Provide the most important <cmath> functions for __half/bfloat16

[fbusato]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

libcu++

Is your feature request related to a problem? Please describe.

<cmath> is widely used to interact with floating-point types, but libcu++ doesn't provide support for __half/bfloat16/fp8

Describe the solution you'd like

Provide at least the following functions:

• isfinite
• isinf
• isnan
• isnormal
• signbit

Describe alternatives you've considered

No response

Additional context

No response

[miscco]

We got them recently, they are here https://github.com/NVIDIA/cccl/blob/main/libcudacxx/include/cuda/std/__cmath/traits.h

[fbusato]

I don't see any documentation related to <cmath>

[miscco]

We should add that

[bernhardmgruber]

Where is the functionality for FP8 types? I don't find it.

[miscco]

I believe we wanted to split this up, because there is currently no support for FP8 in libcu++ whereas support for the 16 bit types is effectively done

[bernhardmgruber]

ok, I fixed the issue title then.

[fbusato]

FP8 in CUDA is more a placeholder for exploiting tensor cores. It is not a really general type as bfloat or half. I don't think we need to support common math operation over FP8.

[bernhardmgruber]

FP8 in CUDA is more a placeholder for exploiting tensor cores. It is not a really general type as bfloat or half. I don't think we need to support common math operation over FP8.

We have ugly hacks in our unit tests that specialize [cuda::]::std::numeric_limits for those types, so we already use that functionality today.

AFAIK, we don't need <cmath> for FP8.