Use IEEE 754-2019 definition of min/max

[calebzulawski]

This doesn't match the existing scalar min/max, but it appears that it would optimize much better.

[programmerjake]

interestingly, x86's SIMD min/max are based on C's a < b ? a : b, where if either a or b is NaN, then b is returned. This includes returning a SNaN if b is a SNaN, and also includes returning b if a and b are zero (of any combination of signs).

https://www.felixcloutier.com/x86/maxps

[calebzulawski]

Are you suggesting that this is still incompatible with the 2019 standard?

[programmerjake]

Are you suggesting that this is still incompatible with the 2019 standard?

yes. x86 min/max is incompatible with IEEE 754-2008 or 2019.

[calebzulawski]

I see.

Do you have any idea what the vrageps mentioned here is? https://grouper.ieee.org/groups/msc/ANSI_IEEE-Std-754-2019/background/minNum_maxNum_Removal_Demotion_v3.pdf

[programmerjake]

Do you have any idea what the vrageps mentioned here is? https://grouper.ieee.org/groups/msc/ANSI_IEEE-Std-754-2019/background/minNum_maxNum_Removal_Demotion_v3.pdf

idk. I wasn't able to find them in the current version of intel's software development manual, so I'm guessing they're proposed instructions that weren't actually implemented, so intel later removed them.

[workingjubilee]

Is this about rust-lang/rust#91079 ?

[calebzulawski]

I didn't know about that issue, but yes it appears to be the same.

[programmerjake]

portable-simd's current min/max is incompatible with ieee minimum/maximum functions because it doesn't correctly handle signed zeros, it's compatible with minnum/maxnum afaict.

[calebzulawski]

That's how the scalar min/max is named, too.