我正在尝试实现一个矢量数学库,该库可以根据处理器支持的功能集生成函数的 SIMD 优化版本。目前我为每个功能集都有一个结构和一个定义回退结构的特征。我在特征集结构上实现每个操作。我一直无法实现回退机制。我现在所拥有的会产生流动错误。
error[E0119]: conflicting implementations of trait `Add<Vec4, Vec4>` for type `Scalar`:
--> src/main.rs:41:1
|
14 | default impl<F: FeatureSet, Lhs, Rhs> Add<Lhs, Rhs> for F where <F as FeatureSet>::Fallback: Add<Lhs, Rhs> {
| ---------------------------------------------------------------------------------------------------------- first implementation here
...
41 | impl Add<Vec4, Vec4> for Scalar {
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ conflicting implementation for `Scalar`
有人对如何解决这个问题有任何建议吗?
示例代码:
#![feature(specialization)]
use std::arch::x86_64 as arch;
trait FeatureSet {
type Fallback;
}
trait Add<Lhs, Rhs> {
type Output;
unsafe fn add(lhs: Lhs, rhs: Rhs) -> Self::Output;
}
// Attempt at implementing fallback to implementation in Fallback type
default impl<F: FeatureSet, Lhs, Rhs> Add<Lhs, Rhs> for F where <F as FeatureSet>::Fallback: Add<Lhs, Rhs> {
type Output = <<F as FeatureSet>::Fallback as Add<Lhs, Rhs>>::Output;
unsafe fn add(lhs: Lhs, rhs: Rhs) -> Self::Output {
<F as FeatureSet>::Fallback::add(lhs, rhs)
}
}
#[repr(C)]
union Vec4 {
scalar: std::mem::ManuallyDrop<[f32; 4]>,
#[allow(unused)]
simd: std::mem::ManuallyDrop<arch::__m128>,
}
impl std::convert::From<[f32; 4]> for Vec4 {
fn from(values: [f32; 4]) -> Self {
Self {
scalar: std::mem::ManuallyDrop::new(values)
}
}
}
struct Scalar;
impl FeatureSet for Scalar {
type Fallback = ();
}
impl Add<Vec4, Vec4> for Scalar {
type Output = Vec4;
unsafe fn add(lhs: Vec4, rhs: Vec4) -> Self::Output {
let mut r: [f32; 4] = std::mem::MaybeUninit::uninit().assume_init();
for i in 0 .. 4 {
r[i] = lhs.scalar[i] + rhs.scalar[i];
}
r.into()
}
}
struct Sse;
impl FeatureSet for Sse {
type Fallback = Scalar;
}
// Should fall back to implementation in FeatureSet::Fallback type when not implemented
impl Add<Vec4, Vec4> for Sse {
type Output = Vec4;
#[inline]
#[target_feature(enable = "sse")]
unsafe fn add(lhs: Vec4, rhs: Vec4) -> Self::Output {
let r = arch::_mm_add_ps(*lhs.simd, *rhs.simd);
Vec4 {
simd: std::mem::ManuallyDrop::new(r)
}
}
}
#[inline(never)]
// This function should be able to be specialized for each feature set for later
// runtime selection
unsafe fn perform_add<F: FeatureSet>(a: Vec4, b: Vec4) -> Vec4 {
F::add(a, b)
}
fn main() {
let a = [0.0, 1.0, 2.0, 3.0].into();
let b = [4.0, 4.0, 5.0, 6.0].into();
unsafe {
let c = perform_add::<Sse>(a, b);
println!("{} {} {} {}", c.scalar[0], c.scalar[1], c.scalar[2], c.scalar[3]);
}
}