@aardworx/wombat.rendering 0.9.11 → 0.9.13

package/src/runtime/hybridScene.ts

@@ -225,7 +225,7 @@ export function compileHybridScene(
     fragmentOutputLayout,
     atlasPool,
     ...(opts.enableFamilyMerge === true ? { enableFamilyMerge: true } : {}),
-    ...(opts.enableDerivedUniforms === true ? { enableDerivedUniforms: true } : {}),
+    ...(opts.enableDerivedUniforms === false ? { enableDerivedUniforms: false } : {}),
   });
 
   // ─── Legacy subset → RenderTree → ScenePass ──────────────────────

package/src/runtime/index.ts

@@ -43,6 +43,14 @@ export { flattenRenderTree } from "./flattenTree.js";
 export { isHeapEligible } from "./heapEligibility.js";
 export { renderObjectToHeapSpec } from "./heapAdapter.js";
 
+// ─── Derived uniforms (§7): a uniform binding is a value (aval/constant) OR a rule ──
+export {
+  derivedUniform, DerivedExpr, type DerivedScope,
+  ruleFromIR, uniformRef, isDerivedRule, hashIR,
+  type DerivedRule, type IRFragment,
+  STANDARD_DERIVED_RULES, isStandardDerivedName,
+} from "./derivedUniforms/index.js";
+
 export {
   compileHybridScene,
   type CompileHybridSceneOptions,

package/src/runtime/renderTask.ts

@@ -255,7 +255,7 @@ class RenderTask implements IRenderTask {
         compileEffect: this.ctx.compileEffect,
         ...(this.ctx.heapEnabled !== undefined ? { heapEnabled: this.ctx.heapEnabled } : {}),
         ...(this.ctx.enableFamilyMerge === true ? { enableFamilyMerge: true } : {}),
-        ...(this.ctx.enableDerivedUniforms === true ? { enableDerivedUniforms: true } : {}),
+        ...(this.ctx.enableDerivedUniforms === false ? { enableDerivedUniforms: false } : {}),
       });
       this._scenes.set(cmd, s);
     }

package/src/runtime/runtime.ts

@@ -76,7 +76,7 @@ export class Runtime {
       })),
       ...(opts.heapEnabled !== undefined ? { heapEnabled: opts.heapEnabled } : {}),
       ...(opts.enableFamilyMerge === true ? { enableFamilyMerge: true } : {}),
-      ...(opts.enableDerivedUniforms === true ? { enableDerivedUniforms: true } : {}),
+      ...(opts.enableDerivedUniforms === false ? { enableDerivedUniforms: false } : {}),
     };
     // `device.lost` is a real-WebGPU promise; mock devices may not
     // expose it. Treat as "never lost" in that case.

package/dist/runtime/derivedUniforms/slotId.d.ts

@@ -1,15 +0,0 @@
-/** Branded u32 — top bit selects heap, low 31 bits are slot index. */
-export type SlotId = number & {
-    readonly __slotId: unique symbol;
-};
-export declare const SlotId: {
-    /** Encode a Constituents-heap slot. */
-    constituent(idx: number): SlotId;
-    /** Encode an Intermediates-heap slot. */
-    intermediate(idx: number): SlotId;
-    isIntermediate(s: SlotId): boolean;
-    index(s: SlotId): number;
-};
-/** Bytes per df32 mat4 slot: 16 entries × 2 floats × 4 bytes. */
-export declare const DF32_MAT4_BYTES = 128;
-//# sourceMappingURL=slotId.d.ts.map

package/dist/runtime/derivedUniforms/slotId.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"slotId.d.ts","sourceRoot":"","sources":["../../../src/runtime/derivedUniforms/slotId.ts"],"names":[],"mappings":"AAWA,sEAAsE;AACtE,MAAM,MAAM,MAAM,GAAG,MAAM,GAAG;IAAE,QAAQ,CAAC,QAAQ,EAAE,OAAO,MAAM,CAAA;CAAE,CAAC;AAInE,eAAO,MAAM,MAAM;IACjB,uCAAuC;qBACtB,MAAM,GAAG,MAAM;IAGhC,yCAAyC;sBACvB,MAAM,GAAG,MAAM;sBAGf,MAAM,GAAG,OAAO;aAGzB,MAAM,GAAG,MAAM;CAGzB,CAAC;AAEF,iEAAiE;AACjE,eAAO,MAAM,eAAe,MAAM,CAAC"}

package/dist/runtime/derivedUniforms/slotId.js

@@ -1,30 +0,0 @@
-// Slot-ID encoding for the §7 uber kernel.
-//
-// One u32 namespace covers both the Constituents heap (df32 trafos
-// uploaded from CPU) and the Intermediates heap (df32 outputs of
-// layer-1 recipes). The high bit selects the heap; the low 31 bits are
-// the linear index within that heap. Each slot holds one mat4 in df32
-// = 16 vec2<f32> = 128 bytes.
-//
-// Uber-kernel side: a single helper reads from either heap based on
-// the flag — see `read_mat4_df` in the WGSL source.
-const FLAG_INTERMEDIATE = 0x80000000;
-export const SlotId = {
-    /** Encode a Constituents-heap slot. */
-    constituent(idx) {
-        return (idx & 0x7FFFFFFF);
-    },
-    /** Encode an Intermediates-heap slot. */
-    intermediate(idx) {
-        return ((idx & 0x7FFFFFFF) | FLAG_INTERMEDIATE) >>> 0;
-    },
-    isIntermediate(s) {
-        return (s & FLAG_INTERMEDIATE) !== 0;
-    },
-    index(s) {
-        return s & 0x7FFFFFFF;
-    },
-};
-/** Bytes per df32 mat4 slot: 16 entries × 2 floats × 4 bytes. */
-export const DF32_MAT4_BYTES = 128;
-//# sourceMappingURL=slotId.js.map

