Adaption and GPU based parallelization of the code TEMDDD for the 3D modelling of CSEM data

We have analyzed and are currently enhancing the finite difference time domain code TEMDDD [1] for the forward modelling of marine CSEM (controlled source EM) data. Changes in the code were implemented to include EM sources into a conductive medium required to account for marine CSEM geometries. Currently, its performancy is increased using the massively parallel programming language CUDA (Compute Unified Device Architecture) to handle the most expensive numerical operations. These are the Spectral Lanczos Decomposition Method (SLDM, s.[2]), Householder tridiagonalization and the estimation of eigenvalues and eigenvectors. The Householder tridiagonalization as well as parts of the eigensolver have been parallelized successfully. As a result the performance has increased by a factor of 10 to 13 for the considered matrix sizes. Parallelization of SLDM and Inverse Iteration have been realized and are currently implemented into the code. The code will be applied to a 3D data set, which was acquired during a research cruise in the WND-project (West Nile Delta).