#[cfg(all(feature = "libm", not(feature = "std")))]
#[allow(
unused_imports,
reason = "Generated backends use different subsets of these helpers"
)]
use crate::math::FloatExt as _;
use crate::{Level, arch_types::ArchTypes, prelude::*, seal::Seal};
use crate::{
f32x4, f32x8, f32x16, f64x2, f64x4, f64x8, i8x16, i8x32, i8x64, i16x8, i16x16, i16x32, i32x4,
i32x8, i32x16, i64x2, i64x4, i64x8, mask8x16, mask8x32, mask8x64, mask16x8, mask16x16,
mask16x32, mask32x4, mask32x8, mask32x16, mask64x2, mask64x4, mask64x8, u8x16, u8x32, u8x64,
u16x8, u16x16, u16x32, u32x4, u32x8, u32x16, u64x2, u64x4, u64x8,
};
#[cfg(target_arch = "x86")]
use core::arch::x86::*;
#[cfg(target_arch = "x86_64")]
use core::arch::x86_64::*;
use core::ops::*;
#[doc = "A token for SSE4.2 intrinsics on `x86` and `x86_64`, representing the x86-64-v2 level."]
#[doc = "# Browsing the documentation"]
#[doc = "The method list on this struct is very verbose."]
#[doc = "A better way to browse the docs is by looking at vector types such as [`u32x4`], [`f32x4`] or [`mask32x4`]."]
#[doc = "They include all the operations listed here, and also provide some additional convenience methods."]
#[derive(Clone, Copy, Debug)]
pub struct Sse4_2 {
_private: (),
}
impl Sse4_2 {
#[doc = "Create a SIMD token proving that the x86-64-v2 features are available."]
#[doc = r""]
#[doc = r" Most users should safely obtain the token from [`Level::new`] instead of calling this function."]
#[doc = r""]
#[doc = r" This function can be called without an `unsafe` block from a function"]
#[doc = r" with all the required target features enabled via the `#[target_feature]` annotation."]
#[doc = r""]
#[doc = r" # Safety"]
#[doc = r""]
#[doc = "When invoking this function through an `unsafe` block, the caller must ensure that the current CPU supports `fxsr`, `sse4.2`, `cmpxchg16b`, `popcnt`."]
#[inline]
#[target_feature(enable = "fxsr,sse4.2,cmpxchg16b,popcnt")]
pub const fn assume_supported() -> Self {
Self { _private: () }
}
}
impl Seal for Sse4_2 {}
impl ExtractToken for Sse4_2 {
type S = Self;
#[inline]
fn token(&self) -> Self {
*self
}
}
impl ArchTypes for Sse4_2 {
type f32x4 = crate::support::Aligned128<__m128>;
type i8x16 = crate::support::Aligned128<__m128i>;
type u8x16 = crate::support::Aligned128<__m128i>;
type mask8x16 = crate::support::Aligned128<__m128i>;
type i16x8 = crate::support::Aligned128<__m128i>;
type u16x8 = crate::support::Aligned128<__m128i>;
type mask16x8 = crate::support::Aligned128<__m128i>;
type i32x4 = crate::support::Aligned128<__m128i>;
type u32x4 = crate::support::Aligned128<__m128i>;
type mask32x4 = crate::support::Aligned128<__m128i>;
type f64x2 = crate::support::Aligned128<__m128d>;
type i64x2 = crate::support::Aligned128<__m128i>;
type u64x2 = crate::support::Aligned128<__m128i>;
type mask64x2 = crate::support::Aligned128<__m128i>;
type f32x8 = crate::support::Aligned256<[__m128; 2usize]>;
type i8x32 = crate::support::Aligned256<[__m128i; 2usize]>;
type u8x32 = crate::support::Aligned256<[__m128i; 2usize]>;
type mask8x32 = crate::support::Aligned256<[__m128i; 2usize]>;
type i16x16 = crate::support::Aligned256<[__m128i; 2usize]>;
type u16x16 = crate::support::Aligned256<[__m128i; 2usize]>;
type mask16x16 = crate::support::Aligned256<[__m128i; 2usize]>;
type i32x8 = crate::support::Aligned256<[__m128i; 2usize]>;
type u32x8 = crate::support::Aligned256<[__m128i; 2usize]>;
type mask32x8 = crate::support::Aligned256<[__m128i; 2usize]>;
type f64x4 = crate::support::Aligned256<[__m128d; 2usize]>;
type i64x4 = crate::support::Aligned256<[__m128i; 2usize]>;
type u64x4 = crate::support::Aligned256<[__m128i; 2usize]>;
type mask64x4 = crate::support::Aligned256<[__m128i; 2usize]>;
type f32x16 = crate::support::Aligned512<[__m128; 4usize]>;
type i8x64 = crate::support::Aligned512<[__m128i; 4usize]>;
type u8x64 = crate::support::Aligned512<[__m128i; 4usize]>;
type mask8x64 = crate::support::Aligned512<[__m128i; 4usize]>;
type i16x32 = crate::support::Aligned512<[__m128i; 4usize]>;
type u16x32 = crate::support::Aligned512<[__m128i; 4usize]>;
type mask16x32 = crate::support::Aligned512<[__m128i; 4usize]>;
type i32x16 = crate::support::Aligned512<[__m128i; 4usize]>;
type u32x16 = crate::support::Aligned512<[__m128i; 4usize]>;
type mask32x16 = crate::support::Aligned512<[__m128i; 4usize]>;
type f64x8 = crate::support::Aligned512<[__m128d; 4usize]>;
type i64x8 = crate::support::Aligned512<[__m128i; 4usize]>;
type u64x8 = crate::support::Aligned512<[__m128i; 4usize]>;
type mask64x8 = crate::support::Aligned512<[__m128i; 4usize]>;
}
impl Simd for Sse4_2 {
type f32s = f32x4<Self>;
type f64s = f64x2<Self>;
type u8s = u8x16<Self>;
type i8s = i8x16<Self>;
type u16s = u16x8<Self>;
type i16s = i16x8<Self>;
type u32s = u32x4<Self>;
type i32s = i32x4<Self>;
type u64s = u64x2<Self>;
type i64s = i64x2<Self>;
type mask8s = mask8x16<Self>;
type mask16s = mask16x8<Self>;
type mask32s = mask32x4<Self>;
type mask64s = mask64x2<Self>;
#[inline(always)]
fn level(self) -> Level {
Level::Sse4_2(self)
}
#[inline]
fn vectorize<F: FnOnce() -> R, R>(self, f: F) -> R {
#[inline]
#[target_feature(enable = "fxsr,sse4.2,cmpxchg16b,popcnt")]
fn vectorize_sse4_2<F: FnOnce() -> R, R>(f: F) -> R {
f()
}
unsafe { vectorize_sse4_2(f) }
}
#[inline(always)]
fn abs_f32x4(self, a: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_andnot_ps(_mm_set1_ps(-0.0), a.into()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn splat_f32x4(self, val: f32) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, val: f32) -> f32x4<Sse4_2> {
_mm_set1_ps(val).simd_into(token)
}
);
kernel(self, val)
}
#[inline(always)]
fn reverse_f32x4(self, a: f32x4<Self>) -> f32x4<Self> {
let indices: u8x16<Self> =
[12, 13, 14, 15, 8, 9, 10, 11, 4, 5, 6, 7, 0, 1, 2, 3].simd_into(self);
self.swizzle_dyn_f32x4(a, indices)
}
#[inline(always)]
fn slide_f32x4<const SHIFT: usize>(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
if SHIFT >= 4usize {
return b;
}
let result = dyn_alignr_128(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 4usize,
);
Bytes::from_bytes(u8x16 {
val: crate::support::Aligned128(result),
simd: self,
})
}
#[inline(always)]
fn compress_f32x4(self, values: f32x4<Self>, mask: mask32x4<Self>) -> f32x4<Self> {
let bits = self.to_bitmask_mask32x4(mask) as usize & 15;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_32[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn compress_merge_f32x4(
self,
values: f32x4<Self>,
mask: mask32x4<Self>,
merge: f32x4<Self>,
) -> f32x4<Self> {
let bits = self.to_bitmask_mask32x4(mask) as usize & 15;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_32[bits].0);
let compressed = self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control);
let inactive = self.simd_lt_i8x16(i8x16::from_bytes(control), i8x16::splat(self, 0));
Bytes::from_bytes(self.select_u8x16(inactive, Bytes::to_bytes(merge), compressed))
}
#[inline(always)]
fn expand_f32x4(self, values: f32x4<Self>, mask: mask32x4<Self>) -> f32x4<Self> {
let bits = self.to_bitmask_mask32x4(mask) as usize & 15;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_32[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn expand_merge_f32x4(
self,
values: f32x4<Self>,
mask: mask32x4<Self>,
merge: f32x4<Self>,
) -> f32x4<Self> {
let bits = self.to_bitmask_mask32x4(mask) as usize & 15;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_32[bits].0);
let expanded = Bytes::from_bytes(self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control));
self.select_f32x4(mask, expanded, merge)
}
#[inline(always)]
fn neg_f32x4(self, a: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_xor_ps(a.into(), _mm_set1_ps(-0.0)).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn sqrt_f32x4(self, a: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_sqrt_ps(a.into()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn approximate_recip_f32x4(self, a: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_rcp_ps(a.into()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn add_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_add_ps(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn sub_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_sub_ps(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn mul_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_mul_ps(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn div_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_div_ps(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn copysign_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
let mask = _mm_set1_ps(-0.0);
_mm_or_ps(_mm_and_ps(mask, b.into()), _mm_andnot_ps(mask, a.into()))
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn reduce_max_f32x4(self, a: f32x4<Self>) -> f32 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> f32 {
let reduced: __m128 = a.into();
let shifted = _mm_castsi128_ps(_mm_srli_si128::<8>(_mm_castps_si128(reduced)));
let reduced = { _mm_max_ps(reduced, shifted) };
let shifted = _mm_castsi128_ps(_mm_srli_si128::<4>(_mm_castps_si128(reduced)));
let reduced = { _mm_max_ps(reduced, shifted) };
_mm_cvtss_f32(reduced)
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_min_f32x4(self, a: f32x4<Self>) -> f32 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> f32 {
let reduced: __m128 = a.into();
let shifted = _mm_castsi128_ps(_mm_srli_si128::<8>(_mm_castps_si128(reduced)));
let reduced = { _mm_min_ps(reduced, shifted) };
let shifted = _mm_castsi128_ps(_mm_srli_si128::<4>(_mm_castps_si128(reduced)));
let reduced = { _mm_min_ps(reduced, shifted) };
_mm_cvtss_f32(reduced)
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_max_precise_f32x4(self, a: f32x4<Self>) -> f32 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> f32 {
let reduced: __m128 = a.into();
let shifted = _mm_castsi128_ps(_mm_srli_si128::<8>(_mm_castps_si128(reduced)));
let reduced = {
let intermediate = _mm_max_ps(reduced, shifted);
let b_is_nan = _mm_cmpunord_ps(shifted, shifted);
_mm_blendv_ps(intermediate, reduced, b_is_nan)
};
let shifted = _mm_castsi128_ps(_mm_srli_si128::<4>(_mm_castps_si128(reduced)));
let reduced = {
let intermediate = _mm_max_ps(reduced, shifted);
let b_is_nan = _mm_cmpunord_ps(shifted, shifted);
_mm_blendv_ps(intermediate, reduced, b_is_nan)
};
_mm_cvtss_f32(reduced)
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_min_precise_f32x4(self, a: f32x4<Self>) -> f32 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> f32 {
let reduced: __m128 = a.into();
let shifted = _mm_castsi128_ps(_mm_srli_si128::<8>(_mm_castps_si128(reduced)));
let reduced = {
let intermediate = _mm_min_ps(reduced, shifted);
let b_is_nan = _mm_cmpunord_ps(shifted, shifted);
_mm_blendv_ps(intermediate, reduced, b_is_nan)
};
let shifted = _mm_castsi128_ps(_mm_srli_si128::<4>(_mm_castps_si128(reduced)));
let reduced = {
let intermediate = _mm_min_ps(reduced, shifted);
let b_is_nan = _mm_cmpunord_ps(shifted, shifted);
_mm_blendv_ps(intermediate, reduced, b_is_nan)
};
_mm_cvtss_f32(reduced)
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_sum_f32x4(self, a: f32x4<Self>) -> f32 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> f32 {
let a: __m128 = a.into();
let adjacent = _mm_add_ps(a, _mm_shuffle_ps::<0b10_11_00_01>(a, a));
_mm_cvtss_f32(_mm_add_ss(adjacent, _mm_movehl_ps(adjacent, adjacent)))
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_product_f32x4(self, a: f32x4<Self>) -> f32 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> f32 {
let a: __m128 = a.into();
let adjacent = _mm_mul_ps(a, _mm_shuffle_ps::<0b10_11_00_01>(a, a));
_mm_cvtss_f32(_mm_mul_ss(adjacent, _mm_movehl_ps(adjacent, adjacent)))
}
);
kernel(self, a)
}
#[inline(always)]
fn max_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_max_ps(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn min_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_min_ps(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn max_precise_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
let intermediate = _mm_max_ps(a.into(), b.into());
let b_is_nan = _mm_cmpunord_ps(b.into(), b.into());
_mm_blendv_ps(intermediate, a.into(), b_is_nan).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn min_precise_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
let intermediate = _mm_min_ps(a.into(), b.into());
let b_is_nan = _mm_cmpunord_ps(b.into(), b.into());
_mm_blendv_ps(intermediate, a.into(), b_is_nan).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_eq_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> mask32x4<Sse4_2> {
_mm_castps_si128(_mm_cmpeq_ps(a.into(), b.into())).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_ne_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> mask32x4<Sse4_2> {
_mm_castps_si128(_mm_cmpneq_ps(a.into(), b.into())).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_lt_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> mask32x4<Sse4_2> {
_mm_castps_si128(_mm_cmplt_ps(a.into(), b.into())).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_le_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> mask32x4<Sse4_2> {
_mm_castps_si128(_mm_cmple_ps(a.into(), b.into())).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_low_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_unpacklo_ps(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_high_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_unpackhi_ps(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_low_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_shuffle_ps::<0b10_00_10_00>(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_high_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>, b: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_shuffle_ps::<0b11_01_11_01>(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn interleave_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> (f32x4<Self>, f32x4<Self>) {
(self.zip_low_f32x4(a, b), self.zip_high_f32x4(a, b))
}
#[inline(always)]
fn deinterleave_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> (f32x4<Self>, f32x4<Self>) {
(self.unzip_low_f32x4(a, b), self.unzip_high_f32x4(a, b))
}
#[inline(always)]
fn mul_add_f32x4(self, a: f32x4<Self>, b: f32x4<Self>, c: f32x4<Self>) -> f32x4<Self> {
a * b + c
}
#[inline(always)]
fn mul_add_precise_f32x4(self, a: f32x4<Self>, b: f32x4<Self>, c: f32x4<Self>) -> f32x4<Self> {
crate::math::mul_add_precise_f32x4(self, a, b, c)
}
#[inline(always)]
fn mul_sub_f32x4(self, a: f32x4<Self>, b: f32x4<Self>, c: f32x4<Self>) -> f32x4<Self> {
a * b - c
}
#[inline(always)]
fn mul_sub_precise_f32x4(self, a: f32x4<Self>, b: f32x4<Self>, c: f32x4<Self>) -> f32x4<Self> {
self.mul_add_precise_f32x4(a, b, -c)
}
#[inline(always)]
fn floor_f32x4(self, a: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_round_ps::<{ _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC }>(a.into())
.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn ceil_f32x4(self, a: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_round_ps::<{ _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC }>(a.into())
.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn round_ties_even_f32x4(self, a: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_round_ps::<{ _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC }>(a.into())
.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn round_f32x4(self, a: f32x4<Self>) -> f32x4<Self> {
let bias = self.splat_f32x4(const { f32::next_down(0.5) });
(a + self.copysign_f32x4(bias, a)).trunc()
}
#[inline(always)]
fn fract_f32x4(self, a: f32x4<Self>) -> f32x4<Self> {
a - self.trunc_f32x4(a)
}
#[inline(always)]
fn trunc_f32x4(self, a: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_round_ps::<{ _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC }>(a.into())
.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn to_degrees_f32x4(self, a: f32x4<Self>) -> f32x4<Self> {
a * const { f32::to_degrees(1.0) }
}
#[inline(always)]
fn to_radians_f32x4(self, a: f32x4<Self>) -> f32x4<Self> {
a * const { f32::to_radians(1.0) }
}
#[inline(always)]
fn is_nan_f32x4(self, a: f32x4<Self>) -> mask32x4<Self> {
a.simd_ne(a)
}
#[inline(always)]
fn is_infinite_f32x4(self, a: f32x4<Self>) -> mask32x4<Self> {
a.abs().simd_eq(f32::INFINITY)
}
#[inline(always)]
fn is_finite_f32x4(self, a: f32x4<Self>) -> mask32x4<Self> {
a.abs().simd_lt(f32::INFINITY)
}
#[inline(always)]
fn is_subnormal_f32x4(self, a: f32x4<Self>) -> mask32x4<Self> {
let i: i32x4<Self> = a.bitcast();
(i & 0x7F80_0000_i32).simd_eq(0) & (i & 0x007F_FFFF_i32).simd_ne(0)
}
#[inline(always)]
fn is_normal_f32x4(self, a: f32x4<Self>) -> mask32x4<Self> {
let i: i32x4<Self> = a.bitcast();
let exp = i & 0x7F80_0000_i32;
exp.simd_ne(0) & exp.simd_ne(0x7F80_0000_i32)
}
#[inline(always)]
fn is_sign_positive_f32x4(self, a: f32x4<Self>) -> mask32x4<Self> {
let i: i32x4<Self> = a.bitcast();
(i & i32::MIN).simd_eq(0)
}
#[inline(always)]
fn is_sign_negative_f32x4(self, a: f32x4<Self>) -> mask32x4<Self> {
!a.is_sign_positive()
}
#[inline(always)]
fn select_f32x4(self, a: mask32x4<Self>, b: f32x4<Self>, c: f32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: mask32x4<Sse4_2>,
b: f32x4<Sse4_2>,
c: f32x4<Sse4_2>,
) -> f32x4<Sse4_2> {
_mm_blendv_ps(c.into(), b.into(), _mm_castsi128_ps(a.into())).simd_into(token)
}
);
kernel(self, a, b, c)
}
#[inline(always)]
fn combine_f32x4(self, a: f32x4<Self>, b: f32x4<Self>) -> f32x8<Self> {
f32x8 {
val: crate::support::Aligned256([a.val.0, b.val.0]),
simd: self,
}
}
#[inline(always)]
fn load_four_interleaved_f32x4(self, src: &[f32; 16usize]) -> [f32x4<Self>; 4usize] {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, src: &[f32; 16usize]) -> [f32x4<Sse4_2>; 4usize] {
let (chunks, []) = src.as_chunks::<4usize>() else {
unreachable!()
};
let v0: __m128 =
crate::transmute::checked_transmute_copy::<[f32; 4usize], __m128>(&chunks[0]);
let v1: __m128 =
crate::transmute::checked_transmute_copy::<[f32; 4usize], __m128>(&chunks[1]);
let v2: __m128 =
crate::transmute::checked_transmute_copy::<[f32; 4usize], __m128>(&chunks[2]);
let v3: __m128 =
crate::transmute::checked_transmute_copy::<[f32; 4usize], __m128>(&chunks[3]);
let tmp0 = _mm_unpacklo_ps(v0, v1);
let tmp1 = _mm_unpackhi_ps(v0, v1);
let tmp2 = _mm_unpacklo_ps(v2, v3);
let tmp3 = _mm_unpackhi_ps(v2, v3);
let out0 = _mm_castpd_ps(_mm_unpacklo_pd(_mm_castps_pd(tmp0), _mm_castps_pd(tmp2)));
let out1 = _mm_castpd_ps(_mm_unpackhi_pd(_mm_castps_pd(tmp0), _mm_castps_pd(tmp2)));
let out2 = _mm_castpd_ps(_mm_unpacklo_pd(_mm_castps_pd(tmp1), _mm_castps_pd(tmp3)));
let out3 = _mm_castpd_ps(_mm_unpackhi_pd(_mm_castps_pd(tmp1), _mm_castps_pd(tmp3)));
[
out0.simd_into(token),
out1.simd_into(token),
out2.simd_into(token),
out3.simd_into(token),
]
}
);
kernel(self, src)
}
#[inline(always)]
fn store_four_interleaved_f32x4(
self,
vectors: [f32x4<Self>; 4usize],
dest: &mut [f32; 16usize],
) -> () {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
vectors: [f32x4<Sse4_2>; 4usize],
dest: &mut [f32; 16usize],
) -> () {
let _ = token;
let v0: __m128 = vectors[0].into();
let v1: __m128 = vectors[1].into();
let v2: __m128 = vectors[2].into();
let v3: __m128 = vectors[3].into();
let tmp0 = _mm_unpacklo_ps(v0, v1);
let tmp1 = _mm_unpackhi_ps(v0, v1);
let tmp2 = _mm_unpacklo_ps(v2, v3);
let tmp3 = _mm_unpackhi_ps(v2, v3);
let out0 = _mm_castpd_ps(_mm_unpacklo_pd(_mm_castps_pd(tmp0), _mm_castps_pd(tmp2)));
let out1 = _mm_castpd_ps(_mm_unpackhi_pd(_mm_castps_pd(tmp0), _mm_castps_pd(tmp2)));
let out2 = _mm_castpd_ps(_mm_unpacklo_pd(_mm_castps_pd(tmp1), _mm_castps_pd(tmp3)));
let out3 = _mm_castpd_ps(_mm_unpackhi_pd(_mm_castps_pd(tmp1), _mm_castps_pd(tmp3)));
let (chunks, []) = dest.as_chunks_mut::<4usize>() else {
unreachable!()
};
crate::transmute::checked_transmute_store::<__m128, [f32; 4usize]>(
out0,
&mut chunks[0],
);
crate::transmute::checked_transmute_store::<__m128, [f32; 4usize]>(
out1,
&mut chunks[1],
);
crate::transmute::checked_transmute_store::<__m128, [f32; 4usize]>(
out2,
&mut chunks[2],
);
crate::transmute::checked_transmute_store::<__m128, [f32; 4usize]>(
out3,
&mut chunks[3],
);
}
);
kernel(self, vectors, dest);
}
#[inline(always)]
fn widen_f32x4(self, a: f32x4<Self>) -> (f64x2<Self>, f64x2<Self>) {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> (f64x2<Sse4_2>, f64x2<Sse4_2>) {
let raw = a.into();
(
_mm_cvtps_pd(raw).simd_into(token),
_mm_cvtps_pd(_mm_movehl_ps(raw, raw)).simd_into(token),
)
}
);
kernel(self, a)
}
#[inline(always)]
fn cvt_u32_f32x4(self, a: f32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> u32x4<Sse4_2> {
let mut converted = _mm_cvttps_epi32(a.into());
let in_range = _mm_cmplt_ps(a.into(), _mm_set1_ps(2147483648.0));
let all_in_range = _mm_movemask_ps(in_range) == 0b1111;
if !all_in_range {
let excess = _mm_sub_ps(a.into(), _mm_set1_ps(2147483648.0));
let excess_converted = _mm_cvttps_epi32(_mm_andnot_ps(in_range, excess));
converted = _mm_add_epi32(converted, excess_converted);
}
converted.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn cvt_u32_precise_f32x4(self, a: f32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> u32x4<Sse4_2> {
let a = _mm_max_ps(a.into(), _mm_setzero_ps());
let mut converted = _mm_cvttps_epi32(a);
let in_range = _mm_cmplt_ps(a, _mm_set1_ps(2147483648.0));
let all_in_range = _mm_movemask_ps(in_range) == 0b1111;
if !all_in_range {
let exceeds_unsigned_range =
_mm_castps_si128(_mm_cmplt_ps(_mm_set1_ps(4294967040.0), a));
let excess = _mm_sub_ps(a, _mm_set1_ps(2147483648.0));
let excess_converted = _mm_cvttps_epi32(_mm_andnot_ps(in_range, excess));
converted = _mm_add_epi32(converted, excess_converted);
converted = _mm_or_si128(converted, exceeds_unsigned_range);
}
converted.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn cvt_i32_f32x4(self, a: f32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> i32x4<Sse4_2> {
_mm_cvttps_epi32(a.into()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn cvt_i32_precise_f32x4(self, a: f32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f32x4<Sse4_2>) -> i32x4<Sse4_2> {
let a = a.into();
let converted = _mm_cvttps_epi32(a);
let positive_overflow =
_mm_castps_si128(_mm_cmple_ps(_mm_set1_ps(2147483648.0), a));
let converted = _mm_xor_si128(converted, positive_overflow);
let is_not_nan = _mm_castps_si128(_mm_cmpord_ps(a, a));
_mm_and_si128(converted, is_not_nan).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn abs_i8x16(self, a: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
_mm_abs_epi8(a.into()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn splat_i8x16(self, val: i8) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, val: i8) -> i8x16<Sse4_2> {
_mm_set1_epi8(val).simd_into(token)
}
);
kernel(self, val)
}
#[inline(always)]
fn reverse_i8x16(self, a: i8x16<Self>) -> i8x16<Self> {
let indices: u8x16<Self> =
[15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0].simd_into(self);
self.swizzle_dyn_i8x16(a, indices)
}
#[inline(always)]
fn slide_i8x16<const SHIFT: usize>(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
if SHIFT >= 16usize {
return b;
}
let result = dyn_alignr_128(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT,
);
Bytes::from_bytes(u8x16 {
val: crate::support::Aligned128(result),
simd: self,
})
}
#[inline(always)]
fn compress_i8x16(self, values: i8x16<Self>, mask: mask8x16<Self>) -> i8x16<Self> {
let mask = mask8x16 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x16(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_i8x16(
self,
values: i8x16<Self>,
mask: mask8x16<Self>,
merge: i8x16<Self>,
) -> i8x16<Self> {
let mask = mask8x16 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x16(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_i8x16(self, values: i8x16<Self>, mask: mask8x16<Self>) -> i8x16<Self> {
let mask = mask8x16 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x16(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_i8x16(
self,
values: i8x16<Self>,
mask: mask8x16<Self>,
merge: i8x16<Self>,
) -> i8x16<Self> {
let mask = mask8x16 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x16(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn count_ones_i8x16(self, a: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
let value = a.into();
let nibble_mask = _mm_set1_epi8(0x0f);
let lookup = _mm_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);
let low = _mm_and_si128(value, nibble_mask);
let high = _mm_and_si128(_mm_srli_epi16::<4>(value), nibble_mask);
let byte_counts = _mm_add_epi8(
_mm_shuffle_epi8(lookup, low),
_mm_shuffle_epi8(lookup, high),
);
byte_counts.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn count_zeros_i8x16(self, a: i8x16<Self>) -> i8x16<Self> {
self.count_ones_i8x16(self.not_i8x16(a))
}
#[inline(always)]
fn add_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
_mm_add_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_add_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
_mm_adds_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn sub_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
_mm_sub_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_sub_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
_mm_subs_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn mul_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
let a = a.into();
let b = b.into();
let low_mask = _mm_set1_epi16(0xFF);
let dst_even = _mm_mullo_epi16(a, b);
let dst_odd = _mm_maddubs_epi16(a, _mm_andnot_si128(low_mask, b));
_mm_or_si128(
_mm_slli_epi16(dst_odd, 8),
_mm_and_si128(dst_even, low_mask),
)
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn and_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
_mm_and_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn or_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
_mm_or_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn xor_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
_mm_xor_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn not_i8x16(self, a: i8x16<Self>) -> i8x16<Self> {
a ^ !0
}
#[inline(always)]
fn shl_i8x16(self, a: i8x16<Self>, shift: u32) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, shift: u32) -> i8x16<Sse4_2> {
let val = a.into();
let shift_count = _mm_cvtsi32_si128(shift.cast_signed());
let mask_byte = 0xff_u32.wrapping_shr(shift) as i8;
let byte_mask = _mm_set1_epi8(mask_byte);
_mm_sll_epi16(_mm_and_si128(val, byte_mask), shift_count).simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shlv_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
[
i8::wrapping_shl(a[0usize], b[0usize] as u32),
i8::wrapping_shl(a[1usize], b[1usize] as u32),
i8::wrapping_shl(a[2usize], b[2usize] as u32),
i8::wrapping_shl(a[3usize], b[3usize] as u32),
i8::wrapping_shl(a[4usize], b[4usize] as u32),
i8::wrapping_shl(a[5usize], b[5usize] as u32),
i8::wrapping_shl(a[6usize], b[6usize] as u32),
i8::wrapping_shl(a[7usize], b[7usize] as u32),
i8::wrapping_shl(a[8usize], b[8usize] as u32),
i8::wrapping_shl(a[9usize], b[9usize] as u32),
i8::wrapping_shl(a[10usize], b[10usize] as u32),
i8::wrapping_shl(a[11usize], b[11usize] as u32),
i8::wrapping_shl(a[12usize], b[12usize] as u32),
i8::wrapping_shl(a[13usize], b[13usize] as u32),
i8::wrapping_shl(a[14usize], b[14usize] as u32),
i8::wrapping_shl(a[15usize], b[15usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn shr_i8x16(self, a: i8x16<Self>, shift: u32) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, shift: u32) -> i8x16<Sse4_2> {
let val = a.into();
let shift_count = _mm_cvtsi32_si128(shift.cast_signed());
let mask_byte = 0xff_u32.wrapping_shr(shift) as i8;
let byte_mask = _mm_set1_epi8(mask_byte);
let shifted = _mm_srl_epi16(val, shift_count);
let result = {
let sign = _mm_cmpgt_epi8(_mm_setzero_si128(), val);
_mm_or_si128(
_mm_and_si128(shifted, byte_mask),
_mm_andnot_si128(byte_mask, sign),
)
};
result.simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shrv_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
[
i8::wrapping_shr(a[0usize], b[0usize] as u32),
i8::wrapping_shr(a[1usize], b[1usize] as u32),
i8::wrapping_shr(a[2usize], b[2usize] as u32),
i8::wrapping_shr(a[3usize], b[3usize] as u32),
i8::wrapping_shr(a[4usize], b[4usize] as u32),
i8::wrapping_shr(a[5usize], b[5usize] as u32),
i8::wrapping_shr(a[6usize], b[6usize] as u32),
i8::wrapping_shr(a[7usize], b[7usize] as u32),
i8::wrapping_shr(a[8usize], b[8usize] as u32),
i8::wrapping_shr(a[9usize], b[9usize] as u32),
i8::wrapping_shr(a[10usize], b[10usize] as u32),
i8::wrapping_shr(a[11usize], b[11usize] as u32),
i8::wrapping_shr(a[12usize], b[12usize] as u32),
i8::wrapping_shr(a[13usize], b[13usize] as u32),
i8::wrapping_shr(a[14usize], b[14usize] as u32),
i8::wrapping_shr(a[15usize], b[15usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn reduce_max_i8x16(self, a: i8x16<Self>) -> i8 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>) -> i8 {
let reduced: __m128i = a.into();
let shifted = _mm_srli_si128::<8>(reduced);
let reduced = { _mm_max_epi8(reduced, shifted) };
let shifted = _mm_srli_si128::<4>(reduced);
let reduced = { _mm_max_epi8(reduced, shifted) };
let shifted = _mm_srli_si128::<2>(reduced);
let reduced = { _mm_max_epi8(reduced, shifted) };
let shifted = _mm_srli_si128::<1>(reduced);
let reduced = { _mm_max_epi8(reduced, shifted) };
{
let lanes: [i8; 16usize] = crate::transmute::checked_transmute_copy(&reduced);
lanes[0]
}
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_min_i8x16(self, a: i8x16<Self>) -> i8 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>) -> i8 {
let reduced: __m128i = a.into();
let shifted = _mm_srli_si128::<8>(reduced);
let reduced = { _mm_min_epi8(reduced, shifted) };
let shifted = _mm_srli_si128::<4>(reduced);
let reduced = { _mm_min_epi8(reduced, shifted) };
let shifted = _mm_srli_si128::<2>(reduced);
let reduced = { _mm_min_epi8(reduced, shifted) };
let shifted = _mm_srli_si128::<1>(reduced);
let reduced = { _mm_min_epi8(reduced, shifted) };
{
let lanes: [i8; 16usize] = crate::transmute::checked_transmute_copy(&reduced);
lanes[0]
}
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_sum_i8x16(self, a: i8x16<Self>) -> i8 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>) -> i8 {
let sum: __m128i = a.into();
let sum = _mm_add_epi8(sum, _mm_srli_si128::<8>(sum));
let sum = _mm_add_epi8(sum, _mm_srli_si128::<4>(sum));
let sum = _mm_add_epi8(sum, _mm_srli_si128::<2>(sum));
let sum = _mm_add_epi8(sum, _mm_srli_si128::<1>(sum));
let lanes: [i8; 16usize] = crate::transmute::checked_transmute_copy(&sum);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_product_i8x16(self, a: i8x16<Self>) -> i8 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>) -> i8 {
let value: __m128i = a.into();
let product = _mm_mullo_epi16(
_mm_unpacklo_epi8(value, value),
_mm_unpackhi_epi8(value, value),
);
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b11_10_11_10>(product));
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b01_01_01_01>(product));
let product = _mm_mullo_epi16(product, _mm_srli_epi32::<16>(product));
