rayzee 6.5.0 → 7.1.0

package/src/Stages/SSRC.js

@@ -12,10 +12,11 @@
 // Textures read:       pathtracer:color, pathtracer:normalDepth, motionVector:screenSpace
 
 import { uniform } from 'three/tsl';
-import { StorageTexture, TextureNode } from 'three/webgpu';
-import { HalfFloatType, RGBAFormat, NearestFilter, LinearFilter } from 'three';
+import { StorageTexture, TextureNode, RenderTarget } from 'three/webgpu';
+import { HalfFloatType, RGBAFormat, NearestFilter, LinearFilter, Box2, Vector2 } from 'three';
 import { RenderStage, StageExecutionMode } from '../Pipeline/RenderStage.js';
 import { buildTemporalPass, buildSpatialPass } from '../TSL/SSRC.js';
+import { MAX_STORAGE_TEXTURE_SIZE } from '../EngineDefaults.js';
 
 export class SSRC extends RenderStage {
 
@@ -51,18 +52,30 @@ export class SSRC extends RenderStage {
 		this._readPass1CacheTexNode = new TextureNode(); // current cache (for spatial pass)
 
 		// ─── StorageTextures (5 total) ───
-		const w = 1, h = 1; // resized on first render
+		// StorageTextures stay at max alloc — see resize crash fix (three.js #33061).
+		const w = 1, h = 1; // RTs/uniforms resized on first render
 
 		// Ping-pong temporal cache: .rgb = radiance, .w = history count
-		this._cacheTexA = this._createStorageTex( w, h, NearestFilter );
-		this._cacheTexB = this._createStorageTex( w, h, NearestFilter );
+		this._cacheTexA = this._createStorageTex( MAX_STORAGE_TEXTURE_SIZE, MAX_STORAGE_TEXTURE_SIZE, NearestFilter );
+		this._cacheTexB = this._createStorageTex( MAX_STORAGE_TEXTURE_SIZE, MAX_STORAGE_TEXTURE_SIZE, NearestFilter );
 
 		// Ping-pong previous-frame normalDepth (for edge-stopping in temporal pass)
-		this._prevNDTexA = this._createStorageTex( w, h, NearestFilter );
-		this._prevNDTexB = this._createStorageTex( w, h, NearestFilter );
+		this._prevNDTexA = this._createStorageTex( MAX_STORAGE_TEXTURE_SIZE, MAX_STORAGE_TEXTURE_SIZE, NearestFilter );
+		this._prevNDTexB = this._createStorageTex( MAX_STORAGE_TEXTURE_SIZE, MAX_STORAGE_TEXTURE_SIZE, NearestFilter );
 
 		// Final output (LinearFilter for Display fragment shader sampling)
-		this._outputTex = this._createStorageTex( w, h, LinearFilter );
+		this._outputTex = this._createStorageTex( MAX_STORAGE_TEXTURE_SIZE, MAX_STORAGE_TEXTURE_SIZE, LinearFilter );
+
+		// Active-region copy target — published downstream (storage tex is over-allocated)
+		this._srcRegion = new Box2( new Vector2( 0, 0 ), new Vector2( 0, 0 ) );
+		this.outputTarget = new RenderTarget( w, h, {
+			type: HalfFloatType,
+			format: RGBAFormat,
+			minFilter: LinearFilter,
+			magFilter: LinearFilter,
+			depthBuffer: false,
+			stencilBuffer: false
+		} );
 
 		// ─── State ───
 		this._currentPingPong = 0; // 0: read B, write A; 1: read A, write B
@@ -99,7 +112,7 @@ export class SSRC extends RenderStage {
 		if ( colorTex?.image ) {
 
 			const { width, height } = colorTex.image;
-			if ( width !== this._cacheTexA.image.width || height !== this._cacheTexA.image.height ) {
+			if ( width !== this.outputTarget.width || height !== this.outputTarget.height ) {
 
 				this.setSize( width, height );
 
@@ -145,8 +158,13 @@ export class SSRC extends RenderStage {
 		// Advance frames-since-reset counter (capped to avoid overflow)
 		this._framesSinceReset.value = Math.min( this._framesSinceReset.value + 1, 9999 );
 
