Acceleration on many-cores CPUs and GPUs

1
Acceleration on many-cores CPUs and GPUs

• Dinesh ManochaLauri Savioja

2
Leveraging multi-core acceleration

• Available on both CPUs and GPUs
• Can be used for signal processing, numerical
  calculations and geometry processing

3
Frustum Tracing Pipeline
Frustum Triangle Intersection
4
Frustum tracing using multiple cores

• Multiple frusta and rays can be easily traced in
  parallel
• Frusta tracing specular reflections and edge
  diffractions
• Ray tracing diffuse reflections
• Can scale linearly with the number of cores
  (tested up to 16 cores)

5
Frustum Tracing Pipeline
Frustum Triangle Intersection
6
Frustum Tracing Results (7 cores)
Theater 54 ?s Factory 174 ?s Game 14K ?s Sibenik 71K ?s City 72K ?s SodaHall 1.5M ?s
diffraction NO NO NO NO YES YES
frusta 56K 40K 206K 198K 80K 108K
time (msec) 33 27 273 598 206 373
7
Frustum Tracing Results (7 cores)
Interactive geometric propagation on complex
scenesChandak et al. 2008
8
Scaling of FastV (Scaling with cores)
Fastest, accurate geometric propagation
algorithmChandak et al. 2009
9
Numerical Acoustics with Adaptive Rectangular
Decomposition on the GPU
Nikunj Raghuvanshi, Brandon Lloyd, Naga K.
Govindaraju, Ming C. Lin Department of
Computer Science, UNC Chapel Hill Microsoft
Corporation
10
Rectangular Decomposition

• Numerical Acoustics can be solved very
  efficiently on a rectangular domain
• Decompose complex domains into rectangles

11
Leveraging GPU for acoustics

• Solution of Wave Equation within each rectangle
  can be done using a Discrete Cosine Transform
  (DCT)
• DCTs can be done using FFT
• Use an efficient FFT implementation on the GPU
• Govindaraju, N. K., Lloyd, B., Dotsenko, Y.,
  Smith, B., and Manferdelli, J. 2008. High
  performance discrete Fourier transforms on
  graphics processors. In Proceedings of the 2008
  ACM/IEEE Conference on Supercomputing

12
FFT on the GPU
13
Performance
Scene Name Volume (m3) Time FDTD (CPU) Time Our Technique (GPU) Speedup

Corridor 375 365 min 4 min 90x
House 1,275 2718 min 13 min 200x
Cathedral 13,650 1 week (projected) 30 min 300 x
14
Conclusion

• Rectangular decomposition leverages GPU FFT
  combined with algorithmic improvements leading to
  100x improvement in performance for numerical
  acoustics

15
GPU-based occlusion and scattering

• Use techniques similar to (reflective) shadow
  mapping
• Compute qualitative occlusion or more physically
  grounded surface integral

16
GPU audio processing

• GPUs can be used for audio processing and
  filtering

17
Case Real-time acoustic radiance transfer
18
(No Transcript)
19
Case continued

• More information inS. Siltanen, T. Lokki, and
  L. Savioja, Frequency domain acoustic radiance
  transfer for real-time auralization,' Acta
  Acustica united with Acustica, vol. 95, no. 1,
  pp. 106-117, 2009.

20
Conclusions

• Multi-core CPUs and many-core GPUs can be used to
  accelerate sound rendering
• It is possible to develop interactive sound
  rendering systems by exploiting the commodity
  parallel hardware