Merge commit 'cff979eec1ac0473fc4960ee6cde462c6aeda824' 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

@ -141,6 +141,11 @@ jobs:
- name: Test (release)
run: cargo test --verbose --target=${{ matrix.target }} --release
- name: Generate docs
run: cargo doc --verbose --target=${{ matrix.target }}
env:
RUSTDOCFLAGS: -Dwarnings
wasm-tests:
name: "wasm (firefox, ${{ matrix.name }})"
runs-on: ubuntu-latest

View File

@ -177,6 +177,9 @@ name = "std_float"
version = "0.1.0"
dependencies = [
"core_simd",
"test_helpers",
"wasm-bindgen",
"wasm-bindgen-test",
]
[[package]]

View File

@ -13,11 +13,12 @@
simd_ffi,
staged_api,
strict_provenance,
prelude_import,
ptr_metadata
)]
#![cfg_attr(
all(
any(target_arch = "aarch64", target_arch = "arm",),
any(target_arch = "aarch64", target_arch = "arm64ec", target_arch = "arm",),
any(
all(target_feature = "v6", not(target_feature = "mclass")),
all(target_feature = "mclass", target_feature = "dsp"),
@ -33,12 +34,21 @@
any(target_arch = "powerpc", target_arch = "powerpc64"),
feature(stdarch_powerpc)
)]
#![cfg_attr(
all(target_arch = "x86_64", target_feature = "avx512f"),
feature(stdarch_x86_avx512)
)]
#![warn(missing_docs, clippy::missing_inline_in_public_items)] // basically all items, really
#![deny(unsafe_op_in_unsafe_fn, clippy::undocumented_unsafe_blocks)]
#![doc(test(attr(deny(warnings))))]
#![allow(internal_features)]
#![unstable(feature = "portable_simd", issue = "86656")]
//! Portable SIMD module.
#[prelude_import]
#[allow(unused_imports)]
use core::prelude::v1::*;
#[path = "mod.rs"]
mod core_simd;
pub use self::core_simd::simd;

View File

@ -34,6 +34,7 @@ mod sealed {
fn eq(self, other: Self) -> bool;
fn to_usize(self) -> usize;
fn max_unsigned() -> u64;
type Unsigned: SimdElement;
@ -78,6 +79,11 @@ macro_rules! impl_element {
self as usize
}
#[inline]
fn max_unsigned() -> u64 {
<$unsigned>::MAX as u64
}
type Unsigned = $unsigned;
const TRUE: Self = -1;

View File

@ -16,7 +16,10 @@ where
#[inline]
pub fn swizzle_dyn(self, idxs: Simd<u8, N>) -> Self {
#![allow(unused_imports, unused_unsafe)]
#[cfg(all(target_arch = "aarch64", target_endian = "little"))]
#[cfg(all(
any(target_arch = "aarch64", target_arch = "arm64ec"),
target_endian = "little"
))]
use core::arch::aarch64::{uint8x8_t, vqtbl1q_u8, vtbl1_u8};
#[cfg(all(
target_arch = "arm",
@ -37,6 +40,7 @@ where
#[cfg(all(
any(
target_arch = "aarch64",
target_arch = "arm64ec",
all(target_arch = "arm", target_feature = "v7")
),
target_feature = "neon",
@ -48,7 +52,7 @@ where
#[cfg(target_feature = "simd128")]
16 => transize(wasm::i8x16_swizzle, self, idxs),
#[cfg(all(
target_arch = "aarch64",
any(target_arch = "aarch64", target_arch = "arm64ec"),
target_feature = "neon",
target_endian = "little"
))]

