Acceleration Data Structures for Ray Tracing Algorithms on the GPU

• Overview
• Project goals
• Tech stack
• Project structure
• Build & run
  • Requirements:
  • Configuration
• Project state
• Potential extensions

Overview

Author: Jakub Votrubec Supervisor: Ivan Šimeček University: Czech Technical University, Faculty of Informatics

This project is part of my master's thesis focused on GPU-based acceleration data structures for ray tracing. The goal is to research, implement, extend, and evaluate BVH construction algorithms based on Parallel Locally-Ordered Clustering (PLOC).

Specifically, the project explores:

• PLOC (Meister et al. 2017 PLOC)
• Extended Morton Codes (EMC) (Vinkler et al. 2017 Extended Morton Codes)
• SOBB (Skewed Oriented Bounding Boxes) (Káčerik et al. 2025 SOBB)

Licence can be found in licence.txt

---

Project goals

• Implement GPU-based BVH construction using PLOC
• Extend PLOC with:
  • Extended Morton Codes (EMC) - two versions
  • SOBB
  • Combined EMC + SOBB approach (novel combination)
• Evaluate and compare all variants:
  • PLOC
  • PLOC + EMC v1
  • PLOC + EMC v2
  • PLOC + SOBB
  • PLOC + EMC v1 + SOBB
  • PLOC + EMC v2 + SOBB
• The main evaluation metrics are:
  • BVH build time
  • Tracing performance
  • BVH cost

---

Tech stack

• CPU code: C++20 (Clang)
• GPU code: CUDA 12 (NVCC)
• Build system: CMake
• Configuration: JSON (nlohmann)
• Scene/model format: OBJ (rapidobj)
• Output image format: PPM
• Metrics export: Console & JSON (nlohmann)
• Benchmark orchestation: Python

---

Project structure

src/
  acceleration/     # CPU acceleration
  app/              # Driver code
  config/           # Configuration definition
  framebuffer/      # Framebuffer struct and logic
  gpu/              # CUDA kernels and GPU-specific code
  io/               # Config/scene loaders, framebuffer/results exporters
  renderer/         # Rendering logic and GPU frontend
  scene/            # Scene and primitive representation
  stats/            # Statistics definition
  util/             # Utilities

third-party/        # External dependencies
res/                # Scenes and metadata
out/                # Output renders
experiments/        # Benchmark orchestration, benchmark results
report/             # Master's thesis report and report source files

b.sh                # Build script
r.sh                # Run script (with sample_config.json)
sample_config.json  # Sample configuration

---

Build & run

1. Using scripts:

   ./b.sh <preset> # choose debug/release preset
   ./r.sh          # sample_config.json is used

2. Directly running the executable:

   ./build/diplodocus <json config path>
   

Requirements:

• Clang 14+
• CUDA Toolkit 12 (NVCC)
• CMake (CMake 3.28+)
• git

Configuration

• To change the configuration you can either use and modify sample_config.json, or create your own config file
• Unspecified values fall back to hardcoded deafults
• You can look up the default values in config/ and the precise json formulation in io/config/)

---

Project state

• CLI framework

  • Config loading (JSON)
  • Scene loading (OBJ)
  • PPM image export
  • Collected metrics export (Console or JSON)

• CPU reference renderer

  • Whitted-style ray tracer
  • Metrics collection

• CPU reference acceleration data structure

  • SAH-based BVH inspired by an article from Wald et al.
  • Metrics collection

• GPU infrastructure

  • Scene representation and memory transfer
  • Ray and configuration structures for the GPU

• GPU ray tracing

  • Whitted-style non-stack ray tracer
  • Metrics collection

• GPU BVH construction

  • PLOC
  • PLOC + EMC v1
  • PLOC + EMC v2
  • PLOC + SOBB
  • PLOC + EMC v1 + SOBB
  • PLOC + EMC v2 + SOBB
  • Metrics collection

• Scenes

  • Cornell Box
  • Cornell Box (sphere variant)
  • Stanford Bunny (in a Cornell Box)

• Extended benchmarking

  • Metrics collection structures
  • Console export
  • JSON export
  • Orechstration using Python with data collection into CSV

---

Potential extensions

• GUI application
• Automatic documentation

---