reze-engine 0.62.1 → 0.64.0

package/src/unity/looks.ts

@@ -0,0 +1,372 @@
+// Native looks: a model dressed in the game's own materials.
+//
+// A look names translated shader variants (host.ts), the shared images they
+// sample (ramps, matcaps), and per material the passes it draws and the values
+// it carries — what a game character's material held, with the model's own
+// texture standing in for its _MainTex. Optionally a CharacterRig, which makes
+// the model a character in the game's sense (character.ts).
+//
+// The frame, as the game's CharacterFeature lays it out after the opaques:
+//   1. every native opaque material's passes (Outline, ForwardBase), in queue
+//      order — after the engine's own opaque bundle, so both share one depth;
+//   2. CharHairShadow, CharFaceShadow, CharHairTransE, each over every
+//      material that has one (the hair's shadow on the face, by stencil);
+//   3. the eyes' see-through-hair passes, Override1 for all, then 2, then 3;
+// and the transparent queue in the engine's transparent phase. Shadows are the
+// engine's: a native material still casts through its pass.
+
+import { NativeHost, type NativeShader, type NativeStream, type NativeValue, type ValueSource } from "./host"
+import type { PassState } from "./state"
+import { UNITY_SKIN_WGSL, tangentsFor } from "./skin"
+import { characterBlock, type CharacterRig } from "./character"
+import { engineToGame, gameToEngine } from "./globals"
+
+export type NativeTexture = {
+  width: number
+  height: number
+  /** RGBA8, rows top first. */
+  data: Uint8Array
+  /** Colour (sRGB) or data (linear) — how the shader expects to read it. */
+  srgb: boolean
+}
+
+export type NativeMaterialSpec = {
+  /** The model's materials this dresses, by name. */
+  materials: string[]
+  /** Unity render queue: above 2500 draws in the transparent phase. */
+  queue: number
+  passes: { lightMode: string; shader: string; state: PassState }[]
+  /** Material properties, by name (colours linear, as Unity uploads them). */
+  values: Record<string, NativeValue>
+  /** Texture slots: a key into the look's images, or "@diffuse" for the
+   *  material's own texture. */
+  textures: Record<string, string>
+}
+
+export type NativeLook = {
+  shaders: NativeShader[]
+  textures: Record<string, NativeTexture>
+  materials: NativeMaterialSpec[]
+  rig?: CharacterRig
+}
+
+/** What the engine lends for one model. */
+export type NativeModel = {
+  name: string
+  vertices: Float32Array
+  indices: Uint32Array
+  vertexBuffer: GPUBuffer
+  jointsBuffer: GPUBuffer
+  weightsBuffer: GPUBuffer
+  skinMatrixBuffer: GPUBuffer
+  indexBuffer: GPUBuffer
+  draws: {
+    materialName: string
+    firstIndex: number
+    count: number
+    diffuse: GPUTextureView
+  }[]
+  hidden: (material: string) => boolean
+  /** The head bone's current skinning matrix, or null. */
+  headSkin: () => ArrayLike<number> | null
+  headRest: [number, number, number]
+  /** Its rendering-layer bits, for a look without a rig. */
+  layers: number
+}
+
+type Install = {
+  model: NativeModel
+  look: NativeLook
+  views: Map<string, GPUTextureView>
+  textures: GPUTexture[]
+  pos: GPUBuffer
+  nrm: GPUBuffer
+  tan: GPUBuffer
+  restTan: GPUBuffer
+  params: GPUBuffer
+  skinBind: GPUBindGroup
+  /** Per draw of the model: the spec that dresses it, if any. */
+  dressed: { draw: NativeModel["draws"][number]; spec: NativeMaterialSpec }[]
+  block: Record<string, NativeValue>
+}
+
+const OPAQUE_MODES = ["ALWAYS", "FORWARDBASE"]
+const CHARACTER_MODES = ["CHARHAIRSHADOW", "CHARFACESHADOW", "CHARHAIRTRANSE"]
+const OVERRIDE_MODES = ["OVERRIDE1", "OVERRIDE2", "OVERRIDE3"]
+
+export class NativeLooks {
+  readonly host: NativeHost
+  private installs = new Map<string, Install>()
+  private skinPipeline: GPUComputePipeline
+  private zero: GPUBuffer
+  private white: GPUBuffer
+  private globals: Record<string, NativeValue> = {}
+  private frameTextures: Record<string, GPUTextureView | undefined> = {}
+  private scale = 8
+
+  constructor(
+    private device: GPUDevice,
+    host: NativeHost,
+  ) {
+    this.host = host
+    this.skinPipeline = device.createComputePipeline({
+      label: "native skinning",
