Improved GPU-based Full Waveform Inversion Algorithm with a Lossy Compression and Multiple CUDA Streams

Full Waveform Inversion (FWI) is a wave-based solution that can establish an accurate velocity model. Such a model is crucial for generating high-fidelity seismic images that aid in the discovery of reservoirs. However, FWI has the disadvantage that it requires considerable computing costs. These computational costs can be dramatically high in case of large-scale models with fine grid sizes. To resolve this issue, we optimize an original FWI workflow more adequate to a graphic processing unit (GPU) system. First, most of FWI functions are written with the compute unified device architecture (CUDA) program language. Second, we employ a GPU-based lossy compression library to reduce the memory requirements for storing source wavefields. It can help us to save wavefields into computer memory instead of hard disk. Then, we modified an equation for the zero-lag cross correlation for the gradient vector calculation in order to utilize multiple CUDA streams. To verify our optimized FWI workflow, we compare the performance speed between the original and our proposed FWI algorithms in the various aspects.