A highly portable heterogeneous implementation of a Poisson solver for flows with one periodic direction

The portability of codes has become a major advantage given the continuous development of new architectures for numerical applications, as well as the progressive incorporation of accelerators in modern supercomputers. Following this trend, we have adopted an algebraic approach in the implementation of a Poisson solver for incompressible flows with one periodic direction. This approach, which basically consists of adapting a reduced set of fundamental operations to any architecture (such as the sparse matrixvector product or the dot product of two vectors), allows us to efficiently port our applications to any heterogeneous supercomputers in an easy manner. More particularly, our three-dimensional solver takes advantage of the existing periodic dimension (by means of a Fourier decomposition) to later execute overlapped data transpositions among devices, which conveniently share the workload with their CPU hosts, and solve the resulting two-dimensional decoupled subsystems.The work has been financially supported by a competitive R+D project (ENE2017- 88697-R) by the Spanish Research Agency. A.A.B is supported by a predoctoral contract (DIN2018-010061) by the Ministry of Science, Innovation and Universities of Spain (MICINN) and X.A.F. by the by the Government of Catalonia (2019FI_B2-00076).Postprint (published version