+		// Copy active region out of the over-allocated StorageTexture into the
+		// right-sized RenderTarget; downstream stages UV-sample the latter.
+		this._srcRegion.max.set( this.outputTarget.width, this.outputTarget.height );
+		this.renderer.copyTextureToTexture( this._outputTex, this.outputTarget.texture, this._srcRegion );
+
 		// Publish final output
-		context.setTexture( 'ssrc:output', this._outputTex );
+		context.setTexture( 'ssrc:output', this.outputTarget.texture );
 
 		// Advance ping-pong
 		this._currentPingPong = 1 - this._currentPingPong;
@@ -163,11 +181,9 @@ export class SSRC extends RenderStage {
 
 		if ( width < 1 || height < 1 ) return;
 
-		this._cacheTexA.setSize( width, height );
-		this._cacheTexB.setSize( width, height );
-		this._prevNDTexA.setSize( width, height );
-		this._prevNDTexB.setSize( width, height );
-		this._outputTex.setSize( width, height );
+		// StorageTextures stay at their max allocation (see constructor).
+		this.outputTarget.setSize( width, height );
+		this.outputTarget.texture.needsUpdate = true;
 
 		this.resW.value = width;
 		this.resH.value = height;
@@ -194,6 +210,7 @@ export class SSRC extends RenderStage {
 		this._prevNDTexA.dispose();
 		this._prevNDTexB.dispose();
 		this._outputTex.dispose();
+		this.outputTarget?.dispose();
 		this._colorTexNode?.dispose();
 		this._ndTexNode?.dispose();
 		this._motionTexNode?.dispose();

package/src/Stages/Variance.js

@@ -1,9 +1,10 @@
 import { Fn, wgslFn, float, int, uint, ivec2, uvec2, uniform, If, max,
 	textureLoad, textureStore, workgroupArray, workgroupBarrier, localId, workgroupId } from 'three/tsl';
-import { TextureNode, StorageTexture } from 'three/webgpu';
-import { FloatType, RGBAFormat, LinearFilter } from 'three';
+import { RenderTarget, TextureNode, StorageTexture } from 'three/webgpu';
+import { FloatType, RGBAFormat, LinearFilter, Box2, Vector2 } from 'three';
 import { RenderStage, StageExecutionMode } from '../Pipeline/RenderStage.js';
 import { luminance } from '../TSL/Common.js';
+import { MAX_STORAGE_TEXTURE_SIZE } from '../EngineDefaults.js';
 
 // ── wgslFn helpers ──────────────────────────────────────────
 
@@ -40,8 +41,7 @@ const temporalAccumulate = /*@__PURE__*/ wgslFn( `
  * WebGPU Variance Estimation Stage (Compute Shader)
  *
  * Computes temporal and spatial variance from the path tracer output.
- * Used by AdaptiveSampling for sampling guidance and by
- * BilateralFilter for variance-guided filtering.
+ * Used by BilateralFilter for variance-guided filtering.
  *
  * Uses compute shader with workgroup shared memory for the 3×3
  * spatial variance computation. Each 8×8 workgroup loads a 10×10
@@ -98,18 +98,34 @@ export class Variance extends RenderStage {
 		// FloatType (f32) required — HalfFloat's ~3.3 decimal digits cause catastrophic
 		// cancellation in (meanSq - mean²) for converged pixels, producing a variance
 		// floor of ~0.0001 that the (frame+1)² scaling amplifies to enormous values.
-		this._storageTexA = new StorageTexture( w, h );
+		// StorageTextures over-allocated at max — defensive against three.js #33061
+		// (TSL compute pipeline keeps stale GPUTextureView after StorageTexture.setSize).
+		this._storageTexA = new StorageTexture( MAX_STORAGE_TEXTURE_SIZE, MAX_STORAGE_TEXTURE_SIZE );
 		this._storageTexA.type = FloatType;
 		this._storageTexA.format = RGBAFormat;
 		this._storageTexA.minFilter = LinearFilter;
 		this._storageTexA.magFilter = LinearFilter;
 
