Expand description
Poor-man’s SIMD
The difference between this implementation and std::simd is that this implementation holds the SIMD register directly, thus it exposes more optimization opportunity to use wide range of instructions available.
Also, it gives us more control, for example, it is likely that we will have f16/bf16 support before the standard library does.
The API are close to std::simd to make migration easier in the future.