package/dist/runtime/derivedUniforms/slotId.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"slotId.js","sourceRoot":"","sources":["../../../src/runtime/derivedUniforms/slotId.ts"],"names":[],"mappings":"AAAA,2CAA2C;AAC3C,EAAE;AACF,mEAAmE;AACnE,iEAAiE;AACjE,uEAAuE;AACvE,sEAAsE;AACtE,8BAA8B;AAC9B,EAAE;AACF,oEAAoE;AACpE,oDAAoD;AAKpD,MAAM,iBAAiB,GAAG,UAAU,CAAC;AAErC,MAAM,CAAC,MAAM,MAAM,GAAG;IACpB,uCAAuC;IACvC,WAAW,CAAC,GAAW;QACrB,OAAO,CAAC,GAAG,GAAG,UAAU,CAAW,CAAC;IACtC,CAAC;IACD,yCAAyC;IACzC,YAAY,CAAC,GAAW;QACtB,OAAO,CAAC,CAAC,GAAG,GAAG,UAAU,CAAC,GAAG,iBAAiB,CAAC,KAAK,CAAW,CAAC;IAClE,CAAC;IACD,cAAc,CAAC,CAAS;QACtB,OAAO,CAAC,CAAC,GAAG,iBAAiB,CAAC,KAAK,CAAC,CAAC;IACvC,CAAC;IACD,KAAK,CAAC,CAAS;QACb,OAAO,CAAC,GAAG,UAAU,CAAC;IACxB,CAAC;CACF,CAAC;AAEF,iEAAiE;AACjE,MAAM,CAAC,MAAM,eAAe,GAAG,GAAG,CAAC"}

