Level 0 Level 1 Level 2 - Linköping Universitymarfa45/egev2020/posters/poster-sub1933.… · 1st...
Transcript of Level 0 Level 1 Level 2 - Linköping Universitymarfa45/egev2020/posters/poster-sub1933.… · 1st...
-
Ratios 180x 4000x 27x 2x 23x 20x 25x
Evaluation of SVDAG-inspired compression and rendering, applied to VTK HyperTreeGrid:
- Compression of scientific volume data
- Surface RayTracing of HyperTreeGrid on Nvidia GPUs
- Reduction of the memory footprint for surface rendering of HyperTreeGrid
-
Evaluation of Mesh Compression and GPU Ray Casting for Tree Based AMR data in VTKAntoine Roche1,2,3 and Jérôme Dubois1,3
1 Université Paris-Saclay, CEA, Laboratoire en Informatique Haute Performance pour le Calcul et la simulation, 91680 Bruyères-le-Châtel, France
2 MSc Student, Université de Reims Champagne-Ardennes, France
3 CEA, DAM, DIF, F-91297 Arpajon, France
✓ Efficient memory-wise for sparse simulation data representation
✓ Advanced filtering available in ParaView / VTK
✓ User-controlled hierarchical data reduction in Paraview / VTK to accelerate analyses [3,4]
✓ Generally efficient surface rendering as an unstructured mesh: N3 volume data - N2 surface data
✓ Distributed parallel computations through whole trees dispatch
References[1] Harel, G., Lekien, J., & Pébaÿ, P.P. (2017). Two New Contributions to the Visualization of AMR Grids: I. Interactive Rendering of Extreme-Scale 2-Dimensional
Grids II. Novel Selection Filters in Arbitrary Dimension. ArXiv, abs/1703.00212.
[2] Harel, G., Lekien, J., & Pébaÿ, P.P. (2017). Visualization and Analysis of Large-Scale, Tree-Based, Adaptive Mesh Refinement Simulations with Arbitrary
Rectilinear Geometry. ArXiv, abs/1702.04852.
[3] Dubois, J., Harel, G., & Lekien, J. (2018). Interactive Visualization and Analysis of High Resolution HPC Simulation Data on a Laptop With VTK. 2018 IEEE
Scientific Visualization Conference (SciVis), 127-141.
[4] DUBOIS J., LEKIEN J.-B.: Highly efficient controlled hierarchical data reduction techniques for interactive visualization of massive simulation data. Eurovis19.
[5] KÄMPE V., SINTORN E., ASSARSSON U.: High resolution sparse voxel dags. ACM Trans. Graph. 32, 4 (July 2013), 101:1–101:13. URL: http://doi.acm.org/10.1145/2461912[6] Dan Dolonius, Erik Sintorn, Viktor Kämpe, Ulf Assarsson. Compressing Color Data for Voxelized Surface Geometry. Extended version. IEEE Transactions on Visualization and Computer Graphics, 2018.
[7] KEN MUSETH. VDB: High-Resolution Sparse Volumes with Dynamic Topology. ACM Transactions on Graphics, Vol. 32, No. 3, Article 27, Publication date: June 2013.[8] Rama Karl Hoetzlein. GVDB: Raytracing Sparse VoxelDatabase Structures on the GPU. High Performance Graphics (2016).[9] WALDI., JOHNSONG. P., AMSTUTZJ., BROWNLEEC.,KNOLLA., JEFFERSJ., GÜNTHERJ., NAVRÁTILP.: OSPRay-A CPUray tracing framework for scientific visualization.IEEEtransactions on visualization and computer graphics 23, 1 (2016), 931–940.
8x6 HyperTrees
One HyperTree
In this work, we evaluated AMR mesh compression for the rendering of SciVis volume data.
