lythom/bvle-voxels
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions
bvle-voxels/CLAUDE.md

202 lines
10 KiB
Markdown
Raw Normal View History

Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
# BVLE Voxels - Prototype de Moteur Voxel Hybride
## Vue d'ensemble
Phase 2.2: MDI rendering with CPU-filled indirect args Replace per-chunk DrawInstanced loop with a single DrawInstancedIndirectCount. CPU fills indirect args buffer with same frustum+backface cull logic as Phase 2.1. Key discoveries: - Wicked Engine command signature includes push constant (20-byte stride, not 16) - SV_VertexID does not reliably include startVertexLocation with ExecuteIndirect - Solution: pack chunkIndex|(faceIndex<<16) in push constant, VS reconstructs quad offset from GPUChunkInfo lookup - No explicit DX12 barriers needed (implicit promotion from COMMON suffices) Also adds voxel_engine_spec.md and updates references from .docx to .md.
2026-03-25 22:07:22 +01:00
Prototype de moteur voxel basé sur **Wicked Engine** (MIT, C++17, DX12/Vulkan) pour valider les performances de rendu sur GPU moderne (AMD RDNA 2+ / Nvidia RTX 3060+). Le document de spécification complet est dans `voxel_engine_spec.md` à la racine du projet.
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
Cible : 60+ fps en 1440p, monde de 512x512x256 voxels visibles.
## Architecture
```
bvle-voxels/
├── CMakeLists.txt # Build CMake racine
├── engine/ # Wicked Engine (clone --depth 1, branche main)
│ └── WickedEngine/shaders/voxel/ # Nos shaders copiés ici pour compilation DXC
├── src/
│ ├── voxel/ # Bibliothèque VoxelEngine (static lib)
│ │ ├── VoxelTypes.h # Types fondamentaux (VoxelData, PackedQuad, MaterialDesc, ChunkPos)
│ │ ├── VoxelWorld.h/.cpp # Monde voxel (hashmap de chunks, génération procédurale)
Phase 5.1: Naive Surface Nets smooth rendering Implement CPU-side Naive Surface Nets for smooth voxel surfaces (SmoothStone, Snow) coexisting with blocky voxels (Grass, Dirt, Stone, Sand). Key features: - SmoothMesher with binary SDF, centroid vertex placement, per-axis boundary clamping to align with blocky grid at smooth↔blocky transitions - Cross-chunk connectivity: PAD=2 SDF grid, vertex range [-1, CHUNK_SIZE), canonical edge ownership (no duplicate triangles, no z-fighting) - Face normals oriented by edge axis+sign (robust with binary SDF, unlike SDF gradient dot or centroid sampling approaches) - Y-axis winding fix: sharing cells have different spatial arrangement, requiring opposite winding from X and Z axes - GPU mesher treats smooth neighbors as solid (no blocky faces toward smooth) - Material blending: primary (smooth-only) + secondary (all counts) per vertex - Dedicated shaders: voxelSmoothVS (vertex pulling t6) + voxelSmoothPS (triplanar + lerp blending between two materials) - Separate render pass with LoadOp::LOAD after voxels+topings - New materials: SmoothStone (mat 6), blocky Stone (mat 3) and Dirt patches added to world generation for boundary testing
2026-03-27 13:03:55 +01:00
│ │ ├── VoxelMesher.h/.cpp # Binary Greedy Mesher CPU + SmoothMesher (Naive Surface Nets)
Refactor: extract VoxelRTManager, DeferredGPUBuffer, decompose VoxelRenderPath - Extract DeferredGPUBuffer utility (staging→dirty→capacity GPU buffer pattern) - Extract VoxelRTManager from VoxelRenderer (~500 lines: BLAS/TLAS, RT shadows+AO) - Decompose VoxelRenderPath into CameraController, AnimationState, VoxelProfiler - Replace toping std::sort with O(n) counting sort by (type, variant) - Update CLAUDE.md architecture docs to reflect new file structure
2026-03-31 13:46:35 +02:00
│ │ ├── VoxelRenderer.h/.cpp# Renderer + VoxelRenderPath (CameraController, AnimationState, VoxelProfiler)
│ │ ├── VoxelRTManager.h/.cpp # Ray tracing: BLAS/TLAS lifecycle, shadows+AO dispatches
│ │ ├── DeferredGPUBuffer.h # Utilitaire staging→dirty→capacity GPU buffer upload
Phase 4.2: GPU toping rendering pipeline + winding/lighting fixes Add instanced rendering for toping bevels: dedicated shaders (voxelTopingVS/PS), PSO, GPU buffers (t4 vertices, t5 instances), per-group DrawInstanced in a separate render pass with LoadOp::LOAD. Fix inverted face winding (emitTri auto-winding condition flipped for CW front-facing), slope normals (use inward direction not outward), and PS lighting (negate sunDirection like voxelPS). Update CLAUDE.md with Phase 4.1/4.2 documentation.