View File

@ -1,5 +1,6 @@
use crate::simd::{
cmp::SimdPartialOrd,
num::SimdUint,
ptr::{SimdConstPtr, SimdMutPtr},
LaneCount, Mask, MaskElement, SupportedLaneCount, Swizzle,
};
@ -262,6 +263,7 @@ where
/// # Panics
///
/// Panics if the slice's length is less than the vector's `Simd::N`.
/// Use `load_or_default` for an alternative that does not panic.
///
/// # Example
///
@ -315,6 +317,143 @@ where
unsafe { self.store(slice.as_mut_ptr().cast()) }
}
/// Reads contiguous elements from `slice`. Elements are read so long as they're in-bounds for
/// the `slice`. Otherwise, the default value for the element type is returned.
///
/// # Examples
/// ```
/// # #![feature(portable_simd)]
/// # #[cfg(feature = "as_crate")] use core_simd::simd;
/// # #[cfg(not(feature = "as_crate"))] use core::simd;
/// # use simd::Simd;
/// let vec: Vec<i32> = vec![10, 11];
///
/// let result = Simd::<i32, 4>::load_or_default(&vec);
/// assert_eq!(result, Simd::from_array([10, 11, 0, 0]));
/// ```
#[must_use]
#[inline]
pub fn load_or_default(slice: &[T]) -> Self
where
T: Default,
{
Self::load_or(slice, Default::default())
}
/// Reads contiguous elements from `slice`. Elements are read so long as they're in-bounds for
/// the `slice`. Otherwise, the corresponding value from `or` is passed through.
///
/// # Examples
/// ```
/// # #![feature(portable_simd)]
/// # #[cfg(feature = "as_crate")] use core_simd::simd;
/// # #[cfg(not(feature = "as_crate"))] use core::simd;
/// # use simd::Simd;
/// let vec: Vec<i32> = vec![10, 11];
/// let or = Simd::from_array([-5, -4, -3, -2]);
///
/// let result = Simd::load_or(&vec, or);
/// assert_eq!(result, Simd::from_array([10, 11, -3, -2]));
/// ```
#[must_use]
#[inline]
pub fn load_or(slice: &[T], or: Self) -> Self {
Self::load_select(slice, Mask::splat(true), or)
}
/// Reads contiguous elements from `slice`. Each element is read from memory if its
/// corresponding element in `enable` is `true`.
///
/// When the element is disabled or out of bounds for the slice, that memory location
/// is not accessed and the corresponding value from `or` is passed through.
///
/// # Examples
/// ```
/// # #![feature(portable_simd)]
/// # #[cfg(feature = "as_crate")] use core_simd::simd;
/// # #[cfg(not(feature = "as_crate"))] use core::simd;
/// # use simd::{Simd, Mask};
/// let vec: Vec<i32> = vec![10, 11, 12, 13, 14, 15, 16, 17, 18];
/// let enable = Mask::from_array([true, true, false, true]);
/// let or = Simd::from_array([-5, -4, -3, -2]);
///
/// let result = Simd::load_select(&vec, enable, or);
/// assert_eq!(result, Simd::from_array([10, 11, -3, 13]));
/// ```
#[must_use]
#[inline]
pub fn load_select_or_default(slice: &[T], enable: Mask<<T as SimdElement>::Mask, N>) -> Self
where
T: Default,
{
Self::load_select(slice, enable, Default::default())
}
/// Reads contiguous elements from `slice`. Each element is read from memory if its
/// corresponding element in `enable` is `true`.
///
/// When the element is disabled or out of bounds for the slice, that memory location
/// is not accessed and the corresponding value from `or` is passed through.
///
/// # Examples
/// ```
/// # #![feature(portable_simd)]
/// # #[cfg(feature = "as_crate")] use core_simd::simd;
/// # #[cfg(not(feature = "as_crate"))] use core::simd;
/// # use simd::{Simd, Mask};
/// let vec: Vec<i32> = vec![10, 11, 12, 13, 14, 15, 16, 17, 18];
/// let enable = Mask::from_array([true, true, false, true]);
/// let or = Simd::from_array([-5, -4, -3, -2]);
///
/// let result = Simd::load_select(&vec, enable, or);
/// assert_eq!(result, Simd::from_array([10, 11, -3, 13]));
/// ```
#[must_use]
#[inline]
pub fn load_select(
slice: &[T],
mut enable: Mask<<T as SimdElement>::Mask, N>,
or: Self,
) -> Self {
enable &= mask_up_to(slice.len());
// SAFETY: We performed the bounds check by updating the mask. &[T] is properly aligned to
// the element.
unsafe { Self::load_select_ptr(slice.as_ptr(), enable, or) }
}
/// Reads contiguous elements from `slice`. Each element is read from memory if its
/// corresponding element in `enable` is `true`.
///
/// When the element is disabled, that memory location is not accessed and the corresponding
/// value from `or` is passed through.
#[must_use]
#[inline]
pub unsafe fn load_select_unchecked(
slice: &[T],
enable: Mask<<T as SimdElement>::Mask, N>,
or: Self,
) -> Self {
let ptr = slice.as_ptr();
// SAFETY: The safety of reading elements from `slice` is ensured by the caller.
