[QUESTION] Can we remove the `isnan` check in the `complex<T>` multiplication operator? about libcudacxx HOT 6 CLOSED

Ran a quick experiment where I modified 11.5 libcu++. Modifications were to fix double cast and comment out isnan checks. This data indicates the isnan checks are still very expensive on performance. 65us vs 110us.

The benchmark is found here: https://github.com/NVIDIA/MatX/blob/main/bench/00_transform/conv.cu

A100 isnan checks, double fix

A100 NO isnan checks, double fix

from libcudacxx.

@griwes @wmaxey @gonzalobg @brycelelbach

Can any of you shine some light on this?

from libcudacxx.

CC @cliffburdick @luitjens

from libcudacxx.

This implementation appears to be inherited from the C99 Standard Annex G

Given that the C++ Standard does not implicitly inherit everything from the C Standard and https://eel.is/c++draft/complex.numbers#general-3 makes no mention of any requirement about anything like this, I am convinced that we can remove this from our implementation and remain conforming to the C++ Standard.

from libcudacxx.

I am wary of changing the current implementation because of non subtle runtime differences it would introduce. Those would only happen in a denormal case but still could potentially silently break user code.

Given that the chance of breaking something is rather slim and users might prefer the performance I would propose to add a feature macro that removes the sanitation of denormals.

Something in the direction LIBCUDACXX_ENABLE_SIMPLIFIED_COMPLEX_OPERATIONS?

Obviously I would like to tackle this in #338

from libcudacxx.

#338 because editing a comment does not link the issue / PR

from libcudacxx.