2026-03-26 17:47:08 +01:00
│ │ └── TopingSystem.h/.cpp # Système de topings (biseaux décoratifs sur faces +Y)
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
│ └── app/
│ └── main.cpp # Point d'entrée Win32 + crash handler SEH
├── shaders/ # Sources HLSL des shaders voxel (copiés dans engine/ au build)
Phase 2.5: GPU meshing production pipeline + perf optimizations (80+ FPS) Replace CPU greedy mesher with GPU compute mesher as default rendering pipeline. Key optimizations identified via CPU profiling (ProfileAccum, 5s averages): - Fused regenerate+pack: parallel noise gen + memcpy in same jobsystem pass (6ms → 0ms) - VoxelData memcpy: sizeof(VoxelData)==2 enables direct memcpy instead of bit-shift loop (28ms → <1ms) - Dirty-skip: GPU dispatch/upload only when chunks change, not every frame - Animation: 2 fBm octaves + no caves in animation mode (54ms → 8ms) - Result: 80-110 FPS with 60Hz terrain animation, 700+ FPS static
2026-03-26 09:05:52 +01:00
│ ├── voxelCommon.hlsli # Root signature et CB partagés (inclus par tous les shaders)
│ ├── voxelVS.hlsl # Vertex shader (vertex pulling, triple-mode: CPU/MDI/GPU mesh)
│ ├── voxelPS.hlsl # Pixel shader (triplanar + lighting)
│ ├── voxelCullCS.hlsl # Compute shader frustum+backface cull (Phase 2.3)
Phase 4.2: GPU toping rendering pipeline + winding/lighting fixes Add instanced rendering for toping bevels: dedicated shaders (voxelTopingVS/PS), PSO, GPU buffers (t4 vertices, t5 instances), per-group DrawInstanced in a separate render pass with LoadOp::LOAD. Fix inverted face winding (emitTri auto-winding condition flipped for CW front-facing), slope normals (use inward direction not outward), and PS lighting (negate sunDirection like voxelPS). Update CLAUDE.md with Phase 4.1/4.2 documentation.
2026-03-26 17:47:08 +01:00
│ ├── voxelMeshCS.hlsl # Compute shader GPU mesher 1×1 (Phase 2.4-2.5)
│ ├── voxelTopingVS.hlsl # Vertex shader topings (instanced vertex pulling, t4/t5)
Phase 5.1: Naive Surface Nets smooth rendering Implement CPU-side Naive Surface Nets for smooth voxel surfaces (SmoothStone, Snow) coexisting with blocky voxels (Grass, Dirt, Stone, Sand). Key features: - SmoothMesher with binary SDF, centroid vertex placement, per-axis boundary clamping to align with blocky grid at smooth↔blocky transitions - Cross-chunk connectivity: PAD=2 SDF grid, vertex range [-1, CHUNK_SIZE), canonical edge ownership (no duplicate triangles, no z-fighting) - Face normals oriented by edge axis+sign (robust with binary SDF, unlike SDF gradient dot or centroid sampling approaches) - Y-axis winding fix: sharing cells have different spatial arrangement, requiring opposite winding from X and Z axes - GPU mesher treats smooth neighbors as solid (no blocky faces toward smooth) - Material blending: primary (smooth-only) + secondary (all counts) per vertex - Dedicated shaders: voxelSmoothVS (vertex pulling t6) + voxelSmoothPS (triplanar + lerp blending between two materials) - Separate render pass with LoadOp::LOAD after voxels+topings - New materials: SmoothStone (mat 6), blocky Stone (mat 3) and Dirt patches added to world generation for boundary testing
2026-03-27 13:03:55 +01:00
│ ├── voxelTopingPS.hlsl # Pixel shader topings (triplanar + directional lighting)
│ ├── voxelSmoothVS.hlsl # Vertex shader smooth Surface Nets (vertex pulling, t6)
Phase 6.1: RT infrastructure — MRT normals + BLAS/TLAS build - Normal render target (R16G16B16A16_SNORM) as MRT SV_TARGET1 in all 3 pixel shaders (voxelPS, voxelTopingPS, voxelSmoothPS) for future RT shadow/AO - BLAS extraction compute shader (voxelBLASExtractCS.hlsl): converts PackedQuad StructuredBuffer to float3 position buffer for DXR BLAS input - Blocky BLAS: single BLAS from all GPU-meshed quads (~1.5M triangles) - Smooth BLAS: single BLAS from smooth vertex buffer directly - TLAS: 2 instances (blocky + smooth), identity transforms, CreateBuffer2 with callback to avoid UpdateBuffer on RAY_TRACING flagged buffers - Fix: Wicked always accesses index_buffer in CreateRaytracingAccelerationStructure via to_internal() even for non-indexed geometry — provide dummy valid buffer
