cubecl-matmul 0.8.1

CubeCL Matrix Multiplication Kernels Engine
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
105
106
107
108
109
110
111
112
113
114
115
116

use crate::MatmulInputHandleRef;
use crate::components::batch::BatchConfig;
use crate::components::{
    InputArg, InputRuntimeArg, MatmulElems, MatmulLineSizes, MatmulSetupError, OutputRuntimeArg,
};
use crate::components::{
    MatmulProblem, MatmulSpec, OutputArg,
    global::args::{ConcreteInputsFactory, ConcreteOutputFactory},
};
use crate::kernels::layered::base::Selection;
use crate::kernels::layered::{Algorithm, launch_with_config};
use cubecl_core::prelude::TensorHandleRef;
use cubecl_core::{Runtime, client::ComputeClient};

/// Select which kernel to launch for the given Algorithm.
///
/// Only works for concrete tensor inputs and output.
#[allow(clippy::result_large_err, clippy::too_many_arguments)]
pub fn launch_kernel_concrete<MS: MatmulSpec, R: Runtime, A: Algorithm>(
    client: &ComputeClient<R::Server>,
    lhs: &MatmulInputHandleRef<'_, R>,
    rhs: &MatmulInputHandleRef<'_, R>,
    out: &TensorHandleRef<'_, R>,
    problem: MatmulProblem,
    line_sizes: MatmulLineSizes,
    plane_dim: u32,
    selection: &Selection<A::SelectionArgs>,
) -> Result<(), MatmulSetupError>
where
    InputArg<MS>: ConcreteInputsFactory,
    OutputArg<MS>: ConcreteOutputFactory,
{
    let elems = MatmulElems::new::<MS::Precision>();

    let mut view_line_sizes = line_sizes;

    if let MatmulInputHandleRef::Quantized { scheme, .. } = lhs {
        view_line_sizes.lhs *= scheme.num_quants() as u8;
    }
    if let MatmulInputHandleRef::Quantized { scheme, .. } = rhs {
        view_line_sizes.rhs *= scheme.num_quants() as u8;
    }

    let selection = match selection {
        Selection::Forced(selection) => selection.clone(),
        Selection::Inferred(args) => {
            A::selection::<R>(client, &problem, plane_dim, &view_line_sizes, elems, args)?
        }
    };
    let config = A::setup::<MS::Precision, R>(client, &problem, &selection, &view_line_sizes)?;
    let cube_count_plan = config.hypercube_config().cube_count_plan(
        &problem,
        client.properties().hardware.max_cube_count.clone(),
    );

    launch_with_config::<MS, R, A>(
        client,
        config.cube_dim(),
        cube_count_plan.resolve(),
        <InputArg<MS> as ConcreteInputsFactory>::create(
            client,
            lhs,
            rhs,
            &selection,
            &problem,
            &line_sizes,
            config,
        ),
        <OutputArg<MS> as ConcreteOutputFactory>::create(
            client,
            out,
            &selection,
            &problem,
            &line_sizes,
            config,
        ),
        cube_count_plan.as_args(),
        config,
    )
}

/// Select which kernel to launch for the given Algorithm.
pub fn launch_kernel_virtual<'a, MS: MatmulSpec, R: Runtime, A: Algorithm>(
    client: &ComputeClient<R::Server>,
    input: InputRuntimeArg<'a, MS, R>,
    output: OutputRuntimeArg<'a, MS, R>,
    problem: MatmulProblem,
    view_line_sizes: MatmulLineSizes,
    plane_dim: u32,
    selection: &Selection<A::SelectionArgs>,
) -> Result<(), MatmulSetupError> {
    let elems = MatmulElems::new::<MS::Precision>();

    let selection = match selection {
        Selection::Forced(selection) => selection.clone(),
        Selection::Inferred(args) => {
            A::selection::<R>(client, &problem, plane_dim, &view_line_sizes, elems, args)?
        }
    };
    let config = A::setup::<MS::Precision, R>(client, &problem, &selection, &view_line_sizes)?;

    let cube_count_plan = config.hypercube_config().cube_count_plan(
        &problem,
        client.properties().hardware.max_cube_count.clone(),
    );

    launch_with_config::<MS, R, A>(
        client,
        config.cube_dim(),
        cube_count_plan.resolve(),
        input,
        output,
        cube_count_plan.as_args(),
        config,
    )
}