@woosh/meep-engine 3.11.2 → 3.12.0

package/src/shade/renderer/volumetrics/shader_volumetrics_build_sun_shadow_volume.js

@@ -1,14 +1,13 @@
 import { CAMERA_UNIFORM_STRUCT } from "../camera/CAMERA_UNIFORM_STRUCT.js";
+import { chunk_texel_coordinate_to_uv } from "../shader/chunk/texture/chunk_texel_coordinate_to_uv.js";
 import { CodeChunk } from "../shader/compiler/CodeChunk.js";
 import { ComputeShader } from "../shader/ComputeShader.js";
 import { ShaderResourceSetDescriptor } from "../shader/resource/ShaderResourceSetDescriptor.js";
+import { u32 } from "../shader/type/WEBGPU_WGSL_PRIMITIVE_TYPES.js";
+import { WebGPUStruct } from "../shader/type/WebGPUStruct.js";
 import { SAMPLER_LINEAR_UPSCALE } from "../texture/SAMPLER_LINEAR_UPSCALE.js";
 import { chunk_volumetrics_sample_froxel_extinction } from "./chunk_volumetrics_sample_froxel_extinction.js";
-import {
-    VOLUMETRICS_SUN_SHADOW_CASCADE_COUNT,
-    VOLUMETRICS_SUN_SHADOW_WORKGROUP_RESOLUTION,
-    volumetrics_sun_shadow_sub_steps
-} from "./VOLUMETRICS_SUN_SHADOW_RESOLUTION.js";
+import { VOLUMETRICS_SUN_SHADOW_WORKGROUP_RESOLUTION } from "./VOLUMETRICS_SUN_SHADOW_RESOLUTION.js";
 import { VOLUMETRICS_METADATA_STRUCT } from "./VOLUMETRICS_METADATA_STRUCT.js";
 
 /**
@@ -30,37 +29,166 @@ import { VOLUMETRICS_METADATA_STRUCT } from "./VOLUMETRICS_METADATA_STRUCT.js";
  * with ringing at sharp density transitions. This engine's medium is already a 3D texture with a
  * defined order along any axis, so the scan is available, exact, and has nothing to ring.
  *
- * One shader per cascade rather than one shader taking a cascade index, because the two things that
- * vary — which cascade's transform to read, and whether there is an outer cascade to start from —
- * both change the *bindings*, and a variant that declares an outer texture it never samples would
- * have to be handed one.
+ * Two shaders, written out: the outermost cascade, whose scan starts at zero, and every cascade
+ * inside it, whose scan starts from the cascade outside. That is a difference in *bindings* — the
+ * second reads a texture the first does not have — so it cannot be data. Which cascade is being
+ * built can be, and is: it arrives as immediate data, so the inner cascades share one pipeline.
  *
- * @param {number} index innermost cascade is 0
- * @returns {ComputeShader}
  * @see chunk_volumetrics_sun_optical_depth for the read side
  * @see build_volumetrics_sun_shadow_transform for where each box is placed, and why it is stable
  */
-function make_shader(index) {
 
-    const seeded = index < VOLUMETRICS_SUN_SHADOW_CASCADE_COUNT - 1;
+/**
+ * Which cascade a dispatch is building.
+ *
+ * Per-dispatch, uniform, read-only and eight bytes, which is the shape immediate data exists for:
+ * no pooled buffer, no staging copy, no bind group.
+ *
+ * `sub_steps` travels with it rather than being derived in the shader because it is a property of
+ * the cascade ladder rather than of the box: {@link volumetrics_sun_shadow_sub_steps} is where the
+ * rule that it is even and at least two lives, and nothing in the metadata lets the shader
+ * re-derive it.
+ *
+ * @type {WebGPUStruct}
+ */
+export const VOLUMETRICS_SUN_SHADOW_BUILD_SETTINGS = WebGPUStruct.from({
+    cascade: u32,
+    sub_steps: u32,
+}, 'VolumetricsSunShadowBuildSettings');
+
+const outermost_resources = new ShaderResourceSetDescriptor();
 
-    const sub_steps = volumetrics_sun_shadow_sub_steps(index);
+outermost_resources.setImmediate("settings", VOLUMETRICS_SUN_SHADOW_BUILD_SETTINGS);
 
