Thiago Ize

Ideally this would not be explicit and instead the compiler would emit a narrowing conversion warning. This would make a float->half operate in the same way as double->half and give...

My proposal was only for doing an `explicit half (float f);`. The other direction `operator float () const;` stays implicit just like before. That means the code above should continue...

There's a related discussion on this in https://github.com/AcademySoftwareFoundation/OpenImageIO/issues/3890#issuecomment-1606538525. In that one lg said: >The solution in this case is simple: just bump the tolerance a teeny bit. The trick is...

All architectures I believe will give the same result provided the compiler generated the same math instructions, which it should if all the archs have the same math instructions enabled...

> I tested out replacing the `madd` with `_mm_fmadd_ss` and it does make the `unit_simd` test pass when compiling with `march=native` for me. However if we go this route then...

I don't know which code path your compiler is taking or exactly how you have it coded up. Can you place some `#pragma message ( "this version is called..." )`...

Also, the llvm comment says 2GB and yet the check is done with `UINT32_MAX` which is 4GB. Is the comment wrong or should the code be changed to `INT32_MAX`? I...

I have approval from Eric Bourque, VP at Autodesk, that all of Autodesk's originally submitted contributions to this project under the 3-Clause BSD License are hereby relicensed to the Apache...

FYI: Arm and x86_64 are going to do bit identical math if you're using full precision math. The only places it would differ are the approximate instructions: `rsqrt` and `rcp`....

> I also considered using the LSB for this, also safe because the allocations are always aligned, the characters will never start on an odd address, so if we "|1"...