@driftengine/core 4.8.7 → 4.10.0

package/src/render/backend/webgpu/flatPass.ts

@@ -12,6 +12,7 @@ import type { LitSwitch, PipelineCache } from './pipelineCache.ts';
 import type { ClothViews } from './clothTextures.ts';
 import { shaderModule } from './shaderModules.ts';
 import { SCENE_ALPHA_COVERS } from '../../sceneCoverage.ts';
+import { REFLECTION_SURFACE_FORMAT } from './reflectionSurfacePass.ts';
 
 /** The material maps the flat pass reads: `2d-array` colour, filtered, not shadow maps. */
 const SURFACE_MAPS: ReadonlySet<string> = new Set([
@@ -98,6 +99,20 @@ export interface StageBindings {
   readonly textures: Readonly<Record<string, TextureBinding>>;
   /** Each pipeline-overridable constant's id, by the GLSL name it was declared under. */
   readonly overrides?: Readonly<Record<string, number>>;
+  /**
+   * The second block, of the uniforms marked `// wgsl:material`: its binding, size and fields.
+   * Absent from a stage that marks none, which is every stage but the lit fragment's.
+   */
+  readonly materialUniforms?: number;
+  readonly materialSize?: number;
+  readonly materialFields?: StageBindings['fields'];
+  /**
+   * The vertex stage's view block, of the uniforms marked `// wgsl:view`: the camera, the light's
+   * matrix and the wind, which a pass settles. Absent from every stage but the lit vertex one.
+   */
+  readonly viewUniforms?: number;
+  readonly viewSize?: number;
+  readonly viewFields?: StageBindings['fields'];
 }
 
 /*
@@ -129,9 +144,16 @@ const VERTEX_BINDINGS = FLAT_BINDINGS.flatVert as unknown as Readonly<
 >;
 
 /** The key for a set of vertex flags: the same rule `variantKey` uses, flags on, sorted, joined. */
-export function flatVertexKey(skinned: boolean, morphed: boolean, instanced = false): string {
+export function flatVertexKey(
+  skinned: boolean,
+  morphed: boolean,
+  instanced = false,
+  /** Whether an instanced variant plays a bone animation. See `shaders/boneAnimation.ts`. */
+  animated = false,
+): string {
   const on: string[] = [];
   /* Alphabetical, because the generator's own `variantKey` sorts and the two keys must agree. */
+  if (animated) on.push('animated');
   if (instanced) on.push('instanced');
   if (morphed) on.push('morphed');
   if (skinned) on.push('skinned');
@@ -143,8 +165,9 @@ export function flatVertexBindings(
   skinned: boolean,
   morphed = false,
   instanced = false,
+  animated = false,
 ): StageBindings {
-  const key = flatVertexKey(skinned, morphed, instanced);
+  const key = flatVertexKey(skinned, morphed, instanced, animated);
   const found = VERTEX_BINDINGS[key];
   if (found === undefined) throw new Error(`flatPass: no vertex bindings for variant "${key}"`);
   return found;
@@ -220,6 +243,8 @@ for (const [key, bindings] of Object.entries(VERTEX_BINDINGS)) {
 }
 
 export const FLAT_VERT_FIELDS = FLAT_BINDINGS.flatVert.none.fields;
+/** The animated instanced variant's block: the instanced fields, and the clip's `uBoneClip`. */
+export const FLAT_ANIMATED_VERT_FIELDS = FLAT_BINDINGS.flatVert['animated+instanced'].fields;
 /**
  * The **largest** variant's block, because one ring serves every draw.
  *
@@ -232,6 +257,55 @@ export const FLAT_VERT_SIZE = Math.max(
 );
 const FLAT_VERT_BINDING = PLAIN_VERTEX.uniforms;
 
