Physically Based Real-time Ray Tracing

1
Physically Based Real-time Ray Tracing

• Ryan Overbeck

2
Overview

• Ray tracing
• Broadest success at solving physically based
  rendering tasks.
• But is slower than Z-Buffer (for primary
  visibility)
• Real-time Ray Tracing
• Recent algorithmic improvements make real-time
  ray tracing possible
• Most of these algorithmic improvements have only
  been demonstrated for primary visibility w. point
  light shadows
• Physically Based Ray Tracing
• Uses Path Tracing / Monte-Carlo ray tracing for
  physical effects
• Consistent algorithmic improvements over the
  years
• Very Slow
• Currently Real-time and Physically Based ray
  tracing algorithms dont work well together

3
Real-time Ray Tracing

• Ray Packets Shoot 4 or 4x4 (or more) rays at a
  time
• Neighboring rays take similar paths through the
  scene
• Use SIMD (ie. SSE) instructions to perform one
  instruction on 4 floating point values at once
• Use interval arithmetic to use a single test to
  represent the entire packet
• Geometry Proxies Frustum Traversal, LCTS
• Use bounding geometry to represent a larger group
  of rays (usually 16x16 128x128)
• Miscellaneous Improvements
• Improved Acceleration Structures,
• Geometry Intersection tests,
• Multi-threading for multi-core architectures,

4
Physically Based Ray Tracing

• Effects
• Image anti-aliasing, Motion Blur
• Complex Lighting (soft shadows, hard shadows,
  environment lighting, ambient lighting,
  inter-reflections)
• Complex Materials (perfect reflection, glossy
  reflection, subsurface scattering,)
• Film / Lens effects (depth of field, bloom, )
• Tools
• Intelligent sampling distributions (Stratified
  Sampling, Importance Sampling, Low discrepancy
  sampling, Poisson Disc Sampling, )
• Adaptive / Interpolated Sampling
• Image / Film processing (filters, tone mapping,)

5
My Goals

• Long Term
• Design algorithms to marry real-time ray tracing
  with physically based ray tracing
• Short Term
• Build a Framework within which I can experiment
• Where to start?
• MLRT fastest real-time ray tracer
• Would need to make it more general
• Highly optimized code very difficult to modify
• PBRT Most general Free physically based ray
  tracer
• Would need to make it fast
• Very nice design easy to modify

6
Progress

• Started overhaul of PBRT
• Tightened up kd-tree traversal code
• Added accelerations for ray vs. triangle
  intersection
• Parallelized PBRT -- multi-core processors
• Tile image space (usually 64x64 tiles)
• statically assign tiles to threads
• Added 4-ray packets for primary rays
• Shoot 4 rays at a time
• Use SIMD SSE to accelerate ray traversal

7
Test Setup

• 512x512 images
• 1 sample/pixel
• PBRT low discrepancy sampling for primary samples
• 2x2 PBRT box filter for image reconstruction
• 3.0GHz Pentium 4 (2 cores)

8
Test Setup

• Scenes

Sponza (66454 Triangles) Sibenik (80479 Triangles) Buddha (1087720 Triangles)
9
Results
Sponza Sibenik Buddha
1 rays/ 1 thread 2.778s 3.266s 3.054s
1 rays/ 2 threads 2.327s 2.532s 3.489s
1 rays/ 4 threads 2.361s 2.534s 4.384s
4 rays/ 1 thread 1.055s 1.041s 1.684s
4 rays/ 2 threads 0.835s 0.775s 1.656s
4 rays/ 4 threads 0.776s 0.819s 1.514s
10
Next Steps

• Tighten up samplers and image filtering
• These two are adding .3 s to 512x512 render time
  (almost as much as ray casting)
• Larger ray packets (esp. 4x4)
• Add Frustum Traversal
• Move to secondary effects
• This will mark the beginning of new research
  territory