@woosh/meep-engine 3.56.0 → 3.57.0

package/src/shade/renderer/fluid/publish/shader_fluid_wind_publish.js

@@ -1,122 +1,142 @@
-import { ComputeShader } from "../../shader/ComputeShader.js";
-import { CodeChunk } from "../../shader/compiler/CodeChunk.js";
-import { ShaderResourceSetDescriptor } from "../../shader/resource/ShaderResourceSetDescriptor.js";
-import { WIND_FIELD_TILE_BORDER, WIND_FIELD_TILE_FORMAT } from "../../wind/WIND_FIELD.js";
-import { chunk_grid_tile_window } from "../../shader/chunk/grid/chunk_grid_tile_window.js";
-import { chunk_fluid_cascade_parent_velocity } from "../coupling/chunk_fluid_cascade_parent.js";
-import { FLUID_MAC_STEP_STRUCT } from "../mac/FLUID_MAC_STEP_STRUCT.js";
-import { FLUID_CASCADE_PARENT_STRUCT } from "../coupling/FLUID_CASCADE_PARENT_STRUCT.js";
-import { chunk_fluid_wind_publish_cell } from "./chunk_fluid_wind_publish_cell.js";
-import { FLUID_WIND_PUBLISH_STRUCT } from "./FLUID_WIND_PUBLISH_STRUCT.js";
-
-/**
- * Invocations along each axis of a workgroup: one per texel of the tile, whose sides are N + 2.
- * @type {number[]}
- */
-export const FLUID_WIND_PUBLISH_GROUP_SIZE = Object.freeze([4, 4, 4]);
-
-const resources = new ShaderResourceSetDescriptor();
-
-resources.createGroup()
-    .addUniform("step", FLUID_MAC_STEP_STRUCT)
-    .addUniform("publish", FLUID_WIND_PUBLISH_STRUCT)
-    .addUniform("parent", FLUID_CASCADE_PARENT_STRUCT)
-    .addTexture("source_u", "unfilterable-float", "3d")
-    .addTexture("source_v", "unfilterable-float", "3d")
-    .addTexture("source_w", "unfilterable-float", "3d")
-    .addTexture("parent_u", "unfilterable-float", "3d")
-    .addTexture("parent_v", "unfilterable-float", "3d")
-    .addTexture("parent_w", "unfilterable-float", "3d")
-    .addStorageTexture("tile", WIND_FIELD_TILE_FORMAT, "3d", "write-only");
-
-const body = CodeChunk.from(
-    //language=WGSL
-    `
-const FLUID_WIND_PUBLISH_BORDER: u32 = ${WIND_FIELD_TILE_BORDER}u;
-
-// the step's end velocity, the persistent set, through its toroidal windows; beyond the window, the
-// cascade's boundary: its parent's identity read as of its record, as its seams read it (the skirt is
-// one face deep, so the read itself, not a step's ghosts)
-fn fluid_wind_publish_face_u(face: vec3<i32>) -> f32 {
-    if (grid_tile_window_contains(face, step.faces_u.size)) {
-        return textureLoad(source_u, grid_tile_window_storage(face, step.faces_u.corner_storage, step.faces_u.period), 0).r;
-    }
-
-    return fluid_parent_face_u(face, step.cells.corner - parent.origin);
-}
-
-fn fluid_wind_publish_face_v(face: vec3<i32>) -> f32 {
-    if (grid_tile_window_contains(face, step.faces_v.size)) {
-        return textureLoad(source_v, grid_tile_window_storage(face, step.faces_v.corner_storage, step.faces_v.period), 0).r;
-    }
-
-    return fluid_parent_face_v(face, step.cells.corner - parent.origin);
-}
-
-fn fluid_wind_publish_face_w(face: vec3<i32>) -> f32 {
-    if (grid_tile_window_contains(face, step.faces_w.size)) {
-        return textureLoad(source_w, grid_tile_window_storage(face, step.faces_w.corner_storage, step.faces_w.period), 0).r;
-    }
-
-    return fluid_parent_face_w(face, step.cells.corner - parent.origin);
-}
-
-@compute @workgroup_size(${FLUID_WIND_PUBLISH_GROUP_SIZE.join(', ')})
-fn main(@builtin(global_invocation_id) id: vec3<u32>) {
-    // the window's cells and a skirt on every side
-    let tile_size = step.cells.size + vec3<u32>(2u * FLUID_WIND_PUBLISH_BORDER);
-
-    if (any(id >= tile_size)) {
