samien
/
three.js
cermin dari https://github.com/mrdoob/three.js.git


			
							12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697989910010110210310410510610710810911011111211311411511611711811912012112212312412512612712812913013113213313413513613713813914014114214314414514614714814915015115215315415515615715815916016116216316416516616716816917017117217317417517617717817918018118218318418518618718818919019119219319419519619719819920020120220320420520620720820921021121221321421521621721821922022122222322422522622722822923023123223323423523623723823924024124224324424524624724824925025125225325425525625725825926026126226326426526626726826927027127227327427527627727827928028128228328428528628728828929029129229329429529629729829930030130230330430530630730830931031131231331431531631731831932032132232332432532632732832933033133233333433533633733833934034134234334434534634734834935035135235335435535635735835936036136236336436536636736836937037137237337437537637737837938038138238338438538638738838939039139239339439539639739839940040140240340440540640740840941041141241341441541641741841942042142242342442542642742842943043143243343443543643743843944044144244344444544644744844945045145245345445545645745845946046146246346446546646746846947047147247347447547647747847948048148248348448548648748848949049149249349449549649749849950050150250350450550650750850951051151251351451551651751851952052152252352452552652752852953053153253353453553653753853954054154254354454554654754854955055155255355455555655755855956056156256356456556656756856957057157257357457557657757857958058158258358458558658758858959059159259359459559659759859960060160260360460560660760860961061161261361461561661761861962062162262362462562662762862963063163263363463563663763863964064164264364464564664764864965065165265365465565665765865966066166266366466566666766866967067167267367467567667767867968068168268368468568668768868969069169269369469569669769869970070170270370470570670770870971071171271371471571671771871972072172272372472572672772872973073173273373473573673773873974074174274374474574674774874975075175275375475575675775875976076176276376476576676776876977077177277377477577677777877978078178278378478578678778878979079179279379479579679779879980080180280380480580680780880981081181281381481581681781881982082182282382482582682782882983083183283383483583683783883984084184284384484584684784884985085185285385485585685785885986086186286386486586686786886987087187287387487587687787887988088188288388488588688788888989089189289389489589689789889990090190290390490590690790890991091191291391491591691791891992092192292392492592692792892993093193293393493593693793893994094194294394494594694794894995095195295395495595695795895996096196296396496596696796896997097197297397497597697797897998098198298398498598698798898999099199299399499599699799899910001001100210031004100510061007100810091010101110121013101410151016101710181019102010211022102310241025102610271028102910301031103210331034103510361037103810391040104110421043104410451046104710481049105010511052105310541055105610571058105910601061106210631064106510661067106810691070107110721073107410751076107710781079108010811082108310841085108610871088108910901091109210931094109510961097109810991100110111021103110411051106110711081109111011111112111311141115111611171118111911201121112211231124112511261127112811291130113111321133113411351136
							<!DOCTYPE html>
<html lang="en">
	<head>
		<title>three.js webgpu - compute rasterizer</title>
		<meta charset="utf-8">
		<meta name="viewport" content="width=device-width, user-scalable=no, minimum-scale=1.0, maximum-scale=1.0">
		<meta property="og:title" content="three.js webgpu - compute rasterizer">
		<meta property="og:type" content="website">
		<meta property="og:url" content="https://threejs.org/examples/webgpu_compute_rasterizer.html">
		<meta property="og:image" content="https://threejs.org/examples/screenshots/webgpu_compute_rasterizer.jpg">
		<link type="text/css" rel="stylesheet" href="example.css">
	</head>
	<body>

		<div id="info">
			<a href="https://threejs.org/" target="_blank" rel="noopener" class="logo-link"></a>

			<div class="title-wrapper">
				<a href="https://threejs.org/" target="_blank" rel="noopener">three.js</a><span>GPU-Driven Compute Rasterizer</span>
			</div>

			<small>Rendering <span id="triangleCount"></span> triangles.</small>
		</div>

		<script type="importmap">
			{
				"imports": {
					"three": "../build/three.webgpu.js",
					"three/webgpu": "../build/three.webgpu.js",
					"three/tsl": "../build/three.tsl.js",
					"three/addons/": "./jsm/"
				}
			}
		</script>

		<script type="module">

			import * as THREE from 'three/webgpu';
			import { Fn, If, Loop, vec4, vec2, uvec4, mat4, uint, float, int, min, max, atomicMax, atomicAdd, atomicStore, atomicLoad, floor, cos, sin, dot, bool, storage, uniform, uniformArray, uv, instanceIndex, vertexIndex, distance, screenSize, time, texture, varyingProperty, sqrt } from 'three/tsl';

			import { OrbitControls } from 'three/addons/controls/OrbitControls.js';
			import { TeapotGeometry } from 'three/addons/geometries/TeapotGeometry.js';

			import { Inspector } from 'three/addons/inspector/Inspector.js';

			import WebGPU from 'three/addons/capabilities/WebGPU.js';

			if ( WebGPU.isAvailable() === false ) {

				document.body.appendChild( WebGPU.getErrorMessage() );

				throw new Error( 'No WebGPU support' );

			}

			let camera, renderer, controls;
			let computeRasterize, computeClear, computeFrustum, computeDispatch, computeHWArgs;
			let quadMesh, hwScene, hwMesh;
			let cameraPos, projScreenMatrixUniform, frustumPlanesUniform, cotHalfFovUniform;

			let screenTriAttr, screenTriAtomic, screenTriRead;
			let screenInstAttr, screenInstBuffer, screenInstRead;
			let maxPixels;

			const rows = 400;
			const cols = 400;
			const instanceCount = rows * cols;

			const MAX_RASTER_SIZE = 16;
			const options = { Mode: 'Meshlet Debug', Rasterizer: 'Both' };