+      layout: "auto",
+      compute: {
+        module: device.createShaderModule({
+          label: "native skinning",
+          code: UNITY_SKIN_WGSL,
+        }),
+        entryPoint: "main",
+      },
+    })
+    this.zero = device.createBuffer({
+      label: "native zero stream",
+      size: 64,
+      usage: GPUBufferUsage.VERTEX,
+    })
+    this.white = device.createBuffer({
+      label: "native white stream",
+      size: 64,
+      usage: GPUBufferUsage.VERTEX,
+      mappedAtCreation: true,
+    })
+    new Float32Array(this.white.getMappedRange()).fill(1)
+    this.white.unmap()
+  }
+
+  get size(): number {
+    return this.installs.size
+  }
+
+  has(model: string): boolean {
+    return this.installs.has(model)
+  }
+
+  /** Is this model's material drawn by its native look? */
+  dresses(model: string, material: string): boolean {
+    const i = this.installs.get(model)
+    return !!i && i.dressed.some((d) => d.draw.materialName === material)
+  }
+
+  install(model: NativeModel, look: NativeLook): void {
+    this.remove(model.name)
+    const d = this.device
+    for (const s of look.shaders) this.host.register(s)
+    const views = new Map<string, GPUTextureView>()
+    const textures: GPUTexture[] = []
+    for (const [key, t] of Object.entries(look.textures)) {
+      const tex = d.createTexture({
+        label: `${model.name}: ${key}`,
+        size: [t.width, t.height],
+        format: t.srgb ? "rgba8unorm-srgb" : "rgba8unorm",
+        usage: GPUTextureUsage.TEXTURE_BINDING | GPUTextureUsage.COPY_DST,
+      })
+      d.queue.writeTexture({ texture: tex }, t.data as Uint8Array<ArrayBuffer>, { bytesPerRow: t.width * 4 }, [t.width, t.height])
+      textures.push(tex)
+      views.set(key, tex.createView())
+    }
+    const n = model.vertices.length / 8
+    const buf = (label: string, floats: number, usage: number) =>
+      d.createBuffer({
+        label: `${model.name}: ${label}`,
+        size: Math.max(16, floats * 4),
+        usage,
+      })
+    const out = GPUBufferUsage.STORAGE | GPUBufferUsage.VERTEX
+    const pos = buf("native position", n * 3, out)
+    const nrm = buf("native normal", n * 3, out)
+    const tan = buf("native tangent", n * 4, out)
+    const restTan = buf("rest tangent", n * 4, GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_DST)
+    d.queue.writeBuffer(restTan, 0, tangentsFor(model.vertices, model.indices) as Float32Array<ArrayBuffer>)
+    const params = d.createBuffer({
+      label: `${model.name}: skin params`,
+      size: 16,
+      usage: GPUBufferUsage.UNIFORM | GPUBufferUsage.COPY_DST,
+    })
+    d.queue.writeBuffer(params, 0, new Uint32Array([n, 0, 0, 0]))
+    const skinBind = d.createBindGroup({
+      label: `${model.name}: native skinning`,
+      layout: this.skinPipeline.getBindGroupLayout(0),
+      entries: [
+        { binding: 0, resource: { buffer: params } },
+        { binding: 1, resource: { buffer: model.vertexBuffer } },
+        { binding: 2, resource: { buffer: model.jointsBuffer } },
+        { binding: 3, resource: { buffer: model.weightsBuffer } },
+        { binding: 4, resource: { buffer: model.skinMatrixBuffer } },
+        { binding: 5, resource: { buffer: restTan } },
+        { binding: 6, resource: { buffer: pos } },
+        { binding: 7, resource: { buffer: nrm } },
+        { binding: 8, resource: { buffer: tan } },
+      ],
+    })
+    const dressed: Install["dressed"] = []
+    for (const draw of model.draws) {
+      const spec = look.materials.find((m) => m.materials.includes(draw.materialName))
+      if (spec) dressed.push({ draw, spec })
+    }
+    this.installs.set(model.name, {
+      model,
+      look,
+      views,
+      textures,
+      pos,
+      nrm,
+      tan,
+      restTan,
+      params,
+      skinBind,
+      dressed,
+      block: {},
+    })
+  }
+
+  remove(name: string): void {
+    const i = this.installs.get(name)
+    if (!i) return
+    this.installs.delete(name)
+    this.host.release(`${name}|`)
+    const bufs = [i.pos, i.nrm, i.tan, i.restTan, i.params]
+    const texs = i.textures
+    // Freed once the GPU is done with the frames that may still name them.