let lanes: [i8; 16usize] = crate::transmute::checked_transmute_copy(&product);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn max_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
_mm_max_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn min_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
_mm_min_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_eq_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> mask8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> mask8x16<Sse4_2> {
_mm_cmpeq_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_lt_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> mask8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> mask8x16<Sse4_2> {
_mm_cmpgt_epi8(b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_le_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> mask8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> mask8x16<Sse4_2> {
_mm_cmpeq_epi8(_mm_min_epi8(a.into(), b.into()), a.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_low_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
_mm_unpacklo_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_high_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
_mm_unpackhi_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_low_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
let mask = _mm_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15);
let t1 = _mm_shuffle_epi8(a.into(), mask);
let t2 = _mm_shuffle_epi8(b.into(), mask);
_mm_unpacklo_epi64(t1, t2).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_high_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>, b: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
let mask = _mm_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15);
let t1 = _mm_shuffle_epi8(a.into(), mask);
let t2 = _mm_shuffle_epi8(b.into(), mask);
_mm_unpackhi_epi64(t1, t2).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn interleave_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> (i8x16<Self>, i8x16<Self>) {
(self.zip_low_i8x16(a, b), self.zip_high_i8x16(a, b))
}
#[inline(always)]
fn deinterleave_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> (i8x16<Self>, i8x16<Self>) {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: i8x16<Sse4_2>,
b: i8x16<Sse4_2>,
) -> (i8x16<Sse4_2>, i8x16<Sse4_2>) {
let mask = _mm_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15);
let a = _mm_shuffle_epi8(a.into(), mask);
let b = _mm_shuffle_epi8(b.into(), mask);
(
_mm_unpacklo_epi64(a, b).simd_into(token),
_mm_unpackhi_epi64(a, b).simd_into(token),
)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn select_i8x16(self, a: mask8x16<Self>, b: i8x16<Self>, c: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: mask8x16<Sse4_2>,
b: i8x16<Sse4_2>,
c: i8x16<Sse4_2>,
) -> i8x16<Sse4_2> {
_mm_blendv_epi8(c.into(), b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b, c)
}
#[inline(always)]
fn combine_i8x16(self, a: i8x16<Self>, b: i8x16<Self>) -> i8x32<Self> {
i8x32 {
val: crate::support::Aligned256([a.val.0, b.val.0]),
simd: self,
}
}
#[inline(always)]
fn neg_i8x16(self, a: i8x16<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>) -> i8x16<Sse4_2> {
_mm_sub_epi8(_mm_setzero_si128(), a.into()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn load_four_interleaved_i8x16(self, src: &[i8; 64usize]) -> [i8x16<Self>; 4usize] {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, src: &[i8; 64usize]) -> [i8x16<Sse4_2>; 4usize] {
let (chunks, []) = src.as_chunks::<16usize>() else {
unreachable!()
};
let v0: __m128i =
crate::transmute::checked_transmute_copy::<[i8; 16usize], __m128i>(&chunks[0]);
let v1: __m128i =
crate::transmute::checked_transmute_copy::<[i8; 16usize], __m128i>(&chunks[1]);
let v2: __m128i =
crate::transmute::checked_transmute_copy::<[i8; 16usize], __m128i>(&chunks[2]);
let v3: __m128i =
crate::transmute::checked_transmute_copy::<[i8; 16usize], __m128i>(&chunks[3]);
let mask = _mm_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15);
let v0 = _mm_shuffle_epi8(v0, mask);
let v1 = _mm_shuffle_epi8(v1, mask);
let v2 = _mm_shuffle_epi8(v2, mask);
let v3 = _mm_shuffle_epi8(v3, mask);
let tmp0 = _mm_unpacklo_epi32(v0, v1);
let tmp1 = _mm_unpackhi_epi32(v0, v1);
let tmp2 = _mm_unpacklo_epi32(v2, v3);
let tmp3 = _mm_unpackhi_epi32(v2, v3);
let out0 = _mm_unpacklo_epi64(tmp0, tmp2);
let out1 = _mm_unpackhi_epi64(tmp0, tmp2);
let out2 = _mm_unpacklo_epi64(tmp1, tmp3);
let out3 = _mm_unpackhi_epi64(tmp1, tmp3);
[
out0.simd_into(token),
out1.simd_into(token),
out2.simd_into(token),
out3.simd_into(token),
]
}
);
kernel(self, src)
}
#[inline(always)]
fn store_four_interleaved_i8x16(
self,
vectors: [i8x16<Self>; 4usize],
dest: &mut [i8; 64usize],
) -> () {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
vectors: [i8x16<Sse4_2>; 4usize],
dest: &mut [i8; 64usize],
) -> () {
let _ = token;
let v0: __m128i = vectors[0].into();
let v1: __m128i = vectors[1].into();
let v2: __m128i = vectors[2].into();
let v3: __m128i = vectors[3].into();
let tmp0 = _mm_unpacklo_epi32(v0, v1);
let tmp1 = _mm_unpackhi_epi32(v0, v1);
let tmp2 = _mm_unpacklo_epi32(v2, v3);
let tmp3 = _mm_unpackhi_epi32(v2, v3);
let out0 = _mm_unpacklo_epi64(tmp0, tmp2);
let out1 = _mm_unpackhi_epi64(tmp0, tmp2);
let out2 = _mm_unpacklo_epi64(tmp1, tmp3);
let out3 = _mm_unpackhi_epi64(tmp1, tmp3);
let mask = _mm_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15);
let out0 = _mm_shuffle_epi8(out0, mask);
let out1 = _mm_shuffle_epi8(out1, mask);
let out2 = _mm_shuffle_epi8(out2, mask);
let out3 = _mm_shuffle_epi8(out3, mask);
let (chunks, []) = dest.as_chunks_mut::<16usize>() else {
unreachable!()
};
crate::transmute::checked_transmute_store::<__m128i, [i8; 16usize]>(
out0,
&mut chunks[0],
);
crate::transmute::checked_transmute_store::<__m128i, [i8; 16usize]>(
out1,
&mut chunks[1],
);
crate::transmute::checked_transmute_store::<__m128i, [i8; 16usize]>(
out2,
&mut chunks[2],
);
crate::transmute::checked_transmute_store::<__m128i, [i8; 16usize]>(
out3,
&mut chunks[3],
);
}
);
kernel(self, vectors, dest);
}
#[inline(always)]
fn widen_i8x16(self, a: i8x16<Self>) -> (i16x8<Self>, i16x8<Self>) {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i8x16<Sse4_2>) -> (i16x8<Sse4_2>, i16x8<Sse4_2>) {
let raw = a.into();
(
_mm_cvtepi8_epi16(raw).simd_into(token),
_mm_cvtepi8_epi16(_mm_srli_si128::<8>(raw)).simd_into(token),
)
}
);
kernel(self, a)
}
#[inline(always)]
fn splat_u8x16(self, val: u8) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, val: u8) -> u8x16<Sse4_2> {
_mm_set1_epi8(val.cast_signed()).simd_into(token)
}
);
kernel(self, val)
}
#[inline(always)]
fn reverse_u8x16(self, a: u8x16<Self>) -> u8x16<Self> {
let indices: u8x16<Self> =
[15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0].simd_into(self);
self.swizzle_dyn_u8x16(a, indices)
}
#[inline(always)]
fn slide_u8x16<const SHIFT: usize>(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
if SHIFT >= 16usize {
return b;
}
let result = dyn_alignr_128(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT,
);
Bytes::from_bytes(u8x16 {
val: crate::support::Aligned128(result),
simd: self,
})
}
#[inline(always)]
fn swizzle_dyn_within_blocks_u8x16(self, a: u8x16<Self>, indices: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, indices: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
let result = _mm_shuffle_epi8(Bytes::to_bytes(a).val.0, indices.into());
Bytes::from_bytes(u8x16 {
val: crate::support::Aligned128(result),
simd: token,
})
}
);
kernel(self, a, indices)
}
#[inline(always)]
fn swizzle_dyn_u8x16(self, a: u8x16<Self>, indices: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, indices: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
let result = _mm_shuffle_epi8(Bytes::to_bytes(a).val.0, indices.into());
Bytes::from_bytes(u8x16 {
val: crate::support::Aligned128(result),
simd: token,
})
}
);
kernel(self, a, indices)
}
#[inline(always)]
fn swizzle_dyn_precise_u8x16(self, a: u8x16<Self>, indices: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, indices: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
let indices = indices.into();
let zeroing_indices = _mm_adds_epu8(indices, _mm_set1_epi8(0x70));
let result = _mm_shuffle_epi8(Bytes::to_bytes(a).val.0, zeroing_indices);
let result_bytes = u8x16 {
val: crate::support::Aligned128(result),
simd: token,
};
Bytes::from_bytes(result_bytes)
}
);
kernel(self, a, indices)
}
#[inline(always)]
fn concat_swizzle_dyn_u8x16(
self,
a: u8x16<Self>,
b: u8x16<Self>,
indices: u8x16<Self>,
) -> u8x16<Self> {
self.concat_swizzle_dyn_precise_u8x16(a, b, indices)
}
#[inline(always)]
fn concat_swizzle_dyn_precise_u8x16(
self,
a: u8x16<Self>,
b: u8x16<Self>,
indices: u8x16<Self>,
) -> u8x16<Self> {
let first_table = Bytes::to_bytes(a);
let second_table = Bytes::to_bytes(b);
let second_table_offset = self.splat_u8x16(16);
let from_first = self.swizzle_dyn_precise_u8x16(first_table, indices);
let from_second = self
.swizzle_dyn_precise_u8x16(second_table, self.sub_u8x16(indices, second_table_offset));
let result_bytes = self.or_u8x16(from_first, from_second);
Bytes::from_bytes(result_bytes)
}
#[inline(always)]
fn compress_u8x16(self, values: u8x16<Self>, mask: mask8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
values: u8x16<Sse4_2>,
mask: mask8x16<Sse4_2>,
) -> u8x16<Sse4_2> {
let mask_bits = token.to_bitmask_mask8x16(mask) as u16;
let low_mask = (usize::from(mask_bits)) & 0xff;
let high_mask = (usize::from(mask_bits)) >> 8;
let low_control = crate::support::COMPRESS_8_CONTROLS[low_mask];
let high_control = crate::support::COMPRESS_8_CONTROLS[high_mask];
let high_control = high_control | 0x0808_0808_0808_0808;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let low_count = low_mask.count_ones() as usize;
let compacted = _mm_shuffle_epi8(values.into(), control);
let splice = crate::support::COMPACT_8_SPLICE_CONTROLS[low_count];
let splice = _mm_set_epi64x((splice >> 64) as i64, splice as i64);
let compressed = _mm_shuffle_epi8(compacted, splice);
compressed.simd_into(token)
}
);
kernel(self, values, mask)
}
#[inline(always)]
fn compress_merge_u8x16(
self,
values: u8x16<Self>,
mask: mask8x16<Self>,
merge: u8x16<Self>,
) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
values: u8x16<Sse4_2>,
mask: mask8x16<Sse4_2>,
merge: u8x16<Sse4_2>,
) -> u8x16<Sse4_2> {
let mask_bits = token.to_bitmask_mask8x16(mask) as u16;
let low_mask = (usize::from(mask_bits)) & 0xff;
let high_mask = (usize::from(mask_bits)) >> 8;
let low_control = crate::support::COMPRESS_8_CONTROLS[low_mask];
let high_control = crate::support::COMPRESS_8_CONTROLS[high_mask];
let high_control = high_control | 0x0808_0808_0808_0808;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let low_count = low_mask.count_ones() as usize;
let compacted = _mm_shuffle_epi8(values.into(), control);
let splice = crate::support::COMPACT_8_SPLICE_CONTROLS[low_count];
let splice = _mm_set_epi64x((splice >> 64) as i64, splice as i64);
let compressed = _mm_shuffle_epi8(compacted, splice);
let count = mask_bits.count_ones() as usize;
let [prefix, _] = crate::transmute::checked_transmute_copy::<_, [__m128i; 2]>(
&crate::support::COMPACT_PREFIX_MASKS[count],
);
_mm_blendv_epi8(merge.into(), compressed, prefix).simd_into(token)
}
);
kernel(self, values, mask, merge)
}
#[inline(always)]
fn expand_u8x16(self, values: u8x16<Self>, mask: mask8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
values: u8x16<Sse4_2>,
mask: mask8x16<Sse4_2>,
) -> u8x16<Sse4_2> {
let mask_bits = token.to_bitmask_mask8x16(mask) as u16;
let low_mask = (usize::from(mask_bits)) & 0xff;
let high_mask = (usize::from(mask_bits)) >> 8;
let low_count = low_mask.count_ones() as u64;
let low_control = crate::support::EXPAND_8_CONTROLS[low_mask];
let high_control = crate::support::EXPAND_8_CONTROLS[high_mask];
let high_base = low_count * 0x0101_0101_0101_0101;
let high_control = high_control + high_base;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let expanded = _mm_shuffle_epi8(values.into(), control);
expanded.simd_into(token)
}
);
kernel(self, values, mask)
}
#[inline(always)]
fn expand_merge_u8x16(
self,
values: u8x16<Self>,
mask: mask8x16<Self>,
merge: u8x16<Self>,
) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
values: u8x16<Sse4_2>,
mask: mask8x16<Sse4_2>,
merge: u8x16<Sse4_2>,
) -> u8x16<Sse4_2> {
let mask_bits = token.to_bitmask_mask8x16(mask) as u16;
let low_mask = (usize::from(mask_bits)) & 0xff;
let high_mask = (usize::from(mask_bits)) >> 8;
let low_count = low_mask.count_ones() as u64;
let low_control = crate::support::EXPAND_8_CONTROLS[low_mask];
let high_control = crate::support::EXPAND_8_CONTROLS[high_mask];
let high_base = low_count * 0x0101_0101_0101_0101;
let high_control = high_control + high_base;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let expanded = _mm_shuffle_epi8(values.into(), control);
_mm_blendv_epi8(merge.into(), expanded, mask.into()).simd_into(token)
}
);
kernel(self, values, mask, merge)
}
#[inline(always)]
fn count_ones_u8x16(self, a: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
let value = a.into();
let nibble_mask = _mm_set1_epi8(0x0f);
let lookup = _mm_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);
let low = _mm_and_si128(value, nibble_mask);
let high = _mm_and_si128(_mm_srli_epi16::<4>(value), nibble_mask);
let byte_counts = _mm_add_epi8(
_mm_shuffle_epi8(lookup, low),
_mm_shuffle_epi8(lookup, high),
);
byte_counts.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn count_zeros_u8x16(self, a: u8x16<Self>) -> u8x16<Self> {
self.count_ones_u8x16(self.not_u8x16(a))
}
#[inline(always)]
fn add_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
_mm_add_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_add_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
_mm_adds_epu8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn sub_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
_mm_sub_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_sub_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
_mm_subs_epu8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn mul_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
let a = a.into();
let b = b.into();
let low_mask = _mm_set1_epi16(0xFF);
let dst_even = _mm_mullo_epi16(a, b);
let dst_odd = _mm_maddubs_epi16(a, _mm_andnot_si128(low_mask, b));
_mm_or_si128(
_mm_slli_epi16(dst_odd, 8),
_mm_and_si128(dst_even, low_mask),
)
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn and_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
_mm_and_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn or_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
_mm_or_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn xor_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
_mm_xor_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn not_u8x16(self, a: u8x16<Self>) -> u8x16<Self> {
a ^ !0
}
#[inline(always)]
fn shl_u8x16(self, a: u8x16<Self>, shift: u32) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, shift: u32) -> u8x16<Sse4_2> {
let val = a.into();
let shift_count = _mm_cvtsi32_si128(shift.cast_signed());
let mask_byte = 0xff_u32.wrapping_shr(shift) as i8;
let byte_mask = _mm_set1_epi8(mask_byte);
_mm_sll_epi16(_mm_and_si128(val, byte_mask), shift_count).simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shlv_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
[
u8::wrapping_shl(a[0usize], b[0usize] as u32),
u8::wrapping_shl(a[1usize], b[1usize] as u32),
u8::wrapping_shl(a[2usize], b[2usize] as u32),
u8::wrapping_shl(a[3usize], b[3usize] as u32),
u8::wrapping_shl(a[4usize], b[4usize] as u32),
u8::wrapping_shl(a[5usize], b[5usize] as u32),
u8::wrapping_shl(a[6usize], b[6usize] as u32),
u8::wrapping_shl(a[7usize], b[7usize] as u32),
u8::wrapping_shl(a[8usize], b[8usize] as u32),
u8::wrapping_shl(a[9usize], b[9usize] as u32),
u8::wrapping_shl(a[10usize], b[10usize] as u32),
u8::wrapping_shl(a[11usize], b[11usize] as u32),
u8::wrapping_shl(a[12usize], b[12usize] as u32),
u8::wrapping_shl(a[13usize], b[13usize] as u32),
u8::wrapping_shl(a[14usize], b[14usize] as u32),
u8::wrapping_shl(a[15usize], b[15usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn shr_u8x16(self, a: u8x16<Self>, shift: u32) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, shift: u32) -> u8x16<Sse4_2> {
let val = a.into();
let shift_count = _mm_cvtsi32_si128(shift.cast_signed());
let mask_byte = 0xff_u32.wrapping_shr(shift) as i8;
let byte_mask = _mm_set1_epi8(mask_byte);
let shifted = _mm_srl_epi16(val, shift_count);
let result = { _mm_and_si128(shifted, byte_mask) };
result.simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shrv_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
[
u8::wrapping_shr(a[0usize], b[0usize] as u32),
u8::wrapping_shr(a[1usize], b[1usize] as u32),
u8::wrapping_shr(a[2usize], b[2usize] as u32),
u8::wrapping_shr(a[3usize], b[3usize] as u32),
u8::wrapping_shr(a[4usize], b[4usize] as u32),
u8::wrapping_shr(a[5usize], b[5usize] as u32),
u8::wrapping_shr(a[6usize], b[6usize] as u32),
u8::wrapping_shr(a[7usize], b[7usize] as u32),
u8::wrapping_shr(a[8usize], b[8usize] as u32),
u8::wrapping_shr(a[9usize], b[9usize] as u32),
u8::wrapping_shr(a[10usize], b[10usize] as u32),
u8::wrapping_shr(a[11usize], b[11usize] as u32),
u8::wrapping_shr(a[12usize], b[12usize] as u32),
u8::wrapping_shr(a[13usize], b[13usize] as u32),
u8::wrapping_shr(a[14usize], b[14usize] as u32),
u8::wrapping_shr(a[15usize], b[15usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn reduce_max_u8x16(self, a: u8x16<Self>) -> u8 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>) -> u8 {
let reduced: __m128i = a.into();
let shifted = _mm_srli_si128::<8>(reduced);
let reduced = { _mm_max_epu8(reduced, shifted) };
let shifted = _mm_srli_si128::<4>(reduced);
let reduced = { _mm_max_epu8(reduced, shifted) };
let shifted = _mm_srli_si128::<2>(reduced);
let reduced = { _mm_max_epu8(reduced, shifted) };
let shifted = _mm_srli_si128::<1>(reduced);
let reduced = { _mm_max_epu8(reduced, shifted) };
{
let lanes: [u8; 16usize] = crate::transmute::checked_transmute_copy(&reduced);
lanes[0]
}
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_min_u8x16(self, a: u8x16<Self>) -> u8 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>) -> u8 {
let reduced: __m128i = a.into();
let shifted = _mm_srli_si128::<8>(reduced);
let reduced = { _mm_min_epu8(reduced, shifted) };
let shifted = _mm_srli_si128::<4>(reduced);
let reduced = { _mm_min_epu8(reduced, shifted) };
let shifted = _mm_srli_si128::<2>(reduced);
let reduced = { _mm_min_epu8(reduced, shifted) };
let shifted = _mm_srli_si128::<1>(reduced);
let reduced = { _mm_min_epu8(reduced, shifted) };
{
let lanes: [u8; 16usize] = crate::transmute::checked_transmute_copy(&reduced);
lanes[0]
}
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_sum_u8x16(self, a: u8x16<Self>) -> u8 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>) -> u8 {
let sum: __m128i = a.into();
let sum = _mm_add_epi8(sum, _mm_srli_si128::<8>(sum));
let sum = _mm_add_epi8(sum, _mm_srli_si128::<4>(sum));
let sum = _mm_add_epi8(sum, _mm_srli_si128::<2>(sum));
let sum = _mm_add_epi8(sum, _mm_srli_si128::<1>(sum));
let lanes: [u8; 16usize] = crate::transmute::checked_transmute_copy(&sum);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_product_u8x16(self, a: u8x16<Self>) -> u8 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>) -> u8 {
let value: __m128i = a.into();
let product = _mm_mullo_epi16(
_mm_unpacklo_epi8(value, value),
_mm_unpackhi_epi8(value, value),
);
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b11_10_11_10>(product));
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b01_01_01_01>(product));
let product = _mm_mullo_epi16(product, _mm_srli_epi32::<16>(product));
let lanes: [u8; 16usize] = crate::transmute::checked_transmute_copy(&product);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn max_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
_mm_max_epu8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn min_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
_mm_min_epu8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_eq_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> mask8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> mask8x16<Sse4_2> {
_mm_cmpeq_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_lt_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> mask8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> mask8x16<Sse4_2> {
{
let sign_bit = _mm_set1_epi8(0x80u8.cast_signed());
let lhs_signed = _mm_xor_si128(b.into(), sign_bit);
let rhs_signed = _mm_xor_si128(a.into(), sign_bit);
_mm_cmpgt_epi8(lhs_signed, rhs_signed)
}
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_le_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> mask8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> mask8x16<Sse4_2> {
_mm_cmpeq_epi8(_mm_min_epu8(a.into(), b.into()), a.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_low_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
_mm_unpacklo_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_high_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
_mm_unpackhi_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_low_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
let mask = _mm_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15);
let t1 = _mm_shuffle_epi8(a.into(), mask);
let t2 = _mm_shuffle_epi8(b.into(), mask);
_mm_unpacklo_epi64(t1, t2).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_high_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>, b: u8x16<Sse4_2>) -> u8x16<Sse4_2> {
let mask = _mm_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15);
let t1 = _mm_shuffle_epi8(a.into(), mask);
let t2 = _mm_shuffle_epi8(b.into(), mask);
_mm_unpackhi_epi64(t1, t2).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn interleave_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> (u8x16<Self>, u8x16<Self>) {
(self.zip_low_u8x16(a, b), self.zip_high_u8x16(a, b))
}
#[inline(always)]
fn deinterleave_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> (u8x16<Self>, u8x16<Self>) {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: u8x16<Sse4_2>,
b: u8x16<Sse4_2>,
) -> (u8x16<Sse4_2>, u8x16<Sse4_2>) {
let mask = _mm_setr_epi8(0, 2, 4, 6, 8, 10, 12, 14, 1, 3, 5, 7, 9, 11, 13, 15);
let a = _mm_shuffle_epi8(a.into(), mask);
let b = _mm_shuffle_epi8(b.into(), mask);
(
_mm_unpacklo_epi64(a, b).simd_into(token),
_mm_unpackhi_epi64(a, b).simd_into(token),
)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn select_u8x16(self, a: mask8x16<Self>, b: u8x16<Self>, c: u8x16<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: mask8x16<Sse4_2>,
b: u8x16<Sse4_2>,
c: u8x16<Sse4_2>,
) -> u8x16<Sse4_2> {
_mm_blendv_epi8(c.into(), b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b, c)
}
#[inline(always)]
fn combine_u8x16(self, a: u8x16<Self>, b: u8x16<Self>) -> u8x32<Self> {
u8x32 {
val: crate::support::Aligned256([a.val.0, b.val.0]),
simd: self,
}
}
#[inline(always)]
fn load_four_interleaved_u8x16(self, src: &[u8; 64usize]) -> [u8x16<Self>; 4usize] {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, src: &[u8; 64usize]) -> [u8x16<Sse4_2>; 4usize] {
let (chunks, []) = src.as_chunks::<16usize>() else {
unreachable!()
};
let v0: __m128i =
crate::transmute::checked_transmute_copy::<[u8; 16usize], __m128i>(&chunks[0]);
let v1: __m128i =
crate::transmute::checked_transmute_copy::<[u8; 16usize], __m128i>(&chunks[1]);
let v2: __m128i =
crate::transmute::checked_transmute_copy::<[u8; 16usize], __m128i>(&chunks[2]);
let v3: __m128i =
crate::transmute::checked_transmute_copy::<[u8; 16usize], __m128i>(&chunks[3]);
let mask = _mm_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15);
let v0 = _mm_shuffle_epi8(v0, mask);
let v1 = _mm_shuffle_epi8(v1, mask);
let v2 = _mm_shuffle_epi8(v2, mask);
let v3 = _mm_shuffle_epi8(v3, mask);
let tmp0 = _mm_unpacklo_epi32(v0, v1);
let tmp1 = _mm_unpackhi_epi32(v0, v1);
let tmp2 = _mm_unpacklo_epi32(v2, v3);
let tmp3 = _mm_unpackhi_epi32(v2, v3);
let out0 = _mm_unpacklo_epi64(tmp0, tmp2);
let out1 = _mm_unpackhi_epi64(tmp0, tmp2);
let out2 = _mm_unpacklo_epi64(tmp1, tmp3);
let out3 = _mm_unpackhi_epi64(tmp1, tmp3);
[
out0.simd_into(token),
out1.simd_into(token),
out2.simd_into(token),
out3.simd_into(token),
]
}
);
kernel(self, src)
}
#[inline(always)]
fn store_four_interleaved_u8x16(
self,
vectors: [u8x16<Self>; 4usize],
dest: &mut [u8; 64usize],
) -> () {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
vectors: [u8x16<Sse4_2>; 4usize],
dest: &mut [u8; 64usize],
) -> () {
let _ = token;
let v0: __m128i = vectors[0].into();
let v1: __m128i = vectors[1].into();
let v2: __m128i = vectors[2].into();
let v3: __m128i = vectors[3].into();
let tmp0 = _mm_unpacklo_epi32(v0, v1);
let tmp1 = _mm_unpackhi_epi32(v0, v1);
let tmp2 = _mm_unpacklo_epi32(v2, v3);
let tmp3 = _mm_unpackhi_epi32(v2, v3);
let out0 = _mm_unpacklo_epi64(tmp0, tmp2);
let out1 = _mm_unpackhi_epi64(tmp0, tmp2);
let out2 = _mm_unpacklo_epi64(tmp1, tmp3);
let out3 = _mm_unpackhi_epi64(tmp1, tmp3);
let mask = _mm_setr_epi8(0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15);
let out0 = _mm_shuffle_epi8(out0, mask);
let out1 = _mm_shuffle_epi8(out1, mask);
let out2 = _mm_shuffle_epi8(out2, mask);
let out3 = _mm_shuffle_epi8(out3, mask);
let (chunks, []) = dest.as_chunks_mut::<16usize>() else {
unreachable!()
};
crate::transmute::checked_transmute_store::<__m128i, [u8; 16usize]>(
out0,
&mut chunks[0],
);
crate::transmute::checked_transmute_store::<__m128i, [u8; 16usize]>(
out1,
&mut chunks[1],
);
crate::transmute::checked_transmute_store::<__m128i, [u8; 16usize]>(
out2,
&mut chunks[2],
);
crate::transmute::checked_transmute_store::<__m128i, [u8; 16usize]>(
out3,
&mut chunks[3],
);
}
);
kernel(self, vectors, dest);
}
#[inline(always)]
fn widen_u8x16(self, a: u8x16<Self>) -> (u16x8<Self>, u16x8<Self>) {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x16<Sse4_2>) -> (u16x8<Sse4_2>, u16x8<Sse4_2>) {
let raw = a.into();
(
_mm_cvtepu8_epi16(raw).simd_into(token),
_mm_cvtepu8_epi16(_mm_srli_si128::<8>(raw)).simd_into(token),
)
}
);
kernel(self, a)
}
#[inline(always)]
fn splat_mask8x16(self, val: bool) -> mask8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, val: bool) -> mask8x16<Sse4_2> {
let val: i8 = if val { !0 } else { 0 };
_mm_set1_epi8(val).simd_into(token)
}
);
kernel(self, val)
}
#[inline(always)]
fn from_bitmask_mask8x16(self, bits: u64) -> mask8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, bits: u64) -> mask8x16<Sse4_2> {
{
let bit_bytes = _mm_cvtsi32_si128(bits as i32);
let bit_bytes = _mm_shuffle_epi8(
bit_bytes,
_mm_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1),
);
let bit_mask =
_mm_setr_epi8(1, 2, 4, 8, 16, 32, 64, -128, 1, 2, 4, 8, 16, 32, 64, -128);
_mm_cmpeq_epi8(_mm_and_si128(bit_bytes, bit_mask), bit_mask)
}
.simd_into(token)
}
);
kernel(self, bits)
}
#[inline(always)]
fn to_bitmask_mask8x16(self, a: mask8x16<Self>) -> u64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask8x16<Sse4_2>) -> u64 {
_mm_movemask_epi8(a.into()) as u32 as u64
}
);
kernel(self, a)
}
#[inline(always)]
fn from_vector_mask8x16(self, vector: i8x16<Self>) -> mask8x16<Self> {
mask8x16 {
val: vector.val,
simd: self,
}
}
#[inline(always)]
fn to_vector_mask8x16(self, a: mask8x16<Self>) -> i8x16<Self> {
i8x16 {
val: a.val,
simd: self,
}
}
#[inline(always)]
fn set_mask8x16(self, a: &mut mask8x16<Self>, index: usize, value: bool) -> () {
assert!(
index < 16usize,
"mask lane index {index} is out of bounds for {} lanes",
16usize
);
let mut lanes: [i8; 16usize] = (*a).into();
lanes[index] = if value { !0 } else { 0 };
*a = lanes.simd_into(self);
}
#[inline(always)]
fn rotate_elements_left_mask8x16<const OFFSET: usize>(
self,
a: mask8x16<Self>,
) -> mask8x16<Self> {
let int = i8x16 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_left::<OFFSET>();
mask8x16 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn rotate_elements_right_mask8x16<const OFFSET: usize>(
self,
a: mask8x16<Self>,
) -> mask8x16<Self> {
let int = i8x16 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_right::<OFFSET>();
mask8x16 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn and_mask8x16(self, a: mask8x16<Self>, b: mask8x16<Self>) -> mask8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask8x16<Sse4_2>, b: mask8x16<Sse4_2>) -> mask8x16<Sse4_2> {
_mm_and_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn or_mask8x16(self, a: mask8x16<Self>, b: mask8x16<Self>) -> mask8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask8x16<Sse4_2>, b: mask8x16<Sse4_2>) -> mask8x16<Sse4_2> {
_mm_or_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn xor_mask8x16(self, a: mask8x16<Self>, b: mask8x16<Self>) -> mask8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask8x16<Sse4_2>, b: mask8x16<Sse4_2>) -> mask8x16<Sse4_2> {
_mm_xor_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn not_mask8x16(self, a: mask8x16<Self>) -> mask8x16<Self> {
self.xor_mask8x16(a, self.splat_mask8x16(true))
}
#[inline(always)]
fn reverse_mask8x16(self, a: mask8x16<Self>) -> mask8x16<Self> {
let lanes = i8x16 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let reversed = self.reverse_i8x16(lanes);
mask8x16 {
val: crate::transmute::checked_transmute_copy(&reversed.val),
simd: self,
}
}
#[inline(always)]
fn select_mask8x16(
self,
a: mask8x16<Self>,
b: mask8x16<Self>,
c: mask8x16<Self>,
) -> mask8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: mask8x16<Sse4_2>,
b: mask8x16<Sse4_2>,
c: mask8x16<Sse4_2>,
) -> mask8x16<Sse4_2> {
_mm_blendv_epi8(c.into(), b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b, c)
}
#[inline(always)]
fn simd_eq_mask8x16(self, a: mask8x16<Self>, b: mask8x16<Self>) -> mask8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask8x16<Sse4_2>, b: mask8x16<Sse4_2>) -> mask8x16<Sse4_2> {
_mm_cmpeq_epi8(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn any_true_mask8x16(self, a: mask8x16<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask8x16<Sse4_2>) -> bool {
_mm_movemask_epi8(a.into()) as u32 != 0
}
);
kernel(self, a)
}
#[inline(always)]
fn all_true_mask8x16(self, a: mask8x16<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask8x16<Sse4_2>) -> bool {
_mm_movemask_epi8(a.into()) as u32 == 0xffff
}
);
kernel(self, a)
}
#[inline(always)]
fn any_false_mask8x16(self, a: mask8x16<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask8x16<Sse4_2>) -> bool {
_mm_movemask_epi8(a.into()) as u32 != 0xffff
}
);
kernel(self, a)
}
#[inline(always)]
fn all_false_mask8x16(self, a: mask8x16<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask8x16<Sse4_2>) -> bool {
_mm_movemask_epi8(a.into()) as u32 == 0
}
);
kernel(self, a)
}
#[inline(always)]
fn combine_mask8x16(self, a: mask8x16<Self>, b: mask8x16<Self>) -> mask8x32<Self> {
mask8x32 {
val: crate::support::Aligned256([a.val.0, b.val.0]),
simd: self,
}
}
#[inline(always)]
