Controlling Memory Consumption of Hierarchical Radiosity with Clustering

1
Controlling Memory Consumption ofHierarchical
Radiosity with Clustering

• Xavier Granier George Drettakis

iMAGIS -GRAVIR/IMAG-INRIA iMAGIS is a joint
project of CNRS/INRIA/UJF/INPG
Graphics Interface 99
2
Overview

• Motivation and Previous Work
• New Framework
• Reduce memory used by links
• Reduce memory used by hierarchy
• Memory Control Mechanism
• Results
• Conclusion

3
Overview

• Motivation and Previous Work
• New Framework
• Reduce memory used by links
• Reduce memory used by hierarchy
• Memory Control Mechanism
• Results
• Conclusion

4
Motivation

• Global illumination of large scenes
• Hierarchical Radiosity with Clustering
  Sil95,SAS94
• fast global illumination
• view-independent
• High memory consumption due to
• Link structure WH97
• Hierarchy structure (clusters, subdivided
  polygons)

5
Hierarchical radiosity
?

• Multi-resolution representation of exchanges
• Clustering

6
Three hierarchy traversals

• Refinement
• build the hierarchy
• create link store information of the exchanges
• Gather
• compute irradiance due to links Ii ?links on i
  FijBkj
• Push-Pull
• Descend hierarchy sum the irradiances
• Leaves reflect entire irradiance
• Returning update the hierarchical representation

7
Hierarchical radiosity
B 0
?1
B E
B 0
?2
8
Refinement
9
Gather
I1 F1E
I I1
10
Push-Pull Descend
I2
I1 I2
I1
11
Push-Pull Return
B1r1I1
B (SAiBi)/A
12
Previous work

• Progressive refinement CCWG88
• Lack of global error control
• Memory used by links
• Study by Willmott and Heckbert WH97
• Getting rid of Links SSSS98
• unshot radiosity
• link cache

13
Shooting Algorithm SSSS98

• Shooting algorithm
• DB0 E (sources)
• DB0 0 (others)
• Theoretically the same as standard algorithm
• Convergence global error

14
Shooting algorithm
B 0 DB 0
?1
B E DB E
B 0 DB 0
?2
15
1st Iteration
B 0 DB 0
?1
B E DB 0
B ?2 I2 DB ?2 I2
?2
16
2nd iteration
B ?1 I1 DB ?1 I1
?1
B E DB 0
B ?2 I2 DB 0
?2
17
3th iteration
B ?1 I1 DB 0
?1
B E DB 0
B ?2 (I2 I3) DB ?2 I3
?2
18
Overview

• Motivation and Previous Work
• New Framework
• Reduce memory used by links
• Reduce memory used by hierarchy
• Memory Control Mechanism
• Results
• Conclusion

19
Goals of New Framework

• Reduce both link and hierarchy memory
  consumption
• Low speed penalty
• Maintain global representation

20
New Framework for Memory Reduction

• Goal
• Keep only links needed in next iteration
• Algorithm
• Unshot radiosity SSSS98 DB Bi1 - Bi
• Merge radiosity steps
• Traverse Link Hierarchy

21
New Framework for Memory Reduction

• Link Hierarchy and Traversal
• Allows reduction of stored links
• Reduce links creation number of links
• Remove links where possible
• Merge radiosity steps
• Facilitates reduction of memory used by radiosity
  hierarchy
• ...allows us to move links higher in the
  hierarchy
• Enables effective memory control mechanism

22
Link hierarchy DS97
Is the father of

• Active link hierarchy leaf
• Passive link hierarchy node

23
Overview

• Motivation and Previous Work
• New Framework
• Reduce memory used by links
• Reduce memory used by hierarchy
• Memory Control Mechanism
• Results
• Conclusion

24
Link Memory Reduction

• Only create links when needed
• Merge Refine and Gather
• Creation Criterion
• First approach Dont store links from sources
• More sophisticated approaches possible
• Control link creation using a cache mechanism
• Predict link utility in future iterations

25
Refine And Gather - schema
source
create
create
create
gather
gather
gather
create and gather
create and gather
refine
refine

• Active link hierarchy leaf
• Passive link hierarchy node

26
Link reduction summary

• New approach reduces memory used by links
• No overall control of memory
• Now, hierarchy uses most of the memory

27
Overview

• Motivation and Previous Work
• New Framework
• Reduce memory used by links
• Reduce memory used by hierarchy
• Memory Control Mechanism
• Results
• Conclusion

28
Hierarchy storage reduction

• Goal reduce memory during refinement
• Otherwise we cannot control overall memory usage
• Full recursion on link hierarchy
• Refine, Gather and Push Pull method
• Hierarchy simplification

29
Refine, Gather And PushPull

• We have to do only one PushPull
• For each hierarchy element
• For each iteration
• Receiver refinement
• Refine gather and push pull on each child
• Source refinement
• Refine gather and push pull on last child
• Else refine gather
• Replace hierarchy on which no links arrive

30
Refine Gather and PushPull
source
create and gather
create and gather push pull
gather
refine and push pull
refine and push pull
gather and push pull
gather and push pull

• Active link hierarchy leaf
• Passive link hierarchy node

31
Refine Gather and PushPull
source
create
replacement

• Active link hierarchy leaf
• Passive link hierarchy node

32
Texture replacement

?
33
Hierarchy reduction summary

• Advantages
• Reduce subdivision due to direct light
• But
• still a memory peak
• Need a mechanism to limit memory

34
Overview

• Motivation and Previous Work
• New Framework
• Reduce memory used by links
• Reduce memory used by hierarchy
• Memory Control Mechanism
• Results
• Conclusion

35
Memory control mechanism

• Control link memory
• Move links higher in the element hierarchy
• Increase texture replacement
• Cache-like test
• Estimate the expected depth of Link hierarchy

36
Overview

• Motivation and Previous Work
• New Framework
• Reduce memory used by links
• Reduce memory used by hierarchy
• Memory Control Mechanism
• Results
• Conclusion

37
Tests scenes
Medium hall 4 blocks - 169 K Polygons
38
Tests scenes
Simple hall 2 blocks - 65 K polygons
complex hall 16 blocks - 676 K polygons
39
Memory used by HRC
40
New algorithm (texture replacement)
41
Cluster reduction
42
Overview

• Motivation and Previous Work
• New Framework
• Reduce memory used by links
• Reduce memory used by hierarchy
• Memory Control Mechanism
• Results
• Conclusion

43
Summary

• New framework for controlling memory
• Using Link Hierarchy
• Merge all steps of the radiosity solution
• Memory control mechanism
• Change gt Use memory where its needed

44
Future Work

• Store part of hierarchy on disk
• load only the parts needed in memory
• Better representation for simplified clusters
• image based rendering
• volumetric primitives
• multi-resolution geometric simplification
• More sophisticated memory control mechanisms
• take the memory of hierarchy into account