wide/

u16x32_.rs

use super::*;

pick! {
  if #[cfg(target_feature="avx512bw")] {
    /// A SIMD vector with 32 elements of type [`u16`].
    ///
    /// See the [crate level documentation] for more information about SIMD
    /// vectors.
    ///
    /// [crate level documentation]: crate
    #[derive(Default, Clone, Copy, PartialEq, Eq)]
    #[repr(C, align(64))]
    pub struct u16x32 { pub(crate) avx512: m512i }
  } else {
    /// A SIMD vector with 32 elements of type [`u16`].
    ///
    /// See the [crate level documentation] for more information about SIMD
    /// vectors.
    ///
    /// [crate level documentation]: crate
    #[derive(Default, Clone, Copy, PartialEq, Eq)]
    #[repr(C, align(64))]
    pub struct u16x32 { pub(crate) a : u16x16, pub(crate) b : u16x16 }
  }
}

impl_simd_uint! {
  unsafe {
    T = u16,
    N = 32,
    Simd = u16x32,
    IntSimd = i16x32,
    T_BITS = 16,
    T_BITS_MUL_2 = 32,
    [
      0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20,
      21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
    ],
    optional_type_x86_inner { X86Inner = __m512i },
    optional_type_arm_inner {},
    optional_type_wasm_inner {},
  }

  #[inline]
  fn not(self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: bitxor_m512i(self.avx512, set_splat_i16_m512i(-1)) }
      } else {
        Self {
          a : self.a.not(),
          b : self.b.not(),
        }
      }
    }
  }

  #[inline]
  fn add(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: add_i16_m512i(self.avx512, rhs.avx512) }
      } else {
        Self {
          a : self.a.add(rhs.a),
          b : self.b.add(rhs.b),
        }
      }
    }
  }

  #[inline]
  fn sub(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: sub_i16_m512i(self.avx512, rhs.avx512) }
      } else {
        Self {
          a : self.a.sub(rhs.a),
          b : self.b.sub(rhs.b),
        }
      }
    }
  }

  #[inline]
  fn mul(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: mul_i16_keep_low_m512i(self.avx512, rhs.avx512) }
      } else {
        Self { a: self.a.mul(rhs.a), b: self.b.mul(rhs.b) }
      }
    }
  }

  #[inline]
  fn bitand(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: bitand_m512i(self.avx512, rhs.avx512) }
      } else {
        Self {
          a : self.a.bitand(rhs.a),
          b : self.b.bitand(rhs.b),
        }
      }
    }
  }

  #[inline]
  fn bitor(self, rhs: Self) -> Self::Output {
    pick! {
    if #[cfg(target_feature="avx512bw")] {
        Self { avx512: bitor_m512i(self.avx512, rhs.avx512) }
      } else {
        Self {
          a : self.a.bitor(rhs.a),
          b : self.b.bitor(rhs.b),
        }
      }
    }
  }

  #[inline]
  fn bitxor(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: bitxor_m512i(self.avx512, rhs.avx512) }
      } else {
        Self {
          a : self.a.bitxor(rhs.a),
          b : self.b.bitxor(rhs.b),
        }
      }
    }
  }

  #[inline]
  fn simd_eq(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: cmp_op_mask_u16_m512i::<{cmp_int_op!(Eq)}>(self.avx512, rhs.avx512) }
      } else {
        Self {
          a : self.a.simd_eq(rhs.a),
          b : self.b.simd_eq(rhs.b),
        }
      }
    }
  }

  #[inline]
  fn simd_ne(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: cmp_op_mask_u16_m512i::<{cmp_int_op!(Ne)}>(self.avx512, rhs.avx512) }
      } else {
        Self {
          a : self.a.simd_ne(rhs.a),
          b : self.b.simd_ne(rhs.b),
        }
      }
    }
  }

  #[inline]
  fn simd_lt(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: cmp_op_mask_u16_m512i::<{cmp_int_op!(Lt)}>(self.avx512, rhs.avx512) }
      } else {
        Self {
          a : rhs.a.simd_gt(self.a),
          b : rhs.b.simd_gt(self.b),
        }
      }
    }
  }