fn widen_mask8x16(self, a: mask8x16<Self>) -> (mask16x8<Self>, mask16x8<Self>) {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask8x16<Sse4_2>) -> (mask16x8<Sse4_2>, mask16x8<Sse4_2>) {
let raw = a.into();
(
_mm_unpacklo_epi8(raw, raw).simd_into(token),
_mm_unpackhi_epi8(raw, raw).simd_into(token),
)
}
);
kernel(self, a)
}
#[inline(always)]
fn abs_i16x8(self, a: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
_mm_abs_epi16(a.into()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn splat_i16x8(self, val: i16) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, val: i16) -> i16x8<Sse4_2> {
_mm_set1_epi16(val).simd_into(token)
}
);
kernel(self, val)
}
#[inline(always)]
fn reverse_i16x8(self, a: i16x8<Self>) -> i16x8<Self> {
let indices: u8x16<Self> =
[14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1].simd_into(self);
self.swizzle_dyn_i16x8(a, indices)
}
#[inline(always)]
fn slide_i16x8<const SHIFT: usize>(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
if SHIFT >= 8usize {
return b;
}
let result = dyn_alignr_128(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 2usize,
);
Bytes::from_bytes(u8x16 {
val: crate::support::Aligned128(result),
simd: self,
})
}
#[inline(always)]
fn compress_i16x8(self, values: i16x8<Self>, mask: mask16x8<Self>) -> i16x8<Self> {
let bits = self.to_bitmask_mask16x8(mask) as usize & 255;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_16[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn compress_merge_i16x8(
self,
values: i16x8<Self>,
mask: mask16x8<Self>,
merge: i16x8<Self>,
) -> i16x8<Self> {
let bits = self.to_bitmask_mask16x8(mask) as usize & 255;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_16[bits].0);
let compressed = self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control);
let inactive = self.simd_lt_i8x16(i8x16::from_bytes(control), i8x16::splat(self, 0));
Bytes::from_bytes(self.select_u8x16(inactive, Bytes::to_bytes(merge), compressed))
}
#[inline(always)]
fn expand_i16x8(self, values: i16x8<Self>, mask: mask16x8<Self>) -> i16x8<Self> {
let bits = self.to_bitmask_mask16x8(mask) as usize & 255;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_16[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn expand_merge_i16x8(
self,
values: i16x8<Self>,
mask: mask16x8<Self>,
merge: i16x8<Self>,
) -> i16x8<Self> {
let bits = self.to_bitmask_mask16x8(mask) as usize & 255;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_16[bits].0);
let expanded = Bytes::from_bytes(self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control));
self.select_i16x8(mask, expanded, merge)
}
#[inline(always)]
fn count_ones_i16x8(self, a: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
let value = a.into();
let nibble_mask = _mm_set1_epi8(0x0f);
let lookup = _mm_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);
let low = _mm_and_si128(value, nibble_mask);
let high = _mm_and_si128(_mm_srli_epi16::<4>(value), nibble_mask);
let byte_counts = _mm_add_epi8(
_mm_shuffle_epi8(lookup, low),
_mm_shuffle_epi8(lookup, high),
);
_mm_maddubs_epi16(byte_counts, _mm_set1_epi8(1)).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn count_zeros_i16x8(self, a: i16x8<Self>) -> i16x8<Self> {
self.count_ones_i16x8(self.not_i16x8(a))
}
#[inline(always)]
fn add_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
_mm_add_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_add_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
_mm_adds_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn sub_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
_mm_sub_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_sub_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
_mm_subs_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn mul_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
_mm_mullo_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn and_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
_mm_and_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn or_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
_mm_or_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn xor_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
_mm_xor_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn not_i16x8(self, a: i16x8<Self>) -> i16x8<Self> {
a ^ !0
}
#[inline(always)]
fn shl_i16x8(self, a: i16x8<Self>, shift: u32) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, shift: u32) -> i16x8<Sse4_2> {
_mm_sll_epi16(a.into(), _mm_cvtsi32_si128(shift.cast_signed())).simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shlv_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
[
i16::wrapping_shl(a[0usize], b[0usize] as u32),
i16::wrapping_shl(a[1usize], b[1usize] as u32),
i16::wrapping_shl(a[2usize], b[2usize] as u32),
i16::wrapping_shl(a[3usize], b[3usize] as u32),
i16::wrapping_shl(a[4usize], b[4usize] as u32),
i16::wrapping_shl(a[5usize], b[5usize] as u32),
i16::wrapping_shl(a[6usize], b[6usize] as u32),
i16::wrapping_shl(a[7usize], b[7usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn shr_i16x8(self, a: i16x8<Self>, shift: u32) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, shift: u32) -> i16x8<Sse4_2> {
_mm_sra_epi16(a.into(), _mm_cvtsi32_si128(shift.cast_signed())).simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shrv_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
[
i16::wrapping_shr(a[0usize], b[0usize] as u32),
i16::wrapping_shr(a[1usize], b[1usize] as u32),
i16::wrapping_shr(a[2usize], b[2usize] as u32),
i16::wrapping_shr(a[3usize], b[3usize] as u32),
i16::wrapping_shr(a[4usize], b[4usize] as u32),
i16::wrapping_shr(a[5usize], b[5usize] as u32),
i16::wrapping_shr(a[6usize], b[6usize] as u32),
i16::wrapping_shr(a[7usize], b[7usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn reduce_max_i16x8(self, a: i16x8<Self>) -> i16 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>) -> i16 {
let reduced: __m128i = a.into();
let shifted = _mm_srli_si128::<8>(reduced);
let reduced = { _mm_max_epi16(reduced, shifted) };
let shifted = _mm_srli_si128::<4>(reduced);
let reduced = { _mm_max_epi16(reduced, shifted) };
let shifted = _mm_srli_si128::<2>(reduced);
let reduced = { _mm_max_epi16(reduced, shifted) };
{
let lanes: [i16; 8usize] = crate::transmute::checked_transmute_copy(&reduced);
lanes[0]
}
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_min_i16x8(self, a: i16x8<Self>) -> i16 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>) -> i16 {
let reduced: __m128i = a.into();
let shifted = _mm_srli_si128::<8>(reduced);
let reduced = { _mm_min_epi16(reduced, shifted) };
let shifted = _mm_srli_si128::<4>(reduced);
let reduced = { _mm_min_epi16(reduced, shifted) };
let shifted = _mm_srli_si128::<2>(reduced);
let reduced = { _mm_min_epi16(reduced, shifted) };
{
let lanes: [i16; 8usize] = crate::transmute::checked_transmute_copy(&reduced);
lanes[0]
}
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_sum_i16x8(self, a: i16x8<Self>) -> i16 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>) -> i16 {
let sum: __m128i = a.into();
let sum = _mm_add_epi16(sum, _mm_srli_si128::<8>(sum));
let sum = _mm_add_epi16(sum, _mm_srli_si128::<4>(sum));
let sum = _mm_add_epi16(sum, _mm_srli_si128::<2>(sum));
let lanes: [i16; 8usize] = crate::transmute::checked_transmute_copy(&sum);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_product_i16x8(self, a: i16x8<Self>) -> i16 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>) -> i16 {
let product: __m128i = a.into();
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b11_10_11_10>(product));
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b01_01_01_01>(product));
let product = _mm_mullo_epi16(product, _mm_srli_epi32::<16>(product));
let lanes: [i16; 8usize] = crate::transmute::checked_transmute_copy(&product);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn max_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
_mm_max_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn min_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
_mm_min_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_eq_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> mask16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> mask16x8<Sse4_2> {
_mm_cmpeq_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_lt_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> mask16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> mask16x8<Sse4_2> {
_mm_cmpgt_epi16(b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_le_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> mask16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> mask16x8<Sse4_2> {
_mm_cmpeq_epi16(_mm_min_epi16(a.into(), b.into()), a.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_low_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
_mm_unpacklo_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_high_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
_mm_unpackhi_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_low_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
let mask = _mm_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15);
let t1 = _mm_shuffle_epi8(a.into(), mask);
let t2 = _mm_shuffle_epi8(b.into(), mask);
_mm_unpacklo_epi64(t1, t2).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_high_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
let mask = _mm_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15);
let t1 = _mm_shuffle_epi8(a.into(), mask);
let t2 = _mm_shuffle_epi8(b.into(), mask);
_mm_unpackhi_epi64(t1, t2).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn interleave_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> (i16x8<Self>, i16x8<Self>) {
(self.zip_low_i16x8(a, b), self.zip_high_i16x8(a, b))
}
#[inline(always)]
fn deinterleave_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> (i16x8<Self>, i16x8<Self>) {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: i16x8<Sse4_2>,
b: i16x8<Sse4_2>,
) -> (i16x8<Sse4_2>, i16x8<Sse4_2>) {
let mask = _mm_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15);
let a = _mm_shuffle_epi8(a.into(), mask);
let b = _mm_shuffle_epi8(b.into(), mask);
(
_mm_unpacklo_epi64(a, b).simd_into(token),
_mm_unpackhi_epi64(a, b).simd_into(token),
)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn select_i16x8(self, a: mask16x8<Self>, b: i16x8<Self>, c: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: mask16x8<Sse4_2>,
b: i16x8<Sse4_2>,
c: i16x8<Sse4_2>,
) -> i16x8<Sse4_2> {
_mm_blendv_epi8(c.into(), b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b, c)
}
#[inline(always)]
fn combine_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i16x16<Self> {
i16x16 {
val: crate::support::Aligned256([a.val.0, b.val.0]),
simd: self,
}
}
#[inline(always)]
fn neg_i16x8(self, a: i16x8<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>) -> i16x8<Sse4_2> {
_mm_sub_epi16(_mm_setzero_si128(), a.into()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn load_four_interleaved_i16x8(self, src: &[i16; 32usize]) -> [i16x8<Self>; 4usize] {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, src: &[i16; 32usize]) -> [i16x8<Sse4_2>; 4usize] {
let (chunks, []) = src.as_chunks::<8usize>() else {
unreachable!()
};
let v0: __m128i =
crate::transmute::checked_transmute_copy::<[i16; 8usize], __m128i>(&chunks[0]);
let v1: __m128i =
crate::transmute::checked_transmute_copy::<[i16; 8usize], __m128i>(&chunks[1]);
let v2: __m128i =
crate::transmute::checked_transmute_copy::<[i16; 8usize], __m128i>(&chunks[2]);
let v3: __m128i =
crate::transmute::checked_transmute_copy::<[i16; 8usize], __m128i>(&chunks[3]);
let mask = _mm_setr_epi8(0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15);
let v0 = _mm_shuffle_epi8(v0, mask);
let v1 = _mm_shuffle_epi8(v1, mask);
let v2 = _mm_shuffle_epi8(v2, mask);
let v3 = _mm_shuffle_epi8(v3, mask);
let tmp0 = _mm_unpacklo_epi32(v0, v1);
let tmp1 = _mm_unpackhi_epi32(v0, v1);
let tmp2 = _mm_unpacklo_epi32(v2, v3);
let tmp3 = _mm_unpackhi_epi32(v2, v3);
let out0 = _mm_unpacklo_epi64(tmp0, tmp2);
let out1 = _mm_unpackhi_epi64(tmp0, tmp2);
let out2 = _mm_unpacklo_epi64(tmp1, tmp3);
let out3 = _mm_unpackhi_epi64(tmp1, tmp3);
[
out0.simd_into(token),
out1.simd_into(token),
out2.simd_into(token),
out3.simd_into(token),
]
}
);
kernel(self, src)
}
#[inline(always)]
fn store_four_interleaved_i16x8(
self,
vectors: [i16x8<Self>; 4usize],
dest: &mut [i16; 32usize],
) -> () {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
vectors: [i16x8<Sse4_2>; 4usize],
dest: &mut [i16; 32usize],
) -> () {
let _ = token;
let v0: __m128i = vectors[0].into();
let v1: __m128i = vectors[1].into();
let v2: __m128i = vectors[2].into();
let v3: __m128i = vectors[3].into();
let tmp0 = _mm_unpacklo_epi32(v0, v1);
let tmp1 = _mm_unpackhi_epi32(v0, v1);
let tmp2 = _mm_unpacklo_epi32(v2, v3);
let tmp3 = _mm_unpackhi_epi32(v2, v3);
let out0 = _mm_unpacklo_epi64(tmp0, tmp2);
let out1 = _mm_unpackhi_epi64(tmp0, tmp2);
let out2 = _mm_unpacklo_epi64(tmp1, tmp3);
let out3 = _mm_unpackhi_epi64(tmp1, tmp3);
let mask = _mm_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15);
let out0 = _mm_shuffle_epi8(out0, mask);
let out1 = _mm_shuffle_epi8(out1, mask);
let out2 = _mm_shuffle_epi8(out2, mask);
let out3 = _mm_shuffle_epi8(out3, mask);
let (chunks, []) = dest.as_chunks_mut::<8usize>() else {
unreachable!()
};
crate::transmute::checked_transmute_store::<__m128i, [i16; 8usize]>(
out0,
&mut chunks[0],
);
crate::transmute::checked_transmute_store::<__m128i, [i16; 8usize]>(
out1,
&mut chunks[1],
);
crate::transmute::checked_transmute_store::<__m128i, [i16; 8usize]>(
out2,
&mut chunks[2],
);
crate::transmute::checked_transmute_store::<__m128i, [i16; 8usize]>(
out3,
&mut chunks[3],
);
}
);
kernel(self, vectors, dest);
}
#[inline(always)]
fn widen_i16x8(self, a: i16x8<Self>) -> (i32x4<Self>, i32x4<Self>) {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>) -> (i32x4<Sse4_2>, i32x4<Sse4_2>) {
let raw = a.into();
(
_mm_cvtepi16_epi32(raw).simd_into(token),
_mm_cvtepi16_epi32(_mm_srli_si128::<8>(raw)).simd_into(token),
)
}
);
kernel(self, a)
}
#[inline(always)]
fn narrow_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i8x16<Sse4_2> {
let mask = _mm_set1_epi16(0xff);
_mm_packus_epi16(_mm_and_si128(a.into(), mask), _mm_and_si128(b.into(), mask))
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_narrow_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i16x8<Sse4_2>, b: i16x8<Sse4_2>) -> i8x16<Sse4_2> {
_mm_packs_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn relaxed_narrow_i16x8(self, a: i16x8<Self>, b: i16x8<Self>) -> i8x16<Self> {
debug_assert!(
a.as_slice()
.iter()
.chain(b.as_slice())
.all(|&value| { value >= i8::MIN as i16 && value <= i8::MAX as i16 }),
"relaxed_narrow inputs must fit in the destination type",
);
self.saturating_narrow_i16x8(a, b)
}
#[inline(always)]
fn splat_u16x8(self, val: u16) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, val: u16) -> u16x8<Sse4_2> {
_mm_set1_epi16(val.cast_signed()).simd_into(token)
}
);
kernel(self, val)
}
#[inline(always)]
fn reverse_u16x8(self, a: u16x8<Self>) -> u16x8<Self> {
let indices: u8x16<Self> =
[14, 15, 12, 13, 10, 11, 8, 9, 6, 7, 4, 5, 2, 3, 0, 1].simd_into(self);
self.swizzle_dyn_u16x8(a, indices)
}
#[inline(always)]
fn slide_u16x8<const SHIFT: usize>(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
if SHIFT >= 8usize {
return b;
}
let result = dyn_alignr_128(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 2usize,
);
Bytes::from_bytes(u8x16 {
val: crate::support::Aligned128(result),
simd: self,
})
}
#[inline(always)]
fn compress_u16x8(self, values: u16x8<Self>, mask: mask16x8<Self>) -> u16x8<Self> {
let bits = self.to_bitmask_mask16x8(mask) as usize & 255;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_16[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn compress_merge_u16x8(
self,
values: u16x8<Self>,
mask: mask16x8<Self>,
merge: u16x8<Self>,
) -> u16x8<Self> {
let bits = self.to_bitmask_mask16x8(mask) as usize & 255;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_16[bits].0);
let compressed = self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control);
let inactive = self.simd_lt_i8x16(i8x16::from_bytes(control), i8x16::splat(self, 0));
Bytes::from_bytes(self.select_u8x16(inactive, Bytes::to_bytes(merge), compressed))
}
#[inline(always)]
fn expand_u16x8(self, values: u16x8<Self>, mask: mask16x8<Self>) -> u16x8<Self> {
let bits = self.to_bitmask_mask16x8(mask) as usize & 255;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_16[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn expand_merge_u16x8(
self,
values: u16x8<Self>,
mask: mask16x8<Self>,
merge: u16x8<Self>,
) -> u16x8<Self> {
let bits = self.to_bitmask_mask16x8(mask) as usize & 255;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_16[bits].0);
let expanded = Bytes::from_bytes(self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control));
self.select_u16x8(mask, expanded, merge)
}
#[inline(always)]
fn count_ones_u16x8(self, a: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
let value = a.into();
let nibble_mask = _mm_set1_epi8(0x0f);
let lookup = _mm_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);
let low = _mm_and_si128(value, nibble_mask);
let high = _mm_and_si128(_mm_srli_epi16::<4>(value), nibble_mask);
let byte_counts = _mm_add_epi8(
_mm_shuffle_epi8(lookup, low),
_mm_shuffle_epi8(lookup, high),
);
_mm_maddubs_epi16(byte_counts, _mm_set1_epi8(1)).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn count_zeros_u16x8(self, a: u16x8<Self>) -> u16x8<Self> {
self.count_ones_u16x8(self.not_u16x8(a))
}
#[inline(always)]
fn add_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
_mm_add_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_add_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
_mm_adds_epu16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn sub_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
_mm_sub_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_sub_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
_mm_subs_epu16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn mul_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
_mm_mullo_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn and_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
_mm_and_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn or_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
_mm_or_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn xor_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
_mm_xor_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn not_u16x8(self, a: u16x8<Self>) -> u16x8<Self> {
a ^ !0
}
#[inline(always)]
fn shl_u16x8(self, a: u16x8<Self>, shift: u32) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, shift: u32) -> u16x8<Sse4_2> {
_mm_sll_epi16(a.into(), _mm_cvtsi32_si128(shift.cast_signed())).simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shlv_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
[
u16::wrapping_shl(a[0usize], b[0usize] as u32),
u16::wrapping_shl(a[1usize], b[1usize] as u32),
u16::wrapping_shl(a[2usize], b[2usize] as u32),
u16::wrapping_shl(a[3usize], b[3usize] as u32),
u16::wrapping_shl(a[4usize], b[4usize] as u32),
u16::wrapping_shl(a[5usize], b[5usize] as u32),
u16::wrapping_shl(a[6usize], b[6usize] as u32),
u16::wrapping_shl(a[7usize], b[7usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn shr_u16x8(self, a: u16x8<Self>, shift: u32) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, shift: u32) -> u16x8<Sse4_2> {
_mm_srl_epi16(a.into(), _mm_cvtsi32_si128(shift.cast_signed())).simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shrv_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
[
u16::wrapping_shr(a[0usize], b[0usize] as u32),
u16::wrapping_shr(a[1usize], b[1usize] as u32),
u16::wrapping_shr(a[2usize], b[2usize] as u32),
u16::wrapping_shr(a[3usize], b[3usize] as u32),
u16::wrapping_shr(a[4usize], b[4usize] as u32),
u16::wrapping_shr(a[5usize], b[5usize] as u32),
u16::wrapping_shr(a[6usize], b[6usize] as u32),
u16::wrapping_shr(a[7usize], b[7usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn reduce_max_u16x8(self, a: u16x8<Self>) -> u16 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>) -> u16 {
let reduced: __m128i = a.into();
let shifted = _mm_srli_si128::<8>(reduced);
let reduced = { _mm_max_epu16(reduced, shifted) };
let shifted = _mm_srli_si128::<4>(reduced);
let reduced = { _mm_max_epu16(reduced, shifted) };
let shifted = _mm_srli_si128::<2>(reduced);
let reduced = { _mm_max_epu16(reduced, shifted) };
{
let lanes: [u16; 8usize] = crate::transmute::checked_transmute_copy(&reduced);
lanes[0]
}
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_min_u16x8(self, a: u16x8<Self>) -> u16 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>) -> u16 {
let reduced: __m128i = a.into();
let shifted = _mm_srli_si128::<8>(reduced);
let reduced = { _mm_min_epu16(reduced, shifted) };
let shifted = _mm_srli_si128::<4>(reduced);
let reduced = { _mm_min_epu16(reduced, shifted) };
let shifted = _mm_srli_si128::<2>(reduced);
let reduced = { _mm_min_epu16(reduced, shifted) };
{
let lanes: [u16; 8usize] = crate::transmute::checked_transmute_copy(&reduced);
lanes[0]
}
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_sum_u16x8(self, a: u16x8<Self>) -> u16 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>) -> u16 {
let sum: __m128i = a.into();
let sum = _mm_add_epi16(sum, _mm_srli_si128::<8>(sum));
let sum = _mm_add_epi16(sum, _mm_srli_si128::<4>(sum));
let sum = _mm_add_epi16(sum, _mm_srli_si128::<2>(sum));
let lanes: [u16; 8usize] = crate::transmute::checked_transmute_copy(&sum);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_product_u16x8(self, a: u16x8<Self>) -> u16 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>) -> u16 {
let product: __m128i = a.into();
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b11_10_11_10>(product));
let product = _mm_mullo_epi16(product, _mm_shuffle_epi32::<0b01_01_01_01>(product));
let product = _mm_mullo_epi16(product, _mm_srli_epi32::<16>(product));
let lanes: [u16; 8usize] = crate::transmute::checked_transmute_copy(&product);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn max_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
_mm_max_epu16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn min_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
_mm_min_epu16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_eq_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> mask16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> mask16x8<Sse4_2> {
_mm_cmpeq_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_lt_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> mask16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> mask16x8<Sse4_2> {
{
let sign_bit = _mm_set1_epi16(0x8000u16.cast_signed());
let lhs_signed = _mm_xor_si128(b.into(), sign_bit);
let rhs_signed = _mm_xor_si128(a.into(), sign_bit);
_mm_cmpgt_epi16(lhs_signed, rhs_signed)
}
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_le_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> mask16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> mask16x8<Sse4_2> {
_mm_cmpeq_epi16(_mm_min_epu16(a.into(), b.into()), a.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_low_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
_mm_unpacklo_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_high_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
_mm_unpackhi_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_low_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
let mask = _mm_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15);
let t1 = _mm_shuffle_epi8(a.into(), mask);
let t2 = _mm_shuffle_epi8(b.into(), mask);
_mm_unpacklo_epi64(t1, t2).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_high_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u16x8<Sse4_2> {
let mask = _mm_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15);
let t1 = _mm_shuffle_epi8(a.into(), mask);
let t2 = _mm_shuffle_epi8(b.into(), mask);
_mm_unpackhi_epi64(t1, t2).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn interleave_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> (u16x8<Self>, u16x8<Self>) {
(self.zip_low_u16x8(a, b), self.zip_high_u16x8(a, b))
}
#[inline(always)]
fn deinterleave_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> (u16x8<Self>, u16x8<Self>) {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: u16x8<Sse4_2>,
b: u16x8<Sse4_2>,
) -> (u16x8<Sse4_2>, u16x8<Sse4_2>) {
let mask = _mm_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15);
let a = _mm_shuffle_epi8(a.into(), mask);
let b = _mm_shuffle_epi8(b.into(), mask);
(
_mm_unpacklo_epi64(a, b).simd_into(token),
_mm_unpackhi_epi64(a, b).simd_into(token),
)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn select_u16x8(self, a: mask16x8<Self>, b: u16x8<Self>, c: u16x8<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: mask16x8<Sse4_2>,
b: u16x8<Sse4_2>,
c: u16x8<Sse4_2>,
) -> u16x8<Sse4_2> {
_mm_blendv_epi8(c.into(), b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b, c)
}
#[inline(always)]
fn combine_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u16x16<Self> {
u16x16 {
val: crate::support::Aligned256([a.val.0, b.val.0]),
simd: self,
}
}
#[inline(always)]
fn load_four_interleaved_u16x8(self, src: &[u16; 32usize]) -> [u16x8<Self>; 4usize] {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, src: &[u16; 32usize]) -> [u16x8<Sse4_2>; 4usize] {
let (chunks, []) = src.as_chunks::<8usize>() else {
unreachable!()
};
let v0: __m128i =
crate::transmute::checked_transmute_copy::<[u16; 8usize], __m128i>(&chunks[0]);
let v1: __m128i =
crate::transmute::checked_transmute_copy::<[u16; 8usize], __m128i>(&chunks[1]);
let v2: __m128i =
crate::transmute::checked_transmute_copy::<[u16; 8usize], __m128i>(&chunks[2]);
let v3: __m128i =
crate::transmute::checked_transmute_copy::<[u16; 8usize], __m128i>(&chunks[3]);
let mask = _mm_setr_epi8(0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15);
let v0 = _mm_shuffle_epi8(v0, mask);
let v1 = _mm_shuffle_epi8(v1, mask);
let v2 = _mm_shuffle_epi8(v2, mask);
let v3 = _mm_shuffle_epi8(v3, mask);
let tmp0 = _mm_unpacklo_epi32(v0, v1);
let tmp1 = _mm_unpackhi_epi32(v0, v1);
let tmp2 = _mm_unpacklo_epi32(v2, v3);
let tmp3 = _mm_unpackhi_epi32(v2, v3);
let out0 = _mm_unpacklo_epi64(tmp0, tmp2);
let out1 = _mm_unpackhi_epi64(tmp0, tmp2);
let out2 = _mm_unpacklo_epi64(tmp1, tmp3);
let out3 = _mm_unpackhi_epi64(tmp1, tmp3);
[
out0.simd_into(token),
out1.simd_into(token),
out2.simd_into(token),
out3.simd_into(token),
]
}
);
kernel(self, src)
}
#[inline(always)]
fn store_four_interleaved_u16x8(
self,
vectors: [u16x8<Self>; 4usize],
dest: &mut [u16; 32usize],
) -> () {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
vectors: [u16x8<Sse4_2>; 4usize],
dest: &mut [u16; 32usize],
) -> () {
let _ = token;
let v0: __m128i = vectors[0].into();
let v1: __m128i = vectors[1].into();
let v2: __m128i = vectors[2].into();
let v3: __m128i = vectors[3].into();
let tmp0 = _mm_unpacklo_epi32(v0, v1);
let tmp1 = _mm_unpackhi_epi32(v0, v1);
let tmp2 = _mm_unpacklo_epi32(v2, v3);
let tmp3 = _mm_unpackhi_epi32(v2, v3);
let out0 = _mm_unpacklo_epi64(tmp0, tmp2);
let out1 = _mm_unpackhi_epi64(tmp0, tmp2);
let out2 = _mm_unpacklo_epi64(tmp1, tmp3);
let out3 = _mm_unpackhi_epi64(tmp1, tmp3);
let mask = _mm_setr_epi8(0, 1, 4, 5, 8, 9, 12, 13, 2, 3, 6, 7, 10, 11, 14, 15);
let out0 = _mm_shuffle_epi8(out0, mask);
let out1 = _mm_shuffle_epi8(out1, mask);
let out2 = _mm_shuffle_epi8(out2, mask);
let out3 = _mm_shuffle_epi8(out3, mask);
let (chunks, []) = dest.as_chunks_mut::<8usize>() else {
unreachable!()
};
crate::transmute::checked_transmute_store::<__m128i, [u16; 8usize]>(
out0,
&mut chunks[0],
);
crate::transmute::checked_transmute_store::<__m128i, [u16; 8usize]>(
out1,
&mut chunks[1],
);
crate::transmute::checked_transmute_store::<__m128i, [u16; 8usize]>(
out2,
&mut chunks[2],
);
crate::transmute::checked_transmute_store::<__m128i, [u16; 8usize]>(
out3,
&mut chunks[3],
);
}
);
kernel(self, vectors, dest);
}
#[inline(always)]
fn widen_u16x8(self, a: u16x8<Self>) -> (u32x4<Self>, u32x4<Self>) {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>) -> (u32x4<Sse4_2>, u32x4<Sse4_2>) {
let raw = a.into();
(
_mm_cvtepu16_epi32(raw).simd_into(token),
_mm_cvtepu16_epi32(_mm_srli_si128::<8>(raw)).simd_into(token),
)
}
);
kernel(self, a)
}
#[inline(always)]
fn narrow_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u8x16<Sse4_2> {
let mask = _mm_set1_epi16(0xff);
_mm_packus_epi16(_mm_and_si128(a.into(), mask), _mm_and_si128(b.into(), mask))
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_narrow_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u16x8<Sse4_2>, b: u16x8<Sse4_2>) -> u8x16<Sse4_2> {
let max = _mm_set1_epi16(u8::MAX as i16);
let a = a.into();
let b = b.into();
let a = _mm_sub_epi16(a, _mm_subs_epu16(a, max));
let b = _mm_sub_epi16(b, _mm_subs_epu16(b, max));
_mm_packus_epi16(a, b).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn relaxed_narrow_u16x8(self, a: u16x8<Self>, b: u16x8<Self>) -> u8x16<Self> {
debug_assert!(
a.as_slice()
.iter()
.chain(b.as_slice())
.all(|&value| { value >= u8::MIN as u16 && value <= u8::MAX as u16 }),
"relaxed_narrow inputs must fit in the destination type",
);
self.narrow_u16x8(a, b)
}
#[inline(always)]
fn splat_mask16x8(self, val: bool) -> mask16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, val: bool) -> mask16x8<Sse4_2> {
let val: i16 = if val { !0 } else { 0 };
_mm_set1_epi16(val).simd_into(token)
}
);
kernel(self, val)
}
#[inline(always)]
fn from_bitmask_mask16x8(self, bits: u64) -> mask16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, bits: u64) -> mask16x8<Sse4_2> {
{
let bit_lanes = _mm_set1_epi16(bits as i16);
let bit_mask = _mm_setr_epi16(1, 2, 4, 8, 16, 32, 64, 128);
_mm_cmpeq_epi16(_mm_and_si128(bit_lanes, bit_mask), bit_mask)
}
.simd_into(token)
}
);
kernel(self, bits)
}
#[inline(always)]
fn to_bitmask_mask16x8(self, a: mask16x8<Self>) -> u64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask16x8<Sse4_2>) -> u64 {
{
let packed = _mm_packs_epi16(a.into(), a.into());
_mm_movemask_epi8(packed) as u8 as u64
}
}
);
kernel(self, a)
}
#[inline(always)]
fn from_vector_mask16x8(self, vector: i16x8<Self>) -> mask16x8<Self> {
mask16x8 {
val: vector.val,
simd: self,
}
}
#[inline(always)]
fn to_vector_mask16x8(self, a: mask16x8<Self>) -> i16x8<Self> {
i16x8 {
val: a.val,
simd: self,
}
}
#[inline(always)]
fn set_mask16x8(self, a: &mut mask16x8<Self>, index: usize, value: bool) -> () {
assert!(
index < 8usize,
"mask lane index {index} is out of bounds for {} lanes",
8usize
);
let mut lanes: [i16; 8usize] = (*a).into();
lanes[index] = if value { !0 } else { 0 };
*a = lanes.simd_into(self);
}
#[inline(always)]
fn rotate_elements_left_mask16x8<const OFFSET: usize>(
self,
a: mask16x8<Self>,
) -> mask16x8<Self> {