+/*
+ * **The view block may be shorter in a variant, never different.** One slot serves every lit draw
+ * of a pass, plain and instanced alike, so every field a variant's view block declares must sit
+ * where the plain variant has it. The instanced variant declares no wind, and its block is the plain
+ * one's first two matrices.
+ */
+const PLAIN_VIEW_FIELDS = PLAIN_VERTEX.viewFields ?? {};
+for (const [key, bindings] of Object.entries(VERTEX_BINDINGS)) {
+  if (bindings.viewUniforms !== PLAIN_VERTEX.viewUniforms) {
+    throw new Error(`flatPass: vertex variant "${key}" binds its view block elsewhere`);
+  }
+  for (const [field, here] of Object.entries(bindings.viewFields ?? {})) {
+    if (PLAIN_VIEW_FIELDS[field]?.offset !== here.offset) {
+      throw new Error(
+        `flatPass: vertex variant "${key}" puts view field ${field} at ${here.offset} where the ` +
+          `plain variant has it at ${PLAIN_VIEW_FIELDS[field]?.offset}.`,
+      );
+    }
+  }
+}
+
+/** The view block's fields, binding and size: the largest variant's, as `FLAT_VERT_SIZE` is. */
+export const FLAT_VIEW_FIELDS = PLAIN_VIEW_FIELDS;
+export const FLAT_VIEW_SIZE = Math.max(
+  ...Object.values(VERTEX_BINDINGS).map((bindings) => bindings.viewSize ?? 0),
+);
+const FLAT_VIEW_BINDING = PLAIN_VERTEX.viewUniforms ?? -1;
+if (FLAT_VIEW_BINDING < 0) throw new Error('flatPass: the lit vertex stage declares no view block');
+
+/**
+ * The most sampled textures and distinct samplers any lit variant carrying the environment probe
+ * binds in its fragment stage, counted from the generated bindings: what a device must allow per
+ * stage before the probe's permutation can be built. A sampler several textures share counts once,
+ * which is what `// wgsl:share` buys.
+ */
+export function probeStageCeilings(): { readonly textures: number; readonly samplers: number } {
+  let textures = 0;
+  let samplers = 0;
+  for (const [variant, bindings] of Object.entries(FLAT_BINDINGS.flatFrag)) {
+    if (!variant.includes('environmentProbe')) continue;
+    const listed = Object.values(
+      (bindings as { textures?: Record<string, { sampler: number }> }).textures ?? {},
+    );
+    textures = Math.max(textures, listed.length);
+    samplers = Math.max(samplers, new Set(listed.map((entry) => entry.sampler)).size);
+  }
+  return { textures, samplers };
+}
+
 /** Everything the generator recorded about one permutation of the fragment stage. */
 export function flatFragmentBindings(variant: FlatVariant): StageBindings {
   const found = FRAGMENT_BINDINGS[variant];
@@ -239,6 +313,46 @@ export function flatFragmentBindings(variant: FlatVariant): StageBindings {
   return found;
 }
 
+/**
+ * The lit fragment stage's two blocks as the renderer keeps them on the CPU: the pass's, then the
+ * material's straight after it, in one scratch, with every field at its offset in that scratch.
+ *
+ * **One scratch rather than two, so no setter has to know which block its field is in.** A setter
+ * writes by name, as it always has; the draw copies the material's bytes — `passSize` onwards — into
+ * the store and the pass's bytes into a ring slot, each only when something has changed it.
+ */
+export interface FlatFragmentScratch {
+  readonly fields: StageBindings['fields'];
+  /** Bytes of the pass's block, which is where the material's starts in the scratch. */
+  readonly passSize: number;
+  readonly passBinding: number;
+  readonly materialSize: number;
+  readonly materialBinding: number;
+}
+
+export function flatFragmentScratch(variant: FlatVariant): FlatFragmentScratch {
+  const bindings = flatFragmentBindings(variant);
+  const { materialUniforms, materialSize, materialFields } = bindings;
+  if (
+    materialUniforms === undefined ||
+    materialSize === undefined ||
+    materialFields === undefined
+  ) {
+    throw new Error(`flatPass: variant "${variant}" has no material block`);
+  }
+  const fields: Record<string, StageBindings['fields'][string]> = { ...bindings.fields };
+  for (const [name, field] of Object.entries(materialFields)) {
+    fields[name] = { ...field, offset: bindings.uniformSize + field.offset };
+  }
+  return {
+    fields,
+    passSize: bindings.uniformSize,
+    passBinding: bindings.uniforms,
+    materialSize,
+    materialBinding: materialUniforms,
+  };
+}
+
 /**
  * The lit fragment stage's constants, set from the cache's switches, keyed by each one's **id**.
  *
@@ -253,6 +367,7 @@ function flatFragmentConstants(
   cache: PipelineCache,
   model: SurfaceModelKind | null,
   skinPass: SkinHalf = 'whole',
+  reflectionSurface = false,
 ): Record<string, number> {
   const overrides = flatFragmentBindings(variant).overrides;
   if (overrides?.['GLASS_SHADOWS'] === undefined) {
@@ -271,6 +386,11 @@ function flatFragmentConstants(
       constants[String(id)] = skinPass !== 'whole' && name === SKIN_HALF_SWITCH[skinPass] ? 1 : 0;
       continue;
     }
+    /* The reflection pass's surface half, set from the key too. See `reflectionSurface.ts`. */
+    if (name === 'REFLECTION_SURFACE') {
+      constants[String(id)] = reflectionSurface ? 1 : 0;
+      continue;
+    }
     if (!(name in cache.litSwitches)) {
       throw new Error(`flatPass: variant "${variant}" declares a switch "${name}" no cache sets`);
     }
@@ -279,6 +399,14 @@ function flatFragmentConstants(
   return constants;
 }
 