2026-03-28 14:48:11 +01:00
│ ├── voxelSmoothPS.hlsl # Pixel shader smooth (triplanar + material blending)
Phase 6.2: RT shadows — inline ray queries with BLAS/TLAS fix Add shadow compute shader (voxelShadowCS.hlsl) that traces rays toward the sun using DXR inline ray queries (RayQuery<>, SM 6.5). Shadows modulate voxelRT_ in-place via RWTexture2D (no extra render target). Key fixes to Phase 6.1 BLAS/TLAS infrastructure: - Sequential index buffer required: Wicked treats IndexCount=0 with non-null IndexBuffer as "0 indexed triangles" → empty BLAS - Memory barriers between BLAS→TLAS→RT: without GPUBarrier::Memory() the TLAS build races with BLAS builds, causing zero ray hits - inverseViewProjection added to VoxelCB for depth reconstruction F5 toggles shadows OFF→ON→DEBUG (red=hit, green=miss, blue=backface).
2026-03-28 20:01:18 +01:00
│ ├── voxelBLASExtractCS.hlsl # Compute shader BLAS position extraction (Phase 6.1)
Phase 6.3: RT ambient occlusion with bilateral blur - 8 cosine-weighted hemisphere rays per pixel (inline ray queries, SM 6.5) - Distance-weighted AO: quadratic falloff (1-hitT/aoRadius)² instead of binary hit/miss - World-space hash seed: voxel coord + tangent-plane frac position (stable, no flicker) - Bilateral blur pipeline: 2-pass separable (H+V), radius 6, depth+normal edge-stopping - 4-pass dispatch: shadow+rawAO → blur H → blur V → apply - AO written to separate R8_UNORM texture, blurred, then applied to color buffer - Debug mode (F5 x3): grayscale AO visualization
2026-03-29 09:31:19 +02:00
│ ├── voxelShadowCS.hlsl # Compute shader RT shadows + raw AO (inline ray queries, Phase 6.2+6.3)
│ ├── voxelAOBlurCS.hlsl # Compute shader bilateral AO blur (separable H/V, Phase 6.3)
Phase 7.1: stylized lighting — hemisphere ambient, colored shadows, rim light, tone mapping Wonderbox-inspired lighting overhaul across all 3 pixel shaders: - Hemisphere ambient (sky blue above, warm brown below) replaces flat ambient - RT shadows lerp toward blue-violet tint instead of plain darkening (factor 0.55) - Rim light (fresnel) with warm golden color on silhouettes (30% on vegetation) - Soft exponential tone mapping + saturation boost in final post-process pass - CB parameters for all lighting values (skyAmbient, groundAmbient, shadowTint, etc.) - Fog color/density centralized from CB instead of hardcoded per-shader - Screenshot mode (CLI "screenshot"): fixed camera, AO convergence, auto-capture - AO noise stability: world-space hash using voxel center + tangent-axis frac position - AO distance-weighted falloff: continuous occlusion values instead of binary hit/miss
2026-03-29 15:00:12 +02:00
│ └── voxelAOApplyCS.hlsl # Compute shader AO apply + tone mapping + saturation (Phase 6.3 + 7)
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
├── CLAUDE.md
└── TROUBLESHOOTING.md # Pièges techniques, debugging, APIs Wicked
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
```
## Build
### Prérequis
- CMake 3.19+ (`winget install Kitware.CMake`)
- Visual Studio 2022 Build Tools (`winget install Microsoft.VisualStudio.2022.BuildTools`)
- Windows SDK 10.0.26100+ (`winget install Microsoft.WindowsSDK.10.0.26100`)
### Commandes
```bash
# Configurer (depuis la racine du projet)
cmake -B build -G "Visual Studio 17 2022" -A x64 -DCMAKE_SYSTEM_VERSION=10.0.26100.0
# Compiler
cmake --build build --config Release --target BVLEVoxels --parallel
# Exécutable produit dans build/Release/BVLEVoxels.exe
```
Le SDK 10.0.26100 est requis car les headers DX12 (`d3dx12_check_feature_support.h`) fournis par Wicked Engine ne sont pas compatibles avec le SDK 22621.
