pub trait CpuFloat:
Copy
+ Default
+ std::ops::Add<Output = Self>
+ std::ops::Sub<Output = Self>
+ std::ops::Mul<Output = Self>
{
#[must_use]
fn mul_add(self, a: Self, b: Self) -> Self;
}
impl CpuFloat for f32 {
#[inline]
fn mul_add(self, a: Self, b: Self) -> Self {
Self::mul_add(self, a, b)
}
}
impl CpuFloat for f64 {
#[inline]
fn mul_add(self, a: Self, b: Self) -> Self {
Self::mul_add(self, a, b)
}
}
#[inline]
pub fn elementwise_add<T: CpuFloat>(a: &[T], b: &[T], output: &mut [T]) {
assert_eq!(a.len(), b.len());
assert_eq!(a.len(), output.len());
for ((o, &ai), &bi) in output.iter_mut().zip(a).zip(b) {
*o = ai + bi;
}
}
#[inline]
pub fn elementwise_mul<T: CpuFloat>(a: &[T], b: &[T], output: &mut [T]) {
assert_eq!(a.len(), b.len());
assert_eq!(a.len(), output.len());
for ((o, &ai), &bi) in output.iter_mut().zip(a).zip(b) {
*o = ai * bi;
}
}
#[inline]
pub fn elementwise_fma<T: CpuFloat>(a: &[T], b: &[T], c: &[T], output: &mut [T]) {
assert_eq!(a.len(), b.len());
assert_eq!(a.len(), c.len());
assert_eq!(a.len(), output.len());
for (((o, &ai), &bi), &ci) in output.iter_mut().zip(a).zip(b).zip(c) {
*o = ai.mul_add(bi, ci);
}
}
#[inline]
pub fn dot_product<T: CpuFloat>(a: &[T], b: &[T]) -> T {
assert_eq!(a.len(), b.len());
a.iter()
.zip(b)
.fold(T::default(), |acc, (&ai, &bi)| acc + ai * bi)
}
#[inline]
pub fn kahan_sum<T: CpuFloat>(input: &[T]) -> T {
let mut sum = T::default();
let mut c = T::default();
for &x in input {
let y = x - c;
let t = sum + y;
c = (t - sum) - y;
sum = t;
}
sum
}
#[inline]
pub fn reduce_sum<T: CpuFloat>(input: &[T]) -> T {
input.iter().fold(T::default(), |acc, &x| acc + x)
}