use coeus_core::{ComputeBackend, CpuAddressableStorage, CpuAddressableStorageMut, Scalar};
use coeus_tensor::Tensor;
#[inline]
pub fn pad<T: Scalar, B: ComputeBackend + Default>(
x: &Tensor<T, B>,
pads: &[(usize, usize)],
value: T,
) -> Result<Tensor<T, B>, B::Error>
where
B::DeviceBuffer<T>: CpuAddressableStorage<T> + CpuAddressableStorageMut<T>,
{
let ndim = x.ndim();
assert_eq!(pads.len(), ndim, "pad: pads.len() must equal tensor ndim");
if pads.iter().all(|&(b, a)| b == 0 && a == 0) {
return Ok(x.clone());
}
let backend = B::default();
let mut out_shape = x.shape_cloned();
for d in 0..ndim {
out_shape[d] += pads[d].0 + pads[d].1;
}
let values = coeus_cpu::pad_values(x.layout(), x.storage().as_slice(), pads, value)
.expect("coeus-cpu pad failed");
Tensor::from_slice_on(out_shape, &values, &backend)
}