cubek-matmul 0.2.0

CubeK: Matrix Multiplication Kernels
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

use crate::launch::{
    ConcreteInputsFactory, ConcreteOutputFactory, InputArg, InputRuntimeArg, MatmulArgs, OutputArg,
    OutputRuntimeArg,
};
use crate::{
    definition::MatmulProblem, definition::MatmulSetupError, definition::MatmulVectorSizes,
    definition::cube_mapping_launch,
};
use crate::{
    routines::LaunchInfo,
    routines::{BlueprintStrategy, Routine},
    {definition::MatmulElems, launch::ConfigRuntimeArg},
};
use cubecl::{
    prelude::TensorBinding,
    {Runtime, client::ComputeClient},
};
use cubek_std::InputBinding;

/// 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<MA: MatmulArgs<Config = ()>, R: Runtime, A: Routine<()>>(
    client: &ComputeClient<R>,
    lhs: InputBinding<R>,
    rhs: InputBinding<R>,
    out: TensorBinding<R>,
    problem: MatmulProblem,
    vector_sizes: MatmulVectorSizes,
    blueprint_strategy: &BlueprintStrategy<(), A>,
    dtypes: &mut MatmulElems,
) -> Result<(), MatmulSetupError>
where
    InputArg<MA>: ConcreteInputsFactory<A>,
    OutputArg<MA>: ConcreteOutputFactory<A>,
{
    let mut view_vector_sizes = vector_sizes;

    if let InputBinding::Quantized { scheme, .. } = lhs {
        view_vector_sizes.lhs *= scheme.num_quants();
    }
    if let InputBinding::Quantized { scheme, .. } = rhs {
        view_vector_sizes.rhs *= scheme.num_quants();
    }

    let device_settings = A::device_settings(client, view_vector_sizes);
    let expand_info = A::expand_blueprint(&problem, &device_settings, blueprint_strategy)?;
    let launch_info = A::prepare(&problem, &device_settings, expand_info)?;

    let input = <InputArg<MA> as ConcreteInputsFactory<A>>::create(
        lhs,
        rhs,
        &launch_info.blueprint,
        &problem,
        &launch_info.vector_sizes,
        dtypes,
    );
    let output = <OutputArg<MA> as ConcreteOutputFactory<A>>::create(
        out,
        &launch_info.blueprint,
        &problem,
        &launch_info.vector_sizes,
        dtypes,
    );

    launch_kernel::<MA, R, A>(client, input, output, (), launch_info)
}

/// Select which kernel to launch for the given Algorithm.
#[allow(clippy::too_many_arguments)]
pub fn launch_kernel_virtual<MA: MatmulArgs, R: Runtime, A: Routine<MA::Config>>(
    client: &ComputeClient<R>,
    input: InputRuntimeArg<MA, R>,
    output: OutputRuntimeArg<MA, R>,
    config: ConfigRuntimeArg<MA, R>,
    problem: MatmulProblem,
    view_vector_sizes: MatmulVectorSizes,
    blueprint_strategy: &BlueprintStrategy<MA::Config, A>,
) -> Result<(), MatmulSetupError> {
    let device_settings = A::device_settings(client, view_vector_sizes);
    let expand_info = A::expand_blueprint(&problem, &device_settings, blueprint_strategy)?;
    let launch_info = A::prepare(&problem, &device_settings, expand_info)?;

    launch_kernel::<MA, R, A>(client, input, output, config, launch_info)
}

/// Select which kernel to launch for the given Algorithm.
#[allow(clippy::too_many_arguments)]
pub fn launch_kernel<MA: MatmulArgs, R: Runtime, A: Routine<MA::Config>>(
    client: &ComputeClient<R>,
    input: InputRuntimeArg<MA, R>,
    output: OutputRuntimeArg<MA, R>,
    config: ConfigRuntimeArg<MA, R>,
    launch_info: LaunchInfo<A::Blueprint>,
) -> Result<(), MatmulSetupError> {
    A::launch::<MA, R>(
        client,
        launch_info.cube_dim,
        launch_info.cube_count_plan.resolve(),
        launch_info.address_type,
        input,
        output,
        config,
        cube_mapping_launch(&launch_info.cube_count_plan),
        launch_info.blueprint,
        &launch_info.dtypes,
        &launch_info.vector_sizes,
    )
}