-    const resources = new ShaderResourceSetDescriptor();
+outermost_resources.createGroup()
+    .addUniform("volumetrics_metadata", VOLUMETRICS_METADATA_STRUCT)
+    .addUniform("camera", CAMERA_UNIFORM_STRUCT)
+    .addSampler("sLinear", "filtering", SAMPLER_LINEAR_UPSCALE)
+    .addTexture("tParticipatingMedia1", "float", "3d")
+    .addStorageTexture("tSunShadowVolume", "rgba16float", "3d", "write-only");
 
-    const group = resources.createGroup()
-        .addUniform("volumetrics_metadata", VOLUMETRICS_METADATA_STRUCT)
-        .addUniform("camera", CAMERA_UNIFORM_STRUCT)
-        .addSampler("sLinear", "filtering", SAMPLER_LINEAR_UPSCALE)
-        .addTexture("tParticipatingMedia1", "float", "3d");
+/**
+ * The outermost cascade. Nothing is outside it, so its scan starts from zero and it binds no outer
+ * volume.
+ *
+ * @type {ComputeShader}
+ */
+export const shader_volumetrics_build_sun_shadow_volume_outermost = ComputeShader.from({
+    label: 'Volumetrics / build sun shadow volume (outermost)',
+    resources: outermost_resources,
+    body: CodeChunk.from(
+        //language=WGSL
+        `
+@compute @workgroup_size(${VOLUMETRICS_SUN_SHADOW_WORKGROUP_RESOLUTION.join(",")})
+fn main(
+    @builtin(global_invocation_id) column : vec3<u32>
+){
+    let resolution = textureDimensions(tSunShadowVolume);
 
-    if (seeded) {
-        group.addTexture("tSunShadowOuter", "float", "3d");
+    if(any(column.xy >= resolution.xy)){
+        // overflow
+        return;
     }
 
-    group.addStorageTexture("tSunShadowVolume", "rgba16float", "3d", "write-only");
+    // The volume's z axis is the direction light TRAVELS, so this column is one sun ray and slice 0
+    // is its end nearest the sun. Walking it in order turns the per-froxel integral into a running
+    // sum: every voxel's optical depth is its predecessor's plus one slab. That is the whole point
+    // of the pass — the same information the 16-step march produces, in O(voxels) instead of
+    // O(froxels * steps), and produced once rather than once per froxel that happens to lie on the
+    // ray.
+    //
+    // One invocation per column rather than a cooperative scan. The dependency chain is one
+    // cascade's depth long and each link is a texture fetch, so a Hillis-Steele or Blinn-Sengupta
+    // scan would trade this pass's perfectly coherent, register-resident accumulator for log(n)
+    // rounds of shared memory and barriers to speed up something that is already bandwidth-bound.
+    // The aerial pass has the same shape for the same reason.
+    //
+    // Immediate data, so this is uniform across the dispatch.
+    let cascade = settings.cascade;
+
+    let uvw_to_ws = volumetrics_metadata.sun_shadow_uvw_to_ws[cascade];
+
+    // What a voxel must hold is the medium's INTEGRAL across its slab, not the medium's value at
+    // its centre, and one point sample per voxel is not that. A fog volume three units thick
+    // crossed by a 2.3-unit lattice has its entire optical depth resting on one or two samples:
+    // whether they land inside it, on its fade, or just outside swings the answer by the whole
+    // shadow, and the per-frame column jitter under them turns that into a strobe.
+    //
+    // So sample the slab several times and store the storage-resolution result. The sampling rate
+    // and the resolution are separate knobs and only the first one has to beat the medium's
+    // features. The count is per cascade — see volumetrics_sun_shadow_sub_steps — so that every
+    // cascade samples at the same world-space spacing rather than at the same count per voxel.
+    let sub_steps = settings.sub_steps;
+
+    let sub_steps_half = sub_steps / 2u;
+
+    let sub_step_ws = volumetrics_metadata.sun_shadow_step_ws[cascade] / f32(sub_steps);
+
+    let w_per_slice = 1.0 / f32(resolution.z);
+    let w_per_sub_step = w_per_slice / f32(sub_steps);
+
+    let column_uv = texel_coordinate_to_uv(vec2<f32>(column.xy), resolution.xy);
 
