mathieu/granite-rust
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Use new bitmask intrinsics with byte arrays

Browse source
This commit is contained in:
Caleb Zulawski 2021-09-13 00:48:26 +00:00 • committed by Jubilee Young
parent 1ce1c645cf
commit 36e198b97a
3 changed files with 9 additions and 39 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

9
crates/core_simd/src/lane_count.rs
View file

@ -15,34 +15,25 @@ impl<const LANES: usize> LaneCount<LANES> {
pub trait SupportedLaneCount: Sealed {
#[doc(hidden)]
type BitMask: Copy + Default + AsRef<[u8]> + AsMut<[u8]>;
#[doc(hidden)]
type IntBitMask;
}
impl<const LANES: usize> Sealed for LaneCount<LANES> {}
impl SupportedLaneCount for LaneCount<1> {
type BitMask = [u8; 1];
type IntBitMask = u8;
}
impl SupportedLaneCount for LaneCount<2> {
type BitMask = [u8; 1];
type IntBitMask = u8;
}
impl SupportedLaneCount for LaneCount<4> {
type BitMask = [u8; 1];
type IntBitMask = u8;
}
impl SupportedLaneCount for LaneCount<8> {
type BitMask = [u8; 1];
type IntBitMask = u8;
}
impl SupportedLaneCount for LaneCount<16> {
type BitMask = [u8; 2];
type IntBitMask = u16;
}
impl SupportedLaneCount for LaneCount<32> {
type BitMask = [u8; 4];
type IntBitMask = u32;
}

20
crates/core_simd/src/masks/bitmask.rs
View file

@ -1,3 +1,4 @@
#![allow(unused_imports)]
use super::MaskElement;
use crate::simd::intrinsics;
use crate::simd::{LaneCount, Simd, SupportedLaneCount};
@ -101,24 +102,17 @@ where
#[inline]
pub fn to_int(self) -> Simd<T, LANES> {
unsafe {
let mask: <LaneCount<LANES> as SupportedLaneCount>::IntBitMask =
core::mem::transmute_copy(&self);
intrinsics::simd_select_bitmask(mask, Simd::splat(T::TRUE), Simd::splat(T::FALSE))
crate::intrinsics::simd_select_bitmask(
self.0,
Simd::splat(T::TRUE),
Simd::splat(T::FALSE),
)
}
}
#[inline]
pub unsafe fn from_int_unchecked(value: Simd<T, LANES>) -> Self {
// TODO remove the transmute when rustc is more flexible
assert_eq!(
core::mem::size_of::<<LaneCount::<LANES> as SupportedLaneCount>::BitMask>(),
core::mem::size_of::<<LaneCount::<LANES> as SupportedLaneCount>::IntBitMask>(),
);
unsafe {
let mask: <LaneCount<LANES> as SupportedLaneCount>::IntBitMask =
intrinsics::simd_bitmask(value);
Self(core::mem::transmute_copy(&mask), PhantomData)
}
unsafe { Self(crate::intrinsics::simd_bitmask(value), PhantomData) }
}
#[cfg(feature = "generic_const_exprs")]

19
crates/core_simd/src/masks/full_masks.rs
View file

@ -106,15 +106,8 @@ where
#[inline]
pub fn to_bitmask(self) -> [u8; LaneCount::<LANES>::BITMASK_LEN] {
unsafe {
// TODO remove the transmute when rustc can use arrays of u8 as bitmasks
assert_eq!(
core::mem::size_of::<<LaneCount::<LANES> as SupportedLaneCount>::IntBitMask>(),
LaneCount::<LANES>::BITMASK_LEN,
);
let bitmask: <LaneCount<LANES> as SupportedLaneCount>::IntBitMask =
intrinsics::simd_bitmask(self.0);
let mut bitmask: [u8; LaneCount::<LANES>::BITMASK_LEN] =
core::mem::transmute_copy(&bitmask);
crate::intrinsics::simd_bitmask(self.0);
// There is a bug where LLVM appears to implement this operation with the wrong
// bit order.
@ -142,15 +135,7 @@ where
}
}
// TODO remove the transmute when rustc can use arrays of u8 as bitmasks
assert_eq!(
core::mem::size_of::<<LaneCount::<LANES> as SupportedLaneCount>::IntBitMask>(),
LaneCount::<LANES>::BITMASK_LEN,
);
let bitmask: <LaneCount<LANES> as SupportedLaneCount>::IntBitMask =
core::mem::transmute_copy(&bitmask);
Self::from_int_unchecked(intrinsics::simd_select_bitmask(
Self::from_int_unchecked(crate::intrinsics::simd_select_bitmask(
bitmask,
Self::splat(true).to_int(),
Self::splat(false).to_int(),