2025 年 12 巻 1 号 p. 113-130
This study examines the accuracy of numerical computations, focusing specifically on built-in elemental and special functions. The IEEE Standard 754 guarantees the accuracy of basic arithmetic operations and square root calculations. However, it does not establish precision criteria for built-in elemental and special functions, such as trigonometric, exponential, logarithmic, and gamma functions. In this paper, we evaluate the accuracy of low-precision elemental and special functions supported by MATLAB, GNU Octave, two C language compilers, CUDA, and NumPy in Python, using an interval enclosure technique. We assessed the accuracy of these functions for all low-precision floating-point inputs represented in FP16, BF16, and FP32. Additionally, we conducted a partial investigation into the accuracy of functions involving FP64 in Java within the aforementioned environment.
