Acceleration of large-scale CGH generation using multi-GPU cluster

Computer generated hologram (CGH) is a promising technology for realizing 3D displays. Large-scale CGH has an advantage that it resolves problems of existing 3D displays. However, the large-scale CGH generation requires a lot of memory space and computation time in proportion to pixel number. Further, in order to use CGH as a display, it needs to be generated in real time, and this is the reason why CGH does not suit to practical use. Computation of CGH is comprised of data-independent operations and current GPU has thousands of processing cores. Thus, acceleration of CGH generation can be expected by using GPU. To accelerate CGH generation processing, we adapt several parallelization and optimization techniques to the CGH program both for single node and multiple ones. The single node optimization techniques include the way of object decomposition, the reduction of data transfer amount between CPU and GPU, the kernel integration, stream processing, and the utilization of multi-GPU parallelism. The multi-node optimization includes inter-node data distribution method. The results show that we have achieved 134.7 times speed-up compared to sequential program execution by CPU.