SHF: Medium: Hierarchical Tuning of Floating-Point Computations

点击翻译按钮获取中文摘要

The project implements methods to improve the resource-efficiency of numerical computations on a variety of computing machines, ranging from supercomputers to mobile devices, by adaptively reducing data-precision based on the needs of the application. Efficiently adapting data-precision permits these machines to run larger computations and also improves the overall performance (including energy consumption) made possible by a combination of reduced computational burden as well as reduced data movement. The intellectual merits of this project are to research and develop the key steps to understand the nature of applications and computing systems, and to suitably minimize computing demands through targeted data-precision adjustment. Broader impacts of the work include training graduate students and releasing tools that the community can employ in future hardware and software product, to help minimize overall energy consumption and improve performance.Unused precision in floating-point computations ends up wasting allocated space in caches, and also causes unnecessary data movement. The technical parts of the project are to identify as well as pursue opportunities for optimally allocating precision, and to efficiently implement such allocation methods in actual codes. In addition to developing new algorithms to tune precision assisted by automated learning methods, the project develops symbolic analysis methods to serve novel roles in floating-point instruction selection and optimization in the form of superoptimizers. These tools will be released to a community of researchers interested in working toward exascale computing, and deploying applications in safety-critical devices. This work represents a synergistic combination of the investigator's skills ranging through high performance computing, formal methods, and compiler technologies.