+    void this.device.queue.onSubmittedWorkDone().then(() => {
+      for (const b of bufs) b.destroy()
+      for (const t of texs) t.destroy()
+    })
+  }
+
+  /**
+   * Before the scene pass: skin every dressed model and compute each
+   * character's block; the frame's Unity globals come computed (engine.ts).
+   */
+  prepare(
+    encoder: GPUCommandEncoder,
+    globals: Record<string, NativeValue>,
+    scale: number,
+    textures: Record<string, GPUTextureView | undefined>,
+  ): void {
+    if (!this.installs.size) return
+    this.scale = scale
+    this.globals = globals
+    this.frameTextures = textures
+    const pass = encoder.beginComputePass({ label: "native skinning" })
+    pass.setPipeline(this.skinPipeline)
+    for (const i of this.installs.values()) {
+      pass.setBindGroup(0, i.skinBind)
+      pass.dispatchWorkgroups(Math.ceil(i.model.vertices.length / 8 / 64))
+      i.block = i.look.rig
+        ? characterBlock(i.look.rig, i.model.headSkin(), i.model.headRest, scale)
+        : {
+            unity_RenderingLayer: new Uint32Array([i.model.layers >>> 0, 0, 0, 0]),
+          }
+    }
+    pass.end()
+  }
+
+  private streams(i: Install, shader: string): NativeStream[] {
+    return this.host.inputs(shader).map(({ semantic }) => {
+      if (semantic === "POSITION")
+        return {
+          buffer: i.pos,
+          offset: 0,
+          stride: 12,
+          format: "float32x3" as const,
+        }
+      if (semantic === "NORMAL")
+        return {
+          buffer: i.nrm,
+          offset: 0,
+          stride: 12,
+          format: "float32x3" as const,
+        }
+      if (semantic === "TANGENT")
+        return {
+          buffer: i.tan,
+          offset: 0,
+          stride: 16,
+          format: "float32x4" as const,
+        }
+      if (semantic === "TEXCOORD0")
+        return {
+          buffer: i.model.vertexBuffer,
+          offset: 24,
+          stride: 32,
+          format: "float32x2" as const,
+        }
+      // A missing vertex colour reads white, as Unity binds it; any other
+      // missing stream reads zero.
+      if (semantic === "COLOR")
+        return {
+          buffer: this.white,
+          offset: 0,
+          stride: 0,
+          format: "float32x4" as const,
+        }
+      return {
+        buffer: this.zero,
+        offset: 0,
+        stride: 0,
+        format: "float32x4" as const,
+      }
+    })
+  }
+
+  private drawPass(pass: GPURenderPassEncoder, i: Install, d: Install["dressed"][number], p: NativeMaterialSpec["passes"][number]): void {
+    if (i.model.hidden(d.draw.materialName)) return
+    const s = this.scale
+    const perDraw: ValueSource = {
+      // The skinned streams are in engine units; the game's world is 1/scale of them.