-		this._storageTexB = new StorageTexture( w, h );
+		this._storageTexB = new StorageTexture( MAX_STORAGE_TEXTURE_SIZE, MAX_STORAGE_TEXTURE_SIZE );
 		this._storageTexB.type = FloatType;
 		this._storageTexB.format = RGBAFormat;
 		this._storageTexB.minFilter = LinearFilter;
 		this._storageTexB.magFilter = LinearFilter;
 
+		this._srcRegion = new Box2( new Vector2( 0, 0 ), new Vector2( 0, 0 ) );
+
+		// Right-sized RenderTarget published to downstream (UV-sampled). The
+		// over-allocated StorageTexture itself must not be published — UV
+		// sampling a 2048 texture would read the wrong region.
+		this._outputTarget = new RenderTarget( w, h, {
+			type: FloatType,
+			format: RGBAFormat,
+			minFilter: LinearFilter,
+			magFilter: LinearFilter,
+			depthBuffer: false,
+			stencilBuffer: false
+		} );
+
 		this.currentMoments = 0; // 0 = write A, read B; 1 = write B, read A
 		this._compiled = false;
 		this._needsWarmReset = false;
@@ -275,8 +291,8 @@ export class Variance extends RenderStage {
 		const img = colorTex.image;
 		if ( img && img.width > 0 && img.height > 0 ) {
 
-			if ( img.width !== this._storageTexA.image.width ||
-				img.height !== this._storageTexA.image.height ) {
+			if ( img.width !== this._outputTarget.width ||
+				img.height !== this._outputTarget.height ) {
 
 				this.setSize( img.width, img.height );
 
@@ -333,8 +349,13 @@ export class Variance extends RenderStage {
 		// Swap for next frame
 		this.currentMoments = 1 - this.currentMoments;
 
-		// Publish (StorageTexture works as regular Texture for downstream sampling)
-		context.setTexture( 'variance:output', writeTarget );
+		// Copy the active region out of the over-allocated StorageTexture into the
+		// right-sized RenderTarget; downstream stages UV-sample the latter.
+		this._srcRegion.max.set( this._outputTarget.width, this._outputTarget.height );
+		this.renderer.copyTextureToTexture( writeTarget, this._outputTarget.texture, this._srcRegion );
+
+		// Publish the RenderTarget (not the over-allocated StorageTexture)
+		context.setTexture( 'variance:output', this._outputTarget.texture );
 
 	}
 
@@ -352,8 +373,9 @@ export class Variance extends RenderStage {
 
 	setSize( width, height ) {
 
-		this._storageTexA.setSize( width, height );
-		this._storageTexB.setSize( width, height );
+		// StorageTextures stay at their max allocation (see constructor).
+		this._outputTarget.setSize( width, height );
+		this._outputTarget.texture.needsUpdate = true;
 		this.resW.value = width;
 		this.resH.value = height;
 
@@ -371,6 +393,7 @@ export class Variance extends RenderStage {
 		this._computeNodeB?.dispose();
 		this._storageTexA?.dispose();
 		this._storageTexB?.dispose();
+		this._outputTarget?.dispose();
 		this._colorTexNode?.dispose();
 		this._readTexNodeA?.dispose();
 		this._readTexNodeB?.dispose();

