burn-cubecl 0.22.0-pre.2

Generic backend that can be compiled just-in-time to any shader language target
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104

use burn_backend::DType;
use burn_backend::cubecl::dtype_to_storage_type;
use cubecl::{
    calculate_cube_count_elemwise,
    prelude::*,
    std::tensor::layout::linear::{LinearView, LinearViewMut},
};

use crate::{
    CubeRuntime,
    kernel::utils::{address_type, broadcast_shape},
    ops::{max_vector_size_many, numeric::empty_device_dtype},
    tensor::CubeTensor,
};

#[cube(launch, address_type = "dynamic")]
fn mask_where_kernel<T: Numeric, B: Int, N: Size>(
    input: LinearView<'_, Vector<T, N>>,
    value: LinearView<'_, Vector<T, N>>,
    mask: LinearView<'_, Vector<B, N>>,
    mut output: LinearViewMut<'_, Vector<T, N>>,
    #[define(T, B)] _dtypes: [StorageType; 2],
) {
    let pos = ABSOLUTE_POS;
    if !output.is_in_bounds(pos) {
        terminate!();
    }

    output.write(
        pos,
        select_many(
            Vector::cast_from(mask.read(pos)),
            value.read(pos),
            input.read(pos),
        ),
    );
}

#[derive(Clone, Copy, Debug)]
/// Define how to run the mask where kernel.
///
/// # Notes
///
/// All assertions should be done before choosing the strategy.
pub enum MaskWhereStrategy {
    /// Don't mutate any input.
    Readonly,
    /// Reuse the lhs tensor inplace.
    InplaceLhs,
    /// Reuse the rhs tensor inplace.
    InplaceRhs,
}

/// Execute the mask where kernel with the given strategy.
pub fn mask_where<R: CubeRuntime>(
    input: CubeTensor<R>,
    mask: CubeTensor<R>,
    value: CubeTensor<R>,
    strategy: MaskWhereStrategy,
    dtype_bool: DType,
) -> CubeTensor<R> {
    let vector_size = max_vector_size_many(&[&input, &mask, &value], input.meta.num_dims() - 1);

    let working_units = input.meta.num_elements() / vector_size as usize;
    let cube_dim = CubeDim::new(&input.client, working_units);
    let cube_count = calculate_cube_count_elemwise(&input.client, working_units, cube_dim);

    let out_shape = broadcast_shape(&[&input, &mask, &value]);

    let output = match strategy {
        MaskWhereStrategy::Readonly => empty_device_dtype(
            input.client.clone(),
            input.device.clone(),
            out_shape,
            input.dtype,
        ),
        MaskWhereStrategy::InplaceLhs => input.clone(),
        MaskWhereStrategy::InplaceRhs => value.clone(),
    };

    let out = match strategy {
        MaskWhereStrategy::Readonly => output.clone().into_linear_view(),
        MaskWhereStrategy::InplaceLhs => output.as_linear_view_alias(0),
        MaskWhereStrategy::InplaceRhs => output.as_linear_view_alias(1),
    };

    mask_where_kernel::launch(
        &output.client,
        cube_count,
        cube_dim,
        address_type!(input, value, mask, output),
        vector_size,
        input.into_linear_view_like(&output),
        value.into_linear_view_like(&output),
        mask.into_linear_view_like(&output),
        out,
        [
            dtype_to_storage_type(output.dtype),
            dtype_to_storage_type(dtype_bool),
        ],
    );

    output
}