### Post-build automatique (CMakeLists.txt)
Le build copie automatiquement :
1. `dxcompiler.dll` → à côté de l'exe (requis pour la compilation runtime des shaders)
2. `shaders/*.hlsl` → `engine/WickedEngine/shaders/voxel/` (pour que `LoadShader` les trouve via `SHADERSOURCEPATH`)
3. `engine/Content/` → à côté de l'exe (assets Wicked Engine)
## Intégration Wicked Engine
### Backend graphique
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
Wicked Engine utilise **DX12 par défaut sur Windows**, Vulkan sur Linux. Les shaders sont écrits en **HLSL** et compilés via DXC. Pour forcer Vulkan sur Windows, passer `"vulkan"` en argument de ligne de commande.
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
### Point d'entrée et architecture de rendu
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
`VoxelRenderPath` hérite de `wi::RenderPath3D`. Le rendu voxel utilise ses propres render targets (`voxelRT_`, `voxelDepth_`) et est exécuté dans `Render()` sur un **command list dédié**. Le résultat est composité dans `Compose()` via `wi::image::Draw()`.
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
**NE JAMAIS créer un render pass dans `Compose()`** : cette méthode est appelée à l'intérieur du render pass du swapchain. Imbriquer des render passes est interdit en D3D12.
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
```
Render() → RenderPath3D::Render() // Wicked rend sa scène
→ device->BeginCommandList() // Nouveau cmd list
→ renderer.render(cmd, ...) // Notre render pass (clear + draw voxels → voxelRT_)
Compose() → RenderPath3D::Compose() // Wicked affiche son résultat
→ wi::image::Draw(voxelRT_) // On overlay nos voxels par-dessus
```
La caméra est gérée manuellement dans `Update()` en écrivant directement `camera->Eye`, `camera->At` (direction LookTo), `camera->Up`.
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
> See `TROUBLESHOOTING.md` for the detailed Wicked API reference table, shader binding pitfalls, DX12 resource state management, and debugging guides.
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
## Détails d'implémentation
### VoxelData (16 bits)
```
[15:8] material ID (256 matériaux)
[7:4] flags (smooth, transparent, emissive, custom)
[3:0] metadata (orientation, variant)
```
### PackedQuad (64 bits = 8 octets par quad)
```
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
[5:0] position X (0-63) [23:18] width (1-32)
[11:6] position Y (0-63) [29:24] height (1-32)
[17:12] position Z (0-63) [32:30] face (0-5)
[40:33] material ID [48:41] blendMatID
[59:49] chunkIndex (11 bits) [63:60] blendEdges (4 bits)
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
```
### Binary Greedy Mesher (CPU, `VoxelMesher.cpp`)
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
Masques binaires par axe, face culling par shift/XOR, greedy merge rectangulaire par tranche de profondeur.
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
### Génération procédurale (`VoxelWorld.cpp`)
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
Perlin noise 3D, fBm 5 octaves (2 en animation), caves 3D, matériaux par altitude. Chunks Y=0..7. Animation 60 Hz via `regenerateAnimated()` parallélisé avec `wi::jobsystem`.
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
### Renderer (`VoxelRenderer.cpp`)
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
- **Triple-mode VS** : CPU path, MDI path, GPU mesh path (défaut)
- **GPU mesh** : compute shader `voxelMeshCS` → barrier UAV→SRV → `DrawInstanced` (readback 1-frame-delay)
- **Vertex pulling** via `SV_VertexID`, pas de vertex buffer classique
- **Per-chunk info** : `StructuredBuffer<GPUChunkInfo>` (80 bytes/chunk)
- **Height-based blending** (Phase 3) : PS lit `voxelDataBuffer` (t3), winner-takes-all heightmap, corner attenuation
- **Render targets propres** : `voxelRT_` (R8G8B8A8) + `voxelDepth_` (D32_FLOAT)
Refactor: extract VoxelRTManager, DeferredGPUBuffer, decompose VoxelRenderPath - Extract DeferredGPUBuffer utility (staging→dirty→capacity GPU buffer pattern) - Extract VoxelRTManager from VoxelRenderer (~500 lines: BLAS/TLAS, RT shadows+AO) - Decompose VoxelRenderPath into CameraController, AnimationState, VoxelProfiler - Replace toping std::sort with O(n) counting sort by (type, variant) - Update CLAUDE.md architecture docs to reflect new file structure
2026-03-31 13:46:35 +02:00
- **CPU profiling** : `VoxelProfiler` (21 `ProfileAccum`, moyennes toutes les 5s)
- **DeferredGPUBuffer** : utilitaire pour buffers GPU avec staging CPU, dirty flag, capacity-based growth (25% headroom)
- **VoxelRTManager** (`VoxelRTManager.h/.cpp`) : gère BLAS/TLAS, dispatches RT shadows+AO, isolé du renderer
- **VoxelRenderPath** décomposé en : `CameraController` (mouvement/souris), `AnimationState` (tick terrain), `VoxelProfiler`
- **Toping sort** : counting sort O(n) par (type, variant) au lieu de `std::sort`
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
## Phases de développement
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
### Phase 1 - Setup et meshing de base [FAIT]
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
Fork Wicked Engine, VoxelWorld procédural, Binary Greedy Mesher CPU (~300K quads), rendu vertex pulling, caméra libre, crash handler SEH.