-        return;
-    }
-
-    // tile texel t holds window-local cell t - border
-    let cell = vec3<i32>(id) - vec3<i32>(i32(FLUID_WIND_PUBLISH_BORDER));
-
-    textureStore(tile, publish.tile_origin + id, vec4<f32>(fluid_wind_publish_cell(cell, step.far_velocity), 0.0));
-}
-`, [
-        chunk_grid_tile_window,
-        chunk_fluid_cascade_parent_velocity,
-        chunk_fluid_wind_publish_cell,
-    ]
-);
-
-/**
- * A cascade's publish into the wind field (GPU_FLUID_CASCADES_PROPOSAL_2026_10_16.md §11, the publish
- * half of pass 6): every texel of its atlas tile, the window's cells and the one-texel skirt around
- * them, gets `chunk_fluid_wind_publish_cell` — the velocity at that cell's centre less the step's far
- * field — as `rgba16float`, alpha 0.
- *
- * The skirt is the cells one past the window. Each takes the window's faces where it has them (along
- * an axis, the boundary face plane) and, for the rest, what the MAC seams answer beyond the window:
- * the parent's identity read as of the cascade's read record (`chunk_fluid_cascade_parent_velocity`),
- * which for the coarsest is the far field. Beside the far field, a skirt cell beside a face of the
- * window takes half that boundary face's deviation on the face's axis and none on the others, and one
- * past an edge or a corner none at all.
- *
- * It writes its own tile and never samples the atlas: the atlas is bound as a write-only storage
- * texture and nothing else.
- *
- * One invocation a texel, in 4 × 4 × 4 workgroups over the tile, dispatched
- * `ceil((N + 2) / 4)` on each axis; each reads its cell's six faces (`textureLoad`) and stores one
- * texel. Standalone until it is fused into the gradient pass (proposal §8 (3)), where the faces are in
- * workgroup memory already.
- *
- * Bindings:
- * - `step`: {@link FLUID_MAC_STEP_STRUCT}: the cascade's windows and `far_velocity`, `v_far(t_s,c)`;
- *   `cells.size` is the window, and the tile's content;
- * - `publish`: {@link FLUID_WIND_PUBLISH_STRUCT}: the tile's texel origin in the atlas;
- * - `parent`: {@link FLUID_CASCADE_PARENT_STRUCT}, and `parent_u`, `parent_v`, `parent_w` the parent's
- *   persistent faces, read for the skirt;
- * - `source_u`, `source_v`, `source_w`: the cascade's persistent velocity after its step, `r32float`
- *   3-D textures, toroidal through `step.faces_*`;
- * - `tile`: written, the atlas, `rgba16float` 3-D storage texture.
- *
- * @type {ComputeShader}
- */
-export const shader_fluid_wind_publish = ComputeShader.from({
-    label: "Fluid / wind publish",
-    body,
-    resources,
-});
+import { ComputeShader } from "../../shader/ComputeShader.js";
+import { CodeChunk } from "../../shader/compiler/CodeChunk.js";
+import { ShaderResourceSetDescriptor } from "../../shader/resource/ShaderResourceSetDescriptor.js";
+import { WIND_FIELD_TILE_BORDER, WIND_FIELD_TILE_FORMAT } from "../../wind/WIND_FIELD.js";
+import { chunk_grid_tile_window } from "../../shader/chunk/grid/chunk_grid_tile_window.js";
+import { chunk_fluid_cascade_parent_velocity } from "../coupling/chunk_fluid_cascade_parent.js";
+import { FLUID_MAC_STEP_STRUCT } from "../mac/FLUID_MAC_STEP_STRUCT.js";
+import { FLUID_CASCADE_PARENT_STRUCT } from "../coupling/FLUID_CASCADE_PARENT_STRUCT.js";
+import { chunk_fluid_wind_publish_cell } from "./chunk_fluid_wind_publish_cell.js";
+import { FLUID_WIND_PUBLISH_STRUCT } from "./FLUID_WIND_PUBLISH_STRUCT.js";
+
+/**
+ * Invocations along each axis of a workgroup: one per texel of the tile, whose sides are N + 2.
