Autotuning Power Consumption and Computation Accuracy using ppOpen-AT
Shouhei Yamanashi, Hisashi Yashiro, Takahiro Katagiri, Toru Nagai, Satoshi Ohshima · 2022
Mixed-precision computation mainly focuses on shortening the execution time, at the expense of accuracy. To achieve speedups for numerical calculation using mixed-precision computation, it is necessary to tune software performance with respect to not only execution speed but also computation accuracy and power consumption. This increases the overall cost of tuning. Autotuning (AT) is one of the candidates among several technologies available for reducing the cost associated with tuning the software performance. In this study, we propose a method for AT to obtain speedups with respect to computation accuracy and power consumption. The proposed AT method uses an AT language that changes computation accuracy of the original code to mixed-precision by combining double and single precisions. Performance evaluation was carried out by using the Fujitsu PRIMEHPC FX1000, which is a “Fugaku” type supercomputer installed at the Information Technology Center, Nagoya University. The proposed method achieved a 1.5x reduction in execution time and energy consumption while retaining reasonable accuracy degradation from the original code of a global cloud resolving model.