@driftengine/core 4.4.2 → 4.5.0

package/src/render/shaders/gpudriven/sunGlass.wgsl.ts

@@ -0,0 +1,139 @@
+/**
+ * The sun through glass for the device: `gpudriven/sunGlass.ts`'s `glassTintAt`, term for term, as
+ * the text the shading pass and the parity entry point both include.
+ *
+ * **The reads are the including module's**, as `SHADE_SHADOW_WGSL`'s are: `glassDepthAt(u, v)` the
+ * nearest pane's depth, `glassTintTexel(u, v)` the tint at the finest level, and
+ * `glassTintAround(u, v, radius)` the tint averaged over a spread of that radius. The shading pass
+ * answers the last from the tint's mip chain; the parity entry answers it with the reference's own
+ * box average, so everything around the reads is checked as it stands. Every constant is the
+ * reference's, interpolated rather than restated.
+ */
+import { FROST_RADIUS_CAP, FROST_SPREAD } from '../../glassShadow.ts';
+import {
+  SHADOW_DEPTH_BIAS,
+  SHADOW_PCF_OFFSETS,
+  SHADOW_TEXEL_SCALE,
+} from '../../gpudriven/shadow.ts';
+import { SHADE_SHADOW_WGSL } from './shadow.wgsl.ts';
+
+export const SUN_GLASS_WGSL = /* wgsl */ `
+struct SunGlassSettings {
+  mapSize: f32,
+  depthSpan: f32,
+  uvPerMetre: f32,
+  taps: f32,
+}
+
+/*
+ * A coarse texel with the open ground's white taken back out, by how milky the pane at the
+ * receiver is: the pane's colour and its share of the texel. glassShadow.ts's unmixTint.
+ */
+fn sunGlassUnmix(mip: vec4<f32>, centreClarity: f32) -> vec4<f32> {
+  if (centreClarity >= 1.0) { return vec4<f32>(mip.xyz, 1.0); }
+  let share = clamp((1.0 - mip.w) / (1.0 - centreClarity), 0.0, 1.0);
+  if (share <= 0.0) { return vec4<f32>(1.0, 1.0, 1.0, 0.0); }
+  let colour = clamp((mip.xyz - vec3<f32>(1.0 - share)) / share, vec3<f32>(0.0), vec3<f32>(1.0));
+  return vec4<f32>(colour, share);
+}
+
+/* What the sun keeps through glass at map position (u, v) and light depth "depth". */
+fn sunGlass(u: f32, v: f32, depth: f32, settings: SunGlassSettings) -> vec3<f32> {
+  if (u < 0.0 || u > 1.0 || v < 0.0 || v > 1.0) { return vec3<f32>(1.0); }
+  let texel = ${SHADOW_TEXEL_SCALE.toFixed(4)} / settings.mapSize;
+  let bias = ${SHADOW_DEPTH_BIAS.toFixed(4)} / settings.depthSpan;
+  let taps = min(u32(settings.taps), ${SHADOW_PCF_OFFSETS.length}u);
+  let clarity = glassTintTexel(u, v).w;
+  let behind = max(depth - glassDepthAt(u, v), 0.0) * settings.depthSpan;
+  let spread = (1.0 - clarity) * ${FROST_SPREAD.toFixed(6)} * behind * settings.uvPerMetre;
+  let radius = min(max(texel, spread), texel * ${FROST_RADIUS_CAP.toFixed(1)});
+
+  var cover = 0.0;
+  var colour = vec3<f32>(0.0);
+  var found = 0.0;
+  for (var i = 0u; i < taps; i = i + 1u) {
+    let offset = SHADOW_TAPS[i];
+    /* The outline: a tap counts the pane only where the receiver is behind it, by the bias. */
+    if (depth > glassDepthAt(u + offset.x * texel, v + offset.y * texel) + bias) {
+      cover = cover + 1.0;
+    }
+    /* The colour: the spread's taps, each an average, unmixed and weighed by the pane's share. */
