1
Fork 0

Merge commit 'cff979eec1' into sync-portable-simd-2024-03-22

This commit is contained in:
Daniel Paoliello 2024-03-22 16:58:39 -07:00
commit 9e0ec251d5
13 changed files with 505 additions and 52 deletions

View file

@ -1,4 +1,3 @@
#![cfg_attr(feature = "as_crate", no_std)] // We are std!
#![cfg_attr(
feature = "as_crate",
feature(core_intrinsics),
@ -44,7 +43,7 @@ use crate::sealed::Sealed;
/// For now this trait is available to permit experimentation with SIMD float
/// operations that may lack hardware support, such as `mul_add`.
pub trait StdFloat: Sealed + Sized {
/// Fused multiply-add. Computes `(self * a) + b` with only one rounding error,
/// Elementwise fused multiply-add. Computes `(self * a) + b` with only one rounding error,
/// yielding a more accurate result than an unfused multiply-add.
///
/// Using `mul_add` *may* be more performant than an unfused multiply-add if the target
@ -57,22 +56,65 @@ pub trait StdFloat: Sealed + Sized {
unsafe { intrinsics::simd_fma(self, a, b) }
}
/// Produces a vector where every lane has the square root value
/// of the equivalently-indexed lane in `self`
/// Produces a vector where every element has the square root value
/// of the equivalently-indexed element in `self`
#[inline]
#[must_use = "method returns a new vector and does not mutate the original value"]
fn sqrt(self) -> Self {
unsafe { intrinsics::simd_fsqrt(self) }
}
/// Returns the smallest integer greater than or equal to each lane.
/// Produces a vector where every element has the sine of the value
/// in the equivalently-indexed element in `self`.
#[must_use = "method returns a new vector and does not mutate the original value"]
fn sin(self) -> Self;
/// Produces a vector where every element has the cosine of the value
/// in the equivalently-indexed element in `self`.
#[must_use = "method returns a new vector and does not mutate the original value"]
fn cos(self) -> Self;
/// Produces a vector where every element has the exponential (base e) of the value
/// in the equivalently-indexed element in `self`.
#[must_use = "method returns a new vector and does not mutate the original value"]
fn exp(self) -> Self;
/// Produces a vector where every element has the exponential (base 2) of the value
/// in the equivalently-indexed element in `self`.
#[must_use = "method returns a new vector and does not mutate the original value"]
fn exp2(self) -> Self;
/// Produces a vector where every element has the natural logarithm of the value
/// in the equivalently-indexed element in `self`.
#[must_use = "method returns a new vector and does not mutate the original value"]
fn ln(self) -> Self;
/// Produces a vector where every element has the logarithm with respect to an arbitrary
/// in the equivalently-indexed elements in `self` and `base`.
#[inline]
#[must_use = "method returns a new vector and does not mutate the original value"]
fn log(self, base: Self) -> Self {
unsafe { intrinsics::simd_div(self.ln(), base.ln()) }
}
/// Produces a vector where every element has the base-2 logarithm of the value
/// in the equivalently-indexed element in `self`.
#[must_use = "method returns a new vector and does not mutate the original value"]
fn log2(self) -> Self;
/// Produces a vector where every element has the base-10 logarithm of the value
/// in the equivalently-indexed element in `self`.
#[must_use = "method returns a new vector and does not mutate the original value"]
fn log10(self) -> Self;
/// Returns the smallest integer greater than or equal to each element.
#[must_use = "method returns a new vector and does not mutate the original value"]
#[inline]
fn ceil(self) -> Self {
unsafe { intrinsics::simd_ceil(self) }
}
/// Returns the largest integer value less than or equal to each lane.
/// Returns the largest integer value less than or equal to each element.
#[must_use = "method returns a new vector and does not mutate the original value"]
#[inline]
fn floor(self) -> Self {
@ -101,46 +143,65 @@ pub trait StdFloat: Sealed + Sized {
impl<const N: usize> Sealed for Simd<f32, N> where LaneCount<N>: SupportedLaneCount {}
impl<const N: usize> Sealed for Simd<f64, N> where LaneCount<N>: SupportedLaneCount {}
// We can safely just use all the defaults.
impl<const N: usize> StdFloat for Simd<f32, N>
where
LaneCount<N>: SupportedLaneCount,
{
/// Returns the floating point's fractional value, with its integer part removed.
#[must_use = "method returns a new vector and does not mutate the original value"]
#[inline]
fn fract(self) -> Self {
self - self.trunc()
macro_rules! impl_float {
{
$($fn:ident: $intrinsic:ident,)*
} => {
impl<const N: usize> StdFloat for Simd<f32, N>
where
LaneCount<N>: SupportedLaneCount,
{
#[inline]
fn fract(self) -> Self {
self - self.trunc()
}
$(
#[inline]
fn $fn(self) -> Self {
unsafe { intrinsics::$intrinsic(self) }
}
)*
}
impl<const N: usize> StdFloat for Simd<f64, N>
where
LaneCount<N>: SupportedLaneCount,
{
#[inline]
fn fract(self) -> Self {
self - self.trunc()
}
$(
#[inline]
fn $fn(self) -> Self {
// https://github.com/llvm/llvm-project/issues/83729
#[cfg(target_arch = "aarch64")]
{
let mut ln = Self::splat(0f64);
for i in 0..N {
ln[i] = self[i].$fn()
}
ln
}
#[cfg(not(target_arch = "aarch64"))]
{
unsafe { intrinsics::$intrinsic(self) }
}
}
)*
}
}
}
impl<const N: usize> StdFloat for Simd<f64, N>
where
LaneCount<N>: SupportedLaneCount,
{
/// Returns the floating point's fractional value, with its integer part removed.
#[must_use = "method returns a new vector and does not mutate the original value"]
#[inline]
fn fract(self) -> Self {
self - self.trunc()
}
}
#[cfg(test)]
mod tests {
use super::*;
use simd::prelude::*;
#[test]
fn everything_works() {
let x = f32x4::from_array([0.1, 0.5, 0.6, -1.5]);
let x2 = x + x;
let _xc = x.ceil();
let _xf = x.floor();
let _xr = x.round();
let _xt = x.trunc();
let _xfma = x.mul_add(x, x);
let _xsqrt = x.sqrt();
let _ = x2.abs() * x2;
}
impl_float! {
sin: simd_fsin,
cos: simd_fcos,
exp: simd_fexp,
exp2: simd_fexp2,
ln: simd_flog,
log2: simd_flog2,
log10: simd_flog10,
}