+/** The material block's binding, which every lit variant declares. */
+function materialBinding(fragment: StageBindings): number {
+  if (fragment.materialUniforms === undefined) {
+    throw new Error('flatPass: the lit fragment stage declares no material block');
+  }
+  return fragment.materialUniforms;
+}
+
 /** The depth format both backends compare with. */
 /* Re-exported so the twenty passes that already import it from here do not each learn a new
    module; `depthConvention.ts` is where the decision lives and why. */
@@ -290,15 +418,19 @@ const VISIBILITY_FRAGMENT = 0x2;
 /**
  * The layout this variant binds against.
  *
- * The vertex block is **dynamic**, because every draw supplies its own model matrix and tint
- * from its own slot of one buffer. The fragment block is not: it holds the lights, the fog
- * and the camera, settled once by `bindMeshPass` and unchanged between draws.
+ * All four blocks are **dynamic**: the vertex block because every draw supplies its own model
+ * matrix and tint from its own slot of one buffer; the fragment stage's pass block — the lights, the
+ * fog, the camera — and the vertex stage's view block because a frame binds more than one pass; and
+ * the material block because a material changes between draws. Their dynamic offsets go in binding
+ * order: the draw's, the pass's, the material's, the view's.
  */
 export function createFlatBindGroupLayout(
   device: GPUDevice,
   variant: FlatVariant,
   skinned = false,
   morphed = false,
+  /** The instanced variant that plays a bone animation: its three vertex textures. */
+  animated = false,
 ): GPUBindGroupLayout {
   const fragment = flatFragmentBindings(variant);
   const entries: GPUBindGroupLayoutEntry[] = [
@@ -307,15 +439,31 @@ export function createFlatBindGroupLayout(
       visibility: VISIBILITY_VERTEX,
       buffer: { type: 'uniform', hasDynamicOffset: true, minBindingSize: FLAT_VERT_SIZE },
     },
+    {
+      binding: FLAT_VIEW_BINDING,
+      visibility: VISIBILITY_VERTEX,
+      /* Dynamic, a slot a pass, for the reason the fragment stage's pass block is. */
+      buffer: { type: 'uniform', hasDynamicOffset: true, minBindingSize: FLAT_VIEW_SIZE },
+    },
     {
       binding: fragment.uniforms,
       visibility: VISIBILITY_FRAGMENT,
       /*
-       * Dynamic, because this block carries the material and a material changes between draws.
-       * See `WebGPURenderer.perFrame` for why one buffer cannot hold two answers.
+       * Dynamic, because a pass's lights and camera change between passes of one frame — a
+       * mirror's, a capture's — and one buffer cannot hold two answers. See `WebGPURenderer.passRing`.
        */
       buffer: { type: 'uniform', hasDynamicOffset: true, minBindingSize: fragment.uniformSize },
     },
+    {
+      binding: materialBinding(fragment),
+      visibility: VISIBILITY_FRAGMENT,
+      /* Dynamic, because a material changes between draws. See `MaterialStore`. */
+      buffer: {
+        type: 'uniform',
+        hasDynamicOffset: true,
+        minBindingSize: fragment.materialSize ?? 0,
+      },
+    },
   ];
 