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
Phase 2.5: GPU meshing production pipeline + perf optimizations (80+ FPS) Replace CPU greedy mesher with GPU compute mesher as default rendering pipeline. Key optimizations identified via CPU profiling (ProfileAccum, 5s averages): - Fused regenerate+pack: parallel noise gen + memcpy in same jobsystem pass (6ms → 0ms) - VoxelData memcpy: sizeof(VoxelData)==2 enables direct memcpy instead of bit-shift loop (28ms → <1ms) - Dirty-skip: GPU dispatch/upload only when chunks change, not every frame - Animation: 2 fBm octaves + no caves in animation mode (54ms → 8ms) - Result: 80-110 FPS with 60Hz terrain animation, 700+ FPS static
2026-03-26 09:05:52 +01:00
### Phase 2 - Performance GPU [FAIT]
Phase 2.2: MDI rendering with CPU-filled indirect args Replace per-chunk DrawInstanced loop with a single DrawInstancedIndirectCount. CPU fills indirect args buffer with same frustum+backface cull logic as Phase 2.1. Key discoveries: - Wicked Engine command signature includes push constant (20-byte stride, not 16) - SV_VertexID does not reliably include startVertexLocation with ExecuteIndirect - Solution: pack chunkIndex|(faceIndex<<16) in push constant, VS reconstructs quad offset from GPUChunkInfo lookup - No explicit DX12 barriers needed (implicit promotion from COMMON suffices) Also adds voxel_engine_spec.md and updates references from .docx to .md.
2026-03-25 22:07:22 +01:00
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
- **2.1** : Mega-buffer + CPU frustum/backface cull + per-face DrawInstanced
- **2.2** : CPU-filled indirect args + DrawInstancedIndirectCount (MDI)
- **2.3** : GPU compute culling (0.006 ms / 168 chunks)
- **2.4** : GPU compute mesher benchmark (CPU 277ms vs GPU 5.3ms)
- **2.5** : GPU meshing production + CPU optimisations (fused regen+pack, memcpy, dirty cache)
- **Résultat** : 80-110 FPS avec animation 60 Hz, 700+ FPS statique
Phase 2.5: GPU meshing production pipeline + perf optimizations (80+ FPS) Replace CPU greedy mesher with GPU compute mesher as default rendering pipeline. Key optimizations identified via CPU profiling (ProfileAccum, 5s averages): - Fused regenerate+pack: parallel noise gen + memcpy in same jobsystem pass (6ms → 0ms) - VoxelData memcpy: sizeof(VoxelData)==2 enables direct memcpy instead of bit-shift loop (28ms → <1ms) - Dirty-skip: GPU dispatch/upload only when chunks change, not every frame - Animation: 2 fBm octaves + no caves in animation mode (54ms → 8ms) - Result: 80-110 FPS with 60Hz terrain animation, 700+ FPS static
2026-03-26 09:05:52 +01:00
Phase 3: PS-based texture blending with winner-takes-all heightmap Replace pre-encoded quad blend data (v1) with per-pixel voxel data lookups in the pixel shader. The PS reads voxelDataBuffer (SRV t3) to find neighbor materials dynamically, enabling 2 independent blend axes, stair-priority neighbor detection, and winner-takes-all heightmap-driven transitions. Key design decisions validated through 6 iterations (see blending_experiments.md): - Winner-takes-all: material with highest heightmap score wins 100% (sharp but organic transitions, not smooth gradient) - Symmetric bias: bias = 0.5 - weight ensures equal chance at border - Subtractive corner attenuation (param=0.80): xAdj = xEdge - saturate(yEdge - 0.80) reduces blend at corners naturally - Blend zone = 0.25 voxels from each edge (50% of face) - Debug mode (F4) visualizes blend zones as colors
2026-03-26 12:14:08 +01:00
### Phase 3 - Texture blending [FAIT]
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
PS-based heightmap blending, winner-takes-all, corner attenuation subtractive. GPU mesh path uniquement.