unsafe { Self::load_select_ptr(ptr, enable, or) }
}
/// Reads contiguous elements starting at `ptr`. Each element is read from memory if its
/// corresponding element in `enable` is `true`.
///
/// When the element is disabled, that memory location is not accessed and the corresponding
/// value from `or` is passed through.
#[must_use]
#[inline]
pub unsafe fn load_select_ptr(
ptr: *const T,
enable: Mask<<T as SimdElement>::Mask, N>,
or: Self,
) -> Self {
// SAFETY: The safety of reading elements through `ptr` is ensured by the caller.
unsafe { core::intrinsics::simd::simd_masked_load(enable.to_int(), ptr, or) }
}
/// Reads from potentially discontiguous indices in `slice` to construct a SIMD vector.
/// If an index is out-of-bounds, the element is instead selected from the `or` vector.
///
@ -493,6 +632,77 @@ where
unsafe { core::intrinsics::simd::simd_gather(or, source, enable.to_int()) }
}
/// Conditionally write contiguous elements to `slice`. The `enable` mask controls
/// which elements are written, as long as they're in-bounds of the `slice`.
/// If the element is disabled or out of bounds, no memory access to that location
/// is made.
///
/// # Examples
/// ```
/// # #![feature(portable_simd)]
/// # #[cfg(feature = "as_crate")] use core_simd::simd;
/// # #[cfg(not(feature = "as_crate"))] use core::simd;
/// # use simd::{Simd, Mask};
/// let mut arr = [0i32; 4];
/// let write = Simd::from_array([-5, -4, -3, -2]);
/// let enable = Mask::from_array([false, true, true, true]);
///
/// write.store_select(&mut arr[..3], enable);
/// assert_eq!(arr, [0, -4, -3, 0]);
/// ```
#[inline]
pub fn store_select(self, slice: &mut [T], mut enable: Mask<<T as SimdElement>::Mask, N>) {
enable &= mask_up_to(slice.len());
// SAFETY: We performed the bounds check by updating the mask. &[T] is properly aligned to
// the element.
unsafe { self.store_select_ptr(slice.as_mut_ptr(), enable) }
}
/// Conditionally write contiguous elements to `slice`. The `enable` mask controls
/// which elements are written.
///
/// # Safety
///
/// Every enabled element must be in bounds for the `slice`.
///
/// # Examples
/// ```
/// # #![feature(portable_simd)]
/// # #[cfg(feature = "as_crate")] use core_simd::simd;
/// # #[cfg(not(feature = "as_crate"))] use core::simd;
/// # use simd::{Simd, Mask};
/// let mut arr = [0i32; 4];
/// let write = Simd::from_array([-5, -4, -3, -2]);
/// let enable = Mask::from_array([false, true, true, true]);
///
/// unsafe { write.store_select_unchecked(&mut arr, enable) };
/// assert_eq!(arr, [0, -4, -3, -2]);
/// ```
#[inline]
pub unsafe fn store_select_unchecked(
self,
slice: &mut [T],
enable: Mask<<T as SimdElement>::Mask, N>,
) {
let ptr = slice.as_mut_ptr();
// SAFETY: The safety of writing elements in `slice` is ensured by the caller.
unsafe { self.store_select_ptr(ptr, enable) }
}
/// Conditionally write contiguous elements starting from `ptr`.
/// The `enable` mask controls which elements are written.
/// When disabled, the memory location corresponding to that element is not accessed.
///
/// # Safety
///
/// Memory addresses for element are calculated [`pointer::wrapping_offset`] and
/// each enabled element must satisfy the same conditions as [`core::ptr::write`].
#[inline]
pub unsafe fn store_select_ptr(self, ptr: *mut T, enable: Mask<<T as SimdElement>::Mask, N>) {
// SAFETY: The safety of writing elements through `ptr` is ensured by the caller.
unsafe { core::intrinsics::simd::simd_masked_store(enable.to_int(), ptr, self) }
}
/// Writes the values in a SIMD vector to potentially discontiguous indices in `slice`.
/// If an index is out-of-bounds, the write is suppressed without panicking.
/// If two elements in the scattered vector would write to the same index
@ -980,3 +1190,37 @@ where
{
type Mask = isize;
}
#[inline]
fn lane_indices<const N: usize>() -> Simd<usize, N>
where
LaneCount<N>: SupportedLaneCount,
{
let mut index = [0; N];
for i in 0..N {
index[i] = i;
}
Simd::from_array(index)
}
#[inline]
fn mask_up_to<M, const N: usize>(len: usize) -> Mask<M, N>
where
LaneCount<N>: SupportedLaneCount,
M: MaskElement,
{
let index = lane_indices::<N>();
let max_value: u64 = M::max_unsigned();
macro_rules! case {
($ty:ty) => {
if N < <$ty>::MAX as usize && max_value as $ty as u64 == max_value {
return index.cast().simd_lt(Simd::splat(len.min(N) as $ty)).cast();
}
};
}
case!(u8);
case!(u16);
case!(u32);
case!(u64);
index.simd_lt(Simd::splat(len)).cast()
}