let int = i16x8 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_left::<OFFSET>();
mask16x8 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn rotate_elements_right_mask16x8<const OFFSET: usize>(
self,
a: mask16x8<Self>,
) -> mask16x8<Self> {
let int = i16x8 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_right::<OFFSET>();
mask16x8 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn and_mask16x8(self, a: mask16x8<Self>, b: mask16x8<Self>) -> mask16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask16x8<Sse4_2>, b: mask16x8<Sse4_2>) -> mask16x8<Sse4_2> {
_mm_and_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn or_mask16x8(self, a: mask16x8<Self>, b: mask16x8<Self>) -> mask16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask16x8<Sse4_2>, b: mask16x8<Sse4_2>) -> mask16x8<Sse4_2> {
_mm_or_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn xor_mask16x8(self, a: mask16x8<Self>, b: mask16x8<Self>) -> mask16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask16x8<Sse4_2>, b: mask16x8<Sse4_2>) -> mask16x8<Sse4_2> {
_mm_xor_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn not_mask16x8(self, a: mask16x8<Self>) -> mask16x8<Self> {
self.xor_mask16x8(a, self.splat_mask16x8(true))
}
#[inline(always)]
fn reverse_mask16x8(self, a: mask16x8<Self>) -> mask16x8<Self> {
let lanes = i16x8 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let reversed = self.reverse_i16x8(lanes);
mask16x8 {
val: crate::transmute::checked_transmute_copy(&reversed.val),
simd: self,
}
}
#[inline(always)]
fn select_mask16x8(
self,
a: mask16x8<Self>,
b: mask16x8<Self>,
c: mask16x8<Self>,
) -> mask16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: mask16x8<Sse4_2>,
b: mask16x8<Sse4_2>,
c: mask16x8<Sse4_2>,
) -> mask16x8<Sse4_2> {
_mm_blendv_epi8(c.into(), b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b, c)
}
#[inline(always)]
fn simd_eq_mask16x8(self, a: mask16x8<Self>, b: mask16x8<Self>) -> mask16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask16x8<Sse4_2>, b: mask16x8<Sse4_2>) -> mask16x8<Sse4_2> {
_mm_cmpeq_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn any_true_mask16x8(self, a: mask16x8<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask16x8<Sse4_2>) -> bool {
_mm_movemask_epi8(a.into()) as u32 != 0
}
);
kernel(self, a)
}
#[inline(always)]
fn all_true_mask16x8(self, a: mask16x8<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask16x8<Sse4_2>) -> bool {
_mm_movemask_epi8(a.into()) as u32 == 0xffff
}
);
kernel(self, a)
}
#[inline(always)]
fn any_false_mask16x8(self, a: mask16x8<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask16x8<Sse4_2>) -> bool {
_mm_movemask_epi8(a.into()) as u32 != 0xffff
}
);
kernel(self, a)
}
#[inline(always)]
fn all_false_mask16x8(self, a: mask16x8<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask16x8<Sse4_2>) -> bool {
_mm_movemask_epi8(a.into()) as u32 == 0
}
);
kernel(self, a)
}
#[inline(always)]
fn combine_mask16x8(self, a: mask16x8<Self>, b: mask16x8<Self>) -> mask16x16<Self> {
mask16x16 {
val: crate::support::Aligned256([a.val.0, b.val.0]),
simd: self,
}
}
#[inline(always)]
fn widen_mask16x8(self, a: mask16x8<Self>) -> (mask32x4<Self>, mask32x4<Self>) {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask16x8<Sse4_2>) -> (mask32x4<Sse4_2>, mask32x4<Sse4_2>) {
let raw = a.into();
(
_mm_unpacklo_epi16(raw, raw).simd_into(token),
_mm_unpackhi_epi16(raw, raw).simd_into(token),
)
}
);
kernel(self, a)
}
#[inline(always)]
fn narrow_mask16x8(self, a: mask16x8<Self>, b: mask16x8<Self>) -> mask8x16<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask16x8<Sse4_2>, b: mask16x8<Sse4_2>) -> mask8x16<Sse4_2> {
_mm_packs_epi16(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn abs_i32x4(self, a: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
_mm_abs_epi32(a.into()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn splat_i32x4(self, val: i32) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, val: i32) -> i32x4<Sse4_2> {
_mm_set1_epi32(val).simd_into(token)
}
);
kernel(self, val)
}
#[inline(always)]
fn reverse_i32x4(self, a: i32x4<Self>) -> i32x4<Self> {
let indices: u8x16<Self> =
[12, 13, 14, 15, 8, 9, 10, 11, 4, 5, 6, 7, 0, 1, 2, 3].simd_into(self);
self.swizzle_dyn_i32x4(a, indices)
}
#[inline(always)]
fn slide_i32x4<const SHIFT: usize>(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
if SHIFT >= 4usize {
return b;
}
let result = dyn_alignr_128(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 4usize,
);
Bytes::from_bytes(u8x16 {
val: crate::support::Aligned128(result),
simd: self,
})
}
#[inline(always)]
fn compress_i32x4(self, values: i32x4<Self>, mask: mask32x4<Self>) -> i32x4<Self> {
let bits = self.to_bitmask_mask32x4(mask) as usize & 15;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_32[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn compress_merge_i32x4(
self,
values: i32x4<Self>,
mask: mask32x4<Self>,
merge: i32x4<Self>,
) -> i32x4<Self> {
let bits = self.to_bitmask_mask32x4(mask) as usize & 15;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_32[bits].0);
let compressed = self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control);
let inactive = self.simd_lt_i8x16(i8x16::from_bytes(control), i8x16::splat(self, 0));
Bytes::from_bytes(self.select_u8x16(inactive, Bytes::to_bytes(merge), compressed))
}
#[inline(always)]
fn expand_i32x4(self, values: i32x4<Self>, mask: mask32x4<Self>) -> i32x4<Self> {
let bits = self.to_bitmask_mask32x4(mask) as usize & 15;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_32[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn expand_merge_i32x4(
self,
values: i32x4<Self>,
mask: mask32x4<Self>,
merge: i32x4<Self>,
) -> i32x4<Self> {
let bits = self.to_bitmask_mask32x4(mask) as usize & 15;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_32[bits].0);
let expanded = Bytes::from_bytes(self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control));
self.select_i32x4(mask, expanded, merge)
}
#[inline(always)]
fn count_ones_i32x4(self, a: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
let value = a.into();
let nibble_mask = _mm_set1_epi8(0x0f);
let lookup = _mm_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);
let low = _mm_and_si128(value, nibble_mask);
let high = _mm_and_si128(_mm_srli_epi16::<4>(value), nibble_mask);
let byte_counts = _mm_add_epi8(
_mm_shuffle_epi8(lookup, low),
_mm_shuffle_epi8(lookup, high),
);
_mm_madd_epi16(
_mm_maddubs_epi16(byte_counts, _mm_set1_epi8(1)),
_mm_set1_epi16(1),
)
.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn count_zeros_i32x4(self, a: i32x4<Self>) -> i32x4<Self> {
self.count_ones_i32x4(self.not_i32x4(a))
}
#[inline(always)]
fn add_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
_mm_add_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_add_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
let a = a.into();
let b = b.into();
let wrapped = _mm_add_epi32(a, b);
let overflow = _mm_xor_si128(_mm_cmpgt_epi32(a, wrapped), b);
let bound = _mm_xor_si128(_mm_srai_epi32::<31>(a), _mm_set1_epi32(i32::MAX));
let result = _mm_blendv_ps(
_mm_castsi128_ps(wrapped),
_mm_castsi128_ps(bound),
_mm_castsi128_ps(overflow),
);
_mm_castps_si128(result).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn sub_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
_mm_sub_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_sub_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
let a = a.into();
let b = b.into();
let wrapped = _mm_sub_epi32(a, b);
let overflow = _mm_xor_si128(_mm_cmpgt_epi32(wrapped, a), b);
let bound = _mm_xor_si128(_mm_srai_epi32::<31>(a), _mm_set1_epi32(i32::MAX));
let result = _mm_blendv_ps(
_mm_castsi128_ps(wrapped),
_mm_castsi128_ps(bound),
_mm_castsi128_ps(overflow),
);
_mm_castps_si128(result).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn mul_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
_mm_mullo_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn and_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
_mm_and_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn or_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
_mm_or_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn xor_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
_mm_xor_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn not_i32x4(self, a: i32x4<Self>) -> i32x4<Self> {
a ^ !0
}
#[inline(always)]
fn shl_i32x4(self, a: i32x4<Self>, shift: u32) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, shift: u32) -> i32x4<Sse4_2> {
_mm_sll_epi32(a.into(), _mm_cvtsi32_si128(shift.cast_signed())).simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shlv_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
[
i32::wrapping_shl(a[0usize], b[0usize] as u32),
i32::wrapping_shl(a[1usize], b[1usize] as u32),
i32::wrapping_shl(a[2usize], b[2usize] as u32),
i32::wrapping_shl(a[3usize], b[3usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn shr_i32x4(self, a: i32x4<Self>, shift: u32) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, shift: u32) -> i32x4<Sse4_2> {
_mm_sra_epi32(a.into(), _mm_cvtsi32_si128(shift.cast_signed())).simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shrv_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
[
i32::wrapping_shr(a[0usize], b[0usize] as u32),
i32::wrapping_shr(a[1usize], b[1usize] as u32),
i32::wrapping_shr(a[2usize], b[2usize] as u32),
i32::wrapping_shr(a[3usize], b[3usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn reduce_max_i32x4(self, a: i32x4<Self>) -> i32 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>) -> i32 {
let reduced: __m128i = a.into();
let shifted = _mm_srli_si128::<8>(reduced);
let reduced = { _mm_max_epi32(reduced, shifted) };
let shifted = _mm_srli_si128::<4>(reduced);
let reduced = { _mm_max_epi32(reduced, shifted) };
{
let lanes: [i32; 4usize] = crate::transmute::checked_transmute_copy(&reduced);
lanes[0]
}
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_min_i32x4(self, a: i32x4<Self>) -> i32 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>) -> i32 {
let reduced: __m128i = a.into();
let shifted = _mm_srli_si128::<8>(reduced);
let reduced = { _mm_min_epi32(reduced, shifted) };
let shifted = _mm_srli_si128::<4>(reduced);
let reduced = { _mm_min_epi32(reduced, shifted) };
{
let lanes: [i32; 4usize] = crate::transmute::checked_transmute_copy(&reduced);
lanes[0]
}
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_sum_i32x4(self, a: i32x4<Self>) -> i32 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>) -> i32 {
let sum: __m128i = a.into();
let sum = _mm_add_epi32(sum, _mm_srli_si128::<8>(sum));
let sum = _mm_add_epi32(sum, _mm_srli_si128::<4>(sum));
let lanes: [i32; 4usize] = crate::transmute::checked_transmute_copy(&sum);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_product_i32x4(self, a: i32x4<Self>) -> i32 {
let mul_level_0: [i32; 2usize] = [
a[0usize].wrapping_mul(a[1usize]),
a[2usize].wrapping_mul(a[3usize]),
];
let mul_level_1: [i32; 1usize] = [mul_level_0[0usize].wrapping_mul(mul_level_0[1usize])];
mul_level_1[0]
}
#[inline(always)]
fn max_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
_mm_max_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn min_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
_mm_min_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_eq_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> mask32x4<Sse4_2> {
_mm_cmpeq_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_lt_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> mask32x4<Sse4_2> {
_mm_cmpgt_epi32(b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_le_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> mask32x4<Sse4_2> {
_mm_cmpeq_epi32(_mm_min_epi32(a.into(), b.into()), a.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_low_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
_mm_unpacklo_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_high_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
_mm_unpackhi_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_low_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
let t1 = _mm_shuffle_epi32::<0b11_01_10_00>(a.into());
let t2 = _mm_shuffle_epi32::<0b11_01_10_00>(b.into());
_mm_unpacklo_epi64(t1, t2).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_high_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
let t1 = _mm_shuffle_epi32::<0b11_01_10_00>(a.into());
let t2 = _mm_shuffle_epi32::<0b11_01_10_00>(b.into());
_mm_unpackhi_epi64(t1, t2).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn interleave_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> (i32x4<Self>, i32x4<Self>) {
(self.zip_low_i32x4(a, b), self.zip_high_i32x4(a, b))
}
#[inline(always)]
fn deinterleave_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> (i32x4<Self>, i32x4<Self>) {
(self.unzip_low_i32x4(a, b), self.unzip_high_i32x4(a, b))
}
#[inline(always)]
fn select_i32x4(self, a: mask32x4<Self>, b: i32x4<Self>, c: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: mask32x4<Sse4_2>,
b: i32x4<Sse4_2>,
c: i32x4<Sse4_2>,
) -> i32x4<Sse4_2> {
_mm_blendv_epi8(c.into(), b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b, c)
}
#[inline(always)]
fn combine_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i32x8<Self> {
i32x8 {
val: crate::support::Aligned256([a.val.0, b.val.0]),
simd: self,
}
}
#[inline(always)]
fn neg_i32x4(self, a: i32x4<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>) -> i32x4<Sse4_2> {
_mm_sub_epi32(_mm_setzero_si128(), a.into()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn load_four_interleaved_i32x4(self, src: &[i32; 16usize]) -> [i32x4<Self>; 4usize] {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, src: &[i32; 16usize]) -> [i32x4<Sse4_2>; 4usize] {
let (chunks, []) = src.as_chunks::<4usize>() else {
unreachable!()
};
let v0: __m128i =
crate::transmute::checked_transmute_copy::<[i32; 4usize], __m128i>(&chunks[0]);
let v1: __m128i =
crate::transmute::checked_transmute_copy::<[i32; 4usize], __m128i>(&chunks[1]);
let v2: __m128i =
crate::transmute::checked_transmute_copy::<[i32; 4usize], __m128i>(&chunks[2]);
let v3: __m128i =
crate::transmute::checked_transmute_copy::<[i32; 4usize], __m128i>(&chunks[3]);
let tmp0 = _mm_unpacklo_epi32(v0, v1);
let tmp1 = _mm_unpackhi_epi32(v0, v1);
let tmp2 = _mm_unpacklo_epi32(v2, v3);
let tmp3 = _mm_unpackhi_epi32(v2, v3);
let out0 = _mm_unpacklo_epi64(tmp0, tmp2);
let out1 = _mm_unpackhi_epi64(tmp0, tmp2);
let out2 = _mm_unpacklo_epi64(tmp1, tmp3);
let out3 = _mm_unpackhi_epi64(tmp1, tmp3);
[
out0.simd_into(token),
out1.simd_into(token),
out2.simd_into(token),
out3.simd_into(token),
]
}
);
kernel(self, src)
}
#[inline(always)]
fn store_four_interleaved_i32x4(
self,
vectors: [i32x4<Self>; 4usize],
dest: &mut [i32; 16usize],
) -> () {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
vectors: [i32x4<Sse4_2>; 4usize],
dest: &mut [i32; 16usize],
) -> () {
let _ = token;
let v0: __m128i = vectors[0].into();
let v1: __m128i = vectors[1].into();
let v2: __m128i = vectors[2].into();
let v3: __m128i = vectors[3].into();
let tmp0 = _mm_unpacklo_epi32(v0, v1);
let tmp1 = _mm_unpackhi_epi32(v0, v1);
let tmp2 = _mm_unpacklo_epi32(v2, v3);
let tmp3 = _mm_unpackhi_epi32(v2, v3);
let out0 = _mm_unpacklo_epi64(tmp0, tmp2);
let out1 = _mm_unpackhi_epi64(tmp0, tmp2);
let out2 = _mm_unpacklo_epi64(tmp1, tmp3);
let out3 = _mm_unpackhi_epi64(tmp1, tmp3);
let (chunks, []) = dest.as_chunks_mut::<4usize>() else {
unreachable!()
};
crate::transmute::checked_transmute_store::<__m128i, [i32; 4usize]>(
out0,
&mut chunks[0],
);
crate::transmute::checked_transmute_store::<__m128i, [i32; 4usize]>(
out1,
&mut chunks[1],
);
crate::transmute::checked_transmute_store::<__m128i, [i32; 4usize]>(
out2,
&mut chunks[2],
);
crate::transmute::checked_transmute_store::<__m128i, [i32; 4usize]>(
out3,
&mut chunks[3],
);
}
);
kernel(self, vectors, dest);
}
#[inline(always)]
fn widen_i32x4(self, a: i32x4<Self>) -> (i64x2<Self>, i64x2<Self>) {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>) -> (i64x2<Sse4_2>, i64x2<Sse4_2>) {
let raw = a.into();
(
_mm_cvtepi32_epi64(raw).simd_into(token),
_mm_cvtepi32_epi64(_mm_srli_si128::<8>(raw)).simd_into(token),
)
}
);
kernel(self, a)
}
#[inline(always)]
fn narrow_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i16x8<Sse4_2> {
let mask = _mm_set1_epi32(0xffff);
_mm_packus_epi32(_mm_and_si128(a.into(), mask), _mm_and_si128(b.into(), mask))
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_narrow_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>, b: i32x4<Sse4_2>) -> i16x8<Sse4_2> {
_mm_packs_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn relaxed_narrow_i32x4(self, a: i32x4<Self>, b: i32x4<Self>) -> i16x8<Self> {
debug_assert!(
a.as_slice()
.iter()
.chain(b.as_slice())
.all(|&value| { value >= i16::MIN as i32 && value <= i16::MAX as i32 }),
"relaxed_narrow inputs must fit in the destination type",
);
self.saturating_narrow_i32x4(a, b)
}
#[inline(always)]
fn cvt_f32_i32x4(self, a: i32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i32x4<Sse4_2>) -> f32x4<Sse4_2> {
_mm_cvtepi32_ps(a.into()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn splat_u32x4(self, val: u32) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, val: u32) -> u32x4<Sse4_2> {
_mm_set1_epi32(val.cast_signed()).simd_into(token)
}
);
kernel(self, val)
}
#[inline(always)]
fn reverse_u32x4(self, a: u32x4<Self>) -> u32x4<Self> {
let indices: u8x16<Self> =
[12, 13, 14, 15, 8, 9, 10, 11, 4, 5, 6, 7, 0, 1, 2, 3].simd_into(self);
self.swizzle_dyn_u32x4(a, indices)
}
#[inline(always)]
fn slide_u32x4<const SHIFT: usize>(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
if SHIFT >= 4usize {
return b;
}
let result = dyn_alignr_128(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 4usize,
);
Bytes::from_bytes(u8x16 {
val: crate::support::Aligned128(result),
simd: self,
})
}
#[inline(always)]
fn compress_u32x4(self, values: u32x4<Self>, mask: mask32x4<Self>) -> u32x4<Self> {
let bits = self.to_bitmask_mask32x4(mask) as usize & 15;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_32[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn compress_merge_u32x4(
self,
values: u32x4<Self>,
mask: mask32x4<Self>,
merge: u32x4<Self>,
) -> u32x4<Self> {
let bits = self.to_bitmask_mask32x4(mask) as usize & 15;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_32[bits].0);
let compressed = self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control);
let inactive = self.simd_lt_i8x16(i8x16::from_bytes(control), i8x16::splat(self, 0));
Bytes::from_bytes(self.select_u8x16(inactive, Bytes::to_bytes(merge), compressed))
}
#[inline(always)]
fn expand_u32x4(self, values: u32x4<Self>, mask: mask32x4<Self>) -> u32x4<Self> {
let bits = self.to_bitmask_mask32x4(mask) as usize & 15;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_32[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn expand_merge_u32x4(
self,
values: u32x4<Self>,
mask: mask32x4<Self>,
merge: u32x4<Self>,
) -> u32x4<Self> {
let bits = self.to_bitmask_mask32x4(mask) as usize & 15;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_32[bits].0);
let expanded = Bytes::from_bytes(self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control));
self.select_u32x4(mask, expanded, merge)
}
#[inline(always)]
fn count_ones_u32x4(self, a: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
let value = a.into();
let nibble_mask = _mm_set1_epi8(0x0f);
let lookup = _mm_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);
let low = _mm_and_si128(value, nibble_mask);
let high = _mm_and_si128(_mm_srli_epi16::<4>(value), nibble_mask);
let byte_counts = _mm_add_epi8(
_mm_shuffle_epi8(lookup, low),
_mm_shuffle_epi8(lookup, high),
);
_mm_madd_epi16(
_mm_maddubs_epi16(byte_counts, _mm_set1_epi8(1)),
_mm_set1_epi16(1),
)
.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn count_zeros_u32x4(self, a: u32x4<Self>) -> u32x4<Self> {
self.count_ones_u32x4(self.not_u32x4(a))
}
#[inline(always)]
fn add_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
_mm_add_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_add_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
let a = a.into();
let b = b.into();
let threshold = _mm_xor_si128(b, _mm_set1_epi32(-1));
_mm_add_epi32(_mm_min_epu32(a, threshold), b).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn sub_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
_mm_sub_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_sub_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
let a = a.into();
let b = b.into();
_mm_sub_epi32(_mm_max_epu32(a, b), b).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn mul_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
_mm_mullo_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn and_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
_mm_and_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn or_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
_mm_or_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn xor_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
_mm_xor_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn not_u32x4(self, a: u32x4<Self>) -> u32x4<Self> {
a ^ !0
}
#[inline(always)]
fn shl_u32x4(self, a: u32x4<Self>, shift: u32) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, shift: u32) -> u32x4<Sse4_2> {
_mm_sll_epi32(a.into(), _mm_cvtsi32_si128(shift.cast_signed())).simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shlv_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
[
u32::wrapping_shl(a[0usize], b[0usize] as u32),
u32::wrapping_shl(a[1usize], b[1usize] as u32),
u32::wrapping_shl(a[2usize], b[2usize] as u32),
u32::wrapping_shl(a[3usize], b[3usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn shr_u32x4(self, a: u32x4<Self>, shift: u32) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, shift: u32) -> u32x4<Sse4_2> {
_mm_srl_epi32(a.into(), _mm_cvtsi32_si128(shift.cast_signed())).simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shrv_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
[
u32::wrapping_shr(a[0usize], b[0usize] as u32),
u32::wrapping_shr(a[1usize], b[1usize] as u32),
u32::wrapping_shr(a[2usize], b[2usize] as u32),
u32::wrapping_shr(a[3usize], b[3usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn reduce_max_u32x4(self, a: u32x4<Self>) -> u32 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>) -> u32 {
let reduced: __m128i = a.into();
let shifted = _mm_srli_si128::<8>(reduced);
let reduced = { _mm_max_epu32(reduced, shifted) };
let shifted = _mm_srli_si128::<4>(reduced);
let reduced = { _mm_max_epu32(reduced, shifted) };
{
let lanes: [u32; 4usize] = crate::transmute::checked_transmute_copy(&reduced);
lanes[0]
}
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_min_u32x4(self, a: u32x4<Self>) -> u32 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>) -> u32 {
let reduced: __m128i = a.into();
let shifted = _mm_srli_si128::<8>(reduced);
let reduced = { _mm_min_epu32(reduced, shifted) };
let shifted = _mm_srli_si128::<4>(reduced);
let reduced = { _mm_min_epu32(reduced, shifted) };
{
let lanes: [u32; 4usize] = crate::transmute::checked_transmute_copy(&reduced);
lanes[0]
}
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_sum_u32x4(self, a: u32x4<Self>) -> u32 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>) -> u32 {
let sum: __m128i = a.into();
let sum = _mm_add_epi32(sum, _mm_srli_si128::<8>(sum));
let sum = _mm_add_epi32(sum, _mm_srli_si128::<4>(sum));
let lanes: [u32; 4usize] = crate::transmute::checked_transmute_copy(&sum);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_product_u32x4(self, a: u32x4<Self>) -> u32 {
let mul_level_0: [u32; 2usize] = [
a[0usize].wrapping_mul(a[1usize]),
a[2usize].wrapping_mul(a[3usize]),
];
let mul_level_1: [u32; 1usize] = [mul_level_0[0usize].wrapping_mul(mul_level_0[1usize])];
mul_level_1[0]
}
#[inline(always)]
fn max_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
_mm_max_epu32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn min_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
_mm_min_epu32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_eq_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> mask32x4<Sse4_2> {
_mm_cmpeq_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_lt_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> mask32x4<Sse4_2> {
{
let sign_bit = _mm_set1_epi32(0x80000000u32.cast_signed());
let lhs_signed = _mm_xor_si128(b.into(), sign_bit);
let rhs_signed = _mm_xor_si128(a.into(), sign_bit);
_mm_cmpgt_epi32(lhs_signed, rhs_signed)
}
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_le_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> mask32x4<Sse4_2> {
_mm_cmpeq_epi32(_mm_min_epu32(a.into(), b.into()), a.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_low_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
_mm_unpacklo_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_high_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
_mm_unpackhi_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_low_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
let t1 = _mm_shuffle_epi32::<0b11_01_10_00>(a.into());
let t2 = _mm_shuffle_epi32::<0b11_01_10_00>(b.into());
_mm_unpacklo_epi64(t1, t2).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_high_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u32x4<Sse4_2> {
let t1 = _mm_shuffle_epi32::<0b11_01_10_00>(a.into());
let t2 = _mm_shuffle_epi32::<0b11_01_10_00>(b.into());
_mm_unpackhi_epi64(t1, t2).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn interleave_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> (u32x4<Self>, u32x4<Self>) {
(self.zip_low_u32x4(a, b), self.zip_high_u32x4(a, b))
}
#[inline(always)]
fn deinterleave_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> (u32x4<Self>, u32x4<Self>) {
(self.unzip_low_u32x4(a, b), self.unzip_high_u32x4(a, b))
}
#[inline(always)]
fn select_u32x4(self, a: mask32x4<Self>, b: u32x4<Self>, c: u32x4<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: mask32x4<Sse4_2>,
b: u32x4<Sse4_2>,
c: u32x4<Sse4_2>,
) -> u32x4<Sse4_2> {
_mm_blendv_epi8(c.into(), b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b, c)
}
#[inline(always)]
fn combine_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u32x8<Self> {
u32x8 {
val: crate::support::Aligned256([a.val.0, b.val.0]),
simd: self,
}
}
#[inline(always)]
fn load_four_interleaved_u32x4(self, src: &[u32; 16usize]) -> [u32x4<Self>; 4usize] {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, src: &[u32; 16usize]) -> [u32x4<Sse4_2>; 4usize] {
let (chunks, []) = src.as_chunks::<4usize>() else {
unreachable!()
};
let v0: __m128i =
crate::transmute::checked_transmute_copy::<[u32; 4usize], __m128i>(&chunks[0]);
let v1: __m128i =
crate::transmute::checked_transmute_copy::<[u32; 4usize], __m128i>(&chunks[1]);
let v2: __m128i =
crate::transmute::checked_transmute_copy::<[u32; 4usize], __m128i>(&chunks[2]);
let v3: __m128i =
crate::transmute::checked_transmute_copy::<[u32; 4usize], __m128i>(&chunks[3]);
let tmp0 = _mm_unpacklo_epi32(v0, v1);
let tmp1 = _mm_unpackhi_epi32(v0, v1);
let tmp2 = _mm_unpacklo_epi32(v2, v3);
let tmp3 = _mm_unpackhi_epi32(v2, v3);
let out0 = _mm_unpacklo_epi64(tmp0, tmp2);
let out1 = _mm_unpackhi_epi64(tmp0, tmp2);
let out2 = _mm_unpacklo_epi64(tmp1, tmp3);
let out3 = _mm_unpackhi_epi64(tmp1, tmp3);
[
out0.simd_into(token),
out1.simd_into(token),
out2.simd_into(token),
out3.simd_into(token),
]
}
);
kernel(self, src)
}
#[inline(always)]
fn store_four_interleaved_u32x4(
self,
vectors: [u32x4<Self>; 4usize],
dest: &mut [u32; 16usize],
) -> () {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
vectors: [u32x4<Sse4_2>; 4usize],
dest: &mut [u32; 16usize],
) -> () {
let _ = token;
let v0: __m128i = vectors[0].into();
let v1: __m128i = vectors[1].into();
let v2: __m128i = vectors[2].into();
let v3: __m128i = vectors[3].into();
let tmp0 = _mm_unpacklo_epi32(v0, v1);
let tmp1 = _mm_unpackhi_epi32(v0, v1);
let tmp2 = _mm_unpacklo_epi32(v2, v3);
let tmp3 = _mm_unpackhi_epi32(v2, v3);
let out0 = _mm_unpacklo_epi64(tmp0, tmp2);
let out1 = _mm_unpackhi_epi64(tmp0, tmp2);
let out2 = _mm_unpacklo_epi64(tmp1, tmp3);
let out3 = _mm_unpackhi_epi64(tmp1, tmp3);
let (chunks, []) = dest.as_chunks_mut::<4usize>() else {
unreachable!()
};
crate::transmute::checked_transmute_store::<__m128i, [u32; 4usize]>(
out0,
&mut chunks[0],
);
crate::transmute::checked_transmute_store::<__m128i, [u32; 4usize]>(
out1,
&mut chunks[1],
);
crate::transmute::checked_transmute_store::<__m128i, [u32; 4usize]>(
out2,
&mut chunks[2],
);
crate::transmute::checked_transmute_store::<__m128i, [u32; 4usize]>(
out3,
&mut chunks[3],
);
}
);
kernel(self, vectors, dest);
}
#[inline(always)]
fn widen_u32x4(self, a: u32x4<Self>) -> (u64x2<Self>, u64x2<Self>) {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>) -> (u64x2<Sse4_2>, u64x2<Sse4_2>) {
let raw = a.into();
(
_mm_cvtepu32_epi64(raw).simd_into(token),
_mm_cvtepu32_epi64(_mm_srli_si128::<8>(raw)).simd_into(token),
)
}
);
kernel(self, a)
}
#[inline(always)]
fn narrow_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u16x8<Sse4_2> {
let mask = _mm_set1_epi32(0xffff);
_mm_packus_epi32(_mm_and_si128(a.into(), mask), _mm_and_si128(b.into(), mask))
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_narrow_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>, b: u32x4<Sse4_2>) -> u16x8<Sse4_2> {
let max = _mm_set1_epi32(u16::MAX as i32);
_mm_packus_epi32(_mm_min_epu32(a.into(), max), _mm_min_epu32(b.into(), max))
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn relaxed_narrow_u32x4(self, a: u32x4<Self>, b: u32x4<Self>) -> u16x8<Self> {
debug_assert!(
a.as_slice()
.iter()
.chain(b.as_slice())
.all(|&value| { value >= u16::MIN as u32 && value <= u16::MAX as u32 }),
"relaxed_narrow inputs must fit in the destination type",
);
self.narrow_u32x4(a, b)
}
#[inline(always)]
fn cvt_f32_u32x4(self, a: u32x4<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u32x4<Sse4_2>) -> f32x4<Sse4_2> {
let a = a.into();
let lo = _mm_blend_epi16::<0xAA>(a, _mm_set1_epi32(0x4B000000));
let hi =
_mm_blend_epi16::<0xAA>(_mm_srli_epi32::<16>(a), _mm_set1_epi32(0x53000000));
let fhi = _mm_sub_ps(
_mm_castsi128_ps(hi),
_mm_set1_ps(f32::from_bits(0x53000080)),
);
let result = _mm_add_ps(_mm_castsi128_ps(lo), fhi);
result.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn splat_mask32x4(self, val: bool) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, val: bool) -> mask32x4<Sse4_2> {
let val: i32 = if val { !0 } else { 0 };
_mm_set1_epi32(val).simd_into(token)
}
);
kernel(self, val)
}
#[inline(always)]
fn from_bitmask_mask32x4(self, bits: u64) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, bits: u64) -> mask32x4<Sse4_2> {
{
let bit_lanes = _mm_set1_epi32(bits as i32);
let bit_mask = _mm_setr_epi32(1, 2, 4, 8);
_mm_cmpeq_epi32(_mm_and_si128(bit_lanes, bit_mask), bit_mask)
}
.simd_into(token)
}
);
kernel(self, bits)
}
#[inline(always)]
