Turbulence Visualization at the Terascale on Desktop PCs Marc - PowerPoint PPT Presentation

Turbulence Visualization at the Terascale on Desktop PCs Marc Treib*, Kai Bürger*, Florian Reichl*, Charles Meneveau + , Alex Szalay + , and Rüdiger Westermann* * Technische Universität München, Munich, Germany + Johns Hopkins University, Baltimore, Maryland, USA computer graphics & visualization

Turbulence Visualization • Local flow features based on velocity gradient (Jacobian) • Vorticity ω • Strain rate tensor S • Rotation rate tensor Ω • Derived scalar metrics • λ 2 , Q Hunt , Δ Chong , Q Ω , Q S , … • Multi-scale velocity field • Hierarchical nature of vortices

Turbulence Visualization: Challenges DNS of Isotropic and MHD Turbulence http://turbulence.pha.jhu.edu • 1024 3 Cartesian grid Vector-valued: • × 3 floats per grid point 12 GB Time-dependent: • × 1000 time steps 12 TB Main challenge: Data handling / streaming!

Turbulence Visualization: Challenges Level-of-Detail not admissible! LOD 0 LOD 1

Compression to the Rescue! • Compress data to reduce memory footprint + bandwidths • Decompress as late as possible • Stream and cache only compressed data • Unpack → render → discard • CUDA implementation for maximum throughput

Compression Scheme • Our choice: DWT + Quantization + Run-Length + Huffman Coding • Typically gives excellent quality per bit • Efficient data-parallel CUDA implementation • Contact me (treib@tum.de) if you want the code! 4 1 1 0 0 4 1 0 0 0 1 DWT+Quant RLE Huffman 2 0 3

Run-Length Coding Decoder Encoder 0 0 4 1 0 0 0 1 2 0 3 Compact A[i] += 1 2 1 4 4 1 1 2 3 7 Incl. scan A[i] -= A[i-1]+1 2 3 7 4 1 1 2 0 3 Scatter 0 0 4 1 0 0 0 1

Huffman Coding - Sequential Encoder 4 1 1 3 2 20 1 2 1 2 3 4 … … 20

Huffman Coding - Parallel Encoder 4 1 1 3 2 20 1 2 1 2 Write codeword lengths 3 3 2 2 4 3 12 2 3 4 Excl. scan … … 0 3 5 7 11 14 26 28 31 20

Huffman Coding - Parallel Decoder Encoder 4 1 1 3 2 20 1 2 Write codeword lengths Decode (sequential per thread) 3 2 2 4 3 12 2 3 Excl. scan 4 1 1 3 2 20 1 2 0 3 5 7 11 14 26 28 31

Compression Quality Original: 96 bpv 3.5 bpv 1.3 bpv

Compression Throughput • Timings for a 256 3 float3 volume on a GTX 580: • Upload+Decompress: ~30 ms (> 6 GB/s of output!) • About 50:50 split between DWT and entropy coding • Compress+Download: ~100 ms • For comparison: Upload or download of uncompressed data: ~35 ms

Preprocessing Pipeline PREPROCESS (per time step) Disk Bricking Compression Storage 3 min 10 s 4 s Velocity Volume Bricks with Overlap Compressed Bricks Compact File 1024 3 x 3 floats 5 3 x 256 3 x 3 floats 5 3 x 4 MB 400 MB

Visualization Pipeline RUNTIME (per time step) CPU GPU … Disk Load Upstream Decomp. Render Display Compressed Bricks ≥ 2 s 4 s 3 s

Advanced Visualization: Fast Preview CPU GPU … Upstream Decompress Compressed Bricks 3 s 2 s Compute Feature … Downstream Compress Compressed Bricks 3 s (Scalar) … Upstream Display Decompress Render 1 s 0.2 s

Interactive Exploration

Multi-Scale Analysis: Filtering • Coarse scale: box-filtered velocity • Large filter radii commonplace (50+) • Need neighboring bricks! 8 4 7 3 6 • Tensor product approach 2 5 • 3 passes (x/y/z) 1 • Per brick: decompress → filter → compress compress

Multi-Scale Analysis: Results Fine-scale iso-surfaces colored by vorticity alignment wrt. coarse scale

Conclusion • System for the visualization of very large turbulence data • Semi-interactive exploration possible on a desktop PC • Visit us at the exhibition! Booth 3958, Hall 4 • Also check out our papers for more details: M. Treib, K. Bürger, F. Reichl, C. Meneveau, A. Szalay, and R. Westermann Turbulence Visualization at the Terascale on Desktop PCs IEEE Transactions on Visualization and Computer Graphics 18(12) (2012) M. Treib, F. Reichl, S. Auer, R. Westermann Interactive Editing of GigaSample Terrain Fields Computer Graphics Forum 31(2) (2012)

Thanks for your attention! Iso-surfaces in fine data (red) only within iso-surfaces in coarse data (gray) Questions?