package/dist/runtime/derivedUniforms/uberKernel.wgsl.d.ts

@@ -1,13 +0,0 @@
-export declare const DERIVED_UBER_KERNEL_WGSL = "\n\n// \u2500\u2500\u2500 Bindings \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n@group(0) @binding(0) var<storage, read>       Constituents: array<vec2<f32>>;\n@group(0) @binding(1) var<storage, read_write> MainHeap:     array<f32>;\n\nstruct Record {\n  recipe_id: u32,\n  in0:       u32,    // constituent slot index\n  in1:       u32,    // constituent slot index, 0 if unused\n  in2:       u32,    // constituent slot index, 0 if unused\n  out_byte:  u32,    // byte offset in MainHeap\n  _pad:      u32,    // bumps stride to 24 (defensively 16-aligned)\n}\n@group(0) @binding(2) var<storage, read>       Records: array<Record>;\n\n// Live record count. arrayLength(Records) reports the GPU buffer's\n// capacity (rounded up by allocator growth), not the live count \u2014\n// trailing slots hold stale data from swap-removed records. Threads\n// past Count.count would read stale records and trample random\n// MainHeap bytes.\nstruct CountUniform { count: u32 }\n@group(0) @binding(3) var<uniform>             Count: CountUniform;\n\n// \u2500\u2500\u2500 df32 primitives \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn split12(a: f32) -> vec2<f32> {\n  let hi = bitcast<f32>(bitcast<u32>(a) & 0xFFFFE000u);\n  let lo = a - hi;\n  return vec2<f32>(hi, lo);\n}\n\nfn two_sum(a: f32, b: f32) -> vec2<f32> {\n  let s  = a + b;\n  let bb = fma(1.0, s, -a);\n  let t1 = fma(1.0, s, -bb);\n  let t2 = fma(1.0, a, -t1);\n  let t3 = fma(1.0, b, -bb);\n  return vec2<f32>(s, t2 + t3);\n}\n\nfn quick_two_sum(a: f32, b: f32) -> vec2<f32> {\n  let s = a + b;\n  let t = fma(1.0, s, -a);\n  return vec2<f32>(s, fma(1.0, b, -t));\n}\n\nfn two_prod(a: f32, b: f32) -> vec2<f32> {\n  let p = a * b;\n  let A = split12(a);\n  let B = split12(b);\n  let err = ((A.x * B.x - p) + A.x * B.y + A.y * B.x) + A.y * B.y;\n  return vec2<f32>(p, err);\n}\n\nfn df_add(a: vec2<f32>, b: vec2<f32>) -> vec2<f32> {\n  let s = two_sum(a.x, b.x);\n  let t = two_sum(a.y, b.y);\n  let s3 = quick_two_sum(s.x, s.y + t.x);\n  return quick_two_sum(s3.x, s3.y + t.y);\n}\n\nfn df_mul(a: vec2<f32>, b: vec2<f32>) -> vec2<f32> {\n  let p      = two_prod(a.x, b.x);\n  let cross1 = fma(a.x, b.y, p.y);\n  let cross  = fma(a.y, b.x, cross1);\n  return quick_two_sum(p.x, cross);\n}\n\n// \u2500\u2500\u2500 Constituents access \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n//\n// One slot = 16 vec2<f32> = one df32 mat4 in row-major order\n// (matching Aardvark M44d convention). Element (row, col) of slot k\n// lives at vec2 index k*16 + row*4 + col.\n\nfn read_entry(slot: u32, row: u32, col: u32) -> vec2<f32> {\n  return Constituents[slot * 16u + row * 4u + col];\n}\n\n// (A \u00B7 B) entry (r, c) in df32, both A and B in constituents.\nfn df_mul_entry(a_slot: u32, b_slot: u32, r: u32, c: u32) -> vec2<f32> {\n  var acc = vec2<f32>(0.0, 0.0);\n  for (var k: u32 = 0u; k < 4u; k = k + 1u) {\n    acc = df_add(acc, df_mul(read_entry(a_slot, r, k), read_entry(b_slot, k, c)));\n  }\n  return acc;\n}\n\n// Compute one full row of (A \u00B7 B), returns 4 df32 entries.\nfn df_mul_row(a_slot: u32, b_slot: u32, r: u32) -> array<vec2<f32>, 4> {\n  var out: array<vec2<f32>, 4>;\n  for (var c: u32 = 0u; c < 4u; c = c + 1u) {\n    out[c] = df_mul_entry(a_slot, b_slot, r, c);\n  }\n  return out;\n}\n\n// Triple product (A \u00B7 B \u00B7 C) entry (r, c). Uses 4 df_muls per entry\n// \u2014 recomputes the inner row each call. Acceptable: at most one\n// such recipe per RO and only on dirty frames.\nfn df_mul3_entry(a_slot: u32, b_slot: u32, c_slot: u32, r: u32, c: u32) -> vec2<f32> {\n  // (A\u00B7B\u00B7C)[r,c] = sum_k (A\u00B7B)[r,k] \u00B7 C[k,c]\n  var acc = vec2<f32>(0.0, 0.0);\n  for (var k: u32 = 0u; k < 4u; k = k + 1u) {\n    let ab_rk = df_mul_entry(a_slot, b_slot, r, k);\n    acc = df_add(acc, df_mul(ab_rk, read_entry(c_slot, k, c)));\n  }\n  return acc;\n}\n\n// \u2500\u2500\u2500 Writers \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n//\n// Mat4: 16 contiguous f32s, row-major.\n// Mat3: 3 rows \u00D7 4-stride (std140 mat3<f32>), 12 f32s; last column\n//       per row left untouched.\n\nfn write_mat4_entry(out_byte: u32, row: u32, col: u32, val: f32) {\n  MainHeap[(out_byte >> 2u) + row * 4u + col] = val;\n}\n\nfn write_mat3_entry(out_byte: u32, row: u32, col: u32, val: f32) {\n  MainHeap[(out_byte >> 2u) + row * 4u + col] = val;\n}\n\n// \u2500\u2500\u2500 Recipe arms \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\nfn run_collapse_mat4(in0: u32, out_byte: u32) {\n  for (var r: u32 = 0u; r < 4u; r = r + 1u) {\n    for (var c: u32 = 0u; c < 4u; c = c + 1u) {\n      let e = read_entry(in0, r, c);\n      write_mat4_entry(out_byte, r, c, e.x + e.y);\n    }\n  }\n}\n\n// NormalMatrix: (M\u207B\u00B9)\u1D40 upper-3\u00D73.\n//   NM[i, j] = MInv[j, i]\nfn run_collapse_mat3_normal(in0: u32, out_byte: u32) {\n  for (var i: u32 = 0u; i < 3u; i = i + 1u) {\n    for (var j: u32 = 0u; j < 3u; j = j + 1u) {\n      let e = read_entry(in0, j, i);\n      write_mat3_entry(out_byte, i, j, e.x + e.y);\n    }\n  }\n}\n\nfn run_dfmul2_collapse(in0: u32, in1: u32, out_byte: u32) {\n  for (var r: u32 = 0u; r < 4u; r = r + 1u) {\n    for (var c: u32 = 0u; c < 4u; c = c + 1u) {\n      let e = df_mul_entry(in0, in1, r, c);\n      write_mat4_entry(out_byte, r, c, e.x + e.y);\n    }\n  }\n}\n\nfn run_dfmul3_collapse(in0: u32, in1: u32, in2: u32, out_byte: u32) {\n  for (var r: u32 = 0u; r < 4u; r = r + 1u) {\n    for (var c: u32 = 0u; c < 4u; c = c + 1u) {\n      let e = df_mul3_entry(in0, in1, in2, r, c);\n      write_mat4_entry(out_byte, r, c, e.x + e.y);\n    }\n  }\n}\n\n// \u2500\u2500\u2500 Entry \u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\n@compute @workgroup_size(64)\nfn main(@builtin(global_invocation_id) gid: vec3<u32>) {\n  if (gid.x >= Count.count) { return; }\n  let r = Records[gid.x];\n  switch r.recipe_id {\n    case 0u: { run_collapse_mat4(r.in0, r.out_byte); }            // ModelTrafo\n    case 1u: { run_collapse_mat4(r.in0, r.out_byte); }            // ModelTrafoInv\n    case 2u: { run_collapse_mat3_normal(r.in0, r.out_byte); }     // NormalMatrix\n    case 3u: { run_dfmul2_collapse(r.in0, r.in1, r.out_byte); }   // ModelViewTrafo\n    case 4u: { run_dfmul2_collapse(r.in0, r.in1, r.out_byte); }   // ModelViewTrafoInv\n    case 5u: { run_dfmul3_collapse(r.in0, r.in1, r.in2, r.out_byte); } // ModelViewProjTrafo\n    case 6u: { run_dfmul3_collapse(r.in0, r.in1, r.in2, r.out_byte); } // ModelViewProjTrafoInv\n    case 7u: { run_collapse_mat4(r.in0, r.out_byte); }            // ViewTrafo\n    case 8u: { run_collapse_mat4(r.in0, r.out_byte); }            // ViewTrafoInv\n    case 9u: { run_dfmul2_collapse(r.in0, r.in1, r.out_byte); }   // ViewProjTrafo\n    case 10u:{ run_dfmul2_collapse(r.in0, r.in1, r.out_byte); }   // ViewProjTrafoInv\n    case 11u:{ run_collapse_mat4(r.in0, r.out_byte); }            // ProjTrafo\n    case 12u:{ run_collapse_mat4(r.in0, r.out_byte); }            // ProjTrafoInv\n    default: { return; }\n  }\n}\n\n";
-/** Records are 5× u32 = 20 bytes. WGSL storage struct alignment of
- *  4 means the array stride is 20 bytes. */
-export declare const RECORD_STRIDE_BYTES = 24;
-export declare const RECORD_STRIDE_U32 = 6;
-export declare const RECORD_FIELD: {
-    readonly recipe_id: 0;
-    readonly in0: 1;
-    readonly in1: 2;
-    readonly in2: 3;
-    readonly out_byte: 4;
-};
-//# sourceMappingURL=uberKernel.wgsl.d.ts.map