			// Buffer visibility packaging configuration
			const TRIANGLE_INDEX_BITS = 14; 			// Bits allocated for triangle index (2^14 = 16384 max triangles)
			const TRIANGLE_INDEX_MASK = 0x3FFF; 		// Bitmask to extract triangle index (14 bits)
			const DEPTH_PRECISION_MAX = 4294967295.0; 	// Maximum value of the 32-bit depth (2^32 - 1)

			const background = new THREE.Color( .1, .1, .1 );

			init();

			async function init() {

				renderer = new THREE.WebGPURenderer();
				renderer.setPixelRatio( window.devicePixelRatio );
				renderer.setSize( window.innerWidth, window.innerHeight );
				renderer.setAnimationLoop( animate );
				renderer.inspector = new Inspector();
				document.body.appendChild( renderer.domElement );

				await renderer.init();

				camera = new THREE.PerspectiveCamera( 50, window.innerWidth / window.innerHeight, .25, 1000000 );
				camera.position.set( 0, 15, 50 );

				controls = new OrbitControls( camera, renderer.domElement );
				controls.target.y = - 1.5;
				controls.enableDamping = true;
				controls.zoomSpeed = .5;
				controls.maxDistance = 1000;
				controls.maxPolarAngle = Math.PI / 2;

				// Generate LOD Geometries
				const lods = [
					{ geometry: new TeapotGeometry( 1, 10 ), error: 0.0 },
					{ geometry: new TeapotGeometry( 1, 8 ), error: 0.005 },
					{ geometry: new TeapotGeometry( 1, 6 ), error: 0.015 },
					{ geometry: new TeapotGeometry( 1, 5 ), error: 0.03 },
					{ geometry: new TeapotGeometry( 1, 4 ), error: 0.06 },
					{ geometry: new TeapotGeometry( 1, 3 ), error: 0.1 },
					{ geometry: new TeapotGeometry( 1, 2 ), error: 0.2 }
				];

				let totalVertices = 0;
				let totalIndices = 0;

				for ( const lod of lods ) {

					const geom = lod.geometry;
					const pos = geom.attributes.position;
					const uvs = geom.attributes.uv;
					const idx = geom.index ? Array.from( geom.index.array ) : Array.from( { length: pos.count }, ( _, i ) => i );

					lod.numVertices = pos.count;
					lod.numTriangles = idx.length / 3;
					lod.vertexOffset = totalVertices;
					lod.indexOffset = totalIndices;
					lod.positions = pos;
					lod.uvs = uvs;
					lod.indices = idx;

					totalVertices += pos.count;
					totalIndices += idx.length;

				}

				const maxTrianglesPerInstance = lods[ 0 ].numTriangles;
				const totalTriangles = rows * cols * maxTrianglesPerInstance;
				document.getElementById( 'triangleCount' ).innerText = new Intl.NumberFormat().format( totalTriangles );

				const vertexArray = new Float32Array( totalVertices * 4 ); // vec4 padded
				const uvArray = new Float32Array( totalVertices * 2 );
				const indexArray = new Uint32Array( totalIndices );
				const meshletTriangleArray = new Uint32Array( totalIndices / 3 ); // 1 meshlet ID per triangle

				let currentMeshletId = 1;

				for ( const lod of lods ) {

					for ( let i = 0; i < lod.numVertices; i ++ ) {

						const vIdx = lod.vertexOffset + i;
						vertexArray[ vIdx * 4 + 0 ] = lod.positions.getX( i );
						vertexArray[ vIdx * 4 + 1 ] = lod.positions.getY( i );
						vertexArray[ vIdx * 4 + 2 ] = lod.positions.getZ( i );
						vertexArray[ vIdx * 4 + 3 ] = 1.0;

						if ( lod.uvs ) {

							uvArray[ vIdx * 2 + 0 ] = lod.uvs.getX( i );
							uvArray[ vIdx * 2 + 1 ] = lod.uvs.getY( i );

						}

					}

					let currentTriCount = 0;
					for ( let i = 0; i < lod.numTriangles; i ++ ) {

						const triIdx = ( lod.indexOffset / 3 ) + i;
						indexArray[ triIdx * 3 + 0 ] = lod.vertexOffset + lod.indices[ i * 3 + 0 ];
						indexArray[ triIdx * 3 + 1 ] = lod.vertexOffset + lod.indices[ i * 3 + 1 ];
						indexArray[ triIdx * 3 + 2 ] = lod.vertexOffset + lod.indices[ i * 3 + 2 ];

						if ( currentTriCount >= 126 ) {

							currentMeshletId ++;
							currentTriCount = 0;

						}

						meshletTriangleArray[ triIdx ] = currentMeshletId;
						currentTriCount ++;

					}

					currentMeshletId ++;

				}

				// Precompute Bounding Spheres for each 64-triangle Chunk (Cluster)
				let totalChunks = 0;
				for ( const lod of lods ) {

					lod.numChunks = Math.ceil( lod.numTriangles / 64 );
					lod.chunkStart = totalChunks;
					totalChunks += lod.numChunks;

				}

				const chunkBoundsData = new Float32Array( totalChunks * 4 ); // vec4: cx, cy, cz, radius
				let currentChunkId = 0;

				for ( const lod of lods ) {

					const positions = lod.positions;
					const indices = lod.indices;

					for ( let c = 0; c < lod.numChunks; c ++ ) {

						const startTri = c * 64;
						const endTri = Math.min( startTri + 64, lod.numTriangles );

						// 1. Calculate Center
						let cx = 0, cy = 0, cz = 0;
						const vertCount = ( endTri - startTri ) * 3;
						for ( let t = startTri; t < endTri; t ++ ) {

							for ( let v = 0; v < 3; v ++ ) {

								const idx = indices[ t * 3 + v ];
								cx += positions.getX( idx );
								cy += positions.getY( idx );
								cz += positions.getZ( idx );

							}

						}

						cx /= vertCount;
						cy /= vertCount;
						cz /= vertCount;

						// 2. Calculate Radius
						let maxDistSq = 0;
						for ( let t = startTri; t < endTri; t ++ ) {

							for ( let v = 0; v < 3; v ++ ) {

								const idx = indices[ t * 3 + v ];
								const dx = positions.getX( idx ) - cx;
								const dy = positions.getY( idx ) - cy;
								const dz = positions.getZ( idx ) - cz;
								const distSq = dx * dx + dy * dy + dz * dz;
								if ( distSq > maxDistSq ) maxDistSq = distSq;

							}

						}

						const radius = Math.sqrt( maxDistSq );

						chunkBoundsData[ currentChunkId * 4 + 0 ] = cx;
						chunkBoundsData[ currentChunkId * 4 + 1 ] = cy;
						chunkBoundsData[ currentChunkId * 4 + 2 ] = cz;
						chunkBoundsData[ currentChunkId * 4 + 3 ] = radius;
						currentChunkId ++;

