Emulation of the FMA in rounded-to-nearest floating-point arithmetic
Stef Graillat, Jean‐Michel Muller · HAL (Le Centre pour la Communication Scientifique Directe) · 2024
We present algorithms that allow to emulate the fused multiply-add (FMA) instruction and the correctly-rounded sum of three numbers (ADD3) in binary floating-point arithmetic, using only rounding-to-nearest floatingpoint additions, multiplications, and comparisons.We then introduce variants of these algorithm that make it possible to compute the error of an ADD3 or FMA operation.