[clang] `__bf16` Type Support Macros

[bassiounix]

I can't find any macro that can indicate the availability of the __bf16 type, all I can find is hardware support macros but no generic type availability macros or anything set in clang or <float.h>.

This is needed for Bfloat16 GSoC project in libc.

If the solution is to gather the hardware support macros available in clang and define a macro for the general type support in <float.h>, I'd like to work on this before April 8th (the start of GSoC).

[jhuber6]

@AaronBallman I'm somewhat curious how we should handle bfloat16 in clang. The type is in the C++ standard, but not in C yet. LLVM lists it as a normal floating point type, so it should be usable, legalization issues notwithstanding https://llvm.org/docs/LangRef.html#floating-point-types.

[bassiounix]

My research indicates that __bf16 is the internal implementation of the type in clang and gcc, which is available in C and C++ as a keyword. It's only a standard in C++23 as std::bfloat16_t and not yet defined inside <stdfloat> in clang. And __bf16 type is supported on all targets we have in clang with no macros to indicate general type support.

[AaronBallman]

I'm a bit confused -- why does libc want to use bfloat16? That's the non-IEEE 754 type (half being the IEEE type which I think maps to _Float16 under a __STDC_IEC_60559_BFP__ implementation).

Regardless, __bf16 is how you get a bfloat16 in Clang, at least for targets which support it.

And __bf16 type is supported on all targets we have in clang with no macros to indicate general type support.

According to our documentation, it's not supported on all targets and never has native support: https://clang.llvm.org/docs/LanguageExtensions.html#half-precision-floating-point and we have TargetInfo::hasBFloat16Type() which returns false by default.

[bassiounix]

why does libc want to use bfloat16?

Given that it's used in ML apps, is a C++23 standard type, has native support on Nvidia and AMD GPUs, and libc provides its independent implementation to CPUs and GPUs for math operations, an implementation for math functions which support bfloat can be justified.

According to our documentation, it's not supported on all targets and never has native support

Yes, and I've been told that it maps mostly to float. However, what I'm looking for is not native support, rather a general type availability macro to indicate whether I can use __bf16 type or not. Something similar to __FLT16_MANT_DIG__, maybe __HAS_BF16__ or something similar.

[AaronBallman]

In InitPreprocessor.cpp we currently have:

  if (TI.hasFloat16Type())
    DefineFloatMacros(Builder, "FLT16", &TI.getHalfFormat(), "F16");

but nothing analogous for hasBFloat16Type(). It would be reasonable to add such a set of prefined macros IMO.

[jhuber6]

In InitPreprocessor.cpp we currently have:

  if (TI.hasFloat16Type())
    DefineFloatMacros(Builder, "FLT16", &TI.getHalfFormat(), "F16");

but nothing analogous for hasBFloat16Type(). It would be reasonable to add such a set of prefined macros IMO.

Then we need to add the appropriate handling to float.h presumably. f16 is guarded by __STDC_WANT_IEC_60559_TYPES_EXT__, I'm unsure what bf16 would use.

[AaronBallman]

Then we need to add the appropriate handling to float.h presumably. f16 is guarded by __STDC_WANT_IEC_60559_TYPES_EXT__, I'm unsure what bf16 would use.

It seems that GCC predefines them with the __BFLT16_ prefix and no non-underscored variants in float.h: https://godbolt.org/z/8GhrjGKdv so we should probably do the same.