					}

				}

				// Upload LOD offsets to GPU (uvec4: triangleStart, numTriangles, chunkStart, 0)
				const lodOffsetsData = new Uint32Array( lods.length * 4 );
				for ( let i = 0; i < lods.length; i ++ ) {

					lodOffsetsData[ i * 4 + 0 ] = lods[ i ].indexOffset / 3;
					lodOffsetsData[ i * 4 + 1 ] = lods[ i ].numTriangles;
					lodOffsetsData[ i * 4 + 2 ] = lods[ i ].chunkStart;

				}

				const lodOffsetsBuffer = storage( new THREE.StorageBufferAttribute( lodOffsetsData, 4 ), 'uvec4', lods.length ).toReadOnly();
				const chunkBoundsBuffer = storage( new THREE.StorageBufferAttribute( chunkBoundsData, 4 ), 'vec4', totalChunks ).toReadOnly();

				// Storage Buffers
				const vertexBuffer = storage( new THREE.StorageBufferAttribute( vertexArray, 4 ), 'vec4', totalVertices ).toReadOnly();
				const uvBuffer = storage( new THREE.StorageBufferAttribute( uvArray, 2 ), 'vec2', totalVertices ).toReadOnly();
				const indexBuffer = storage( new THREE.StorageBufferAttribute( indexArray, 1 ), 'uint', totalIndices ).toReadOnly();
				const meshletIdBuffer = storage( new THREE.StorageBufferAttribute( meshletTriangleArray, 1 ), 'uint', totalIndices / 3 ).toReadOnly();

				const materialModeUniform = uniform( 0, 'uint' );

				const textureMap = new THREE.TextureLoader().load( 'textures/uv_grid_directx.jpg' );
				textureMap.colorSpace = THREE.SRGBColorSpace;
				textureMap.wrapS = THREE.RepeatWrapping;
				textureMap.wrapT = THREE.RepeatWrapping;

				const timeScale = uniform( 1.0 );

				const parameterGroup = renderer.inspector.createParameters( 'Parameters' );
				parameterGroup.add( options, 'Mode', { 'Meshlet Debug': 'Meshlet Debug', 'Texture': 'Texture' } ).addEventListener( 'change', ( e ) => {

					materialModeUniform.value = e.value === 'Texture' ? 1 : 0;

				} );

				parameterGroup.add( options, 'Rasterizer', { 'SW Only': 'SW Only', 'HW Only': 'HW Only', 'Both': 'Both' } );

				parameterGroup.add( timeScale, 'value', 0.0, 1.0 ).name( 'Animation Speed' );

				// Visibility buffers
				// screenTri: 32-bit depth — the closest fragment wins via atomicMax
				// screenInst: payload — instId (18 high bits) | megaTriangleIndex (14 low bits),
				// written by the depth test winner (best effort; not atomic with the depth update)
				createScreenBuffers();

				const staticInstanceData = new Float32Array( instanceCount * 4 );
				let dataIndex = 0;

				for ( let i = 0; i < rows; i ++ ) {

					for ( let j = 0; j < cols; j ++ ) {

						staticInstanceData[ dataIndex ++ ] = ( i - rows / 2 ) * 4.0;
						staticInstanceData[ dataIndex ++ ] = - 1;
						staticInstanceData[ dataIndex ++ ] = ( j - cols / 2 ) * 4.0;
						staticInstanceData[ dataIndex ++ ] = 1.0; // scale

					}

				}

				const instanceDataBuffer = storage( new THREE.StorageBufferAttribute( staticInstanceData, 4 ), 'vec4', instanceCount );

				const instanceWorldData = new Float32Array( instanceCount * 16 );
				const instanceMvpData = new Float32Array( instanceCount * 16 );

				const instanceWorldAttr = new THREE.StorageBufferAttribute( instanceWorldData, 16 );
				const instanceMvpAttr = new THREE.StorageBufferAttribute( instanceMvpData, 16 );

				const instanceWorldBuffer = storage( instanceWorldAttr, 'mat4', instanceCount );
				const instanceMvpBuffer = storage( instanceMvpAttr, 'mat4', instanceCount );
				const instanceWorldRead = storage( instanceWorldAttr, 'mat4', instanceCount ).toReadOnly();

				const workQueueCountData = new Uint32Array( 1 );
				const workQueueCountAttr = new THREE.StorageBufferAttribute( workQueueCountData, 1 );
				const workQueueCountAtomic = storage( workQueueCountAttr, 'uint', 1 ).toAtomic();
				const workQueueCountRead = storage( workQueueCountAttr, 'uint', 1 ).toReadOnly();

				const dispatchData = new Uint32Array( 3 );
				const dispatchAttr = new THREE.IndirectStorageBufferAttribute( dispatchData, 3 );
				const dispatchBuffer = storage( dispatchAttr, 'uint', 3 );

				// Work queue budget — one item is a 64-triangle chunk of one visible instance
				const MAX_WORK_ITEMS = 2820000;
				const workQueueData = new Uint32Array( MAX_WORK_ITEMS * 4 );
				const workQueueBuffer = storage( new THREE.StorageBufferAttribute( workQueueData, 4 ), 'uvec4', MAX_WORK_ITEMS );

				// HW Rasterizer Buffers (for large triangles that exceed SW raster budget)
				const MAX_HW_TRIANGLES = 100000;

				// HW queue: index 0 is atomic counter, indices 1..MAX store payload32
				const hwQueueData = new Uint32Array( MAX_HW_TRIANGLES + 1 );
				const hwQueueAttr = new THREE.StorageBufferAttribute( hwQueueData, 1 );
				const hwQueueAtomic = storage( hwQueueAttr, 'uint', MAX_HW_TRIANGLES + 1 ).toAtomic();
				const hwQueueRead = storage( hwQueueAttr, 'uint', MAX_HW_TRIANGLES + 1 ).toReadOnly();

				// Draw indirect buffer: vertexCount, instanceCount, firstVertex, firstInstance
				const hwDrawData = new Uint32Array( 4 );
				const hwDrawAttr = new THREE.IndirectStorageBufferAttribute( hwDrawData, 4 );
				const hwDrawBuffer = storage( hwDrawAttr, 'uint', 4 );

				projScreenMatrixUniform = uniform( new THREE.Matrix4() );
				frustumPlanesUniform = uniformArray( [
					new THREE.Vector4(), new THREE.Vector4(), new THREE.Vector4(),
					new THREE.Vector4(), new THREE.Vector4(), new THREE.Vector4()
				], 'vec4' );
				cameraPos = uniform( new THREE.Vector3() );
				cotHalfFovUniform = uniform( 1.0 );
				const pixelErrorThresholdUniform = uniform( 4.0 );
				const maxRasterSizeUniform = uniform( MAX_RASTER_SIZE, 'int' ); // Max bounding box size in pixels for SW rasterizer