+    let pane = sunGlassUnmix(glassTintAround(u + offset.x * radius, v + offset.y * radius, radius), clarity);
+    colour = colour + pane.xyz * pane.w;
+    found = found + pane.w;
+  }
+  let mixed = select(vec3<f32>(1.0), colour / max(found, 1e-6), found > 0.0);
+  return mix(vec3<f32>(1.0), mixed, cover / f32(max(taps, 1u)));
+}
+`;
+
+/**
+ * `sunGlass` over maps held in storage buffers, read as the reference reads them, so the whole
+ * lookup can be compared with `glassTintAt` rather than only the arithmetic between the reads.
+ *
+ * Four floats a case in — u, v, the receiver's depth, one spare — and four out, the tint and one
+ * spare. The config is the map's size, the depth span, map units a metre, and the tap count.
+ */
+export const SUN_GLASS_PARITY_WGSL = /* wgsl */ `
+@group(0) @binding(0) var<storage, read> depths: array<f32>;
+@group(0) @binding(1) var<storage, read> tints: array<f32>;
+@group(0) @binding(2) var<storage, read> cases: array<f32>;
+@group(0) @binding(3) var<storage, read> config: array<f32>;
+@group(0) @binding(4) var<storage, read_write> results: array<f32>;
+
+/* The opaque lookup's own read, which this entry never calls: included only for its tap table. */
+fn shadowDepthAt(u: f32, v: f32) -> f32 { return 1.0; }
+${SHADE_SHADOW_WGSL}
+
+fn glassIndex(u: f32, v: f32) -> u32 {
+  let size = config[0];
+  let x = u32(clamp(floor(u * size), 0.0, size - 1.0));
+  let y = u32(clamp(floor(v * size), 0.0, size - 1.0));
+  return y * u32(size) + x;
+}
+
+fn glassDepthAt(u: f32, v: f32) -> f32 {
+  return depths[glassIndex(u, v)];
+}
+
+fn glassTintTexel(u: f32, v: f32) -> vec4<f32> {
+  let at = glassIndex(u, v) * 4u;
+  return vec4<f32>(tints[at], tints[at + 1u], tints[at + 2u], tints[at + 3u]);
+}
+
+/* The reference's coarse read: the average of every texel within the radius. */
+fn glassTintAround(u: f32, v: f32, radius: f32) -> vec4<f32> {
+  let size = config[0];
+  let reach = radius * size;
+  let x0 = u32(max(0.0, floor(u * size - reach)));
+  let x1 = u32(min(size - 1.0, floor(u * size + reach)));
+  let y0 = u32(max(0.0, floor(v * size - reach)));
+  let y1 = u32(min(size - 1.0, floor(v * size + reach)));
+  var sum = vec4<f32>(0.0);
+  var n = 0.0;
+  for (var y = y0; y <= y1; y = y + 1u) {
+    for (var x = x0; x <= x1; x = x + 1u) {
+      let at = (y * u32(size) + x) * 4u;
+      sum = sum + vec4<f32>(tints[at], tints[at + 1u], tints[at + 2u], tints[at + 3u]);
+      n = n + 1.0;
+    }
+  }
+  return sum / max(n, 1.0);
+}
+${SUN_GLASS_WGSL}
+
+@compute @workgroup_size(64)
+fn main(@builtin(global_invocation_id) id: vec3<u32>) {
+  if (id.x * 4u + 3u >= arrayLength(&cases)) { return; }
+  var settings: SunGlassSettings;
+  settings.mapSize = config[0];
+  settings.depthSpan = config[1];
+  settings.uvPerMetre = config[2];
+  settings.taps = config[3];
+  let at = id.x * 4u;
+  let tint = sunGlass(cases[at], cases[at + 1u], cases[at + 2u], settings);
+  results[at] = tint.x;
+  results[at + 1u] = tint.y;
+  results[at + 2u] = tint.z;
+}
+`;