package/dist/runtime/derivedUniforms/uberKernel.wgsl.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"uberKernel.wgsl.d.ts","sourceRoot":"","sources":["../../../src/runtime/derivedUniforms/uberKernel.wgsl.ts"],"names":[],"mappings":"AAkBA,eAAO,MAAM,wBAAwB,y0QA4LpC,CAAC;AAEF;4CAC4C;AAC5C,eAAO,MAAM,mBAAmB,KAAK,CAAC;AACtC,eAAO,MAAM,iBAAiB,IAAI,CAAC;AAEnC,eAAO,MAAM,YAAY;;;;;;CAMf,CAAC"}

package/dist/runtime/derivedUniforms/uberKernel.wgsl.js

@@ -1,218 +0,0 @@
-// §7 derived-uniforms uber kernel — single flat dispatch over leaves.
-//
-// Records are static-ish (only edited on RO add/remove). Each thread
-// processes one record:
-//   1. Load record (recipe_id, in0, in1, in2, out_byte).
-//   2. Switch on recipe_id, compute inline (df_mul + collapse), write
-//      f32 mat4/mat3 to MainHeap at out_byte.
-//
-// No dirty test on the kernel: when constituents don't change, the
-// kernel rewrites identical bits. Saves a CPU writeBuffer + a binding
-// per dispatch. GPU does ~64 FLOPs per record × N records — trivial at
-// the scales this runs at.
-//
-// df32 primitives mirror examples/df32-precision/. Two opacity hooks:
-//   - bitcast<f32>(bits & MASK) for Veltkamp split (integer ops, no
-//     algebraic identity).
-//   - fma(1.0, x, -y) for compensated subtractions.
-export const DERIVED_UBER_KERNEL_WGSL = /* wgsl */ `
-
-// ─── Bindings ─────────────────────────────────────────────────────
-@group(0) @binding(0) var<storage, read>       Constituents: array<vec2<f32>>;
-@group(0) @binding(1) var<storage, read_write> MainHeap:     array<f32>;
-
-struct Record {
-  recipe_id: u32,
-  in0:       u32,    // constituent slot index
-  in1:       u32,    // constituent slot index, 0 if unused
-  in2:       u32,    // constituent slot index, 0 if unused
-  out_byte:  u32,    // byte offset in MainHeap
-  _pad:      u32,    // bumps stride to 24 (defensively 16-aligned)
-}
-@group(0) @binding(2) var<storage, read>       Records: array<Record>;
-
-// Live record count. arrayLength(Records) reports the GPU buffer's
-// capacity (rounded up by allocator growth), not the live count —
-// trailing slots hold stale data from swap-removed records. Threads
-// past Count.count would read stale records and trample random
-// MainHeap bytes.
-struct CountUniform { count: u32 }
-@group(0) @binding(3) var<uniform>             Count: CountUniform;
-
-// ─── df32 primitives ──────────────────────────────────────────────
-
-fn split12(a: f32) -> vec2<f32> {
-  let hi = bitcast<f32>(bitcast<u32>(a) & 0xFFFFE000u);
-  let lo = a - hi;
-  return vec2<f32>(hi, lo);
-}
-
-fn two_sum(a: f32, b: f32) -> vec2<f32> {
-  let s  = a + b;
-  let bb = fma(1.0, s, -a);
-  let t1 = fma(1.0, s, -bb);
-  let t2 = fma(1.0, a, -t1);
-  let t3 = fma(1.0, b, -bb);
-  return vec2<f32>(s, t2 + t3);
-}
-
-fn quick_two_sum(a: f32, b: f32) -> vec2<f32> {
-  let s = a + b;
-  let t = fma(1.0, s, -a);
-  return vec2<f32>(s, fma(1.0, b, -t));
-}
-
-fn two_prod(a: f32, b: f32) -> vec2<f32> {
-  let p = a * b;
-  let A = split12(a);
-  let B = split12(b);
-  let err = ((A.x * B.x - p) + A.x * B.y + A.y * B.x) + A.y * B.y;
-  return vec2<f32>(p, err);
-}
-
-fn df_add(a: vec2<f32>, b: vec2<f32>) -> vec2<f32> {
-  let s = two_sum(a.x, b.x);
-  let t = two_sum(a.y, b.y);
-  let s3 = quick_two_sum(s.x, s.y + t.x);
-  return quick_two_sum(s3.x, s3.y + t.y);
-}
-
-fn df_mul(a: vec2<f32>, b: vec2<f32>) -> vec2<f32> {
-  let p      = two_prod(a.x, b.x);
-  let cross1 = fma(a.x, b.y, p.y);
-  let cross  = fma(a.y, b.x, cross1);
-  return quick_two_sum(p.x, cross);
-}
-
-// ─── Constituents access ──────────────────────────────────────────
-//
-// One slot = 16 vec2<f32> = one df32 mat4 in row-major order
-// (matching Aardvark M44d convention). Element (row, col) of slot k
-// lives at vec2 index k*16 + row*4 + col.
-
-fn read_entry(slot: u32, row: u32, col: u32) -> vec2<f32> {
-  return Constituents[slot * 16u + row * 4u + col];
-}
-
-// (A · B) entry (r, c) in df32, both A and B in constituents.
-fn df_mul_entry(a_slot: u32, b_slot: u32, r: u32, c: u32) -> vec2<f32> {
-  var acc = vec2<f32>(0.0, 0.0);
-  for (var k: u32 = 0u; k < 4u; k = k + 1u) {
-    acc = df_add(acc, df_mul(read_entry(a_slot, r, k), read_entry(b_slot, k, c)));
-  }
-  return acc;
