Reverse time migration in three dimensions has two key bottlenecks - wavefield computation and IO limitations due to mass data transfer. Wave propagation and correlation in three dimensions imposes large computational constraints, both in terms of the number of floating point operations and the size of objects that need to be allocated. Furthermore, wavefields must be constantly written and read from disk since memory cannot possibly hold them all, this causes conventional RTM to become IO bound quickly. To address the computation requirements GPU propagation kernels are used to greatly reduce the computation time for source side modeling, achieving speeds in excess of 2.5 gigapoints calculated per second. Additionally, data handling requirements are vastly reduced by imposing pseudo-random boundaries on the velocity field, allowing time reversable source propagation. Achieving time reversable source propagation alleviates the requirement of checkpointing or boundary reinjection; this minimal data transfer results in GPU on-device operations becoming further accelerated.
![[pdf]](icons/pdf.png){kind=icon}