mikros/rust
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Add footnote

Browse Source
This commit is contained in:
Caleb Zulawski 2023-07-28 21:38:58 -04:00
parent 6e8d21ee76
commit 3da60554e5
1 changed files with 3 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
crates/core_simd/src/core_simd_docs.md
View File

@ -30,8 +30,10 @@ Instead, they map to a reasonable implementation of the operation for the target
Consistency between targets is not compromised to use faster or fewer instructions.
In some cases, `std::arch` will provide a faster function that has slightly different behavior than the `std::simd` equivalent.
For example, [`_mm_min_ps`](`core::arch::x86_64::_mm_min_ps`) can be slightly faster than [`SimdFloat::simd_min`], but does not conform to the IEEE standard also used by [`f32::min`].
For example, [`_mm_min_ps`](`core::arch::x86_64::_mm_min_ps`)[^1] can be slightly faster than [`SimdFloat::simd_min`], but does not conform to the IEEE standard also used by [`f32::min`].
When necessary, [`Simd<T, N>`] can be converted to the types provided by `std::arch` to make use of target-specific functions.
Many targets simply don't have SIMD, or don't support SIMD for a particular element type.
In those cases, regular scalar operations are generated instead.
[^1]: `_mm_min_ps(x, y)` is equivalent to `x.simd_lt(y).select(x, y)`