cl_intel_bfloat16_conversions

[bashbaug]

This PR adds the cl_intel_bfloat16_conversions specification that adds built-in functions to convert between 32-bit floating point values and bfloat16 values.

The related SPIR-V extension draft for bfloat16 conversions can be found here:

https://github.com/intel/llvm/blob/sycl/sycl/doc/design/spirv-extensions/SPV_INTEL_bfloat16_conversion.asciidoc

I will keep this as a draft PR and extension version 0.9.0 initially to allow for last minute changes before publication, but I believe the writeup is complete and I would appreciate any feedback. Thanks!

[alycm]

LGTM, just a few typos.

[bashbaug]

Thanks, that's embarassing. I'm not sure if my spell checker was too aggressive or not aggressive enough! I've accepted all three of the spelling fixes.

[alycm]

Some comments that are not part of my review because they are technical comments on a shipping vendor extension, but leaving them here for posterity should this ever be considered more widely.

Please note that this extension currently does not introduce a bfloat16 type to OpenCL C and instead the built-in functions convert to or from a ushort 16-bit unsigned integer type with a bit pattern that represents a bfloat16 value.

Not adding a new type is interesting as there is precedence for a storage-only floating-point type in OpenCL (half when cl_khr_fp16 is not supported). The only advantages that I can think of doing so is reduced implementation work, and opening the possibility for using integer arithmetic operations on the type, i.e. bit-level manipulation of the bfloat16.

Because bfloat16 ends with a number this does lead to awkward function names like intel_convert_bfloat1616_as_ushort16, but the awkward-ness is preferable to the ambiguity without the vector size suffix.

If we decide to add a true bfloat16 type we should consider other names that do not end in a number (bfloat16_t?).

Yes, the paired numbers aren't great for reading, it requires both knowing the bfloat16 is a type, and that type**n** is the vector convention to be able to read something like bloat162.