View File

@ -24,7 +24,7 @@ mod x86;
#[cfg(target_arch = "wasm32")]
mod wasm32;
#[cfg(any(target_arch = "aarch64", target_arch = "arm",))]
#[cfg(any(target_arch = "aarch64", target_arch = "arm64ec", target_arch = "arm",))]
mod arm;
#[cfg(any(target_arch = "powerpc", target_arch = "powerpc64"))]

View File

@ -4,12 +4,13 @@ use crate::simd::*;
#[cfg(target_arch = "arm")]
use core::arch::arm::*;
#[cfg(target_arch = "aarch64")]
#[cfg(any(target_arch = "aarch64", target_arch = "arm64ec"))]
use core::arch::aarch64::*;
#[cfg(all(
any(
target_arch = "aarch64",
target_arch = "arm64ec",
all(target_arch = "arm", target_feature = "v7"),
),
target_endian = "little"
@ -69,7 +70,10 @@ mod simd32 {
from_transmute! { unsafe Simd<i8, 4> => int8x4_t }
}
#[cfg(target_arch = "aarch64")]
#[cfg(all(
any(target_arch = "aarch64", target_arch = "arm64ec"),
target_endian = "little"
))]
mod aarch64 {
use super::neon::*;
use super::*;

View File

@ -0,0 +1,35 @@
#![feature(portable_simd)]
use core_simd::simd::prelude::*;
#[cfg(target_arch = "wasm32")]
use wasm_bindgen_test::*;
#[cfg(target_arch = "wasm32")]
wasm_bindgen_test_configure!(run_in_browser);
#[test]
#[cfg_attr(target_arch = "wasm32", wasm_bindgen_test)]
fn masked_load_store() {
let mut arr = [u8::MAX; 7];
u8x4::splat(0).store_select(&mut arr[5..], Mask::from_array([false, true, false, true]));
// write to index 8 is OOB and dropped
assert_eq!(arr, [255u8, 255, 255, 255, 255, 255, 0]);
u8x4::from_array([0, 1, 2, 3]).store_select(&mut arr[1..], Mask::splat(true));
assert_eq!(arr, [255u8, 0, 1, 2, 3, 255, 0]);
// read from index 8 is OOB and dropped
assert_eq!(
u8x4::load_or(&arr[4..], u8x4::splat(42)),
u8x4::from_array([3, 255, 0, 42])
);
assert_eq!(
u8x4::load_select(
&arr[4..],
Mask::from_array([true, false, true, true]),
u8x4::splat(42)
),
u8x4::from_array([3, 42, 0, 42])
);
}

View File

@ -1,6 +1,6 @@
#![feature(portable_simd)]
use core::{fmt, ops::RangeInclusive};
use test_helpers::{self, biteq, make_runner, prop_assert_biteq};
use test_helpers::{biteq, make_runner, prop_assert_biteq};
fn swizzle_dyn_scalar_ver<const N: usize>(values: [u8; N], idxs: [u8; N]) -> [u8; N] {
let mut array = [0; N];

View File

@ -8,6 +8,13 @@ edition = "2021"
[dependencies]
core_simd = { path = "../core_simd", default-features = false }
[dev-dependencies.test_helpers]
path = "../test_helpers"
[target.'cfg(target_arch = "wasm32")'.dev-dependencies]
wasm-bindgen = "0.2"
wasm-bindgen-test = "0.3"
[features]
default = ["as_crate"]
as_crate = []

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,
}

View File

@ -0,0 +1,74 @@
#![feature(portable_simd)]
macro_rules! unary_test {
{ $scalar:tt, $($func:tt),+ } => {
test_helpers::test_lanes! {
$(
fn $func<const LANES: usize>() {
test_helpers::test_unary_elementwise(
&core_simd::simd::Simd::<$scalar, LANES>::$func,
&$scalar::$func,
&|_| true,
)
}
)*
}
}
}
macro_rules! binary_test {
{ $scalar:tt, $($func:tt),+ } => {
test_helpers::test_lanes! {
$(
fn $func<const LANES: usize>() {
test_helpers::test_binary_elementwise(
&core_simd::simd::Simd::<$scalar, LANES>::$func,
&$scalar::$func,
&|_, _| true,
)
}
)*
}
}
}
macro_rules! ternary_test {
{ $scalar:tt, $($func:tt),+ } => {
test_helpers::test_lanes! {
$(
fn $func<const LANES: usize>() {
test_helpers::test_ternary_elementwise(
&core_simd::simd::Simd::<$scalar, LANES>::$func,
&$scalar::$func,
&|_, _, _| true,
)
}
)*
}
}
}
macro_rules! impl_tests {
{ $scalar:tt } => {
mod $scalar {
use std_float::StdFloat;
unary_test! { $scalar, sqrt, sin, cos, exp, exp2, ln, log2, log10, ceil, floor, round, trunc }
binary_test! { $scalar, log }
ternary_test! { $scalar, mul_add }
test_helpers::test_lanes! {
fn fract<const LANES: usize>() {
test_helpers::test_unary_elementwise_flush_subnormals(
&core_simd::simd::Simd::<$scalar, LANES>::fract,
&$scalar::fract,
&|_| true,
)
}
}
}
}
}
impl_tests! { f32 }
impl_tests! { f64 }