 |
| Titel |
Forward and adjoint spectral-element simulations of seismic wave propagation using hardware accelerators |
| VerfasserIn |
Daniel Peter, Brice Videau, Kevin Pouget, Dimitri Komatitsch |
| Konferenz |
EGU General Assembly 2015
|
| Medientyp |
Artikel
|
| Sprache |
Englisch
|
| Digitales Dokument |
PDF |
| Erschienen |
In: GRA - Volume 17 (2015) |
| Datensatznummer |
250113661
|
| Publikation (Nr.) |
 EGU/EGU2015-13875.pdf |
|
|
|
|
|
| Zusammenfassung |
| Improving the resolution of tomographic images is crucial to answer important questions on
the nature of Earth’s subsurface structure and internal processes. Seismic tomography is the
most prominent approach where seismic signals from ground-motion records are used to infer
physical properties of internal structures such as compressional- and shear-wave speeds,
anisotropy and attenuation. Recent advances in regional- and global-scale seismic inversions
move towards full-waveform inversions which require accurate simulations of seismic wave
propagation in complex 3D media, providing access to the full 3D seismic wavefields.
However, these numerical simulations are computationally very expensive and need
high-performance computing (HPC) facilities for further improving the current state of
knowledge.
During recent years, many-core architectures such as graphics processing units (GPUs)
have been added to available large HPC systems. Such GPU-accelerated computing together
with advances in multi-core central processing units (CPUs) can greatly accelerate scientific
applications. There are mainly two possible choices of language support for GPU cards, the
CUDA programming environment and OpenCL language standard. CUDA software
development targets NVIDIA graphic cards while OpenCL was adopted mainly by AMD
graphic cards. In order to employ such hardware accelerators for seismic wave propagation
simulations, we incorporated a code generation tool BOAST into an existing spectral-element
code package SPECFEM3D_GLOBE. This allows us to use meta-programming of
computational kernels and generate optimized source code for both CUDA and OpenCL
languages, running simulations on either CUDA or OpenCL hardware accelerators. We show
here applications of forward and adjoint seismic wave propagation on CUDA/OpenCL GPUs,
validating results and comparing performances for different simulations and hardware
usages. |
|
|
|
|
|
|