package/src/TSL/CompactKernel.js

@@ -0,0 +1,110 @@
+/**
+ * CompactKernel.js — wavefront stream compaction: active rays → dense index array for next bounce (256×1, 1D).
+ */
+
+import {
+	Fn, uint, select,
+	If,
+	instanceIndex,
+	atomicAdd, atomicLoad,
+	subgroupExclusiveAdd, subgroupAdd, subgroupBroadcast,
+	Return,
+} from 'three/tsl';
+
+import { readRayBounceFlags } from '../Processor/PackedRayBuffer.js';
+import { RAY_FLAG, COUNTER } from '../Processor/QueueManager.js';
+
+const WG_SIZE = 256;
+
+export function buildCompactKernel( params ) {
+
+	const {
+		rayBufferRO,
+		activeIndicesReadRO,
+		activeIndicesWriteRW,
+		counters,
+		currentActiveCount,
+	} = params;
+
+	const computeFn = Fn( () => {
+
+		const threadIdx = instanceIndex;
+
+		// ACTIVE_RAY_COUNT is zeroed before compact, so the dense-list length comes from ENTERING_COUNT.
+		const bound = counters ? atomicLoad( counters.element( uint( COUNTER.ENTERING_COUNT ) ) ) : currentActiveCount;
+		If( threadIdx.greaterThanEqual( bound ), () => {
+
+			Return();
+
+		} );
+
+		const rayID = activeIndicesReadRO.element( threadIdx );
+
+		const flags = readRayBounceFlags( rayBufferRO, rayID );
+
+		If( flags.bitAnd( uint( RAY_FLAG.ACTIVE ) ).notEqual( uint( 0 ) ), () => {
+
+			const writeIdx = atomicAdd( counters.element( uint( COUNTER.ACTIVE_RAY_COUNT ) ), uint( 1 ) );
+			activeIndicesWriteRW.element( writeIdx ).assign( rayID );
+
+		} );
+
+	} );
+
+	return computeFn;
+
+}
+
+/**
+ * Subgroup prefix-sum compaction: one global atomicAdd per subgroup instead of per survivor.
+ * Requires renderer.hasFeature('subgroups'); control flow must stay uniform (no divergent Return).
+ */
+export function buildCompactSubgroupKernel( params ) {
+
+	const {
+		rayBufferRO,
+		activeIndicesReadRO,
+		activeIndicesWriteRW,
+		counters,
+		currentActiveCount,
+	} = params;
+
+	const computeFn = Fn( () => {
+
+		const threadIdx = instanceIndex;
+		const bound = counters ? atomicLoad( counters.element( uint( COUNTER.ENTERING_COUNT ) ) ) : currentActiveCount;
+
+		// No early Return: all lanes must reach the subgroup ops; out-of-range lanes contribute 0 and read stale-but-in-capacity slots.
+		const inRange = threadIdx.lessThan( bound );
+		const rayID = activeIndicesReadRO.element( threadIdx );
+		const flags = readRayBounceFlags( rayBufferRO, rayID );
+		const isActive = inRange.and( flags.bitAnd( uint( RAY_FLAG.ACTIVE ) ).notEqual( uint( 0 ) ) );
+		const activeU = select( isActive, uint( 1 ), uint( 0 ) );
+
+		// .toVar() materializes the subgroup ops at uniform control flow; inlining into the divergent If(isActive) write is rejected by WGSL.
+		const localOffset = subgroupExclusiveAdd( activeU ).toVar();
+		const sgCount = subgroupAdd( activeU ).toVar();
+
+		// laneId via exclusiveAdd(1) since TSL lacks subgroup_invocation_id; lane 0 does the single per-subgroup atomicAdd.
+		const laneId = subgroupExclusiveAdd( uint( 1 ) ).toVar();
+		const base = uint( 0 ).toVar();
+		If( laneId.equal( uint( 0 ) ), () => {
+
+			base.assign( atomicAdd( counters.element( uint( COUNTER.ACTIVE_RAY_COUNT ) ), sgCount ) );
+
+		} );
+		const sgBase = subgroupBroadcast( base, uint( 0 ) ).toVar();
+
+		If( isActive, () => {
+
+			activeIndicesWriteRW.element( sgBase.add( localOffset ) ).assign( rayID );
+
+		} );
+
+	} );
+
+	return computeFn;
+
+}
+
+export { WG_SIZE as COMPACT_WG_SIZE };