package/src/render/shaders/lightVolume.ts

@@ -47,6 +47,7 @@
  *   halves of one frame disagreeing about the same light.
  */
 
+import { SUN_DYNAMIC_LAYER, SUN_PEELED_LAYER, SUN_STATIC_LAYER } from '../shadowMap.ts';
 import { resolveConditionals } from './conditionals.ts';
 
 export const LIGHT_VOLUME_VERT = `#version 300 es
@@ -159,9 +160,8 @@ uniform vec3 uDustOffset;
 uniform float uSunShadow;
 uniform mat4 uLightViewProj;
 uniform float uShadowMapSize;
-uniform sampler2D uStaticShadowMap;
-uniform sampler2D uPeeledShadowMap;
-uniform sampler2D uDynamicShadowMap;
+/* The sun's static, moving and peeled maps, as layers of one array: see shadowMap.ts. */
+uniform highp sampler2DArray uSunShadows;
 uniform int uPeeledShadowEnabled;
 #endif
 
@@ -324,11 +324,11 @@ float sunReach(vec3 worldPos) {
   if (edgeFade <= 0.0) return 1.0;
 
   float compare = p.z - BIAS;
-  float blocked = occlusion(textureLod(uStaticShadowMap, p.xy, 0.0).r, compare);
+  float blocked = occlusion(textureLod(uSunShadows, vec3(p.xy, ${SUN_STATIC_LAYER}.0), 0.0).r, compare);
   if (uPeeledShadowEnabled != 0) {
-    blocked = max(blocked, occlusion(textureLod(uPeeledShadowMap, p.xy, 0.0).r, compare));
+    blocked = max(blocked, occlusion(textureLod(uSunShadows, vec3(p.xy, ${SUN_PEELED_LAYER}.0), 0.0).r, compare));
   }
-  blocked = max(blocked, occlusion(textureLod(uDynamicShadowMap, p.xy, 0.0).r, compare));
+  blocked = max(blocked, occlusion(textureLod(uSunShadows, vec3(p.xy, ${SUN_DYNAMIC_LAYER}.0), 0.0).r, compare));
   return mix(1.0, 1.0 - blocked, uSunShadow * edgeFade);
 }
 #endif

package/src/render/shaders/octahedralResolve.ts

@@ -12,6 +12,30 @@ void main() {
   gl_Position = vec4(p * 2.0 - 1.0, 0.0, 1.0);
 }`;
 
+/**
+ * Which cube face owns this octahedral texel, and where in that face it looks: `d`, `forward` and
+ * `faceUv`, with every texel another face owns discarded. **One text for both resolves**, depth
+ * and tint, because the tie-breaking has to agree with itself as well as with the CPU: a sliver
+ * of texels given to one face's depth and another face's colour is a tinted seam.
+ */
+const FACE_REGION = `  vec3 d = octDecode(gl_FragCoord.xy / uEdge);
+
+  /*
+   * Which face owns this texel. The tie-breaking is x, then y, then z, and it has to match the
+   * CPU's exactly or a sliver of texels along a boundary belongs to two faces or to none.
+   */
+  vec3 a = abs(d);
+  int face;
+  if (a.x >= a.y && a.x >= a.z) face = d.x >= 0.0 ? 0 : 1;
+  else if (a.y >= a.z) face = d.y >= 0.0 ? 2 : 3;
+  else face = d.z >= 0.0 ? 4 : 5;
+  if (face != uFaceIndex) discard;
+
+  /* lookAt looks down -z, so a direction inside this face has a negative z here. */
+  vec3 v = uFaceRotation * d;
+  float forward = -v.z;
+  vec2 faceUv = (v.xy / forward) * 0.5 + 0.5;`;
+
 /**
  * One cube face into its region of an octahedral map, as linear radial distance.
  *
@@ -46,23 +70,7 @@ uniform float uEdge;
 ${OCTAHEDRAL_GLSL}
 
 void main() {
-  vec3 d = octDecode(gl_FragCoord.xy / uEdge);
-
-  /*
-   * Which face owns this texel. The tie-breaking is x, then y, then z, and it has to match the
-   * CPU's exactly or a sliver of texels along a boundary belongs to two faces or to none.
-   */
-  vec3 a = abs(d);
-  int face;
-  if (a.x >= a.y && a.x >= a.z) face = d.x >= 0.0 ? 0 : 1;
-  else if (a.y >= a.z) face = d.y >= 0.0 ? 2 : 3;
-  else face = d.z >= 0.0 ? 4 : 5;
-  if (face != uFaceIndex) discard;
-
-  /* lookAt looks down -z, so a direction inside this face has a negative z here. */
-  vec3 v = uFaceRotation * d;
-  float forward = -v.z;
-  vec2 faceUv = (v.xy / forward) * 0.5 + 0.5;
+${FACE_REGION}
 
   /*
    * textureLod at level zero rather than plain texture. The scratch carries one storage level and
@@ -81,3 +89,28 @@ void main() {
   /* d is a unit vector, so forward is its cosine to the axis and this is the true radius. */
   gl_FragDepth = clamp((faceLocalZ / forward) / uFar, 0.0, 1.0);
 }`;
+
+/**
+ * One face of a lamp's glass tint into its region of the tint layer: the depth resolve's region,
+ * and a colour copied rather than a depth decoded.
+ *
+ * **Nearest, at level zero**, like the depth beside it: the tint's outline has to land on the
+ * glass depth's texels, and a filtered copy would smear a pane's colour a texel past its edge.
+ */
+export const OCTAHEDRAL_RESOLVE_TINT_FRAG = `#version 300 es
+precision highp float;
+
+/** The face the bake just tinted, as colour. */
+uniform highp sampler2D uFace;
+uniform mat3 uFaceRotation;
+uniform int uFaceIndex;
+uniform float uEdge;
+
+out vec4 outTint;
+
+${OCTAHEDRAL_GLSL}
+
+void main() {
+${FACE_REGION}
+  outTint = textureLod(uFace, faceUv, 0.0);
+}`;