				// Compute Clear
				computeClear = Fn( () => {

					atomicStore( screenTriAtomic.element( instanceIndex ), uint( 0 ) );
					screenInstBuffer.element( instanceIndex ).assign( uint( 0 ) );

					If( instanceIndex.equal( 0 ), () => {

						atomicStore( workQueueCountAtomic.element( 0 ), uint( 0 ) );
						atomicStore( hwQueueAtomic.element( 0 ), uint( 0 ) );

					} );

				} )().compute( maxPixels, [ 256 ] ).setName( 'Compute Clear' );

				// Compute Frustum (GPU Culling, LOD & Work Allocation)
				computeFrustum = Fn( () => {

					const data = instanceDataBuffer.element( instanceIndex );
					const pos = data.xyz;
					const scale = data.w;
					const i = float( instanceIndex );

					// Rotation
					const rotY = time.mul( timeScale ).add( i );
					const c = cos( rotY );
					const s = sin( rotY );

					// Compose MatrixWorld
					const matrixWorld = mat4(
						vec4( c.mul( scale ), 0.0, s.mul( scale ), 0.0 ),
						vec4( 0.0, scale, 0.0, 0.0 ),
						vec4( s.negate().mul( scale ), 0.0, c.mul( scale ), 0.0 ),
						vec4( pos, 1.0 )
					);

					const visible = bool( true ).toVar();
					const radius = scale.mul( 2.0 ); // bounding sphere radius

					// Frustum culling using the 6 extracted world-space planes
					Loop( { start: 0, end: 6 }, ( { i: planeIndex } ) => {

						const plane = frustumPlanesUniform.element( planeIndex );
						const dist = dot( plane.xyz, pos ).add( plane.w );

						If( dist.lessThan( radius.negate() ), () => {

							visible.assign( false );

						} );

					} );

					If( visible, () => {

						const distToCamera = distance( cameraPos, pos );

						// Precompute projection factor once (Screen-Space Projected Error)
						// pixelError = cotHalfFov * errorWorld / dist * screenH / 2
						const pixelFactor = cotHalfFovUniform.div( max( 0.01, distToCamera ) ).mul( float( screenSize.y ) ).div( 2.0 );

						const lodLevel = uint( 0 ).toVar();

						let lodSelection = null;
						for ( let i = lods.length - 1; i > 0; i -- ) {

							const checkLod = float( lods[ i ].error ).mul( scale ).mul( pixelFactor ).lessThanEqual( pixelErrorThresholdUniform );

							if ( lodSelection === null ) {

								lodSelection = If( checkLod, () => {

									lodLevel.assign( i );

								} );

							} else {

								lodSelection = lodSelection.ElseIf( checkLod, () => {

									lodLevel.assign( i );

								} );

							}

						}

						const lodData = lodOffsetsBuffer.element( lodLevel );
						const lodTriStart = lodData.x;
						const lodNumTriangles = lodData.y;
						const lodChunkStart = lodData.z;

						// Calculate Work Items (64 triangles per item)
						const workItems = lodNumTriangles.add( 63 ).div( 64 );

						// Evaluate each Chunk (Cluster)
						Loop( { name: 'cIdx', type: 'uint', start: uint( 0 ), end: workItems, condition: '<' }, ( { cIdx: chunkIndex } ) => {

							const globalChunkId = lodChunkStart.add( uint( chunkIndex ) );
							const chunkBounds = chunkBoundsBuffer.element( globalChunkId );
							const chunkCenterLocal = chunkBounds.xyz;
							const chunkRadiusLocal = chunkBounds.w;

							// Transform chunk bounding sphere to world space and store as var to prevent inlining
							const chunkCenterWorld = matrixWorld.mul( vec4( chunkCenterLocal, 1.0 ) ).xyz.toVar();
							const chunkRadiusWorld = chunkRadiusLocal.mul( scale ).toVar();

							const chunkVisible = bool( true ).toVar();

							// Frustum cull the chunk
							Loop( { name: 'pIdx', start: 0, end: 6 }, ( { pIdx: planeIndex } ) => {

								const plane = frustumPlanesUniform.element( planeIndex );
								const chunkDist = dot( plane.xyz, chunkCenterWorld ).add( plane.w );

								If( chunkDist.lessThan( chunkRadiusWorld.negate() ), () => {

									chunkVisible.assign( false );

								} );

							} );

							If( chunkVisible, () => {

								const itemIndex = atomicAdd( workQueueCountAtomic.element( 0 ), 1 );

								// uvec4( instanceIndex, triangleStart, lodNumTriangles, chunkIndex )
								workQueueBuffer.element( itemIndex ).assign(
									uvec4( instanceIndex, lodTriStart, lodNumTriangles, uint( chunkIndex ) )
								);

							} );

						} );

						// Store transform for this instance
						instanceWorldBuffer.element( instanceIndex ).assign( matrixWorld );
						instanceMvpBuffer.element( instanceIndex ).assign( projScreenMatrixUniform.mul( matrixWorld ) );

					} );

				} )().compute( instanceCount ).setName( 'Compute Frustum' );

				// Compute Dispatch (Indirect arguments)
				computeDispatch = Fn( () => {

					const totalWorkgroups = workQueueCountRead.element( 0 );

					const maxDim = uint( 65535 );

					// Split totalWorkgroups into 2D dispatch if it exceeds 65535
					const dispatchX = min( totalWorkgroups, maxDim );
					const dispatchY = totalWorkgroups.add( maxDim ).sub( 1 ).div( maxDim );

					dispatchBuffer.element( 0 ).assign( dispatchX );
					dispatchBuffer.element( 1 ).assign( dispatchY );
					dispatchBuffer.element( 2 ).assign( 1 );

				} )().compute( 1 ).setName( 'Compute Dispatch' );

				// Edge function for barycentric coordinates
				const edgeFunction = Fn( ( [ a, b, c ] ) => {

