Hybrid CPU-GPU computation of adjoint derivatives in time domain

摘要: For the solution of large-scale inverse scattering problems - in either acoustic or electromagnetic domain gradient based optimization approaches are a method choice, especially when derivatives regarding parameter interest can be obtained from adjoint fields [1], [2]. Gradients effectively computed using an approach where direct and integrated opposite temporal direction. This yielding high memory consumption, reduced computation gradients checkpointing recomputation states checkpoint is choice. We propose use graphics processing units (GPU) to accelerate by solving problem on GPU CPU computer. The implementation pipelining CUDA streams pinned masks transfer between host allows for at only little more than twice time problem.