Automatic Tunable Floating-Point Format Selection in a RISC-V Ecosystem
摘要
In this work, we address variable precision computation that can be adjusted depending on the characteristics and the specification of the algorithm. We base our method on the Tunable Floating-Point (TFP) format, which allows to change format in floating-point units to tune the precision and the dynamic range to specific parts of the algorithm. We propose an automatic method to change the TFP format during the computation, based on a quality function, which is embedded in the execution of the arithmetic kernel.