package/src/TSL/DebugKernel.js

@@ -0,0 +1,98 @@
+/**
+ * DebugKernel.js — wavefront debug visualization (16×16, 2D screen-space dispatch).
+ *
+ * Single-pass primary-ray debug viz for visMode 1-10 (mode 11 = NaN/Inf is a FinalWrite
+ * post-branch on the accumulated color, handled there). Generates a camera ray per pixel and
+ * delegates to the renderer-agnostic TraceDebugMode for the per-mode color; mode 9 (stratified
+ * sample pattern) is computed inline. Writes the color directly to the output (no accumulation).
+ */
+
+import {
+	Fn, float, vec2, vec4, int, uint, uvec2,
+	If, textureStore,
+	localId, workgroupId,
+} from 'three/tsl';
+
+import { generateRayFromCamera } from './BVHTraversal.js';
+import { Ray } from './Struct.js';
+import { TraceDebugMode } from './Debugger.js';
+import { pcgHash, getStratifiedSample } from './Random.js';
+
+const WG_SIZE = 16;
+
+export function buildDebugKernel( params ) {
+
+	const {
+		writeColorTex, writeNDTex, writeAlbedoTex,
+		resolution, renderWidth, renderHeight,
+		cameraWorldMatrix, cameraProjectionMatrixInverse, cameraProjectionMatrix, cameraViewMatrix,
+		enableDOF, focalLength, aperture, focusDistance, sceneScale, apertureScale, anamorphicRatio,
+		bvhBuffer, triangleBuffer, materialBuffer,
+		envTexture, environmentMatrix, environmentIntensity, enableEnvironmentLight,
+		visMode, debugVisScale,
+		albedoMaps, normalMaps, bumpMaps, metalnessMaps, roughnessMaps, emissiveMaps,
+		frame,
+		samplesPerPass = 1,
+	} = params;
+
+	const computeFn = Fn( () => {
+
+		const gx = int( workgroupId.x ).mul( WG_SIZE ).add( int( localId.x ) );
+		const gy = int( workgroupId.y ).mul( WG_SIZE ).add( int( localId.y ) );
+
+		If( gx.lessThan( renderWidth ).and( gy.lessThan( renderHeight ) ), () => {
+
+			const pixelCoord = vec2( float( gx ).add( 0.5 ), float( gy ).add( 0.5 ) );
+			const pixelIndex = gy.mul( int( resolution.x ) ).add( gx );
+			const seed = pcgHash( { state: uint( pixelIndex ).add( uint( 1 ) ) } ).toVar();
+
+			// Center-pixel primary ray (no AA jitter — debug viz wants a stable, sharp image).
+			const screenPosition = pixelCoord.div( resolution ).mul( 2.0 ).sub( 1.0 ).toVar();
+			screenPosition.y.assign( screenPosition.y.negate() );
+
+			const ray = Ray.wrap( generateRayFromCamera(
+				screenPosition, seed,
+				cameraWorldMatrix, cameraProjectionMatrixInverse,
+				enableDOF, focalLength, aperture, focusDistance, sceneScale, apertureScale, anamorphicRatio,
+			) );
+
+			const color = vec4( 1.0, 0.0, 1.0, 1.0 ).toVar();
+
+			// Mode 9: visualize the stratified AA-jitter pattern (R,G = jitter).
+			If( visMode.equal( int( 9 ) ), () => {
+
+				// Use the real per-frame sample count so >1 SPP shows the stratified lattice (totalRays≤1 → plain random).
+				const jitter = getStratifiedSample( pixelCoord, int( 0 ), int( samplesPerPass ), seed, resolution, frame );
+				color.assign( vec4( jitter, 1.0, 1.0 ) );
+
+			} ).Else( () => {
+
+				// Modes 1-8, 10 — shared per-mode debug color (primary-ray trace + counters).
+				color.assign( TraceDebugMode(
+					ray.origin, ray.direction,
+					bvhBuffer, triangleBuffer, materialBuffer,
+					envTexture, environmentMatrix, environmentIntensity, enableEnvironmentLight,
+					visMode, debugVisScale,
+					pixelCoord, resolution,
+					albedoMaps, normalMaps, bumpMaps, metalnessMaps, roughnessMaps, emissiveMaps,
+					cameraProjectionMatrix, cameraViewMatrix,
+					frame,
+				) );
+
+			} );
+
+			const uintCoord = uvec2( uint( gx ), uint( gy ) );
+			textureStore( writeColorTex, uintCoord, color ).toWriteOnly();
+			// Benign MRT so the denoiser/display never read stale normal/albedo on a debug frame.
+			textureStore( writeNDTex, uintCoord, vec4( 0.5, 0.5, 1.0, 1.0 ) ).toWriteOnly();
+			textureStore( writeAlbedoTex, uintCoord, vec4( color.xyz, 1.0 ) ).toWriteOnly();
+
+		} );
+
+	} );
+
+	return computeFn;
+
+}
+
+export { WG_SIZE as DEBUG_WG_SIZE };

package/src/TSL/Environment.js

@@ -1,4 +1,4 @@
-import { Fn, wgslFn, vec2, vec4, float, int, If, texture, dot, sin, sqrt, floor, fract, min, max, mix, clamp } from 'three/tsl';
+import { Fn, wgslFn, vec2, vec4, ivec2, float, int, If, texture, dot, sin, sqrt, floor, fract, min, max, mix, clamp } from 'three/tsl';
 
 import { REC709_LUMINANCE_COEFFICIENTS } from './Common.js';
 