					// (c.y - a.y) * (b.x - a.x) - (c.x - a.x) * (b.y - a.y)
					return c.y.sub( a.y ).mul( b.x.sub( a.x ) ).sub( c.x.sub( a.x ).mul( b.y.sub( a.y ) ) );

				} );

				// Compute Rasterizer
				computeRasterize = Fn( () => {

					const totalWorkgroups = workQueueCountRead.element( 0 );
					const totalThreads = totalWorkgroups.mul( 64 );

					If( instanceIndex.lessThan( totalThreads ), () => {

						const workItemId = instanceIndex.div( 64 );
						const localTriangleIndex = instanceIndex.mod( 64 );

						const workItem = workQueueBuffer.element( workItemId );
						const instId = workItem.x;
						const lodTriStart = workItem.y;
						const lodNumTriangles = workItem.z;
						const chunkIndex = workItem.w;

						const globalTriangleIndex = chunkIndex.mul( 64 ).add( localTriangleIndex );

						If( globalTriangleIndex.lessThan( lodNumTriangles ), () => {

							const megaTriangleIndex = lodTriStart.add( globalTriangleIndex );
							const indexOffset = megaTriangleIndex.mul( 3 );

							const i0 = indexBuffer.element( indexOffset );
							const i1 = indexBuffer.element( indexOffset.add( 1 ) );
							const i2 = indexBuffer.element( indexOffset.add( 2 ) );

							const v0 = vertexBuffer.element( i0 );
							const v1 = vertexBuffer.element( i1 );
							const v2 = vertexBuffer.element( i2 );

							const instMvpMatrix = instanceMvpBuffer.element( instId );

							// MVP
							const p0 = instMvpMatrix.mul( v0 );
							const p1 = instMvpMatrix.mul( v1 );
							const p2 = instMvpMatrix.mul( v2 );

							// Near plane clipping
							If( p0.w.greaterThan( 0.0 ).and( p1.w.greaterThan( 0.0 ) ).and( p2.w.greaterThan( 0.0 ) ), () => {

								const ndc0 = p0.xyz.div( p0.w );
								const ndc1 = p1.xyz.div( p1.w );
								const ndc2 = p2.xyz.div( p2.w );

								// Early Backface Culling in NDC
								const areaNdc = edgeFunction( ndc0, ndc1, ndc2 );

								If( areaNdc.greaterThan( 0.0 ), () => {

									// NDC guard: skip triangles entirely outside clip volume
									const ndcMinX = min( ndc0.x, min( ndc1.x, ndc2.x ) );
									const ndcMaxX = max( ndc0.x, max( ndc1.x, ndc2.x ) );
									const ndcMinY = min( ndc0.y, min( ndc1.y, ndc2.y ) );
									const ndcMaxY = max( ndc0.y, max( ndc1.y, ndc2.y ) );

									If( ndcMaxX.greaterThan( - 1.0 ).and( ndcMinX.lessThan( 1.0 ) ).and( ndcMaxY.greaterThan( - 1.0 ) ).and( ndcMinY.lessThan( 1.0 ) ), () => {

										// Map to screen coordinates
										const w = screenSize.x;
										const h = screenSize.y;
										const s0 = ndc0.xy.add( 1.0 ).mul( 0.5 ).mul( vec2( w, h ) );
										const s1 = ndc1.xy.add( 1.0 ).mul( 0.5 ).mul( vec2( w, h ) );
										const s2 = ndc2.xy.add( 1.0 ).mul( 0.5 ).mul( vec2( w, h ) );

										// Bounding Box
										const minX = max( 0.0, min( s0.x, min( s1.x, s2.x ) ) );
										const maxX = min( w.sub( 1.0 ), max( s0.x, max( s1.x, s2.x ) ) );
										const minY = max( 0.0, min( s0.y, min( s1.y, s2.y ) ) );
										const maxY = min( h.sub( 1.0 ), max( s0.y, max( s1.y, s2.y ) ) );

										const startX = int( floor( minX ) );
										const endX = int( floor( maxX ) );
										const startY = int( floor( minY ) );
										const endY = int( floor( maxY ) );

										// Big triangle guard: skip triangles larger than maxRasterSize
										// This is the key performance safeguard — software rasterizers
										// should only handle small triangles. Large triangles cause O(n²)
										// pixel iteration per thread, which kills performance when close.
										const bbWidth = endX.sub( startX );
										const bbHeight = endY.sub( startY );

										// Compute payload32 for HW path (full precision)
										// payload32: instId (18 bits) | megaTriangleIndex (14 bits)
										const payload32 = instId.shiftLeft( TRIANGLE_INDEX_BITS ).bitOr( megaTriangleIndex.bitAnd( TRIANGLE_INDEX_MASK ) );

										// Sub-pixel / Valid bounds rejection + big triangle guard
										If( startX.lessThanEqual( endX ).and( startY.lessThanEqual( endY ) ).and( bbWidth.lessThanEqual( maxRasterSizeUniform ) ).and( bbHeight.lessThanEqual( maxRasterSizeUniform ) ), () => {

											const area = edgeFunction( s0, s1, s2 );

											const stepX_w0 = s1.y.sub( s2.y );
											const stepY_w0 = s2.x.sub( s1.x );

											const stepX_w1 = s2.y.sub( s0.y );
											const stepY_w1 = s0.x.sub( s2.x );

											const stepX_w2 = s0.y.sub( s1.y );
											const stepY_w2 = s1.x.sub( s0.x );

											// Top-Left rule check for each edge to guarantee watertightness
											const isTopLeft0 = stepX_w0.lessThan( 0.0 ).or( stepX_w0.equal( 0.0 ).and( stepY_w0.greaterThan( 0.0 ) ) );
											const isTopLeft1 = stepX_w1.lessThan( 0.0 ).or( stepX_w1.equal( 0.0 ).and( stepY_w1.greaterThan( 0.0 ) ) );
											const isTopLeft2 = stepX_w2.lessThan( 0.0 ).or( stepX_w2.equal( 0.0 ).and( stepY_w2.greaterThan( 0.0 ) ) );

											const bias0 = isTopLeft0.select( 0.0, - 1e-5 );
											const bias1 = isTopLeft1.select( 0.0, - 1e-5 );
											const bias2 = isTopLeft2.select( 0.0, - 1e-5 );

											const pStart = vec2( float( startX ).add( 0.5 ), float( startY ).add( 0.5 ) );

											const row_w0 = edgeFunction( s1, s2, pStart ).toVar();
											const row_w1 = edgeFunction( s2, s0, pStart ).toVar();
											const row_w2 = edgeFunction( s0, s1, pStart ).toVar();

											row_w0.addAssign( bias0 );
											row_w1.addAssign( bias1 );
											row_w2.addAssign( bias2 );

