Smooth MixedResolution GPU Volume Rendering

1
Smooth Mixed-Resolution GPU Volume Rendering

• Johanna Beyer1, Markus Hadwiger1,
• Torsten Möller2, Laura Fritz1

1 VRVis Research Center, Austria 2 Simon Fraser
University, Canada
2
Motivation

• Texture memory limits volume size in GPU volume
  rendering
• Bricking and multi-resolution techniques
• Simple down-sampling skip samples
• ? No pre-filtering leads to aliasing
• Sample placement often restricted
• ? Filters with odd number of taps (1, 3, )
• Mixing of resolution levels
• ? Adjacent bricks of different resolution
• Often discontinuities at brick borders
• Or original samples modified

3
Smooth multi-resolution interpolation

• Combine bricks of different resolution
• Pre-filtering with even number of taps (2, 4, )
• Efficient texture packing with single-pass
  raycasting
• Continuity at brick borders

4
Flat vs. Hierarchical Bricking
hierarchical
flat
flat

• Culling rate higher for flat bricking

5
2-Level Flat Bricking

• Subdivide volume into equally sized bricks (e.g.
  32³) and duplicate voxels at brick boundary
  (322)³

• Store all bricks in a single large cache texture

6
2-Level Flat Bricking

• Construct lower resolution bricks (161)³

• Texture packing
• 8 low-res bricks fit into one high-res brick
  (34³)

7
Texture packing

• Determine active bricks via culling
• Pack active bricks into single 3D cache texture
• Lookup from virtual volume to physical cache
  texture
• Layout texture
• RGB address translation
• A multi-resolution scale

8
Volume Rendering (1)

• Lookup from virtual volume space into cache
  texture

from layout texture

• Brick resolution transparent to shader!

9
Volume Rendering (2)

• Smooth C0-continuous interpolation
• Always use hardware-native tri-linear
  interpolation
• Adjust texture coordinates for interpolation in
  resolution level boundary
• Necessary steps
• During raycasting Warping of texture coordinates
• After culling update Cache Fixup

10
Smooth Interpolation 2D Case

• Transition region that requires modification is
  band of 0.5 voxels inside the brick of higher
  resolution
• Two different interpolation functions in 2D
• trapezoids and triangles

• Modify texture coords for hardware-native
  tri-linear interpolation

• Special attention at corners
  ? primitive type
  depending on surrounding bricks

11
Smooth Interpolation 2D Case

• Triangle and trapezoid must be mapped to square
  for hardware bi-linear interpolation

12
Smooth Interpolation 3D Case

• In 3D three different primitives for
  interpolation
• Truncated pyramid (two trapezoid projections)
• Pyramid (two triangle projections)
• One trapezoidal and one triangular projection

13
Cache Fixup

• Only performed when cache is updated
• Fix duplicated border of high-res bricks that
  are adjacent to low-res bricks in virtual volume
  space

14
Shader Modification

• Determine if sample position is
• (1) within the 0.5 texel border of a high-res
  brick, and (2) adjacent to a low-res brick
• Determine primitive type via two 2D projections
• Warp texture coordinates in 2D projections and
  combine to 3D

15
Results Performance

• Smooth vs. Discontinuous mixed-resolution DVR

Core 2 Duo, 3 GHz, 4GB RAM, NVidia Geforce GTX
280, viewport 512 x 512
16
Results Visual (2)
17
Results Visual (1)
18
Summary

• Mixed-resolution volume rendering
• Efficient texture packing scheme, fast lookup in
  shader

• Downsampling positions shifted by half a voxel
  for more flexibility in choice of downsampling
  filters
• C0-continuous transitions between resolution
  levels
• Warping of texture coords for hardware-native
  tri-linear interpolation

19
Conclusion

• Texture Packing
• Single-pass rendering combined with mixed
  resolutions
• - Additional address look-ups
• Smooth interpolation
• Full hardware-native filtering
• One desired texture fetch stays one real
  texture fetch
• No additional texture fetches in density volume
• Scales well with the number of volumes
  (multi-volume rendering)
• - Texture coordinate warping relatively complex
• Downsampling
• No gaps between bricks even with simple shader
• - Low resolution bricks still visible, the actual
  border is very thin

20
Thanks for your attention!

• Acknowledgements
• FFG Austrian Research Promotion Agency
• Agfa Healthcare Vienna

• Contact
• Website http//medvis.vrvis.at
• E-mail beyer_at_vrvis.at

21
(No Transcript)
22
Results Performance

• Bricked vs. Unbricked volume rendering
• Core 2 Duo, 3 GHz, 4GB RAM, NVidia Geforce GTX
  280, viewport 512 x 512

23
Flat Bricking

• Higher memory overhead when using sample
  replication at each brick boundary
• Number of bricks constant, less scalable
• Much more efficient culling
• Fine-grained LOD or shader selection

24
2-Level Flat Bricking - Example

• Brick cache 512x512x1024 (256 mega voxels)
• Brick size 34³, 15x15x30 bricks in cache
• Possible dataset size
• High resolution 480x480x960 (211 mega voxels)
• Low resolution 960x960x1920 (1.6 giga voxels)
• ?Actually much higher due to culled bricks!
• Three Levels Brick size 36³, 14x14x28 bricks in
  cache
• High resolution 448x448x896 (172 mega voxels)
• Middle resolution 896x896x1792 (1.3 giga voxels)
• Low resolution 1792x1792x3584 (10.7 giga voxels)

25
(No Transcript)
26
Brick Cache Fixup

• Implemented on GPU
• Reverse layout texture
  (lookup from brick cache
  into virtual volume space)
• Rasterize all marked high-res bricks
  slice-by-slice, perform fixup on each slice
  separately
• for each slice of brick that needs updating
• if fragment on brick boundary
• perform rev. lookup and fetch sample from adj.
  brick
• if neighbor sample is from a low-res brick
• overwrite high-res sample with nearest-neighbor
  low-res sample
• Copy 2D slice buffer back into 3D cache texture