  #[inline]
  fn simd_gt(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: cmp_op_mask_u16_m512i::<{cmp_int_op!(Nle)}>(self.avx512, rhs.avx512) }
      } else {
        Self {
          a : self.a.simd_gt(rhs.a),
          b : self.b.simd_gt(rhs.b),
        }
      }
    }
  }

  #[inline]
  fn simd_le(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: cmp_op_mask_u16_m512i::<{cmp_int_op!(Le)}>(self.avx512, rhs.avx512) }
      } else {
        Self {
          a : self.a.simd_le(rhs.a),
          b : self.b.simd_le(rhs.b),
        }
      }
    }
  }

  #[inline]
  fn simd_ge(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: cmp_op_mask_u16_m512i::<{cmp_int_op!(Nlt)}>(self.avx512, rhs.avx512) }
      } else {
        Self {
          a : self.a.simd_ge(rhs.a),
          b : self.b.simd_ge(rhs.b),
        }
      }
    }
  }

  #[inline]
  pub fn reduce_add(self) -> u16 {
    let array: [u16x16; 2] = cast(self);
    (array[0] + array[1]).reduce_add()
  }

  #[inline]
  pub fn reduce_mul(self) -> u16 {
    let array: [u16x16; 2] = cast(self);
    (array[0] * array[1]).reduce_mul()
  }

  #[inline]
  pub fn bitselect(self, if_one: Self, if_zero: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self {
          avx512: bitor_m512i(
            bitand_m512i(if_one.avx512, self.avx512),
            bitandnot_m512i(self.avx512, if_zero.avx512),
          ),
        }
      } else {
        Self {
          a: self.a.bitselect(if_one.a, if_zero.a),
          b: self.b.bitselect(if_one.b, if_zero.b),
        }
      }
    }
  }

  #[inline]
  fn select(self, if_true: Self, if_false: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: blend_varying_i8_m512i(if_false.avx512,if_true.avx512,movepi8_mask_m512i(self.avx512)) }
      } else {
        Self {
          a : self.a.select(if_true.a, if_false.a),
          b : self.b.select(if_true.b, if_false.b),
        }
      }
    }
  }

  #[inline]
  pub fn to_bitmask(self) -> u32 {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        // use f16 move_mask since it is the same size as i16
        movepi16_mask_m512i(self.avx512) as u32
      } else {
        self.a.to_bitmask() | (self.b.to_bitmask() << 16)
      }
    }
  }

  #[inline]
  pub fn any(self) -> bool {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        movepi16_mask_m512i(self.avx512) != 0
      } else {
        (self.a | self.b).any()
      }
    }
  }

  #[inline]
  pub fn all(self) -> bool {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        movepi16_mask_m512i(self.avx512) == 0xFFFFFFFF
      } else {
        (self.a & self.b).all()
      }
    }
  }

  #[inline]
  pub fn shuffle(self, indices: u16x32) -> Self {
    pick! {
      if #[cfg(all(target_feature = "avx512bw"))] {
        #[cfg(target_arch = "x86")]
        use core::arch::x86::_mm512_permutexvar_epi16;
        #[cfg(target_arch = "x86_64")]
        use core::arch::x86_64::_mm512_permutexvar_epi16;
        // TODO(safe_arch): Add `_mm512_permutexvar_epi16`
        Self { avx512: unsafe { m512i(_mm512_permutexvar_epi16(indices.avx512.0, self.avx512.0)) } }
      } else {
        let self_halfs = cast::<u16x32, [u16x16; 2]>(self);
        let [indices_a, indices_b] = cast::<u16x32, [u16x16; 2]>(indices);

        cast([self_halfs.shuffle(indices_a), self_halfs.shuffle(indices_b)])
      }
    }
  }

  #[inline]
  pub fn shuffle_zeroing(self, indices: u16x32) -> Self {
    pick! {
      if #[cfg(all(target_feature = "avx512bw"))] {
        self.shuffle(indices) & indices.simd_lt(32)
      } else {
        let self_halfs = cast::<u16x32, [u16x16; 2]>(self);
        let [indices_a, indices_b] = cast::<u16x32, [u16x16; 2]>(indices);

        cast([self_halfs.shuffle_zeroing(indices_a), self_halfs.shuffle_zeroing(indices_b)])
      }
    }
  }