											// Incremental Z Math (ALU Optimization)
											const b0_start = row_w0.div( area );
											const b1_start = row_w1.div( area );
											const b2_start = row_w2.div( area );
											const row_z = b0_start.mul( ndc0.z ).add( b1_start.mul( ndc1.z ) ).add( b2_start.mul( ndc2.z ) ).toVar();

											const stepX_z = stepX_w0.div( area ).mul( ndc0.z ).add( stepX_w1.div( area ).mul( ndc1.z ) ).add( stepX_w2.div( area ).mul( ndc2.z ) );
											const stepY_z = stepY_w0.div( area ).mul( ndc0.z ).add( stepY_w1.div( area ).mul( ndc1.z ) ).add( stepY_w2.div( area ).mul( ndc2.z ) );

											Loop( { name: 'y', type: 'int', start: startY, end: endY, condition: '<=' }, ( { y } ) => {

												const w0 = row_w0.toVar();
												const w1 = row_w1.toVar();
												const w2 = row_w2.toVar();
												const z = row_z.toVar();

												Loop( { name: 'x', type: 'int', start: startX, end: endX, condition: '<=' }, ( { x } ) => {

													If( w0.greaterThanEqual( 0.0 ).and( w1.greaterThanEqual( 0.0 ) ).and( w2.greaterThanEqual( 0.0 ) ), () => {

														If( z.greaterThanEqual( 0.0 ).and( z.lessThanEqual( 1.0 ) ), () => {

															// Calculate 32-bit depth value (fourth-root distribution to maximize depth precision)
															const depth32 = uint( sqrt( sqrt( float( 1.0 ).sub( z ) ) ).mul( DEPTH_PRECISION_MAX ) );

															const pixelIndex = uint( y ).mul( uint( screenSize.x ) ).add( uint( x ) );

															// Early depth pre-check: skip atomicMax if pixel already has closer fragment
															const currentDepth = atomicLoad( screenTriAtomic.element( pixelIndex ) );
															If( depth32.greaterThan( currentDepth ), () => {

																// Atomic depth test
																const prevDepth = atomicMax( screenTriAtomic.element( pixelIndex ), depth32 );

																// If we successfully wrote the closest depth, write the payload
																If( depth32.greaterThan( prevDepth ), () => {

																	screenInstBuffer.element( pixelIndex ).assign( payload32 );

																} );

															} );

														} );

													} );

													w0.addAssign( stepX_w0 );
													w1.addAssign( stepX_w1 );
													w2.addAssign( stepX_w2 );
													z.addAssign( stepX_z );

												} );

												row_w0.addAssign( stepY_w0 );
												row_w1.addAssign( stepY_w1 );
												row_w2.addAssign( stepY_w2 );
												row_z.addAssign( stepY_z );

											} );

										} ).Else( () => {

											// Big triangle → enqueue for HW rasterization
											If( startX.lessThanEqual( endX ).and( startY.lessThanEqual( endY ) ), () => {

												const hwCount = atomicAdd( hwQueueAtomic.element( 0 ), 1 );
												const hwSlot = hwCount.add( 1 );
												atomicStore( hwQueueAtomic.element( hwSlot ), payload32 );

											} );

										} );

									} );

								} ); // End Early Backface Culling

							} ); // End Near Plane Clipping

						} ); // End globalTriangleIndex bounds check

					} ); // End instanceIndex bounds check

				} )().compute( dispatchAttr ).setName( 'Compute Rasterize' );

				// Compute HW Draw Indirect Args
				computeHWArgs = Fn( () => {

					const hwCount = atomicLoad( hwQueueAtomic.element( 0 ) );

					// Non-indexed draw: vertexCount = hwCount * 3 (3 verts per triangle)
					hwDrawBuffer.element( 0 ).assign( hwCount.mul( 3 ) ); // vertexCount
					hwDrawBuffer.element( 1 ).assign( uint( 1 ) ); // instanceCount
					hwDrawBuffer.element( 2 ).assign( uint( 0 ) ); // firstVertex
					hwDrawBuffer.element( 3 ).assign( uint( 0 ) ); // firstInstance

				} )().compute( 1 ).setName( 'Compute HW Args' );

				// Hash function for meshlet colors (shared between HW mesh and fullscreen quad)
				const hashColor = Fn( ( [ id_in ] ) => {

					let id = uint( id_in ).toVar();
					id = id.mul( uint( 747796405 ) ).add( uint( 289559509 ) );
					id = id.shiftRight( 16 ).bitXor( id ).mul( uint( 277803737 ) );
					id = id.shiftRight( 16 ).bitXor( id );

					const r = float( id.bitAnd( uint( 255 ) ) ).div( 255.0 );
					const g = float( id.shiftRight( 8 ).bitAnd( uint( 255 ) ) ).div( 255.0 );
					const b = float( id.shiftRight( 16 ).bitAnd( uint( 255 ) ) ).div( 255.0 );

					return vec4( r.mul( 0.8 ).add( 0.2 ), g.mul( 0.8 ).add( 0.2 ), b.mul( 0.8 ).add( 0.2 ), 1.0 );

				} );

				// HW Rasterizer Mesh (renders big triangles via GPU hardware pipeline)
				// Unlike the SW rasterizer which writes to an atomic screen buffer,
				// the HW mesh renders directly with real colors and hardware depth testing.
				// It renders AFTER the fullscreen quad, overlaying HW-rasterized triangles.
				{

					// Geometry: dummy positions, vertex count driven by indirect draw
					const hwGeometry = new THREE.BufferGeometry();
					hwGeometry.setAttribute( 'position', new THREE.Float32BufferAttribute( new Float32Array( MAX_HW_TRIANGLES * 3 * 3 ), 3 ) );
					hwGeometry.setIndirect( hwDrawAttr );
					hwGeometry.boundingSphere = new THREE.Sphere().set( new THREE.Vector3(), Infinity );

					// Varying to pass payload and UVs from vertex to fragment
					const vPayload = varyingProperty( 'uint', 'vPayload' );
					const vUv = varyingProperty( 'vec2', 'vUv' );

					const hwMaterial = new THREE.NodeMaterial();
					hwMaterial.depthWrite = true;
					hwMaterial.depthTest = true;

