coeus_ops/shape/concat_split_stack/
split.rs1use coeus_core::{ComputeBackend, CpuAddressableStorage, CpuAddressableStorageMut, Scalar};
5use coeus_tensor::Tensor;
6
7#[inline]
15pub fn split<T: Scalar, B: ComputeBackend + Default>(
16 x: &Tensor<T, B>,
17 chunk_size: usize,
18 dim: usize,
19) -> Result<Vec<Tensor<T, B>>, B::Error>
20where
21 B::DeviceBuffer<T>: CpuAddressableStorage<T> + CpuAddressableStorageMut<T>,
22{
23 assert!(chunk_size > 0, "split: chunk_size must be > 0");
24 let ndim = x.ndim();
25 assert!(
26 dim < ndim,
27 "split: dim {dim} out of range for {ndim}D tensor"
28 );
29
30 let backend = B::default();
31 let dim_size = x.shape()[dim];
32 let sizes: Vec<_> = (0..dim_size)
33 .step_by(chunk_size)
34 .map(|start| (dim_size - start).min(chunk_size))
35 .collect();
36 let (flat, offsets) = coeus_cpu::split_values(x.layout(), x.storage().as_slice(), dim, &sizes)
37 .expect("coeus-cpu split failed");
38
39 offsets
40 .windows(2)
41 .zip(sizes)
42 .map(|(bounds, chunk_dim)| {
43 let mut out_shape = x.shape_cloned();
44 out_shape[dim] = chunk_dim;
45 Tensor::from_slice_on(out_shape, &flat[bounds[0]..bounds[1]], &backend)
46 })
47 .collect()
48}