@@ -74,7 +74,7 @@ export const sampleEquirect = Fn( ( [ environment, direction, environmentMatrix,
 // Exact implementation from three-gpu-pathtracer
 export const sampleEquirectProbability = Fn( ( [
 	environment,
-	envCDFBuffer,
+	envCDFTexture,
 	environmentMatrix,
 	environmentIntensity,
 	envTotalSum,
@@ -84,9 +84,8 @@ export const sampleEquirectProbability = Fn( ( [
 	colorOutput
 ] ) => {
 
-	// Packed CDF layout: [marginal (envResolution.y floats) | conditional (envResolution.x * envResolution.y floats)]
-	// The conditional offset equals the marginal length, which is envResolution.y.
-	const condOffset = int( envResolution.y ).toVar();
+	// CDF texture layout: (W+1)×H R32F — conditional[cy*W+cx] at texel (cx,cy); marginal[cy] at column W.
+	const cdfMarginalCol = int( envResolution.x ).toVar();
 
 	// Sample marginal CDF for V coordinate (1D, linear interpolation)
 	const marginalSize = envResolution.y;
@@ -94,7 +93,11 @@ export const sampleEquirectProbability = Fn( ( [
 	const mI0 = int( floor( mIdx ) );
 	const mI1 = min( mI0.add( 1 ), int( marginalSize ).sub( 1 ) );
 	const mFrac = fract( mIdx );
-	const v = mix( envCDFBuffer.element( mI0 ), envCDFBuffer.element( mI1 ), mFrac ).toVar();
+	const v = mix(
+		envCDFTexture.load( ivec2( cdfMarginalCol, mI0 ) ).x,
+		envCDFTexture.load( ivec2( cdfMarginalCol, mI1 ) ).x,
+		mFrac,
+	).toVar();
 
 	// Sample conditional CDF for U coordinate (2D grid, bilinear interpolation)
 	const condW = envResolution.x;
@@ -107,11 +110,10 @@ export const sampleEquirectProbability = Fn( ( [
 	const cy1 = min( cy0.add( 1 ), int( condH ).sub( 1 ) );
 	const fx = fract( cxf );
 	const fy = fract( cyf );
-	const condWi = int( condW );
-	const v00 = envCDFBuffer.element( condOffset.add( cy0.mul( condWi ).add( cx0 ) ) );
-	const v10 = envCDFBuffer.element( condOffset.add( cy0.mul( condWi ).add( cx1 ) ) );
-	const v01 = envCDFBuffer.element( condOffset.add( cy1.mul( condWi ).add( cx0 ) ) );
-	const v11 = envCDFBuffer.element( condOffset.add( cy1.mul( condWi ).add( cx1 ) ) );
+	const v00 = envCDFTexture.load( ivec2( cx0, cy0 ) ).x;
+	const v10 = envCDFTexture.load( ivec2( cx1, cy0 ) ).x;
+	const v01 = envCDFTexture.load( ivec2( cx0, cy1 ) ).x;
+	const v11 = envCDFTexture.load( ivec2( cx1, cy1 ) ).x;
 	const u = mix( mix( v00, v10, fx ), mix( v01, v11, fx ), fy ).toVar();
 
 	const uv = vec2( u, v ).toVar();

