lightweight-simd 0.1.0

// Most or all contents generated by AI (model glm-5.3)

//! Lane-wise operators, float functions, comparisons, and masks for the vector
//! types ([`Simd`] and its [`aligned`](crate::aligned) siblings).

use core::ops::*;
#[cfg(any(feature = "use_libm_fma", target_arch = "aarch64", target_feature = "fma"))]
use num_traits::MulAdd;
use num_traits::{Float, Num, NumAssignOps};

use crate::{Aligned4, Aligned8, Aligned16, Aligned32, Aligned64, Simd};

impl<const N: usize> Simd<bool, N> {
    /// Lane-wise select: picks lanes of `if_true` where the mask is `true` and
    /// lanes of `if_false` where it is `false`.
    #[inline(always)]
    pub fn select<T: Copy>(self, if_true: Simd<T, N>, if_false: Simd<T, N>) -> Simd<T, N> {
        let mut out = if_true.0;
        for ((o, f), m) in out.iter_mut().zip(if_false.0).zip(self.0) {
            if !m {
                *o = f;
            }
        }
        Simd(out)
    }

    /// Returns `true` if any lane is `true`.
    #[inline(always)]
    pub fn any_true(self) -> bool {
        self.0.iter().any(|b| *b)
    }

    /// Returns `true` if all lanes are `true`.
    #[inline(always)]
    pub fn all_true(self) -> bool {
        self.0.iter().all(|b| *b)
    }
}

#[duplicate::duplicate_item(
    _mod_        _vec_;
    [simd_impls] [Simd];
    [impls_a4]   [Aligned4];
    [impls_a8]   [Aligned8];
    [impls_a16]  [Aligned16];
    [impls_a32]  [Aligned32];
    [impls_a64]  [Aligned64];
)]
mod _mod_ {
    use super::*;

    /* #region arithmetic operators */

    #[duplicate::duplicate_item(
        _trait_ _method_;
        [Add]  [add] ;
        [Sub]  [sub] ;
        [Mul]  [mul] ;
        [Div]  [div] ;
    )]
    impl<T, const N: usize> _trait_ for _vec_<T, N>
    where
        T: Num + Copy,
    {
        type Output = Self;

        #[inline(always)]
        fn _method_(self, rhs: Self) -> Self {
            let mut out = self.0;
            for (o, r) in out.iter_mut().zip(rhs.0) {
                *o = _trait_::_method_(*o, r);
            }
            Self(out)
        }
    }

    #[duplicate::duplicate_item(
        _trait_ _method_;
        [Add]  [add] ;
        [Sub]  [sub] ;
        [Mul]  [mul] ;
        [Div]  [div] ;
    )]
    impl<T, const N: usize> _trait_<T> for _vec_<T, N>
    where
        T: Num + Copy,
    {
        type Output = Self;

        #[inline(always)]
        fn _method_(self, rhs: T) -> Self {
            let mut out = self.0;
            for o in out.iter_mut() {
                *o = _trait_::_method_(*o, rhs);
            }
            Self(out)
        }
    }

    #[duplicate::duplicate_item(
        _trait_ _method_;
        [AddAssign] [add_assign];
        [SubAssign] [sub_assign];
        [MulAssign] [mul_assign];
        [DivAssign] [div_assign];
    )]
    impl<T, const N: usize> _trait_ for _vec_<T, N>
    where
        T: NumAssignOps + Copy,
    {
        #[inline(always)]
        fn _method_(&mut self, rhs: Self) {
            for (o, r) in self.0.iter_mut().zip(rhs.0) {
                _trait_::_method_(o, r);
            }
        }
    }

    #[duplicate::duplicate_item(
        _trait_ _method_;
        [AddAssign] [add_assign];
        [SubAssign] [sub_assign];
        [MulAssign] [mul_assign];
        [DivAssign] [div_assign];
    )]
    impl<T, const N: usize> _trait_<T> for _vec_<T, N>
    where
        T: NumAssignOps + Copy,
    {
        #[inline(always)]
        fn _method_(&mut self, rhs: T) {
            for o in self.0.iter_mut() {
                _trait_::_method_(o, rhs);
            }
        }
    }

    impl<T, const N: usize> Neg for _vec_<T, N>
    where
        T: Neg<Output = T> + Copy,
    {
        type Output = Self;

        #[inline(always)]
        fn neg(self) -> Self {
            let mut out = self.0;
            for (o, x) in out.iter_mut().zip(self.0) {
                *o = Neg::neg(x);
            }
            Self(out)
        }
    }

    /* #endregion */

    /* #region bitwise and shift operators */

    #[duplicate::duplicate_item(
        _trait_ _method_;
        [BitAnd] [bitand];
        [BitOr]  [bitor];
        [BitXor] [bitxor];
    )]
    impl<T, const N: usize> _trait_ for _vec_<T, N>
    where
        T: _trait_<Output = T> + Copy,
    {
        type Output = Self;

