Project: Vulkan RT Path Tracer

• Parent: PathTracer Learning
• Phase: PathTracer Learning - Phase 3 - GPU and Vulkan

Goal

• Port the CPU path tracer to Vulkan ray tracing
• Integrate with Godot’s rendering pipeline

Prerequisites

• PathTracer Learning - Project - CPU Path Tracer complete
• Vulkan SDK installed
• GPU with RT support (NVIDIA RTX or AMD RDNA2+)

Milestones

• Milestone 1: Vulkan setup + triangle in RT
  • Initialize Vulkan with RT extensions
  • Create BLAS from a single triangle
  • Create TLAS with one instance
  • Write minimal rgen + rchit + rmiss shaders
  • Display result in a window
• Milestone 2: Full scene + BVH
  • Load OBJ, create BLAS per mesh
  • Create TLAS with all instances
  • Pass vertex/index data to closest-hit shader via device address
  • Reconstruct hit position and normal in shader
• Milestone 3: Path tracing on GPU
  • Implement iterative path tracing in rgen shader
  • Cosine-weighted sampling in GLSL
  • Temporal accumulation (ping-pong buffers)
  • Display accumulated result
• Milestone 4: Full PBR + NEE
  • GGX BRDF in GLSL
  • Shadow rays for direct lighting
  • MIS weighting
• Milestone 5: Godot integration
  • Hook into Godot’s _render_scene()
  • Use Godot’s RenderingDevice for all Vulkan calls
  • Extract mesh data from Godot’s scene graph
  • See PathTracer Learning - Phase 4 - Godot Internals

Key Technical Challenges

• Vertex data access in closest-hit shader
  • Use GL_EXT_buffer_reference to access vertex buffer by device address
  • Store per-instance buffer addresses in a storage buffer
  • Index by gl_InstanceCustomIndexEXT
• Material data
  • Pack PBR parameters into a compact struct
  • Store in per-instance data buffer
  • Sample textures using gl_InstanceCustomIndexEXT as texture array index
• Random number generation
  • PCG hash or Xorshift for per-pixel, per-sample RNG
  • Seed with gl_LaunchIDEXT.xy and frame counter

  uint pcg(uint v) {
      uint state = v * 747796405u + 2891336453u;
      uint word = ((state >> ((state >> 28u) + 4u)) ^ state) * 277803737u;
      return (word >> 22u) ^ word;
  }
  float rand(inout uint seed) {
      seed = pcg(seed);
      return float(seed) / float(0xFFFFFFFFu);
  }

Performance Targets

• 1080p @ 60fps with 1 spp + denoising
• BLAS build: < 1ms per frame for dynamic objects
• TLAS rebuild: < 0.5ms per frame