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
Phase 4.2: GPU toping rendering pipeline + winding/lighting fixes Add instanced rendering for toping bevels: dedicated shaders (voxelTopingVS/PS), PSO, GPU buffers (t4 vertices, t5 instances), per-group DrawInstanced in a separate render pass with LoadOp::LOAD. Fix inverted face winding (emitTri auto-winding condition flipped for CW front-facing), slope normals (use inward direction not outward), and PS lighting (negate sunDirection like voxelPS). Update CLAUDE.md with Phase 4.1/4.2 documentation.
2026-03-26 17:47:08 +01:00
### Phase 4 - Toping [EN COURS]
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
- **4.1** [FAIT] : TopingSystem infrastructure, 4-bit adjacency, priority-based, stone wedges + grass tufts
- **4.2** [FAIT] : Shaders dédiés, vertex pulling instancié, half-Lambert + translucency vegetation
- **4.3** [A FAIRE] : Plus de types, LOD, animation vent, compute shader instances
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
Phase 5.1: Naive Surface Nets smooth rendering Implement CPU-side Naive Surface Nets for smooth voxel surfaces (SmoothStone, Snow) coexisting with blocky voxels (Grass, Dirt, Stone, Sand). Key features: - SmoothMesher with binary SDF, centroid vertex placement, per-axis boundary clamping to align with blocky grid at smooth↔blocky transitions - Cross-chunk connectivity: PAD=2 SDF grid, vertex range [-1, CHUNK_SIZE), canonical edge ownership (no duplicate triangles, no z-fighting) - Face normals oriented by edge axis+sign (robust with binary SDF, unlike SDF gradient dot or centroid sampling approaches) - Y-axis winding fix: sharing cells have different spatial arrangement, requiring opposite winding from X and Z axes - GPU mesher treats smooth neighbors as solid (no blocky faces toward smooth) - Material blending: primary (smooth-only) + secondary (all counts) per vertex - Dedicated shaders: voxelSmoothVS (vertex pulling t6) + voxelSmoothPS (triplanar + lerp blending between two materials) - Separate render pass with LoadOp::LOAD after voxels+topings - New materials: SmoothStone (mat 6), blocky Stone (mat 3) and Dirt patches added to world generation for boundary testing
2026-03-27 13:03:55 +01:00
### Phase 5 - Rendu smooth [EN COURS]
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
- **5.1** [FAIT] : Naive Surface Nets CPU, SDF binaire, cross-chunk connectivity, smooth/blocky boundary
- **5.2** [FAIT] : Smooth vertex normals, geometric normals triplanar, optimisations CPU (560ms → 17ms)
- **5.3** [A FAIRE] : GPU compute Surface Nets
- **5.4** [A FAIRE] : SDF lissé, LOD, pipeline asynchrone
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
Phase 6.1: RT infrastructure — MRT normals + BLAS/TLAS build - Normal render target (R16G16B16A16_SNORM) as MRT SV_TARGET1 in all 3 pixel shaders (voxelPS, voxelTopingPS, voxelSmoothPS) for future RT shadow/AO - BLAS extraction compute shader (voxelBLASExtractCS.hlsl): converts PackedQuad StructuredBuffer to float3 position buffer for DXR BLAS input - Blocky BLAS: single BLAS from all GPU-meshed quads (~1.5M triangles) - Smooth BLAS: single BLAS from smooth vertex buffer directly - TLAS: 2 instances (blocky + smooth), identity transforms, CreateBuffer2 with callback to avoid UpdateBuffer on RAY_TRACING flagged buffers - Fix: Wicked always accesses index_buffer in CreateRaytracingAccelerationStructure via to_internal() even for non-indexed geometry — provide dummy valid buffer
2026-03-28 14:48:11 +01:00
### Phase 6 - Ray tracing hybride [EN COURS]
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
- **6.1** [FAIT] : Normal RT (MRT), BLAS extraction CS, 3 BLAS (blocky+smooth+topings), TLAS
- **6.2** [FAIT] : RT shadows (3 jittered rays, TMin adaptatif, colored shadows)
- **6.3** [FAIT] : RT AO (4 cosine-weighted rays, IGN + Cranley-Patterson, temporal accumulation, bilateral blur)
- **6.4** [A FAIRE] : Fallback shadow maps + SSAO
