use super::super::kernel::elementwise_unary;
use crate::backend_ops::{BackendOps, UnaryOp};
use coeus_core::{CpuAddressableStorage, CpuAddressableStorageMut, Float};
use coeus_tensor::Tensor;
#[inline]
pub fn glu<T: Float, B: BackendOps<T> + Default>(
input: &Tensor<T, B>,
dim: usize,
backend: &B,
) -> Result<Tensor<T, B>, B::Error>
where
B::DeviceBuffer<T>: CpuAddressableStorage<T> + CpuAddressableStorageMut<T>,
{
let ndim = input.ndim();
assert!(
dim < ndim,
"glu: dim {dim} out of bounds for {ndim}D tensor"
);
let dim_size = input.shape()[dim];
assert!(
dim_size.is_multiple_of(2),
"glu: dim {dim} size {dim_size} must be even"
);
let half = dim_size / 2;
let mut parts = crate::shape::split(input, half, dim);
assert_eq!(parts.len(), 2);
let b_part = parts
.pop()
.expect("invariant: split into 2 parts asserted above, second part present");
let a_part = parts
.pop()
.expect("invariant: split into 2 parts asserted above, first part present");
let gate = elementwise_unary(&b_part, backend, UnaryOp::Sigmoid)?;
Ok(crate::binary::mul(&a_part, &gate, backend))
}