Memory Efficient Acceleration Structures and Techniques for CPUbased Volume Raycasting of Large Data

1
Memory Efficient Acceleration Structures and
Techniques for CPU-based Volume Raycasting of
Large Data

• S. Grimm, S. Bruckner, A. Kanitsar and E. Gröller
• Institute of Computer Graphics and Algorithms
• Vienna University of Technology
• Vienna, Austria

2
Motivation (1/3)

• Direct Volume Rendering
• Important tool in
• medical environments
• CT angiography
• run-offs (gt 1000 slices)
• are clinical practice
• Scanner resolutions are
• getting higher (1024x1024 per slice)

Efficient data memory layout essential!
3
Motivation (2/3) - Goals

• Interactive rendering/handling of large datasets
  up to 1 GB
• Support of heterogeneous PC hardware environment

• Test hardware specifications
• Notebook
• Pentium M 1.6 GHz
• 1 GB RAM
• GeForce 4 GO (32 MB)

Smart combination and modification of known
methods!
4
Motivation (3/3)

• Hierarchy of successively larger but slower
  memory technology
• Avoid frequent access to slower levels
• Exploit spatial and temporal locality

Memory hierarchy
Hard disk
Main memory
L2 cache
L1 cache
CPU
5
Outline

• Memory Layout Data Processing Scheme
• Gradient Caching
• Empty Space Skipping
• Parallelization Features
• Results

6
Linear Memory Layout (1/2)
2D Slice
15
12
13
14
8
9
10
11
4
5
6
7
3
0
1
2
Memory storage order
0
1
2
3
4
5
6
7
7
Linear Memory Layout (2/2)
Volume
Store volume as a stack of 2D images (slices)
View dependent cache behavior!
Rays
8
Bricked Memory Layout (1/2)
2D Slice
15
12
13
14
8
9
10
11
4
5
6
7
3
0
1
2
Memory storage order
0
1
4
5
2
3
6
7
9
Bricked Memory Layout (2/2)
Volume
Store volume as a set of equally sized cubes
(bricks)
Constant cache behavior!
Rays
10
Bricked-wise Processing
Volume
Processing of all resample locations is done
brick-wise
High Cache Coherence!
Rays
11
Outline

• Memory Layout Data Processing Scheme
• Gradient Caching
• Empty Space Skipping
• Parallelization Features
• Results

12
Gradient Caching (1/3)

• Pre-computed gradients
• ? High performance
• ? For sufficient quality, memory requirements are
  at least doubled
• Compute gradients on-the-fly
• ? Calculation expensive
• ? No additional storage requirement

13
Gradient Caching (2/3)
Cell

• To accelerate calculation ? Caching of gradients
• Brick-wise traversal allows to use a brick-sized
  gradient cache which can be re-used for each brick

14
Gradient Caching (3/3)
Volume
Gradient cache

• One brick-sized gradient cache
• Constant very small memory requirement

Rays
15
Outline

• Memory Layout Data Processing Scheme
• Gradient Caching
• Empty Space Skipping
• Parallelization Features
• Results

16
Empty Space Skipping (brick-level)

• Min-Max info contained in brick used for
  discarding empty regions
• Template based brick projection to rasterize
  depth values
• In software, very fast for orthographic
  projections

17
Empty Space Skipping (octree-level)

• Each brick contains three-level octree
• Caching of classification information
• Stored in linearized octree using hierarchy
  compression
• Octree goes down to 4x4x4 voxels
• Template based projection

Min-Max and classification caching increase the
memory requirements by approx. 4
18
Cell Invisibility Cache (1/2)
Example ray
Skipped by octree
Not skipped by octree
19
Cell Invisibility Cache (2/2)
NO
Re-sampling Gradient- Estimation Compositing Shadi
ng
Classi- fication
Advance ray
Visible
YES
YES
CIC
Visible
NO
CIC increase the memory requirements by approx.
6
20
Empty Space Skipping

• Project all non-transparent bricks onto image
  plane to find first entry points of rays
• For finer resolution, use a min-max octree per
  brick and project the octree
• Cell Invisibility Cache

All these acceleration techniques increase the
memory requirements by just 10
21
Outline

• Memory Layout Data Processing Scheme
• Gradient Caching
• Empty Space Skipping
• Parallelization Features
• Results

22
Parallelization / Hyper Threading
Law and Yagel 1996
Log. CPU 1
Advancing Ray-front
Log. CPU 2
Phy. CPU 1
Phy. CPU 2
Log. CPU 1
1D Screen
Log. CPU 2
23
Features
View aligned and axis aligned cutting planes
High quality
Multiple segmented object and Transfer-functions
Transfer-functions on clipping planes
24
Outline

• Memory Layout Data Processing Scheme
• Gradient Caching
• Empty Space Skipping
• Parallelization Features
• Results

25
Results (1/3) - Bricking
Linear vs. bricked memory layout
4
Optimal brick size
Speedup factor
3
Speedup 2.8
2
cache thrashing bricking overhead
linearvolumelayout
1
8
64
512
4096
1
32768
Brick size in kilo-byte
Cache size 512 KB
26
Results (2/3) Gradient Caching
Speedup 3.4
Speedup 2.7
Pentium M 1.6 GHz 1 GB RAM
27
Results (3/3) - Performance
Pentium M 1.6 GHz 1 GB RAM
28
Conclusions

• Sub second frame rates for large datasets on a
  standard notebook
• Fully interactive volume visualization of large
  data on commodity hardware is within reach
• Alternative memory layouts are the key to
  handling large datasets

29
Questions?
Visible Male(587 x 341 x 1878)
Intel Pentium M 1600 MHz(software capture)
30
Thank you for your attention
Sponsored by
Institute of Computer Graphic and Algorithms
Tiani MedGraph AG