VIZARD II: A Reconfigurable Interactive Volume Rendering System

1
VIZARD II A Reconfigurable Interactive Volume
Rendering System
M. Meißner, U. Kanus, G. Wetekam, J. Hirche, A.
Ehlert, W. Straßer, M. Doggett, P. Forthmann,
and R. Proksa

• WSI / GRIS, Universität Tübingen
• Philips Research, Hamburg
• Viatronix
• ATI

2
Overview

• Motivation
• Features
• System
• RPU
• Results Future Work
• Acknowledgments

3
Motivation
Need for interactive frame rates, the highest
possible image quality, interactive change of
parameters, and different rendering modes.
Learn from PC graphics ( from keynote) Software
(CPU) scales only with Moores Law but hardware
can additionally scale by implementing the full
algorithm (pipeline) in hardware and possibly
parallel pipelines (SIMD)
4
Motivation
However, hardware (ASIC) is expensive, pays out
only for large quantities, and change/ adaption
in design requires new fabrication.
Thus, combine the performance of special purpose
hardware with (almost) the flexibility of a
software solution using reconfigurable logic
(FPGA).
Board can be re-used for other applications, e.g.
volume reconstruction (Philips)
5
Overview

• Motivation
• Features
• System
• RPU
• Results Future Work
• Acknowledgments

6
Features

• Rays are cast from viewplane
• Flexible sampling rate
• Trilinear interpolation
• Post-classification
• Per sample Phong shading
• High precision compositing
• Acceleration Techniques
• Geometry based space leaping
• Early ray termination

7
Features

• Ensure high(est) image quality by
• unlimited ray-casting including parallel
  and perspective projections
• complex gradient filters mandatory for
  correct GM (? gradient as per voxel property)
  Left Original volume rendering Right
  Central difference, GM (gt1),
  result depends on object
  orientation due to non symmetric
  gradient filter

8
Features

• Ensure high(est) image quality by
• full classification with material properties
  on a per sample base
• additional interval based classification mode
  (pre-integration) and accurate combination
  with Phong shading Meissner et al. GI2002

9
Overview

• Motivation
• Features
• System
• RPU
• Results Future Work
• Acknowledgments

10
System

• Memory interface
• Trilinear interpolation requires eight Voxels
  in parallel. Ideally, one would like to have
  eight memory devices
• For more flexibility, use exchangeable DIMMs
  instead of single SDRAM devices
• DIMM modules are large in physical size and
  come in 64 data bits, thus use four eight
  would not fit anyway - and store two voxel
  values (32 bit each) in one entry.

11
System

• Cubic memory organization Volume is
  partitioned in sub-cubes and stored in linear
  memory (4 DIMMs, 64 bit)

DIMM0
DIMM1
DIMM2
DIMM3
32x32x32 Voxel fit into Caches of 4 DIMMs
--gt ideal access time (e.g. 10ns)
12
System

• Non blocking access using prefetching Delay
  of page x-ings can be overlapped in time.
  Doggett et al. HWW1999

No
Time
Yes
13
System
PCI card (long!)
Host computer
14
System
DIMMs
Power supplyand converters
SRAM
Reconfigurable chip (FPGA)
DSP
PCI interface chip
15
Overview

• Motivation
• Features
• System
• RPU
• Results Future Work
• Acknowledgments

16
RPU

• Central unit of VIZARD II is the RPU

RPU
17
RPU Compositing
Compositing _at_ 100 MHz is a challenge
Opacity Ai1 Ai (1-Ai) As Inverse,
multiply and subsequent Add of two 16 bit values
is not feasible in 10ns Use Threading One
compositing unit,
multiple processed rays Hesser VG 99
18
RPU Compositing

• Iteratively process n rays instead of one

gt Eight rays allow frequency of 100 MHz Better
memory efficiency (overall less page x-ings)
Higher efficiency of early ray termination(overal
l pipeline latency is divided by number of
rays) ? Use early ray group termination
19
Overview

• Motivation
• Features
• System
• RPU
• Results Future Work
• Acknowledgments

20
Results

• Physical Design (2 million gates)

Memory-Controller(s)
CU SU
BU
DSP-I/O
RC AU
TIU
21
Results
22
Results

• DEMO

23
Results

• VIZARDII VolumePro
  TexMap
• Pipelines,_at_MHz 1_at_50 4_at_250
  4_at_300
• Speed 3-7 gt 30
  lt10
• Perspective yes (no) (yes)
• Precision 16 12
  8-10
• iso-surface yes no
  (yes)
• Pre-integrated yes no
  (yes)
• ERT Co yes yes
  (no)

24
Results

• Flexibility Dual use of the same board!!!
• Philips Research developed a design for volume
  reconstruction from projected images (C-arm CT,
  currently at 50 MHz)
• 1. Reconstruction in 1 min (SW gt 30 min)
• 2. Subsequently interactive volume
  rendering of reconstructed data

25
Future Work

• On-chip ray setup
• Remove current bottleneck of system
• Space Leaping
• Content based real-time check of
  contributing 163 subcubes (50x per second)
  Meissner et al. VG2001
• Non-photorealistic rendering
• Technical drawings

26
Overview

• Motivation
• Features
• System
• RPU
• Results Future Work
• Acknowledgments

27
Acknowledgements

• Work
• Fellows and many many students
• Funding
• German Research Council (DFG), grant 382

28
Conclusion

• Graphics Hardware