-}
-
-// Compute one full row of (A · B), returns 4 df32 entries.
-fn df_mul_row(a_slot: u32, b_slot: u32, r: u32) -> array<vec2<f32>, 4> {
-  var out: array<vec2<f32>, 4>;
-  for (var c: u32 = 0u; c < 4u; c = c + 1u) {
-    out[c] = df_mul_entry(a_slot, b_slot, r, c);
-  }
-  return out;
-}
-
-// Triple product (A · B · C) entry (r, c). Uses 4 df_muls per entry
-// — recomputes the inner row each call. Acceptable: at most one
-// such recipe per RO and only on dirty frames.
-fn df_mul3_entry(a_slot: u32, b_slot: u32, c_slot: u32, r: u32, c: u32) -> vec2<f32> {
-  // (A·B·C)[r,c] = sum_k (A·B)[r,k] · C[k,c]
-  var acc = vec2<f32>(0.0, 0.0);
-  for (var k: u32 = 0u; k < 4u; k = k + 1u) {
-    let ab_rk = df_mul_entry(a_slot, b_slot, r, k);
-    acc = df_add(acc, df_mul(ab_rk, read_entry(c_slot, k, c)));
-  }
-  return acc;
-}
-
-// ─── Writers ──────────────────────────────────────────────────────
-//
-// Mat4: 16 contiguous f32s, row-major.
-// Mat3: 3 rows × 4-stride (std140 mat3<f32>), 12 f32s; last column
-//       per row left untouched.
-
-fn write_mat4_entry(out_byte: u32, row: u32, col: u32, val: f32) {
-  MainHeap[(out_byte >> 2u) + row * 4u + col] = val;
-}
-
-fn write_mat3_entry(out_byte: u32, row: u32, col: u32, val: f32) {
-  MainHeap[(out_byte >> 2u) + row * 4u + col] = val;
-}
-
-// ─── Recipe arms ──────────────────────────────────────────────────
-
-fn run_collapse_mat4(in0: u32, out_byte: u32) {
-  for (var r: u32 = 0u; r < 4u; r = r + 1u) {
-    for (var c: u32 = 0u; c < 4u; c = c + 1u) {
-      let e = read_entry(in0, r, c);
-      write_mat4_entry(out_byte, r, c, e.x + e.y);
-    }
-  }
-}
-
-// NormalMatrix: (M⁻¹)ᵀ upper-3×3.
-//   NM[i, j] = MInv[j, i]
-fn run_collapse_mat3_normal(in0: u32, out_byte: u32) {
-  for (var i: u32 = 0u; i < 3u; i = i + 1u) {
-    for (var j: u32 = 0u; j < 3u; j = j + 1u) {
-      let e = read_entry(in0, j, i);
-      write_mat3_entry(out_byte, i, j, e.x + e.y);
-    }
-  }
-}
-
-fn run_dfmul2_collapse(in0: u32, in1: u32, out_byte: u32) {
-  for (var r: u32 = 0u; r < 4u; r = r + 1u) {
-    for (var c: u32 = 0u; c < 4u; c = c + 1u) {
-      let e = df_mul_entry(in0, in1, r, c);
-      write_mat4_entry(out_byte, r, c, e.x + e.y);
-    }
-  }
-}
-
-fn run_dfmul3_collapse(in0: u32, in1: u32, in2: u32, out_byte: u32) {
-  for (var r: u32 = 0u; r < 4u; r = r + 1u) {
-    for (var c: u32 = 0u; c < 4u; c = c + 1u) {
-      let e = df_mul3_entry(in0, in1, in2, r, c);
-      write_mat4_entry(out_byte, r, c, e.x + e.y);
-    }
-  }
-}
-
-// ─── Entry ────────────────────────────────────────────────────────
-
-@compute @workgroup_size(64)
-fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-  if (gid.x >= Count.count) { return; }
-  let r = Records[gid.x];
-  switch r.recipe_id {
-    case 0u: { run_collapse_mat4(r.in0, r.out_byte); }            // ModelTrafo
-    case 1u: { run_collapse_mat4(r.in0, r.out_byte); }            // ModelTrafoInv
-    case 2u: { run_collapse_mat3_normal(r.in0, r.out_byte); }     // NormalMatrix
-    case 3u: { run_dfmul2_collapse(r.in0, r.in1, r.out_byte); }   // ModelViewTrafo
-    case 4u: { run_dfmul2_collapse(r.in0, r.in1, r.out_byte); }   // ModelViewTrafoInv
-    case 5u: { run_dfmul3_collapse(r.in0, r.in1, r.in2, r.out_byte); } // ModelViewProjTrafo
-    case 6u: { run_dfmul3_collapse(r.in0, r.in1, r.in2, r.out_byte); } // ModelViewProjTrafoInv
-    case 7u: { run_collapse_mat4(r.in0, r.out_byte); }            // ViewTrafo
-    case 8u: { run_collapse_mat4(r.in0, r.out_byte); }            // ViewTrafoInv
-    case 9u: { run_dfmul2_collapse(r.in0, r.in1, r.out_byte); }   // ViewProjTrafo
-    case 10u:{ run_dfmul2_collapse(r.in0, r.in1, r.out_byte); }   // ViewProjTrafoInv
-    case 11u:{ run_collapse_mat4(r.in0, r.out_byte); }            // ProjTrafo
-    case 12u:{ run_collapse_mat4(r.in0, r.out_byte); }            // ProjTrafoInv
-    default: { return; }
-  }
-}
-
-`;
-/** Records are 5× u32 = 20 bytes. WGSL storage struct alignment of
- *  4 means the array stride is 20 bytes. */
-export const RECORD_STRIDE_BYTES = 24;
-export const RECORD_STRIDE_U32 = 6;
-export const RECORD_FIELD = {
-    recipe_id: 0,
-    in0: 1,
-    in1: 2,
-    in2: 3,
-    out_byte: 4,
-};
-//# sourceMappingURL=uberKernel.wgsl.js.map

