This repository contains a GPU-accelerated path tracer used for rendering glTF and Collada scenes. The host-side application loads geometry and materials with Assimp, builds a CUDA BVH, and launches a custom renderer that supports temporal and spatial resampling via ReSTIR-GI. Output images are saved as OpenEXR files with color, normal, and depth layers to make post-processing and denoising straightforward.

The master branch is the reference implementation that targets a desktop NVIDIA GPU. A separate vortex branch contains an alternative backend for the Muon/Vortex architecture. See docs/README_vortex.md for the workflow on that branch.

---

Quick Start

# Configure (Release build by default)
cmake -S . -B build

# or enable debug logging / CUDA line info
cmake -S . -B build -DBUILD_DEBUG=ON

# Build the renderer (ninja or make depending on your generator)
cmake --build build -j

# Render a still image to build/output.exr using a bundled scene
./build/pathtracer -l 4 -m 8 -R 0 -o 1 -f build/output.exr gltf/Chess.glb

Generated images are written to the path you pass with -f. The OpenEXR file stores RGB in R/G/B, surface normals in NX/NY/NZ, and the linear depth in Z.

After rendering a sequence (-g), consider running the optional batch_denoise.py tool together with Intel Open Image Denoise and ffmpeg to denoise and assemble a video.

---

Repository Layout

Path	Purpose
src/host/	Command-line application, renderer orchestration, asset loading and EXR output.
src/target/	CUDA kernels, sampling utilities, and the GPU path tracer implementation.
src/scene/	Geometry, BVH, light, and material utilities shared by host and device.
src/util/	Third-party helper libraries (stb, tinygltf, json, RNG, ReSTIR helpers, etc.).
dae/	Sample Collada scenes kept from the original CS184 starter kit.
gltf/	Ready-to-render glTF assets used when launching the renderer from the CLI.
bakes/	(Optional) Pre-baked textures or lighting data; not tracked in Git.
build/	CMake build tree. Use cmake -S . -B build to populate.

---

Dependencies

Install the following before building:

• CUDA Toolkit (11.8 or newer recommended) with a compatible NVIDIA GPU.
• CMake 3.8+ and either Ninja or Make.
• A C++17 compiler (GCC 9+, Clang 10+, or MSVC 2019+).
• Assimp development package (provides assimp::assimp).
• OpenEXR development package for EXR output.

Optional but useful:

• ffmpeg for assembling PNG/EXR sequences into videos.
• Intel Open Image Denoise (oidnDenoise) when running batch_denoise.py.
• Python 3.9+ with a standard library only (for batch_denoise.py).

Install the CUDA Toolkit via the NVIDIA packages that match your driver version.

---

Building

1. Configure a build directory:

   cmake -S . -B build [-DBUILD_DEBUG=ON]

   BUILD_DEBUG=ON switches to a Debug configuration, enables CUDA line info, and turns on verbose renderer logging.

2. Compile:

   cmake --build build -j

   The resulting executable is build/pathtracer (or build/Debug/pathtracer when using the multi-config generators on Windows).

---

Running the Renderer

Command-line usage:

Usage: pathtracer [options] <scene-file>

  -l  <INT>        Number of samples per area light
  -g  <INT>        Number of frames to generate (1 for a still)
  -R  <INT>        Enable ReSTIR-GI (0 = off, 1 = on)
  -d  <INT>        Enable debug logging (0 = off, 1 = on)
  -m  <INT>        Maximum ray depth
  -o  <INT>        Accumulate indirect bounces (1) or only use the last bounce (0)
  -f  <PATH>       Output EXR filename
  -r  <W> <H>      Resolution when running headless mode
  -h               Show this help message

Typical workflows:

• One-off render

  ./build/pathtracer -l 8 -m 12 -o 1 -R 1 -f build/chess_restir.exr -r 1920 1080 gltf/ChessCubeSkybox.glb

  This writes an EXR with color, normals, and depth. Open it in an EXR viewer

• Turntable animation

  ./build/pathtracer -l 4 -m 6 -g 120 -f build/chess_turntable.exr -r 1280 720 gltf/Chess.glb

  When -g > 1 the camera performs a full 360° rotation; frames are stored as <prefix>0000.exr, <prefix>0001.exr, …

• Debugging

  ./build/pathtracer -d 1 -o 0 -m 2 -f build/debug_bunny.exr gltf/bunny.glb

  Debug output prints kernel configuration, BVH build time, and ray statistics.

Outputs

Each EXR pixel stores:

• R/G/B: ACES-tonemapped radiance.
• NX/NY/NZ: Shading normal in world space.
• Z: Camera-space depth.

---

Assets & Scenes

• glTF scenes: curated test assets under gltf/. Use these to validate the CUDA backend quickly.
• Collada scenes: legacy CS184 scenes under dae/. Helpful for CPU reference renders or regression testing.

./build/pathtracer -f build/custom.exr ~/assets/my_scene.glb

---

TODO

• PBRT support
• Better rendering algorithms (Area-ReSTIR, etc)
• More complex scenes
• Tight integration with open image denoise