### Phase 7 - Stylized Lighting [EN COURS]
- **7.1** [FAIT] : Hemisphere ambient, colored shadows, rim light, tone mapping + saturation, screenshot mode
Phase 7.1: stylized lighting — hemisphere ambient, colored shadows, rim light, tone mapping Wonderbox-inspired lighting overhaul across all 3 pixel shaders: - Hemisphere ambient (sky blue above, warm brown below) replaces flat ambient - RT shadows lerp toward blue-violet tint instead of plain darkening (factor 0.55) - Rim light (fresnel) with warm golden color on silhouettes (30% on vegetation) - Soft exponential tone mapping + saturation boost in final post-process pass - CB parameters for all lighting values (skyAmbient, groundAmbient, shadowTint, etc.) - Fog color/density centralized from CB instead of hardcoded per-shader - Screenshot mode (CLI "screenshot"): fixed camera, AO convergence, auto-capture - AO noise stability: world-space hash using voxel center + tangent-axis frac position - AO distance-weighted falloff: continuous occlusion values instead of binary hit/miss
2026-03-29 15:00:12 +02:00
Phase 2.5: GPU meshing production pipeline + perf optimizations (80+ FPS) Replace CPU greedy mesher with GPU compute mesher as default rendering pipeline. Key optimizations identified via CPU profiling (ProfileAccum, 5s averages): - Fused regenerate+pack: parallel noise gen + memcpy in same jobsystem pass (6ms → 0ms) - VoxelData memcpy: sizeof(VoxelData)==2 enables direct memcpy instead of bit-shift loop (28ms → <1ms) - Dirty-skip: GPU dispatch/upload only when chunks change, not every frame - Animation: 2 fBm octaves + no caves in animation mode (54ms → 8ms) - Result: 80-110 FPS with 60Hz terrain animation, 700+ FPS static
2026-03-26 09:05:52 +01:00
## Métriques cibles et résultats
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
Phase 6.1: RT infrastructure — MRT normals + BLAS/TLAS build - Normal render target (R16G16B16A16_SNORM) as MRT SV_TARGET1 in all 3 pixel shaders (voxelPS, voxelTopingPS, voxelSmoothPS) for future RT shadow/AO - BLAS extraction compute shader (voxelBLASExtractCS.hlsl): converts PackedQuad StructuredBuffer to float3 position buffer for DXR BLAS input - Blocky BLAS: single BLAS from all GPU-meshed quads (~1.5M triangles) - Smooth BLAS: single BLAS from smooth vertex buffer directly - TLAS: 2 instances (blocky + smooth), identity transforms, CreateBuffer2 with callback to avoid UpdateBuffer on RAY_TRACING flagged buffers - Fix: Wicked always accesses index_buffer in CreateRaytracingAccelerationStructure via to_internal() even for non-indexed geometry — provide dummy valid buffer
2026-03-28 14:48:11 +01:00
| Métrique | Cible | Résultat (Ryzen 7 9800X3D + RX 9070 XT) |
Phase 2.5: GPU meshing production pipeline + perf optimizations (80+ FPS) Replace CPU greedy mesher with GPU compute mesher as default rendering pipeline. Key optimizations identified via CPU profiling (ProfileAccum, 5s averages): - Fused regenerate+pack: parallel noise gen + memcpy in same jobsystem pass (6ms → 0ms) - VoxelData memcpy: sizeof(VoxelData)==2 enables direct memcpy instead of bit-shift loop (28ms → <1ms) - Dirty-skip: GPU dispatch/upload only when chunks change, not every frame - Animation: 2 fBm octaves + no caves in animation mode (54ms → 8ms) - Result: 80-110 FPS with 60Hz terrain animation, 700+ FPS static
2026-03-26 09:05:52 +01:00
|----------|-------|---------------------------------------|
Fix VRAM leak: capacity-based BLAS/TLAS + deferred toping BLAS upload Per-frame CreateRaytracingAccelerationStructure calls during F3 animation caused VRAM explosion (especially toping BLAS at ~23M vertices). Now all 3 BLASes use capacity-based allocation with 25% headroom — only recreated when vertex count exceeds capacity, otherwise just BuildRaytracingAS with updated desc.vertex_count. TLAS only recreated when instance count changes. Also adds deferred toping BLAS position upload via UpdateBuffer in Render() (topingBLASDirty_ flag), enabling toping shadows to update during animation. Split CLAUDE.md into CLAUDE.md + TROUBLESHOOTING.md for maintainability.