        #[inline(always)]
        fn _method_(self, rhs: Self) -> Self {
            let mut out = self.0;
            for (o, r) in out.iter_mut().zip(rhs.0) {
                *o = _trait_::_method_(*o, r);
            }
            Self(out)
        }
    }

    #[duplicate::duplicate_item(
        _trait_ _method_;
        [BitAndAssign] [bitand_assign];
        [BitOrAssign]  [bitor_assign];
        [BitXorAssign] [bitxor_assign];
    )]
    impl<T, const N: usize> _trait_ for _vec_<T, N>
    where
        T: _trait_ + Copy,
    {
        #[inline(always)]
        fn _method_(&mut self, rhs: Self) {
            for (o, r) in self.0.iter_mut().zip(rhs.0) {
                _trait_::_method_(o, r);
            }
        }
    }

    impl<T, const N: usize> Not for _vec_<T, N>
    where
        T: Not<Output = T> + Copy,
    {
        type Output = Self;

        #[inline(always)]
        fn not(self) -> Self {
            let mut out = self.0;
            for (o, x) in out.iter_mut().zip(self.0) {
                *o = Not::not(x);
            }
            Self(out)
        }
    }

    #[duplicate::duplicate_item(
        _trait_ _method_;
        [Shl] [shl];
        [Shr] [shr];
    )]
    impl<T, const N: usize> _trait_<usize> for _vec_<T, N>
    where
        T: _trait_<usize, Output = T> + Copy,
    {
        type Output = Self;

        #[inline(always)]
        fn _method_(self, rhs: usize) -> Self {
            let mut out = self.0;
            for o in out.iter_mut() {
                *o = _trait_::_method_(*o, rhs);
            }
            Self(out)
        }
    }

    /* #endregion */

    /* #region fused multiply-add */
    //
    // The fused path is selected when the compilation target has hardware
    // FMA (`target_feature = "fma"`, which follows `-C target-cpu`, i.e. any
    // x86-64-v3+ or `native` build; on aarch64 scalar/vector FMA is baseline)
    // or when the `use_libm_fma` feature pins it on. Everything else —
    // including architectures the predicate does not know — gets the fast
    // separated form, never the slow libm software `fma`. See docs/adr/0005.

    #[cfg(any(feature = "use_libm_fma", target_arch = "aarch64", target_feature = "fma"))]
    impl<T, const N: usize> _vec_<T, N>
    where
        T: MulAdd<Output = T> + Copy,
    {
        /// Lane-wise fused multiply-add: `self * b + c` with a single
        /// rounding per lane.
        ///
        /// This is the default on targets with hardware FMA (x86 built with
        /// `-C target-cpu` of `x86-64-v3` or higher, `native`, and all
        /// aarch64), where it lowers to one FMA instruction per register
        /// with no extra flags — and is also faster than a separated
        /// multiply and add there. The non-default cargo feature
        /// `use_libm_fma` selects this path on *every* target, accepting a
        /// correctly-rounded but slow libm `fma` call (~4.6x a separated
        /// mul+add) where hardware FMA is absent.
        #[inline(always)]
        pub fn mul_add(self, b: Self, c: Self) -> Self {
            let mut out = self.0;
            for ((o, b), c) in out.iter_mut().zip(b.0).zip(c.0) {
                *o = MulAdd::mul_add(*o, b, c);
            }
            Self(out)
        }

        /// Lane-wise fused `*self = b * c + *self`, by the receiver.
        ///
        /// Note that the operand order differs from [`mul_add`](Self::mul_add):
        /// the accumulator is the receiver, matching `self += b * c`.
        #[inline(always)]
        pub fn fma_from(&mut self, b: Self, c: Self) {
            for ((b, c), o) in b.0.iter().zip(c.0.iter()).zip(self.0.iter_mut()) {
                *o = MulAdd::mul_add(*b, *c, *o);
            }
        }
    }

    #[cfg(not(any(feature = "use_libm_fma", target_arch = "aarch64", target_feature = "fma")))]
    impl<T, const N: usize> _vec_<T, N>
    where
        T: Mul<Output = T> + Add<Output = T> + Copy,
    {
        /// Lane-wise multiply-add: `self * b + c`, evaluated as a separate
        /// multiply and add (two roundings per lane).
        ///
        /// This is the fallback on targets without hardware FMA (generic
        /// x86-64, `x86-64-v2`), where it avoids the slow software-fused
        /// libm `fma` calls; it also fuses into a single hardware FMA
        /// instruction when compiled with `-C llvm-args=-fp-contract=fast`.
        /// Enable the `use_libm_fma` feature to get the single-rounding
        /// fused form on such targets instead.
        #[inline(always)]
        pub fn mul_add(self, b: Self, c: Self) -> Self {
            let mut out = self.0;
            for ((o, b), c) in out.iter_mut().zip(b.0).zip(c.0) {
                *o = *o * b + c;
            }
            Self(out)
        }

