[SYCL][Doc] Add isnan to bfloat16 math doc

[jinge90]

Signed-off-by: jinge90 [email protected]

[jinge90]

Hi, @gmlueck and @JackAKirk
I added "isnan" for bfloat16 in https://github.com/intel/llvm/blob/sycl/sycl/include/sycl/ext/oneapi/experimental/bfloat16_math.hpp#L36
This PR adds it in bfloat16 math doc. However, I didn't add "isnan" for sycl::marray<sycl::ext::oneapi::bfloat16, N> and wanted to consult you what the prototype should look like?
For bfloat16, isnan prototype is:
bool isnan(sycl::ext::oneapi::bfloat16);
For sycl::marray<sycl::ext::oneapi::bfloat16, N>, I expect the prototype to look like:
sycl::marray<bool, N> isnan(sycl::marray<sycl::ext::oneapi::bfloat16, N>);
Each bool value in return value corresponds each bf16 value in input param.

In CUDA math, NV has "hisnan2": https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____BFLOAT162__COMPARISON.html#group__CUDA__MATH____BFLOAT162__COMPARISON_1gf613f74f4c0220e6d85b91b086e2788a
hisnan2 return turn value is a __nv_bfloat162 with the corresponding nv_bfloat16 results set to 1.0 for NaN, 0.0 otherwise.

So, what's your idea? Do we need to align with NV math? If so, the return type should be sycl::marray<sycl::ext::oneapi::bfloat16, N>.

Thanks very much.

[JackAKirk]

I expect the prototype to look like: sycl::marray<bool, N> isnan(sycl::marray<sycl::ext::oneapi::bfloat16, N>); Each bool value in return value corresponds each bf16 value in input param.

In CUDA math, NV has "hisnan2": https://docs.nvidia.com/cuda/cuda-math-api/group__CUDA__MATH____BFLOAT162__COMPARISON.html#group__CUDA__MATH____BFLOAT162__COMPARISON_1gf613f74f4c0220e6d85b91b086e2788a hisnan2 return turn value is a __nv_bfloat162 with the corresponding nv_bfloat16 results set to 1.0 for NaN, 0.0 otherwise.

So, what's your idea? Do we need to align with NV math? If so, the return type should be sycl::marray<sycl::ext::oneapi::bfloat16, N>.

Thanks very much.

Hmm well spotted. My guess is they have only done this because they don't have a corresponding __nv_bool2 to __nv_bfloat162. My instinct is to agree with you that in the SYCL case it should return sycl::marray<bool, N>.

[gmlueck]

My instinct is to agree with you that in the SYCL case it should return sycl::marray<bool, N>.

Yes, I agree too.

[JackAKirk]

LGTM

[jinge90]

Hi, @intel/dpcpp-specification-reviewers and @gmlueck
Kind ping~
Thanks very much.

[jinge90]

Hi, @gmlueck
Is there anything else in this PR which needs update?

Thanks very much.

[gmlueck]

Is there anything else in this PR which needs update?

No, LGTM!