[C++] Enable hardware support for arrow::util::Float16 on GCC and Clang

[andishgar]

Describe the enhancement requested

arrow::util::Float16 is a software simulation of the IEEE 754 float16 (2008) standard. However, some architectures provide hardware support, such as ARM and x86_64 AVX512-FP16 microarchitectures. Furthermore, GCC and Clang provide support via _Float16. However, there is currently no support in MSVC.

Additionally, I found that Google Highway provides such support.

@kou, do you think it is worth investigating further and enabling this for Arrow?

Component(s)

C++

[Arbaaz123676]

Hi @andishgar, I would like to work on this issue.

[andishgar]

@Arbaaz123676, I'm just a regular contributor, and this issue is merely a proposal. It needs to be verified by an Arrow core developer.

[kou]

Can we benchmark the current arrow::util::Float16 and the hardware float16?

[kou]

BTW, it seems that C++23 will have std::float16_t: https://en.cppreference.com/w/cpp/types/floating-point.html

[andishgar]

BTW, it seems that C++23 will have std::float16_t: https://en.cppreference.com/w/cpp/types/floating-point.html

regarding this, As I mentioned before link support for float16 is completely optional, and MSVC does not currently implement it.

[kou]

Oh, sorry. I missed it.

[andishgar]

Can we benchmark the current arrow::util::Float16 and the hardware float16?

Several notes:

1. I do not have any hardware that supports the f16p instruction, so I cannot conduct benchmarks.
2. I checked the assembly code generated hereand here, and I found hardware instructions for FP16 operations.
3. I plan to enable arithmetic operations for +, -, *, and /.
4. I think that even if we conduct a benchmark and arrow::util::Float16 performs significantly better than GCC and Clang, it is still better to rely on compiler support and wait for further improvements from the compilers.

@kou, what is your opinion?