Generic Mesh Refinement on GPU

1
Generic Mesh Refinement on GPU

• Tamy Boubekeur Christophe Schlick
• LaBRI INRIA CNRS
• University of Bordeaux

2
Today GPU Features

• Combination of Vertex and Fragment Shaders
• Dedicated hardware primitives (VBO, FBO)
• C-like programming language (GLSL, CG, HLSL)
• gt Most of the rendering task is on GPU

3
Today GPU Rendering

• Per-pixel illumination
• Hard and Soft Shadows
• HDR images ATI 2003
• Global Illumination Effects NVIDIA 2005
• gt Near-Realistic real-time rendering

4
Realistic Image

• Means realistic lighting of course, but also
• Complex objects (many polygons)
• Details (at the pixel resolution)
• Usual solution
• Shape complexity smooth and/or detailled
  surfaces
• Appearance complexity high-resolution
  multi-texturing fragments shaders

5
Fair Surface Representation

• Multi-scale representation for real-time
  graphics
• Light representation for animation/physics.
• On-the-fly geometric refinement for silhouette
  and contours.
• Details at rasterization time.
• One acceptable solution
• CPU Low resolution clever mesh, carefully
  designed.
• CPU/GPU Medium resolution displacement map.
• GPU High resolution appearance map (normal,
  color).

6
Fair Surface Representation

• Multi-scale representation for real-time
  graphics
• Light representation for animation/physics.
• On-the-fly geometric refinement for silhouette
  and contours.
• Details at rasterization time.
• One acceptable solution
• CPU Low resolution clever mesh, carefully
  designed.
• CPU/GPU Medium resolution displacement map.
• GPU High resolution appearance map (normal,
  color).

7
Previous Work

• Subdivision Surfaces
• Many schemes Catmull 1978Loop 1987Zorin
  2000Stam 2003
• Local control Biermann 2001
• Hardware evaluation Bishoff 2000Bolz 2002
• GPU implementation Bolz 2003Shiue 2005
• Local Refinement
• Curved PN Triangles Vlachos 2001Chung 2003
• Scalar Tagged Meshes Boubekeur 2005

8
Dynamic Refinement
Frame Rendering
Graphics Bus
Refinement
Coarse Mesh
Tessellated Mesh
Refined Mesh
Tessellator
Displacement
RENDERING
CPU
GPU
Non-programmable Graphics Hardware
9
Dynamic Refinement
Frame Rendering
Graphics Bus
Refinement
Coarse Mesh
Tessellated Mesh
Refined Mesh
Tessellator
Displacement
RENDERING
CPU
GPU
Programmable Graphics Hardware
10
Our goal
Virtual GPU Tessellator Unit on todays GPU
Graphics Bus
Refinement
Coarse Mesh
Tessellated Mesh
Refined Mesh
Tessellator
Displacement
RENDERING
CPU
GPU

• Dynamic coarse meshes on CPU
• Per-polygon tessellation on GPU
• Local computation of refined vertices
• General enough solution for various kinds of
  local displacement

11
Our Approach

• Storing only ONE triangle fully tessellated the
  Refinement
• Pattern (RP), directly on the GPU memory
• Configuring the GPU with the coarse triangle
  parameters
• Rendering the RP instead of each coarse triangle
• Conforming the RP to the current GPU
   configuration 

12
GPU Refinement Pattern

• The only one drawn primitive
• A collection of triangle strips, pre-generated
  and stored on the GPU
• Problem
• How to transform efficiently the RP toward the
  current coarse triangle ?
• How to resolve this problem in a GPU-friendly
  fashion (ie. for each vertex of the RP
  independently) ?

13
GPU Refinement Pattern

• Solution
• Encoding the barycentric coordinates u, v, w as
  position of the RP vertices.
• Using this parameterization to interpolate
  parameters (position, normal, color, textures
  coordinates, etc) of the current triangle in the
  vertex shader

14
Displacement computation

• Performed by the Vertex Shader for each refined
  vertex.
• 3D-based displacement with interpolated position.
• 2D-based displacement with interpolated texture
  coordinates gt texturing meshes with meshes.

15
Level Of Detail

• Discrete LOD by simply storing a set of RP at
  different resolution on the GPU memory
• Switching on a per-object basis between the
  different level of detail
• Continuous LOD by interpolation between 2
  consecutive level of details.