  #[inline]
  pub fn shuffle_wrapping(self, indices: u16x32) -> Self {
    pick! {
      if #[cfg(all(target_feature = "avx512bw"))] {
        // `avx512` shuffle intrinsics are wrapping
        self.shuffle(indices)
      } else {
        let self_halfs = cast::<u16x32, [u16x16; 2]>(self);
        let [indices_a, indices_b] = cast::<u16x32, [u16x16; 2]>(indices);

        cast([self_halfs.shuffle_wrapping(indices_a), self_halfs.shuffle_wrapping(indices_b)])
      }
    }
  }

  #[inline]
  fn shuffle(self: [u16x32; 2], indices: u16x32) -> u16x32 {
    pick! {
      if #[cfg(all(target_feature = "avx512bw"))] {
        #[cfg(target_arch = "x86")]
        use core::arch::x86::_mm512_permutex2var_epi16;
        #[cfg(target_arch = "x86_64")]
        use core::arch::x86_64::_mm512_permutex2var_epi16;
        // TODO(safe_arch): add `_mm512_permutex2var_epi16`.
        u16x32 {
          avx512: unsafe {
            m512i(_mm512_permutex2var_epi16(self[0].avx512.0, indices.avx512.0, self[1].avx512.0))
          },
        }
      } else {
        self[0].shuffle_zeroing(indices) | self[1].shuffle_zeroing(indices - 32)
      }
    }
  }

  #[inline]
  fn shuffle_zeroing(self: [u16x32; 2], indices: u16x32) -> u16x32 {
    pick! {
      if #[cfg(all(target_feature = "avx512bw", target_feature = "avx512vl"))] {
        self.shuffle(indices) & indices.simd_lt(64)
      } else {
        self.shuffle(indices)
      }
    }
  }

  #[inline]
  fn shuffle_wrapping(self: [u16x32; 2], indices: u16x32) -> u16x32 {
    pick! {
      if #[cfg(all(target_feature = "avx512bw"))] {
        // `avx512` shuffle intrinsics are wrapping
        self.shuffle(indices)
      } else {
        self.shuffle(indices & 63)
      }
    }
  }

  #[inline]
  fn shuffle(self: [u16x32; 3], indices: u16x32) -> u16x32 {
    [self[0], self[1]].shuffle_zeroing(indices) | self[2].shuffle_zeroing(indices - 64)
  }

  #[inline]
  fn shuffle_zeroing(self: [u16x32; 3], indices: u16x32) -> u16x32 {
    self.shuffle(indices)
  }

  #[inline]
  fn shuffle_wrapping(self: [u16x32; 3], indices: u16x32) -> u16x32 {
    self.shuffle(indices % 96)
  }

  #[inline]
  fn shuffle(self: [u16x32; 4], indices: u16x32) -> u16x32 {
    [self[0], self[1]].shuffle_zeroing(indices) | [self[2], self[3]].shuffle_zeroing(indices - 64)
  }

  #[inline]
  fn shuffle_zeroing(self: [u16x32; 4], indices: u16x32) -> u16x32 {
    self.shuffle(indices)
  }

  #[inline]
  fn shuffle_wrapping(self: [u16x32; 4], indices: u16x32) -> u16x32 {
    self.shuffle(indices & 127)
  }

  ///
  /// Currently this function is never accelerated.
  #[inline]
  pub fn transpose(data: [Self; 32]) -> [Self; 32] {
    // Can this be optimized?

    #[inline(always)]
    fn transpose_column(data: &[u16x32; 32], index: usize) -> u16x32 {
      u16x32::new([
        data[0].as_array()[index],
        data[1].as_array()[index],
        data[2].as_array()[index],
        data[3].as_array()[index],
        data[4].as_array()[index],
        data[5].as_array()[index],
        data[6].as_array()[index],
        data[7].as_array()[index],
        data[8].as_array()[index],
        data[9].as_array()[index],
        data[10].as_array()[index],
        data[11].as_array()[index],
        data[12].as_array()[index],
        data[13].as_array()[index],
        data[14].as_array()[index],
        data[15].as_array()[index],
        data[16].as_array()[index],
        data[17].as_array()[index],
        data[18].as_array()[index],
        data[19].as_array()[index],
        data[20].as_array()[index],
        data[21].as_array()[index],
        data[22].as_array()[index],
        data[23].as_array()[index],
        data[24].as_array()[index],
        data[25].as_array()[index],
        data[26].as_array()[index],
        data[27].as_array()[index],
        data[28].as_array()[index],
        data[29].as_array()[index],
        data[30].as_array()[index],
        data[31].as_array()[index],
      ])
    }

