Stencil Routed A-Buffer

1
Stencil Routed A-Buffer

• Kevin Myers and Louis Bavoil
• NVIDIA

2
Our Cool Thing
3
What is it?

• A-Buffer
• Simply a list of fragments per-pixel
• The A-buffer, an antialiased hidden surface
  method Carpenter 84
• Related Work
• Depth Peeling Mammen 89 Everitt 01
• k-Buffer Bavoil et al. 07

4
Why do I need this?

• Often want more than nearest
• Alpha blending
• Volume rendering
• Collision detection
• Refraction and caustics
• Global illumination

5
Why is it hard?

• GPUs optimized to capture nearest layer
• Z buffering and early z test
• Fine for most real-time lighting models
• Wasteful if not rendering front to back

6
Things that dont work

• Blending cant just turn of z-buffering
• Most operations non-commutative
• MRT
• Cant direct output
• Reading what youre writing
• Hazardous
• Multi-Layer Depth Peeling via Fragment Sort
  Liu et al. 06
• k-Buffer Bavoil et al. 07

7
A-Buffer

• A list of fragments per-pixel
• Anything on the GPU that resembles this?
• MSAA
• A list of samples per-pixel
• Samples store coverage

8
MSAA in review

• Multisampled Antialiasing
• Fragments are rasterized at a higher res
• 8xMSAA 8 x aliased resolution
• Pixel shader is run once per-pixel
• Frame buffer storage is at sample resolution

9
Say What?

• MSAA samples A-Buffer pixels??
• MSAA sample patterns dont help
• Need all MSAA samples at pixel center

10
Line up your Sub-samples

• Turn off multisampling
• Still render to an MSAA buffer
• Pixel shader output bloats to all sub-samples
• BOOL D3D10_RASTERIZER_DESCMultisampleEnable
• Now writing 8 samples per pixel
• All have the same value!!

11
Bloating Your Pixel

• Applause?
• Meets the definition
• List of fragments per-pixel
• Not exactly what we want
• Each item contains same value
• Next fragment will clobber the entire list
• Need to update one entry in the list
• Once and only once

12
Stencil Routing
Stencil always increments Stencil passes when 4
13
Stencil Routing

• First introduced by Purcell et al 2003
• Did not work for general rasterization
• Tile aligned points
• Fat point is spread across four pixels
• Four pixels get same value
• Stencil allows one pixel to update

14
Stencil Routing and MSAA

• Stencil always operates at sample res
• Regardless of MultisampleEnable state
• DX10 Spec
• Use sub-samples to route
• Allows any pixel shader output to be routed
• Arbitrary primitives

15
Stencil Routing and MSAA
16
A Stencil Test That Works

• StencilFunc
• D3D10_COMPARISON_EQUAL
• StencilRef
• 2
• More on this later
• StencilPassOp and StencilFailOp
• D3D10_STENCIL_OP_DECR_SAT

17
Initializing Stencil

• Clear stencil buffer to pass value ( 2 )
• Initializes sample 0 to 2
• Use SampleMask to selectively update
• Stencil set to replace with refrence value

18
Why start at 2?

• When all sub-samples are written
• Most stencil values will be 0
• Except the last one written
• Last sample written stencil 1
• When overflow occurs
• All stencil values will be 0

19
Occlusion Query Test
Pixel did not overflow
Pixel overflowed
20
Handling Overflow

• Set sample mask to last sample updated
• Draw full screen quad
• Issue an occlusion query
• Set stencil to pass if stencil 0
• Check occlusion query
• Sample pass count overflow count

21
Handling Overflow

• Occlusion query
• Good
• Very fast
• Allows for dynamic A-Buffer sizing
• Bad
• Requires some CPU intervention
• Ideally A-Buffer size is fixed

22
Demo
Demo Time!
23
Secrets of the Dragon

• Single A-Buffer
• RG32F
• R is packed color
• G is depth
• Saves on texture loads
• Post process sort
• 8 fragment per-pixel bitonic sort
• Additional fragments, insertion sort

24
8800 GTX Performance
Alpha Blended Stanford Dragon
8 Layers Depth Peeling 8xABuffer ABuffer Speedup
640x480 30.9 164 5.3
800x600 30.4 139 4.6
1024x768 29.5 110 3.7
1280x960 28.1 81.4 2.9
1600x1200 26.2 54.9 2.1

16 Layers Depth Peeling 8xABuffer ABuffer Speedup
640x480 15.5 76.7 4.9
800x600 15.3 63.0 4.1
1024x768 14.7 48.0 3.3
1280x960 14.1 34.6 2.5
1600x1200 13.3 23.1 1.7
25
LimitsDOH!

• 254 layers of depth max
• 8-bit stencil ( 255 1 for overflow bit )
• If you do this call us cause thats crazy
• Fragments at same depth
• Must be handled in post-process
• MSAA

26
Summary

• Stencil Routed A-Buffer
• Ideally suited for complex geometries
• Much faster than depth peeling
• A-buffer can be dynamically resized
• Use an occlusion query
• Best to pre-determine size

27
Future Work

• Render target arrays
• Each target has its own stencil buffer
• Target replaces sub-sample
• Or augments sub-sample
• arrays MSAA level in one CPU pass
• With dx10 saturates 254 layers
• Use instancing for additional GPU passes

28
Thanks for all the fish

• Claudio Silva, Steven Callahan, Joao Comba, Aaron
  Lefohn, Cass Everitt, Peach Myers

29
The last slide

• ?
• kmyers_at_nvidia.com
• lbavoil_at_nvidia.com