@woosh/meep-engine 3.60.9 → 3.61.1

package/src/shade/renderer/shader/chunk/subgroup/chunk_subgroup_bitset_or.js

@@ -0,0 +1,73 @@
+import { UINT32_MAX } from "../../../../../core/binary/UINT32_MAX.js";
+import { WebGPUExtensionType } from "../../../../descriptor/WebGPUExtensionType.js";
+import { CodeChunk } from "../../compiler/CodeChunk.js";
+
+/**
+ * The index {@link chunk_subgroup_bitset_or} skips.
+ *
+ * @type {number}
+ */
+export const SUBGROUP_BITSET_OR_NONE = UINT32_MAX;
+
+/**
+ * Set bit `index` of a bitset, one atomic per distinct word the subgroup holds: the lanes holding
+ * the first outstanding word take a round, OR together the bits they hold in it — or skip that when
+ * they all hold the same bit — one of them sets them with one `atomicOr`, and the rounds repeat
+ * until no word is left.
+ *
+ * `fn subgroup_bitset_or(bits: ptr<storage, array<atomic<u32>>, read_write>, index: u32)`
+ *
+ * Bit `index` is bit `index & 31` of word `index >> 5`. An index of {@link SUBGROUP_BITSET_OR_NONE},
+ * or one whose word is past the end of `bits`, sets nothing. Every lane of the subgroup calls it,
+ * from subgroup-uniform control flow.
+ *
+ * The atomic is made by a lane of the round, so a lane that holds nothing never stands in for one
+ * that does — in a fragment shader, a helper invocation given {@link SUBGROUP_BITSET_OR_NONE} can
+ * never be the one that writes, and its write would be discarded.
+ *
+ * A subgroup runs one round per distinct word it holds, so the loop is bounded by its lane count.
+ * Indices that follow the screen — a material per pixel — come in runs, and a subgroup ordinarily
+ * holds one or two.
+ *
+ * @see chunk_subgroup_keyed_append for the counting twin
+ * @type {CodeChunk}
+ */
+export const chunk_subgroup_bitset_or = CodeChunk.from(
+    //language=WGSL
+    `
+// every lane runs the loop until no lane holds a word; the rounds run on the lanes holding one
+@diagnostic(off, subgroup_uniformity)
+fn subgroup_bitset_or(bits: ptr<storage, array<atomic<u32>>, read_write>, index: u32) {
+    const NONE = ${SUBGROUP_BITSET_OR_NONE}u;
+
+    var outstanding = select(index >> 5u, NONE, index == NONE);
+    let bit = 1u << (index & 31u);
+
+    loop {
+        if (!subgroupAny(outstanding != NONE)) {
+            break;
+        }
+
+        if (outstanding != NONE) {
+            let round_word = subgroupBroadcastFirst(outstanding);
+
+            if (outstanding == round_word) {
+                var round_bits = bit;
+
+                // a run of one index needs no reduction
+                if (!subgroupAll(bit == subgroupBroadcastFirst(bit))) {
+                    round_bits = subgroupOr(bit);
+                }
+
+                if (subgroupElect() && round_word < arrayLength(bits)) {
+                    atomicOr(&(*bits)[round_word], round_bits);
+                }
+
+                outstanding = NONE;
+            }
+        }
+    }
+}
+`);
+
+chunk_subgroup_bitset_or.addExtension(WebGPUExtensionType.Subgroups);

package/src/shade/wgsl/emulator/WGSLJavaScriptCompiler.js

@@ -3232,7 +3232,14 @@ function inferBuiltinReturnType(name, args) {
     // Both gathers return the four texels of a 2x2 footprint. Saying so here matters more than
     // for most builtins: the shadow chunks immediately swizzle the result (`samples.wzxy`), and
     // an unknown type degrades swizzle codegen downstream.
-    if (name === 'textureGather' || name === 'textureGatherCompare') {
+    if (name === 'textureGather') {
+        const first = args[0] !== undefined ? args[0].type : null;
+        const texture = first !== null && first.kind === 'texture' ? first : (args[1] !== undefined ? args[1].type : null);
+        const scalar = texture !== null && texture.kind === 'texture' ? texture.scalar : 'f32';
+        return { kind: 'vec', size: 4, scalar: scalar === 'u32' || /uint$/.test(scalar) ? 'u32'
+            : scalar === 'i32' || /sint$/.test(scalar) ? 'i32' : 'f32' };
+    }
+    if (name === 'textureGatherCompare') {
         return { kind: 'vec', size: 4, scalar: 'f32' };
     }
     if (name === 'textureDimensions') {