package/dist/runtime/derivedUniforms/uberKernel.wgsl.js.map

@@ -1 +0,0 @@
-{"version":3,"file":"uberKernel.wgsl.js","sourceRoot":"","sources":["../../../src/runtime/derivedUniforms/uberKernel.wgsl.ts"],"names":[],"mappings":"AAAA,sEAAsE;AACtE,EAAE;AACF,qEAAqE;AACrE,wBAAwB;AACxB,yDAAyD;AACzD,sEAAsE;AACtE,8CAA8C;AAC9C,EAAE;AACF,mEAAmE;AACnE,sEAAsE;AACtE,uEAAuE;AACvE,2BAA2B;AAC3B,EAAE;AACF,sEAAsE;AACtE,oEAAoE;AACpE,2BAA2B;AAC3B,oDAAoD;AAEpD,MAAM,CAAC,MAAM,wBAAwB,GAAG,UAAU,CAAC;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;CA4LlD,CAAC;AAEF;4CAC4C;AAC5C,MAAM,CAAC,MAAM,mBAAmB,GAAG,EAAE,CAAC;AACtC,MAAM,CAAC,MAAM,iBAAiB,GAAG,CAAC,CAAC;AAEnC,MAAM,CAAC,MAAM,YAAY,GAAG;IAC1B,SAAS,EAAE,CAAC;IACZ,GAAG,EAAQ,CAAC;IACZ,GAAG,EAAQ,CAAC;IACZ,GAAG,EAAQ,CAAC;IACZ,QAAQ,EAAG,CAAC;CACJ,CAAC"}

package/src/runtime/derivedUniforms/slotId.ts

@@ -1,35 +0,0 @@
-// Slot-ID encoding for the §7 uber kernel.
-//
-// One u32 namespace covers both the Constituents heap (df32 trafos
-// uploaded from CPU) and the Intermediates heap (df32 outputs of
-// layer-1 recipes). The high bit selects the heap; the low 31 bits are
-// the linear index within that heap. Each slot holds one mat4 in df32
-// = 16 vec2<f32> = 128 bytes.
-//
-// Uber-kernel side: a single helper reads from either heap based on
-// the flag — see `read_mat4_df` in the WGSL source.
-
-/** Branded u32 — top bit selects heap, low 31 bits are slot index. */
-export type SlotId = number & { readonly __slotId: unique symbol };
-
-const FLAG_INTERMEDIATE = 0x80000000;
-
-export const SlotId = {
-  /** Encode a Constituents-heap slot. */
-  constituent(idx: number): SlotId {
-    return (idx & 0x7FFFFFFF) as SlotId;
-  },
-  /** Encode an Intermediates-heap slot. */
-  intermediate(idx: number): SlotId {
-    return ((idx & 0x7FFFFFFF) | FLAG_INTERMEDIATE) >>> 0 as SlotId;
-  },
-  isIntermediate(s: SlotId): boolean {
-    return (s & FLAG_INTERMEDIATE) !== 0;
-  },
-  index(s: SlotId): number {
-    return s & 0x7FFFFFFF;
-  },
-};
-
-/** Bytes per df32 mat4 slot: 16 entries × 2 floats × 4 bytes. */
-export const DF32_MAT4_BYTES = 128;

