raytracing scientific data in
play

RAYTRACING SCIENTIFIC DATA IN (R) NVIDIA OPTIX WITH TM GVDB - PowerPoint PPT Presentation

April 4-7, 2016 | Silicon Valley RAYTRACING SCIENTIFIC DATA IN (R) NVIDIA OPTIX WITH TM GVDB SPARSE VOLUMES Rama Karl Hoetzlein, April 6 th 2016 CONVERGENCE Video Motion Games Pictures CONVERGENCE Physical Sciences Video


  1. April 4-7, 2016 | Silicon Valley RAYTRACING SCIENTIFIC DATA IN (R) NVIDIA OPTIX WITH TM GVDB SPARSE VOLUMES Rama Karl Hoetzlein, April 6 th 2016

  2. CONVERGENCE Video Motion Games Pictures

  3. CONVERGENCE Physical Sciences Video Mathematics Motion SpaceX Games Pictures Data Visualization

  4. CONVERGENCE Physical Sciences Video Mathematics Motion SpaceX Games Pictures GVDB Data Visualization ProVis / CAD NVIDIA OptiX

  5. OUTLINE - Science - SpaceX - Volume Visualization - Motion Pictures – OpenVDB - Provis Graphics – NVIDIA OptiX - GVDB S parse Volumes - Results

  6. SPACEX Wavelet Simulation Source Data S paceX Dragon Capsule Re-entry 3D S parse Volume data: 3864 x 944 x 464 Adam Lichtl, S tephen Jones, GPU Technology Conference 2015

  7. SPACEX Data Preparation Data Visualization Wavelet Cloud

  8. A Numerical S tudy of High-Pressure Oxygen/ Methane Mixing and Combustion of a S hear Coaxial Inj ector, Nan Zong & Vigor Y ang, AIAA 2005 NVIDIA INDEX HPC Visualization Using NVIDIA Index S ee show floor! Tom-Michael Thamm, Christopher Lux, Marc Nienhaus

  9. VOLUME VISUALIZATION Dense Volumes 16 x 16 = 256 data values

  10. VOLUME VISUALIZATION Dense Volumes 16 x 16 = 256 data values

  11. VOLUME VISUALIZATION Dense Volumes

  12. VOLUME VISUALIZATION Dense Volumes 8 empty steps 5 sample steps

  13. VOLUME VISUALIZATION Sparse Volumes 52 data values (instead of 256!) 2 DDA skip steps 5 sample steps

  14. Value Atlas VOLUME VISUALIZATION Topology

  15. VOLUME VISUALIZATION Tilemap Topology Octree Topology Many levels Only two levels Every level divided in half Map has uniform divisions

  16. SPACEX Data Preparation with OpenVDB OpenVDB Grid Wavelet Cloud Sparse, Spatial Grid Hierarchy. Ideal for GPU Simulation. Efficent storage and visualization. Difficult to directly visualize.

  17. MOTION PICTURES: OPENVDB

  18. MOTION PICTURES: OPENVDB S parse Hierarchy of Grids Novel topology for very large volumes. Multi-core CPU Design VDB: High-Resolution Sparse Volumes with Dynamic Topology Ken Muset h , Dreamworks Animation. Transactions on Graphics, 2013

  19. OPENVDB VDB Topology Key features: Many levels Can store Each level is a grid very large volumes with only a few levels. Each level has its own resolution Efficient to traverse, since every level is a grid. e.g. top = 4x4 mid = 3x3 brick = 4x4

  20. OPENVDB VS AMR S parse Hierarchy of Voxel Grids OpenVDB: Adaptive Mesh Refinement (AMR): 1) Children are strictly contained 1) Children are strictly contained 2) Spatial extent is a voxel of parent 2) Spatial extent is free in parent 3) All siblings have same resolution 3) Siblings can have different res

  21. SUMMARY OF GOALS 1. Load and Visualize S parse Grids 2. Using the VDB Data S tructure 3. ..with.. Efficient Rendering on GPU 4. ..and.. High qualit y Rayt racing

  22. Multi-Display GRID SDK Virtual Reality GPUDirect for Video Video Codec S DK VR Works 1 VIRTUAL REALITY DISPLA Y LIVE VIDEO VXGI Iray SDK OptiX GVDB MDL SDK vMaterials 3 RENDERING MATERIALS NVIDIA DesignWorks™ - Advanced technologies for design applications https:/ / developer.nvidia.com/ designworks

  23. Multi-Display GRID SDK Virtual Reality GPUDirect for Video Video Codec S DK VR Works 1 VIRTUAL REALITY DISPLA Y LIVE VIDEO * NEW * VXGI Iray SDK OptiX GVDB MDL SDK vMaterials 3 RENDERING MATERIALS NVIDIA DesignWorks™ - Advanced technologies for design applications https:/ / developer.nvidia.com/ designworks

  24. GTC 2016 Talks: S 6643, S 6244 OptiX: A General Purpose Ray Tracing Engine, S teven Parker, NVIDIA ACM Transactions on Graphics, Vol. 29, No. 4, July 2010 S 6258, S 6844, S 6261, S 6320

  25. User Application User launch describe Application Mesh Hit Ray Shade Program Program OptiX Acceleration Scheduler API Structures JIT Compiler Launch! GPU

  26. NVIDIA GVDB SPARSE VOLUMES Data Visualization Efficient GPU Structures, GVDB Raytracing and Compute NVIDIA OptiX OpenVDB High Quality Raytracing CPU Format and Layout

