Rounding and Type Change methods

[Lokathor]

Before we can begin implementing a few of the more advanced functions, we'll need a few more of the basic utility functions for manipulating values.

For floats:

• round to nearest whole number (output is the same float type)
• ceiling / round to infinity
• floor / round to negative infinity
• truncate / round to zero
• fraction (remove the whole-number part)
• round to integer (output is the same-bit signed integer type)

For signed ints:

• round to float (output is the same-bit float type)

[calebzulawski]

Any idea what the type converting functions should be called? There doesn't seem to be an equivalent for f32 or f64...

[Lokathor]

in wide i call them round and then round_int.

basically https://doc.rust-lang.org/core/arch/x86_64/fn._mm_cvtps_epi32.html

[calebzulawski]

Probably relevant: https://doc.rust-lang.org/std/intrinsics/fn.float_to_int_unchecked.html

There actually is an equivalent in std but I missed it. This intrinsic should actually work on vectors as well.

[Lokathor]

I think that cvtps_epi32 returns a value for any value, even if the value is not the same number. I think that nans produce i32::MIN or something like that.

I'd hate to add more ways to get undef, but i suppose it could be done if that's what the people clamor for.

[calebzulawski]

Well, on a failure it raises a fp exception still (NEON's vcvt does this as well). I think the function should be unsafe because we the output is undefined on those lanes (but if we specify it's only undefined on those lanes, I think it's easy enough to use, even if it's unsafe?)