DAG Construction Erik Sintorn Chalmers University of Technology

  • Slides: 47
Download presentation
DAG Construction Erik Sintorn Chalmers University of Technology, Sweden

DAG Construction Erik Sintorn Chalmers University of Technology, Sweden

Outline • Content • Voxelization and building an SVO • Efficient SVO to DAG

Outline • Content • Voxelization and building an SVO • Efficient SVO to DAG compression, first version • Efficient SVO to DAG compression, second version

Parallel primitives in CUDA • Our GPU Implementation is based on standard parallel primitives

Parallel primitives in CUDA • Our GPU Implementation is based on standard parallel primitives • Stream Compaction • Sorting • Prefix Sum • A number of libraries exist for doing these things • THRUST (http: //docs. nvidia. com/cuda/thrust/index. html) • CUDPP (http: //cudpp. github. io/) • CUB (https: //nvlabs. github. io/cub/)

Voxelization • Render geometry to voxel-grid sized rendertarget • Render triangles from X, Y

Voxelization • Render geometry to voxel-grid sized rendertarget • Render triangles from X, Y or Z axis that gives largest area • Single pass, choose axis in geometry shader • Avoid cracks with GL_NV_conservative_raster • Output long list of fragments as Morton order integers • Simple tutorial at https: //developer. nvidia. com/content/basics-gpu-voxelization

Build SVO in CUDA • parent coord child index

Build SVO in CUDA • parent coord child index

Build SVO in CUDA (leaf level) • Sort all voxels (and remove identical) 001010|001

Build SVO in CUDA (leaf level) • Sort all voxels (and remove identical) 001010|001 001010|010 001010|011 001011|001 001011|100 001100|100

Build SVO in CUDA (leaf level) • Sort all voxels (and remove identical) 001010|001

Build SVO in CUDA (leaf level) • Sort all voxels (and remove identical) 001010|001 001010|010 001010|011 001011|001 001011|100 010100|100 • Mark first child of all parents 1 0 0 1

Build SVO in CUDA (leaf level) • Sort all voxels (and remove identical) 001010|001

Build SVO in CUDA (leaf level) • Sort all voxels (and remove identical) 001010|001 001010|010 001010|011 001011|001 001011|100 010100|100 • Mark first child of all parents 1 0 0 1 1 2 • Prefix sum gives position of parents 0 0 0 1

Build SVO in CUDA (leaf level) • Sort all voxels (and remove identical) 001010|001

Build SVO in CUDA (leaf level) • Sort all voxels (and remove identical) 001010|001 001010|010 001010|011 001011|001 001011|100 010100|100 • Mark first child of all parents 1 0 0 1 1 2 • Prefix sum gives position of parents 0 0 0 1 • Build leaf level cords and childmasks 001010 001011 010100 00001110 00010000

Build SVO in CUDA (internal levels) • Start from coords in last pass 001|010

Build SVO in CUDA (internal levels) • Start from coords in last pass 001|010 001|011 010|100 • Mark first child of all parents 1 0 1 • Prefix sum gives position of parents 0 0 1 • Build next level cords childmasks and pointer 001 010 00001100|0 00010000|2

Creating a DAG from an SVO 1111 001 0110 005 1101 007 1111 010

Creating a DAG from an SVO 1111 001 0110 005 1101 007 1111 010 014

Creating a DAG from an SVO Level: 0 1111 001 1 0110 005 1101

Creating a DAG from an SVO Level: 0 1111 001 1 0110 005 1101 007 1111 010 014 2

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 2 -

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 2 - |000|001| - 1101 002|003| - |004 1111 005|006|007|008 0110 - |009|010| -

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001|

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001| - 1101 002|003| - |004 1111 005|006|007|008 0110 - |009|010| - 2 original idx: 0 1 2 3 4 5 6 7 8 9 10

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001|

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001| - 1101 002|003| - |004 1111 005|006|007|008 0110 - |009|010| - 2 original idx: 0 9 1 10 2 8 3 4 6 5 7

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001|

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001| - 1101 002|003| - |004 1111 005|006|007|008 0110 - |009|010| - 2 original idx: 0 9 1 10 2 8 3 4 6 5 7 sorted idx: 0 2 4 6 7 9 8 10 5 1 3

