62 lines (38 loc) · 1.42 KB

`wide` Changelog

1.3.0

Fixes the behaviour of f32x16 functions is_finite and round_int. They previously gave incorrect output in some cases.
Added more f32x16 functionality.

1.2.0

added reduce operations and dot to i16x32

1.1.2

Use native NEON intrinsics for f32x4::blend and f64x2::blend on aarch64, improving performance by using a single vbslq instruction instead of the generic 3-operation fallback.

1.1.1

Further improvements to the Neg impls on non-x86 targets.

1.1.0

add i64x2::min, i64x2::max.
add u64x2::min, u64x2::max.
add u64x8::min.

1.0.3

fix floating point negation edge cases, which also allows optimizations in more cases.

1.0.2

fix edge case where wide was using the wrong avx512 sub-features to select when safe_arch functions could be called, causing build errors.

1.0.1

initial stable version

0.8.3

to_bitmask implemented for all unsigned int types.

0.8.2

Fixed additional build errors on wasm targets.

0.8.1

Fixed type errors in the fallback implementations of some methods that completely prevented compilation of the crate on some targets.

0.8.0

Breaking:
- move_mask returns unsigned values now.
- move_mask and many other methods renamed to better align with the standard library portable simd types. See issue 209 for details.