diff options
Diffstat (limited to 'library/portable-simd/crates/core_simd/src/masks')
3 files changed, 662 insertions, 0 deletions
diff --git a/library/portable-simd/crates/core_simd/src/masks/bitmask.rs b/library/portable-simd/crates/core_simd/src/masks/bitmask.rs new file mode 100644 index 000000000..365ecc0a3 --- /dev/null +++ b/library/portable-simd/crates/core_simd/src/masks/bitmask.rs @@ -0,0 +1,246 @@ +#![allow(unused_imports)] +use super::MaskElement; +use crate::simd::intrinsics; +use crate::simd::{LaneCount, Simd, SupportedLaneCount, ToBitMask}; +use core::marker::PhantomData; + +/// A mask where each lane is represented by a single bit. +#[repr(transparent)] +pub struct Mask<T, const LANES: usize>( + <LaneCount<LANES> as SupportedLaneCount>::BitMask, + PhantomData<T>, +) +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount; + +impl<T, const LANES: usize> Copy for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ +} + +impl<T, const LANES: usize> Clone for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ + fn clone(&self) -> Self { + *self + } +} + +impl<T, const LANES: usize> PartialEq for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ + fn eq(&self, other: &Self) -> bool { + self.0.as_ref() == other.0.as_ref() + } +} + +impl<T, const LANES: usize> PartialOrd for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ + fn partial_cmp(&self, other: &Self) -> Option<core::cmp::Ordering> { + self.0.as_ref().partial_cmp(other.0.as_ref()) + } +} + +impl<T, const LANES: usize> Eq for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ +} + +impl<T, const LANES: usize> Ord for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ + fn cmp(&self, other: &Self) -> core::cmp::Ordering { + self.0.as_ref().cmp(other.0.as_ref()) + } +} + +impl<T, const LANES: usize> Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + pub fn splat(value: bool) -> Self { + let mut mask = <LaneCount<LANES> as SupportedLaneCount>::BitMask::default(); + if value { + mask.as_mut().fill(u8::MAX) + } else { + mask.as_mut().fill(u8::MIN) + } + if LANES % 8 > 0 { + *mask.as_mut().last_mut().unwrap() &= u8::MAX >> (8 - LANES % 8); + } + Self(mask, PhantomData) + } + + #[inline] + #[must_use = "method returns a new bool and does not mutate the original value"] + pub unsafe fn test_unchecked(&self, lane: usize) -> bool { + (self.0.as_ref()[lane / 8] >> (lane % 8)) & 0x1 > 0 + } + + #[inline] + pub unsafe fn set_unchecked(&mut self, lane: usize, value: bool) { + unsafe { + self.0.as_mut()[lane / 8] ^= ((value ^ self.test_unchecked(lane)) as u8) << (lane % 8) + } + } + + #[inline] + #[must_use = "method returns a new vector and does not mutate the original value"] + pub fn to_int(self) -> Simd<T, LANES> { + unsafe { + intrinsics::simd_select_bitmask(self.0, Simd::splat(T::TRUE), Simd::splat(T::FALSE)) + } + } + + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + pub unsafe fn from_int_unchecked(value: Simd<T, LANES>) -> Self { + unsafe { Self(intrinsics::simd_bitmask(value), PhantomData) } + } + + #[cfg(feature = "generic_const_exprs")] + #[inline] + #[must_use = "method returns a new array and does not mutate the original value"] + pub fn to_bitmask_array<const N: usize>(self) -> [u8; N] { + assert!(core::mem::size_of::<Self>() == N); + + // Safety: converting an integer to an array of bytes of the same size is safe + unsafe { core::mem::transmute_copy(&self.0) } + } + + #[cfg(feature = "generic_const_exprs")] + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + pub fn from_bitmask_array<const N: usize>(bitmask: [u8; N]) -> Self { + assert!(core::mem::size_of::<Self>() == N); + + // Safety: converting an array of bytes to an integer of the same size is safe + Self(unsafe { core::mem::transmute_copy(&bitmask) }, PhantomData) + } + + #[inline] + pub fn to_bitmask_integer<U>(self) -> U + where + super::Mask<T, LANES>: ToBitMask<BitMask = U>, + { + // Safety: these are the same types + unsafe { core::mem::transmute_copy(&self.0) } + } + + #[inline] + pub fn from_bitmask_integer<U>(bitmask: U) -> Self + where + super::Mask<T, LANES>: ToBitMask<BitMask = U>, + { + // Safety: these are the same types + unsafe { Self(core::mem::transmute_copy(&bitmask), PhantomData) } + } + + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + pub fn