   /** What each sampler binding was declared as, so a shared one is declared once and agrees. */
@@ -462,12 +610,43 @@ export function createFlatBindGroupLayout(
     }
   }
 
+  /* A bone animation's places, turns and the batch's clocks, read by index as the palette is. */
+  if (animated) {
+    for (const binding of animationBindings()) {
+      entries.push({
+        binding,
+        visibility: VISIBILITY_VERTEX,
+        texture: { sampleType: 'unfilterable-float' },
+      });
+    }
+  }
+
   return device.createBindGroupLayout({
-    label: `flat.layout:${variant}|${flatVertexKey(skinned, morphed)}`,
+    label: `flat.layout:${variant}|${flatVertexKey(skinned, morphed, animated, animated)}`,
     entries,
   });
 }
 
+/** The three textures a bone animation binds, in the order `AnimationViews` holds them. */
+function animationBindings(): [number, number, number] {
+  const textures = flatVertexBindings(false, false, true, true).textures;
+  const at = (name: string): number => {
+    const found = textures[name];
+    if (found === undefined) {
+      throw new Error(`flatPass: the animated vertex shader declares no ${name}`);
+    }
+    return found.texture;
+  };
+  return [at('uBonePlaces'), at('uBoneTurns'), at('uInstanceClocks')];
+}
+
+/** What an animated draw binds besides its material: the clip's two textures and its batch's clocks. */
+export interface AnimationViews {
+  readonly places: GPUTextureView;
+  readonly turns: GPUTextureView;
+  readonly clocks: GPUTextureView;
+}
+
 /**
  * Bind the buffers and one view per declared texture.
  *
@@ -482,7 +661,9 @@ export function createFlatBindGroup(
   layout: GPUBindGroupLayout,
   variant: FlatVariant,
   perDraw: GPUBuffer,
-  perFrame: GPUBuffer,
+  passes: GPUBuffer,
+  materials: GPUBuffer,
+  views: GPUBuffer,
   resolve: (name: string) => { view: GPUTextureView; sampler: GPUSampler },
   /** The palette view for a skinned bind group, or null for an unskinned one. */
   palette: GPUTextureView | null = null,
@@ -490,6 +671,8 @@ export function createFlatBindGroup(
   deltas: GPUTextureView | null = null,
   /** A skinned bind group's cloth views — the draw's, or the stand-in where it sets none. */
   cloth: ClothViews | null = null,
+  /** An animated instanced group's views. See `AnimationViews`. */
+  animation: AnimationViews | null = null,
 ): GPUBindGroup {
   const fragment = flatFragmentBindings(variant);
   const entries: GPUBindGroupEntry[] = [
@@ -497,7 +680,12 @@ export function createFlatBindGroup(
       binding: FLAT_VERT_BINDING,
       resource: { buffer: perDraw, size: FLAT_VERT_SIZE },
     },
-    { binding: fragment.uniforms, resource: { buffer: perFrame, size: fragment.uniformSize } },
+    { binding: FLAT_VIEW_BINDING, resource: { buffer: views, size: FLAT_VIEW_SIZE } },
+    { binding: fragment.uniforms, resource: { buffer: passes, size: fragment.uniformSize } },
+    {
+      binding: materialBinding(fragment),
+      resource: { buffer: materials, size: fragment.materialSize ?? 0 },
+    },
   ];
 
   /* The same dedupe the layout does, and for the same reason: the shadow bindings share one
@@ -536,8 +724,15 @@ export function createFlatBindGroup(
     entries.push({ binding: restAt, resource: cloth.rest });
   }
 
