3 ms·

I think TFA is too quick to dismiss the fadd_fast intrinsic as "incredibly dangerous". In many cases you do know that your numbers are not infinities nor NaNs.

• SuchAnonMuchWow 2y ago
  It doesn't really makes sense for kahan summation, as the compiler would just make it similar to a naive summation because the errors terms would be zero under the assumptions of fadd_fast. 2sum would also break. This is exactly what you are loosing when using fadd_fast: fine control over the errors terms of floating point operations that do matter in a lot of cases. An other thing you are loosing is reproducibility: depending on the machine, compiler version, etc, your program will compute differently and for example may switch from linear to quadratic errors terms when you recompile. It could be the difference between a numerical algorithm converging or not, this kind of things.