package/src/render/shaders/recon/lateDepth.wgsl.ts

@@ -0,0 +1,47 @@
+/**
+ * The render's depth, taken up to the output size for the draws that land after the upscale.
+ *
+ * **The nearest of the four render texels around each output pixel**, so a translucent draw is
+ * never let through an opaque edge the render did cover. What it gives up: where a translucent
+ * surface crosses an opaque silhouette, the edge steps at the render's resolution.
+ *
+ * Hand-written, and built by concatenation rather than a template literal: the reconstruction's
+ * passes have no WebGL2 twin (the AGENTS.md exception for them), and a `${}` inside WGSL source is
+ * the thing that rule forbids.
+ */
+import { REVERSED_DEPTH } from '../../depthConvention.ts';
+
+/** The nearer of two stored depths: the larger in a reversed buffer. */
+const NEARER = REVERSED_DEPTH ? 'max' : 'min';
+/** Where the search for the nearest starts: the far plane. */
+const FARTHEST = REVERSED_DEPTH ? '0.0' : '1.0';
+
+export const LATE_DEPTH_WGSL =
+  '@group(0) @binding(0) var source: texture_depth_2d;\n' +
+  '@group(0) @binding(1) var<uniform> sizes: vec4<f32>;\n' +
+  '\n' +
+  '@vertex\n' +
+  'fn lateDepthVert(@builtin(vertex_index) index: u32) -> @builtin(position) vec4<f32> {\n' +
+  '  let x = f32((index << 1u) & 2u);\n' +
+  '  let y = f32(index & 2u);\n' +
+  '  return vec4<f32>(x * 2.0 - 1.0, 1.0 - y * 2.0, 0.0, 1.0);\n' +
+  '}\n' +
+  '\n' +
+  '@fragment\n' +
+  'fn lateDepthFrag(@builtin(position) at: vec4<f32>) -> @builtin(frag_depth) f32 {\n' +
+  '  let centre = at.xy * sizes.xy / sizes.zw - vec2<f32>(0.5);\n' +
+  '  let base = vec2<i32>(floor(centre));\n' +
+  '  let top = vec2<i32>(sizes.xy) - vec2<i32>(1);\n' +
+  '  var nearest = ' +
+  FARTHEST +
+  ';\n' +
+  '  for (var y = 0; y < 2; y += 1) {\n' +
+  '    for (var x = 0; x < 2; x += 1) {\n' +
+  '      let texel = clamp(base + vec2<i32>(x, y), vec2<i32>(0), top);\n' +
+  '      nearest = ' +
+  NEARER +
+  '(nearest, textureLoad(source, texel, 0));\n' +
+  '    }\n' +
+  '  }\n' +
+  '  return nearest;\n' +
+  '}\n';

package/src/render/shaders/recon/motion.wgsl.ts

@@ -88,6 +88,90 @@ fn motionUv(clip: vec4<f32>) -> vec2<f32> {
   return clip.xy / clip.w * 0.5 + vec2<f32>(0.5);
 }
 
