Hardwareaccelerated Rendering of Antialiased Shadows With Shadow Maps

1
Hardware-accelerated Rendering of Antialiased
Shadows With Shadow Maps

• Stefan Brabec and Hans-Peter Seidel
• Max-Planck-Institut für Informatik
• Saarbrücken, Germany

2
Overview

• Introduction
• Shadow mapping using graphics hardware
• Hardware-based percentage closer filtering
• A faster approach
• Results
• Conclusions

3
Introduction

• Hardware-accelerated rendering mostlydominated
  by local effects
• Triangle through-put
• Fast lighting texturing
• Most important global effect are shadows
• Two main classes of shadow algorithms
• Shadow Volumes Crow 77
• Shadow Maps Williams 78

4
Shadow Mapping

• Sampling based approach
• Generate depth map from light source position
• Store depth values of frontmost pixels
• Final rendering pass
• Foreach pixel seen by the camera
• Transform to light source coordinate system
• Compare corresponding entry in depth mapwith
  transformed camera depth

5
Shadow Mapping

• Shadow Maps
• Make their way to hardware
• SGI InfiniteReality, NVIDIA GeForce3, ATI
• Implemented as special texture map operation
• Compare value at (s/q,t/q) with r/q -gt result 1
  or 0
• Can also be implemented using standard OpenGL

6
Shadow Mapping

• Pros
• Does not depend on scene geometry
• Can be implemented using graphics-hardware
• Cons
• Sampling problems (aliasing)
• High resolution depth maps
• Filtering

7
Shadow Mapping

• Percentage closer filtering Reeves 87
• Filtering depth values makes no sense
• Perform shadow test before filtering

8
Shadow Mapping

• Our Shadow Mapping Approach
• Render scene as seen by light source
• Encode depth values in alpha channel
• Project this texture into the final scene
• Subtract depth values to determine lit pixels

9
Shadow Mapping

• 1D ramp texturebrings depth tocolor values

10
Hardware-based PCF

• How can Reeves scheme be adopted for
  hardware-based shadow mapping ?
• Idea
• Generate multi-channel depth mapwhere each entry
  contains also valuesof neighborhood pixels

11
Shadow Mapping

• Problem 1D ramp texture can only use 256 depth
  values because of frame buffer depth
• Only for scenes with moderate complexity
• But 16 bits possible with NVIDIAsregister
  combiners
• Use two color channels (high low 8 bits)
• See NVIDIA website for more details

12
Hardware-based PCF

• Multi-channel shadow map
• Use RGBA instead of alpha channel only
• 4 values to sample a 2x2 region
• Increases effective shadow map resolution by a
  factor of 2in each dimension
• Shadow map generation
• Render scene four times where in each pass
• One channel (R,G,B or A) is selected
• Image-plane is jittered (stratified sampling)
• Copy RGBA image to texture

13
Hardware-based PCF

• Stratified sampling and pixel packing

14
Hardware-based PCF

• Computing the percentage of shadowing
• Final pass using all four channels
• R camera_z light_z_sample_1
• G camera_z light_z_sample_2
• B camera_z light_z_sample_3
• A camera_z light_z_sample_4
• Channel is in shadow if value gt 0

15
Hardware-based PCF

• Computing the percentage of shadowing
• Implemented using OpenGL Imaging Subset

64,63 corresponds to 25 shadow Sums up to 255
for full shadow
16
Hardware-based PCF

• Larger filter sizes possible
• Split up into parts of max. 4 components
• Sum up shadow contributions using
  theaccumulation buffer
• Problem
• Number of rendering passes (scene geometry)not
  really practical for dynamic scenes

17
Fast PCF

• Problem
• Previous algorithm needs four passesonly for
  shadow map generation
• Unsuitable for very complex scenes
• Fast PCF
• Use only one rendering pass and
  collectneighborhood pixels

18
Fast PCF

• Fast PCF filtering and pixel packing

19
Fast PCF

• Collect pixels using convolution

• Currently the only hardware-acceleratedoperation
  to include information from adjacent pixels

20
Fast PCF

• Assemble pixels using neighborhood channels

• Center pixel consists of RGBA values taken from
  lower-left 2x2 region
• Use larger (3x3) filter for best hardwaresupport

21
Fast PCF

• Texture coordinate offset
• Shift center pixel to be centered in 2x2 region

22
Fast PCF

• Fast PCF only uses one rendering passfor shadow
  map generation
• Suitable for larger filter sizes
• Example 4x4 footprint
• Four passes which will sample the lower right,
  lower left, upper right and upper left2x2 region
• Results in 4520 shadowing levels
• Four shadow map textures used

23
Results
without filtering
24
Results
normal PCF (filter size 2x2)
25
Results
fast PCF (filter size 2x2)
26
Results
real PCF vs. fast PCF
27
Results
fast,multipass PCF (filter size 4x4)
28
Results

• Timings Overview

Machine SGI Octane VPro/8Image Resolution
800x600Shadow Map 512x512Polygons 7000
Frame rate
Generation
Shadowing
No filtering
30 fps
1 pass
2 levels
PCF 2x2
10 fps
4 passes
5 levels
Fast PCF 2x2
20 fps
1 pass
5 levels
PCF 4x4
lt 0.5 fps
16 passes
20 levels
Fast PCF 4x4
5 fps
4 passes
20 levels
29
no filtering
30
fast PCF (filter size 2x2)
31
Conclusions

• Benefits
• Good-looking, anti-aliased shadows
• Hardware-based (fast)
• Standard OpenGL
• Drawbacks
• OpenGL Imaging Subset not everywhere available
• Depth to color mapping needs more than 8 bits
  per channel for complex scenes

32
Conclusions

• Future Work
• Now only constant filter sizes
• Better mipmap-like method to compute filter
  sizeusing the projected area

33
Thank You !
Visit us at http//www.mpi-sb.mpg.de Max-Planck-In
stitut für Informatik Computer Graphics Group