What traits abstract over what?

[workingjubilee]

We've discussed potentially implementing trait abstractions over SIMD vectors in the basic API because it's pretty unwieldy to handle every single possible TxW combo, and we've mentioned we will need to leave some traits in separate semi-official crates, e.g. a simd-complex, but we need to think about what abstractions we want to offer where.

Discussion partly here: https://rust-lang.zulipchat.com/#narrow/stream/257879-project-portable-simd/topic/Semi-official.20(non-std.3A.3Asimd).20crates

[workingjubilee]

of particular note, what seems like a strong consensus:

I'd be fine with just recording an official stance that "SIMD types should, by default, implement ops traits vertically", and then leaving more traits to be added to core::num later on. Perhaps a "we also propose the following traits but they can be accepted separately from the main SIMD work"
—@Lokathor

[workingjubilee]

Somewhat obvious if one reviews the code so far, but so far we've been inclined towards a strong absence of traits in the public API, with the occasional attempt to use them cleverly for bounds. Basically there isn't a lot of reason in std to trait over u32 and u8 or u16 and i16, so we've been mostly also not. This may change, we'll see!

[bjorn3]

Simd code would more likely to want to abstract over different lane sizes for optimal speed on different processors than scalar code would want to abstract over scalar size.

[calebzulawski]

To add on to Jubilee's justification, I think it's not so much that there is no desire to abstract over SIMD vectors, but there isn't much need to do so in core. I think this is similar to why num is a separate crate. I envision a crate similar to num (or even part of it) for SIMD.