+/*
+ * The skinned stage: the joints this frame and last frame, read from two palette textures of the
+ * frame's ring. The same four weighted products in the same order as the scene's own skinning, so
+ * the depth this stage produces is the depth the scene wrote.
+ */
+@group(1) @binding(0) var palette: texture_2d<f32>;
+@group(1) @binding(1) var previousPalette: texture_2d<f32>;
+
+fn jointOf(source: texture_2d<f32>, index: i32) -> mat4x4<f32> {
+  let x = index * 4;
+  return mat4x4<f32>(
+    textureLoad(source, vec2<i32>(x, 0), 0),
+    textureLoad(source, vec2<i32>(x + 1, 0), 0),
+    textureLoad(source, vec2<i32>(x + 2, 0), 0),
+    textureLoad(source, vec2<i32>(x + 3, 0), 0));
+}
+
+fn skinOf(source: texture_2d<f32>, joints: vec4<f32>, weights: vec4<f32>) -> mat4x4<f32> {
+  return jointOf(source, i32(joints.x)) * weights.x
+       + jointOf(source, i32(joints.y)) * weights.y
+       + jointOf(source, i32(joints.z)) * weights.z
+       + jointOf(source, i32(joints.w)) * weights.w;
+}
+
+@vertex
+fn motionSkinnedVert(
+  @location(0) position: vec3<f32>,
+  @location(11) joints: vec4<f32>,
+  @location(12) weights: vec4<f32>,
+) -> Varyings {
+  let world = draw.model * (skinOf(palette, joints, weights) * vec4<f32>(position, 1.0));
+  let wasWorld =
+    draw.previousModel * (skinOf(previousPalette, joints, weights) * vec4<f32>(position, 1.0));
+  var out: Varyings;
+  out.clip = frame.raster * world;
+  out.now = frame.viewProj * world;
+  out.was = frame.previousViewProj * wasWorld;
+  return out;
+}
+
+/*
+ * The rewritten stage: a mesh whose vertices were rewritten this frame reads last frame's rows from
+ * a second buffer of the same stride, so each vertex carries both of its positions.
+ */
+@vertex
+fn motionDynamicVert(
+  @location(0) position: vec3<f32>,
+  @location(1) previousPosition: vec3<f32>,
+) -> Varyings {
+  let world = draw.model * vec4<f32>(position, 1.0);
+  let wasWorld = draw.previousModel * vec4<f32>(previousPosition, 1.0);
+  var out: Varyings;
+  out.clip = frame.raster * world;
+  out.now = frame.viewProj * world;
+  out.was = frame.previousViewProj * wasWorld;
+  return out;
+}
+
+/*
+ * The instanced stage: each instance's matrix this frame and last frame, read from the batch's
+ * placement and the copy of last frame's it kept. Slot for slot, so an instance is paired with
+ * itself; the per-draw matrices are not read.
+ */
+@vertex
+fn motionInstancedVert(
+  @location(0) position: vec3<f32>,
+  @location(2) m0: vec4<f32>,
+  @location(3) m1: vec4<f32>,
+  @location(4) m2: vec4<f32>,
+  @location(5) m3: vec4<f32>,
+  @location(6) p0: vec4<f32>,
+  @location(7) p1: vec4<f32>,
+  @location(8) p2: vec4<f32>,
+  @location(9) p3: vec4<f32>,
+) -> Varyings {
+  let world = mat4x4<f32>(m0, m1, m2, m3) * vec4<f32>(position, 1.0);
+  let wasWorld = mat4x4<f32>(p0, p1, p2, p3) * vec4<f32>(position, 1.0);
+  var out: Varyings;
+  out.clip = frame.raster * world;
+  out.now = frame.viewProj * world;
+  out.was = frame.previousViewProj * wasWorld;
+  return out;
+}
+
 @fragment
 fn motionFrag(in: Varyings) -> @location(0) vec4<f32> {
   /*

package/src/render/shadowCasters.ts

@@ -1,3 +1,4 @@
+import { resolveGlass, type GlassOptions, type ResolvedGlass } from './glass.ts';
 import type { ReadonlyMat4 } from 'gl-matrix';
 import type { InstanceData } from './instancedMesh.ts';
 import type { InstancedHandle, MeshHandle, ScatterHandle } from './backend/api.ts';
@@ -41,7 +42,33 @@ import type { SurfaceTextureHandle } from './backend/api.ts';
  * outwards, so a required parameter added in a minor stops every consumer's sink and every test
  * double compiling. A sink that ignores it behaves exactly as it did.
  */
-export type SceneCasterMaterial = SurfaceMaterial<SurfaceTextureHandle> | null;
+export type SceneCasterMaterial =
+  | (SurfaceMaterial<SurfaceTextureHandle> & {
+      /**
+       * The caster is glass: it lets light through, so it writes no depth into any shadow map and a
+       * lamp behind it lights what is beyond. The colour replay into a mirror or a probe still draws
+       * it. With `RenderQuality.glassShadows` on it casts a coloured shadow instead: see
+       * `glassShadow.ts`.
+       */
+      readonly glass?: GlassOptions;
+    })
+  | null;
+
+/**
+ * Whether a caster writes depth into a shadow map: everything but glass. One line in each depth
+ * sink, so the answer is the same in every shadow a backend draws.
+ */
+export function castsDepth(material: SceneCasterMaterial | undefined): boolean {
+  return (material?.glass?.transmission ?? 0) <= 0;
+}
+
+/**
+ * The caster's glass, resolved into `out`, and whether it is glass at all — the rule `castsDepth`
+ * states from the other side, so a caster is exactly one of the two.
+ */
+export function glassOf(material: SceneCasterMaterial | undefined, out: ResolvedGlass): boolean {
+  return resolveGlass(material?.glass, out);
+}
 
 export interface ShadowCasterSink {
   /** A rigid mesh at a model transform: the world, a prop, a static piece of scenery. */