fn to_bitmask_mask32x4(self, a: mask32x4<Self>) -> u64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask32x4<Sse4_2>) -> u64 {
_mm_movemask_ps(_mm_castsi128_ps(a.into())) as u32 as u64
}
);
kernel(self, a)
}
#[inline(always)]
fn from_vector_mask32x4(self, vector: i32x4<Self>) -> mask32x4<Self> {
mask32x4 {
val: vector.val,
simd: self,
}
}
#[inline(always)]
fn to_vector_mask32x4(self, a: mask32x4<Self>) -> i32x4<Self> {
i32x4 {
val: a.val,
simd: self,
}
}
#[inline(always)]
fn set_mask32x4(self, a: &mut mask32x4<Self>, index: usize, value: bool) -> () {
assert!(
index < 4usize,
"mask lane index {index} is out of bounds for {} lanes",
4usize
);
let mut lanes: [i32; 4usize] = (*a).into();
lanes[index] = if value { !0 } else { 0 };
*a = lanes.simd_into(self);
}
#[inline(always)]
fn rotate_elements_left_mask32x4<const OFFSET: usize>(
self,
a: mask32x4<Self>,
) -> mask32x4<Self> {
let int = i32x4 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_left::<OFFSET>();
mask32x4 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn rotate_elements_right_mask32x4<const OFFSET: usize>(
self,
a: mask32x4<Self>,
) -> mask32x4<Self> {
let int = i32x4 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_right::<OFFSET>();
mask32x4 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn and_mask32x4(self, a: mask32x4<Self>, b: mask32x4<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask32x4<Sse4_2>, b: mask32x4<Sse4_2>) -> mask32x4<Sse4_2> {
_mm_and_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn or_mask32x4(self, a: mask32x4<Self>, b: mask32x4<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask32x4<Sse4_2>, b: mask32x4<Sse4_2>) -> mask32x4<Sse4_2> {
_mm_or_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn xor_mask32x4(self, a: mask32x4<Self>, b: mask32x4<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask32x4<Sse4_2>, b: mask32x4<Sse4_2>) -> mask32x4<Sse4_2> {
_mm_xor_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn not_mask32x4(self, a: mask32x4<Self>) -> mask32x4<Self> {
self.xor_mask32x4(a, self.splat_mask32x4(true))
}
#[inline(always)]
fn reverse_mask32x4(self, a: mask32x4<Self>) -> mask32x4<Self> {
let lanes = i32x4 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let reversed = self.reverse_i32x4(lanes);
mask32x4 {
val: crate::transmute::checked_transmute_copy(&reversed.val),
simd: self,
}
}
#[inline(always)]
fn select_mask32x4(
self,
a: mask32x4<Self>,
b: mask32x4<Self>,
c: mask32x4<Self>,
) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: mask32x4<Sse4_2>,
b: mask32x4<Sse4_2>,
c: mask32x4<Sse4_2>,
) -> mask32x4<Sse4_2> {
_mm_blendv_epi8(c.into(), b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b, c)
}
#[inline(always)]
fn simd_eq_mask32x4(self, a: mask32x4<Self>, b: mask32x4<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask32x4<Sse4_2>, b: mask32x4<Sse4_2>) -> mask32x4<Sse4_2> {
_mm_cmpeq_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn any_true_mask32x4(self, a: mask32x4<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask32x4<Sse4_2>) -> bool {
_mm_movemask_ps(_mm_castsi128_ps(a.into())) as u32 != 0
}
);
kernel(self, a)
}
#[inline(always)]
fn all_true_mask32x4(self, a: mask32x4<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask32x4<Sse4_2>) -> bool {
_mm_movemask_ps(_mm_castsi128_ps(a.into())) as u32 == 0b1111
}
);
kernel(self, a)
}
#[inline(always)]
fn any_false_mask32x4(self, a: mask32x4<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask32x4<Sse4_2>) -> bool {
_mm_movemask_ps(_mm_castsi128_ps(a.into())) as u32 != 0b1111
}
);
kernel(self, a)
}
#[inline(always)]
fn all_false_mask32x4(self, a: mask32x4<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask32x4<Sse4_2>) -> bool {
_mm_movemask_ps(_mm_castsi128_ps(a.into())) as u32 == 0
}
);
kernel(self, a)
}
#[inline(always)]
fn combine_mask32x4(self, a: mask32x4<Self>, b: mask32x4<Self>) -> mask32x8<Self> {
mask32x8 {
val: crate::support::Aligned256([a.val.0, b.val.0]),
simd: self,
}
}
#[inline(always)]
fn widen_mask32x4(self, a: mask32x4<Self>) -> (mask64x2<Self>, mask64x2<Self>) {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask32x4<Sse4_2>) -> (mask64x2<Sse4_2>, mask64x2<Sse4_2>) {
let raw = a.into();
(
_mm_unpacklo_epi32(raw, raw).simd_into(token),
_mm_unpackhi_epi32(raw, raw).simd_into(token),
)
}
);
kernel(self, a)
}
#[inline(always)]
fn narrow_mask32x4(self, a: mask32x4<Self>, b: mask32x4<Self>) -> mask16x8<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask32x4<Sse4_2>, b: mask32x4<Sse4_2>) -> mask16x8<Sse4_2> {
_mm_packs_epi32(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn abs_f64x2(self, a: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_andnot_pd(_mm_set1_pd(-0.0), a.into()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn splat_f64x2(self, val: f64) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, val: f64) -> f64x2<Sse4_2> {
_mm_set1_pd(val).simd_into(token)
}
);
kernel(self, val)
}
#[inline(always)]
fn reverse_f64x2(self, a: f64x2<Self>) -> f64x2<Self> {
let indices: u8x16<Self> =
[8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7].simd_into(self);
self.swizzle_dyn_f64x2(a, indices)
}
#[inline(always)]
fn slide_f64x2<const SHIFT: usize>(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
if SHIFT >= 2usize {
return b;
}
let result = dyn_alignr_128(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 8usize,
);
Bytes::from_bytes(u8x16 {
val: crate::support::Aligned128(result),
simd: self,
})
}
#[inline(always)]
fn compress_f64x2(self, values: f64x2<Self>, mask: mask64x2<Self>) -> f64x2<Self> {
let bits = self.to_bitmask_mask64x2(mask) as usize & 3;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_64[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn compress_merge_f64x2(
self,
values: f64x2<Self>,
mask: mask64x2<Self>,
merge: f64x2<Self>,
) -> f64x2<Self> {
let bits = self.to_bitmask_mask64x2(mask) as usize & 3;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_64[bits].0);
let compressed = self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control);
let inactive = self.simd_lt_i8x16(i8x16::from_bytes(control), i8x16::splat(self, 0));
Bytes::from_bytes(self.select_u8x16(inactive, Bytes::to_bytes(merge), compressed))
}
#[inline(always)]
fn expand_f64x2(self, values: f64x2<Self>, mask: mask64x2<Self>) -> f64x2<Self> {
let bits = self.to_bitmask_mask64x2(mask) as usize & 3;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_64[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn expand_merge_f64x2(
self,
values: f64x2<Self>,
mask: mask64x2<Self>,
merge: f64x2<Self>,
) -> f64x2<Self> {
let bits = self.to_bitmask_mask64x2(mask) as usize & 3;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_64[bits].0);
let expanded = Bytes::from_bytes(self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control));
self.select_f64x2(mask, expanded, merge)
}
#[inline(always)]
fn neg_f64x2(self, a: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_xor_pd(a.into(), _mm_set1_pd(-0.0)).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn sqrt_f64x2(self, a: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_sqrt_pd(a.into()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn approximate_recip_f64x2(self, a: f64x2<Self>) -> f64x2<Self> {
1.0 / a
}
#[inline(always)]
fn add_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_add_pd(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn sub_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_sub_pd(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn mul_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_mul_pd(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn div_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_div_pd(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn copysign_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
let mask = _mm_set1_pd(-0.0);
_mm_or_pd(_mm_and_pd(mask, b.into()), _mm_andnot_pd(mask, a.into()))
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn reduce_max_f64x2(self, a: f64x2<Self>) -> f64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>) -> f64 {
let reduced: __m128d = a.into();
let shifted = _mm_castsi128_pd(_mm_srli_si128::<8>(_mm_castpd_si128(reduced)));
let reduced = { _mm_max_pd(reduced, shifted) };
_mm_cvtsd_f64(reduced)
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_min_f64x2(self, a: f64x2<Self>) -> f64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>) -> f64 {
let reduced: __m128d = a.into();
let shifted = _mm_castsi128_pd(_mm_srli_si128::<8>(_mm_castpd_si128(reduced)));
let reduced = { _mm_min_pd(reduced, shifted) };
_mm_cvtsd_f64(reduced)
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_max_precise_f64x2(self, a: f64x2<Self>) -> f64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>) -> f64 {
let reduced: __m128d = a.into();
let shifted = _mm_castsi128_pd(_mm_srli_si128::<8>(_mm_castpd_si128(reduced)));
let reduced = {
let intermediate = _mm_max_pd(reduced, shifted);
let b_is_nan = _mm_cmpunord_pd(shifted, shifted);
_mm_blendv_pd(intermediate, reduced, b_is_nan)
};
_mm_cvtsd_f64(reduced)
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_min_precise_f64x2(self, a: f64x2<Self>) -> f64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>) -> f64 {
let reduced: __m128d = a.into();
let shifted = _mm_castsi128_pd(_mm_srli_si128::<8>(_mm_castpd_si128(reduced)));
let reduced = {
let intermediate = _mm_min_pd(reduced, shifted);
let b_is_nan = _mm_cmpunord_pd(shifted, shifted);
_mm_blendv_pd(intermediate, reduced, b_is_nan)
};
_mm_cvtsd_f64(reduced)
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_sum_f64x2(self, a: f64x2<Self>) -> f64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>) -> f64 {
let a: __m128d = a.into();
_mm_cvtsd_f64(_mm_add_sd(a, _mm_unpackhi_pd(a, a)))
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_product_f64x2(self, a: f64x2<Self>) -> f64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>) -> f64 {
let a: __m128d = a.into();
_mm_cvtsd_f64(_mm_mul_sd(a, _mm_unpackhi_pd(a, a)))
}
);
kernel(self, a)
}
#[inline(always)]
fn max_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_max_pd(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn min_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_min_pd(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn max_precise_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
let intermediate = _mm_max_pd(a.into(), b.into());
let b_is_nan = _mm_cmpunord_pd(b.into(), b.into());
_mm_blendv_pd(intermediate, a.into(), b_is_nan).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn min_precise_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
let intermediate = _mm_min_pd(a.into(), b.into());
let b_is_nan = _mm_cmpunord_pd(b.into(), b.into());
_mm_blendv_pd(intermediate, a.into(), b_is_nan).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_eq_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> mask64x2<Sse4_2> {
_mm_castpd_si128(_mm_cmpeq_pd(a.into(), b.into())).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_ne_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> mask64x2<Sse4_2> {
_mm_castpd_si128(_mm_cmpneq_pd(a.into(), b.into())).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_lt_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> mask64x2<Sse4_2> {
_mm_castpd_si128(_mm_cmplt_pd(a.into(), b.into())).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_le_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> mask64x2<Sse4_2> {
_mm_castpd_si128(_mm_cmple_pd(a.into(), b.into())).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_low_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_unpacklo_pd(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_high_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_unpackhi_pd(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_low_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_shuffle_pd::<0b00>(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_high_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_shuffle_pd::<0b11>(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn interleave_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> (f64x2<Self>, f64x2<Self>) {
(self.zip_low_f64x2(a, b), self.zip_high_f64x2(a, b))
}
#[inline(always)]
fn deinterleave_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> (f64x2<Self>, f64x2<Self>) {
(self.unzip_low_f64x2(a, b), self.unzip_high_f64x2(a, b))
}
#[inline(always)]
fn mul_add_f64x2(self, a: f64x2<Self>, b: f64x2<Self>, c: f64x2<Self>) -> f64x2<Self> {
a * b + c
}
#[inline(always)]
fn mul_add_precise_f64x2(self, a: f64x2<Self>, b: f64x2<Self>, c: f64x2<Self>) -> f64x2<Self> {
crate::math::mul_add_precise_f64x2(self, a, b, c)
}
#[inline(always)]
fn mul_sub_f64x2(self, a: f64x2<Self>, b: f64x2<Self>, c: f64x2<Self>) -> f64x2<Self> {
a * b - c
}
#[inline(always)]
fn mul_sub_precise_f64x2(self, a: f64x2<Self>, b: f64x2<Self>, c: f64x2<Self>) -> f64x2<Self> {
self.mul_add_precise_f64x2(a, b, -c)
}
#[inline(always)]
fn floor_f64x2(self, a: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_round_pd::<{ _MM_FROUND_TO_NEG_INF | _MM_FROUND_NO_EXC }>(a.into())
.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn ceil_f64x2(self, a: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_round_pd::<{ _MM_FROUND_TO_POS_INF | _MM_FROUND_NO_EXC }>(a.into())
.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn round_ties_even_f64x2(self, a: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_round_pd::<{ _MM_FROUND_TO_NEAREST_INT | _MM_FROUND_NO_EXC }>(a.into())
.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn round_f64x2(self, a: f64x2<Self>) -> f64x2<Self> {
let bias = self.splat_f64x2(const { f64::next_down(0.5) });
(a + self.copysign_f64x2(bias, a)).trunc()
}
#[inline(always)]
fn fract_f64x2(self, a: f64x2<Self>) -> f64x2<Self> {
a - self.trunc_f64x2(a)
}
#[inline(always)]
fn trunc_f64x2(self, a: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>) -> f64x2<Sse4_2> {
_mm_round_pd::<{ _MM_FROUND_TO_ZERO | _MM_FROUND_NO_EXC }>(a.into())
.simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn to_degrees_f64x2(self, a: f64x2<Self>) -> f64x2<Self> {
a * const { f64::to_degrees(1.0) }
}
#[inline(always)]
fn to_radians_f64x2(self, a: f64x2<Self>) -> f64x2<Self> {
a * const { f64::to_radians(1.0) }
}
#[inline(always)]
fn is_nan_f64x2(self, a: f64x2<Self>) -> mask64x2<Self> {
a.simd_ne(a)
}
#[inline(always)]
fn is_infinite_f64x2(self, a: f64x2<Self>) -> mask64x2<Self> {
a.abs().simd_eq(f64::INFINITY)
}
#[inline(always)]
fn is_finite_f64x2(self, a: f64x2<Self>) -> mask64x2<Self> {
a.abs().simd_lt(f64::INFINITY)
}
#[inline(always)]
fn is_subnormal_f64x2(self, a: f64x2<Self>) -> mask64x2<Self> {
let i: i64x2<Self> = a.bitcast();
(i & 0x7FF0_0000_0000_0000_i64).simd_eq(0) & (i & 0x000F_FFFF_FFFF_FFFF_i64).simd_ne(0)
}
#[inline(always)]
fn is_normal_f64x2(self, a: f64x2<Self>) -> mask64x2<Self> {
let i: i64x2<Self> = a.bitcast();
let exp = i & 0x7FF0_0000_0000_0000_i64;
exp.simd_ne(0) & exp.simd_ne(0x7FF0_0000_0000_0000_i64)
}
#[inline(always)]
fn is_sign_positive_f64x2(self, a: f64x2<Self>) -> mask64x2<Self> {
let i: i64x2<Self> = a.bitcast();
(i & i64::MIN).simd_eq(0)
}
#[inline(always)]
fn is_sign_negative_f64x2(self, a: f64x2<Self>) -> mask64x2<Self> {
!a.is_sign_positive()
}
#[inline(always)]
fn select_f64x2(self, a: mask64x2<Self>, b: f64x2<Self>, c: f64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: mask64x2<Sse4_2>,
b: f64x2<Sse4_2>,
c: f64x2<Sse4_2>,
) -> f64x2<Sse4_2> {
_mm_blendv_pd(c.into(), b.into(), _mm_castsi128_pd(a.into())).simd_into(token)
}
);
kernel(self, a, b, c)
}
#[inline(always)]
fn combine_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f64x4<Self> {
f64x4 {
val: crate::support::Aligned256([a.val.0, b.val.0]),
simd: self,
}
}
#[inline(always)]
fn load_four_interleaved_f64x2(self, src: &[f64; 8usize]) -> [f64x2<Self>; 4usize] {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, src: &[f64; 8usize]) -> [f64x2<Sse4_2>; 4usize] {
let (chunks, []) = src.as_chunks::<2usize>() else {
unreachable!()
};
let v0: __m128d =
crate::transmute::checked_transmute_copy::<[f64; 2usize], __m128d>(&chunks[0]);
let v1: __m128d =
crate::transmute::checked_transmute_copy::<[f64; 2usize], __m128d>(&chunks[1]);
let v2: __m128d =
crate::transmute::checked_transmute_copy::<[f64; 2usize], __m128d>(&chunks[2]);
let v3: __m128d =
crate::transmute::checked_transmute_copy::<[f64; 2usize], __m128d>(&chunks[3]);
let out0 = _mm_unpacklo_pd(v0, v2);
let out1 = _mm_unpackhi_pd(v0, v2);
let out2 = _mm_unpacklo_pd(v1, v3);
let out3 = _mm_unpackhi_pd(v1, v3);
[
out0.simd_into(token),
out1.simd_into(token),
out2.simd_into(token),
out3.simd_into(token),
]
}
);
kernel(self, src)
}
#[inline(always)]
fn store_four_interleaved_f64x2(
self,
vectors: [f64x2<Self>; 4usize],
dest: &mut [f64; 8usize],
) -> () {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
vectors: [f64x2<Sse4_2>; 4usize],
dest: &mut [f64; 8usize],
) -> () {
let _ = token;
let v0: __m128d = vectors[0].into();
let v1: __m128d = vectors[1].into();
let v2: __m128d = vectors[2].into();
let v3: __m128d = vectors[3].into();
let out0 = _mm_unpacklo_pd(v0, v1);
let out1 = _mm_unpacklo_pd(v2, v3);
let out2 = _mm_unpackhi_pd(v0, v1);
let out3 = _mm_unpackhi_pd(v2, v3);
let (chunks, []) = dest.as_chunks_mut::<2usize>() else {
unreachable!()
};
crate::transmute::checked_transmute_store::<__m128d, [f64; 2usize]>(
out0,
&mut chunks[0],
);
crate::transmute::checked_transmute_store::<__m128d, [f64; 2usize]>(
out1,
&mut chunks[1],
);
crate::transmute::checked_transmute_store::<__m128d, [f64; 2usize]>(
out2,
&mut chunks[2],
);
crate::transmute::checked_transmute_store::<__m128d, [f64; 2usize]>(
out3,
&mut chunks[3],
);
}
);
kernel(self, vectors, dest);
}
#[inline(always)]
fn narrow_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>, b: f64x2<Sse4_2>) -> f32x4<Sse4_2> {
let low = _mm_cvtpd_ps(a.into());
let high = _mm_cvtpd_ps(b.into());
_mm_movelh_ps(low, high).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_narrow_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f32x4<Self> {
self.narrow_f64x2(a, b)
}
#[inline(always)]
fn relaxed_narrow_f64x2(self, a: f64x2<Self>, b: f64x2<Self>) -> f32x4<Self> {
self.narrow_f64x2(a, b)
}
#[inline(always)]
fn cvt_u64_f64x2(self, a: f64x2<Self>) -> u64x2<Self> {
[a[0usize] as u64, a[1usize] as u64].simd_into(self)
}
#[inline(always)]
fn cvt_u64_precise_f64x2(self, a: f64x2<Self>) -> u64x2<Self> {
[a[0usize] as u64, a[1usize] as u64].simd_into(self)
}
#[inline(always)]
#[cfg(target_arch = "x86_64")]
fn cvt_i64_f64x2(self, a: f64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: f64x2<Sse4_2>) -> i64x2<Sse4_2> {
let a = a.into();
let low = _mm_cvttsd_si64(a);
let high = _mm_cvttsd_si64(_mm_unpackhi_pd(a, a));
_mm_set_epi64x(high, low).simd_into(token)
}
);
kernel(self, a)
}
#[cfg(target_arch = "x86")]
fn cvt_i64_f64x2(self, a: f64x2<Self>) -> i64x2<Self> {
[a[0usize] as i64, a[1usize] as i64].simd_into(self)
}
#[inline(always)]
fn cvt_i64_precise_f64x2(self, a: f64x2<Self>) -> i64x2<Self> {
[a[0usize] as i64, a[1usize] as i64].simd_into(self)
}
#[inline(always)]
fn abs_i64x2(self, a: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
let a = a.into();
let mask = _mm_cmpgt_epi64(_mm_setzero_si128(), a);
_mm_sub_epi64(_mm_xor_si128(a, mask), mask).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn splat_i64x2(self, val: i64) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, val: i64) -> i64x2<Sse4_2> {
_mm_set1_epi64x(val).simd_into(token)
}
);
kernel(self, val)
}
#[inline(always)]
fn reverse_i64x2(self, a: i64x2<Self>) -> i64x2<Self> {
let indices: u8x16<Self> =
[8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7].simd_into(self);
self.swizzle_dyn_i64x2(a, indices)
}
#[inline(always)]
fn slide_i64x2<const SHIFT: usize>(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
if SHIFT >= 2usize {
return b;
}
let result = dyn_alignr_128(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 8usize,
);
Bytes::from_bytes(u8x16 {
val: crate::support::Aligned128(result),
simd: self,
})
}
#[inline(always)]
fn compress_i64x2(self, values: i64x2<Self>, mask: mask64x2<Self>) -> i64x2<Self> {
let bits = self.to_bitmask_mask64x2(mask) as usize & 3;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_64[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn compress_merge_i64x2(
self,
values: i64x2<Self>,
mask: mask64x2<Self>,
merge: i64x2<Self>,
) -> i64x2<Self> {
let bits = self.to_bitmask_mask64x2(mask) as usize & 3;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_64[bits].0);
let compressed = self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control);
let inactive = self.simd_lt_i8x16(i8x16::from_bytes(control), i8x16::splat(self, 0));
Bytes::from_bytes(self.select_u8x16(inactive, Bytes::to_bytes(merge), compressed))
}
#[inline(always)]
fn expand_i64x2(self, values: i64x2<Self>, mask: mask64x2<Self>) -> i64x2<Self> {
let bits = self.to_bitmask_mask64x2(mask) as usize & 3;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_64[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn expand_merge_i64x2(
self,
values: i64x2<Self>,
mask: mask64x2<Self>,
merge: i64x2<Self>,
) -> i64x2<Self> {
let bits = self.to_bitmask_mask64x2(mask) as usize & 3;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_64[bits].0);
let expanded = Bytes::from_bytes(self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control));
self.select_i64x2(mask, expanded, merge)
}
#[inline(always)]
fn count_ones_i64x2(self, a: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
let value = a.into();
let nibble_mask = _mm_set1_epi8(0x0f);
let lookup = _mm_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);
let low = _mm_and_si128(value, nibble_mask);
let high = _mm_and_si128(_mm_srli_epi16::<4>(value), nibble_mask);
let byte_counts = _mm_add_epi8(
_mm_shuffle_epi8(lookup, low),
_mm_shuffle_epi8(lookup, high),
);
_mm_sad_epu8(byte_counts, _mm_setzero_si128()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn count_zeros_i64x2(self, a: i64x2<Self>) -> i64x2<Self> {
self.count_ones_i64x2(self.not_i64x2(a))
}
#[inline(always)]
fn add_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
_mm_add_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_add_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
let a = a.into();
let b = b.into();
let wrapped = _mm_add_epi64(a, b);
let overflow = _mm_xor_si128(_mm_cmpgt_epi64(a, wrapped), b);
let a_sign = _mm_srai_epi32::<31>(_mm_shuffle_epi32::<0xf5>(a));
let bound = _mm_xor_si128(a_sign, _mm_set1_epi64x(i64::MAX));
let result = _mm_blendv_pd(
_mm_castsi128_pd(wrapped),
_mm_castsi128_pd(bound),
_mm_castsi128_pd(overflow),
);
_mm_castpd_si128(result).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn sub_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
_mm_sub_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_sub_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
let a = a.into();
let b = b.into();
let wrapped = _mm_sub_epi64(a, b);
let overflow = _mm_xor_si128(_mm_cmpgt_epi64(wrapped, a), b);
let a_sign = _mm_srai_epi32::<31>(_mm_shuffle_epi32::<0xf5>(a));
let bound = _mm_xor_si128(a_sign, _mm_set1_epi64x(i64::MAX));
let result = _mm_blendv_pd(
_mm_castsi128_pd(wrapped),
_mm_castsi128_pd(bound),
_mm_castsi128_pd(overflow),
);
_mm_castpd_si128(result).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn mul_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
let a = a.into();
let b = b.into();
let a_high = _mm_shuffle_epi32::<0xf5>(a);
let b_high = _mm_shuffle_epi32::<0xf5>(b);
let cross = _mm_add_epi64(_mm_mul_epu32(a_high, b), _mm_mul_epu32(a, b_high));
let low = _mm_mul_epu32(a, b);
_mm_add_epi64(low, _mm_slli_epi64::<32>(cross)).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn and_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
_mm_and_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn or_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
_mm_or_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn xor_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
_mm_xor_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn not_i64x2(self, a: i64x2<Self>) -> i64x2<Self> {
a ^ !0
}
#[inline(always)]
fn shl_i64x2(self, a: i64x2<Self>, shift: u32) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, shift: u32) -> i64x2<Sse4_2> {
_mm_sll_epi64(a.into(), _mm_cvtsi32_si128(shift.cast_signed())).simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shlv_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
[
i64::wrapping_shl(a[0usize], b[0usize] as u32),
i64::wrapping_shl(a[1usize], b[1usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn shr_i64x2(self, a: i64x2<Self>, shift: u32) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, shift: u32) -> i64x2<Sse4_2> {
let value = a.into();
let count = _mm_cvtsi32_si128(shift.cast_signed());
let shifted_bias = _mm_srl_epi64(_mm_set1_epi64x(i64::MIN), count);
let shifted = _mm_srl_epi64(value, count);
_mm_sub_epi64(_mm_xor_si128(shifted, shifted_bias), shifted_bias).simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shrv_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
[
i64::wrapping_shr(a[0usize], b[0usize] as u32),
i64::wrapping_shr(a[1usize], b[1usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn reduce_max_i64x2(self, a: i64x2<Self>) -> i64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>) -> i64 {
let lanes: [i64; 2] = a.into();
lanes[0].max(lanes[1])
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_min_i64x2(self, a: i64x2<Self>) -> i64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>) -> i64 {
let lanes: [i64; 2] = a.into();
lanes[0].min(lanes[1])
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_sum_i64x2(self, a: i64x2<Self>) -> i64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>) -> i64 {
let sum: __m128i = a.into();
let sum = _mm_add_epi64(sum, _mm_srli_si128::<8>(sum));
let lanes: [i64; 2usize] = crate::transmute::checked_transmute_copy(&sum);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_product_i64x2(self, a: i64x2<Self>) -> i64 {
let mul_level_0: [i64; 1usize] = [a[0usize].wrapping_mul(a[1usize])];
mul_level_0[0]
}
#[inline(always)]
fn max_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
let a = a.into();
let b = b.into();
let mask = _mm_cmpgt_epi64(b, a);
let result = _mm_blendv_pd(
_mm_castsi128_pd(a),
_mm_castsi128_pd(b),
_mm_castsi128_pd(mask),
);
_mm_castpd_si128(result).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn min_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
let a = a.into();
let b = b.into();
let mask = _mm_cmpgt_epi64(a, b);
let result = _mm_blendv_pd(
_mm_castsi128_pd(a),
_mm_castsi128_pd(b),
_mm_castsi128_pd(mask),
);
_mm_castpd_si128(result).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_eq_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> mask64x2<Sse4_2> {
_mm_cmpeq_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_lt_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> mask64x2<Sse4_2> {
_mm_cmpgt_epi64(b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_le_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> mask64x2<Sse4_2> {
_mm_xor_si128(_mm_cmpgt_epi64(a.into(), b.into()), _mm_set1_epi64x(-1))
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_low_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
_mm_unpacklo_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_high_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
_mm_unpackhi_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_low_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
_mm_unpacklo_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_high_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
_mm_unpackhi_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn interleave_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> (i64x2<Self>, i64x2<Self>) {
(self.zip_low_i64x2(a, b), self.zip_high_i64x2(a, b))
}
#[inline(always)]
fn deinterleave_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> (i64x2<Self>, i64x2<Self>) {
(self.unzip_low_i64x2(a, b), self.unzip_high_i64x2(a, b))
}
#[inline(always)]
fn select_i64x2(self, a: mask64x2<Self>, b: i64x2<Self>, c: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: mask64x2<Sse4_2>,
b: i64x2<Sse4_2>,
c: i64x2<Sse4_2>,
) -> i64x2<Sse4_2> {
_mm_blendv_epi8(c.into(), b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b, c)
}
#[inline(always)]
fn combine_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i64x4<Self> {
i64x4 {
val: crate::support::Aligned256([a.val.0, b.val.0]),
simd: self,
}
}
#[inline(always)]
fn neg_i64x2(self, a: i64x2<Self>) -> i64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>) -> i64x2<Sse4_2> {
_mm_sub_epi64(_mm_setzero_si128(), a.into()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn load_four_interleaved_i64x2(self, src: &[i64; 8usize]) -> [i64x2<Self>; 4usize] {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, src: &[i64; 8usize]) -> [i64x2<Sse4_2>; 4usize] {
let (chunks, []) = src.as_chunks::<2usize>() else {
unreachable!()
};
let v0: __m128i =
crate::transmute::checked_transmute_copy::<[i64; 2usize], __m128i>(&chunks[0]);
let v1: __m128i =
crate::transmute::checked_transmute_copy::<[i64; 2usize], __m128i>(&chunks[1]);
let v2: __m128i =
crate::transmute::checked_transmute_copy::<[i64; 2usize], __m128i>(&chunks[2]);
let v3: __m128i =
crate::transmute::checked_transmute_copy::<[i64; 2usize], __m128i>(&chunks[3]);
let out0 = _mm_unpacklo_epi64(v0, v2);
let out1 = _mm_unpackhi_epi64(v0, v2);
let out2 = _mm_unpacklo_epi64(v1, v3);
let out3 = _mm_unpackhi_epi64(v1, v3);
[
out0.simd_into(token),
out1.simd_into(token),
out2.simd_into(token),
out3.simd_into(token),
]
}
);
kernel(self, src)
}
#[inline(always)]
fn store_four_interleaved_i64x2(
self,
vectors: [i64x2<Self>; 4usize],
dest: &mut [i64; 8usize],
) -> () {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
vectors: [i64x2<Sse4_2>; 4usize],
dest: &mut [i64; 8usize],
) -> () {
let _ = token;
let v0: __m128i = vectors[0].into();
let v1: __m128i = vectors[1].into();
let v2: __m128i = vectors[2].into();
let v3: __m128i = vectors[3].into();
let out0 = _mm_unpacklo_epi64(v0, v1);
let out1 = _mm_unpacklo_epi64(v2, v3);
let out2 = _mm_unpackhi_epi64(v0, v1);
let out3 = _mm_unpackhi_epi64(v2, v3);
let (chunks, []) = dest.as_chunks_mut::<2usize>() else {
unreachable!()
};
crate::transmute::checked_transmute_store::<__m128i, [i64; 2usize]>(
out0,
&mut chunks[0],
);
crate::transmute::checked_transmute_store::<__m128i, [i64; 2usize]>(
out1,
&mut chunks[1],
);
crate::transmute::checked_transmute_store::<__m128i, [i64; 2usize]>(
out2,
&mut chunks[2],
);
crate::transmute::checked_transmute_store::<__m128i, [i64; 2usize]>(
out3,
&mut chunks[3],
);
}
);
kernel(self, vectors, dest);
}
#[inline(always)]
fn narrow_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i32x4<Sse4_2> {
_mm_castps_si128(_mm_shuffle_ps::<0x88>(
