-
Phase 2 — CPU Ray Tracing
- Build a software path tracer before touching Vulkan. Understanding the algorithm on CPU makes the GPU version much clearer.
- Reference: “Ray Tracing in One Weekend” series by Peter Shirley
- Parent: PathTracer Learning
-
2.1 Camera and Ray Generation
-
2.2 Ray Definition and Scene Intersection
- PathTracer Learning - Concept - Ray Definition
- A ray is a parametric line:
P(t) = origin + t * direction
t > 0 means in front of the origint_min and t_max define the valid interval (avoids self-intersection)
- PathTracer Learning - Concept - Ray-Triangle Intersection
- Möller–Trumbore algorithm — the standard for real-time and offline rendering
- Returns
t, u, v barycentric coordinates
- Used to interpolate normals, UVs, and other vertex attributes
- PathTracer Learning - Concept - AABB
- Axis-Aligned Bounding Box — the building block of BVH
- Slab method intersection: test 3 pairs of parallel planes
t_enter = max(t_x_min, t_y_min, t_z_min), t_exit = min(t_x_max, t_y_max, t_z_max)- Hit if
t_enter <= t_exit && t_exit > 0
-
2.3 Acceleration Structures
- PathTracer Learning - Concept - BVH Construction
- Bounding Volume Hierarchy — tree of AABBs
- Naive O(N) per ray → BVH O(log N) per ray
- SAH (Surface Area Heuristic) for optimal splits
- Binned SAH: O(N) per level, K=32 bins
- PathTracer Learning - Concept - BVH Traversal
- Recursive descent: test AABB, if hit recurse into children
- Leaf nodes contain actual triangles
- Ordered traversal: visit closer child first for early exit
- Iterative traversal with explicit stack (GPU-friendly)
-
2.4 Radiometry and Shading
-
2.5 Materials and BRDF
- PathTracer Learning - Concept - BRDF
- Bidirectional Reflectance Distribution Function
f_r(ω_i, ω_o) — ratio of reflected radiance to incident irradiance- Must be energy-conserving:
∫ f_r cos(θ) dω ≤ 1
- Must be reciprocal:
f_r(ω_i, ω_o) = f_r(ω_o, ω_i)
- PathTracer Learning - Concept - Microfacet Theory
- GGX NDF, Smith G term — the theory behind rough specular surfaces
- Why
α = roughness² (perceptual remapping)
- PathTracer Learning - Concept - Fresnel Effect
- Schlick approximation:
F(θ) = F0 + (1-F0)(1-cosθ)^5
- Metals vs dielectrics:
F0 interpretation
- Why everything is more reflective at grazing angles
- Lambertian (diffuse) BRDF
f_r = albedo / π- Scatters light equally in all directions
- Combined with cosine-weighted sampling:
f_r * cos(θ) / p(ω) = albedo
- The
π and cos(θ) cancel perfectly — very clean implementation
- GGX Microfacet BRDF
f_r = D(h) * G(ω_i, ω_o) * F(ω_o, h) / (4 * dot(N, ω_i) * dot(N, ω_o))D — Normal Distribution Function (GGX/Trowbridge-Reitz)G — Geometric shadowing/masking (Smith)F — Fresnel term (Schlick approximation)- GGX NDF:
D(h) = α² / (π * (dot(N,h)² * (α²-1) + 1)²)
- Disney Principled BRDF
- Artist-friendly:
baseColor, metallic, roughness, specular, anisotropic, sheen, clearcoat, transmission
- Combines diffuse + specular + clearcoat + sheen layers
- Used in Godot’s PBR material system
-
2.6 Path Tracing Core Loop
-
2.7 Image Output and Tone Mapping
-
2.8 Accumulation and Convergence
- PathTracer Learning - Concept - Temporal Accumulation
- Average N samples over time:
color = (prev * (N-1) + new) / N
- Converges to ground truth as N → ∞
- Fireflies
- Extremely bright pixels from rare high-energy paths
- Caused by low-probability paths with high contribution
- Mitigation: clamp radiance, use MIS, better importance sampling
- Convergence rate
- Monte Carlo converges at
O(1/√N) — need 4x samples to halve noise
- Importance sampling reduces variance without changing convergence rate
- Denoising (DLSS, OIDN) compensates for slow convergence