Mixed precision techniques in numerical simulations

Discrete Element Methods (DEM) simulate the interaction of large numbers of rigid, incompressible objects with each other. Mainstream DEM codes focus on analytical shapes (spheres, ellipsoids, ...) to streamline the identification of contacts between objects. This step nevertheless continues to dominate the simulation time. We study mixed precision, heterogeneous programming and data structure engineering techniques to tackle these performance hotspots plus to enable simulations to use nonanalytical, i.e. triangulated shapes.

We study how we can separate data storage and exchange formats from compute formats such that we reduce the memory that needs to be accessed to check collisions. Next we investigate the potential for heterogeneous computing to accelerate simulations by offloading compute intense operations to GPUs while trying to hide/avoid the data transfer overheads of moving data between host and device memory. Alongside the work on heterogeneous compute there will be investigations into time stepping and integration schemes that complement the communication capabilities of modern heterogeneous systems. All of these steps are realised with different precision formats and combine different precisions and precision algorithms for different substeps. Lastly, we will explore what a programming language extension looks like to enable programmers to more easily take advantage of any mixed precision benefits we develop, i.e. we try to generalise our ideas to make them beneficial for more application areas.