_mm_castsi128_ps(a.into()),
_mm_castsi128_ps(b.into()),
))
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_narrow_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>, b: i64x2<Sse4_2>) -> i32x4<Sse4_2> {
let a = a.into();
let b = b.into();
let low = _mm_castps_si128(_mm_shuffle_ps::<0x88>(
_mm_castsi128_ps(a),
_mm_castsi128_ps(b),
));
let high = _mm_castps_si128(_mm_shuffle_ps::<0xdd>(
_mm_castsi128_ps(a),
_mm_castsi128_ps(b),
));
let low_sign = _mm_srai_epi32::<31>(low);
let fits = _mm_cmpeq_epi32(high, low_sign);
let high_sign = _mm_srai_epi32::<31>(high);
let bound = _mm_xor_si128(high_sign, _mm_set1_epi32(i32::MAX));
_mm_blendv_epi8(bound, low, fits).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn relaxed_narrow_i64x2(self, a: i64x2<Self>, b: i64x2<Self>) -> i32x4<Self> {
debug_assert!(
a.as_slice()
.iter()
.chain(b.as_slice())
.all(|&value| { value >= i32::MIN as i64 && value <= i32::MAX as i64 }),
"relaxed_narrow inputs must fit in the destination type",
);
self.narrow_i64x2(a, b)
}
#[inline(always)]
fn cvt_f64_i64x2(self, a: i64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: i64x2<Sse4_2>) -> f64x2<Sse4_2> {
let a = a.into();
let low = _mm_blend_epi16::<0xcc>(a, _mm_set1_epi64x(0x4330_0000_0000_0000));
let high = _mm_srli_epi64::<32>(a);
let high = _mm_xor_si128(high, _mm_set1_epi64x(4985484789646622720i64));
let high = _mm_sub_pd(
_mm_castsi128_pd(high),
_mm_set1_pd(f64::from_bits(4985484789647671296u64)),
);
_mm_add_pd(_mm_castsi128_pd(low), high).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn splat_u64x2(self, val: u64) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, val: u64) -> u64x2<Sse4_2> {
_mm_set1_epi64x(val.cast_signed()).simd_into(token)
}
);
kernel(self, val)
}
#[inline(always)]
fn reverse_u64x2(self, a: u64x2<Self>) -> u64x2<Self> {
let indices: u8x16<Self> =
[8, 9, 10, 11, 12, 13, 14, 15, 0, 1, 2, 3, 4, 5, 6, 7].simd_into(self);
self.swizzle_dyn_u64x2(a, indices)
}
#[inline(always)]
fn slide_u64x2<const SHIFT: usize>(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
if SHIFT >= 2usize {
return b;
}
let result = dyn_alignr_128(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 8usize,
);
Bytes::from_bytes(u8x16 {
val: crate::support::Aligned128(result),
simd: self,
})
}
#[inline(always)]
fn compress_u64x2(self, values: u64x2<Self>, mask: mask64x2<Self>) -> u64x2<Self> {
let bits = self.to_bitmask_mask64x2(mask) as usize & 3;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_64[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn compress_merge_u64x2(
self,
values: u64x2<Self>,
mask: mask64x2<Self>,
merge: u64x2<Self>,
) -> u64x2<Self> {
let bits = self.to_bitmask_mask64x2(mask) as usize & 3;
let control = u8x16::simd_from(self, crate::support::compact_128::COMPRESS_64[bits].0);
let compressed = self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control);
let inactive = self.simd_lt_i8x16(i8x16::from_bytes(control), i8x16::splat(self, 0));
Bytes::from_bytes(self.select_u8x16(inactive, Bytes::to_bytes(merge), compressed))
}
#[inline(always)]
fn expand_u64x2(self, values: u64x2<Self>, mask: mask64x2<Self>) -> u64x2<Self> {
let bits = self.to_bitmask_mask64x2(mask) as usize & 3;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_64[bits].0);
Bytes::from_bytes(self.swizzle_dyn_precise_u8x16(Bytes::to_bytes(values), control))
}
#[inline(always)]
fn expand_merge_u64x2(
self,
values: u64x2<Self>,
mask: mask64x2<Self>,
merge: u64x2<Self>,
) -> u64x2<Self> {
let bits = self.to_bitmask_mask64x2(mask) as usize & 3;
let control = u8x16::simd_from(self, crate::support::compact_128::EXPAND_64[bits].0);
let expanded = Bytes::from_bytes(self.swizzle_dyn_u8x16(Bytes::to_bytes(values), control));
self.select_u64x2(mask, expanded, merge)
}
#[inline(always)]
fn count_ones_u64x2(self, a: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
let value = a.into();
let nibble_mask = _mm_set1_epi8(0x0f);
let lookup = _mm_setr_epi8(0, 1, 1, 2, 1, 2, 2, 3, 1, 2, 2, 3, 2, 3, 3, 4);
let low = _mm_and_si128(value, nibble_mask);
let high = _mm_and_si128(_mm_srli_epi16::<4>(value), nibble_mask);
let byte_counts = _mm_add_epi8(
_mm_shuffle_epi8(lookup, low),
_mm_shuffle_epi8(lookup, high),
);
_mm_sad_epu8(byte_counts, _mm_setzero_si128()).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn count_zeros_u64x2(self, a: u64x2<Self>) -> u64x2<Self> {
self.count_ones_u64x2(self.not_u64x2(a))
}
#[inline(always)]
fn add_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
_mm_add_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_add_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
let a = a.into();
let b = b.into();
let wrapped = _mm_add_epi64(a, b);
let sign_bias = _mm_set1_epi64x(i64::MIN);
let overflow = _mm_cmpgt_epi64(
_mm_xor_si128(a, sign_bias),
_mm_xor_si128(wrapped, sign_bias),
);
_mm_or_si128(wrapped, overflow).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn sub_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
_mm_sub_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_sub_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
let a = a.into();
let b = b.into();
let wrapped = _mm_sub_epi64(a, b);
let sign_bias = _mm_set1_epi64x(i64::MIN);
let no_borrow =
_mm_cmpgt_epi64(_mm_xor_si128(a, sign_bias), _mm_xor_si128(b, sign_bias));
_mm_and_si128(wrapped, no_borrow).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn mul_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
let a = a.into();
let b = b.into();
let a_high = _mm_shuffle_epi32::<0xf5>(a);
let b_high = _mm_shuffle_epi32::<0xf5>(b);
let cross = _mm_add_epi64(_mm_mul_epu32(a_high, b), _mm_mul_epu32(a, b_high));
let low = _mm_mul_epu32(a, b);
_mm_add_epi64(low, _mm_slli_epi64::<32>(cross)).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn and_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
_mm_and_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn or_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
_mm_or_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn xor_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
_mm_xor_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn not_u64x2(self, a: u64x2<Self>) -> u64x2<Self> {
a ^ !0
}
#[inline(always)]
fn shl_u64x2(self, a: u64x2<Self>, shift: u32) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, shift: u32) -> u64x2<Sse4_2> {
_mm_sll_epi64(a.into(), _mm_cvtsi32_si128(shift.cast_signed())).simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shlv_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
[
u64::wrapping_shl(a[0usize], b[0usize] as u32),
u64::wrapping_shl(a[1usize], b[1usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn shr_u64x2(self, a: u64x2<Self>, shift: u32) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, shift: u32) -> u64x2<Sse4_2> {
_mm_srl_epi64(a.into(), _mm_cvtsi32_si128(shift.cast_signed())).simd_into(token)
}
);
kernel(self, a, shift)
}
#[inline(always)]
fn shrv_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
[
u64::wrapping_shr(a[0usize], b[0usize] as u32),
u64::wrapping_shr(a[1usize], b[1usize] as u32),
]
.simd_into(self)
}
#[inline(always)]
fn reduce_max_u64x2(self, a: u64x2<Self>) -> u64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>) -> u64 {
let lanes: [u64; 2] = a.into();
lanes[0].max(lanes[1])
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_min_u64x2(self, a: u64x2<Self>) -> u64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>) -> u64 {
let lanes: [u64; 2] = a.into();
lanes[0].min(lanes[1])
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_sum_u64x2(self, a: u64x2<Self>) -> u64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>) -> u64 {
let sum: __m128i = a.into();
let sum = _mm_add_epi64(sum, _mm_srli_si128::<8>(sum));
let lanes: [u64; 2usize] = crate::transmute::checked_transmute_copy(&sum);
lanes[0]
}
);
kernel(self, a)
}
#[inline(always)]
fn reduce_product_u64x2(self, a: u64x2<Self>) -> u64 {
let mul_level_0: [u64; 1usize] = [a[0usize].wrapping_mul(a[1usize])];
mul_level_0[0]
}
#[inline(always)]
fn max_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
let a = a.into();
let b = b.into();
let mask = {
let sign_bit = _mm_set1_epi64x(0x8000000000000000u64.cast_signed());
let lhs_signed = _mm_xor_si128(b, sign_bit);
let rhs_signed = _mm_xor_si128(a, sign_bit);
_mm_cmpgt_epi64(lhs_signed, rhs_signed)
};
let result = _mm_blendv_pd(
_mm_castsi128_pd(a),
_mm_castsi128_pd(b),
_mm_castsi128_pd(mask),
);
_mm_castpd_si128(result).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn min_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
let a = a.into();
let b = b.into();
let mask = {
let sign_bit = _mm_set1_epi64x(0x8000000000000000u64.cast_signed());
let lhs_signed = _mm_xor_si128(a, sign_bit);
let rhs_signed = _mm_xor_si128(b, sign_bit);
_mm_cmpgt_epi64(lhs_signed, rhs_signed)
};
let result = _mm_blendv_pd(
_mm_castsi128_pd(a),
_mm_castsi128_pd(b),
_mm_castsi128_pd(mask),
);
_mm_castpd_si128(result).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_eq_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> mask64x2<Sse4_2> {
_mm_cmpeq_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_lt_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> mask64x2<Sse4_2> {
{
let sign_bit = _mm_set1_epi64x(0x8000000000000000u64.cast_signed());
let lhs_signed = _mm_xor_si128(b.into(), sign_bit);
let rhs_signed = _mm_xor_si128(a.into(), sign_bit);
_mm_cmpgt_epi64(lhs_signed, rhs_signed)
}
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn simd_le_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> mask64x2<Sse4_2> {
_mm_xor_si128(
{
let sign_bit = _mm_set1_epi64x(0x8000000000000000u64.cast_signed());
let lhs_signed = _mm_xor_si128(a.into(), sign_bit);
let rhs_signed = _mm_xor_si128(b.into(), sign_bit);
_mm_cmpgt_epi64(lhs_signed, rhs_signed)
},
_mm_set1_epi64x(-1),
)
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_low_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
_mm_unpacklo_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn zip_high_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
_mm_unpackhi_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_low_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
_mm_unpacklo_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn unzip_high_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u64x2<Sse4_2> {
_mm_unpackhi_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn interleave_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> (u64x2<Self>, u64x2<Self>) {
(self.zip_low_u64x2(a, b), self.zip_high_u64x2(a, b))
}
#[inline(always)]
fn deinterleave_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> (u64x2<Self>, u64x2<Self>) {
(self.unzip_low_u64x2(a, b), self.unzip_high_u64x2(a, b))
}
#[inline(always)]
fn select_u64x2(self, a: mask64x2<Self>, b: u64x2<Self>, c: u64x2<Self>) -> u64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: mask64x2<Sse4_2>,
b: u64x2<Sse4_2>,
c: u64x2<Sse4_2>,
) -> u64x2<Sse4_2> {
_mm_blendv_epi8(c.into(), b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b, c)
}
#[inline(always)]
fn combine_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u64x4<Self> {
u64x4 {
val: crate::support::Aligned256([a.val.0, b.val.0]),
simd: self,
}
}
#[inline(always)]
fn load_four_interleaved_u64x2(self, src: &[u64; 8usize]) -> [u64x2<Self>; 4usize] {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, src: &[u64; 8usize]) -> [u64x2<Sse4_2>; 4usize] {
let (chunks, []) = src.as_chunks::<2usize>() else {
unreachable!()
};
let v0: __m128i =
crate::transmute::checked_transmute_copy::<[u64; 2usize], __m128i>(&chunks[0]);
let v1: __m128i =
crate::transmute::checked_transmute_copy::<[u64; 2usize], __m128i>(&chunks[1]);
let v2: __m128i =
crate::transmute::checked_transmute_copy::<[u64; 2usize], __m128i>(&chunks[2]);
let v3: __m128i =
crate::transmute::checked_transmute_copy::<[u64; 2usize], __m128i>(&chunks[3]);
let out0 = _mm_unpacklo_epi64(v0, v2);
let out1 = _mm_unpackhi_epi64(v0, v2);
let out2 = _mm_unpacklo_epi64(v1, v3);
let out3 = _mm_unpackhi_epi64(v1, v3);
[
out0.simd_into(token),
out1.simd_into(token),
out2.simd_into(token),
out3.simd_into(token),
]
}
);
kernel(self, src)
}
#[inline(always)]
fn store_four_interleaved_u64x2(
self,
vectors: [u64x2<Self>; 4usize],
dest: &mut [u64; 8usize],
) -> () {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
vectors: [u64x2<Sse4_2>; 4usize],
dest: &mut [u64; 8usize],
) -> () {
let _ = token;
let v0: __m128i = vectors[0].into();
let v1: __m128i = vectors[1].into();
let v2: __m128i = vectors[2].into();
let v3: __m128i = vectors[3].into();
let out0 = _mm_unpacklo_epi64(v0, v1);
let out1 = _mm_unpacklo_epi64(v2, v3);
let out2 = _mm_unpackhi_epi64(v0, v1);
let out3 = _mm_unpackhi_epi64(v2, v3);
let (chunks, []) = dest.as_chunks_mut::<2usize>() else {
unreachable!()
};
crate::transmute::checked_transmute_store::<__m128i, [u64; 2usize]>(
out0,
&mut chunks[0],
);
crate::transmute::checked_transmute_store::<__m128i, [u64; 2usize]>(
out1,
&mut chunks[1],
);
crate::transmute::checked_transmute_store::<__m128i, [u64; 2usize]>(
out2,
&mut chunks[2],
);
crate::transmute::checked_transmute_store::<__m128i, [u64; 2usize]>(
out3,
&mut chunks[3],
);
}
);
kernel(self, vectors, dest);
}
#[inline(always)]
fn narrow_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u32x4<Sse4_2> {
_mm_castps_si128(_mm_shuffle_ps::<0x88>(
_mm_castsi128_ps(a.into()),
_mm_castsi128_ps(b.into()),
))
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn saturating_narrow_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>, b: u64x2<Sse4_2>) -> u32x4<Sse4_2> {
let a = a.into();
let b = b.into();
let low = _mm_castps_si128(_mm_shuffle_ps::<0x88>(
_mm_castsi128_ps(a),
_mm_castsi128_ps(b),
));
let high = _mm_castps_si128(_mm_shuffle_ps::<0xdd>(
_mm_castsi128_ps(a),
_mm_castsi128_ps(b),
));
let zero = _mm_setzero_si128();
let fits = _mm_cmpeq_epi32(high, zero);
let ones = _mm_cmpeq_epi32(zero, zero);
_mm_or_si128(low, _mm_xor_si128(fits, ones)).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn relaxed_narrow_u64x2(self, a: u64x2<Self>, b: u64x2<Self>) -> u32x4<Self> {
debug_assert!(
a.as_slice()
.iter()
.chain(b.as_slice())
.all(|&value| { value >= u32::MIN as u64 && value <= u32::MAX as u64 }),
"relaxed_narrow inputs must fit in the destination type",
);
self.narrow_u64x2(a, b)
}
#[inline(always)]
fn cvt_f64_u64x2(self, a: u64x2<Self>) -> f64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u64x2<Sse4_2>) -> f64x2<Sse4_2> {
let a = a.into();
let low = _mm_blend_epi16::<0xcc>(a, _mm_set1_epi64x(0x4330_0000_0000_0000));
let high = _mm_srli_epi64::<32>(a);
let high = _mm_xor_si128(high, _mm_set1_epi64x(4985484787499139072i64));
let high = _mm_sub_pd(
_mm_castsi128_pd(high),
_mm_set1_pd(f64::from_bits(4985484787500187648u64)),
);
_mm_add_pd(_mm_castsi128_pd(low), high).simd_into(token)
}
);
kernel(self, a)
}
#[inline(always)]
fn splat_mask64x2(self, val: bool) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, val: bool) -> mask64x2<Sse4_2> {
let val: i64 = if val { !0 } else { 0 };
_mm_set1_epi64x(val).simd_into(token)
}
);
kernel(self, val)
}
#[inline(always)]
fn from_bitmask_mask64x2(self, bits: u64) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, bits: u64) -> mask64x2<Sse4_2> {
{
let bit_lanes = _mm_set1_epi64x(bits.cast_signed());
let bit_mask = _mm_set_epi64x(2, 1);
_mm_cmpeq_epi64(_mm_and_si128(bit_lanes, bit_mask), bit_mask)
}
.simd_into(token)
}
);
kernel(self, bits)
}
#[inline(always)]
fn to_bitmask_mask64x2(self, a: mask64x2<Self>) -> u64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask64x2<Sse4_2>) -> u64 {
_mm_movemask_pd(_mm_castsi128_pd(a.into())) as u32 as u64
}
);
kernel(self, a)
}
#[inline(always)]
fn from_vector_mask64x2(self, vector: i64x2<Self>) -> mask64x2<Self> {
mask64x2 {
val: vector.val,
simd: self,
}
}
#[inline(always)]
fn to_vector_mask64x2(self, a: mask64x2<Self>) -> i64x2<Self> {
i64x2 {
val: a.val,
simd: self,
}
}
#[inline(always)]
fn set_mask64x2(self, a: &mut mask64x2<Self>, index: usize, value: bool) -> () {
assert!(
index < 2usize,
"mask lane index {index} is out of bounds for {} lanes",
2usize
);
let mut lanes: [i64; 2usize] = (*a).into();
lanes[index] = if value { !0 } else { 0 };
*a = lanes.simd_into(self);
}
#[inline(always)]
fn rotate_elements_left_mask64x2<const OFFSET: usize>(
self,
a: mask64x2<Self>,
) -> mask64x2<Self> {
let int = i64x2 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_left::<OFFSET>();
mask64x2 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn rotate_elements_right_mask64x2<const OFFSET: usize>(
self,
a: mask64x2<Self>,
) -> mask64x2<Self> {
let int = i64x2 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_right::<OFFSET>();
mask64x2 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn and_mask64x2(self, a: mask64x2<Self>, b: mask64x2<Self>) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask64x2<Sse4_2>, b: mask64x2<Sse4_2>) -> mask64x2<Sse4_2> {
_mm_and_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn or_mask64x2(self, a: mask64x2<Self>, b: mask64x2<Self>) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask64x2<Sse4_2>, b: mask64x2<Sse4_2>) -> mask64x2<Sse4_2> {
_mm_or_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn xor_mask64x2(self, a: mask64x2<Self>, b: mask64x2<Self>) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask64x2<Sse4_2>, b: mask64x2<Sse4_2>) -> mask64x2<Sse4_2> {
_mm_xor_si128(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn not_mask64x2(self, a: mask64x2<Self>) -> mask64x2<Self> {
self.xor_mask64x2(a, self.splat_mask64x2(true))
}
#[inline(always)]
fn reverse_mask64x2(self, a: mask64x2<Self>) -> mask64x2<Self> {
let lanes = i64x2 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let reversed = self.reverse_i64x2(lanes);
mask64x2 {
val: crate::transmute::checked_transmute_copy(&reversed.val),
simd: self,
}
}
#[inline(always)]
fn select_mask64x2(
self,
a: mask64x2<Self>,
b: mask64x2<Self>,
c: mask64x2<Self>,
) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
a: mask64x2<Sse4_2>,
b: mask64x2<Sse4_2>,
c: mask64x2<Sse4_2>,
) -> mask64x2<Sse4_2> {
_mm_blendv_epi8(c.into(), b.into(), a.into()).simd_into(token)
}
);
kernel(self, a, b, c)
}
#[inline(always)]
fn simd_eq_mask64x2(self, a: mask64x2<Self>, b: mask64x2<Self>) -> mask64x2<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask64x2<Sse4_2>, b: mask64x2<Sse4_2>) -> mask64x2<Sse4_2> {
_mm_cmpeq_epi64(a.into(), b.into()).simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn any_true_mask64x2(self, a: mask64x2<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask64x2<Sse4_2>) -> bool {
_mm_movemask_pd(_mm_castsi128_pd(a.into())) as u32 != 0
}
);
kernel(self, a)
}
#[inline(always)]
fn all_true_mask64x2(self, a: mask64x2<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask64x2<Sse4_2>) -> bool {
_mm_movemask_pd(_mm_castsi128_pd(a.into())) as u32 == 0b11
}
);
kernel(self, a)
}
#[inline(always)]
fn any_false_mask64x2(self, a: mask64x2<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask64x2<Sse4_2>) -> bool {
_mm_movemask_pd(_mm_castsi128_pd(a.into())) as u32 != 0b11
}
);
kernel(self, a)
}
#[inline(always)]
fn all_false_mask64x2(self, a: mask64x2<Self>) -> bool {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask64x2<Sse4_2>) -> bool {
_mm_movemask_pd(_mm_castsi128_pd(a.into())) as u32 == 0
}
);
kernel(self, a)
}
#[inline(always)]
fn combine_mask64x2(self, a: mask64x2<Self>, b: mask64x2<Self>) -> mask64x4<Self> {
mask64x4 {
val: crate::support::Aligned256([a.val.0, b.val.0]),
simd: self,
}
}
#[inline(always)]
fn narrow_mask64x2(self, a: mask64x2<Self>, b: mask64x2<Self>) -> mask32x4<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask64x2<Sse4_2>, b: mask64x2<Sse4_2>) -> mask32x4<Sse4_2> {
_mm_castps_si128(_mm_shuffle_ps::<0x88>(
_mm_castsi128_ps(a.into()),
_mm_castsi128_ps(b.into()),
))
.simd_into(token)
}
);
kernel(self, a, b)
}
#[inline(always)]
fn slide_f32x8<const SHIFT: usize>(self, a: f32x8<Self>, b: f32x8<Self>) -> f32x8<Self> {
if SHIFT >= 8usize {
return b;
}
let result = cross_block_alignr_128x2(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 4usize,
);
Bytes::from_bytes(u8x32 {
val: crate::support::Aligned256(result),
simd: self,
})
}
#[inline(always)]
fn compress_f32x8(self, values: f32x8<Self>, mask: mask32x8<Self>) -> f32x8<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_f32x8(
self,
values: f32x8<Self>,
mask: mask32x8<Self>,
merge: f32x8<Self>,
) -> f32x8<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_f32x8(self, values: f32x8<Self>, mask: mask32x8<Self>) -> f32x8<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_f32x8(
self,
values: f32x8<Self>,
mask: mask32x8<Self>,
merge: f32x8<Self>,
) -> f32x8<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn simd_ne_f32x8(self, a: f32x8<Self>, b: f32x8<Self>) -> mask32x8<Self> {
let (a0, a1) = self.split_f32x8(a);
let (b0, b1) = self.split_f32x8(b);
self.combine_mask32x4(self.simd_ne_f32x4(a0, b0), self.simd_ne_f32x4(a1, b1))
}
#[inline(always)]
fn is_nan_f32x8(self, a: f32x8<Self>) -> mask32x8<Self> {
a.simd_ne(a)
}
#[inline(always)]
fn is_infinite_f32x8(self, a: f32x8<Self>) -> mask32x8<Self> {
a.abs().simd_eq(f32::INFINITY)
}
#[inline(always)]
fn is_finite_f32x8(self, a: f32x8<Self>) -> mask32x8<Self> {
a.abs().simd_lt(f32::INFINITY)
}
#[inline(always)]
fn is_subnormal_f32x8(self, a: f32x8<Self>) -> mask32x8<Self> {
let i: i32x8<Self> = a.bitcast();
(i & 0x7F80_0000_i32).simd_eq(0) & (i & 0x007F_FFFF_i32).simd_ne(0)
}
#[inline(always)]
fn is_normal_f32x8(self, a: f32x8<Self>) -> mask32x8<Self> {
let i: i32x8<Self> = a.bitcast();
let exp = i & 0x7F80_0000_i32;
exp.simd_ne(0) & exp.simd_ne(0x7F80_0000_i32)
}
#[inline(always)]
fn is_sign_positive_f32x8(self, a: f32x8<Self>) -> mask32x8<Self> {
let i: i32x8<Self> = a.bitcast();
(i & i32::MIN).simd_eq(0)
}
#[inline(always)]
fn is_sign_negative_f32x8(self, a: f32x8<Self>) -> mask32x8<Self> {
!a.is_sign_positive()
}
#[inline(always)]
fn combine_f32x8(self, a: f32x8<Self>, b: f32x8<Self>) -> f32x16<Self> {
f32x16 {
val: crate::support::Aligned512([a.val.0[0], a.val.0[1], b.val.0[0], b.val.0[1]]),
simd: self,
}
}
#[inline(always)]
fn split_f32x8(self, a: f32x8<Self>) -> (f32x4<Self>, f32x4<Self>) {
(
f32x4 {
val: crate::support::Aligned128(a.val.0[0]),
simd: self,
},
f32x4 {
val: crate::support::Aligned128(a.val.0[1]),
simd: self,
},
)
}
#[inline(always)]
fn slide_i8x32<const SHIFT: usize>(self, a: i8x32<Self>, b: i8x32<Self>) -> i8x32<Self> {
if SHIFT >= 32usize {
return b;
}
let result = cross_block_alignr_128x2(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT,
);
Bytes::from_bytes(u8x32 {
val: crate::support::Aligned256(result),
simd: self,
})
}
#[inline(always)]
fn compress_i8x32(self, values: i8x32<Self>, mask: mask8x32<Self>) -> i8x32<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_i8x32(
self,
values: i8x32<Self>,
mask: mask8x32<Self>,
merge: i8x32<Self>,
) -> i8x32<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_i8x32(self, values: i8x32<Self>, mask: mask8x32<Self>) -> i8x32<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_i8x32(
self,
values: i8x32<Self>,
mask: mask8x32<Self>,
merge: i8x32<Self>,
) -> i8x32<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn combine_i8x32(self, a: i8x32<Self>, b: i8x32<Self>) -> i8x64<Self> {
i8x64 {
val: crate::support::Aligned512([a.val.0[0], a.val.0[1], b.val.0[0], b.val.0[1]]),
simd: self,
}
}
#[inline(always)]
fn split_i8x32(self, a: i8x32<Self>) -> (i8x16<Self>, i8x16<Self>) {
(
i8x16 {
val: crate::support::Aligned128(a.val.0[0]),
simd: self,
},
i8x16 {
val: crate::support::Aligned128(a.val.0[1]),
simd: self,
},
)
}
#[inline(always)]
fn slide_u8x32<const SHIFT: usize>(self, a: u8x32<Self>, b: u8x32<Self>) -> u8x32<Self> {
if SHIFT >= 32usize {
return b;
}
let result = cross_block_alignr_128x2(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT,
);
Bytes::from_bytes(u8x32 {
val: crate::support::Aligned256(result),
simd: self,
})
}
#[inline(always)]
fn swizzle_dyn_u8x32(self, a: u8x32<Self>, indices: u8x32<Self>) -> u8x32<Self> {
self.swizzle_dyn_precise_u8x32(a, indices)
}
#[inline(always)]
fn swizzle_dyn_precise_u8x32(self, a: u8x32<Self>, indices: u8x32<Self>) -> u8x32<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x32<Sse4_2>, indices: u8x32<Sse4_2>) -> u8x32<Sse4_2> {
let bytes = Bytes::to_bytes(a);
let (table_low, table_high) = token.split_u8x32(bytes);
let (indices_low, indices_high) = token.split_u8x32(indices);
let high_table_offset = token.splat_u8x16(16);
let output_low_from_low = token.swizzle_dyn_precise_u8x16(table_low, indices_low);
let output_low_from_high = token.swizzle_dyn_precise_u8x16(
table_high,
token.sub_u8x16(indices_low, high_table_offset),
);
let output_low = token.or_u8x16(output_low_from_low, output_low_from_high);
let output_high_from_low = token.swizzle_dyn_precise_u8x16(table_low, indices_high);
let output_high_from_high = token.swizzle_dyn_precise_u8x16(
table_high,
token.sub_u8x16(indices_high, high_table_offset),
);
let output_high = token.or_u8x16(output_high_from_low, output_high_from_high);
let result_bytes = token.combine_u8x16(output_low, output_high);
Bytes::from_bytes(result_bytes)
}
);
kernel(self, a, indices)
}
#[inline(always)]
fn concat_swizzle_dyn_u8x32(
self,
a: u8x32<Self>,
b: u8x32<Self>,
indices: u8x32<Self>,
) -> u8x32<Self> {
self.concat_swizzle_dyn_precise_u8x32(a, b, indices)
}
#[inline(always)]
fn concat_swizzle_dyn_precise_u8x32(
self,
a: u8x32<Self>,
b: u8x32<Self>,
indices: u8x32<Self>,
) -> u8x32<Self> {
let first_table = Bytes::to_bytes(a);
let second_table = Bytes::to_bytes(b);
let second_table_offset = self.splat_u8x32(32);
let from_first = self.swizzle_dyn_precise_u8x32(first_table, indices);
let from_second = self
.swizzle_dyn_precise_u8x32(second_table, self.sub_u8x32(indices, second_table_offset));
let result_bytes = self.or_u8x32(from_first, from_second);
Bytes::from_bytes(result_bytes)
}
#[inline(always)]
fn compress_u8x32(self, values: u8x32<Self>, mask: mask8x32<Self>) -> u8x32<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
values: u8x32<Sse4_2>,
mask: mask8x32<Sse4_2>,
) -> u8x32<Sse4_2> {
let mask_bits = token.to_bitmask_mask8x32(mask);
let mut output = [0u8; 32];
let mut output_lane = 0;
let block_bits = (mask_bits & 0xffff) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_control = crate::support::COMPRESS_8_CONTROLS[low_mask];
let high_control = crate::support::COMPRESS_8_CONTROLS[high_mask];
let high_control = high_control | 0x0808_0808_0808_0808;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let low_count = low_mask.count_ones() as usize;
let compacted = _mm_shuffle_epi8(values.val.0[0], control);
let splice = crate::support::COMPACT_8_SPLICE_CONTROLS[low_count];
let splice = _mm_set_epi64x((splice >> 64) as i64, splice as i64);
let compressed = _mm_shuffle_epi8(compacted, splice);
crate::transmute::checked_transmute_store(
compressed,
output[output_lane..].first_chunk_mut::<16>().unwrap(),
);
output_lane += block_bits.count_ones() as usize;
let block_bits = (mask_bits >> 16 & 0xffff) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_control = crate::support::COMPRESS_8_CONTROLS[low_mask];
let high_control = crate::support::COMPRESS_8_CONTROLS[high_mask];
let high_control = high_control | 0x0808_0808_0808_0808;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let low_count = low_mask.count_ones() as usize;
let compacted = _mm_shuffle_epi8(values.val.0[1], control);
let splice = crate::support::COMPACT_8_SPLICE_CONTROLS[low_count];
let splice = _mm_set_epi64x((splice >> 64) as i64, splice as i64);
let compressed = _mm_shuffle_epi8(compacted, splice);
crate::transmute::checked_transmute_store(
compressed,
output[output_lane..].first_chunk_mut::<16>().unwrap(),
);
u8x32::simd_from(token, output)
}
);
kernel(self, values, mask)
}
#[inline(always)]
fn compress_merge_u8x32(
self,
values: u8x32<Self>,
mask: mask8x32<Self>,
merge: u8x32<Self>,
) -> u8x32<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
values: u8x32<Sse4_2>,
mask: mask8x32<Sse4_2>,
merge: u8x32<Sse4_2>,
) -> u8x32<Sse4_2> {
let mask_bits = token.to_bitmask_mask8x32(mask);
let mut output = [0u8; 32];
let mut output_lane = 0;
let block_bits = (mask_bits & 0xffff) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_control = crate::support::COMPRESS_8_CONTROLS[low_mask];
let high_control = crate::support::COMPRESS_8_CONTROLS[high_mask];
let high_control = high_control | 0x0808_0808_0808_0808;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let low_count = low_mask.count_ones() as usize;
let compacted = _mm_shuffle_epi8(values.val.0[0], control);
let splice = crate::support::COMPACT_8_SPLICE_CONTROLS[low_count];
let splice = _mm_set_epi64x((splice >> 64) as i64, splice as i64);
let compressed = _mm_shuffle_epi8(compacted, splice);
crate::transmute::checked_transmute_store(
compressed,
output[output_lane..].first_chunk_mut::<16>().unwrap(),
);
output_lane += block_bits.count_ones() as usize;
let block_bits = (mask_bits >> 16 & 0xffff) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_control = crate::support::COMPRESS_8_CONTROLS[low_mask];
let high_control = crate::support::COMPRESS_8_CONTROLS[high_mask];
let high_control = high_control | 0x0808_0808_0808_0808;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let low_count = low_mask.count_ones() as usize;
let compacted = _mm_shuffle_epi8(values.val.0[1], control);
let splice = crate::support::COMPACT_8_SPLICE_CONTROLS[low_count];
let splice = _mm_set_epi64x((splice >> 64) as i64, splice as i64);
let compressed = _mm_shuffle_epi8(compacted, splice);
crate::transmute::checked_transmute_store(
compressed,
output[output_lane..].first_chunk_mut::<16>().unwrap(),
);
output_lane += block_bits.count_ones() as usize;
let remaining = output_lane.saturating_sub(0).min(16);
let [prefix, _] = crate::transmute::checked_transmute_copy::<_, [__m128i; 2]>(
&crate::support::COMPACT_PREFIX_MASKS[remaining],
);
let compressed = crate::transmute::checked_transmute_copy(
output[0..].first_chunk::<16>().unwrap(),
);
let blended = _mm_blendv_epi8(merge.val.0[0], compressed, prefix);
crate::transmute::checked_transmute_store(
blended,
output[0..].first_chunk_mut::<16>().unwrap(),
);
let remaining = output_lane.saturating_sub(16).min(16);
let [prefix, _] = crate::transmute::checked_transmute_copy::<_, [__m128i; 2]>(
&crate::support::COMPACT_PREFIX_MASKS[remaining],
);
let compressed = crate::transmute::checked_transmute_copy(
output[16..].first_chunk::<16>().unwrap(),
);
let blended = _mm_blendv_epi8(merge.val.0[1], compressed, prefix);
crate::transmute::checked_transmute_store(
blended,
output[16..].first_chunk_mut::<16>().unwrap(),
);
u8x32::simd_from(token, output)
}
);
kernel(self, values, mask, merge)
}
#[inline(always)]
fn expand_u8x32(self, values: u8x32<Self>, mask: mask8x32<Self>) -> u8x32<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
values: u8x32<Sse4_2>,
mask: mask8x32<Sse4_2>,
) -> u8x32<Sse4_2> {