+      unity_ObjectToWorld: engineToGame(s),
+      unity_WorldToObject: gameToEngine(s),
+      unity_WorldTransformParams: [0, 0, 0, 1],
+      unity_LightData: [0, 0, 0, 0],
+      unity_LightIndices: [
+        [0, 0, 0, 0],
+        [0, 0, 0, 0],
+      ],
+      unity_LODFade: [1, 1, 0, 0],
+    }
+    const textures: Record<string, GPUTextureView | undefined> = {
+      ...this.frameTextures,
+    }
+    for (const [slot, key] of Object.entries(d.spec.textures)) textures[slot] = key === "@diffuse" ? d.draw.diffuse : i.views.get(key)
+    this.host.draw(pass, {
+      shader: p.shader,
+      state: p.state,
+      streams: this.streams(i, p.shader),
+      index: {
+        buffer: i.model.indexBuffer,
+        format: "uint32",
+        offset: d.draw.firstIndex * 4,
+        count: d.draw.count,
+      },
+      sources: [perDraw, i.block, d.spec.values, this.globals],
+      textures,
+      key: `${i.model.name}|${d.draw.materialName}|${p.lightMode}|${p.shader}`,
+      pass: "main",
+    })
+  }
+
+  /** Every dressed material whose queue is in range, in queue order. */
+  private each(opaque: boolean): { i: Install; d: Install["dressed"][number] }[] {
+    const out: { i: Install; d: Install["dressed"][number] }[] = []
+    for (const i of this.installs.values()) for (const d of i.dressed) if (d.spec.queue <= 2500 === opaque) out.push({ i, d })
+    return out.sort((a, b) => a.d.spec.queue - b.d.spec.queue)
+  }
+
+  /** After the engine's opaque phase: the opaques, then the character passes. */
+  drawOpaque(pass: GPURenderPassEncoder): void {
+    if (!this.installs.size) return
+    const list = this.each(true)
+    for (const { i, d } of list) for (const p of d.spec.passes) if (OPAQUE_MODES.includes(p.lightMode)) this.drawPass(pass, i, d, p)
+    for (const mode of [...CHARACTER_MODES, ...OVERRIDE_MODES])
+      for (const { i, d } of list) for (const p of d.spec.passes) if (p.lightMode === mode) this.drawPass(pass, i, d, p)
+  }
+
+  /** In the engine's transparent phase. */
+  drawTransparent(pass: GPURenderPassEncoder): void {
+    if (!this.installs.size) return
+    for (const { i, d } of this.each(false))
+      for (const p of d.spec.passes) if (OPAQUE_MODES.includes(p.lightMode)) this.drawPass(pass, i, d, p)
+  }
+}

package/src/unity/skin.ts

@@ -0,0 +1,128 @@
+// Skinned vertex streams for native (game-shader) materials.
+//
+// The engine skins in the vertex stage of its own material shaders, from the
+// bone palette (skinMats) and the post-morph vertex buffer. A game shader skins
+// nothing: Unity hands it world-space-ready POSITION / NORMAL / TANGENT streams
+// from its SkinnedMeshRenderer. So a model that wears native materials is
+// skinned once per frame here, by compute, into three plain vertex buffers the
+// game shaders read as Unity's — the same streams the compare page feeds them.
+//
+// Inputs are the engine's own buffers, untouched: the interleaved vertex
+// buffer (position, normal, uv — 8 floats, morphs already applied), the joints
+// (uint16x4, two words a vertex), the weights (unorm8x4, one word) and the
+// skin matrices. Tangents are not in the engine's vertex format; they are
+// derived once at install (tangentsFor) and skinned here with the normal.
+
+export const UNITY_SKIN_WGSL = /* wgsl */ `
+struct Params { count: u32, _a: u32, _b: u32, _c: u32 }
+@group(0) @binding(0) var<uniform> P: Params;
+@group(0) @binding(1) var<storage, read> verts: array<f32>;      // pos(3) nrm(3) uv(2)
+@group(0) @binding(2) var<storage, read> joints: array<u32>;     // uint16x4: two words
+@group(0) @binding(3) var<storage, read> weights: array<u32>;    // unorm8x4: one word
+@group(0) @binding(4) var<storage, read> skinMats: array<mat4x4f>;
+@group(0) @binding(5) var<storage, read> tangents: array<f32>;   // rest tangent(4)
+@group(0) @binding(6) var<storage, read_write> outPos: array<f32>;
+@group(0) @binding(7) var<storage, read_write> outNrm: array<f32>;
+@group(0) @binding(8) var<storage, read_write> outTan: array<f32>;
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) id: vec3u) {
+  let v = id.x;
+  if (v >= P.count) { return; }
+  let p = vec4f(verts[v * 8u], verts[v * 8u + 1u], verts[v * 8u + 2u], 1.0);
+  let n = vec3f(verts[v * 8u + 3u], verts[v * 8u + 4u], verts[v * 8u + 5u]);
+  let t = vec4f(tangents[v * 4u], tangents[v * 4u + 1u], tangents[v * 4u + 2u], tangents[v * 4u + 3u]);
+  let j01 = joints[v * 2u];
+  let j23 = joints[v * 2u + 1u];
+  let j = vec4u(j01 & 0xffffu, j01 >> 16u, j23 & 0xffffu, j23 >> 16u);
+  var w = unpack4x8unorm(weights[v]);
+  // Renormalised as the material vertex stage does: PMX ships unnormalised
+  // weights on extras, and a sum of zero is the bind pose.
