[−][src]Crate wide
A crate to help you go wide.
Specifically, this crate has data types for blocks of primitives packed together and used as a single unit. This works very well with SIMD/vector hardware of various targets. Both in terms of explicit SIMD usage and also in terms of allowing LLVM's auto-vectorizer to do its job.
All SIMD usage is on a best effort basis. Results will vary based on target, optimization level, method, and if you're using a Nightly compiler or not. Otherwise you get a "fallback" implementation, which will just do the normal computation on each lane individually.
- Note: The crate will auto-detect if you're using Nightly and take
advantage of it. You don't do anything on your part. Activate the
always_use_stablefeature if you'd like to suppress this effect such as for testing purposes.
What About packed_simd?
Compared to the
packed_simd RFC efforts,
this crate is less concerned with complete coverage of all possible
intrinsics and being totally generic across all widths. Instead, I focus on
having a very simple, easy to understand setup that avoids generics and
tries to just be plain and obvious at all times. The goal is that using a
wide type should be as close as possible to using the scalar version of the
same type. Some function designed for f32 inputs and outputs should "just
work" when you change it to f32x4 inputs and outputs.
Also, packed_simd is Nightly-only, whereas this crate works on Stable.
Even on Stable this crate will give you reasonable levels of SIMD just
from LLVM's auto-vectorizer being pretty good at its job when you give it
code that it recognizes.
When packed_simd eventually makes it into Stable it might make this
crate obsolete. However, in September of 2019 I asked the packed_simd
folks if there was any kind of ETA, 6 months, 12 months, or more, and they
just said "no ETA". So I'm not gonna wait around for packed_simd.
Modules
| consts | Various |
Structs
| f32x4 |