2026-03-30 21:37:39 +02:00
| FPS 1440p | > 60 fps | 80-110 FPS (anim blocky), 700+ FPS (statique) |
| FPS anim smooth+topings | > 15 fps | 17 FPS (smooth+topings+blocky anim 60Hz) |
| Meshing GPU (blocky) | < 200 us/chunk | ~0.6 us/chunk (0.1ms / 171 chunks) |
| Meshing CPU (smooth) | < 30ms | 17ms (parallele, 648 chunks) |
| Memoire GPU | < 500 Mo | ~30 Mo |
| Draw calls | < 100 | 1 (GPU mesh) ou 1 (MDI) |
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
## Conventions
- Namespaces : tout le code voxel est dans `namespace voxel`
- Chunks : 32x32x32, configurable via `CHUNK_SIZE`
- Coordonnées : Y = haut, monde infini en X/Z, hashmap sparse
Phase 5.1: Naive Surface Nets smooth rendering Implement CPU-side Naive Surface Nets for smooth voxel surfaces (SmoothStone, Snow) coexisting with blocky voxels (Grass, Dirt, Stone, Sand). Key features: - SmoothMesher with binary SDF, centroid vertex placement, per-axis boundary clamping to align with blocky grid at smooth↔blocky transitions - Cross-chunk connectivity: PAD=2 SDF grid, vertex range [-1, CHUNK_SIZE), canonical edge ownership (no duplicate triangles, no z-fighting) - Face normals oriented by edge axis+sign (robust with binary SDF, unlike SDF gradient dot or centroid sampling approaches) - Y-axis winding fix: sharing cells have different spatial arrangement, requiring opposite winding from X and Z axes - GPU mesher treats smooth neighbors as solid (no blocky faces toward smooth) - Material blending: primary (smooth-only) + secondary (all counts) per vertex - Dedicated shaders: voxelSmoothVS (vertex pulling t6) + voxelSmoothPS (triplanar + lerp blending between two materials) - Separate render pass with LoadOp::LOAD after voxels+topings - New materials: SmoothStone (mat 6), blocky Stone (mat 3) and Dirt patches added to world generation for boundary testing
2026-03-27 13:03:55 +01:00
- Matériaux : palette de 256, index 0 = air (vide), 1=grass, 2=dirt, 3=stone (blocky), 4=sand, 5=snow (smooth), 6=smoothstone (smooth)
Phase 2: GPU-driven voxel rendering pipeline Mega-buffer architecture replacing per-chunk GPU buffers: - Single StructuredBuffer<PackedQuad> for all chunks (2M quads, 16 MB) - StructuredBuffer<GPUChunkInfo> with per-chunk metadata (position, quad offsets, face groups) - VS reads chunk info via push constants (b999) for driver-safe chunk indexing - CPU frustum culling with wi::primitive::Frustum + AABB per chunk - Quads sorted by face direction in greedy mesher (faceOffsets/faceCounts) - GPU frustum + backface cull compute shader (voxelCullCS.hlsl) - GPU binary mesher compute shader baseline (voxelMeshCS.hlsl) - Indirect draw buffers and timestamp query infrastructure - README with build instructions and project architecture
2026-03-25 14:24:05 +01:00
- Faces : 0=+X, 1=-X, 2=+Y, 3=-Y, 4=+Z, 5=-Z
Phase 5.1: Naive Surface Nets smooth rendering Implement CPU-side Naive Surface Nets for smooth voxel surfaces (SmoothStone, Snow) coexisting with blocky voxels (Grass, Dirt, Stone, Sand). Key features: - SmoothMesher with binary SDF, centroid vertex placement, per-axis boundary clamping to align with blocky grid at smooth↔blocky transitions - Cross-chunk connectivity: PAD=2 SDF grid, vertex range [-1, CHUNK_SIZE), canonical edge ownership (no duplicate triangles, no z-fighting) - Face normals oriented by edge axis+sign (robust with binary SDF, unlike SDF gradient dot or centroid sampling approaches) - Y-axis winding fix: sharing cells have different spatial arrangement, requiring opposite winding from X and Z axes - GPU mesher treats smooth neighbors as solid (no blocky faces toward smooth) - Material blending: primary (smooth-only) + secondary (all counts) per vertex - Dedicated shaders: voxelSmoothVS (vertex pulling t6) + voxelSmoothPS (triplanar + lerp blending between two materials) - Separate render pass with LoadOp::LOAD after voxels+topings - New materials: SmoothStone (mat 6), blocky Stone (mat 3) and Dirt patches added to world generation for boundary testing
2026-03-27 13:03:55 +01:00
- Smooth flag : `FLAG_SMOOTH = 0x1` dans VoxelData flags — active Surface Nets au lieu du rendu blocky
Reference in a new issue Copy permalink