We achieved great improvements over native HTG rendering scheme :
- Reduction of memory footprint by ratios of 3-7x
- Load data for preview at full scale 50-100x faster
- Make it possible to render data on GPU, originally impossible
Level 0 Level 1 Level 2
Advanced filtering User-controlled hierarchical data reduction
✓ High compression ratios of octree on surface Computer Graphics scenes
✓ Very efficient surface RayTracing on Nvidia GPUs using CUDA
Computer Graphics – Compressed Octree (SVDAG) [5]Scientific Visualization - VTK HyperTreeGrid [1,2]
✓ Extreme compression: number of octrees nodes for scientific visualization volume data
✓ Enables the interactive surface rendering for some HyperTreeGrid models on the GPU,
otherwise not possible with current native VTK rendering
✓ Almost instant preview of data at full simulation resolution instead of dozens of seconds
Results - Combined Data Reduction
SVDAG-inspired compression on a quad-tree (visible nodes only)
4x less leaf nodes - 2x less including parent nodes
Proposal
HyperTreeGrid and advanced filtering Compression and RayTracing
Surface RayTracing
✓ Implementation of a HyperTreeGrid to SVDAG conversion
✓ Evaluation of compression and rendering of HyperTreeGrid surface and volume models as SVDAG
✓ Benchmark of an export of HyperTreeGrid as SVDAG to storage for preview purposes
Contributions
1 HyperTreeGrid 48 HyperTreesGrid of
48 SVDAGsGrid of
1 SVDAG
Grid of
16 SVDAGs
X X X X
Grid of
4 SVDAGs
1st iteration merge merge merge
3,1E+082,2E+08
4,6E+07
4,0E+067,0E+06
1,0E+08
1,0E+07
1,7E+06
5,1E+04
1,7E+06 2,0E+06
3,0E+05
5,0E+06
4,0E+05
1,0E+04
1,0E+05
1,0E+06
1,0E+07
1,0E+08
1,0E+09
Asteroid FuelAssembly Sphere Lucy Dragon Hairball Bunny
Compression of the trees – number of nodes HyperTreeGrid
SVDAG
2 2,5 49
252
1005
4774
147 149 155173
217351 933
3247
0
1 000
2 000
3 000
4 000
5 000
6 000
7 000
8 000
32 64 128 256 512 1024 2048 4096
Vid
eo
mem
ory
footp
rin
tin
Mbyte
s–
Max 8
192 M
Byte
s
Grid resolution – number of voxels per 3D axis
GPU memory footprint of Hairball modelvoxelized at different resolutions
HyperTreeGrid
SVDAG
Out of
memory
HyperTreeGrid SVDAG HyperTreeGrid SVDAG
110s
14GB
1s
1,8GB25s
6,5GB
0,5s
1,8GB
Conclusion
In future work, we plan to:
- investigate attribute compression [6], for HyperTreeGrid volume data
- Explore other AMR representations such as VDB [7]
- Evaluate volume rendering of HyperTreeGrid data on
- Nvidia GPUs with GVDB [8]
- x86 CPUs with Intel OSPRay [9].
Drastic reduction of:
- Time to first picture from disk, 50-100x
- Overall memory footprint, 3-7x
Ideal for previewing data at full scale
Not possible to further filter previewed data
Volume models Surface models
Rendering for the Visualization Toolkit (VTK) HyperTreeGrid: memory consumption can be very large
Generation of an unstructured mesh with as much surface data as volume data for extreme cases
For same mesh, surface representation is on average 10x more costly in memory than HyperTreeGrid
In extreme cases, surface representation does not fit on a GPU
Problem
Voxel Grid 323 643 1283 2563 5123 10243 20483 40963
GPU memory
(MB, Max 8192)2 2,5 49 252 1005 4774
Out of
Memory
Out of
Memory
Difference between
10243 (left) and 40963
(right) grid resolution
for Hairball model
48x
HypertreeGrid volume rendered on a Nvidia GPU with GVDBGTX 1050, 4Go VRAM
Future Work
http://doi.acm.org/10.1145/2461912