Computational Science Initiative Event

The performance portability is increasingly difficult in the Exascale era because of growing architectural diversity and a proliferation of programming models for parallelism, both within a single supercomputers and across a range of computers. I describe the approach taken in the Grid software package for Lattice QCD, and hope that some of the techniques are more generally applicable, at very least to those problems with locally structured cartesian grids. Specifically, I describe a C++ approach giving good performance for both SIMD and SIMT compilation with a single source for high performance kernels. Bio: Prof Peter Boyle joined Brookhaven National Laboratory in 2019. Previously he has been with the University of Edinburgh from 2000 until 2019. His research is in the field of Lattice QCD, and he has worked with IBM and Columbia University on the QCDOC and BlueGene/Q supercomputer designs. He has recently worked with Intel on both codesign with QCD and with machine learning. He has authored seven patents in various aspects of computer and memory system design.