<f32>, radius: f32) -> vec3<f32> {
+  let d = p - centre;
+  let d2 = dot(d, d);
+  if (d2 >= radius * radius || d2 < 1e-18) {
+    return p;
+  }
+  return centre + d * (radius / sqrt(d2));
+}
+
+/* applyMaxDistance, applyStops (back, then front) and applyColliders, in that order. */
+@compute @workgroup_size(64)
+fn clothLimit(@builtin(global_invocation_id) id: vec3<u32>) {
+  let i = id.x;
+  if (i >= c.count || inverseMass(i) == 0.0) {
+    return;
+  }
+  let n = c.count;
+  let at = stateAt(POSITION, i);
+  var p = load(at);
+  let t = skinnedAt(i);
+  /* A max distance below zero is none: the CPU's Infinity, which a device may not be handed. */
+  let maxDistance = statics[n + i];
+  if ((c.limits & 1u) != 0u && maxDistance >= 0.0) {
+    let limit = maxDistance * c.maxDistanceScale;
+    let d = p - t;
+    let d2 = dot(d, d);
+    if (d2 > limit * limit) {
+      p = t + d * (limit / sqrt(d2));
+    }
+  }
+  let normal = poseOf(6u * n + i * 3u);
+  if ((c.limits & 2u) != 0u) {
+    let radius = statics[2u * n + i * 2u + 1u];
+    if (radius > 0.0) {
+      p = pushOut(p, t - normal * (statics[2u * n + i * 2u] + radius), radius);
+    }
+  }
+  if ((c.limits & 4u) != 0u) {
+    let radius = statics[4u * n + i * 2u + 1u];
+    if (radius > 0.0) {
+      p = pushOut(p, t + normal * (statics[4u * n + i * 2u] + radius), radius);
+    }
+  }
+  let pad = statics[6u * n + i] + c.margin;
+  let ends = c.joints * 16u;
+  let radii = ends + c.colliders * 6u;
+  for (var k = 0u; k < c.colliders; k++) {
+    let a = vec3<f32>(frame[ends + k * 6u], frame[ends + k * 6u + 1u], frame[ends + k * 6u + 2u]);
+    let e = vec3<f32>(frame[ends + k * 6u + 3u], frame[ends + k * 6u + 4u], frame[ends + k * 6u + 5u]) - a;
+    let e2 = dot(e, e);
+    var s = 0.0;
+    if (e2 > 1e-18) {
+      s = clamp(dot(p - a, e) / e2, 0.0, 1.0);
+    }
+    let r0 = frame[radii + k * 2u];
+    let r1 = frame[radii + k * 2u + 1u];
+    p = pushOut(p, a + e * s, r0 + (r1 - r0) * s + pad);
+  }
+  store(at, p);
+}
+
+/* finishParticles. */
+@compute @workgroup_size(64)
+fn clothFinish(@builtin(global_invocation_id) id: vec3<u32>) {
+  let i = id.x;
+  if (i >= c.count) {
+    return;
+  }
+  let vAt = stateAt(VELOCITY, i);
+  if (inverseMass(i) == 0.0) {
+    store(vAt, vec3<f32>(0.0));
+    return;
+  }
+  store(vAt, (load(stateAt(POSITION, i)) - load(stateAt(PREVIOUS, i))) * c.invH);
+}
+
+/* SkinnedCloth.blend: after a reset, the simulation's share grows over the blend's steps. */
+@compute @workgroup_size(64)
+fn clothBlend(@builtin(global_invocation_id) id: vec3<u32>) {
+  let i = id.x;
+  if (i >= c.count) {
+    return;
+  }
+  let t = skinnedAt(i);
+  let at = stateAt(POSITION, i);
+  store(at, t + (load(at) - t) * pace.blend);
+  let vAt = stateAt(VELOCITY, i);
+  store(vAt, load(vAt) * pace.blend);
+}
+
+@group(1) @binding(0) var output: texture_storage_2d<rgba32float, write>;
+
+/* SkinnedCloth.interpolate, into the particle texture the vertex stage reads. */
+@compute @workgroup_size(64)
+fn clothPublish(@builtin(global_invocation_id) id: vec3<u32>) {
+  let i = id.x;
+  if (i >= c.count) {
+    return;
+  }
+  let a = load(stateAt(START, i));
+  let p = a + (load(stateAt(POSITION, i)) - a) * pace.alpha;
+  textureStore(output, vec2<u32>(i % c.width, i / c.width), vec4<f32>(p, 0.0));
+}
+`;

package/dist/render/backend/webgpu/skinScatterPass.d.ts

@@ -0,0 +1,51 @@
+import type { DrawCommand } from './drawCommand.ts';
+import type { PipelineCache } from './pipelineCache.ts';
+/** The format the diffuse lands in: half floats, because it is linear light before any grade. */
+export declare const SKIN_TARGET_FORMAT: GPUTextureFormat;
+export declare class SkinScatterPass {
+    private readonly device;
+    private readonly layout;
+    private readonly uniforms;
+    private readonly staging;
+    private readonly linear;
+    private readonly nearest;
+    private readonly commands;
+    /** Each profile's scatter distance per channel, in metres: four floats a profile. */
+    private readonly profiles;
+    private diffuse;
+    private diffuseMsaa;
+    private across;
+    private diffuseView;
+    private diffuseMsaaView;
+    private acrossView;
+    private groupAcross;
+    private groupDown;
+    private groupsDepth;
+    constructor(device: GPUDevice);
+    /** Whether any skin was kept for this frame. */
+    get pending(): boolean;
+    /** A command for one skin draw's diffuse half, filled by the caller. */
+    take(): DrawCommand;
+    /** What one profile scatters, per channel, in metres: the last material to name it says. */
+    setProfile(index: number, r: number, g: number, b: number): void;
+    /** The targets at the frame's size and sample count, made again only when either changes. */
+    size(width: number, height: number, samples: number): void;
+    /**
+     * Every kept draw's diffuse into the target, against the frame's own depth, which it reads and
+     * does not write. Called with no pass open on `encoder`.
+     */
+    drawDiffuse(encoder: GPUCommandEncoder, depth: GPUTextureView, issue: (pass: GPURenderPassEncoder, command: DrawCommand) => void, timestampWrites: GPURenderPassTimestampWrites | undefined): void;
+    /**
+     * Across into the second target, then down and added into the frame's colour — `frame` the
+     * attachment its pass draws into, `resolve` where a multisampled one resolves, `samples` its
+     * count. `projection` is the frame's, `height` the frame's height in pixels.
+     */
+    spread(encoder: GPUCommandEncoder, pipelines: PipelineCache, depth: GPUTextureView, frame: GPUTextureView, resolve: GPUTextureView | null, samples: number, projection: ArrayLike<number>, inverseProjection: ArrayLike<number>, timestampWrites: () => GPURenderPassTimestampWrites | undefined): void;
+    /** Spent at the frame's start: the kept draws were this frame's. */
+    reset(): void;
+    dispose(): void;
+    private group;
+    /** Both axes' blocks: the step along each, and what they share. Column-major matrices. */
+    private writeUniforms;
+    private releaseTargets;
+}