package/src/shade/wgsl/emulator/WGSLLanguageCore.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"WGSLLanguageCore.d.ts","sourceRoot":"","sources":["../../../../../src/shade/wgsl/emulator/WGSLLanguageCore.js"],"names":[],"mappings":";IAk3DA;;;;;;;;;;;;;OAaG;IACH,+DAyFC"}
+{"version":3,"file":"WGSLLanguageCore.d.ts","sourceRoot":"","sources":["../../../../../src/shade/wgsl/emulator/WGSLLanguageCore.js"],"names":[],"mappings":";IAs3DA;;;;;;;;;;;;;OAaG;IACH,+DAyFC"}

package/src/shade/wgsl/emulator/WGSLLanguageCore.js

@@ -1347,7 +1347,7 @@ export const WGSLLanguageCore = {
             const tz = w * t.size[2] + oz;
             texel_coord = dim === 1 ? [tx] : (dim === 2 ? [tx, ty] : [tx, ty, tz]);
         }
-        const result = new Float32Array(4);
+        const result = t.data instanceof Uint32Array || t.data instanceof Int32Array ? new t.data.constructor(4) : new Float32Array(4);
         cpu_texture_sample(result, sampler ?? FALLBACK_LINEAR_CLAMP_SAMPLER, t, texel_coord);
         return result;
     },
@@ -1504,7 +1504,7 @@ export const WGSLLanguageCore = {
 
         if (extra_b !== undefined) { offset_x = extra_b[0]; offset_y = extra_b[1]; }
 
-        const result = new Float32Array(4);
+        const result = t.data instanceof Uint32Array || t.data instanceof Int32Array ? new t.data.constructor(4) : new Float32Array(4);
 
         cpu_texture_gather(
             result,
@@ -1582,7 +1582,7 @@ export const WGSLLanguageCore = {
             throw new Error("textureGatherCompare: the bound sampler has no 'compare' function — WGSL requires a sampler_comparison here. Bind a SamplerDescriptor with `compare` set (the engine's shadow path uses SHADOWMAP_COMPARE_SAMPLER_DESCRIPTOR, `compare: \"greater\"` for reverse-Z).");
         }
 
-        const result = new Float32Array(4);
+        const result = t.data instanceof Uint32Array || t.data instanceof Int32Array ? new t.data.constructor(4) : new Float32Array(4);
 
         cpu_texture_gather(
             result,
@@ -1773,6 +1773,10 @@ export const WGSLLanguageCore = {
     subgroupOr(value) { return value; },
     subgroupXor(value) { return value; },
 
+    // Votes over a subgroup of one are that one's predicate.
+    subgroupAll(predicate) { return predicate; },
+    subgroupAny(predicate) { return predicate; },
+
     // Scans: exclusive of the only lane is the identity, inclusive of it is its value.
     subgroupExclusiveAdd() { return 0; },
     subgroupExclusiveMul() { return 1; },

package/src/shade/renderer/global_illumination/brick4/cpu/brick4_sh3_color_encode.d.ts

@@ -1,2 +0,0 @@
-export function brick4_sh3_color_encode(output: any, output_offset: any, input: any, input_offset: any): void;
-//# sourceMappingURL=brick4_sh3_color_encode.d.ts.map

package/src/shade/renderer/global_illumination/brick4/cpu/brick4_sh3_color_encode.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"brick4_sh3_color_encode.d.ts","sourceRoot":"","sources":["../../../../../../../src/shade/renderer/global_illumination/brick4/cpu/brick4_sh3_color_encode.js"],"names":[],"mappings":"AAAA,8GAKC"}

package/src/shade/renderer/global_illumination/brick4/cpu/brick4_sh3_color_encode.js

@@ -1,6 +0,0 @@
-export function brick4_sh3_color_encode(
-    output, output_offset,
-    input, input_offset,
-) {
-    throw new Error("Method not implemented.");
-}