Speaker
Dr
Guido Cossu
(Edinburgh University)
Description
We discuss the development of a new data parallel library targeted to any problem requiring solving partial differential equations on a regular grid. The main motivation in the development is the proliferation of the hierarchies of parallelism in modern architectures that requires constant adaptations of large parts of the codebases for each new architecture like the new Intel Knights Landing. Our code is mainly using high level C++, localising the architecture dependence in a small part of the code and guaranteeing performance by expression templates and a organization of memory structures that automatically follows the underlying vector units. We review the status and design choices with a particular stress on the lattice QCD performance of the code.
Primary author
Dr
Guido Cossu
(Edinburgh University)
