I’m surprised not to see fused multiply-add here:
FMA(x,y,z) ~ x += y*z
You see it all over the place in numerical computing and deep learning. FMA(x,y,z) ~ x += y*z
You see it all over the place in numerical computing and deep learning.If you separate the operations like that, the intermediate result is rounded. The final result is then often different from the result of the FMA, which doesn't do the intermediate rounding.