let input = <[u8; 32]>::from(values);
let mut output = [0u8; 32];
let mut input_lane = 0;
let block_bits = _mm_movemask_epi8(mask.val.0[0]) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_count = low_mask.count_ones() as u64;
let low_control = crate::support::EXPAND_8_CONTROLS[low_mask];
let high_control = crate::support::EXPAND_8_CONTROLS[high_mask];
let high_base = low_count * 0x0101_0101_0101_0101;
let high_control = high_control + high_base;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let packed = crate::transmute::checked_transmute_copy(
input[input_lane..].first_chunk::<16>().unwrap(),
);
let expanded = _mm_shuffle_epi8(packed, control);
let result = expanded;
crate::transmute::checked_transmute_store(
result,
output[0..].first_chunk_mut::<16>().unwrap(),
);
input_lane += block_bits.count_ones() as usize;
let block_bits = _mm_movemask_epi8(mask.val.0[1]) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_count = low_mask.count_ones() as u64;
let low_control = crate::support::EXPAND_8_CONTROLS[low_mask];
let high_control = crate::support::EXPAND_8_CONTROLS[high_mask];
let high_base = low_count * 0x0101_0101_0101_0101;
let high_control = high_control + high_base;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let packed = crate::transmute::checked_transmute_copy(
input[input_lane..].first_chunk::<16>().unwrap(),
);
let expanded = _mm_shuffle_epi8(packed, control);
let result = expanded;
crate::transmute::checked_transmute_store(
result,
output[16..].first_chunk_mut::<16>().unwrap(),
);
u8x32::simd_from(token, output)
}
);
kernel(self, values, mask)
}
#[inline(always)]
fn expand_merge_u8x32(
self,
values: u8x32<Self>,
mask: mask8x32<Self>,
merge: u8x32<Self>,
) -> u8x32<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
values: u8x32<Sse4_2>,
mask: mask8x32<Sse4_2>,
merge: u8x32<Sse4_2>,
) -> u8x32<Sse4_2> {
let input = <[u8; 32]>::from(values);
let mut output = [0u8; 32];
let mut input_lane = 0;
let block_bits = _mm_movemask_epi8(mask.val.0[0]) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_count = low_mask.count_ones() as u64;
let low_control = crate::support::EXPAND_8_CONTROLS[low_mask];
let high_control = crate::support::EXPAND_8_CONTROLS[high_mask];
let high_base = low_count * 0x0101_0101_0101_0101;
let high_control = high_control + high_base;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let packed = crate::transmute::checked_transmute_copy(
input[input_lane..].first_chunk::<16>().unwrap(),
);
let expanded = _mm_shuffle_epi8(packed, control);
let result = _mm_blendv_epi8(merge.val.0[0], expanded, mask.val.0[0]);
crate::transmute::checked_transmute_store(
result,
output[0..].first_chunk_mut::<16>().unwrap(),
);
input_lane += block_bits.count_ones() as usize;
let block_bits = _mm_movemask_epi8(mask.val.0[1]) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_count = low_mask.count_ones() as u64;
let low_control = crate::support::EXPAND_8_CONTROLS[low_mask];
let high_control = crate::support::EXPAND_8_CONTROLS[high_mask];
let high_base = low_count * 0x0101_0101_0101_0101;
let high_control = high_control + high_base;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let packed = crate::transmute::checked_transmute_copy(
input[input_lane..].first_chunk::<16>().unwrap(),
);
let expanded = _mm_shuffle_epi8(packed, control);
let result = _mm_blendv_epi8(merge.val.0[1], expanded, mask.val.0[1]);
crate::transmute::checked_transmute_store(
result,
output[16..].first_chunk_mut::<16>().unwrap(),
);
u8x32::simd_from(token, output)
}
);
kernel(self, values, mask, merge)
}
#[inline(always)]
fn combine_u8x32(self, a: u8x32<Self>, b: u8x32<Self>) -> u8x64<Self> {
u8x64 {
val: crate::support::Aligned512([a.val.0[0], a.val.0[1], b.val.0[0], b.val.0[1]]),
simd: self,
}
}
#[inline(always)]
fn split_u8x32(self, a: u8x32<Self>) -> (u8x16<Self>, u8x16<Self>) {
(
u8x16 {
val: crate::support::Aligned128(a.val.0[0]),
simd: self,
},
u8x16 {
val: crate::support::Aligned128(a.val.0[1]),
simd: self,
},
)
}
#[inline(always)]
fn from_vector_mask8x32(self, vector: i8x32<Self>) -> mask8x32<Self> {
mask8x32 {
val: vector.val,
simd: self,
}
}
#[inline(always)]
fn to_vector_mask8x32(self, a: mask8x32<Self>) -> i8x32<Self> {
i8x32 {
val: a.val,
simd: self,
}
}
#[inline(always)]
fn set_mask8x32(self, a: &mut mask8x32<Self>, index: usize, value: bool) -> () {
assert!(
index < 32usize,
"mask lane index {index} is out of bounds for {} lanes",
32usize
);
let mut lanes: [i8; 32usize] = (*a).into();
lanes[index] = if value { !0 } else { 0 };
*a = lanes.simd_into(self);
}
#[inline(always)]
fn rotate_elements_left_mask8x32<const OFFSET: usize>(
self,
a: mask8x32<Self>,
) -> mask8x32<Self> {
let int = i8x32 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_left::<OFFSET>();
mask8x32 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn rotate_elements_right_mask8x32<const OFFSET: usize>(
self,
a: mask8x32<Self>,
) -> mask8x32<Self> {
let int = i8x32 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_right::<OFFSET>();
mask8x32 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn combine_mask8x32(self, a: mask8x32<Self>, b: mask8x32<Self>) -> mask8x64<Self> {
mask8x64 {
val: crate::support::Aligned512([a.val.0[0], a.val.0[1], b.val.0[0], b.val.0[1]]),
simd: self,
}
}
#[inline(always)]
fn split_mask8x32(self, a: mask8x32<Self>) -> (mask8x16<Self>, mask8x16<Self>) {
(
mask8x16 {
val: crate::support::Aligned128(a.val.0[0]),
simd: self,
},
mask8x16 {
val: crate::support::Aligned128(a.val.0[1]),
simd: self,
},
)
}
#[inline(always)]
fn slide_i16x16<const SHIFT: usize>(self, a: i16x16<Self>, b: i16x16<Self>) -> i16x16<Self> {
if SHIFT >= 16usize {
return b;
}
let result = cross_block_alignr_128x2(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 2usize,
);
Bytes::from_bytes(u8x32 {
val: crate::support::Aligned256(result),
simd: self,
})
}
#[inline(always)]
fn compress_i16x16(self, values: i16x16<Self>, mask: mask16x16<Self>) -> i16x16<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_i16x16(
self,
values: i16x16<Self>,
mask: mask16x16<Self>,
merge: i16x16<Self>,
) -> i16x16<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_i16x16(self, values: i16x16<Self>, mask: mask16x16<Self>) -> i16x16<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_i16x16(
self,
values: i16x16<Self>,
mask: mask16x16<Self>,
merge: i16x16<Self>,
) -> i16x16<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn combine_i16x16(self, a: i16x16<Self>, b: i16x16<Self>) -> i16x32<Self> {
i16x32 {
val: crate::support::Aligned512([a.val.0[0], a.val.0[1], b.val.0[0], b.val.0[1]]),
simd: self,
}
}
#[inline(always)]
fn split_i16x16(self, a: i16x16<Self>) -> (i16x8<Self>, i16x8<Self>) {
(
i16x8 {
val: crate::support::Aligned128(a.val.0[0]),
simd: self,
},
i16x8 {
val: crate::support::Aligned128(a.val.0[1]),
simd: self,
},
)
}
#[inline(always)]
fn slide_u16x16<const SHIFT: usize>(self, a: u16x16<Self>, b: u16x16<Self>) -> u16x16<Self> {
if SHIFT >= 16usize {
return b;
}
let result = cross_block_alignr_128x2(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 2usize,
);
Bytes::from_bytes(u8x32 {
val: crate::support::Aligned256(result),
simd: self,
})
}
#[inline(always)]
fn compress_u16x16(self, values: u16x16<Self>, mask: mask16x16<Self>) -> u16x16<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_u16x16(
self,
values: u16x16<Self>,
mask: mask16x16<Self>,
merge: u16x16<Self>,
) -> u16x16<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_u16x16(self, values: u16x16<Self>, mask: mask16x16<Self>) -> u16x16<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_u16x16(
self,
values: u16x16<Self>,
mask: mask16x16<Self>,
merge: u16x16<Self>,
) -> u16x16<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn combine_u16x16(self, a: u16x16<Self>, b: u16x16<Self>) -> u16x32<Self> {
u16x32 {
val: crate::support::Aligned512([a.val.0[0], a.val.0[1], b.val.0[0], b.val.0[1]]),
simd: self,
}
}
#[inline(always)]
fn split_u16x16(self, a: u16x16<Self>) -> (u16x8<Self>, u16x8<Self>) {
(
u16x8 {
val: crate::support::Aligned128(a.val.0[0]),
simd: self,
},
u16x8 {
val: crate::support::Aligned128(a.val.0[1]),
simd: self,
},
)
}
#[inline(always)]
fn to_bitmask_mask16x16(self, a: mask16x16<Self>) -> u64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask16x16<Sse4_2>) -> u64 {
{
let packed = _mm_packs_epi16(a.val.0[0], a.val.0[1]);
_mm_movemask_epi8(packed) as u32 as u64
}
}
);
kernel(self, a)
}
#[inline(always)]
fn from_vector_mask16x16(self, vector: i16x16<Self>) -> mask16x16<Self> {
mask16x16 {
val: vector.val,
simd: self,
}
}
#[inline(always)]
fn to_vector_mask16x16(self, a: mask16x16<Self>) -> i16x16<Self> {
i16x16 {
val: a.val,
simd: self,
}
}
#[inline(always)]
fn set_mask16x16(self, a: &mut mask16x16<Self>, index: usize, value: bool) -> () {
assert!(
index < 16usize,
"mask lane index {index} is out of bounds for {} lanes",
16usize
);
let mut lanes: [i16; 16usize] = (*a).into();
lanes[index] = if value { !0 } else { 0 };
*a = lanes.simd_into(self);
}
#[inline(always)]
fn rotate_elements_left_mask16x16<const OFFSET: usize>(
self,
a: mask16x16<Self>,
) -> mask16x16<Self> {
let int = i16x16 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_left::<OFFSET>();
mask16x16 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn rotate_elements_right_mask16x16<const OFFSET: usize>(
self,
a: mask16x16<Self>,
) -> mask16x16<Self> {
let int = i16x16 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_right::<OFFSET>();
mask16x16 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn combine_mask16x16(self, a: mask16x16<Self>, b: mask16x16<Self>) -> mask16x32<Self> {
mask16x32 {
val: crate::support::Aligned512([a.val.0[0], a.val.0[1], b.val.0[0], b.val.0[1]]),
simd: self,
}
}
#[inline(always)]
fn split_mask16x16(self, a: mask16x16<Self>) -> (mask16x8<Self>, mask16x8<Self>) {
(
mask16x8 {
val: crate::support::Aligned128(a.val.0[0]),
simd: self,
},
mask16x8 {
val: crate::support::Aligned128(a.val.0[1]),
simd: self,
},
)
}
#[inline(always)]
fn slide_i32x8<const SHIFT: usize>(self, a: i32x8<Self>, b: i32x8<Self>) -> i32x8<Self> {
if SHIFT >= 8usize {
return b;
}
let result = cross_block_alignr_128x2(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 4usize,
);
Bytes::from_bytes(u8x32 {
val: crate::support::Aligned256(result),
simd: self,
})
}
#[inline(always)]
fn compress_i32x8(self, values: i32x8<Self>, mask: mask32x8<Self>) -> i32x8<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_i32x8(
self,
values: i32x8<Self>,
mask: mask32x8<Self>,
merge: i32x8<Self>,
) -> i32x8<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_i32x8(self, values: i32x8<Self>, mask: mask32x8<Self>) -> i32x8<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_i32x8(
self,
values: i32x8<Self>,
mask: mask32x8<Self>,
merge: i32x8<Self>,
) -> i32x8<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn combine_i32x8(self, a: i32x8<Self>, b: i32x8<Self>) -> i32x16<Self> {
i32x16 {
val: crate::support::Aligned512([a.val.0[0], a.val.0[1], b.val.0[0], b.val.0[1]]),
simd: self,
}
}
#[inline(always)]
fn split_i32x8(self, a: i32x8<Self>) -> (i32x4<Self>, i32x4<Self>) {
(
i32x4 {
val: crate::support::Aligned128(a.val.0[0]),
simd: self,
},
i32x4 {
val: crate::support::Aligned128(a.val.0[1]),
simd: self,
},
)
}
#[inline(always)]
fn slide_u32x8<const SHIFT: usize>(self, a: u32x8<Self>, b: u32x8<Self>) -> u32x8<Self> {
if SHIFT >= 8usize {
return b;
}
let result = cross_block_alignr_128x2(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 4usize,
);
Bytes::from_bytes(u8x32 {
val: crate::support::Aligned256(result),
simd: self,
})
}
#[inline(always)]
fn compress_u32x8(self, values: u32x8<Self>, mask: mask32x8<Self>) -> u32x8<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_u32x8(
self,
values: u32x8<Self>,
mask: mask32x8<Self>,
merge: u32x8<Self>,
) -> u32x8<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_u32x8(self, values: u32x8<Self>, mask: mask32x8<Self>) -> u32x8<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_u32x8(
self,
values: u32x8<Self>,
mask: mask32x8<Self>,
merge: u32x8<Self>,
) -> u32x8<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn combine_u32x8(self, a: u32x8<Self>, b: u32x8<Self>) -> u32x16<Self> {
u32x16 {
val: crate::support::Aligned512([a.val.0[0], a.val.0[1], b.val.0[0], b.val.0[1]]),
simd: self,
}
}
#[inline(always)]
fn split_u32x8(self, a: u32x8<Self>) -> (u32x4<Self>, u32x4<Self>) {
(
u32x4 {
val: crate::support::Aligned128(a.val.0[0]),
simd: self,
},
u32x4 {
val: crate::support::Aligned128(a.val.0[1]),
simd: self,
},
)
}
#[inline(always)]
fn from_vector_mask32x8(self, vector: i32x8<Self>) -> mask32x8<Self> {
mask32x8 {
val: vector.val,
simd: self,
}
}
#[inline(always)]
fn to_vector_mask32x8(self, a: mask32x8<Self>) -> i32x8<Self> {
i32x8 {
val: a.val,
simd: self,
}
}
#[inline(always)]
fn set_mask32x8(self, a: &mut mask32x8<Self>, index: usize, value: bool) -> () {
assert!(
index < 8usize,
"mask lane index {index} is out of bounds for {} lanes",
8usize
);
let mut lanes: [i32; 8usize] = (*a).into();
lanes[index] = if value { !0 } else { 0 };
*a = lanes.simd_into(self);
}
#[inline(always)]
fn rotate_elements_left_mask32x8<const OFFSET: usize>(
self,
a: mask32x8<Self>,
) -> mask32x8<Self> {
let int = i32x8 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_left::<OFFSET>();
mask32x8 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn rotate_elements_right_mask32x8<const OFFSET: usize>(
self,
a: mask32x8<Self>,
) -> mask32x8<Self> {
let int = i32x8 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_right::<OFFSET>();
mask32x8 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn combine_mask32x8(self, a: mask32x8<Self>, b: mask32x8<Self>) -> mask32x16<Self> {
mask32x16 {
val: crate::support::Aligned512([a.val.0[0], a.val.0[1], b.val.0[0], b.val.0[1]]),
simd: self,
}
}
#[inline(always)]
fn split_mask32x8(self, a: mask32x8<Self>) -> (mask32x4<Self>, mask32x4<Self>) {
(
mask32x4 {
val: crate::support::Aligned128(a.val.0[0]),
simd: self,
},
mask32x4 {
val: crate::support::Aligned128(a.val.0[1]),
simd: self,
},
)
}
#[inline(always)]
fn slide_f64x4<const SHIFT: usize>(self, a: f64x4<Self>, b: f64x4<Self>) -> f64x4<Self> {
if SHIFT >= 4usize {
return b;
}
let result = cross_block_alignr_128x2(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 8usize,
);
Bytes::from_bytes(u8x32 {
val: crate::support::Aligned256(result),
simd: self,
})
}
#[inline(always)]
fn compress_f64x4(self, values: f64x4<Self>, mask: mask64x4<Self>) -> f64x4<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_f64x4(
self,
values: f64x4<Self>,
mask: mask64x4<Self>,
merge: f64x4<Self>,
) -> f64x4<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_f64x4(self, values: f64x4<Self>, mask: mask64x4<Self>) -> f64x4<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_f64x4(
self,
values: f64x4<Self>,
mask: mask64x4<Self>,
merge: f64x4<Self>,
) -> f64x4<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn simd_ne_f64x4(self, a: f64x4<Self>, b: f64x4<Self>) -> mask64x4<Self> {
let (a0, a1) = self.split_f64x4(a);
let (b0, b1) = self.split_f64x4(b);
self.combine_mask64x2(self.simd_ne_f64x2(a0, b0), self.simd_ne_f64x2(a1, b1))
}
#[inline(always)]
fn is_nan_f64x4(self, a: f64x4<Self>) -> mask64x4<Self> {
a.simd_ne(a)
}
#[inline(always)]
fn is_infinite_f64x4(self, a: f64x4<Self>) -> mask64x4<Self> {
a.abs().simd_eq(f64::INFINITY)
}
#[inline(always)]
fn is_finite_f64x4(self, a: f64x4<Self>) -> mask64x4<Self> {
a.abs().simd_lt(f64::INFINITY)
}
#[inline(always)]
fn is_subnormal_f64x4(self, a: f64x4<Self>) -> mask64x4<Self> {
let i: i64x4<Self> = a.bitcast();
(i & 0x7FF0_0000_0000_0000_i64).simd_eq(0) & (i & 0x000F_FFFF_FFFF_FFFF_i64).simd_ne(0)
}
#[inline(always)]
fn is_normal_f64x4(self, a: f64x4<Self>) -> mask64x4<Self> {
let i: i64x4<Self> = a.bitcast();
let exp = i & 0x7FF0_0000_0000_0000_i64;
exp.simd_ne(0) & exp.simd_ne(0x7FF0_0000_0000_0000_i64)
}
#[inline(always)]
fn is_sign_positive_f64x4(self, a: f64x4<Self>) -> mask64x4<Self> {
let i: i64x4<Self> = a.bitcast();
(i & i64::MIN).simd_eq(0)
}
#[inline(always)]
fn is_sign_negative_f64x4(self, a: f64x4<Self>) -> mask64x4<Self> {
!a.is_sign_positive()
}
#[inline(always)]
fn combine_f64x4(self, a: f64x4<Self>, b: f64x4<Self>) -> f64x8<Self> {
f64x8 {
val: crate::support::Aligned512([a.val.0[0], a.val.0[1], b.val.0[0], b.val.0[1]]),
simd: self,
}
}
#[inline(always)]
fn split_f64x4(self, a: f64x4<Self>) -> (f64x2<Self>, f64x2<Self>) {
(
f64x2 {
val: crate::support::Aligned128(a.val.0[0]),
simd: self,
},
f64x2 {
val: crate::support::Aligned128(a.val.0[1]),
simd: self,
},
)
}
#[inline(always)]
fn slide_i64x4<const SHIFT: usize>(self, a: i64x4<Self>, b: i64x4<Self>) -> i64x4<Self> {
if SHIFT >= 4usize {
return b;
}
let result = cross_block_alignr_128x2(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 8usize,
);
Bytes::from_bytes(u8x32 {
val: crate::support::Aligned256(result),
simd: self,
})
}
#[inline(always)]
fn compress_i64x4(self, values: i64x4<Self>, mask: mask64x4<Self>) -> i64x4<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_i64x4(
self,
values: i64x4<Self>,
mask: mask64x4<Self>,
merge: i64x4<Self>,
) -> i64x4<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_i64x4(self, values: i64x4<Self>, mask: mask64x4<Self>) -> i64x4<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_i64x4(
self,
values: i64x4<Self>,
mask: mask64x4<Self>,
merge: i64x4<Self>,
) -> i64x4<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn combine_i64x4(self, a: i64x4<Self>, b: i64x4<Self>) -> i64x8<Self> {
i64x8 {
val: crate::support::Aligned512([a.val.0[0], a.val.0[1], b.val.0[0], b.val.0[1]]),
simd: self,
}
}
#[inline(always)]
fn split_i64x4(self, a: i64x4<Self>) -> (i64x2<Self>, i64x2<Self>) {
(
i64x2 {
val: crate::support::Aligned128(a.val.0[0]),
simd: self,
},
i64x2 {
val: crate::support::Aligned128(a.val.0[1]),
simd: self,
},
)
}
#[inline(always)]
fn slide_u64x4<const SHIFT: usize>(self, a: u64x4<Self>, b: u64x4<Self>) -> u64x4<Self> {
if SHIFT >= 4usize {
return b;
}
let result = cross_block_alignr_128x2(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 8usize,
);
Bytes::from_bytes(u8x32 {
val: crate::support::Aligned256(result),
simd: self,
})
}
#[inline(always)]
fn compress_u64x4(self, values: u64x4<Self>, mask: mask64x4<Self>) -> u64x4<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_u64x4(
self,
values: u64x4<Self>,
mask: mask64x4<Self>,
merge: u64x4<Self>,
) -> u64x4<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_u64x4(self, values: u64x4<Self>, mask: mask64x4<Self>) -> u64x4<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x32(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_u64x4(
self,
values: u64x4<Self>,
mask: mask64x4<Self>,
merge: u64x4<Self>,
) -> u64x4<Self> {
let mask = mask8x32 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x32(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn combine_u64x4(self, a: u64x4<Self>, b: u64x4<Self>) -> u64x8<Self> {
u64x8 {
val: crate::support::Aligned512([a.val.0[0], a.val.0[1], b.val.0[0], b.val.0[1]]),
simd: self,
}
}
#[inline(always)]
fn split_u64x4(self, a: u64x4<Self>) -> (u64x2<Self>, u64x2<Self>) {
(
u64x2 {
val: crate::support::Aligned128(a.val.0[0]),
simd: self,
},
u64x2 {
val: crate::support::Aligned128(a.val.0[1]),
simd: self,
},
)
}
#[inline(always)]
fn from_vector_mask64x4(self, vector: i64x4<Self>) -> mask64x4<Self> {
mask64x4 {
val: vector.val,
simd: self,
}
}
#[inline(always)]
fn to_vector_mask64x4(self, a: mask64x4<Self>) -> i64x4<Self> {
i64x4 {
val: a.val,
simd: self,
}
}
#[inline(always)]
fn set_mask64x4(self, a: &mut mask64x4<Self>, index: usize, value: bool) -> () {
assert!(
index < 4usize,
"mask lane index {index} is out of bounds for {} lanes",
4usize
);
let mut lanes: [i64; 4usize] = (*a).into();
lanes[index] = if value { !0 } else { 0 };
*a = lanes.simd_into(self);
}
#[inline(always)]
fn rotate_elements_left_mask64x4<const OFFSET: usize>(
self,
a: mask64x4<Self>,
) -> mask64x4<Self> {
let int = i64x4 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_left::<OFFSET>();
mask64x4 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn rotate_elements_right_mask64x4<const OFFSET: usize>(
self,
a: mask64x4<Self>,
) -> mask64x4<Self> {
let int = i64x4 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_right::<OFFSET>();
mask64x4 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn combine_mask64x4(self, a: mask64x4<Self>, b: mask64x4<Self>) -> mask64x8<Self> {
mask64x8 {
val: crate::support::Aligned512([a.val.0[0], a.val.0[1], b.val.0[0], b.val.0[1]]),
simd: self,
}
}
#[inline(always)]
fn split_mask64x4(self, a: mask64x4<Self>) -> (mask64x2<Self>, mask64x2<Self>) {
(
mask64x2 {
val: crate::support::Aligned128(a.val.0[0]),
simd: self,
},
mask64x2 {
val: crate::support::Aligned128(a.val.0[1]),
simd: self,
},
)
}
#[inline(always)]
fn slide_f32x16<const SHIFT: usize>(self, a: f32x16<Self>, b: f32x16<Self>) -> f32x16<Self> {
if SHIFT >= 16usize {
return b;
}
let result = cross_block_alignr_128x4(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 4usize,
);
Bytes::from_bytes(u8x64 {
val: crate::support::Aligned512(result),
simd: self,
})
}
#[inline(always)]
fn compress_f32x16(self, values: f32x16<Self>, mask: mask32x16<Self>) -> f32x16<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_f32x16(
self,
values: f32x16<Self>,
mask: mask32x16<Self>,
merge: f32x16<Self>,
) -> f32x16<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_f32x16(self, values: f32x16<Self>, mask: mask32x16<Self>) -> f32x16<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_f32x16(
self,
values: f32x16<Self>,
mask: mask32x16<Self>,
merge: f32x16<Self>,
) -> f32x16<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn simd_ne_f32x16(self, a: f32x16<Self>, b: f32x16<Self>) -> mask32x16<Self> {
let (a0, a1) = self.split_f32x16(a);
let (b0, b1) = self.split_f32x16(b);
self.combine_mask32x8(self.simd_ne_f32x8(a0, b0), self.simd_ne_f32x8(a1, b1))
}
#[inline(always)]
fn is_nan_f32x16(self, a: f32x16<Self>) -> mask32x16<Self> {
a.simd_ne(a)
}
#[inline(always)]
fn is_infinite_f32x16(self, a: f32x16<Self>) -> mask32x16<Self> {
a.abs().simd_eq(f32::INFINITY)
}
#[inline(always)]
fn is_finite_f32x16(self, a: f32x16<Self>) -> mask32x16<Self> {
a.abs().simd_lt(f32::INFINITY)
}
#[inline(always)]
fn is_subnormal_f32x16(self, a: f32x16<Self>) -> mask32x16<Self> {
let i: i32x16<Self> = a.bitcast();
(i & 0x7F80_0000_i32).simd_eq(0) & (i & 0x007F_FFFF_i32).simd_ne(0)
}
#[inline(always)]
fn is_normal_f32x16(self, a: f32x16<Self>) -> mask32x16<Self> {
let i: i32x16<Self> = a.bitcast();
let exp = i & 0x7F80_0000_i32;
exp.simd_ne(0) & exp.simd_ne(0x7F80_0000_i32)
}
#[inline(always)]
fn is_sign_positive_f32x16(self, a: f32x16<Self>) -> mask32x16<Self> {
let i: i32x16<Self> = a.bitcast();
(i & i32::MIN).simd_eq(0)
}
#[inline(always)]
fn is_sign_negative_f32x16(self, a: f32x16<Self>) -> mask32x16<Self> {
!a.is_sign_positive()
}
#[inline(always)]
fn split_f32x16(self, a: f32x16<Self>) -> (f32x8<Self>, f32x8<Self>) {
(
f32x8 {
val: crate::support::Aligned256([a.val.0[0], a.val.0[1]]),
simd: self,
},
f32x8 {
val: crate::support::Aligned256([a.val.0[2], a.val.0[3]]),
simd: self,
},
)
}
#[inline(always)]
fn slide_i8x64<const SHIFT: usize>(self, a: i8x64<Self>, b: i8x64<Self>) -> i8x64<Self> {
if SHIFT >= 64usize {
return b;
}
let result = cross_block_alignr_128x4(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT,
);
Bytes::from_bytes(u8x64 {
val: crate::support::Aligned512(result),
simd: self,
})
}
#[inline(always)]
fn compress_i8x64(self, values: i8x64<Self>, mask: mask8x64<Self>) -> i8x64<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_i8x64(
self,
values: i8x64<Self>,
mask: mask8x64<Self>,
merge: i8x64<Self>,
) -> i8x64<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_i8x64(self, values: i8x64<Self>, mask: mask8x64<Self>) -> i8x64<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_i8x64(
self,
values: i8x64<Self>,
mask: mask8x64<Self>,
merge: i8x64<Self>,
) -> i8x64<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn split_i8x64(self, a: i8x64<Self>) -> (i8x32<Self>, i8x32<Self>) {
(
i8x32 {
val: crate::support::Aligned256([a.val.0[0], a.val.0[1]]),
simd: self,
},
i8x32 {
val: crate::support::Aligned256([a.val.0[2], a.val.0[3]]),
simd: self,
},
)
}
#[inline(always)]
fn slide_u8x64<const SHIFT: usize>(self, a: u8x64<Self>, b: u8x64<Self>) -> u8x64<Self> {
if SHIFT >= 64usize {
return b;
}
let result = cross_block_alignr_128x4(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT,
);
Bytes::from_bytes(u8x64 {
val: crate::support::Aligned512(result),
simd: self,
})
}
#[inline(always)]
fn swizzle_dyn_u8x64(self, a: u8x64<Self>, indices: u8x64<Self>) -> u8x64<Self> {
self.swizzle_dyn_precise_u8x64(a, indices)
}
#[inline(always)]
fn swizzle_dyn_precise_u8x64(self, a: u8x64<Self>, indices: u8x64<Self>) -> u8x64<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: u8x64<Sse4_2>, indices: u8x64<Sse4_2>) -> u8x64<Sse4_2> {
let bytes = Bytes::to_bytes(a);
let (table_low, table_high) = token.split_u8x64(bytes);
let (indices_low, indices_high) = token.split_u8x64(indices);
let high_table_offset = token.splat_u8x32(32);
let output_low_from_low = token.swizzle_dyn_precise_u8x32(table_low, indices_low);
let output_low_from_high = token.swizzle_dyn_precise_u8x32(
table_high,
token.sub_u8x32(indices_low, high_table_offset),
);
let output_low = token.or_u8x32(output_low_from_low, output_low_from_high);
let output_high_from_low = token.swizzle_dyn_precise_u8x32(table_low, indices_high);
let output_high_from_high = token.swizzle_dyn_precise_u8x32(
table_high,
token.sub_u8x32(indices_high, high_table_offset),
);
let output_high = token.or_u8x32(output_high_from_low, output_high_from_high);
let result_bytes = token.combine_u8x32(output_low, output_high);
Bytes::from_bytes(result_bytes)
}
);
kernel(self, a, indices)
}
#[inline(always)]
fn concat_swizzle_dyn_u8x64(
self,
a: u8x64<Self>,
b: u8x64<Self>,
indices: u8x64<Self>,
) -> u8x64<Self> {
self.concat_swizzle_dyn_precise_u8x64(a, b, indices)
}
#[inline(always)]
fn concat_swizzle_dyn_precise_u8x64(
self,
a: u8x64<Self>,
b: u8x64<Self>,
indices: u8x64<Self>,
) -> u8x64<Self> {
let first_table = Bytes::to_bytes(a);
let second_table = Bytes::to_bytes(b);
let second_table_offset = self.splat_u8x64(64);
let from_first = self.swizzle_dyn_precise_u8x64(first_table, indices);
let from_second = self
.swizzle_dyn_precise_u8x64(second_table, self.sub_u8x64(indices, second_table_offset));
let result_bytes = self.or_u8x64(from_first, from_second);
Bytes::from_bytes(result_bytes)
}
#[inline(always)]
fn compress_u8x64(self, values: u8x64<Self>, mask: mask8x64<Self>) -> u8x64<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
values: u8x64<Sse4_2>,
mask: mask8x64<Sse4_2>,
) -> u8x64<Sse4_2> {
let mask_bits = token.to_bitmask_mask8x64(mask);
let mut output = [0u8; 64];
let mut output_lane = 0;
let block_bits = (mask_bits & 0xffff) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_control = crate::support::COMPRESS_8_CONTROLS[low_mask];
let high_control = crate::support::COMPRESS_8_CONTROLS[high_mask];
let high_control = high_control | 0x0808_0808_0808_0808;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let low_count = low_mask.count_ones() as usize;
let compacted = _mm_shuffle_epi8(values.val.0[0], control);
let splice = crate::support::COMPACT_8_SPLICE_CONTROLS[low_count];
let splice = _mm_set_epi64x((splice >> 64) as i64, splice as i64);
let compressed = _mm_shuffle_epi8(compacted, splice);
crate::transmute::checked_transmute_store(
compressed,
output[output_lane..].first_chunk_mut::<16>().unwrap(),
);
output_lane += block_bits.count_ones() as usize;
let block_bits = (mask_bits >> 16 & 0xffff) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_control = crate::support::COMPRESS_8_CONTROLS[low_mask];
let high_control = crate::support::COMPRESS_8_CONTROLS[high_mask];
let high_control = high_control | 0x0808_0808_0808_0808;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let low_count = low_mask.count_ones() as usize;
let compacted = _mm_shuffle_epi8(values.val.0[1], control);
let splice = crate::support::COMPACT_8_SPLICE_CONTROLS[low_count];
let splice = _mm_set_epi64x((splice >> 64) as i64, splice as i64);
let compressed = _mm_shuffle_epi8(compacted, splice);
crate::transmute::checked_transmute_store(
compressed,
output[output_lane..].first_chunk_mut::<16>().unwrap(),
);
output_lane += block_bits.count_ones() as usize;
let block_bits = (mask_bits >> 32 & 0xffff) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_control = crate::support::COMPRESS_8_CONTROLS[low_mask];
let high_control = crate::support::COMPRESS_8_CONTROLS[high_mask];
let high_control = high_control | 0x0808_0808_0808_0808;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let low_count = low_mask.count_ones() as usize;
let compacted = _mm_shuffle_epi8(values.val.0[2], control);
let splice = crate::support::COMPACT_8_SPLICE_CONTROLS[low_count];
let splice = _mm_set_epi64x((splice >> 64) as i64, splice as i64);
let compressed = _mm_shuffle_epi8(compacted, splice);
crate::transmute::checked_transmute_store(
compressed,
output[output_lane..].first_chunk_mut::<16>().unwrap(),
);
output_lane += block_bits.count_ones() as usize;
let block_bits = (mask_bits >> 48 & 0xffff) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_control = crate::support::COMPRESS_8_CONTROLS[low_mask];
let high_control = crate::support::COMPRESS_8_CONTROLS[high_mask];
let high_control = high_control | 0x0808_0808_0808_0808;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let low_count = low_mask.count_ones() as usize;
let compacted = _mm_shuffle_epi8(values.val.0[3], control);
let splice = crate::support::COMPACT_8_SPLICE_CONTROLS[low_count];
let splice = _mm_set_epi64x((splice >> 64) as i64, splice as i64);
let compressed = _mm_shuffle_epi8(compacted, splice);
crate::transmute::checked_transmute_store(
compressed,
output[output_lane..].first_chunk_mut::<16>().unwrap(),
);
u8x64::simd_from(token, output)
}
);
kernel(self, values, mask)
}
#[inline(always)]
fn compress_merge_u8x64(
self,
values: u8x64<Self>,
mask: mask8x64<Self>,
merge: u8x64<Self>,
) -> u8x64<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
values: u8x64<Sse4_2>,
mask: mask8x64<Sse4_2>,
merge: u8x64<Sse4_2>,
) -> u8x64<Sse4_2> {
let mask_bits = token.to_bitmask_mask8x64(mask);
let mut output = [0u8; 64];
let mut output_lane = 0;
let block_bits = (mask_bits & 0xffff) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_control = crate::support::COMPRESS_8_CONTROLS[low_mask];
let high_control = crate::support::COMPRESS_8_CONTROLS[high_mask];
let high_control = high_control | 0x0808_0808_0808_0808;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let low_count = low_mask.count_ones() as usize;
let compacted = _mm_shuffle_epi8(values.val.0[0], control);
let splice = crate::support::COMPACT_8_SPLICE_CONTROLS[low_count];
let splice = _mm_set_epi64x((splice >> 64) as i64, splice as i64);
