HardwareAssisted CSG Rendering

1
Hardware-Assisted CSG Rendering

• Fabiano Romeiro, Luiz Velho, Luiz Henrique de
  Figueiredo

Harvard University
IMPA
IMPA
2
Motivation

• CSG Hierarchical modeling of complex objects
• Objects generated by boolean combinations of
  simple primitives

3
Motivation

• Interactivity is important
• Makes design easier

• However it is dificult
• Scalability

4
Previous Approaches

• Boundary representations
• Depth-layering (Goldfeather)
• Recently algorithms that use GPU resources
  (Wiegand, Rappoport, Stewart, Guha, etc)
• Blist, Depth-peeling (Hable and Rossignac)

5
Motivation

• Recent approaches try to reach interactivity on
  ever more complex models

• Current best performing ones are bandwidth
  limited (depth layering, depth peeling, etc)

• GOAL
• Make more use of GPU processing power
• Shift bottleneck from bandwith to GPU instruction
  throughput

6
Method Overview

• Perform spatial subdivision of the CSG object on
  the CPU until locally it is simple enough to be
  ray-traced on the GPU.

• Locally simple enough Surface defined by a
  single primitive or boolean operation of two
  primitives

7
Method Overview Subdivision Step

• Modified octree (sequenced kd-tree) used as
  subdivision structure

Real subdivision example
Illustrative subdivision example in 2D
8
Method Overview Subdivision Step

• Trade-off Increasing threshold increases
  performance but introduces noticeable artifacts

9
Method Overview Ray-tracing Step

• Octree is traversed in recursive front-to-back
  manner
• As leaf cells are reached different action is
  taken according to cell classification

10
Method Overview Ray-tracing Step

• As faces are rendered pixel shader runs for every
  point of the faces

b) Boolean operation of 2 primitives Shader
a) Single primitive Shader
o
o
o
o
11
Results

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

12
Results
CK200
CK500
CK1000
WIDGET
13
Conclusions

• Results comparable performance-wise to current
  top performing methods
• For models with up to 500 primitives mostly
  GPU instruction throughput limited
• With newer GPUs, we expect our method to surpass
  other methods on such models