+  if (animation !== null) {
+    const [places, turns, clocks] = animationBindings();
+    entries.push({ binding: places, resource: animation.places });
+    entries.push({ binding: turns, resource: animation.turns });
+    entries.push({ binding: clocks, resource: animation.clocks });
+  }
+
   return device.createBindGroup({
-    label: `flat.bindGroup:${variant}|${flatVertexKey(palette !== null, deltas !== null)}`,
+    label: `flat.bindGroup:${variant}|${flatVertexKey(palette !== null, deltas !== null, animation !== null, animation !== null)}`,
     layout,
     entries,
   });
@@ -609,6 +804,13 @@ export function flatPipeline(
   model: SurfaceModelKind | null = null,
   /** Which half of a skin it draws. The key must carry it: `|ss`, `|sd` or `|sa`. See `SkinHalf`. */
   skinPass: SkinHalf = 'whole',
+  /**
+   * The reflection pass's surface half: the draw again, against its own depth, into the two
+   * targets `reflectionSurfacePass.ts` holds. The key must carry it: `|rs`.
+   */
+  reflectionSurface = false,
+  /** An instanced pipeline that plays a bone animation. The key must carry it: `|anim`. */
+  animated = false,
 ): GPURenderPipeline {
   return cache.get(
     key,
@@ -633,6 +835,8 @@ export function flatPipeline(
         cloth,
         model,
         skinPass,
+        reflectionSurface,
+        animated,
       ),
     true,
   );
@@ -765,10 +969,14 @@ function flatDescriptor(
   cloth = false,
   model: SurfaceModelKind | null = null,
   skinPass: SkinHalf = 'whole',
+  reflectionSurface = false,
+  /** An instanced pipeline that plays a bone animation. See `shaders/boneAnimation.ts`. */
+  animated = false,
 ): GPURenderPipelineDescriptor {
   const offset = depthOffsetForLayer(depthLayer);
-  /* The diffuse and colour halves are the same surface again, into the blur's targets. */
-  const diffuseAlone = skinPass === 'diffuse' || skinPass === 'albedo';
+  /* The diffuse and colour halves are the same surface again, into the blur's targets; and so is
+     the reflection pass's surface half, into its two. */
+  const diffuseAlone = skinPass === 'diffuse' || skinPass === 'albedo' || reflectionSurface;
   return {
     label: key,
     layout: device.createPipelineLayout({ bindGroupLayouts: [layout] }),
@@ -779,13 +987,13 @@ function flatDescriptor(
        * layout validates, draws, and is wrong, exactly as this file's header says.
        */
       module: shaderModule(device, {
-        label: `flat.vert:${flatVertexKey(skinned, morphed, instanced)}`,
-        code: FLAT_VERT_WGSL[flatVertexKey(skinned, morphed, instanced)] ?? '',
+        label: `flat.vert:${flatVertexKey(skinned, morphed, instanced, animated)}`,
+        code: FLAT_VERT_WGSL[flatVertexKey(skinned, morphed, instanced, animated)] ?? '',
       }),
       entryPoint: 'main',
       buffers: vertexBufferLayouts(present, instanced),
       constants: vertexConstants(
-        flatVertexBindings(skinned, morphed, instanced),
+        flatVertexBindings(skinned, morphed, instanced, animated),
         skinEight,
         cloth,
         model === 'lightmap',
@@ -798,31 +1006,33 @@ function flatDescriptor(
       }),
       entryPoint: 'main',
       /* Every variant declares it, so every lit pipeline may set it: see `PipelineCache`. */
-      constants: flatFragmentConstants(variant, cache, model, skinPass),
-      targets: [
-        oitTarget(oit) ??
-          (translucent === 'additive'
-            ? {
-                format: cache.format,
-                blend: {
-                  color: { srcFactor: 'src-alpha', dstFactor: 'one', operation: 'add' },
-                  alpha: { srcFactor: 'zero', dstFactor: 'one', operation: 'add' },
-                },
-              }
-            : translucent
-              ? {
-                  format: cache.format,
-                  blend: {
-                    color: {
-                      srcFactor: 'src-alpha',
-                      dstFactor: 'one-minus-src-alpha',
-                      operation: 'add',
+      constants: flatFragmentConstants(variant, cache, model, skinPass, reflectionSurface),
+      targets: reflectionSurface
+        ? [null, { format: REFLECTION_SURFACE_FORMAT }, { format: REFLECTION_SURFACE_FORMAT }]
+        : [
+            oitTarget(oit) ??
+              (translucent === 'additive'
+                ? {
+                    format: cache.format,
+                    blend: {
+                      color: { srcFactor: 'src-alpha', dstFactor: 'one', operation: 'add' },
+                      alpha: { srcFactor: 'zero', dstFactor: 'one', operation: 'add' },
                     },
-                    alpha: SCENE_ALPHA_COVERS,
-                  },
-                }
-              : { format: cache.format }),
-      ],
+                  }
+                : translucent
+                  ? {
+                      format: cache.format,
+                      blend: {
+                        color: {
+                          srcFactor: 'src-alpha',
+                          dstFactor: 'one-minus-src-alpha',
+                          operation: 'add',
+                        },
+                        alpha: SCENE_ALPHA_COVERS,
+                      },
+                    }
+                  : { format: cache.format }),
+          ],
     },
     primitive: { topology: 'triangle-list', cullMode: doubleSided ? 'none' : 'back' },
     /* Coverage only where a sample count can carry it: WebGPU refuses it at one sample. */

