Optimization of two Jacobi Smoother Kernels by Domain-Specific Program Transformation

Our aim is to apply program transformations to stencil codes, in order to yield highest possible performance. We observe memory bandwidth as a major limitation in stencil code performance. We conducted a small study in which we applied optimizing transformations to two Jacobi smoother kernels: one 3D 1st-grade 7-point stencil and one 3D 3rdgrade 19-point stencil. To obtain highest performance, the optimizations have to be customized for the execution platform at hand. We illustrate this by experiments on two x86 architectures and one BlueGene/Q architecture. A compiler with specic knowledge about stencil codes and execution platforms should be able to apply our transformations automatically. We are working towards such a compiler in the DFG-funded project ExaStencils.