burn_cubecl/kernel/cast/
bool_cast.rs

1use crate::{
2    CubeRuntime,
3    kernel::utils::address_type,
4    ops::{max_vector_size, numeric::empty_device_dtype},
5    tensor::CubeTensor,
6};
7use burn_backend::TensorMetadata;
8use burn_std::DType;
9use cubecl::{
10    CubeDim, calculate_cube_count_elemwise, num_traits::One, prelude::*,
11    std::tensor::layout::linear::LinearView,
12};
13
14#[cube(launch_unchecked, address_type = "dynamic")]
15fn bool_cast_kernel<B: Int, T: Numeric, N: Size>(
16    input: &LinearView<Vector<B, N>>,
17    output: &mut LinearView<Vector<T, N>, ReadWrite>,
18    #[define(B, T)] _dtypes: [StorageType; 2],
19) {
20    if !output.is_in_bounds(ABSOLUTE_POS) {
21        terminate!();
22    }
23
24    output[ABSOLUTE_POS] = Vector::cast_from(input[ABSOLUTE_POS] & Vector::one());
25}
26
27/// Cast a bool tensor to the given element type.
28///
29/// This alternative to cast is necessary because bool are represented as u32 or u8
30/// where any non-zero value means true. Depending how it was created
31/// it may hold an uncanny bit combination. Naively casting it would not
32/// necessarily yield 0 or 1.
33pub fn bool_cast<R: CubeRuntime>(tensor: CubeTensor<R>, out_dtype: DType) -> CubeTensor<R> {
34    let output = empty_device_dtype(
35        tensor.client.clone(),
36        tensor.device.clone(),
37        tensor.shape(),
38        out_dtype,
39    );
40
41    let vector_size = max_vector_size(&tensor);
42    let num_elems = tensor.meta.num_elements();
43    let working_units = num_elems / vector_size as usize;
44    let cube_dim = CubeDim::new(&tensor.client, working_units);
45    let cube_count = calculate_cube_count_elemwise(&tensor.client, working_units, cube_dim);
46
47    let dtype = tensor.dtype;
48
49    unsafe {
50        bool_cast_kernel::launch_unchecked(
51            &output.client,
52            cube_count,
53            cube_dim,
54            address_type!(tensor, output),
55            vector_size,
56            tensor.into_linear_view(),
57            output.clone().into_linear_view(),
58            [dtype.into(), out_dtype.into()],
59        )
60    };
61
62    output
63}
burn_cubecl/kernel/cast/bool_cast.rs

burn_cubecl/kernel/cast/
bool_cast.rs