    [
      transpose_column(&data, 0),
      transpose_column(&data, 1),
      transpose_column(&data, 2),
      transpose_column(&data, 3),
      transpose_column(&data, 4),
      transpose_column(&data, 5),
      transpose_column(&data, 6),
      transpose_column(&data, 7),
      transpose_column(&data, 8),
      transpose_column(&data, 9),
      transpose_column(&data, 10),
      transpose_column(&data, 11),
      transpose_column(&data, 12),
      transpose_column(&data, 13),
      transpose_column(&data, 14),
      transpose_column(&data, 15),
      transpose_column(&data, 16),
      transpose_column(&data, 17),
      transpose_column(&data, 18),
      transpose_column(&data, 19),
      transpose_column(&data, 20),
      transpose_column(&data, 21),
      transpose_column(&data, 22),
      transpose_column(&data, 23),
      transpose_column(&data, 24),
      transpose_column(&data, 25),
      transpose_column(&data, 26),
      transpose_column(&data, 27),
      transpose_column(&data, 28),
      transpose_column(&data, 29),
      transpose_column(&data, 30),
      transpose_column(&data, 31),
    ]
  }

  #[inline]
  fn shl(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        // Mask `rhs` to 15 to match `wrapping_shl`.
        let rhs = bitand_m512i(rhs.avx512, set_splat_i16_m512i(15));
        Self { avx512: shl_each_u16_m512i(self.avx512, rhs) }
      } else {
        let [self_a, self_b]: [u16x16; 2] = cast(self);
        let [rhs_a, rhs_b]: [u16x16; 2] = cast(rhs);

        cast([self_a << rhs_a, self_b << rhs_b])
      }
    }
  }

  #[inline]
  fn shl(self, rhs: u32) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        // Use `rhs % 16` to perform wrapping shift and not unbounded shift.
        #[expect(clippy::suspicious_arithmetic_impl)]
        let shift = rhs as u16 & 15;
        Self { avx512: shl_all_u16_m512i(self.avx512, shift) }
      } else {
        Self {
          a : self.a.shl(rhs),
          b : self.b.shl(rhs),
        }
      }
    }
  }

  #[inline]
  fn shr(self, rhs: Self) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        // Mask `rhs` to 15 to match `wrapping_shr`.
        let rhs = bitand_m512i(rhs.avx512, set_splat_i16_m512i(15));
        Self { avx512: shr_each_u16_m512i(self.avx512, rhs) }
      } else {
        let [self_a, self_b]: [u16x16; 2] = cast(self);
        let [rhs_a, rhs_b]: [u16x16; 2] = cast(rhs);

        cast([self_a >> rhs_a, self_b >> rhs_b])
      }
    }
  }

  #[inline]
  fn shr(self, rhs: u32) -> Self::Output {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        // Use `rhs % 16` to perform wrapping shift and not unbounded shift.
        #[expect(clippy::suspicious_arithmetic_impl)]
        let shift = rhs as u16 & 15;
        Self { avx512: shr_all_u16_m512i(self.avx512, shift) }
      } else {
        Self {
          a : self.a.shr(rhs),
          b : self.b.shr(rhs),
        }
      }
    }
  }

  #[inline]
  pub fn max(self, rhs: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: max_u16_m512i(self.avx512, rhs.avx512) }
      } else {
        Self {
          a: self.a.max(rhs.a),
          b: self.b.max(rhs.b),
        }
      }
    }
  }

  #[inline]
  pub fn min(self, rhs: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: min_u16_m512i(self.avx512, rhs.avx512) }
      } else {
        Self {
          a: self.a.min(rhs.a),
          b: self.b.min(rhs.b),
        }
      }
    }
  }

  #[inline]
  pub fn reduce_max(self) -> u16 {
    let array: [u16x16; 2] = cast(self);
    array[0].max(array[1]).reduce_max()
  }