package/src/runtime/derivedUniforms/uberKernel.wgsl.ts

@@ -1,220 +0,0 @@
-// §7 derived-uniforms uber kernel — single flat dispatch over leaves.
-//
-// Records are static-ish (only edited on RO add/remove). Each thread
-// processes one record:
-//   1. Load record (recipe_id, in0, in1, in2, out_byte).
-//   2. Switch on recipe_id, compute inline (df_mul + collapse), write
-//      f32 mat4/mat3 to MainHeap at out_byte.
-//
-// No dirty test on the kernel: when constituents don't change, the
-// kernel rewrites identical bits. Saves a CPU writeBuffer + a binding
-// per dispatch. GPU does ~64 FLOPs per record × N records — trivial at
-// the scales this runs at.
-//
-// df32 primitives mirror examples/df32-precision/. Two opacity hooks:
-//   - bitcast<f32>(bits & MASK) for Veltkamp split (integer ops, no
-//     algebraic identity).
-//   - fma(1.0, x, -y) for compensated subtractions.
-
-export const DERIVED_UBER_KERNEL_WGSL = /* wgsl */ `
-
-// ─── Bindings ─────────────────────────────────────────────────────
-@group(0) @binding(0) var<storage, read>       Constituents: array<vec2<f32>>;
-@group(0) @binding(1) var<storage, read_write> MainHeap:     array<f32>;
-
-struct Record {
-  recipe_id: u32,
-  in0:       u32,    // constituent slot index
-  in1:       u32,    // constituent slot index, 0 if unused
-  in2:       u32,    // constituent slot index, 0 if unused
-  out_byte:  u32,    // byte offset in MainHeap
-  _pad:      u32,    // bumps stride to 24 (defensively 16-aligned)
-}
-@group(0) @binding(2) var<storage, read>       Records: array<Record>;
-
-// Live record count. arrayLength(Records) reports the GPU buffer's
-// capacity (rounded up by allocator growth), not the live count —
-// trailing slots hold stale data from swap-removed records. Threads
-// past Count.count would read stale records and trample random
-// MainHeap bytes.
-struct CountUniform { count: u32 }
-@group(0) @binding(3) var<uniform>             Count: CountUniform;
-
-// ─── df32 primitives ──────────────────────────────────────────────
-
-fn split12(a: f32) -> vec2<f32> {
-  let hi = bitcast<f32>(bitcast<u32>(a) & 0xFFFFE000u);
-  let lo = a - hi;
-  return vec2<f32>(hi, lo);
-}
-
-fn two_sum(a: f32, b: f32) -> vec2<f32> {
-  let s  = a + b;
-  let bb = fma(1.0, s, -a);
-  let t1 = fma(1.0, s, -bb);
-  let t2 = fma(1.0, a, -t1);
-  let t3 = fma(1.0, b, -bb);
-  return vec2<f32>(s, t2 + t3);
-}
-
-fn quick_two_sum(a: f32, b: f32) -> vec2<f32> {
-  let s = a + b;
-  let t = fma(1.0, s, -a);
-  return vec2<f32>(s, fma(1.0, b, -t));
-}
-
-fn two_prod(a: f32, b: f32) -> vec2<f32> {
-  let p = a * b;
-  let A = split12(a);
-  let B = split12(b);
-  let err = ((A.x * B.x - p) + A.x * B.y + A.y * B.x) + A.y * B.y;
-  return vec2<f32>(p, err);
-}
-
-fn df_add(a: vec2<f32>, b: vec2<f32>) -> vec2<f32> {
-  let s = two_sum(a.x, b.x);
-  let t = two_sum(a.y, b.y);
-  let s3 = quick_two_sum(s.x, s.y + t.x);
-  return quick_two_sum(s3.x, s3.y + t.y);
-}
-
-fn df_mul(a: vec2<f32>, b: vec2<f32>) -> vec2<f32> {
-  let p      = two_prod(a.x, b.x);
-  let cross1 = fma(a.x, b.y, p.y);
-  let cross  = fma(a.y, b.x, cross1);
-  return quick_two_sum(p.x, cross);
-}
-
-// ─── Constituents access ──────────────────────────────────────────
-//
-// One slot = 16 vec2<f32> = one df32 mat4 in row-major order
-// (matching Aardvark M44d convention). Element (row, col) of slot k
-// lives at vec2 index k*16 + row*4 + col.
-
-fn read_entry(slot: u32, row: u32, col: u32) -> vec2<f32> {
-  return Constituents[slot * 16u + row * 4u + col];
-}
-
-// (A · B) entry (r, c) in df32, both A and B in constituents.
-fn df_mul_entry(a_slot: u32, b_slot: u32, r: u32, c: u32) -> vec2<f32> {
-  var acc = vec2<f32>(0.0, 0.0);
-  for (var k: u32 = 0u; k < 4u; k = k + 1u) {
-    acc = df_add(acc, df_mul(read_entry(a_slot, r, k), read_entry(b_slot, k, c)));
-  }
-  return acc;
-}
-
-// Compute one full row of (A · B), returns 4 df32 entries.
-fn df_mul_row(a_slot: u32, b_slot: u32, r: u32) -> array<vec2<f32>, 4> {
-  var out: array<vec2<f32>, 4>;