package/src/TSL/ExtendKernel.js

@@ -0,0 +1,75 @@
+/**
+ * ExtendKernel.js — wavefront BVH traversal (256×1, 1D ray-parallel dispatch).
+ */
+
+import {
+	Fn, uint,
+	If,
+	instanceIndex,
+	atomicLoad,
+	Return,
+} from 'three/tsl';
+
+import { traverseBVH } from './BVHTraversal.js';
+import { Ray, HitInfo } from './Struct.js';
+import {
+	readRayOrigin, readRayDirection, readMediumStack,
+	writeHitPacked,
+} from '../Processor/PackedRayBuffer.js';
+import { COUNTER } from '../Processor/QueueManager.js';
+
+const WG_SIZE = 256;
+
+export function buildExtendKernel( params ) {
+
+	const {
+		bvhBuffer, triangleBuffer, materialBuffer,
+		rayBufferRO,
+		hitBufferRW,
+		activeIndicesRO,
+		counters,
+		maxRayCount,
+	} = params;
+
+	const computeFn = Fn( () => {
+
+		const threadIdx = instanceIndex;
+
+		// kernels bound on ENTERING_COUNT so an over-sized (margin) dispatch is safe.
+		const bound = counters ? atomicLoad( counters.element( uint( COUNTER.ENTERING_COUNT ) ) ) : maxRayCount;
+		If( threadIdx.greaterThanEqual( bound ), () => {
+
+			Return();
+
+		} );
+
+		const rayID = activeIndicesRO.element( threadIdx );
+
+		const origin = readRayOrigin( rayBufferRO, rayID ).toVar();
+		const direction = readRayDirection( rayBufferRO, rayID ).toVar();
+
+		const ray = Ray( { origin, direction } );
+
+		// insideMedium bypasses front/back culling so the ray can hit a glass/SSS back-facing boundary.
+		const insideMedium = readMediumStack( rayBufferRO, rayID ).stackDepth.greaterThan( uint( 0 ) );
+		const hitInfo = HitInfo.wrap( traverseBVH(
+			ray, bvhBuffer, triangleBuffer, insideMedium,
+		) ).toVar();
+
+		writeHitPacked(
+			hitBufferRW, rayID,
+			hitInfo.dst,
+			uint( hitInfo.triangleIndex ),
+			hitInfo.uv.x, hitInfo.uv.y,
+			hitInfo.normal,
+			uint( hitInfo.materialIndex ),
+			uint( hitInfo.meshIndex ),
+		);
+
+	} );
+
+	return computeFn;
+
+}
+
+export { WG_SIZE as EXTEND_WG_SIZE };