        /// Lane-wise `*self = b * c + *self`, by the receiver.
        ///
        /// Note that the operand order differs from [`mul_add`](Self::mul_add):
        /// the accumulator is the receiver, matching `self += b * c`. The
        /// evaluation strategy is the same as [`mul_add`](Self::mul_add)'s.
        #[inline(always)]
        pub fn fma_from(&mut self, b: Self, c: Self) {
            for ((b, c), o) in b.0.iter().zip(c.0.iter()).zip(self.0.iter_mut()) {
                *o = *b * *c + *o;
            }
        }
    }

    /* #endregion */

    /* #region float functions */

    impl<T, const N: usize> _vec_<T, N>
    where
        T: Float + Copy,
    {
        #[duplicate::duplicate_item(
            _name_;
            [abs];
            [sqrt];
            [recip];
            [floor];
            [ceil];
            [round];
            [trunc];
            [fract];
        )]
        #[doc = concat!("Lane-wise [`Float::", stringify!(_name_), "`].")]
        #[inline(always)]
        pub fn _name_(self) -> Self {
            Self(self.0.map(|x| x._name_()))
        }

        /// Lane-wise minimum; NaN handling follows the element's scalar
        /// [`Float::min`].
        #[inline(always)]
        pub fn min(self, other: Self) -> Self {
            let mut out = self.0;
            for (o, r) in out.iter_mut().zip(other.0) {
                *o = o.min(r);
            }
            Self(out)
        }

        /// Lane-wise maximum; NaN handling follows the element's scalar
        /// [`Float::max`].
        #[inline(always)]
        pub fn max(self, other: Self) -> Self {
            let mut out = self.0;
            for (o, r) in out.iter_mut().zip(other.0) {
                *o = o.max(r);
            }
            Self(out)
        }

        /// Lane-wise [`Float::copysign`].
        #[inline(always)]
        pub fn copysign(self, sign: Self) -> Self {
            let mut out = self.0;
            for (o, r) in out.iter_mut().zip(sign.0) {
                *o = o.copysign(r);
            }
            Self(out)
        }
    }

    impl<const N: usize> _vec_<f32, N> {
        /// Lane-wise [`f32::round_ties_even`].
        ///
        /// Not available generically through [`num_traits::Float`], so this is
        /// implemented only for the builtin float types.
        #[inline(always)]
        pub fn round_ties_even(self) -> Self {
            Self(self.0.map(f32::round_ties_even))
        }
    }

    impl<const N: usize> _vec_<f64, N> {
        /// Lane-wise [`f64::round_ties_even`].
        ///
        /// Not available generically through [`num_traits::Float`], so this is
        /// implemented only for the builtin float types.
        #[inline(always)]
        pub fn round_ties_even(self) -> Self {
            Self(self.0.map(f64::round_ties_even))
        }
    }

    /* #endregion */

    /* #region comparisons */

    impl<T, const N: usize> _vec_<T, N>
    where
        T: PartialOrd + Copy,
    {
        /// Lane-wise `self < rhs`.
        #[inline(always)]
        pub fn simd_lt(self, rhs: Self) -> Simd<bool, N> {
            Simd(core::array::from_fn(|i| self.0[i] < rhs.0[i]))
        }

        /// Lane-wise `self <= rhs`.
        #[inline(always)]
        pub fn simd_le(self, rhs: Self) -> Simd<bool, N> {
            Simd(core::array::from_fn(|i| self.0[i] <= rhs.0[i]))
        }

        /// Lane-wise `self > rhs`.
        #[inline(always)]
        pub fn simd_gt(self, rhs: Self) -> Simd<bool, N> {
            Simd(core::array::from_fn(|i| self.0[i] > rhs.0[i]))
        }

        /// Lane-wise `self >= rhs`.
        #[inline(always)]
        pub fn simd_ge(self, rhs: Self) -> Simd<bool, N> {
            Simd(core::array::from_fn(|i| self.0[i] >= rhs.0[i]))
        }
    }

    impl<T, const N: usize> _vec_<T, N>
    where
        T: PartialEq + Copy,
    {
        /// Lane-wise `self == rhs`.
        #[inline(always)]
        pub fn simd_eq(self, rhs: Self) -> Simd<bool, N> {
            Simd(core::array::from_fn(|i| self.0[i] == rhs.0[i]))
        }

        /// Lane-wise `self != rhs`.
        #[inline(always)]
        pub fn simd_ne(self, rhs: Self) -> Simd<bool, N> {
            Simd(core::array::from_fn(|i| self.0[i] != rhs.0[i]))
        }
    }

    /* #endregion */
}