GPU/CPU Collabration Computing On Full Waveform Inversion And Its Application

With the development of computer science and ripening of related theory, full waveform inversion has become a hot subject in geophysics. In the last decade, graphic processing unit developed rapidly and general purpose GPU programming concept matured. GPU was applied broadly in geophysics and its algorithms are often mixed with matrixes and loops, which can be dealt more efficiently by parallelism. CPU needs to work in coordination with GPU to meet challenges on tasks with increasing complexity. Full waveform inversion contains many aspects and its framework is a mathematical optimization problem of many details. In order to introduce parallel computation, the parallel and serial part of the program must be treated differently. The host is responsible for overall framework and serial sections while GPU deals with the parallel parts, in this way GPU/CPU coordination is achieved. This thesis is concentrated on the following points.(1)The concept of general purpose GPU.programming is checked in this paper. By introducing CUDA, methods of program rewriting, storage usages, memory read and write bottlenecks and optimizations are covered. The form of memory organization and distribution of operation units are studied.(2)The thesis studies wave equation of acoustic approximation. Based on Torantola theory about full waveform inversion, this thesis conducts the conjugate gradient method. It also studies the idea of using wave field residual to reconstruct model parameters and improves methods about convergence and numerical matching relating to inversion algorithm. Iversion program is tested on different models and the result is analysised.(3)The forward modeling and inverse-time propagation are focused on. The wave equation and earth model are discretized. By using main idea of parallel program reconstruction and combining the characteristics of the two functions, the two kernels are constructed and then built into the inversion program framework. Some adjustment is made to the heterogeneous program to ensure correctness.(4)On windows7 platform, tests of muti-layer model are conducted. Using Nvidia Quadro NVS3000 M graphic card, parallel programs are tested. And relevant information like SM occupancy rate, bottleneck of parallel program are checked andsome opitimizations are done.