package/src/TSL/FinalWriteKernel.js

@@ -0,0 +1,121 @@
+/**
+ * FinalWriteKernel.js — wavefront final output: temporal accumulation + MRT StorageTexture writes (16×16, 2D).
+ */
+
+import {
+	Fn, wgslFn, float, vec2, vec4, int, uint, uvec2,
+	If, mix, select, texture, textureStore,
+	localId, workgroupId,
+} from 'three/tsl';
+
+import {
+	readRayRadiance, readGBuffer, gbDecodeNormalDepth, gbDecodeAlbedo,
+} from '../Processor/PackedRayBuffer.js';
+
+const WG_SIZE = 16;
+
+// Debug mode 11: NaN/Inf detector — red where the accumulated color is NaN/Inf, black elsewhere.
+const nanInfToRed = /*@__PURE__*/ wgslFn( `
+	fn nanInfToRed( c: vec3f ) -> vec3f {
+		let isNan = c.x != c.x || c.y != c.y || c.z != c.z;
+		let isInf = abs( c.x ) > 1e30f || abs( c.y ) > 1e30f || abs( c.z ) > 1e30f;
+		if ( isNan || isInf ) { return vec3f( 1.0f, 0.0f, 0.0f ); }
+		return vec3f( 0.0f );
+	}
+` );
+
+export function buildFinalWriteKernel( params ) {
+
+	const {
+		rayBufferRO, gBufferRO,
+		writeColorTex, writeNDTex, writeAlbedoTex,
+		resolution, frame,
+		enableAccumulation, hasPreviousAccumulated, accumulationAlpha, cameraIsMoving,
+		transparentBackground,
+		prevAccumTexture, prevNormalDepthTexture, prevAlbedoTexture,
+		renderWidth, renderHeight,
+		// Multi-sample: average S sample-slots per pixel (slot = pixel + k*w*h, w*h from the resolution uniform).
+		samplesPerPass = 1,
+		visMode,
+	} = params;
+
+	const S = samplesPerPass | 0;
+
+	const computeFn = Fn( () => {
+
+		const gx = int( workgroupId.x ).mul( WG_SIZE ).add( int( localId.x ) );
+		const gy = int( workgroupId.y ).mul( WG_SIZE ).add( int( localId.y ) );
+
+		If( gx.lessThan( renderWidth ).and( gy.lessThan( renderHeight ) ), () => {
+
+			const pixelIndex = gy.mul( int( resolution.x ) ).add( gx );
+			const rayID = uint( pixelIndex );
+
+			// Average the S sub-samples; MRT (normal/depth/albedo) from sub-sample 0.
+			const sampleColor = ( () => {
+
+				if ( S <= 1 ) return readRayRadiance( rayBufferRO, rayID );
+				const acc = readRayRadiance( rayBufferRO, rayID ).toVar();
+				const mrps = uint( resolution.x ).mul( uint( resolution.y ) ).toVar(); // w*h from the resolution uniform, not baked
+				for ( let k = 1; k < S; k ++ ) {
+
+					acc.addAssign( readRayRadiance( rayBufferRO, rayID.add( uint( k ).mul( mrps ) ) ) );
+
+				}
+
+				acc.assign( acc.div( float( S ) ) );
+				return acc;
+
+			} )();
+			// MRT comes from the per-pixel G-buffer (rayID == pixelIndex here, i.e. sub-sample 0). Half-packed: decode.
+			const gbuf = readGBuffer( gBufferRO, rayID );
+			const normalDepth = gbDecodeNormalDepth( gbuf );
+			const albedoID = vec4( gbDecodeAlbedo( gbuf ), 0.0 );
+
+			const finalColor = sampleColor.xyz.toVar();
+			const finalNormalDepth = normalDepth.toVar();
+			const finalAlbedo = albedoID.xyz.toVar();
+			const outputAlpha = select( transparentBackground, sampleColor.w, float( 1.0 ) ).toVar();
+
+			const pixelCoord = vec2( float( gx ).add( 0.5 ), float( gy ).add( 0.5 ) );
+			const prevUV = pixelCoord.div( resolution );
+
+			// visMode 11 (NaN/Inf) bypasses accumulation (megakernel parity main_TSL_PathTracer.js:355) so the
+			// detector runs on each frame's fresh color — else mix() propagates a transient NaN and it stays red forever.
+			If( enableAccumulation.and( cameraIsMoving.not() ).and( frame.greaterThan( uint( 0 ) ) ).and( hasPreviousAccumulated ).and( visMode.notEqual( int( 11 ) ) ), () => {
+
+				const prevAccumSample = texture( prevAccumTexture, prevUV, 0 ).toVar();
+
+				finalColor.assign( mix( prevAccumSample.xyz, sampleColor.xyz, accumulationAlpha ) );
+				finalNormalDepth.assign( mix( texture( prevNormalDepthTexture, prevUV, 0 ), finalNormalDepth, accumulationAlpha ) );
+				finalAlbedo.assign( mix( texture( prevAlbedoTexture, prevUV, 0 ).xyz, finalAlbedo, accumulationAlpha ) );
+
+				If( transparentBackground, () => {
+
+					outputAlpha.assign( mix( prevAccumSample.w, sampleColor.w, accumulationAlpha ) );
+
+				} );
+
+			} );
+
+			// Debug mode 11: flag NaN/Inf on the accumulated color (red on NaN/Inf, black elsewhere).
+			If( visMode.equal( int( 11 ) ), () => {
+
+				finalColor.assign( nanInfToRed( finalColor ) );
+
+			} );
+
+			const uintCoord = uvec2( uint( gx ), uint( gy ) );
+			textureStore( writeColorTex, uintCoord, vec4( finalColor, outputAlpha ) ).toWriteOnly();
+			textureStore( writeNDTex, uintCoord, finalNormalDepth ).toWriteOnly();
+			textureStore( writeAlbedoTex, uintCoord, vec4( finalAlbedo, 1.0 ) ).toWriteOnly();
+
+		} );
+
+	} );
+
+	return computeFn;
+
+}
+
+export { WG_SIZE as FINALWRITE_WG_SIZE };