Structure-preserving Smoothing for Seismic Amplitude Data by Anisotropic Diffusion using GPGPU

Transcription

1 GPU Technology Conference 2016 April, 4-7 San Jose, CA, USA Structure-preserving Smoothing for Seismic Amplitude Data by Anisotropic Diffusion using GPGPU Joner Duarte

2 Outline Introduction Why is noise attenuation important? Objective Related Works Structure-preserving Smoothing Method Parallel Approach Results Conclusion 2

3 Introduction Why is noise attenuation important? Original seismic data crossline 905 of F3 block 1 Zoom area from the input data and the corresponding filtered image. 3 1 Volume of the Netherlands offshore F3 block downloaded from the Opendtect website

4 Introduction Why is noise attenuation important? Preserve relevant structural features is fundamental 4 Original image Gaussian filter with radius 1.0 Proposed filter

5 Introduction Why is noise attenuation important? Seismic attributes have been used to accelerate interpretations, and their results are directly related to the quality of the seismic data. 5 Input seismic data inline 640 of the Kerry data 1 1 Kerry data from New Zealand provided for use by New Zealand Petroleum and Minerals Zoom area of the input data and the corresponding semblance attribute Filtered data and the corresponding semblance attribute

6 Introduction Why is noise attenuation important? Curvature attributes are widely used to improve seismic interpretation process. The results of curvature attribute can be directly improved by our iterative noise attenuation process. Original Filtered 3x 6 Zoomed areas of time slice 396ms of F3 Block data Original Filtered 3x

7 Introduction Objective Remove noise preserving relevant details such as structural and stratigraphic discontinuities in interactive time. 7

8 Related Works [1990] Perona and Malik, Scale-Space and Edge Detection Using Anisotropic Diffusion [2002] Hocker and Fehmers, Fast structural interpretation with structure-oriented filtering [2009] Dave Hale, Structure-oriented smoothing and semblance [2011] Baddari et al., Seismic Noise Attenuation by Means of an Anisotropic Non-linear Diffusion Filter 8

9 Structure-preserving Smoothing Method Input Volume Do you wish iterate again? STEP 1: Computing Seismic Attributes STEP 2: Applying The Anisotropic Diffusion Filter Filtered Volume Overview Pampanelli P., Seismic amplitude smoothing by anisotropic diffusion preserving structural features, 2015 Use the instantaneous phase as horizon indicator attribute Two steps Compute the constraints Assemble and solve the system of equations Iterable 9

10 Structure-preserving Smoothing Method Input Volume STEP 1: Computing Seismic Attributes STEP 2: Applying The Anisotropic Diffusion Filter Compute Hilbert Transform Y t Compute Seismic Complex Trace Z t = X t + iy t Do you wish iterate again? Filtered Volume 10

11 Structure-preserving Smoothing Method Input Volume STEP 1: Computing Seismic Attributes STEP 2: Applying The Anisotropic Diffusion Filter Compute Horizon Indicator Attribute¹ Φ = δφ δx = 1 X 2 + Y 2 δφ δy = 1 X 2 + Y 2 δφ δx, δφ δy δy X δx δy X δy Y δx δx Y δx δy 11 Do you wish iterate again? Filtered Volume where Φ is the instantaneous phase attribute 1 Silva, P.M. et al., Horizon indicator attributes and applications. SEG Technical Program Expanded Abstracts 2012.

12 Structure-preserving Smoothing Method Input Volume STEP 1: Computing Seismic Attributes Compute Horizon Indicator Attribute¹ Φ = δφ δx, δφ δy STEP 2: Applying The Anisotropic Diffusion Filter instantaneous phase gradient Φ Φ Do you wish iterate again? Filtered Volume horizon 12

13 Structure-preserving Smoothing Method Input Volume STEP 1: Computing Seismic Attributes Compute Fault Attribute¹ F max = MAX X. Φ, Y. Φ Do you wish iterate again? STEP 2: Applying The Anisotropic Diffusion Filter Filtered Volume where X represents the normalized vector of the amplitude gradient and F max 1, 1 1 Pampaneli et al., A new fault-enhancement attribute based on first order directional derivatives of complex trace, EAGE,