					// Vertex shader: vertex pulling from HW queue
					hwMaterial.positionNode = Fn( () => {

						// vertexIndex: 0,1,2, 3,4,5, 6,7,8, ...
						const triIndex = vertexIndex.div( 3 ); // which triangle in HW queue
						const localVert = vertexIndex.mod( 3 ); // which vertex (0, 1, 2)

						const payload32 = hwQueueRead.element( triIndex.add( 1 ) );
						const instId = payload32.shiftRight( TRIANGLE_INDEX_BITS );
						const megaTriIdx = payload32.bitAnd( TRIANGLE_INDEX_MASK );

						// Fetch actual vertex index from the mega index buffer
						const vertGlobalIdx = indexBuffer.element( megaTriIdx.mul( 3 ).add( localVert ) );
						const v = vertexBuffer.element( vertGlobalIdx );

						// Transform to world space
						const worldPos = instanceWorldRead.element( instId ).mul( v );

						const uvVal = uvBuffer.element( vertGlobalIdx );
						vUv.assign( uvVal );

						vPayload.assign( payload32 );

						return worldPos.xyz;

					} )();

					// Fragment shader: directly output final color (no storage buffer writes)
					hwMaterial.fragmentNode = Fn( () => {

						const payload32 = vPayload;
						const instId = payload32.shiftRight( TRIANGLE_INDEX_BITS );
						const megaTriangleIndex = payload32.bitAnd( TRIANGLE_INDEX_MASK );

						const outColor = vec4( 0.0 ).toVar();

						If( materialModeUniform.equal( 0 ), () => {

							const meshletId = meshletIdBuffer.element( megaTriangleIndex ).add( instId.mul( 1000 ) );
							outColor.assign( hashColor( meshletId ) );

						} ).Else( () => {

							// Hardware interpolated UV!
							outColor.assign( texture( textureMap, vUv ) );

						} );

						return outColor;

					} )();

					hwMesh = new THREE.Mesh( hwGeometry, hwMaterial );
					hwMesh.frustumCulled = false;

					hwScene = new THREE.Scene();
					hwScene.add( hwMesh );

				}


				// Fullscreen Presentation Pass
				const material = new THREE.NodeMaterial();
				material.depthWrite = true;

				// Shared screen-coordinate helper
				const getPixelIndex = () => {

					const screenX = uint( floor( uv().x.mul( screenSize.x ) ) );
					const screenY = uint( floor( uv().y.oneMinus().mul( screenSize.y ) ) );
					return screenY.mul( uint( screenSize.x ) ).add( screenX );

				};

				// Output depth from the SW rasterizer so HW mesh can depth test against it
				material.depthNode = Fn( () => {

					const pixelIndex = getPixelIndex();

					// Read 32-bit depth from buffer
					const depth32 = screenTriRead.element( pixelIndex );

					// Reconstruct NDC Z from non-linear depth32 (fourth-root distribution)
					const y = float( depth32 ).div( DEPTH_PRECISION_MAX );
					const y2 = y.mul( y );
					const v = y2.mul( y2 ); // raise to the fourth power (y^4) to get original v
					return float( 1.0 ).sub( v );

				} )();

				material.colorNode = Fn( () => {

					const pixelIndex = getPixelIndex();

					// Single buffer read — check for background immediately (using 32-bit depth)
					const depth32 = screenTriRead.element( pixelIndex );

					// Background color for pixels with no geometry
					const outColor = vec4( background, 1.0 ).toVar();

					If( depth32.greaterThan( 0 ), () => {

						// Read the single packed payload
						const payload32 = screenInstRead.element( pixelIndex );
						const megaTriangleIndex = payload32.bitAnd( TRIANGLE_INDEX_MASK );
						const instId = payload32.shiftRight( TRIANGLE_INDEX_BITS );

						// Visibility Buffer: Fetch exact vertices and UVs
						const i0 = indexBuffer.element( megaTriangleIndex.mul( 3 ).add( 0 ) );
						const i1 = indexBuffer.element( megaTriangleIndex.mul( 3 ).add( 1 ) );
						const i2 = indexBuffer.element( megaTriangleIndex.mul( 3 ).add( 2 ) );

						const v0 = vertexBuffer.element( i0 );
						const v1 = vertexBuffer.element( i1 );
						const v2 = vertexBuffer.element( i2 );

						const t_uv0 = uvBuffer.element( i0 );
						const t_uv1 = uvBuffer.element( i1 );
						const t_uv2 = uvBuffer.element( i2 );

						// Project Vertices to Screen Space
						const mvpMatrix = instanceMvpBuffer.element( instId );

						const p0 = mvpMatrix.mul( v0 );
						const p1 = mvpMatrix.mul( v1 );
						const p2 = mvpMatrix.mul( v2 );

						const ndc0 = p0.xyz.div( p0.w );
						const ndc1 = p1.xyz.div( p1.w );
						const ndc2 = p2.xyz.div( p2.w );

						const w = screenSize.x;
						const h = screenSize.y;
						const s0 = ndc0.xy.add( 1.0 ).mul( 0.5 ).mul( vec2( w, h ) );
						const s1 = ndc1.xy.add( 1.0 ).mul( 0.5 ).mul( vec2( w, h ) );
						const s2 = ndc2.xy.add( 1.0 ).mul( 0.5 ).mul( vec2( w, h ) );

						const p = vec2( uv().x.mul( screenSize.x ), uv().y.oneMinus().mul( screenSize.y ) );

						// Compute Barycentrics
						const area = edgeFunction( s0, s1, s2 );
						const w0 = edgeFunction( s1, s2, p );
						const w1 = edgeFunction( s2, s0, p );
						const w2 = edgeFunction( s0, s1, p );

						// Guard against division by zero for safe execution
						const safeArea = area.equal( 0.0 ).select( 1.0, area );
						const b0 = w0.div( safeArea );
						const b1 = w1.div( safeArea );
						const b2 = w2.div( safeArea );

