A Modular Approach to Performance, Portability and Productivity for 3D Wave Models

The HPC hardware landscape is growing increasingly complex in order to meet demands in scientific computing for greater performance. In recent years there has been an explosion of parallel devices coming on to the scene: GPUs, Xeon Phis and FPGAs to name but a few examples. As of writing, even the current leading supercomputer, Sunway TianhuLight, uses its own bespoke on-chip accelerators[12]. Available programming models, however, lag behind and are not currently able to provide the necessary tools for running scientific codes across platforms in ways that are performant, portable and productive.
This environment creates a plethora of challenges for computational scientists of which we focus on two: first the need for a high level of productivity for codes that still get good performance and second consistently getting good performance across platforms - the “performance portability” problem. Existing solutions tend to be either not productive but provide good performance or focus on high-level abstractions requiring heuristics to get good performance (often which are tied to particular platforms). While some current approaches raise the productivity level, they are often trying to solve the same problems over and over or trying to solve too many issues for a niche domain. In addition, many of these approaches have only been tested on simplistic benchmarks, which can lose critical functionality of real-world simulation codes. We instead propose a modular approach using existing frameworks to target these issues separately: a high-level DSL to target the productivity problem compiling into an IR language which addresses the performance portability problem.
Our previous research has shown that the development of more productive and performance portable codes for room acoustics simulations is possible. Preliminary results using the intermediary parallel language lift[16] confirm that this framework is capable of handling complex stencils. Further developing lift and targeting existing stencil-focused DSLs will create a simple, modularized approach which harnesses and expands existing functionality instead of trying to reinvent the wheel. This modular approach can then be used as an example to extend to other physical simulations using similar algorithms.