Memory efficient 3D reverse time migration

Conclusions

The goal of this investigation was to try and reduce the RTM bottlenecks of 3D wavefield computation and memory transfers. The former of these was vastly reduced by using a series of GPU oriented kernels to perform wave propagation in a massively parallel sense, a total speed up of 30x was seen. The memory transfer data handling problem within RTM was diminished by using increasingly random boundaries on the velocity model to produce time-reversable wavefield modeling, alleviating the method from the intensely IO bound technique of checkpointing. A simple 2.5D single reflector model showed an encouraging RTM result, which demonstrates that random boundaries in 3D do indeed provide results comparable to checkpointing; work is underway to obtain detailed images of more complex velocity structures.