						// Perspective correct UV interpolation (32-bit floats!)
						const z_inv = b0.div( p0.w ).add( b1.div( p1.w ) ).add( b2.div( p2.w ) );
						const safeZInv = z_inv.equal( 0.0 ).select( 1.0, z_inv );
						const b0_p = b0.div( p0.w ).div( safeZInv );
						const b1_p = b1.div( p1.w ).div( safeZInv );
						const b2_p = b2.div( p2.w ).div( safeZInv );

						const uv_interp = t_uv0.mul( b0_p ).add( t_uv1.mul( b1_p ) ).add( t_uv2.mul( b2_p ) );

						// Compute screen-space derivatives analytically (extremely clean, no helper fragment issues)
						const dw0_dx = s2.y.sub( s1.y );
						const dw1_dx = s0.y.sub( s2.y );
						const dw2_dx = s1.y.sub( s0.y );

						const dw0_dy = s1.x.sub( s2.x );
						const dw1_dy = s2.x.sub( s0.x );
						const dw2_dy = s0.x.sub( s1.x );

						const q0 = float( 1.0 ).div( p0.w );
						const q1 = float( 1.0 ).div( p1.w );
						const q2 = float( 1.0 ).div( p2.w );

						const sum_w_q = w0.mul( q0 ).add( w1.mul( q1 ) ).add( w2.mul( q2 ) );
						const safe_sum_w_q = sum_w_q.equal( 0.0 ).select( 1.0, sum_w_q );

						const dUvDx = (
							dw0_dx.mul( q0 ).mul( t_uv0.sub( uv_interp ) )
								.add( dw1_dx.mul( q1 ).mul( t_uv1.sub( uv_interp ) ) )
								.add( dw2_dx.mul( q2 ).mul( t_uv2.sub( uv_interp ) ) )
						).div( safe_sum_w_q );

						const dUvDy = (
							dw0_dy.mul( q0 ).mul( t_uv0.sub( uv_interp ) )
								.add( dw1_dy.mul( q1 ).mul( t_uv1.sub( uv_interp ) ) )
								.add( dw2_dy.mul( q2 ).mul( t_uv2.sub( uv_interp ) ) )
						).div( safe_sum_w_q );

						If( materialModeUniform.equal( 0 ), () => {

							const meshletId = meshletIdBuffer.element( megaTriangleIndex ).add( instId.mul( 1000 ) );
							outColor.assign( hashColor( meshletId ) );

						} ).Else( () => {

							outColor.assign( texture( textureMap, uv_interp ).grad( dUvDx, dUvDy ) );

						} );

					} );

					return outColor;

				} )();

				quadMesh = new THREE.QuadMesh( material );

				window.addEventListener( 'resize', onWindowResize );

			}

			function createScreenBuffers() {

				const size = new THREE.Vector2();
				renderer.getDrawingBufferSize( size );
				const newMaxPixels = size.x * size.y;

				if ( newMaxPixels === maxPixels ) return;

				maxPixels = newMaxPixels;

				if ( screenTriAttr ) screenTriAttr.dispose();
				if ( screenInstAttr ) screenInstAttr.dispose();

				const screenTriData = new Uint32Array( maxPixels );
				screenTriAttr = new THREE.StorageBufferAttribute( screenTriData, 1 );

				const screenInstData = new Uint32Array( maxPixels );
				screenInstAttr = new THREE.StorageBufferAttribute( screenInstData, 1 );

				if ( screenTriAtomic === undefined ) {

					screenTriAtomic = storage( screenTriAttr, 'uint', maxPixels ).toAtomic();
					screenTriRead = storage( screenTriAttr, 'uint', maxPixels ).toReadOnly();

					screenInstBuffer = storage( screenInstAttr, 'uint', maxPixels );
					screenInstRead = storage( screenInstAttr, 'uint', maxPixels ).toReadOnly();

				} else {

					screenTriAtomic.value = screenTriAttr;
					screenTriAtomic.bufferCount = maxPixels;

					screenTriRead.value = screenTriAttr;
					screenTriRead.bufferCount = maxPixels;

					screenInstBuffer.value = screenInstAttr;
					screenInstBuffer.bufferCount = maxPixels;

					screenInstRead.value = screenInstAttr;
					screenInstRead.bufferCount = maxPixels;

					computeClear.count = maxPixels;
					computeClear.dispose();

					computeRasterize.dispose();
					computeFrustum.dispose();
					computeDispatch.dispose();
					computeHWArgs.dispose();

					quadMesh.material.dispose();
					hwMesh.material.dispose();

				}

			}

			function onWindowResize() {

				camera.aspect = window.innerWidth / window.innerHeight;
				camera.updateProjectionMatrix();

				renderer.setSize( window.innerWidth, window.innerHeight );

				createScreenBuffers();

			}

			const frustum = new THREE.Frustum();
			const projScreenMatrix = new THREE.Matrix4();
			const cameraInverse = new THREE.Matrix4();

			function animate() {

				controls.update();

				camera.updateMatrixWorld();

				cameraInverse.copy( camera.matrixWorld ).invert();
				projScreenMatrix.multiplyMatrices( camera.projectionMatrix, cameraInverse );
				frustum.setFromProjectionMatrix( projScreenMatrix );

				// Update GPU uniforms
				projScreenMatrixUniform.value.copy( projScreenMatrix );
				cameraPos.value.copy( camera.position );
				cotHalfFovUniform.value = camera.projectionMatrix.elements[ 5 ];

				// Pack frustum planes into the uniform array
				const planes = frustum.planes;
				const planesArray = frustumPlanesUniform.array;
				for ( let i = 0; i < 6; i ++ ) {

					const p = planes[ i ];
					planesArray[ i ].set( p.normal.x, p.normal.y, p.normal.z, p.constant );

				}

				// Compute & Render
				renderer.compute( computeClear );
				renderer.compute( computeFrustum );
				renderer.compute( computeDispatch );
				renderer.compute( computeRasterize );
				renderer.compute( computeHWArgs );

				const rasterMode = options.Rasterizer;

				// SW presentation (fullscreen quad reads atomic buffer)
				if ( rasterMode === 'SW Only' || rasterMode === 'Both' ) {

					quadMesh.render( renderer );

				}

				// HW mesh renders with real depth testing + colors
				if ( rasterMode === 'HW Only' || rasterMode === 'Both' ) {

					hwScene.background = ( rasterMode === 'HW Only' ) ? background : null;
					renderer.autoClear = ( rasterMode === 'HW Only' );
					renderer.render( hwScene, camera );
					renderer.autoClear = true;

				}

			}

		</script>
	</body>
</html>