Using heterogeneous computing for scattering prediction in scenarios with several source configurations

In this work, we present a tool for solving large scattering problems with several acoustic source configurations. These problems entail a large matrix multiplication where the matrices must be generated on demand so that problems can be solved using systems with less memory than that required to store the whole matrices. We have analysed and developed different versions: one based on multiple matrix-vector products, two different approaches built on tiled matrix multiplication, and one heterogeneous implementation for using a GPU and a Xeon Phi simultaneously. To test these implementations, we have used different devices: multicore CPUs, a Xeon Phi accelerator, and a Tesla GPU. When compared to our initial work, the peak speedup of the new solutions is $$25\times $$25× for CPU, $$17\times $$17× for Phi, $$20\times $$20× for GPU, and $$20\times $$20× for the heterogeneous GPU + Phi implementation. Finally, it is worth mentioning that the tool presented in this work can be adapted and applied to other fields whenever the problem to solve requires a large matrix multiplication where the elements must be generated on demand (e.g. the inverse scattering problem in electromagnetics).