  27. NVIDIA GVDB SPARSE VOLUMES What is GVDB? VDB Sparse Voxel Structures entirely on GPU Efficient Direct Raytracing Compatible with OpenVDB files and layout Integrated with NVIDIA OptiX Compute Operations on Sparse Volumes

  28. NVIDIA GVDB SPARSE VOLUMES VDB Topology Additional features: Run-time tree configuration Dynamic topology Efficient construction S ame t opology as OpenVDB

  29. Brick Voxels NVIDIA GVDB SPARSE VOLUMES Value Atlas GPU VDB Topology VDB Topology CPU

  30. RESULTS 624 x 304 x 304 (133 MB/ frame) A Numerical S tudy of High-Pressure Oxygen/ Methane Mixing and Combustion of a S hear Coaxial Inj ector, Nan Zong & Vigor Y ang, AIAA 2005

  31. A Numerical S tudy of High-Pressure Oxygen/ Methane Mixing and Combustion of a S hear Coaxial Inj ector, Nan Zong & Vigor Y ang, AIAA 2005

  32. A Numerical S tudy of High-Pressure Oxygen/ Methane Mixing and Combustion of a S hear Coaxial Inj ector, Nan Zong & Vigor Y ang, AIAA 2005

  33. RESULTS Scalability Testing Fluid simulations from S mooth Particle Hydrodynamics 1,000,000 Particle simulation on GPU using Fluids v.3, http:/ / fluids3.com Generated volumes at multiple resolutions by resampling velocity field Low-res: 512 x 180 x 256, 50 MB/ frame, 60 GB total Med-res: 1024 x 360 x 512, 338 MB/ frame, 408 GB total High-res: 2048 x 720 x 1024, 2.4 GB/ frame, 3 TB total

  34. RESULTS Source Data: 1 million Particles Simulated at 20 fps

  35. RESULTS Low Resolution High Resolution 512 x 180 x 256 2048 x 720 x 1024 # of Leaves: 616,444 # of Leaves: 12,131 File size: 50 MB/ frame File size: 2400 MB/ frame GPU Resampling: 16 seconds/ frame GPU Resampling: 804 seconds/ frame Load Atlas: 5420 msec/ frame Load Atlas: 450 msec/ frame Raytracing: 27 msec/ frame Raytracing: 64 msec/ frame 38 fps 15 fps

  36. RESULTS Low Resolution High Resolution 512 x 180 x 256 2048 x 720 x 1024 48x more data # of Leaves: 616,444 # of Leaves: 12,131 File size: 50 MB/ frame File size: 2400 MB/ frame GPU Resampling: 16 seconds/ frame GPU Resampling: 804 seconds/ frame 2x time Load Atlas: 5420 msec/ frame Load Atlas: 450 msec/ frame Raytracing: 27 msec/ frame Raytracing: 64 msec/ frame 38 fps 15 fps

  37. INTERACTIVE RENDERING 16 fps 1 ray / pixel 68.2 67.4 19 fps 58.1 54.3 26 fps Time 43.2 (ms) 38.0 35 fps 28.2 26.4 9.8 9.6 7.4 6.3 Empty Volume Isoval S kipping Render Render Low Res 512 Medium Res 1024 High Res 1536 Highest Res 2048

  38. NVIDIA GVDB SPARSE VOLUMES Data Management j ump OpenVDB direct down check check Sparse data .vdb bit bit sample resample OptiX Rendered .raw Dense data GVDB Images voxelized Polygonal .obj data .vbx .vdb .vbx Time Series .vdb .vbx .vdb Fast cache .vdb format

  39. NVIDIA GVDB SPARSE VOLUMES Data Example: Simulate, Process and Raytrace of 1200 frames Render Processing S imulat e check bit 4 seconds/ frame 1 min 6 hours Total: 1 hours 9 min 46 min S PH Resample Points Read VBX Raytrace Write VBX to 1024 3 Grid Simulation Rendering: On GPU On GPU Disk t ot al S parse resample / w Fluids v.3 530 GB 1200 frames in <1 hour wit h shadows and scat t ering 192 rays / pixel

  40. User Application User launch describe Application Mesh Hit Ray Shade Program Program OptiX Acceleration Scheduler API Structures JIT Compiler Launch! GPU

  41. User Application User * NEW * launch describe describe Application Mesh Hit GVDB S hade GVDB Hit Ray Shade Program Program Program Program GVDB Structure OptiX Acceleration Scheduler API Structures JIT Compiler Launch! GPU

  42. Poly-to-Voxel Scattering

  43. Poly-to-Poly Scattering

  44. NVIDIA GVDB SPARSE VOLUMES Interested in GVDB? Let us know! http:/ / developer.nvidia.com/ gvdb Rama Karl Hoetzlein NVIDIA

  45. THANKS! TO.. S paceX Stephen Jones Victor X. Guerrero Andre Kessler Ken Museth, Dreamworks Animation NVIDIA NVIDIA IndeX Tom Fogal Christopher Lux Tristan Lorach Marc Nienhaus Craig Fullman Tom-Michael Thamm Steven Parker This research used resources of the Oak Ridge Leadership Computing Facility, which is a DOE Office of S cience User Facility supported under Contract DE-AC05-00OR22725

  46. April 4-7, 2016 | Silicon Valley THANK YOU JOIN THE NVIDIA DEVELOPER PROGRAM AT developer.nvidia.com/ j oin developer.nvidia.com/ gvdb GVDB INTEREST SURVEY

Recommend


More recommend