-    const body = CodeChunk.from(
+    // The outermost cascade's box contains the whole frustum, and the medium exists only inside
+    // the frustum, so the slab its scan starts on is empty and zero is the whole of the truth about
+    // what is between it and the sun.
+    var optical_depth = vec3<f32>(0.0);
+
+    for(var slice = 0u; slice < resolution.z; slice++){
+
+        let voxel = vec3<u32>(column.xy, slice);
+
+        let w_slice = f32(slice) * w_per_slice;
+
+        for(var sub = 0u; sub < sub_steps; sub++){
+
+            // Halfway through the slab, so the value stored at slice k is the integral from the
+            // box's sun-side face to the voxel's CENTRE. Linear interpolation between two stored
+            // values then reproduces the true integral exactly wherever extinction is constant
+            // across the pair. Storing the integral to the far face instead — a right-endpoint sum,
+            // which is what the march this replaces does — biases every reading by half a slab.
+            if(sub == sub_steps_half){
+                textureStore(tSunShadowVolume, voxel, vec4<f32>(optical_depth, 0.0));
+            }
+
+            let w = w_slice + (f32(sub) + 0.5) * w_per_sub_step;
+
+            let position_ws = (uvw_to_ws * vec4<f32>(column_uv, w, 1.0)).xyz;
+
+            optical_depth += volumetrics_sample_froxel_extinction(position_ws) * sub_step_ws;
+        }
+    }
+}
+    `, [
+            CAMERA_UNIFORM_STRUCT.declaration_chunk,
+            chunk_texel_coordinate_to_uv,
+            chunk_volumetrics_sample_froxel_extinction,
+        ]
+    )
+});
+
+const seeded_resources = new ShaderResourceSetDescriptor();
+
+seeded_resources.setImmediate("settings", VOLUMETRICS_SUN_SHADOW_BUILD_SETTINGS);
+
+seeded_resources.createGroup()
+    .addUniform("volumetrics_metadata", VOLUMETRICS_METADATA_STRUCT)
+    .addUniform("camera", CAMERA_UNIFORM_STRUCT)
+    .addSampler("sLinear", "filtering", SAMPLER_LINEAR_UPSCALE)
+    .addTexture("tParticipatingMedia1", "float", "3d")
+    .addTexture("tSunShadowOuter", "float", "3d")
+    .addStorageTexture("tSunShadowVolume", "rgba16float", "3d", "write-only");
+
+/**
+ * Every cascade but the outermost, seeded from the one outside it. Which one that is arrives as
+ * immediate data, so the whole inner ladder shares this pipeline.
+ *
+ * @type {ComputeShader}
+ */
+export const shader_volumetrics_build_sun_shadow_volume_seeded = ComputeShader.from({
+    label: 'Volumetrics / build sun shadow volume (seeded)',
+    resources: seeded_resources,
+    body: CodeChunk.from(
         //language=WGSL
         `
 @compute @workgroup_size(${VOLUMETRICS_SUN_SHADOW_WORKGROUP_RESOLUTION.join(",")})
@@ -86,9 +214,11 @@ fn main(
     // scan would trade this pass's perfectly coherent, register-resident accumulator for log(n)
     // rounds of shared memory and barriers to speed up something that is already bandwidth-bound.
     // The aerial pass has the same shape for the same reason.
-    const CASCADE = ${index}u;
+    //
+    // Immediate data, so this is uniform across the dispatch.
+    let cascade = settings.cascade;
 
-    let uvw_to_ws = volumetrics_metadata.sun_shadow_uvw_to_ws[CASCADE];
+    let uvw_to_ws = volumetrics_metadata.sun_shadow_uvw_to_ws[cascade];
 
     // What a voxel must hold is the medium's INTEGRAL across its slab, not the medium's value at
     // its centre, and one point sample per voxel is not that. A fog volume three units thick
@@ -100,17 +230,18 @@ fn main(
     // and the resolution are separate knobs and only the first one has to beat the medium's
     // features. The count is per cascade — see volumetrics_sun_shadow_sub_steps — so that every
     // cascade samples at the same world-space spacing rather than at the same count per voxel.
-    const SUB_STEPS = ${sub_steps}u;
+    let sub_steps = settings.sub_steps;
 