convert<U>(self) -> Mask<U, LANES> + where + U: MaskElement, + { + // Safety: bitmask layout does not depend on the element width + unsafe { core::mem::transmute_copy(&self) } + } + + #[inline] + #[must_use = "method returns a new bool and does not mutate the original value"] + pub fn any(self) -> bool { + self != Self::splat(false) + } + + #[inline] + #[must_use = "method returns a new bool and does not mutate the original value"] + pub fn all(self) -> bool { + self == Self::splat(true) + } +} + +impl<T, const LANES: usize> core::ops::BitAnd for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, + <LaneCount<LANES> as SupportedLaneCount>::BitMask: AsRef<[u8]> + AsMut<[u8]>, +{ + type Output = Self; + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + fn bitand(mut self, rhs: Self) -> Self { + for (l, r) in self.0.as_mut().iter_mut().zip(rhs.0.as_ref().iter()) { + *l &= r; + } + self + } +} + +impl<T, const LANES: usize> core::ops::BitOr for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, + <LaneCount<LANES> as SupportedLaneCount>::BitMask: AsRef<[u8]> + AsMut<[u8]>, +{ + type Output = Self; + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + fn bitor(mut self, rhs: Self) -> Self { + for (l, r) in self.0.as_mut().iter_mut().zip(rhs.0.as_ref().iter()) { + *l |= r; + } + self + } +} + +impl<T, const LANES: usize> core::ops::BitXor for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ + type Output = Self; + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + fn bitxor(mut self, rhs: Self) -> Self::Output { + for (l, r) in self.0.as_mut().iter_mut().zip(rhs.0.as_ref().iter()) { + *l ^= r; + } + self + } +} + +impl<T, const LANES: usize> core::ops::Not for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ + type Output = Self; + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + fn not(mut self) -> Self::Output { + for x in self.0.as_mut() { + *x = !*x; + } + if LANES % 8 > 0 { + *self.0.as_mut().last_mut().unwrap() &= u8::MAX >> (8 - LANES % 8); + } + self + } +} diff --git a/library/portable-simd/crates/core_simd/src/masks/full_masks.rs b/library/portable-simd/crates/core_simd/src/masks/full_masks.rs new file mode 100644 index 000000000..adf0fcbea --- /dev/null +++ b/library/portable-simd/crates/core_simd/src/masks/full_masks.rs @@ -0,0 +1,323 @@ +//! Masks that take up full SIMD vector registers. + +use super::MaskElement; +use crate::simd::intrinsics; +use crate::simd::{LaneCount, Simd, SupportedLaneCount, ToBitMask}; + +#[cfg(feature = "generic_const_exprs")] +use crate::simd::ToBitMaskArray; + +#[repr(transparent)] +pub struct Mask<T, const LANES: usize>(Simd<T, LANES>) +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount; + +impl<T, const LANES: usize> Copy for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ +} + +impl<T, const LANES: usize> Clone for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + fn clone(&self) -> Self { + *self + } +} + +impl<T, const LANES: usize> PartialEq for Mask<T, LANES> +where + T: MaskElement + PartialEq, + LaneCount<LANES>: SupportedLaneCount, +{ + fn eq(&self, other: &Self) -> bool { + self.0.eq(&other.0) + } +} + +impl<T, const LANES: usize> PartialOrd for Mask<T, LANES> +where + T: MaskElement + PartialOrd, + LaneCount<LANES>: SupportedLaneCount, +{ + fn partial_cmp(&self, other: &Self) -> Option<core::cmp::Ordering> { + self.0.partial_cmp(&other.0) + } +} + +impl<T, const LANES: usize> Eq for Mask<T, LANES> +where + T: MaskElement + Eq, + LaneCount<LANES>: SupportedLaneCount, +{ +} + +impl<T, const LANES: usize> Ord for Mask<T, LANES> +where + T: MaskElement + Ord, + LaneCount<LANES>: SupportedLaneCount, +{ + fn cmp(&self, other: &Self) -> core::cmp::Ordering { + self.0.cmp(&other.0) + } +} + +// Used for bitmask bit order workaround +pub(crate) trait ReverseBits { + // Reverse the least significant `n` bits of `self`. + // (Remaining bits must be 0.) + fn reverse_bits(self, n: usize) -> Self; +} + +macro_rules! impl_reverse_bits { + { $($int:ty),* } => { + $( + impl ReverseBits for $int { + #[inline(always)] + fn reverse_bits(self, n: usize) -> Self { + let rev = <$int>::reverse_bits(self); + let bitsize = core::mem::size_of::<$int>() * 8; + if n < bitsize { + // Shift things back to the right + rev >> (bitsize - n) + } else { + rev + } + } + } + )* + } +} + +impl_reverse_bits! { u8, u16, u32, u64 } + +impl<T, const LANES: usize> Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + pub fn splat(value: bool) -> Self { + Self(Simd::splat(if value { T::TRUE } else { T::FALSE })) + } + + #[inline] + #[must_use = "method returns a new bool and does not mutate the original value"] + pub unsafe fn test_unchecked(&self, lane: usize) -> bool { + T::eq(self.0[lane], T::TRUE) + } + + #[inline] + pub unsafe fn set_unchecked(&mut self, lane: usize, value: bool) { + self.0[lane] = if value { T::TRUE } else { T::FALSE } + } + + #[inline] + #[must_use = "method returns a new vector and does not mutate the original value"] + pub fn to_int(self) -> Simd<T, LANES> { + self.0 + } + + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + pub unsafe fn from_int_unchecked(value: Simd<T, LANES>) -> Self { + Self(value) + } + + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + pub fn convert<U>(self) -> Mask<U, LANES> + where + U: MaskElement, + { + // Safety: masks are simply integer vectors of 0 and -1, and we can cast the element type. + unsafe { Mask(intrinsics::simd_cast(self.0)) } + } + + #[cfg(feature = "generic_const_exprs")] + #[inline] + #[must_use = "method returns a new array and does not mutate the original value"] + pub fn to_bitmask_array<const N: usize>(self) -> [u8; N] + where + super::Mask<T, LANES>: ToBitMaskArray, + [(); <super::Mask<T, LANES> as ToBitMaskArray>::BYTES]: Sized, + { + assert_eq!(<super::Mask<T, LANES> as ToBitMaskArray>::BYTES, N); + + // Safety: N is the correct bitmask size + unsafe { + // Compute the bitmask + let bitmask: [u8; <super::Mask<T, LANES> as ToBitMaskArray>::BYTES] = + intrinsics::simd_bitmask(self.0); + + // Transmute to the return type, previously asserted to be the same size + let mut bitmask: [u8; N] = core::mem::transmute_copy(&bitmask); + + // LLVM assumes bit order should match endianness + if cfg!(target_endian = "big") { + for x in bitmask.as_mut() { + *x = x.reverse_bits(); + } + }; + + bitmask + } + } + + #[cfg(feature = "generic_const_exprs")] + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + pub fn from_bitmask_array<const N: usize>(mut bitmask: [u8; N]) -> Self + where + super::Mask<T, LANES>: ToBitMaskArray, + [(); <super::Mask<T, LANES> as ToBitMaskArray>::BYTES]: Sized, + { + assert_eq!(<super::Mask<T, LANES> as ToBitMaskArray>::BYTES, N); + + // Safety: N is the correct bitmask size + unsafe { + // LLVM assumes bit order should match endianness + if cfg!(target_endian = "big") { + for x in bitmask.as_mut() { + *x = x.reverse_bits(); + } + } + + // Transmute to the bitmask type, previously asserted to be the same size + let bitmask: [u8; <super::Mask<T, LANES> as ToBitMaskArray>::BYTES] = + core::mem::transmute_copy(&bitmask); + + // Compute the regular mask + Self::from_int_unchecked(intrinsics::simd_select_bitmask( + bitmask, + Self::splat(true).to_int(), + Self::splat(false).to_int(), + )) + } + } + + #[inline] + pub(crate) fn to_bitmask_integer<U: ReverseBits>(self) -> U + where + super::Mask<T, LANES>: ToBitMask<BitMask = U>, + { + // Safety: U is required to be the appropriate bitmask type + let bitmask: U = unsafe { intrinsics::simd_bitmask(self.0) }; + + // LLVM assumes bit order should match endianness + if cfg!(target_endian = "big") { + bitmask.reverse_bits(LANES) + } else { + bitmask + } + } + + #[inline] + pub(crate) fn from_bitmask_integer<U: ReverseBits>(bitmask: U) -> Self + where + super::Mask<T, LANES>: ToBitMask<BitMask = U>, + { + // LLVM assumes bit order should match endianness + let bitmask = if cfg!(target_endian = "big") { + bitmask.reverse_bits(LANES) + } else { + bitmask + }; + + // Safety: U is required to be the appropriate bitmask type + unsafe { + Self::from_int_unchecked(intrinsics::simd_select_bitmask( + bitmask, + Self::splat(true).to_int(), + Self::splat(false).to_int(), + )) + } + } + + #[inline] + #[must_use = "method returns a new bool and does not mutate the original value"] + pub fn any(self) -> bool { + // Safety: use `self` as an integer vector + unsafe { intrinsics::simd_reduce_any(self.to_int()) } + } + + #[inline] + #[must_use = "method returns a new vector and does not mutate the original value"] + pub fn all(self) -> bool { + // Safety: use `self` as an integer vector + unsafe { intrinsics::simd_reduce_all(self.to_int()) } + } +} + +impl<T, const LANES: usize> core::convert::From<Mask<T, LANES>> for Simd<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ + fn from(value: Mask<T, LANES>) -> Self { + value.0 + } +} + +impl<T, const LANES: usize> core::ops::BitAnd for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ + type Output = Self; + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + fn bitand(self, rhs: Self) -> Self { + // Safety: `self` is an integer vector + unsafe { Self(intrinsics::simd_and(self.0, rhs.0)) } + } +} + +impl<T, const LANES: usize> core::ops::BitOr for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ + type Output = Self; + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + fn bitor(self, rhs: Self) -> Self { + // Safety: `self` is an integer vector + unsafe { Self(intrinsics::simd_or(self.0, rhs.0)) } + } +} + +impl<T, const LANES: usize> core::ops::BitXor for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ + type Output = Self; + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + fn bitxor(self, rhs: Self) -> Self { + // Safety: `self` is an integer vector + unsafe { Self(intrinsics::simd_xor(self.0, rhs.0)) } + } +} + +impl<T, const LANES: usize> core::ops::Not for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ + type Output = Self; + #[inline] + #[must_use = "method returns a new mask and does not mutate the original value"] + fn not(self) -> Self::Output { + Self::splat(true) ^ self + } +} diff --git a/library/portable-simd/crates/core_simd/src/masks/to_bitmask.rs b/library/portable-simd/crates/core_simd/src/masks/to_bitmask.rs new file mode 100644 index 000000000..65d3ce9be --- /dev/null +++ b/library/portable-simd/crates/core_simd/src/masks/to_bitmask.rs @@ -0,0 +1,93 @@ +use super::{mask_impl, Mask, MaskElement}; +use crate::simd::{LaneCount, SupportedLaneCount}; + +mod sealed { + pub trait Sealed {} +} +pub use sealed::Sealed; + +impl<T, const LANES: usize> Sealed for Mask<T, LANES> +where + T: MaskElement, + LaneCount<LANES>: SupportedLaneCount, +{ +} + +/// Converts masks to and from integer bitmasks. +/// +/// Each bit of the bitmask corresponds to a mask lane, starting with the LSB. +pub trait ToBitMask: Sealed { + /// The integer bitmask type. + type BitMask; + + /// Converts a mask to a bitmask. + fn to_bitmask(self) -> Self::BitMask; + + /// Converts a bitmask to a mask. + fn from_bitmask(bitmask: Self::BitMask) -> Self; +} + +/// Converts masks to and from byte array bitmasks. +/// +/// Each bit of the bitmask corresponds to a mask lane, starting with the LSB of the first byte. +#[cfg(feature = "generic_const_exprs")] +pub trait ToBitMaskArray: Sealed { + /// The length of the bitmask array. + const BYTES: usize; + + /// Converts a mask to a bitmask. + fn to_bitmask_array(self) -> [u8; Self::BYTES]; + + /// Converts a bitmask to a mask. + fn from_bitmask_array(bitmask: [u8; Self::BYTES]) -> Self; +} + +macro_rules! impl_integer_intrinsic { + { $(impl ToBitMask<BitMask=$int:ty> for Mask<_, $lanes:literal>)* } => { + $( + impl<T: MaskElement> ToBitMask for Mask<T, $lanes> { + type BitMask = $int; + + fn to_bitmask(self) -> $int { + self.0.to_bitmask_integer() + } + + fn from_bitmask(bitmask: $int) -> Self { + Self(mask_impl::Mask::from_bitmask_integer(bitmask)) + } + } + )* + } +} + +impl_integer_intrinsic! { + impl ToBitMask<BitMask=u8> for Mask<_, 1> + impl ToBitMask<BitMask=u8> for Mask<_, 2> + impl ToBitMask<BitMask=u8> for Mask<_, 4> + impl ToBitMask<BitMask=u8> for Mask<_, 8> + impl ToBitMask<BitMask=u16> for Mask<_, 16> + impl ToBitMask<BitMask=u32> for Mask<_, 32> + impl ToBitMask<BitMask=u64> for Mask<_, 64> +} + +/// Returns the minimum numnber of bytes in a bitmask with `lanes` lanes. +#[cfg(feature = "generic_const_exprs")] +pub const fn bitmask_len(lanes: usize) -> usize { + (lanes + 7) / 8 +} + +#[cfg(feature = "generic_const_exprs")] +impl<T: MaskElement, const LANES: usize> ToBitMaskArray for Mask<T, LANES> +where + LaneCount<LANES>: SupportedLaneCount, +{ + const BYTES: usize = bitmask_len(LANES); + + fn to_bitmask_array(self) -> [u8; Self::BYTES] { + self.0.to_bitmask_array() + } + + fn from_bitmask_array(bitmask: [u8; Self::BYTES]) -> Self { + Mask(mask_impl::Mask::from_bitmask_array(bitmask)) + } +} |