cubecl_std/tensor/
identity.rs

1use cubecl::frontend::TensorHandleRef;
2use cubecl::prelude::*;
3use cubecl::tensor_line_size_parallel;
4use cubecl_core as cubecl;
5
6use super::TensorHandle;
7
8#[cube(launch_unchecked)]
9fn identity_kernel<C: Numeric>(output: &mut Tensor<Line<C>>, gap: u32) {
10    let pos_x = ABSOLUTE_POS_X * output.line_size();
11    if ABSOLUTE_POS_Y < output.shape(0) && pos_x < output.shape(1) {
12        let mut line = Line::empty(output.line_size()).fill(C::from_int(0));
13        let offs_y = ABSOLUTE_POS_Y * output.stride(0);
14
15        let start_pos = offs_y + pos_x;
16        let mut offset = 0;
17        while offset < output.line_size() {
18            let remainder = (start_pos + offset) % gap;
19            if remainder % gap == 0 {
20                line[offset] = C::from_int(1);
21                offset += gap;
22            } else {
23                offset += gap - remainder;
24            }
25        }
26        output[start_pos / output.line_size()] = line;
27    }
28}
29
30/// Launch identity matrix kernel.
31/// Ensure output is a [`TensorHandle`] containing a square matrix.
32/// output will contain the identity matrix.
33pub fn launch<R: Runtime, C: Numeric>(
34    client: &ComputeClient<R::Server, R::Channel>,
35    output: &TensorHandle<R, C>,
36) {
37    launch_ref::<R, C>(client, &output.as_ref());
38}
39
40/// Launch identity matrix kernel by ref.
41/// Ensure output is a [`TensorHandleRef`] containing a square matrix.
42/// output will contain the identity matrix.
43pub fn launch_ref<R: Runtime, C: Numeric>(
44    client: &ComputeClient<R::Server, R::Channel>,
45    output: &TensorHandleRef<R>,
46) {
47    assert_eq!(2, output.shape.len(), "input should be a matrix");
48    assert_eq!(
49        output.shape[0], output.shape[1],
50        "input should be a square matrix"
51    );
52
53    let vectorization_factor = tensor_line_size_parallel(
54        R::supported_line_sizes().iter().cloned(),
55        output.shape,
56        output.strides,
57        1,
58    );
59
60    let cube_dim = CubeDim::default();
61    let lines_x = output.shape[1] as u32 / vectorization_factor as u32;
62    let cube_count_x = lines_x.div_ceil(cube_dim.x);
63    let cube_count_y = (output.shape[0] as u32).div_ceil(cube_dim.y);
64    let cube_count = CubeCount::new_2d(cube_count_x, cube_count_y);
65
66    unsafe {
67        identity_kernel::launch_unchecked::<C, R>(
68            client,
69            cube_count,
70            cube_dim,
71            TensorArg::from_raw_parts::<C>(
72                output.handle,
73                output.strides,
74                output.shape,
75                vectorization_factor,
76            ),
77            ScalarArg::new(output.strides[0] as u32 + 1),
78        );
79    }
80}
cubecl_std/tensor/identity.rs

cubecl_std/tensor/
identity.rs
