high performance computing on graphics processing units: hgpu.org

The expanding hardware diversity in high performance computing adds enormous complexity to scientific software development. Developers who aim to write maintainable software have two options: 1) To use a so-called data locality abstraction that handles portability internally, thereby, performance-productivity becomes a trade off. Such abstractions usually come in the form of libraries, domain-specific languages, and run-time systems. 2) To use generic programming where performance, productivity and portability are subject to software design. In the direction of the second, this work describes a design approach that allows the integration of low-level and verbose programming tools into high-level generic algorithms based on template meta-programming in C++. This enables the development of performance-portable applications targeting host-device computer architectures, such as CPUs and GPUs. With a suitable design in place, the extensibility of generic algorithms to new hardware becomes a well defined procedure that can be developed in isolation from other parts of the code. That allows scientific software to be maintainable and efficient in a period of diversifying hardware in HPC. As proof of concept, a finite-difference modelling algorithm for the acoustic wave equation is developed and benchmarked using roofline model analysis on Intel Xeon Gold 6248 CPU, Nvidia Tesla V100 GPU, and AMD MI100 GPU.