  #[inline]
  pub fn reduce_min(self) -> u16 {
    let array: [u16x16; 2] = cast(self);
    array[0].min(array[1]).reduce_min()
  }

  #[inline]
  pub fn unbounded_shl(self, rhs: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: shl_each_u16_m512i(self.avx512, rhs.avx512) }
      } else {
        let [self_a, self_b] = cast::<u16x32, [u16x16; 2]>(self);
        let [rhs_a, rhs_b] = cast::<u16x32, [u16x16; 2]>(rhs);

        cast([self_a.unbounded_shl(rhs_a), self_b.unbounded_shl(rhs_b)])
      }
    }
  }

  #[inline]
  pub fn unbounded_shl_scalar(self, rhs: u32) -> Self {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        // `u32 as u16` truncates the higher half so we need to manually
        // saturate.
        Self { avx512: shl_all_u16_m512i(self.avx512, rhs.min(u16::MAX as u32) as u16) }
      } else {
        Self {
          a: self.a.unbounded_shl_scalar(rhs),
          b: self.b.unbounded_shl_scalar(rhs),
        }
      }
    }
  }

  #[inline]
  pub fn unbounded_shr(self, rhs: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: shr_each_u16_m512i(self.avx512, rhs.avx512) }
      } else {
        let [self_a, self_b] = cast::<u16x32, [u16x16; 2]>(self);
        let [rhs_a, rhs_b] = cast::<u16x32, [u16x16; 2]>(rhs);

        cast([self_a.unbounded_shr(rhs_a), self_b.unbounded_shr(rhs_b)])
      }
    }
  }

  #[inline]
  pub fn unbounded_shr_scalar(self, rhs: u32) -> Self {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        // `u32 as u16` truncates the higher half so we need to manually
        // saturate.
        Self { avx512: shr_all_u16_m512i(self.avx512, rhs.min(u16::MAX as u32) as u16) }
      } else {
        Self {
          a: self.a.unbounded_shr_scalar(rhs),
          b: self.b.unbounded_shr_scalar(rhs),
        }
      }
    }
  }

  #[inline]
  pub fn saturating_add(self, rhs: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: add_saturating_u16_m512i(self.avx512, rhs.avx512) }
      } else {
        Self {
          a: self.a.saturating_add(rhs.a),
          b: self.b.saturating_add(rhs.b),
        }
      }
    }
  }

  #[inline]
  pub fn saturating_sub(self, rhs: Self) -> Self {
    pick! {
      if #[cfg(target_feature="avx512bw")] {
        Self { avx512: sub_saturating_u16_m512i(self.avx512, rhs.avx512) }
      } else {
        Self {
          a: self.a.saturating_sub(rhs.a),
          b: self.b.saturating_sub(rhs.b),
        }
      }
    }
  }

  #[inline]
  pub fn overflowing_mul(self, rhs: Self) -> (Self, Self) {
    let (low, high) = self.mul_keep_low_high(rhs);
    let overflow = high.simd_ne(Self::ZERO);
    (low, overflow)
  }

  optional_fn_widening_mul {
    // Cannot have `widening_mul` because there is no `u32x32` type.
  }

  #[inline]
  pub fn mul_keep_low_high(self, rhs: Self) -> (Self, Self) {
    // x86 has no `_mm512_mul_epu16` intrinsic so there is no `avx512`
    // optimization.

    let [self_a, self_b] = cast::<u16x32, [u16x16; 2]>(self);
    let [rhs_a, rhs_b] = cast::<u16x32, [u16x16; 2]>(rhs);

    let result_a = self_a.mul_keep_low_high(rhs_a);
    let result_b = self_b.mul_keep_low_high(rhs_b);
    (cast([result_a.0, result_b.0]), cast([result_a.1, result_b.1]))
  }

  #[inline]
  pub fn mul_keep_high(self, rhs: Self) -> Self {
    // x86 has no `_mm512_mul_epu16` intrinsic so there is no `avx512`
    // optimization.

    let [self_a, self_b] = cast::<u16x32, [u16x16; 2]>(self);
    let [rhs_a, rhs_b] = cast::<u16x32, [u16x16; 2]>(rhs);

    cast([self_a.mul_keep_high(rhs_a), self_b.mul_keep_high(rhs_b)])
  }

  optional_fn_deserialize {}
}