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001|

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001| - 1101 002|003| - |004 1111 005|006|007|008 0110 - |009|010| - 2 original idx: 0 9 1 10 2 8 3 4 6 5 7 sorted idx: 0 2 4 6 7 9 8 10 5 1 3 first unique: 1 0 1 0 0 1 1

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001|

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001| - 1101 002|003| - |004 1111 005|006|007|008 0110 - |009|010| - 2 original idx: 0 9 1 10 2 8 3 4 6 5 7 sorted idx: 0 2 4 6 7 9 8 10 5 1 3 first unique: 1 0 1 0 0 1 1 compact idx: 0 0 1 1 2 2 3 3 3 4 5

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001|

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001| - 1101 002|003| - |004 1111 005|006|007|008 0110 - |009|010| - 2 original idx: 0 9 1 10 2 8 3 4 6 5 7 sorted idx: 0 2 4 6 7 9 8 10 5 1 3 first unique: 1 0 1 0 0 1 1 compact idx: 0 0 1 1 2 2 3 3 3 4 5

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001|

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001| - 1101 002|003| - |004 1111 005|006|007|008 0110 |009|010| - |000|010| 2 original idx: 0 9 1 10 2 8 3 4 6 5 7 sorted idx: 0 2 4 6 7 9 8 10 5 1 3 first unique: 1 0 1 0 0 1 1 compact idx: 0 0 1 1 2 2 3 3 3 4 5

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 2 -

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 2 - |000|001| - 1101 002|000| - |000 1111 004|000|005|002 0110 - |000|001| -

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001|

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001| 0 2 1101 002|000| - |000 1111 004|000|005|002 0110 1 2 - |000|001| 3

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001|

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 - |000|001| 0 2 0110 - |000|001| 3 1101 002|000| - |000 1111 004|000|005|002 1 2

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 2 -

Creating a DAG from an SVO Level: 0 1111 000|001|002|003 1 0110 2 - |000|001| - 0110 - |000|001| - 1101 002|000| - |000 1111 004|000|005|002 0 3 1 2 0 2 3 1 1 0 0 1 2

Creating a DAG from an SVO Level: 0 1111 000|001|002|001 1 0110 2 -

Creating a DAG from an SVO Level: 0 1111 000|001|002|001 1 0110 2 - |000|001| - 1101 002|000| - |000 1111 004|000|005|002

Creating a DAG from an SVO Level: 0 1111 000|001|002|001 5 words 1 0110

Creating a DAG from an SVO Level: 0 1111 000|001|002|001 5 words 1 0110 - |000|001| - 1101 002|000| - |000 1111 004|000|005|002 4 words 3 words 5 words 2 1 word 1 word

Creating a DAG from an SVO Level: 0 1111 000|001|002|001 0 1 0110 -

Creating a DAG from an SVO Level: 0 1111 000|001|002|001 0 1 0110 - |000|001| - 1101 002|000| - |000 1111 004|000|005|002 5 8 12 2 17 18 19 20 21 22

Creating a DAG from an SVO Level: 0 1111 005|008|012|005 0 1 0110 -

Creating a DAG from an SVO Level: 0 1111 005|008|012|005 0 1 0110 - |017|018| - 1101 019|017| - |017 1111 021|017|021|019 5 8 12 2 17 18 19 20 21 22

Creating a DAG from an SVO 1111 005|008|012|005 0110 017|018 1101 019|017 1111 021|017|021|019

Creating a DAG from an SVO 1111 005|008|012|005 0110 017|018 1101 019|017 1111 021|017|021|019 Issues: • Using eight int 32 numbers as a sorting key is expensive • Since we start from the bottom of the SVO, the processing time will be dependent on the size of the leaf level

Creating a DAG from an SVO, Alternative Version

Creating a DAG from an SVO, Alternative Version

Creating a DAG from an SVO, Alternative Version

Creating a DAG from an SVO, Alternative Version

Creating a DAG from an SVO, Alternative Version

Creating a DAG from an SVO, Alternative Version

Creating a DAG from an SVO, Alternative Version

Creating a DAG from an SVO, Alternative Version

Creating a DAG from an SVO, Alternative Version

Creating a DAG from an SVO, Alternative Version

Creating a DAG from an SVO, Alternative Version

Creating a DAG from an SVO, Alternative Version

Compressing Large DAGs • Epic Citadel • 128 k x 128 k voxels •

Compressing Large DAGs • Epic Citadel • 128 k x 128 k voxels • DAG size ~1 GB • Pointerless (!) SVO 5. 1 GB

