implementing challenging seismic attributes on the gpu...

®

Porting Makes You Stronger Implementing Challenging Seismic Attributes

on the GPU

Jon Marbach,

Rich Welch,

Steve Dominguez,

Pate Motter

TerraSpark Geosciences

© 2013 TerraSpark Geosciences, LLC. All Rights Reserved.

• Talk overview

• Introductions

• Algorithms Ported

• Development Strategy

• Supporting GPU Acceleration in a Commercial Product

• Performance

• “Porting Makes You Stronger”

• Workflow Impact

• Conclusions

© 2013 TerraSpark Geosciences, LLC. All Rights Reserved. Slide 2

Implementing Seismic Attributes on the GPU

• Who am I?

• Software Architect / Engineer, not a

Geoscientist

• Couldn’t tell a Horst from a Graben

• Background in graphics (with OpenGL)

• I love GPUs!

• An OpenCL user

• Not a CUDA expert

• Under pressure to get results quickly

• I go for quick-wins

• Might not know every trick in the book



• Who is TerraSpark?

• Maker of the seismic interpretation

package Insight Earth™

• Workstation-class application

• Specialize in interpretation of

Faults, Salt, Stratigraphy, and

Shale Plays

• Time-saving tools for interpreters

• “True 3D” interpretation approach

• Compute-intensive attribute calculations

• Visualization-oriented



• Why GPU Acceleration?

• 3D “Image” Data

• Algorithms are image-

processing inspired

• Algorithms are data-parallel

• Algorithms tend to “gather”

many neighboring samples

• Can (sometimes) leverage

fast GPU interpolation



• 4 Key Interpretation Algorithms / Volumes



Horizon Orientation

(Dip/Strike)

Curvature

Fault Enhance Volume

Structurally-Oriented

Noise Filtering

• Horizon Orientation (Dip/Strike)

• Why is it important?

• Guides other “Structurally Oriented” attribute

calculations

• Indicates trends in the volume

• Why is it compute-intensive?

• Calculates gradient structure tensors

• Relies on an eigensystem solver

• Smooths tensors over a large area

• Why accelerate it?

• CPU runtimes on the order of tens of minutes

• Creates workflow bottlenecks



• Structurally Oriented Statistical Filtering


• Removes noise while preserving steeply

dipping features


• Well… it’s not, really but:

• Computes a unique sampling field for each

voxel

• Collects interpolated samples


• CPU runtimes on the order of minutes

• Very frequently used


GPU Acceleration for Interpretation

• Volumetric Curvature


• Can reveal subtle stratigraphic features

• Can reveal small-scale fracturing


• Performs fine-grained radial sampling

• Operates on vector data

• Interpolates samples


• CPU runtimes on the order of tens of minutes

• Increase turn-around time on parameter testing



• Fault Enhance (Automated Fault Extraction, aka “AFE”)


• Identifies and eases interpretation

of potentially 1000s of faults

• Huge savings of interpreter time


• Samples large neighborhoods

of voxels in 3D (a radial search)

• Interpolates samples in 3D


• CPU runtimes on the order of hours or days!

• Sensitive to parameterization, but iteration impractical without GPU



• Fault Enhance (Automated Fault Extraction, aka “AFE”)



Time-slice Section

Shown in inverted colormap for comparison with coherence

• Development/Robustness Strategy

• Start small, build infrastructure, apply knowledge/tools to tougher problems

• Target GPUs with specific feature set

• Fermi and Kepler

• Target wide range of hardware

• Handle lowest common denominator (1GB VRAM)

• Only accelerate “worst-offenders” first, expand out later

• Goal: “It just works (and produces accurate results)”



• Supporting a GPU Application commercially

• Most issues are hardware / driver / environment (expected)

• Strategy appears to be paying off



• Single GPU Speedup vs CPU (8 cores @2.4 GHz)



• Multi-GPU Speedup vs CPU (8 cores @2.4 GHz)



• Single Kepler GPU Speedup vs CPU (8 cores @2.4 GHz)



• Single Kepler GPU Speedup vs CPU (after Kepler tuning)



• Porting Makes You Stronger

• Just because it’s C, doesn’t mean it’s right

• Found problems with every single algorithm

• Porting forces a hard look at the existing CPU code and its behavior

• In one case, improved the CPU implementation speed by 3x

• One algorithm (AFE) was entirely re-written

• Improved accuracy of most algorithms for both CPU and GPU

• Stronger Team

• Stronger Product



Input Volume

Size

Insight Earth

1.7 AFE CPU

Runtime

Insight Earth

1.8 AFE CPU

Runtime

1.8 CPU

Speedup

over 1.7

1.8 AFE GPU

Runtime -

Turbo AFE

Speedup

over Insight

Earth 1.7

• Impact on overall Fault Extraction workflow time

• This includes automatic extraction of fault surfaces from volume



• Conclusions

• Seismic attribute calculations are a “no-brainer”

for GPU acceleration

• Performance gains are well worth the effort

• Porting to the GPU can improve CPU code too

• GPU Acceleration brings server-class

performance to seismic interpreters’ workstations

• Small investment in hardware == big

improvement in interpreter productivity (and

happiness!?)

• Available commercially from TerraSpark today



• Questions?



®

Porting Makes You Stronger Implementing Challenging Seismic

Attributes on the GPU

Jon Marbach

Rich Welch

Steve Dominguez

Pate Motter



Special Thanks to our research sponsors:

Apache, BHP, BP, Chevron, ConocoPhillips, Repsol, and Stone Energy

Hardware support provided by: Advanced HPC, HP, Nvidia

®

Jon Marbach

Rich Welch

Steve Dominguez

Pate Motter


Porting Makes You Stronger Implementing Challenging Seismic

Attributes on the GPU


Special Thanks to our research sponsors:

Apache, BHP, BP, Chevron, ConocoPhillips, Repsol, and Stone Energy

Hardware support provided by: Advanced HPC, HP, Nvidia

implementing challenging seismic attributes on the gpu...

Documents