package/src/render/backend/webgpu/instanced.ts

@@ -4,6 +4,8 @@ import { destroySlot } from './instanceCullPass.ts';
 import type { CullSlot } from './instanceCullPass.ts';
 import type { MeshInstances } from '../../instances.ts';
 import type { GpuMesh } from './buffers.ts';
+import { packInstanceClocks } from '../../boneAnimation.ts';
+import { GpuInstanceClocks, type GpuBoneAnimation } from './boneAnimations.ts';
 
 /** VERTEX | COPY_DST, spelled the way every other pass here spells it. */
 const USAGE_VERTEX = 0x0020 | 0x0008;
@@ -36,10 +38,34 @@ export class GpuInstancedBatch {
   slotFrame = -1;
   slotsTaken = 0;
 
-  constructor(device: GPUDevice, mesh: GpuMesh, capacity: number, label: string, cull = false) {
+  /** The bone animation every instance plays, and its clocks; null for a batch that plays none. */
+  readonly animation: GpuBoneAnimation | null;
+  readonly clocks: GpuInstanceClocks | null;
+  /**
+   * An animated batch's own groups — its lit twins by the material group they were built beside, its
+   * shadow groups by what they cut with — and the renderer's group epoch they were built in, after
+   * which every one is dropped. See `WebGPURenderer.animatedGroup`.
+   */
+  readonly litGroups = new Map<GPUBindGroup, GPUBindGroup>();
+  readonly depthGroups = new Map<unknown, GPUBindGroup>();
+  readonly depthPeelGroups = new Map<unknown, GPUBindGroup>();
+  groupEpoch = -1;
+
+  constructor(
+    device: GPUDevice,
+    mesh: GpuMesh,
+    capacity: number,
+    label: string,
+    cull = false,
+    animation: GpuBoneAnimation | null = null,
+  ) {
     this.mesh = mesh;
     this.capacity = capacity;
+    /* An animated batch is culled whole and never by instance: see `submitInstanced`. */
     this.cull = cull;
+    this.animation = animation;
+    this.clocks =
+      animation === null ? null : new GpuInstanceClocks(device, capacity, `${label}.clocks`);
     this.staging = new Float32Array(capacity * INSTANCE_FLOATS);
     this.buffer = device.createBuffer({
       /*
@@ -83,11 +109,16 @@ export class GpuInstancedBatch {
     if (count === 0) return;
     packInstances(data, this.staging);
     queue.writeBuffer(this.buffer, 0, this.staging, 0, count * INSTANCE_FLOATS);
+    if (this.clocks !== null) {
+      packInstanceClocks(data, this.clocks.staging);
+      this.clocks.upload(queue, count);
+    }
   }
 
   dispose(): void {
     this.previousBuffer?.destroy();
     this.buffer.destroy();
+    this.clocks?.dispose();
     for (const slot of this.slots) destroySlot(slot);
   }
 }