Wide format floating-point math libraries

Math libraries are considered an extension of the hardware, providing common mathematical functions not supplied in hardware. Users expect these functions to be as accurate as the basic machine instructions, whether or not these expectations are always justified. Presented in this paper are the performance and accuracy evaluations of eleven transcendental functions found in 64 and 128 bit floating-point formats in math libraries on the CRAYY-MP, the IBM 3090E/VF, the Convex C-240, the Hewlett Packard 9000/720 and the IBM System/6000. Both architecture and algorithms are shown to impact the results.