Hardware-Assisted CSG Rendering

1
Hardware-Assisted CSG Rendering

• Fabiano Romeiro, Luiz Velho, Luiz
  Henrique de Figueiredo

Harvard University
IMPA
IMPA
2
Theme

• CSG (Constructive Solid Geometry)
• Hierarchical modeling of complex objects
• gt Generated by combinations of simple primitives

3
Motivation

• Interactivity is important
• Makes design easier
• Handle complex objects

• However, it is difficult
• Scalability

4
Previous Work

• Initial Approaches (CPU)
• Boundary representations
• Depth-layering
• Goldfeather et al.
• Recent Approaches (GPU)
• Stencil Buffer
• Wiegand, Rappoport, Stewart
• Depth-peeling
• Blister, Hable and Rossignac

5
Our Approach

• Current Algorithms are bandwidth limited (i.e.,
  depth layering, depth peeling, etc)

• GOAL
• Make better use of GPU processing power
• Shift bottleneck
• Bandwidth GPU Instruction throughput

6
Overview

• Perform spatial subdivision on the CPU until CSG
  expression is locally simple enough to render on
  the GPU.

• Locally Simple
• Surface within a cell is
• Single primitive
• Boolean operation of two primitives

7
Algorithm

• Subdivision Step
• Create Hierarchical Spatial Decompositionbased
  on the CSG expression
• Ray Tracing Step
• Traverse Subdivision Structure andRender each
  Cell

8
Subdivision Step

• Subdivision Structure
• Modified octree (sequenced kd-tree)

Illustrative Subdivision Example in 2D
9
Real Subdivision Example
10
Visual Trade-Off

• Quality x Speed
• Increasing threshold gives better performance
  but introduces visible artifacts

1 pixel
20 pixels
OBS No make-up!
11
Ray-Tracing on the GPU

• Draw faces of Objects Proxy Bounding Volume
• Use Vertex and Pixel shadersto trace rays on the
  GPU

12
Ray-Tracing Step

• Octree Traversal
• Recursive front-to-back order
• At leaf cells
• Take action according to cell classification

13
CSG Ray-Tracing

• Two Types of Cells

a) Single primitive shader
b) Boolean operation of 2 primitives
o
o
o
o
14
Results

• Measurements performed in 2 configurations
• C1 Pentium-M 1.4GHz, Geforce FX Go5200
• C2 Athlon 64 3800, Geforce 6800 GT
• To evaluate performance we measured
• S Subdivision time (CPU)
• TR Traversal and rendering time (GPU)
• Ts Setup time (Bandwidth)

15
Results
CK200
CK500
CK1000
WIDGET
Model S S TS TS TR TR
Model C1 C2 C1 C2 C1 C2
WIDGET 17 12 9 5 2262 59
CK200 223 163 84 42 6331 346
CK500 1232 888 432 213 16046 1514
CK1000 3705 2658 1265 626 34975 4324
16
Demo
17
Conclusions

• Performance comparable to best current methods
• For models with 500 primitives
• GPU instruction throughput limited
• Next generation GPUs
• Expect to surpass other methods on complex models