Impact of fastmath?

[stuartarchibald]

Thanks for producing these benchmarks, it's great to see these comparisons made. I was wondering about the impact of fastmath, this seems to be enabled here for Fortran: https://github.com/paugier/nbabel/blob/e68441a717bd984f712938299b092d9f63431272/fortran/Makefile#L11 and here for C++: https://github.com/paugier/nbabel/blob/e68441a717bd984f712938299b092d9f63431272/cpp/Makefile#L17 and transonic/pythran here: https://github.com/paugier/nbabel/blob/e68441a717bd984f712938299b092d9f63431272/py/Makefile#L30

Does removing this flag from the above, or indeed adding it to Julia, Rust and Numba change the results?

[paugier]

Thanks for the interest and the remark. Indeed fastmath has an effect. I added it for Numba in https://github.com/paugier/nbabel/commit/bfe5e3a077acf91a6e42f8b6dffa8c4d0ea0598f

[jfalcou]

Beware that fastmath doesn't guarantee IEEE compliance and may impact the numerical quality of the results. I would advise not to use it.

[paugier]

fastmath was used when possible for the different implementations. So it's kind of fair. We could also choose to disable it for all implementations.