LinearMap with explicit SIMD support?

[tower120]

I don't know how much this would fit into this crate, but LinearMap could significantly win from SIMD search.
I have implementation of similar container that store keys in wide::simd packs. It works on top of "the usual" Vec, but it would be
not so hard to port it to heapless::Vec.

Alternatively - it is possible to implement SIMD search on existent LinearMap for supported SIMD types... But as far as I
know - only desktop CPUs have zero-to-none penalty for loading from non-aligned memory to register. And looks like you
aim for embedded hardware... So you probably "have" to explicitly use SIMD types as key packs.

I understand that simd types is not stable in Rust, and wide crate supports only desktop hardware... But as a suggested solution for that: we could have simple custom SIMD types/traits which will initially points to wide implementation. Than they could be expanded by community on as-need basis...

[k5602]

I don't know how much this would fit into this crate, but LinearMap could significantly win from SIMD search. I have implementation of similar container that store keys in wide::simd packs. It works on top of "the usual" Vec, but it would be not so hard to port it to heapless::Vec.

Alternatively - it is possible to implement SIMD search on existent LinearMap for supported SIMD types... But as far as I know - only desktop CPUs have zero-to-none penalty for loading from non-aligned memory to register. And looks like you aim for embedded hardware... So you probably "have" to explicitly use SIMD types as key packs.

I understand that simd types is not stable in Rust, and wide crate supports only desktop hardware... But as a suggested solution for that: we could have simple custom SIMD types/traits which will initially points to wide implementation. Than they could be expanded by community on as-need basis...

i agree i will start a research and design for this,
i've felt it in my usecase too so i will begin the work.

[tower120]

@k5602 What you gonna use for SIMD operations abstraction?

[k5602]

@k5602 What you gonna use for SIMD operations abstraction?

i'm still in researching and trial and error phase but for now what i got
i design SimdKey with lanes that should match target's natural width and
implemented by types that support batch equality comparison.
and for the backing will be between

1. core::simd is the right thing as longterm option
2. raw core::arch is the answer for now where SIMD, and the scalar impl as the always-available floor, Verbose but can get handled
3. wide is desktop-only

what do you think?

[tower120]

I don't know - I have working implementation with abstraction over wide and abstraction over Contiguous container (so it can be SmallVec, Vec, ArrayVec, etc). Which is actually rather a SortedLinearMap...
I need it to work only for desktop and mobile...
I just thought it could be useful in heapless...

[k5602]

I don't know - I have working implementation with abstraction over wide and abstraction over Contiguous container (so it can be SmallVec, Vec, ArrayVec, etc). Which is actually rather a SortedLinearMap... I need it to work only for desktop and mobile... I just thought it could be useful in heapless...

good thought but ig the whole idea of this that it can be used in embedded devices, i see core::arch is the solution for now so it can be mobile across all platforms.