Computational cost is a major factor that inhibits the practicalapplication of 3D depth migration.We have developed afast parallel scheme to speed up 3D wave-equation depth migrationon a parallel computing device, i.e., on graphics processingunits GPUs. The third-order optimized generalized-screen propagator is used to take advantage of the builtinsoftware implementation of the fast Fourier transform. Thepropagator is coded as a sequence of kernels that can becalled from the computer host for each frequency component.Moving the wavefield extrapolation for each depth levelto the GPUs allows handling a large 3D velocity model, butthis scheme can be speeded up to a limited degree over theCPU implementation because of the low-bandwidth datatransfer between host and device. We have created furtherspeedup in this extrapolation scheme byminimizing the lowbandwidthdata transfer, which is done by storing the 3D velocitymodel and imaged data in the device memory, and reducinghalf the memory demand by compressing the 3D velocitymodel and imaged data using integer arrays instead offloat arrays. By incorporating a 2D tapered function, timeshiftpropagator, and scaling of the inverse Fourier transforminto a compact kernel, the computation time is reduced greatly.Three-dimensional impulse responses and synthetic dataexamples have demonstrated that the GPU-based Fourier migrationtypically is 25 to 40 times faster than the CPU-basedimplementation. It enables us to image complex media using3D depth migration with little concern for computationalcost. The macrovelocity model can be built in amuch shorterturnaround time.
展开▼