-  for (var c: u32 = 0u; c < 4u; c = c + 1u) {
-    out[c] = df_mul_entry(a_slot, b_slot, r, c);
-  }
-  return out;
-}
-
-// Triple product (A · B · C) entry (r, c). Uses 4 df_muls per entry
-// — recomputes the inner row each call. Acceptable: at most one
-// such recipe per RO and only on dirty frames.
-fn df_mul3_entry(a_slot: u32, b_slot: u32, c_slot: u32, r: u32, c: u32) -> vec2<f32> {
-  // (A·B·C)[r,c] = sum_k (A·B)[r,k] · C[k,c]
-  var acc = vec2<f32>(0.0, 0.0);
-  for (var k: u32 = 0u; k < 4u; k = k + 1u) {
-    let ab_rk = df_mul_entry(a_slot, b_slot, r, k);
-    acc = df_add(acc, df_mul(ab_rk, read_entry(c_slot, k, c)));
-  }
-  return acc;
-}
-
-// ─── Writers ──────────────────────────────────────────────────────
-//
-// Mat4: 16 contiguous f32s, row-major.
-// Mat3: 3 rows × 4-stride (std140 mat3<f32>), 12 f32s; last column
-//       per row left untouched.
-
-fn write_mat4_entry(out_byte: u32, row: u32, col: u32, val: f32) {
-  MainHeap[(out_byte >> 2u) + row * 4u + col] = val;
-}
-
-fn write_mat3_entry(out_byte: u32, row: u32, col: u32, val: f32) {
-  MainHeap[(out_byte >> 2u) + row * 4u + col] = val;
-}
-
-// ─── Recipe arms ──────────────────────────────────────────────────
-
-fn run_collapse_mat4(in0: u32, out_byte: u32) {
-  for (var r: u32 = 0u; r < 4u; r = r + 1u) {
-    for (var c: u32 = 0u; c < 4u; c = c + 1u) {
-      let e = read_entry(in0, r, c);
-      write_mat4_entry(out_byte, r, c, e.x + e.y);
-    }
-  }
-}
-
-// NormalMatrix: (M⁻¹)ᵀ upper-3×3.
-//   NM[i, j] = MInv[j, i]
-fn run_collapse_mat3_normal(in0: u32, out_byte: u32) {
-  for (var i: u32 = 0u; i < 3u; i = i + 1u) {
-    for (var j: u32 = 0u; j < 3u; j = j + 1u) {
-      let e = read_entry(in0, j, i);
-      write_mat3_entry(out_byte, i, j, e.x + e.y);
-    }
-  }
-}
-
-fn run_dfmul2_collapse(in0: u32, in1: u32, out_byte: u32) {
-  for (var r: u32 = 0u; r < 4u; r = r + 1u) {
-    for (var c: u32 = 0u; c < 4u; c = c + 1u) {
-      let e = df_mul_entry(in0, in1, r, c);
-      write_mat4_entry(out_byte, r, c, e.x + e.y);
-    }
-  }
-}
-
-fn run_dfmul3_collapse(in0: u32, in1: u32, in2: u32, out_byte: u32) {
-  for (var r: u32 = 0u; r < 4u; r = r + 1u) {
-    for (var c: u32 = 0u; c < 4u; c = c + 1u) {
-      let e = df_mul3_entry(in0, in1, in2, r, c);
-      write_mat4_entry(out_byte, r, c, e.x + e.y);
-    }
-  }
-}
-
-// ─── Entry ────────────────────────────────────────────────────────
-
-@compute @workgroup_size(64)
-fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
-  if (gid.x >= Count.count) { return; }
-  let r = Records[gid.x];
-  switch r.recipe_id {
-    case 0u: { run_collapse_mat4(r.in0, r.out_byte); }            // ModelTrafo
-    case 1u: { run_collapse_mat4(r.in0, r.out_byte); }            // ModelTrafoInv
-    case 2u: { run_collapse_mat3_normal(r.in0, r.out_byte); }     // NormalMatrix
-    case 3u: { run_dfmul2_collapse(r.in0, r.in1, r.out_byte); }   // ModelViewTrafo
-    case 4u: { run_dfmul2_collapse(r.in0, r.in1, r.out_byte); }   // ModelViewTrafoInv
-    case 5u: { run_dfmul3_collapse(r.in0, r.in1, r.in2, r.out_byte); } // ModelViewProjTrafo
-    case 6u: { run_dfmul3_collapse(r.in0, r.in1, r.in2, r.out_byte); } // ModelViewProjTrafoInv
-    case 7u: { run_collapse_mat4(r.in0, r.out_byte); }            // ViewTrafo
-    case 8u: { run_collapse_mat4(r.in0, r.out_byte); }            // ViewTrafoInv
-    case 9u: { run_dfmul2_collapse(r.in0, r.in1, r.out_byte); }   // ViewProjTrafo
-    case 10u:{ run_dfmul2_collapse(r.in0, r.in1, r.out_byte); }   // ViewProjTrafoInv
-    case 11u:{ run_collapse_mat4(r.in0, r.out_byte); }            // ProjTrafo
-    case 12u:{ run_collapse_mat4(r.in0, r.out_byte); }            // ProjTrafoInv
-    default: { return; }
-  }
-}
-
-`;
-
-/** Records are 5× u32 = 20 bytes. WGSL storage struct alignment of
- *  4 means the array stride is 20 bytes. */
-export const RECORD_STRIDE_BYTES = 24;
-export const RECORD_STRIDE_U32 = 6;
-
-export const RECORD_FIELD = {
-  recipe_id: 0,
-  in0:       1,
-  in1:       2,
-  in2:       3,
-  out_byte:  4,
-} as const;