let compressed = _mm_shuffle_epi8(compacted, splice);
crate::transmute::checked_transmute_store(
compressed,
output[output_lane..].first_chunk_mut::<16>().unwrap(),
);
output_lane += block_bits.count_ones() as usize;
let block_bits = (mask_bits >> 16 & 0xffff) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_control = crate::support::COMPRESS_8_CONTROLS[low_mask];
let high_control = crate::support::COMPRESS_8_CONTROLS[high_mask];
let high_control = high_control | 0x0808_0808_0808_0808;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let low_count = low_mask.count_ones() as usize;
let compacted = _mm_shuffle_epi8(values.val.0[1], control);
let splice = crate::support::COMPACT_8_SPLICE_CONTROLS[low_count];
let splice = _mm_set_epi64x((splice >> 64) as i64, splice as i64);
let compressed = _mm_shuffle_epi8(compacted, splice);
crate::transmute::checked_transmute_store(
compressed,
output[output_lane..].first_chunk_mut::<16>().unwrap(),
);
output_lane += block_bits.count_ones() as usize;
let block_bits = (mask_bits >> 32 & 0xffff) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_control = crate::support::COMPRESS_8_CONTROLS[low_mask];
let high_control = crate::support::COMPRESS_8_CONTROLS[high_mask];
let high_control = high_control | 0x0808_0808_0808_0808;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let low_count = low_mask.count_ones() as usize;
let compacted = _mm_shuffle_epi8(values.val.0[2], control);
let splice = crate::support::COMPACT_8_SPLICE_CONTROLS[low_count];
let splice = _mm_set_epi64x((splice >> 64) as i64, splice as i64);
let compressed = _mm_shuffle_epi8(compacted, splice);
crate::transmute::checked_transmute_store(
compressed,
output[output_lane..].first_chunk_mut::<16>().unwrap(),
);
output_lane += block_bits.count_ones() as usize;
let block_bits = (mask_bits >> 48 & 0xffff) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_control = crate::support::COMPRESS_8_CONTROLS[low_mask];
let high_control = crate::support::COMPRESS_8_CONTROLS[high_mask];
let high_control = high_control | 0x0808_0808_0808_0808;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let low_count = low_mask.count_ones() as usize;
let compacted = _mm_shuffle_epi8(values.val.0[3], control);
let splice = crate::support::COMPACT_8_SPLICE_CONTROLS[low_count];
let splice = _mm_set_epi64x((splice >> 64) as i64, splice as i64);
let compressed = _mm_shuffle_epi8(compacted, splice);
crate::transmute::checked_transmute_store(
compressed,
output[output_lane..].first_chunk_mut::<16>().unwrap(),
);
output_lane += block_bits.count_ones() as usize;
let remaining = output_lane.saturating_sub(0).min(16);
let [prefix, _] = crate::transmute::checked_transmute_copy::<_, [__m128i; 2]>(
&crate::support::COMPACT_PREFIX_MASKS[remaining],
);
let compressed = crate::transmute::checked_transmute_copy(
output[0..].first_chunk::<16>().unwrap(),
);
let blended = _mm_blendv_epi8(merge.val.0[0], compressed, prefix);
crate::transmute::checked_transmute_store(
blended,
output[0..].first_chunk_mut::<16>().unwrap(),
);
let remaining = output_lane.saturating_sub(16).min(16);
let [prefix, _] = crate::transmute::checked_transmute_copy::<_, [__m128i; 2]>(
&crate::support::COMPACT_PREFIX_MASKS[remaining],
);
let compressed = crate::transmute::checked_transmute_copy(
output[16..].first_chunk::<16>().unwrap(),
);
let blended = _mm_blendv_epi8(merge.val.0[1], compressed, prefix);
crate::transmute::checked_transmute_store(
blended,
output[16..].first_chunk_mut::<16>().unwrap(),
);
let remaining = output_lane.saturating_sub(32).min(16);
let [prefix, _] = crate::transmute::checked_transmute_copy::<_, [__m128i; 2]>(
&crate::support::COMPACT_PREFIX_MASKS[remaining],
);
let compressed = crate::transmute::checked_transmute_copy(
output[32..].first_chunk::<16>().unwrap(),
);
let blended = _mm_blendv_epi8(merge.val.0[2], compressed, prefix);
crate::transmute::checked_transmute_store(
blended,
output[32..].first_chunk_mut::<16>().unwrap(),
);
let remaining = output_lane.saturating_sub(48).min(16);
let [prefix, _] = crate::transmute::checked_transmute_copy::<_, [__m128i; 2]>(
&crate::support::COMPACT_PREFIX_MASKS[remaining],
);
let compressed = crate::transmute::checked_transmute_copy(
output[48..].first_chunk::<16>().unwrap(),
);
let blended = _mm_blendv_epi8(merge.val.0[3], compressed, prefix);
crate::transmute::checked_transmute_store(
blended,
output[48..].first_chunk_mut::<16>().unwrap(),
);
u8x64::simd_from(token, output)
}
);
kernel(self, values, mask, merge)
}
#[inline(always)]
fn expand_u8x64(self, values: u8x64<Self>, mask: mask8x64<Self>) -> u8x64<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
values: u8x64<Sse4_2>,
mask: mask8x64<Sse4_2>,
) -> u8x64<Sse4_2> {
let input = <[u8; 64]>::from(values);
let mut output = [0u8; 64];
let mut input_lane = 0;
let block_bits = _mm_movemask_epi8(mask.val.0[0]) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_count = low_mask.count_ones() as u64;
let low_control = crate::support::EXPAND_8_CONTROLS[low_mask];
let high_control = crate::support::EXPAND_8_CONTROLS[high_mask];
let high_base = low_count * 0x0101_0101_0101_0101;
let high_control = high_control + high_base;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let packed = crate::transmute::checked_transmute_copy(
input[input_lane..].first_chunk::<16>().unwrap(),
);
let expanded = _mm_shuffle_epi8(packed, control);
let result = expanded;
crate::transmute::checked_transmute_store(
result,
output[0..].first_chunk_mut::<16>().unwrap(),
);
input_lane += block_bits.count_ones() as usize;
let block_bits = _mm_movemask_epi8(mask.val.0[1]) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_count = low_mask.count_ones() as u64;
let low_control = crate::support::EXPAND_8_CONTROLS[low_mask];
let high_control = crate::support::EXPAND_8_CONTROLS[high_mask];
let high_base = low_count * 0x0101_0101_0101_0101;
let high_control = high_control + high_base;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let packed = crate::transmute::checked_transmute_copy(
input[input_lane..].first_chunk::<16>().unwrap(),
);
let expanded = _mm_shuffle_epi8(packed, control);
let result = expanded;
crate::transmute::checked_transmute_store(
result,
output[16..].first_chunk_mut::<16>().unwrap(),
);
input_lane += block_bits.count_ones() as usize;
let block_bits = _mm_movemask_epi8(mask.val.0[2]) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_count = low_mask.count_ones() as u64;
let low_control = crate::support::EXPAND_8_CONTROLS[low_mask];
let high_control = crate::support::EXPAND_8_CONTROLS[high_mask];
let high_base = low_count * 0x0101_0101_0101_0101;
let high_control = high_control + high_base;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let packed = crate::transmute::checked_transmute_copy(
input[input_lane..].first_chunk::<16>().unwrap(),
);
let expanded = _mm_shuffle_epi8(packed, control);
let result = expanded;
crate::transmute::checked_transmute_store(
result,
output[32..].first_chunk_mut::<16>().unwrap(),
);
input_lane += block_bits.count_ones() as usize;
let block_bits = _mm_movemask_epi8(mask.val.0[3]) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_count = low_mask.count_ones() as u64;
let low_control = crate::support::EXPAND_8_CONTROLS[low_mask];
let high_control = crate::support::EXPAND_8_CONTROLS[high_mask];
let high_base = low_count * 0x0101_0101_0101_0101;
let high_control = high_control + high_base;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let packed = crate::transmute::checked_transmute_copy(
input[input_lane..].first_chunk::<16>().unwrap(),
);
let expanded = _mm_shuffle_epi8(packed, control);
let result = expanded;
crate::transmute::checked_transmute_store(
result,
output[48..].first_chunk_mut::<16>().unwrap(),
);
u8x64::simd_from(token, output)
}
);
kernel(self, values, mask)
}
#[inline(always)]
fn expand_merge_u8x64(
self,
values: u8x64<Self>,
mask: mask8x64<Self>,
merge: u8x64<Self>,
) -> u8x64<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(
token: Sse4_2,
values: u8x64<Sse4_2>,
mask: mask8x64<Sse4_2>,
merge: u8x64<Sse4_2>,
) -> u8x64<Sse4_2> {
let input = <[u8; 64]>::from(values);
let mut output = [0u8; 64];
let mut input_lane = 0;
let block_bits = _mm_movemask_epi8(mask.val.0[0]) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_count = low_mask.count_ones() as u64;
let low_control = crate::support::EXPAND_8_CONTROLS[low_mask];
let high_control = crate::support::EXPAND_8_CONTROLS[high_mask];
let high_base = low_count * 0x0101_0101_0101_0101;
let high_control = high_control + high_base;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let packed = crate::transmute::checked_transmute_copy(
input[input_lane..].first_chunk::<16>().unwrap(),
);
let expanded = _mm_shuffle_epi8(packed, control);
let result = _mm_blendv_epi8(merge.val.0[0], expanded, mask.val.0[0]);
crate::transmute::checked_transmute_store(
result,
output[0..].first_chunk_mut::<16>().unwrap(),
);
input_lane += block_bits.count_ones() as usize;
let block_bits = _mm_movemask_epi8(mask.val.0[1]) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_count = low_mask.count_ones() as u64;
let low_control = crate::support::EXPAND_8_CONTROLS[low_mask];
let high_control = crate::support::EXPAND_8_CONTROLS[high_mask];
let high_base = low_count * 0x0101_0101_0101_0101;
let high_control = high_control + high_base;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let packed = crate::transmute::checked_transmute_copy(
input[input_lane..].first_chunk::<16>().unwrap(),
);
let expanded = _mm_shuffle_epi8(packed, control);
let result = _mm_blendv_epi8(merge.val.0[1], expanded, mask.val.0[1]);
crate::transmute::checked_transmute_store(
result,
output[16..].first_chunk_mut::<16>().unwrap(),
);
input_lane += block_bits.count_ones() as usize;
let block_bits = _mm_movemask_epi8(mask.val.0[2]) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_count = low_mask.count_ones() as u64;
let low_control = crate::support::EXPAND_8_CONTROLS[low_mask];
let high_control = crate::support::EXPAND_8_CONTROLS[high_mask];
let high_base = low_count * 0x0101_0101_0101_0101;
let high_control = high_control + high_base;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let packed = crate::transmute::checked_transmute_copy(
input[input_lane..].first_chunk::<16>().unwrap(),
);
let expanded = _mm_shuffle_epi8(packed, control);
let result = _mm_blendv_epi8(merge.val.0[2], expanded, mask.val.0[2]);
crate::transmute::checked_transmute_store(
result,
output[32..].first_chunk_mut::<16>().unwrap(),
);
input_lane += block_bits.count_ones() as usize;
let block_bits = _mm_movemask_epi8(mask.val.0[3]) as usize;
let low_mask = (block_bits) & 0xff;
let high_mask = (block_bits) >> 8;
let low_count = low_mask.count_ones() as u64;
let low_control = crate::support::EXPAND_8_CONTROLS[low_mask];
let high_control = crate::support::EXPAND_8_CONTROLS[high_mask];
let high_base = low_count * 0x0101_0101_0101_0101;
let high_control = high_control + high_base;
let control = _mm_set_epi64x(high_control.cast_signed(), low_control.cast_signed());
let packed = crate::transmute::checked_transmute_copy(
input[input_lane..].first_chunk::<16>().unwrap(),
);
let expanded = _mm_shuffle_epi8(packed, control);
let result = _mm_blendv_epi8(merge.val.0[3], expanded, mask.val.0[3]);
crate::transmute::checked_transmute_store(
result,
output[48..].first_chunk_mut::<16>().unwrap(),
);
u8x64::simd_from(token, output)
}
);
kernel(self, values, mask, merge)
}
#[inline(always)]
fn split_u8x64(self, a: u8x64<Self>) -> (u8x32<Self>, u8x32<Self>) {
(
u8x32 {
val: crate::support::Aligned256([a.val.0[0], a.val.0[1]]),
simd: self,
},
u8x32 {
val: crate::support::Aligned256([a.val.0[2], a.val.0[3]]),
simd: self,
},
)
}
#[inline(always)]
fn from_bitmask_mask8x64(self, bits: u64) -> mask8x64<Self> {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, bits: u64) -> mask8x64<Sse4_2> {
{
let bit_bytes = _mm_set1_epi64x(bits.cast_signed());
let bit_mask =
_mm_setr_epi8(1, 2, 4, 8, 16, 32, 64, -128, 1, 2, 4, 8, 16, 32, 64, -128);
mask8x64 {
val: crate::support::Aligned512([
{
let bit_bytes = _mm_shuffle_epi8(
bit_bytes,
_mm_setr_epi8(0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1),
);
_mm_cmpeq_epi8(_mm_and_si128(bit_bytes, bit_mask), bit_mask)
},
{
let bit_bytes = _mm_shuffle_epi8(
bit_bytes,
_mm_setr_epi8(2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3),
);
_mm_cmpeq_epi8(_mm_and_si128(bit_bytes, bit_mask), bit_mask)
},
{
let bit_bytes = _mm_shuffle_epi8(
bit_bytes,
_mm_setr_epi8(4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5),
);
_mm_cmpeq_epi8(_mm_and_si128(bit_bytes, bit_mask), bit_mask)
},
{
let bit_bytes = _mm_shuffle_epi8(
bit_bytes,
_mm_setr_epi8(6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7),
);
_mm_cmpeq_epi8(_mm_and_si128(bit_bytes, bit_mask), bit_mask)
},
]),
simd: token,
}
}
}
);
kernel(self, bits)
}
#[inline(always)]
fn from_vector_mask8x64(self, vector: i8x64<Self>) -> mask8x64<Self> {
mask8x64 {
val: vector.val,
simd: self,
}
}
#[inline(always)]
fn to_vector_mask8x64(self, a: mask8x64<Self>) -> i8x64<Self> {
i8x64 {
val: a.val,
simd: self,
}
}
#[inline(always)]
fn set_mask8x64(self, a: &mut mask8x64<Self>, index: usize, value: bool) -> () {
assert!(
index < 64usize,
"mask lane index {index} is out of bounds for {} lanes",
64usize
);
let mut lanes: [i8; 64usize] = (*a).into();
lanes[index] = if value { !0 } else { 0 };
*a = lanes.simd_into(self);
}
#[inline(always)]
fn rotate_elements_left_mask8x64<const OFFSET: usize>(
self,
a: mask8x64<Self>,
) -> mask8x64<Self> {
let int = i8x64 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_left::<OFFSET>();
mask8x64 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn rotate_elements_right_mask8x64<const OFFSET: usize>(
self,
a: mask8x64<Self>,
) -> mask8x64<Self> {
let int = i8x64 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_right::<OFFSET>();
mask8x64 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn split_mask8x64(self, a: mask8x64<Self>) -> (mask8x32<Self>, mask8x32<Self>) {
(
mask8x32 {
val: crate::support::Aligned256([a.val.0[0], a.val.0[1]]),
simd: self,
},
mask8x32 {
val: crate::support::Aligned256([a.val.0[2], a.val.0[3]]),
simd: self,
},
)
}
#[inline(always)]
fn slide_i16x32<const SHIFT: usize>(self, a: i16x32<Self>, b: i16x32<Self>) -> i16x32<Self> {
if SHIFT >= 32usize {
return b;
}
let result = cross_block_alignr_128x4(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 2usize,
);
Bytes::from_bytes(u8x64 {
val: crate::support::Aligned512(result),
simd: self,
})
}
#[inline(always)]
fn compress_i16x32(self, values: i16x32<Self>, mask: mask16x32<Self>) -> i16x32<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_i16x32(
self,
values: i16x32<Self>,
mask: mask16x32<Self>,
merge: i16x32<Self>,
) -> i16x32<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_i16x32(self, values: i16x32<Self>, mask: mask16x32<Self>) -> i16x32<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_i16x32(
self,
values: i16x32<Self>,
mask: mask16x32<Self>,
merge: i16x32<Self>,
) -> i16x32<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn split_i16x32(self, a: i16x32<Self>) -> (i16x16<Self>, i16x16<Self>) {
(
i16x16 {
val: crate::support::Aligned256([a.val.0[0], a.val.0[1]]),
simd: self,
},
i16x16 {
val: crate::support::Aligned256([a.val.0[2], a.val.0[3]]),
simd: self,
},
)
}
#[inline(always)]
fn slide_u16x32<const SHIFT: usize>(self, a: u16x32<Self>, b: u16x32<Self>) -> u16x32<Self> {
if SHIFT >= 32usize {
return b;
}
let result = cross_block_alignr_128x4(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 2usize,
);
Bytes::from_bytes(u8x64 {
val: crate::support::Aligned512(result),
simd: self,
})
}
#[inline(always)]
fn compress_u16x32(self, values: u16x32<Self>, mask: mask16x32<Self>) -> u16x32<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_u16x32(
self,
values: u16x32<Self>,
mask: mask16x32<Self>,
merge: u16x32<Self>,
) -> u16x32<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_u16x32(self, values: u16x32<Self>, mask: mask16x32<Self>) -> u16x32<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_u16x32(
self,
values: u16x32<Self>,
mask: mask16x32<Self>,
merge: u16x32<Self>,
) -> u16x32<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn split_u16x32(self, a: u16x32<Self>) -> (u16x16<Self>, u16x16<Self>) {
(
u16x16 {
val: crate::support::Aligned256([a.val.0[0], a.val.0[1]]),
simd: self,
},
u16x16 {
val: crate::support::Aligned256([a.val.0[2], a.val.0[3]]),
simd: self,
},
)
}
#[inline(always)]
fn to_bitmask_mask16x32(self, a: mask16x32<Self>) -> u64 {
crate::kernel!(
#[inline(always)]
fn kernel(token: Sse4_2, a: mask16x32<Sse4_2>) -> u64 {
{
let lo = _mm_packs_epi16(a.val.0[0], a.val.0[1]);
let hi = _mm_packs_epi16(a.val.0[2], a.val.0[3]);
let lo = _mm_movemask_epi8(lo) as u32 as u64;
let hi = _mm_movemask_epi8(hi) as u32 as u64;
lo | (hi << 16usize)
}
}
);
kernel(self, a)
}
#[inline(always)]
fn from_vector_mask16x32(self, vector: i16x32<Self>) -> mask16x32<Self> {
mask16x32 {
val: vector.val,
simd: self,
}
}
#[inline(always)]
fn to_vector_mask16x32(self, a: mask16x32<Self>) -> i16x32<Self> {
i16x32 {
val: a.val,
simd: self,
}
}
#[inline(always)]
fn set_mask16x32(self, a: &mut mask16x32<Self>, index: usize, value: bool) -> () {
assert!(
index < 32usize,
"mask lane index {index} is out of bounds for {} lanes",
32usize
);
let mut lanes: [i16; 32usize] = (*a).into();
lanes[index] = if value { !0 } else { 0 };
*a = lanes.simd_into(self);
}
#[inline(always)]
fn rotate_elements_left_mask16x32<const OFFSET: usize>(
self,
a: mask16x32<Self>,
) -> mask16x32<Self> {
let int = i16x32 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_left::<OFFSET>();
mask16x32 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn rotate_elements_right_mask16x32<const OFFSET: usize>(
self,
a: mask16x32<Self>,
) -> mask16x32<Self> {
let int = i16x32 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_right::<OFFSET>();
mask16x32 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn split_mask16x32(self, a: mask16x32<Self>) -> (mask16x16<Self>, mask16x16<Self>) {
(
mask16x16 {
val: crate::support::Aligned256([a.val.0[0], a.val.0[1]]),
simd: self,
},
mask16x16 {
val: crate::support::Aligned256([a.val.0[2], a.val.0[3]]),
simd: self,
},
)
}
#[inline(always)]
fn slide_i32x16<const SHIFT: usize>(self, a: i32x16<Self>, b: i32x16<Self>) -> i32x16<Self> {
if SHIFT >= 16usize {
return b;
}
let result = cross_block_alignr_128x4(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 4usize,
);
Bytes::from_bytes(u8x64 {
val: crate::support::Aligned512(result),
simd: self,
})
}
#[inline(always)]
fn compress_i32x16(self, values: i32x16<Self>, mask: mask32x16<Self>) -> i32x16<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_i32x16(
self,
values: i32x16<Self>,
mask: mask32x16<Self>,
merge: i32x16<Self>,
) -> i32x16<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_i32x16(self, values: i32x16<Self>, mask: mask32x16<Self>) -> i32x16<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_i32x16(
self,
values: i32x16<Self>,
mask: mask32x16<Self>,
merge: i32x16<Self>,
) -> i32x16<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn split_i32x16(self, a: i32x16<Self>) -> (i32x8<Self>, i32x8<Self>) {
(
i32x8 {
val: crate::support::Aligned256([a.val.0[0], a.val.0[1]]),
simd: self,
},
i32x8 {
val: crate::support::Aligned256([a.val.0[2], a.val.0[3]]),
simd: self,
},
)
}
#[inline(always)]
fn slide_u32x16<const SHIFT: usize>(self, a: u32x16<Self>, b: u32x16<Self>) -> u32x16<Self> {
if SHIFT >= 16usize {
return b;
}
let result = cross_block_alignr_128x4(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 4usize,
);
Bytes::from_bytes(u8x64 {
val: crate::support::Aligned512(result),
simd: self,
})
}
#[inline(always)]
fn compress_u32x16(self, values: u32x16<Self>, mask: mask32x16<Self>) -> u32x16<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_u32x16(
self,
values: u32x16<Self>,
mask: mask32x16<Self>,
merge: u32x16<Self>,
) -> u32x16<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_u32x16(self, values: u32x16<Self>, mask: mask32x16<Self>) -> u32x16<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_u32x16(
self,
values: u32x16<Self>,
mask: mask32x16<Self>,
merge: u32x16<Self>,
) -> u32x16<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn split_u32x16(self, a: u32x16<Self>) -> (u32x8<Self>, u32x8<Self>) {
(
u32x8 {
val: crate::support::Aligned256([a.val.0[0], a.val.0[1]]),
simd: self,
},
u32x8 {
val: crate::support::Aligned256([a.val.0[2], a.val.0[3]]),
simd: self,
},
)
}
#[inline(always)]
fn from_vector_mask32x16(self, vector: i32x16<Self>) -> mask32x16<Self> {
mask32x16 {
val: vector.val,
simd: self,
}
}
#[inline(always)]
fn to_vector_mask32x16(self, a: mask32x16<Self>) -> i32x16<Self> {
i32x16 {
val: a.val,
simd: self,
}
}
#[inline(always)]
fn set_mask32x16(self, a: &mut mask32x16<Self>, index: usize, value: bool) -> () {
assert!(
index < 16usize,
"mask lane index {index} is out of bounds for {} lanes",
16usize
);
let mut lanes: [i32; 16usize] = (*a).into();
lanes[index] = if value { !0 } else { 0 };
*a = lanes.simd_into(self);
}
#[inline(always)]
fn rotate_elements_left_mask32x16<const OFFSET: usize>(
self,
a: mask32x16<Self>,
) -> mask32x16<Self> {
let int = i32x16 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_left::<OFFSET>();
mask32x16 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn rotate_elements_right_mask32x16<const OFFSET: usize>(
self,
a: mask32x16<Self>,
) -> mask32x16<Self> {
let int = i32x16 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_right::<OFFSET>();
mask32x16 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn split_mask32x16(self, a: mask32x16<Self>) -> (mask32x8<Self>, mask32x8<Self>) {
(
mask32x8 {
val: crate::support::Aligned256([a.val.0[0], a.val.0[1]]),
simd: self,
},
mask32x8 {
val: crate::support::Aligned256([a.val.0[2], a.val.0[3]]),
simd: self,
},
)
}
#[inline(always)]
fn slide_f64x8<const SHIFT: usize>(self, a: f64x8<Self>, b: f64x8<Self>) -> f64x8<Self> {
if SHIFT >= 8usize {
return b;
}
let result = cross_block_alignr_128x4(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 8usize,
);
Bytes::from_bytes(u8x64 {
val: crate::support::Aligned512(result),
simd: self,
})
}
#[inline(always)]
fn compress_f64x8(self, values: f64x8<Self>, mask: mask64x8<Self>) -> f64x8<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_f64x8(
self,
values: f64x8<Self>,
mask: mask64x8<Self>,
merge: f64x8<Self>,
) -> f64x8<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_f64x8(self, values: f64x8<Self>, mask: mask64x8<Self>) -> f64x8<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_f64x8(
self,
values: f64x8<Self>,
mask: mask64x8<Self>,
merge: f64x8<Self>,
) -> f64x8<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn simd_ne_f64x8(self, a: f64x8<Self>, b: f64x8<Self>) -> mask64x8<Self> {
let (a0, a1) = self.split_f64x8(a);
let (b0, b1) = self.split_f64x8(b);
self.combine_mask64x4(self.simd_ne_f64x4(a0, b0), self.simd_ne_f64x4(a1, b1))
}
#[inline(always)]
fn is_nan_f64x8(self, a: f64x8<Self>) -> mask64x8<Self> {
a.simd_ne(a)
}
#[inline(always)]
fn is_infinite_f64x8(self, a: f64x8<Self>) -> mask64x8<Self> {
a.abs().simd_eq(f64::INFINITY)
}
#[inline(always)]
fn is_finite_f64x8(self, a: f64x8<Self>) -> mask64x8<Self> {
a.abs().simd_lt(f64::INFINITY)
}
#[inline(always)]
fn is_subnormal_f64x8(self, a: f64x8<Self>) -> mask64x8<Self> {
let i: i64x8<Self> = a.bitcast();
(i & 0x7FF0_0000_0000_0000_i64).simd_eq(0) & (i & 0x000F_FFFF_FFFF_FFFF_i64).simd_ne(0)
}
#[inline(always)]
fn is_normal_f64x8(self, a: f64x8<Self>) -> mask64x8<Self> {
let i: i64x8<Self> = a.bitcast();
let exp = i & 0x7FF0_0000_0000_0000_i64;
exp.simd_ne(0) & exp.simd_ne(0x7FF0_0000_0000_0000_i64)
}
#[inline(always)]
fn is_sign_positive_f64x8(self, a: f64x8<Self>) -> mask64x8<Self> {
let i: i64x8<Self> = a.bitcast();
(i & i64::MIN).simd_eq(0)
}
#[inline(always)]
fn is_sign_negative_f64x8(self, a: f64x8<Self>) -> mask64x8<Self> {
!a.is_sign_positive()
}
#[inline(always)]
fn split_f64x8(self, a: f64x8<Self>) -> (f64x4<Self>, f64x4<Self>) {
(
f64x4 {
val: crate::support::Aligned256([a.val.0[0], a.val.0[1]]),
simd: self,
},
f64x4 {
val: crate::support::Aligned256([a.val.0[2], a.val.0[3]]),
simd: self,
},
)
}
#[inline(always)]
fn slide_i64x8<const SHIFT: usize>(self, a: i64x8<Self>, b: i64x8<Self>) -> i64x8<Self> {
if SHIFT >= 8usize {
return b;
}
let result = cross_block_alignr_128x4(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 8usize,
);
Bytes::from_bytes(u8x64 {
val: crate::support::Aligned512(result),
simd: self,
})
}
#[inline(always)]
fn compress_i64x8(self, values: i64x8<Self>, mask: mask64x8<Self>) -> i64x8<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_i64x8(
self,
values: i64x8<Self>,
mask: mask64x8<Self>,
merge: i64x8<Self>,
) -> i64x8<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_i64x8(self, values: i64x8<Self>, mask: mask64x8<Self>) -> i64x8<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_i64x8(
self,
values: i64x8<Self>,
mask: mask64x8<Self>,
merge: i64x8<Self>,
) -> i64x8<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn split_i64x8(self, a: i64x8<Self>) -> (i64x4<Self>, i64x4<Self>) {
(
i64x4 {
val: crate::support::Aligned256([a.val.0[0], a.val.0[1]]),
simd: self,
},
i64x4 {
val: crate::support::Aligned256([a.val.0[2], a.val.0[3]]),
simd: self,
},
)
}
#[inline(always)]
fn slide_u64x8<const SHIFT: usize>(self, a: u64x8<Self>, b: u64x8<Self>) -> u64x8<Self> {
if SHIFT >= 8usize {
return b;
}
let result = cross_block_alignr_128x4(
self,
Bytes::to_bytes(b).val.0,
Bytes::to_bytes(a).val.0,
SHIFT * 8usize,
);
Bytes::from_bytes(u8x64 {
val: crate::support::Aligned512(result),
simd: self,
})
}
#[inline(always)]
fn compress_u64x8(self, values: u64x8<Self>, mask: mask64x8<Self>) -> u64x8<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn compress_merge_u64x8(
self,
values: u64x8<Self>,
mask: mask64x8<Self>,
merge: u64x8<Self>,
) -> u64x8<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.compress_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn expand_u64x8(self, values: u64x8<Self>, mask: mask64x8<Self>) -> u64x8<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_u8x64(Bytes::to_bytes(values), mask))
}
#[inline(always)]
fn expand_merge_u64x8(
self,
values: u64x8<Self>,
mask: mask64x8<Self>,
merge: u64x8<Self>,
) -> u64x8<Self> {
let mask = mask8x64 {
val: crate::transmute::checked_transmute_copy(&mask.val),
simd: self,
};
Bytes::from_bytes(self.expand_merge_u8x64(
Bytes::to_bytes(values),
mask,
Bytes::to_bytes(merge),
))
}
#[inline(always)]
fn split_u64x8(self, a: u64x8<Self>) -> (u64x4<Self>, u64x4<Self>) {
(
u64x4 {
val: crate::support::Aligned256([a.val.0[0], a.val.0[1]]),
simd: self,
},
u64x4 {
val: crate::support::Aligned256([a.val.0[2], a.val.0[3]]),
simd: self,
},
)
}
#[inline(always)]
fn from_vector_mask64x8(self, vector: i64x8<Self>) -> mask64x8<Self> {
mask64x8 {
val: vector.val,
simd: self,
}
}
#[inline(always)]
fn to_vector_mask64x8(self, a: mask64x8<Self>) -> i64x8<Self> {
i64x8 {
val: a.val,
simd: self,
}
}
#[inline(always)]
fn set_mask64x8(self, a: &mut mask64x8<Self>, index: usize, value: bool) -> () {
assert!(
index < 8usize,
"mask lane index {index} is out of bounds for {} lanes",
8usize
);
let mut lanes: [i64; 8usize] = (*a).into();
lanes[index] = if value { !0 } else { 0 };
*a = lanes.simd_into(self);
}
#[inline(always)]
fn rotate_elements_left_mask64x8<const OFFSET: usize>(
self,
a: mask64x8<Self>,
) -> mask64x8<Self> {
let int = i64x8 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_left::<OFFSET>();
mask64x8 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn rotate_elements_right_mask64x8<const OFFSET: usize>(
self,
a: mask64x8<Self>,
) -> mask64x8<Self> {
let int = i64x8 {
val: crate::transmute::checked_transmute_copy(&a.val),
simd: self,
};
let rotated = int.rotate_elements_right::<OFFSET>();
mask64x8 {
val: crate::transmute::checked_transmute_copy(&rotated.val),
simd: self,
}
}
#[inline(always)]
fn split_mask64x8(self, a: mask64x8<Self>) -> (mask64x4<Self>, mask64x4<Self>) {
(
mask64x4 {
val: crate::support::Aligned256([a.val.0[0], a.val.0[1]]),
simd: self,
},
mask64x4 {
val: crate::support::Aligned256([a.val.0[2], a.val.0[3]]),
simd: self,
},
)
}
}
crate::kernel!(
#[doc = r" This is a version of the `alignr` intrinsic that takes a non-const shift argument. The shift is still"]
#[doc = r" expected to be constant in practice, so the match statement will be optimized out. This exists because"]
#[doc = r" Rust doesn't currently let you do math on const generics."]
#[inline(always)]
fn dyn_alignr_128(token: Sse4_2, a: __m128i, b: __m128i, shift: usize) -> __m128i {
match shift {
0usize => _mm_alignr_epi8::<0i32>(a, b),
1usize => _mm_alignr_epi8::<1i32>(a, b),
2usize => _mm_alignr_epi8::<2i32>(a, b),
3usize => _mm_alignr_epi8::<3i32>(a, b),
4usize => _mm_alignr_epi8::<4i32>(a, b),
5usize => _mm_alignr_epi8::<5i32>(a, b),
6usize => _mm_alignr_epi8::<6i32>(a, b),
7usize => _mm_alignr_epi8::<7i32>(a, b),
8usize => _mm_alignr_epi8::<8i32>(a, b),
9usize => _mm_alignr_epi8::<9i32>(a, b),
10usize => _mm_alignr_epi8::<10i32>(a, b),
11usize => _mm_alignr_epi8::<11i32>(a, b),
12usize => _mm_alignr_epi8::<12i32>(a, b),
13usize => _mm_alignr_epi8::<13i32>(a, b),
14usize => _mm_alignr_epi8::<14i32>(a, b),
15usize => _mm_alignr_epi8::<15i32>(a, b),
_ => unreachable!(),
}
}
);
crate::kernel!(
#[doc = r" Concatenates `b` and `a` (each N blocks) and extracts N blocks starting at byte offset `shift_bytes`."]
#[doc = r" Extracts from [b : a] (b in low bytes, a in high bytes), matching `alignr` semantics."]
#[inline(always)]
fn cross_block_alignr_128x2(
token: Sse4_2,
a: [__m128i; 2usize],
b: [__m128i; 2usize],
shift_bytes: usize,
) -> [__m128i; 2usize] {
[
{
let [lo, hi] =
crate::support::cross_block_slide_blocks_at(&b, &a, 0usize, shift_bytes);
dyn_alignr_128(token, hi, lo, shift_bytes % 16)
},
{
let [lo, hi] =
crate::support::cross_block_slide_blocks_at(&b, &a, 1usize, shift_bytes);
dyn_alignr_128(token, hi, lo, shift_bytes % 16)
},
]
}
);
crate::kernel!(
#[doc = r" Concatenates `b` and `a` (each N blocks) and extracts N blocks starting at byte offset `shift_bytes`."]
#[doc = r" Extracts from [b : a] (b in low bytes, a in high bytes), matching `alignr` semantics."]
#[inline(always)]
fn cross_block_alignr_128x4(
token: Sse4_2,
a: [__m128i; 4usize],
b: [__m128i; 4usize],
shift_bytes: usize,
) -> [__m128i; 4usize] {
[
{
let [lo, hi] =
crate::support::cross_block_slide_blocks_at(&b, &a, 0usize, shift_bytes);
dyn_alignr_128(token, hi, lo, shift_bytes % 16)
},
{
let [lo, hi] =
crate::support::cross_block_slide_blocks_at(&b, &a, 1usize, shift_bytes);
dyn_alignr_128(token, hi, lo, shift_bytes % 16)
},
{
let [lo, hi] =
crate::support::cross_block_slide_blocks_at(&b, &a, 2usize, shift_bytes);
dyn_alignr_128(token, hi, lo, shift_bytes % 16)
},
{
let [lo, hi] =
crate::support::cross_block_slide_blocks_at(&b, &a, 3usize, shift_bytes);
dyn_alignr_128(token, hi, lo, shift_bytes % 16)
},
]
}
);