Achieving mixed precision computing with the help of domain specific libraries

As scientific applications in HPC are increasingly dependent on floating point number-based operations, the demand is emerging for tools to obtain better performance by utilising the characteristics of different representations. A particularly suitable solution for this purpose is mixed precision computing, where some part of an application can be transformed to lower precision, thus covering less memory and producing faster operation execution, while having as small precision loss as possible. In this project, we propose to extend with automatic mixed precision transformations the OPS and OP2: two domain specific libraries for the execution of structured and unstructured grid applications. These libraries also can be viewed as instantiations of the Access-Execute descriptor programming model, thus allowing the examination of interactions between kernels. As a baseline test, we hand tuned a CFD mini application to explore the possible performance gain through mixed precision computing in OP2. By halving the size of the most accessed data set, we achieved a 1.13x (1.1x) speedup on CPUs (and GPUs), compared to the speedup of a fully reduced sized execution: 1.76x (1.44x). Future tasks of the project are 1. to determine the precision loss and whether it is acceptable; 2. to develop strategies to choose candidates for precision lowering; 3. to estimate the expected performance gain.

---

Contributors:

• Bálint Siklósi (Pázmány Péter Catholic University - Hungary)
• Gihan Mudalige (University of Warwick)
• István Reguly (Pázmány Péter Catholic University - Hungary)