This crate proposes an experimental way to use SIMD intrinsics reasonably safely, using the new target_feature 1.1 feature in Rust, recently stabilized.
A much earlier version of this crate experimented with an approach that tried to accomplish safety in safe Rust as of 2018, using types that witnessed the SIMD capability of the CPU. There is a blog post, Towards fearless SIMD, that wrote up the experiment. That approach couldn't quite be made to work, but was an interesting exploration at the time. A practical development along roughly similar lines is the pulp crate.
Some code has been cut and pasted from the half crate, which is released under identical license. That crate is also an optional dependency, for more full f16 support. This dependency will be removed when f16 stabilizes.
For more discussion about this crate, see Towards fearless SIMD, 7 years later. A planned future direction is to autogenerate the the SIMD types and methods, rather than having to maintain a significant amount of boilerplate code.
The SIMD types in this crate are a thin newtype around the corresponding array, for example f32x4 is a newtype for [f32; 4], and also contains a zero-sized token representing a witness to the CPU level. These types are in the crate root and have a number of methods, including loading and storing, that do not require intrinsics. The SIMD types are aligned, but this only affects storage.
A central idea is "levels," which represent a set of target features. Each level has a corresponding module. Each module for a level has a number of submodules, one for each type (though with an underscore instead of x to avoid name collision), with a large number of free functions for SIMD operations that operate on that type.
On aarch64, the levels are neon and fp16. The fp16 level implies neon.
On x86-64, the currently supported level is avx2. This is actually short for the x86-64-v3 microarchitecture level, which corresponds roughly to Haswell. The avx512 level is also planned, which is x86-64-v4.
This crate was inspired by pulp, std::simd, among others in the Rust ecosystem, though makes many decisions differently. It benefited from conversations with Luca Versari, though he is not responsible for any of the mistakes or bad decisions.
The half code was mentioned above. The proc macro was strongly inspired by the safe-arch-macro in rbrotli-enc.