High-Order Difference Forward Modeling And Reverse-Time Migration Based On CUDA

In recent years, with the development of the computer processing capability andthe emergence of some new parallel devices, in particular the improvement ofgeneral-purpose computing ability and programmability of GPU(graphics processor),which brings new choice and hope to accelerate the scientific computing. Comparedwith CPU, GPU has natural parallel structure, and lower power consumption, lowercost, simpler management. At present, computing platform of CUDA(ComputeUnified Device Architecture) which launched by NVIDIA company based on its GPUis the most impressive. Image and video processing, computational biology andchemistry, fluid mechanics simulation, CT image representation, and ray tracing havebeen realized computation speedup on GPU based on CUDA. As it is, this papertransplants the high-order difference forward modeling and reverse-time migration bythe CPU to the GPU to accelerate the high-order difference forward modeling and thereverse-time migration based on CUDA.First, this paper expounds the related theory of CUDA briefly. First order stressand velocity of elastic wave equation is deduced, and the first order stress andvelocity of elastic wave equation is solved by using the staggered grids and high orderdifference, discrete difference scheme of the wave field forward continuation andreverse continuation is deduced. The intrinsic problems of numerical simulation offinite difference method (stability and dispersion as well as the artificial boundaryreflections) are carried on thorough analysis. The frequency dispersion phenomenoncan be suppressed through the adoption of staggered grid technology, model trialshows that improving the spatial difference order appropriately can also suppress thedispersion; The PML (perfectly matched layer) absorbing boundary condition is veryeffect to absorb the boundary reflection wave field. The realization process isexpounded in detail, and the corresponding discrete difference scheme of the firstOrder stress and velocity of elastic wave equation is given. The absorbing boundary c-ondition is also very suitable for parallel computing on GPU.Aiming at solving the problem of the storage requirements in reverse-timemigration, this paper puts forward an effective internal storage strategy, which caneffectively solve the problem of storage in reverse-time migration and the insufficient memory of GPU. By this storage strategy, the source wave field can be reconstructedthrough reverse-time continuation by using the stored effective internal wave field;the mechanism of low-frequency noise in reverse-time migration is analyzed bymodel calculation, then the poynting vector in the field of electromagnetic and themethod of Laplace filter in image processing is introduced into the noise suppression,compared with other denoising methods, the two kinds of denoising methods are moresuitable for parallel computing on GPU. According to the above theories and methods,the implementation process of the algorithm of the high-order difference forwardmodeling and reverse-time migration based on CUDA is expounded.Finally, through the trial of different models by writing programming, thecharacteristics and propagation of elastic wave field are analyzed, and the noisesuppression effect of the two methods based on different models is compared, whichproves that the theory and the algorithm of high-order difference forward modelingand reverse-time migration based on CUDA in this paper are correct. In addition,Computation time based on different platforms is compared, which fully proves thatthe high-order difference forward modeling and the reverse-time migrationbased on CUDA can significantly improve the computing speed.