-    let sub_step_ws = volumetrics_metadata.sun_shadow_step_ws[CASCADE] / f32(SUB_STEPS);
+    let sub_steps_half = sub_steps / 2u;
 
-    let column_uv = vec2<f32>(
-        (f32(column.x) + 0.5) / f32(resolution.x),
-        (f32(column.y) + 0.5) / f32(resolution.y)
-    );
+    let sub_step_ws = volumetrics_metadata.sun_shadow_step_ws[cascade] / f32(sub_steps);
 
-${seeded
-            ? `    // Seeded from the cascade outside this one, at the world position where this column enters
+    let w_per_slice = 1.0 / f32(resolution.z);
+    let w_per_sub_step = w_per_slice / f32(sub_steps);
+
+    let column_uv = texel_coordinate_to_uv(vec2<f32>(column.xy), resolution.xy);
+
+    // Seeded from the cascade outside this one, at the world position where this column enters
     // this box. Everything between the sun and that plane has already been integrated out there, at
     // that cascade's resolution, so the scan here starts from it rather than from zero and an inner
     // cascade is a refinement of the outer rather than a truncation of the medium. Enshrouded
@@ -121,30 +252,28 @@ ${seeded
     // sub-frustum of it.
     let entry_ws = (uvw_to_ws * vec4<f32>(column_uv, 0.0, 1.0)).xyz;
 
-    let outer_uvw = (volumetrics_metadata.sun_shadow_ws_to_uvw[CASCADE + 1u] * vec4<f32>(entry_ws, 1.0)).xyz;
+    let outer_uvw = (volumetrics_metadata.sun_shadow_ws_to_uvw[cascade + 1u] * vec4<f32>(entry_ws, 1.0)).xyz;
 
-    var optical_depth = textureSampleLevel(tSunShadowOuter, sLinear, saturate(outer_uvw), 0.0).rgb;`
-            : `    // The outermost cascade's box contains the whole frustum, and the medium exists only inside
-    // the frustum, so the slab its scan starts on is empty and zero is the whole of the truth about
-    // what is between it and the sun.
-    var optical_depth = vec3<f32>(0.0);`}
+    var optical_depth = textureSampleLevel(tSunShadowOuter, sLinear, saturate(outer_uvw), 0.0).rgb;
 
     for(var slice = 0u; slice < resolution.z; slice++){
 
         let voxel = vec3<u32>(column.xy, slice);
 
-        for(var sub = 0u; sub < SUB_STEPS; sub++){
+        let w_slice = f32(slice) * w_per_slice;
+
+        for(var sub = 0u; sub < sub_steps; sub++){
 
             // Halfway through the slab, so the value stored at slice k is the integral from the
             // box's sun-side face to the voxel's CENTRE. Linear interpolation between two stored
             // values then reproduces the true integral exactly wherever extinction is constant
             // across the pair. Storing the integral to the far face instead — a right-endpoint sum,
             // which is what the march this replaces does — biases every reading by half a slab.
-            if(sub == SUB_STEPS / 2u){
+            if(sub == sub_steps_half){
                 textureStore(tSunShadowVolume, voxel, vec4<f32>(optical_depth, 0.0));
             }
 
-            let w = (f32(slice) + (f32(sub) + 0.5) / f32(SUB_STEPS)) / f32(resolution.z);
+            let w = w_slice + (f32(sub) + 0.5) * w_per_sub_step;
 
             let position_ws = (uvw_to_ws * vec4<f32>(column_uv, w, 1.0)).xyz;
 
@@ -154,21 +283,8 @@ ${seeded
 }
     `, [
             CAMERA_UNIFORM_STRUCT.declaration_chunk,
+            chunk_texel_coordinate_to_uv,
             chunk_volumetrics_sample_froxel_extinction,
         ]
-    );
-
-    return ComputeShader.from({
-        label: `Volumetrics / build sun shadow volume ${index}`,
-        resources,
-        body
-    });
-}
-
-/**
- * One build shader per cascade, innermost first.
- *
- * @type {ComputeShader[]}
- */
-export const shader_volumetrics_build_sun_shadow_volume =
-    Array.from({ length: VOLUMETRICS_SUN_SHADOW_CASCADE_COUNT }, (_, index) => make_shader(index));
+    )
+});

