CUDA-based acceleration of full waveform inversion on GPU

Computational cost and storage requirement are the main obstacles that inhibit the research and practical application of full waveform inversion (FWI). We have developed a fast parallel scheme to speed up FWI on graphics processing unit (GPU),which is a parallel computing device, via CUDA(an acronym for Compute Unified Device Architecture), developed by NVIDA and used as the programming environment. In this parallel scheme, to avoid frequent and low-bandwidth data transfer between host memory and device memory, almost the entire computing task, including propagator and backpropagator, are coded as a sequence of kernel functions that can be called from the compute host for each iterative inversion. The random boundaries conditions are used when propagating source wavefield to solve the storage requirement so that we do not have to save any additional wavefield data and the noise introduced into final inversion image is so weak that can be ignored due to iterations. To test our algorithm, we implement the FWI on Personal Computer (PC) with GTX480 GPU to reconstruct the Marmousi velocity model using synthetic data generated by the finite-difference time domain code. This numerical test indicates that the GPU-based FWI typically is 80 times faster than the CPU-based implementation.