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
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146

use std::{
    cmp::{max, min},
    fmt::Display,
};

use cubek_std::cube_count::{CubeCountPlan, CubeCountStrategy, GlobalOrder, HypercubeBlueprint};

use crate::{
    components::batch::{
        BatchMatmulFamily, CheckBounds,
        gemv_plane_parallel::{GemvKind, GemvPlaneParallelBlueprint, GemvPlaneParallelFamily},
    },
    definition::{MatmulElems, MatmulProblem, MatmulSetupError},
    routines::{
        BlueprintStrategy, DeviceSettings, ExpandInfo, LaunchInfo, Routine, num_concurrent_planes,
    },
};

pub struct GemvPlaneParallelRoutine {}

#[derive(Default, Clone)]
pub struct GemvPlaneParallelStrategy {
    pub target_num_planes: Option<usize>,
}

impl Display for GemvPlaneParallelStrategy {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        write!(f, "_{:?}", self.target_num_planes)
    }
}

impl Routine<()> for GemvPlaneParallelRoutine {
    type Strategy = GemvPlaneParallelStrategy;
    type BatchMatmul = GemvPlaneParallelFamily;
    type Blueprint = <Self::BatchMatmul as BatchMatmulFamily<()>>::Blueprint;
    type Config = <Self::BatchMatmul as BatchMatmulFamily<()>>::Config;

    fn expand_blueprint<R: cubecl::Runtime>(
        problem: &MatmulProblem,
        device_settings: &DeviceSettings<R>,
        strategy: &BlueprintStrategy<(), Self>,
    ) -> Result<ExpandInfo<Self::Blueprint>, MatmulSetupError> {
        let dtypes = MatmulElems::from_globals(&problem.global_dtypes);
        let properties = device_settings.client.properties();

        match strategy {
            BlueprintStrategy::Forced(blueprint) => Ok(ExpandInfo {
                blueprint: blueprint.clone(),
                dtypes,
            }),
            BlueprintStrategy::Inferred(strategy) => {
                let target_num_planes = match strategy.target_num_planes {
                    Some(num_planes) => num_planes,
                    None => num_concurrent_planes(&properties.hardware),
                };

                let kind = GemvKind::from_problem(problem)?;
                let tile_dim =
                    device_settings.plane_dim as usize * device_settings.vector_sizes.rhs;
                let num_planes = match kind {
                    GemvKind::MatVecRowMajor | GemvKind::VecMatColMajor => {
                        // For tile swizzle
                        max(1, min(target_num_planes, problem.k / tile_dim))
                    }
                    GemvKind::VecMatRowMajor | GemvKind::MatVecColMajor => {
                        // For within tile
                        max(1, min(target_num_planes, tile_dim))
                    }
                };

                let num_parallel_problems = match kind {
                    GemvKind::VecMatColMajor => problem.n,
                    GemvKind::VecMatRowMajor => problem.n / tile_dim,
                    GemvKind::MatVecRowMajor => problem.m,
                    GemvKind::MatVecColMajor => problem.m / tile_dim,
                };
                let check_bounds = if num_parallel_problems.is_multiple_of(num_planes) {
                    CheckBounds::None
                } else {
                    CheckBounds::Terminate
                };

                let blueprint = GemvPlaneParallelBlueprint {
                    dtypes: dtypes.clone(),
                    num_planes,
                    tile_dim,
                    hypercube_blueprint: HypercubeBlueprint::builder()
                        .cube_count_strategy(CubeCountStrategy::Flattened)
                        .global_order(GlobalOrder::RowMajor)
                        .build(),
                    kind,
                    check_bounds,
                };

                Ok(ExpandInfo { blueprint, dtypes })
            }
        }
    }

    fn prepare<R: cubecl::Runtime>(
        problem: &MatmulProblem,
        device_settings: &DeviceSettings<R>,
        expand_info: ExpandInfo<Self::Blueprint>,
    ) -> Result<LaunchInfo<Self::Blueprint>, MatmulSetupError> {
        let ExpandInfo { blueprint, dtypes } = expand_info;

        Self::validate_blueprint(
            &device_settings.client,
            &blueprint,
            problem,
            &dtypes,
            &device_settings.vector_sizes,
        )?;

        let cube_dim = Self::BatchMatmul::cubedim_resource(
            &blueprint,
            &dtypes,
            &device_settings.vector_sizes,
        )?
        .to_cube_dim(device_settings.plane_dim)?;

        let num_parallel_problems = match blueprint.kind {
            GemvKind::VecMatColMajor => problem.n,
            GemvKind::VecMatRowMajor => problem.n / blueprint.tile_dim,
            GemvKind::MatVecRowMajor => problem.m,
            GemvKind::MatVecColMajor => problem.m / blueprint.tile_dim,
        };

        let working_cubes = num_parallel_problems.div_ceil(blueprint.num_planes);

        let cube_count_plan = CubeCountPlan::from_blueprint(
            &blueprint.hypercube_blueprint,
            (working_cubes as u32, 1, problem.num_batches() as u32).into(),
            &device_settings.max_cube_count,
        );

        Ok(LaunchInfo {
            blueprint,
            dtypes,
            cube_dim,
            cube_count_plan,
            address_type: problem.address_type,
            vector_sizes: device_settings.vector_sizes,
        })
    }
}