14 Structure-preserving Smoothing Method Input Volume STEP 1: Computing Seismic Attributes STEP 2: Applying The Anisotropic Diffusion Filter Diffusion Tensor D = Φ Φ T Fault Preserving Factor ε = F max 2 Do you wish iterate again? Filtered Volume 14

15 Structure-preserving Smoothing Method Input Volume STEP 1: Computing Seismic Attributes The Anisotropic Diffusion Filter Solving the linear system Ax = b of size m x m, where m = w x h Stencil 3x3 STEP 2: Applying The Anisotropic Diffusion Filter δu δt = εd u Do you wish iterate again? Filtered Volume u n+1 x,y = u n x,y + t ε x,y n D n x,y n+1 u x,y 15

16 Parallel Approach Compute Seismic Attributes (STEP 1) CPU FFTW3 OpenMP 1 loop for each trace (Hilbert Transform) 1 loop with 1 inner loop Memory used in this step Input seismic section Hilbert Transform Y(t) Imaginary part of complex trace Horizon Atribute Instantaneous phase gradient GPU CUFFT 2 kernels (Hilbert Transform) 1 kernel to compute attributes 1 thread per sample Fault Attribute (ε) dxx dxy dyy 16 Diffusion tensor

17 Parallel Approach Assemble and Solve the System of Equations (STEP 2) System solver 50~70% of total time of one iteration Libraries for solving sparse linear systems Sparse matrix format: CSR (Compressed Sparse Row) Memory used in this step 31x input image size A x b A x = b 19x Input section size Input Input Input + 1x + 1x + 10x = section size section size section size 31x Input section size CSR Filtered image Input image NUMERICAL METHOD 17

18 Parallel Approach Numerical Methods Conjugate gradient (CG) Symmetric linear systems Symmetric and positive-definite matrix Ax = b => A T Ax = A T b Stop criteria Error tolerance: 10-4 Limit of 100 iterations No preconditioner used Biconjugate gradient stabilized (BiCGStab) Generalized minimal residual (GMRES) 18

19 Parallel Approach Libraries Intel MKL 11.3 CUSP ViennaCL There are other libraries that we may try in the future (MUMPS, cusolver, PARALUTION, etc) 19

20 Results HP Z820 Workstation 64 GB RAM Intel(R) Xeon(R) E cores (12 threads) Inline 240 of the volume of the Netherlands offshore F3 block¹ 951 x 462 Tesla k40 12 GB 2880 processors Linear system size: x Volume of the Netherlands offshore F3 block downloaded from the Opendtect website 20

21 Results Inline 240 of the volume of the Netherlands offshore F3 block Original slice Filtered Slice with 3 Iterations 21

22 Results Inline 240 of the volume of the Netherlands offshore F3 block Original slice 1 iteration 3 iterations 5 iterations 10 iterations 22

23 Results Number of iterations MKL (ms) CG CUSP (ms) ViennaCL (ms) x 16x 14x 12x 10x 8x 6x 4x 2x GPU Speedup using CG Iterations MKL CUSP ViennaCL 23

24 Results Number of iterations MKL (ms) BiCGSTAB CUSP (ms) ViennaCL (ms) x 16x 14x 12x 10x 8x 6x 4x 2x GPU Speedup using BiCGStab Iterations MKL CUSP ViennaCL 24

25 Results Number of iterations MKL (ms) GMRES CUSP (ms) ViennaCL (ms) x 16x 14x 12x 10x 8x 6x 4x 2x GPU Speedup using GMRES Iterations MKL CUSP ViennaCL 25

26 Time (ms) Time (ms) Results Filtering time with 1 iteration Filtering time with 10 iterations CG BiCGStab GMRES CG BiCGStab GMRES MKL ViennaCL CUSP MKL ViennaCL CUSP 26

27 Conclusions - Method that attenuates noise preserving structural features - Improves interpretation eficiency and also can enhance the results of other seismic attributes - Fast enough to be interactive 27

28 Conclusions - It can be used on demand during the interpretation process and also be integrated with tools like Cintiq tablet - CUSP library proved to be very fast for our problem with a high level interface that makes it very simple to use 28

29 Acknowledgements Questions??? 29