1. Ray-casting in VolumePro™ 1000
Yin Wu, Vishal Bhatia, Hugh Lauer, Larry Seiler

2. VolumePro™ 1000 Summary
Second generation real-time volume rendering accelerator
Ray-casting at 109 samples per second
Ray-per-pixel image quality
Translucent & opaque embedded polygons
8-, 16-, & 32-bit voxels (up to four fields)
Geometry-based space leaping, early ray termination …

3. The Challenge in Ray-casting Performance vs. Image Quality
Shear-Warp
Traverse & resample data in memory order.
Warp needed for final image.
Fast
Efficient memory access
VolumePro 500
Image Quality
2nd resampling
No embedded geometry
Full Image order
Traverse & resample data in pixel order
Image Quality
No 2nd resampling
Embedded geometry
Performance
Memory accesses similar to random access.

4. VolumePro 1000 Ray-casting
Rays through pixels on image plane
Image quality equiv. to full image order
No 2nd resampling

5. VolumePro 1000 Ray-casting
Rays through pixels on image plane …
Samples organized in planes parallel to faces of the volume
Traverse & process data in memory order
Maximize memory performance

6. xy-image order — 2 parts (aka shear-image order)
Voxel processing part
Traverse data slice-by-slice in memory order
Read voxels for each slice of samples
Voxel-oriented processing (e.g., gradient estimation)
Store in on-chip buffers
Sample processing part
Define sample points where rays intersect slices
Traverse & interpolate on-chip buffer in pixel order
Sample-oriented processing
e.g., illumination, filtering, depth testing, compositing
Output to image

7. VolumePro 1000 ray-casting (Additional optimizations)
Section: rays assoc. with tile of image plane
Minimize on-chip buffer space
All slices of a section processed before next section
Enables early ray termination
Mini-block and stamp organized memory
Burst accesses to 222 voxels or 22 pixels
From VolumePro 500
Skewed across eight memory channels
Parallel access to 8 adjacent mini-blocks or stamps in any dimension
From VolumePro 500, Cube-4 (SUNY Stony Brook)

8. Block Diagram Sequencer Pipelines 250 MHz PCI bus Interface 33-66 MHz
Voxels (organized as mini-blocks)
16-node SIMD processor
Sequencer
Voxel processing
PCI bus Interface MHz
control
On-chip Slice buffers
Memory Interface eight channels 16-bit DDR SDRAM
( MHz)
Sample processing
Sample processing
control
Pixels (organized as stamps)

9. Technical summary Four 250 MHz pipelines, 109 samples per second
Trilinear interpolations on seven channels
Four colors or voxel fields
Three gradient components
Classification of (up to) four voxel fields
Phong illumination calculation
25-30 individual visibility tests
Alpha correction, gradient magnitude modulation
Perspective rendering (in Shear-Warp)

10. What next? Is 109 samples/second enough? Major items yet to be done
1 megapixel at 15 fps  ~66 samples/ray (average)
Not very many, even with aggressive space leaping
Major items yet to be done
Content-based space-leaping
Faster memory, pipelines, and Sequencer
Perspective volume rendering
More programmability

11. Pretty Pictures

14. Supplementary Slides

15. Embedding Surfaces in Volumes
Allows inserting pointers, medical prostheses, or geological data markers into the volume
Also supports arbitrary clipping regions, all in real time
Each ray is cast in multiple segments, between surfaces specified by depth buffers
Surface rendering performed in 3D graphics chip using OpenGL
Front clip bounds
Trans- lucent surface
So how do we do it? First, we render the polygon data using OpenGL or Direct3D on the standard 3D graphics chip.
This produces one or more layers of depth and image data. The example at the right shows an opaque background behind a closed translucent surface that produces a front layer and a back layer. A fourth depth layer specified a portion of the volume that we want to clip out.
Opaque background

16. Embedding Surfaces in Volumes
Allows inserting pointers, medical prostheses, or geological data markers into the volume
Also supports arbitrary clipping regions, all in real time
Each ray is cast in multiple segments, between surfaces specified by depth buffers
Surface rendering performed in 3D graphics chip using OpenGL
Next, we start compositing samples along the rays, until they reach another depth-image layer.
We then composite the surface colors into the rays...

17. Embedding Surfaces in Volumes
Allows inserting pointers, medical prostheses, or geological data markers into the volume
Also supports arbitrary clipping regions, all in real time
Each ray is cast in multiple segments, between surfaces specified by depth buffers
Surface rendering performed in 3D graphics chip using OpenGL
... And continue compositing samples along the rays until they reach the next surface.

18. Embedding Surfaces in Volumes
Allows inserting pointers, medical prostheses, or geological data markers into the volume
Also supports arbitrary clipping regions, all in real time
Each ray is cast in multiple segments, between surfaces specified by depth buffers
Surface rendering performed in 3D graphics chip using OpenGL
Finally, when all of the rays have reached the opaque background, we can stop ray-casting.
TURN ON DEMO: TEAPOT with LOBSTER, play with CROP PLANES