+  let s = w.x + w.y + w.z + w.w;
+  w = select(vec4f(1.0, 0.0, 0.0, 0.0), w / s, s > 0.0001);
+  var sp = vec3f(0.0);
+  var sn = vec3f(0.0);
+  var st = vec3f(0.0);
+  for (var k = 0u; k < 4u; k++) {
+    let m = skinMats[j[k]];
+    let r = mat3x3f(m[0].xyz, m[1].xyz, m[2].xyz);
+    sp += (m * p).xyz * w[k];
+    sn += (r * n) * w[k];
+    st += (r * t.xyz) * w[k];
+  }
+  // Degenerate blends (opposing bones at 50/50) cancel to zero; up, not NaN.
+  let ln = dot(sn, sn);
+  let lt = dot(st, st);
+  sn = select(vec3f(0.0, 1.0, 0.0), sn * inverseSqrt(ln), ln > 1e-12);
+  st = select(vec3f(1.0, 0.0, 0.0), st * inverseSqrt(lt), lt > 1e-12);
+  outPos[v * 3u] = sp.x; outPos[v * 3u + 1u] = sp.y; outPos[v * 3u + 2u] = sp.z;
+  outNrm[v * 3u] = sn.x; outNrm[v * 3u + 1u] = sn.y; outNrm[v * 3u + 2u] = sn.z;
+  outTan[v * 4u] = st.x; outTan[v * 4u + 1u] = st.y; outTan[v * 4u + 2u] = st.z; outTan[v * 4u + 3u] = t.w;
+}
+`
+
+/**
+ * Per-vertex tangents (xyz, w = handedness) from positions, normals and uvs,
+ * accumulated per triangle and orthogonalised against the normal — the
+ * standard construction, which is what a mesh imported without tangents gets
+ * in Unity too. `verts` is the engine's interleaved layout (8 floats).
+ */
+export function tangentsFor(verts: Float32Array, indices: Uint32Array | Uint16Array): Float32Array {
+  const n = verts.length / 8
+  const tan = new Float64Array(n * 3)
+  const bit = new Float64Array(n * 3)
+  for (let i = 0; i + 2 < indices.length; i += 3) {
+    const a = indices[i],
+      b = indices[i + 1],
+      c = indices[i + 2]
+    const p = (v: number, k: number) => verts[v * 8 + k]
+    const e1 = [p(b, 0) - p(a, 0), p(b, 1) - p(a, 1), p(b, 2) - p(a, 2)]
+    const e2 = [p(c, 0) - p(a, 0), p(c, 1) - p(a, 1), p(c, 2) - p(a, 2)]
+    const du1 = p(b, 6) - p(a, 6),
+      dv1 = p(b, 7) - p(a, 7)
+    const du2 = p(c, 6) - p(a, 6),
+      dv2 = p(c, 7) - p(a, 7)
+    const det = du1 * dv2 - du2 * dv1
+    if (Math.abs(det) < 1e-12) continue
+    const r = 1 / det
+    for (const v of [a, b, c]) {
+      for (let k = 0; k < 3; k++) {
+        tan[v * 3 + k] += (e1[k] * dv2 - e2[k] * dv1) * r
+        bit[v * 3 + k] += (e2[k] * du1 - e1[k] * du2) * r
+      }
+    }
+  }
+  const out = new Float32Array(n * 4)
+  for (let v = 0; v < n; v++) {
+    const nx = verts[v * 8 + 3],
+      ny = verts[v * 8 + 4],
+      nz = verts[v * 8 + 5]
+    let tx = tan[v * 3],
+      ty = tan[v * 3 + 1],
+      tz = tan[v * 3 + 2]
+    const d = tx * nx + ty * ny + tz * nz
+    tx -= nx * d
+    ty -= ny * d
+    tz -= nz * d
+    let len = Math.hypot(tx, ty, tz)
+    if (len < 1e-12) {
+      // Any direction perpendicular to the normal: a surface with no uv
+      // gradient has no tangent to speak of, and a zero one is NaN downstream.
+      const ax = Math.abs(nx) < 0.9 ? 1 : 0
+      tx = ax ? 0 : -nz
+      ty = ax ? nz : 0
+      tz = ax ? -ny : nx
+      len = Math.hypot(tx, ty, tz) || 1
+    }
+    tx /= len
+    ty /= len
+    tz /= len
+    // Handedness: whether the bitangent agrees with n × t.
+    const cx = ny * tz - nz * ty,
+      cy = nz * tx - nx * tz,
+      cz = nx * ty - ny * tx
+    const w = cx * bit[v * 3] + cy * bit[v * 3 + 1] + cz * bit[v * 3 + 2] < 0 ? -1 : 1
+    out.set([tx, ty, tz, w], v * 4)
+  }
+  return out
+}