+ * @type {number[]}
+ */
+export const FLUID_WIND_PUBLISH_GROUP_SIZE = Object.freeze([4, 4, 4]);
+
+const resources = new ShaderResourceSetDescriptor();
+
+resources.createGroup()
+    .addUniform("step", FLUID_MAC_STEP_STRUCT)
+    .addUniform("publish", FLUID_WIND_PUBLISH_STRUCT)
+    .addUniform("parent", FLUID_CASCADE_PARENT_STRUCT)
+    .addTexture("source_u", "unfilterable-float", "3d")
+    .addTexture("source_v", "unfilterable-float", "3d")
+    .addTexture("source_w", "unfilterable-float", "3d")
+    .addTexture("parent_u", "unfilterable-float", "3d")
+    .addTexture("parent_v", "unfilterable-float", "3d")
+    .addTexture("parent_w", "unfilterable-float", "3d")
+    .addStorageTexture("tile", WIND_FIELD_TILE_FORMAT, "3d", "write-only");
+
+const body = CodeChunk.from(
+    //language=WGSL
+    `
+const FLUID_WIND_PUBLISH_BORDER: u32 = ${WIND_FIELD_TILE_BORDER}u;
+
+// the step's end velocity, the persistent set, through its toroidal windows; beyond the window, the
+// cascade's boundary: its parent's identity read as of its record, as its seams read it (the skirt is
+// one face deep, so the read itself, not a step's ghosts)
+fn fluid_wind_publish_face_u(face: vec3<i32>) -> f32 {
+    if (grid_tile_window_contains(face, step.faces_u.size)) {
+        return textureLoad(source_u, grid_tile_window_storage(face, step.faces_u.corner_storage, step.faces_u.period), 0).r;
+    }
+
+    return fluid_parent_face_u(face, step.cells.corner - parent.origin);
+}
+
+fn fluid_wind_publish_face_v(face: vec3<i32>) -> f32 {
+    if (grid_tile_window_contains(face, step.faces_v.size)) {
+        return textureLoad(source_v, grid_tile_window_storage(face, step.faces_v.corner_storage, step.faces_v.period), 0).r;
+    }
+
+    return fluid_parent_face_v(face, step.cells.corner - parent.origin);
+}
+
+fn fluid_wind_publish_face_w(face: vec3<i32>) -> f32 {
+    if (grid_tile_window_contains(face, step.faces_w.size)) {
+        return textureLoad(source_w, grid_tile_window_storage(face, step.faces_w.corner_storage, step.faces_w.period), 0).r;
+    }
+
+    return fluid_parent_face_w(face, step.cells.corner - parent.origin);
+}
+
+@compute @workgroup_size(${FLUID_WIND_PUBLISH_GROUP_SIZE.join(', ')})
+fn main(@builtin(global_invocation_id) id: vec3<u32>) {
+    // the window's cells and a skirt on every side
+    let tile_size = step.cells.size + vec3<u32>(2u * FLUID_WIND_PUBLISH_BORDER);
+
+    if (any(id >= tile_size)) {
+        return;
+    }
+
+    // tile texel t holds window-local cell t - border
+    let cell = vec3<i32>(id) - vec3<i32>(i32(FLUID_WIND_PUBLISH_BORDER));
+    var deviation: vec3<f32>;
+
+    if (grid_tile_window_contains(cell, step.cells.size)) {
+        let su0 = grid_tile_window_storage(cell, step.faces_u.corner_storage, step.faces_u.period);
+        let su1_x = select(su0.x + 1u, 0u, su0.x + 1u == step.faces_u.period.x);
+        let sv0 = grid_tile_window_storage(cell, step.faces_v.corner_storage, step.faces_v.period);
+        let sv1_y = select(sv0.y + 1u, 0u, sv0.y + 1u == step.faces_v.period.y);