package/src/shade/renderer/volumetrics/taa/shader_volumetrics_taa.d.ts

@@ -1,5 +1,3 @@
-export const chunk_volumetrics_froxel_coord_to_uvw: CodeChunk;
 export const shader_volumetrics_taa: ComputeShader;
-import { CodeChunk } from "../../shader/compiler/CodeChunk.js";
 import { ComputeShader } from "../../shader/ComputeShader.js";
 //# sourceMappingURL=shader_volumetrics_taa.d.ts.map

package/src/shade/renderer/volumetrics/taa/shader_volumetrics_taa.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"shader_volumetrics_taa.d.ts","sourceRoot":"","sources":["../../../../../../src/shade/renderer/volumetrics/taa/shader_volumetrics_taa.js"],"names":[],"mappings":"AA0FA,8DAsCC;AA+LD,mDAIG;0BA1TuB,oCAAoC;8BAChC,+BAA+B"}
+{"version":3,"file":"shader_volumetrics_taa.d.ts","sourceRoot":"","sources":["../../../../../../src/shade/renderer/volumetrics/taa/shader_volumetrics_taa.js"],"names":[],"mappings":"AA0UA,mDAIG;8BApU2B,+BAA+B"}

package/src/shade/renderer/volumetrics/taa/shader_volumetrics_taa.js

@@ -88,10 +88,23 @@ fn texture_sample_blackman_harris_3d(
     [chunk_filter_blackman_harris229]
 );
 
-export const chunk_volumetrics_froxel_coord_to_uvw = CodeChunk.from(
+/**
+ * Froxel coordinate -> a sample point for `frustum_sample_point`, NOT a froxel texture coordinate.
+ *
+ * That distinction is the whole reason this is not `volumetrics_position_world_to_froxel_uvw`
+ * turned around. There `w` is slice space -- `log2(fma(depth, x, y)) * z`, corrected by the
+ * cell-averaged content offset -- because it addresses the 3D texture. Here `w` is a LINEAR
+ * fraction of the near..far range, because its only consumer is `frustum_sample_point`, which
+ * mixes the near and far planes linearly. The two agree at the endpoints and nowhere between, so
+ * feeding this `w` to a froxel textureSample reads the wrong slab.
+ *
+ * @see volumetrics_position_world_to_froxel_uvw
+ * @type {CodeChunk}
+ */
+const chunk_volumetrics_froxel_coord_to_frustum_uvw = CodeChunk.from(
     //language=WGSL
     `
-fn volumetrics_froxel_coord_to_uvw(
+fn volumetrics_froxel_coord_to_frustum_uvw(
     coord: vec3<u32>,
     res: vec3<u32>,
     camera: ${CAMERA_UNIFORM_STRUCT.wgsl_ref},
@@ -139,7 +152,7 @@ fn world_pos_from_froxel(
     cluster_params: vec3f,
     frustum: array<vec4<f32>, 6>
 ) -> vec3f {
-    let uvw = volumetrics_froxel_coord_to_uvw(coord, res, camera, cluster_params);
+    let uvw = volumetrics_froxel_coord_to_frustum_uvw(coord, res, camera, cluster_params);
 
     // Reconstruct the froxel-centre world position. frustum_sample_point intersects the
     // interpolated L/R, B/T, N/F planes; X maps left->left and Z maps near->far directly, so only
@@ -150,7 +163,7 @@ fn world_pos_from_froxel(
     return frustum_sample_point(frustum, vec3f(uvw.x, 1.0 - uvw.y, uvw.z));
 }
 `, [
-        chunk_volumetrics_froxel_coord_to_uvw,
+        chunk_volumetrics_froxel_coord_to_frustum_uvw,
         chunk_frustum_sample_point,
     ]);
 
@@ -312,8 +325,6 @@ fn main(@builtin(global_invocation_id) voxel_id: vec3<u32>) {
         chunk_ndc_to_uv,
         chunk_volumetric_history_bounding_box,
         chunk_aabb4_normalized_dist,
-        chunk_volumetrics_froxel_coord_to_uvw,
-        chunk_frustum_sample_point,
     ]
 );