Implement simd_saturating_{add,sub}

This commit is contained in:
bjorn3 2022-09-05 16:13:36 +00:00
parent 0980596271
commit 782b5fe7ac
3 changed files with 39 additions and 106 deletions

View File

@ -11,110 +11,6 @@ Subject: [PATCH] Disable unsupported tests
crates/core_simd/tests/masks.rs | 3 --- crates/core_simd/tests/masks.rs | 3 ---
5 files changed, 20 insertions(+), 3 deletions(-) 5 files changed, 20 insertions(+), 3 deletions(-)
diff --git a/crates/core_simd/src/elements/int.rs b/crates/core_simd/src/elements/int.rs
index 9b8c37e..ea95f08 100644
--- a/crates/core_simd/src/elements/int.rs
+++ b/crates/core_simd/src/elements/int.rs
@@ -11,6 +11,7 @@ pub trait SimdInt: Copy + Sealed {
/// Scalar type contained by this SIMD vector type.
type Scalar;
+ /*
/// Lanewise saturating add.
///
/// # Examples
@@ -45,6 +46,7 @@ pub trait SimdInt: Copy + Sealed {
/// assert_eq!(unsat, Simd::from_array([1, MAX, MIN, 0]));
/// assert_eq!(sat, Simd::from_array([MIN, MIN, MIN, 0]));
fn saturating_sub(self, second: Self) -> Self;
+ */
/// Lanewise absolute value, implemented in Rust.
/// Every lane becomes its absolute value.
@@ -61,6 +63,7 @@ pub trait SimdInt: Copy + Sealed {
/// ```
fn abs(self) -> Self;
+ /*
/// Lanewise saturating absolute value, implemented in Rust.
/// As abs(), except the MIN value becomes MAX instead of itself.
///
@@ -96,6 +99,7 @@ pub trait SimdInt: Copy + Sealed {
/// assert_eq!(sat, Simd::from_array([MAX, 2, -3, MIN + 1]));
/// ```
fn saturating_neg(self) -> Self;
+ */
/// Returns true for each positive lane and false if it is zero or negative.
fn is_positive(self) -> Self::Mask;
@@ -199,6 +203,7 @@ macro_rules! impl_trait {
type Mask = Mask<<$ty as SimdElement>::Mask, LANES>;
type Scalar = $ty;
+ /*
#[inline]
fn saturating_add(self, second: Self) -> Self {
// Safety: `self` is a vector
@@ -210,6 +215,7 @@ macro_rules! impl_trait {
// Safety: `self` is a vector
unsafe { intrinsics::simd_saturating_sub(self, second) }
}
+ */
#[inline]
fn abs(self) -> Self {
@@ -218,6 +224,7 @@ macro_rules! impl_trait {
(self^m) - m
}
+ /*
#[inline]
fn saturating_abs(self) -> Self {
// arith shift for -1 or 0 mask based on sign bit, giving 2s complement
@@ -230,6 +237,7 @@ macro_rules! impl_trait {
fn saturating_neg(self) -> Self {
Self::splat(0).saturating_sub(self)
}
+ */
#[inline]
fn is_positive(self) -> Self::Mask {
diff --git a/crates/core_simd/src/elements/uint.rs b/crates/core_simd/src/elements/uint.rs
index 21e7e76..0d6dee2 100644
--- a/crates/core_simd/src/elements/uint.rs
+++ b/crates/core_simd/src/elements/uint.rs
@@ -6,6 +6,7 @@ pub trait SimdUint: Copy + Sealed {
/// Scalar type contained by this SIMD vector type.
type Scalar;
+ /*
/// Lanewise saturating add.
///
/// # Examples
@@ -40,6 +41,7 @@ pub trait SimdUint: Copy + Sealed {
/// assert_eq!(unsat, Simd::from_array([3, 2, 1, 0]));
/// assert_eq!(sat, Simd::splat(0));
fn saturating_sub(self, second: Self) -> Self;
+ */
/// Returns the sum of the lanes of the vector, with wrapping addition.
fn reduce_sum(self) -> Self::Scalar;
@@ -78,6 +80,7 @@ macro_rules! impl_trait {
{
type Scalar = $ty;
+ /*
#[inline]
fn saturating_add(self, second: Self) -> Self {
// Safety: `self` is a vector
@@ -89,6 +92,7 @@ macro_rules! impl_trait {
// Safety: `self` is a vector
unsafe { intrinsics::simd_saturating_sub(self, second) }
}
+ */
#[inline]
fn reduce_sum(self) -> Self::Scalar {
diff --git a/crates/core_simd/src/masks/full_masks.rs b/crates/core_simd/src/masks/full_masks.rs diff --git a/crates/core_simd/src/masks/full_masks.rs b/crates/core_simd/src/masks/full_masks.rs
index adf0fcb..e7e657e 100644 index adf0fcb..e7e657e 100644
--- a/crates/core_simd/src/masks/full_masks.rs --- a/crates/core_simd/src/masks/full_masks.rs

View File

@ -84,6 +84,30 @@ fn simd_for_each_lane<'tcx>(
} }
} }
fn simd_pair_for_each_lane_typed<'tcx>(
fx: &mut FunctionCx<'_, '_, 'tcx>,
x: CValue<'tcx>,
y: CValue<'tcx>,
ret: CPlace<'tcx>,
f: &dyn Fn(&mut FunctionCx<'_, '_, 'tcx>, CValue<'tcx>, CValue<'tcx>) -> CValue<'tcx>,
) {
assert_eq!(x.layout(), y.layout());
let layout = x.layout();
let (lane_count, _lane_ty) = layout.ty.simd_size_and_type(fx.tcx);
let (ret_lane_count, _ret_lane_ty) = ret.layout().ty.simd_size_and_type(fx.tcx);
assert_eq!(lane_count, ret_lane_count);
for lane_idx in 0..lane_count {
let x_lane = x.value_lane(fx, lane_idx);
let y_lane = y.value_lane(fx, lane_idx);
let res_lane = f(fx, x_lane, y_lane);
ret.place_lane(fx, lane_idx).write_cvalue(fx, res_lane);
}
}
fn simd_pair_for_each_lane<'tcx>( fn simd_pair_for_each_lane<'tcx>(
fx: &mut FunctionCx<'_, '_, 'tcx>, fx: &mut FunctionCx<'_, '_, 'tcx>,
x: CValue<'tcx>, x: CValue<'tcx>,

View File

@ -730,9 +730,22 @@ pub(super) fn codegen_simd_intrinsic_call<'tcx>(
ret.write_cvalue(fx, res); ret.write_cvalue(fx, res);
} }
sym::simd_saturating_add | sym::simd_saturating_sub => {
intrinsic_args!(fx, args => (x, y); intrinsic);
let bin_op = match intrinsic {
sym::simd_saturating_add => BinOp::Add,
sym::simd_saturating_sub => BinOp::Sub,
_ => unreachable!(),
};
// FIXME use vector instructions when possible
simd_pair_for_each_lane_typed(fx, x, y, ret, &|fx, x_lane, y_lane| {
crate::num::codegen_saturating_int_binop(fx, bin_op, x_lane, y_lane)
});
}
// simd_arith_offset // simd_arith_offset
// simd_saturating_add
// simd_saturating_sub
// simd_scatter // simd_scatter
// simd_gather // simd_gather
// simd_select_bitmask // simd_select_bitmask