+        let sw0 = grid_tile_window_storage(cell, step.faces_w.corner_storage, step.faces_w.period);
+        let sw1_z = select(sw0.z + 1u, 0u, sw0.z + 1u == step.faces_w.period.z);
+
+        let faces = vec3<f32>(
+            textureLoad(source_u, su0, 0).r + textureLoad(source_u, vec3<u32>(su1_x, su0.y, su0.z), 0).r,
+            textureLoad(source_v, sv0, 0).r + textureLoad(source_v, vec3<u32>(sv0.x, sv1_y, sv0.z), 0).r,
+            textureLoad(source_w, sw0, 0).r + textureLoad(source_w, vec3<u32>(sw0.x, sw0.y, sw1_z), 0).r,
+        );
+
+        deviation = 0.5 * faces - step.far_velocity;
+    } else {
+        deviation = fluid_wind_publish_cell(cell, step.far_velocity);
+    }
+
+    textureStore(tile, publish.tile_origin + id, vec4<f32>(deviation, 0.0));
+}
+`, [
+        chunk_grid_tile_window,
+        chunk_fluid_cascade_parent_velocity,
+        chunk_fluid_wind_publish_cell,
+    ]
+);
+
+/**
+ * A cascade's publish into the wind field (GPU_FLUID_CASCADES_PROPOSAL_2026_10_16.md §11, the publish
+ * half of pass 6): every texel of its atlas tile, the window's cells and the one-texel skirt around
+ * them, gets `chunk_fluid_wind_publish_cell` — the velocity at that cell's centre less the step's far
+ * field — as `rgba16float`, alpha 0.
+ *
+ * The skirt is the cells one past the window. Each takes the window's faces where it has them (along
+ * an axis, the boundary face plane) and, for the rest, what the MAC seams answer beyond the window:
+ * the parent's identity read as of the cascade's read record (`chunk_fluid_cascade_parent_velocity`),
+ * which for the coarsest is the far field. Beside the far field, a skirt cell beside a face of the
+ * window takes half that boundary face's deviation on the face's axis and none on the others, and one
+ * past an edge or a corner none at all.
+ *
+ * It writes its own tile and never samples the atlas: the atlas is bound as a write-only storage
+ * texture and nothing else.
+ *
+ * One invocation a texel, in 4 × 4 × 4 workgroups over the tile, dispatched
+ * `ceil((N + 2) / 4)` on each axis; each reads its cell's six faces (`textureLoad`) and stores one
+ * texel. Standalone until it is fused into the gradient pass (proposal §8 (3)), where the faces are in
+ * workgroup memory already.
+ *
+ * Bindings:
+ * - `step`: {@link FLUID_MAC_STEP_STRUCT}: the cascade's windows and `far_velocity`, `v_far(t_s,c)`;
+ *   `cells.size` is the window, and the tile's content;
+ * - `publish`: {@link FLUID_WIND_PUBLISH_STRUCT}: the tile's texel origin in the atlas;
+ * - `parent`: {@link FLUID_CASCADE_PARENT_STRUCT}, and `parent_u`, `parent_v`, `parent_w` the parent's
+ *   persistent faces, read for the skirt;
+ * - `source_u`, `source_v`, `source_w`: the cascade's persistent velocity after its step, `r32float`
+ *   3-D textures, toroidal through `step.faces_*`;
+ * - `tile`: written, the atlas, `rgba16float` 3-D storage texture.
+ *
+ * @type {ComputeShader}
+ */
+export const shader_fluid_wind_publish = ComputeShader.from({
+    label: "Fluid / wind publish",
+    body,
+    resources,
+});