Performance Optimization in Accelerator Libraries using Parameter Tuning

This thesis contributes to the field of performance-portable computation by introducing an auto-tuningframework in alpaka 3, a C++ library that enables platform-independent heterogeneous accelerator pro-gramming. It attempts to enhance alpaka’s performance portability by maximizing resource utilizationacross hardware platforms and minimizing the time consumption of manual tuning for individual prob-lem descriptions. More specifically, the proposed auto-tuning mechanism allows dynamically adjustingperformance-critical parameters during production runs. The tuning-mechanism is platform-aware andtracks performance measurements and tuning characteristics with respect to the underlying architecture.alpaka 3 implements a special programming model that allows domain-specific problem descriptionsindependent of the number of physical workers. This model enables performance-critical parameters,such as grid size and thread-block size, to be subject to the optimization process without violating thecorrectness of the algorithm. While this provides a general mechanism for tuning alpaka 3 kernels, theproposed tuning framework goes further by allowing both custom runtime and compile-time parametersto be involved in the tuning process. This offers users a fine-grained approach to optimizing specificalgorithms.