RayTracer

An ongoing multithreaded offline ray tracer (CPU-based), written in C++, focusing on rendering, spatial acceleration structures, and multithreaded performance optimization.

(Note: This is an active Work in Progress. Current implementation covers high-performance spatial sorting and temporal rendering, with upcoming plans to integrate pure Monte Carlo integration and Probability Density Functions (PDFs).)

Current State

Figure 1 - Evolution of the Ray Tracer.

---

Figure 2 - Temporal Rendering (Motion Blur) integrated with BVH and hardware concurrency.

Performance Benchmarks: Algorithmic vs. Hardware Scaling

The following benchmarks were recorded rendering the high-density Book 1 final scene (approx. 500 primitives) to demonstrate the performance impact of algorithmic spatial sorting compared to raw hardware multithreading.

Figure 3 - High-density benchmarking scene.

---

Benchmarks recorded on an Intel Core Ultra 7 155H, with VisualStudio 2022 being configured on Release mode with -O2 compiler optimisation. To maximize performance without being bottlenecked by hardware scheduling, thread count was dynamically set via std::thread::hardware_concurrency. For optimal results on this specific architecture, utilizing the 6 Performance (P) cores is recommended over the Efficiency (E) or Low-Power (LP) cores.)

Algorithm	Sequential	Multithreaded
Linear Intersection - $O(N)$	3638.43 seconds (1 hour 38 seconds 430 milliseconds)	499.92 seconds (8 minutes 19 seconds 920 milliseconds)
BVH Traversal - $O(\log{N})$	450.33 seconds (7 minutes 30 seconds 330 milliseconds)	84.64 seconds (1 minute 24 seconds 640 milliseconds)

Key Takeaways:

1. Algorithmic Optimization > Hardware Brute Force: A sequential BVH implementation (~450s) actively outperformed a fully multithreaded brute-force approach (~500s).
2. Hardware Scaling: After the $O(\log{N})$ spatial optimization, multithreading yielded a highly efficient ~5.3x performance speedup (~450s down to ~84s).
3. Total Optimization: The combination of acceleration structure and hardware concurrency reduced the render time from over an hour to just 1.5 minutes, providing ~43x overall speedup.

---

Future Roadmap

• Surface Area Heuristic (SAH) for optimized BVH node splitting.
• Monte Carlo Integration and Importance Sampling.
• Participating Media (Volumetrics/Fog).
• Texture mapping and Perlin noise generation.

Citations

• Built following the foundational mathematics outlined in Peter Shirley's Ray Tracing in One Weekend and Ray Tracing: The Next Week books.