16
Drawing algorithm

• Draw (CoarseMesh M, level L)
• GPURPbindLOD (L)
• for all Triangle T of M do
• GPUplaceAttributeOnGPU (T)
• GPURPdraw ()

17
Drawing algorithm

• Draw (CoarseMesh M, level L)
• GPURPbindLOD (L)
• for all Triangle T of M do
• GPUplaceAttributeOnGPU (T)
• GPURPdraw ()

Refinement Shader
1 Attribute Interpolation
2Displacement Computation
3 Output refined vertex
18
Implementation

• Requires only Vertex Shader 1 compatible GPUs
  (all programmable GPUs)
• No use of Fragment Shaders, no multi-pass, no
  render-to-texture, etc.
• Small vertex shader code
• RP stored as triangle strips in a VBO

19
Workload balance

• Software side/CPU workload
• Initialization of the RP (set of RP for LOD).
• High level modification of the coarse mesh.
• Realtime graphics bus load
• Upload of the coarse per-triangle parameters
  through uniforms variables (positions, normals,
  colors, additional local parameters)
• Drawing call of RP (VBO) at the chosen LOD
• Note Nb draw calls Nb coarse triangles

20
Workload balance

• GPU additional workload
• At the beginning of the Vertex Program, for the
  RP-based interpolation (i.e. the tessellation)
• Essentially, the procedural displacement
  computation or query (displacement map with
  Vertex Shaders 3.0)
• Note the RP-based interpolation masks the
• memory latency for texture access from the vertex
• shader NVIDIA 2004

21
Procedural Displacement
12 coarse triangle frequency control 786 432
on GPU triangles

• Concept
• Object Simple Base Mesh Procedural
  Function(s)
• Deep refinement required for high frequencies
• ex a.sin(fP)
• Full use of GPU (GPU-limited for deep refinements)

22
Curved PN Triangles

• Local triangle smoothing with cubic Bezier
  patches
• Point-Normal Construction
• Linear or quadratic normal field

ATI/Valchos et al. 2004

• Interpolatory refinement
• No dynamic topology to manage
• Still non ordered triangle stream
• Usually enough for real-time rendering

23
Curved PN Triangles

• With our approach
• Each coarse triangle is provided to GPU with its
  Bezier coefficients
• b300, b030,b003, b210,b201,b021,
  b120,b102,b012, b111
• After tessellation, the displacement is computed
  with coefficients in the vertex program

Note coefficients can be computed on the GPU and
stored as textures.
24
Scalar Tagged Mesh
EG 2005

• Enriched Meshes for real-time applications
• PN Triangles Scalar tag control on a per-vertex
  basis
• Surface refinement driven by scalar tags

25
Scalar Tagged Mesh
EG 2005

• Shape factors sharpness, tension, bias,
• Sharp creases locally defined
• Bezier coefficient classification around each
  vertex
• Intuitive behavior

26
Scalar Tagged Mesh
EG 2005

• Procedural Normal Map
• Procedural Displacement Map
• With our approach still coarse
  triangle/coefficient transmission

Note branching needed in our implementation
27
Performance
Simple tessellation of triangular meshes.

• no full resolution topology generation/storage
• For dynamic objects, between 1 and 3 orders of
  magnitude faster than a CPU refinement

28
Advantages

• Negligible graphics bus bandwidth
• Full use of GPU for deep refinement
• Easy to implement
• Easy to integrate as API extension (OpenGL driver
  implementation)
• Drawing of objects that would not fit in memory

29
Drawbacks

• Attributes transfer not optimal (uniform
  variables)
• Need of a middleware support for full GPU
  utilization at low tessellation rate
• Today more interesting for deep refinement (gt6x6)

30
Alternative

• Non triangular RP
• Non polygonal rendering (point-based surfaces)
• Non regular RP
• Alternative RP parameterization
• Alternative interpolation

31
Conclusion

• A low cost Hardware Tessellator Unit
• Generic method for hardware surface tessellation
  an displacement
• An hardware implementation of Curved PN Triangles
  for ALL programmable GPUs
• Very deep tessellation rate for procedural
  displacements
• Easy to implement
• Work on todays GPU/no additional hardware

32
Conclusion

• Refinement Shader
• RP-based interpolation Tessellation
• Function evaluation/query Displacement
• Output

33
Current and Future Work

• Symmetry analysis for rough view-dependent LOD
• Refinement with high order continuity using
  constrained meshes (vertex degree) Bolz2002
• Semi-automatic conversion of larges meshes to GPU
  Fair Representation
• New local smoothing algorithms (ST-Meshes)

34
Thank you for your attention !
Demo
http//www.labri.fr/boubek