Auto-tuning Mixed-precision Computation by Specifying Multiple Regions
Xuanzhengbo Ren, Masatoshi Kawai, Tetsuya Hoshino, Takahiro Katagiri, Toru Nagai · 2023
Mixed-precision computation is a promising method for substantially increasing the speed of numerical computations. However, using mixed-precision data is a double-edged sword. Although it can improve the computational performance, the reduction in precision brings more uncertainties and errors. It is necessary to determine which variables can be represented with a lower-precision format without affecting the accuracy of the results. Hence, much effort is spent on selecting appropriate variables while considering the execution time and numerical accuracy. Auto-tuning (AT) is one of several technologies that can assist in eliminating this intensive work. In this study, we investigated an AT strategy for the “Blocks” directive in the auto-tuning language ppOpen-AT to tune multiple regions of a program and evaluated the effectiveness. A benchmark program of the nonhydrostatic icosahedral atmospheric model (NICAM), which is a global cloud resolving model, was considered as a study case. Experimental results indicated that when a single part of the program could perform well in the mixed-precision computation, a combination achieved a better performance. When used on the Flow Type I Subsystem (The Fujitsu PRIMEHPC FX1000), this method achieved almost 1.27× speedup compared with the NICAM benchmark program using all double-precision data.