Compressing Large DAGs • Create a dense top of the tree:

Compressing Large DAGs • Create a dense top of the tree:

Compressing Large DAGs • Voxelize the first child

Compressing Large DAGs • Voxelize the first child

Compressing Large DAGs • Create a DAG

Compressing Large DAGs • Create a DAG

Compressing Large DAGs • Voxelize and Compress a second child

Compressing Large DAGs • Voxelize and Compress a second child

Compressing Large DAGs • End up with a partially reduced DAG

Compressing Large DAGs • End up with a partially reduced DAG

Merging one level Level 2 Level 3 Level 4

Merging one level Level 2 Level 3 Level 4

Merging one level Level 2 Level 3 Level 4

Merging one level Level 2 Level 3 Level 4

Merging one level Level 2 Level 3 Level 4

Merging one level Level 2 Level 3 Level 4

Merging one level Level 2 Level 3 Level 4

Merging one level Level 2 Level 3 Level 4

Merging one level Level 2 Level 3 Level 4

Merging one level Level 2 Level 3 Level 4

Merging one level Level 2 Level 3 Level 4

Merging one level Level 2 Level 3 Level 4

DAG Construction Erik Sintorn Chalmers University of TechnologyDAG Construction Erik Sintorn Chalmers University of Technology
DAG Construction Erik Sintorn Chalmers University of TechnologyDAG Construction Erik Sintorn Chalmers University of Technology
DAG SINTERKLAASJE Dag Sinterklaasje dag dag Dag dagDAG SINTERKLAASJE Dag Sinterklaasje dag dag Dag dag
Chalmers University of Technology Chalmers University of TechnologyChalmers University of Technology Chalmers University of Technology
Advanced Computer Graphics Introduction Erik Sintorn erik sintornchalmersAdvanced Computer Graphics Introduction Erik Sintorn erik sintornchalmers
Advanced Computer Graphics Introduction Erik Sintorn erik sintornchalmersAdvanced Computer Graphics Introduction Erik Sintorn erik sintornchalmers
Advanced Computer Graphics Introduction Erik Sintorn erik sintornchalmersAdvanced Computer Graphics Introduction Erik Sintorn erik sintornchalmers
Advanced Computer Graphics Introduction Erik Sintorn erik sintornchalmersAdvanced Computer Graphics Introduction Erik Sintorn erik sintornchalmers
STIFTELSEN CHALMERS INDUSTRITEKNIK Social media and HEPTech CHALMERSSTIFTELSEN CHALMERS INDUSTRITEKNIK Social media and HEPTech CHALMERS
Introduktion till Chalmers ITsystem Kim Tabrizi Chalmers LindholmenIntroduktion till Chalmers ITsystem Kim Tabrizi Chalmers Lindholmen
God dag HVILKEN DAG ER DET I DAGGod dag HVILKEN DAG ER DET I DAG
Chalmers University of Technology Gas turbine technology MTFChalmers University of Technology Gas turbine technology MTF
Chalmers University of Technology LCA as input toChalmers University of Technology LCA as input to
Chalmers University of Technology Power Estimation Flex SocChalmers University of Technology Power Estimation Flex Soc
Chalmers University of Technology Analysis of thermoacoustic propertiesChalmers University of Technology Analysis of thermoacoustic properties
Chalmers University of Technology Use of Latent VariablesChalmers University of Technology Use of Latent Variables
Chalmers University of Technology SiteDependent Electromagnetic Effects inChalmers University of Technology SiteDependent Electromagnetic Effects in
Chalmers University of Technology REDUCTION POTENTIAL OF ROADChalmers University of Technology REDUCTION POTENTIAL OF ROAD
Chalmers University of Technology Lecture 4 Ideal cyclesChalmers University of Technology Lecture 4 Ideal cycles
Chalmers University of Technology Dynamic Mode Decomposition appliedChalmers University of Technology Dynamic Mode Decomposition applied
Chalmers University of Technology SiteDependent Electromagnetic Effects inChalmers University of Technology SiteDependent Electromagnetic Effects in
Chalmers University of Technology CFD MODELING OF SYNGASChalmers University of Technology CFD MODELING OF SYNGAS
Chalmers University of Technology Metrics and stabilization ofChalmers University of Technology Metrics and stabilization of
Chalmers University of Technology Lecture 8 Axial turbinesChalmers University of Technology Lecture 8 Axial turbines