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
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136

use cubecl_core::{Runtime, client::ComputeClient};

use crate::{
    components::{
        MatmulElems, MatmulLineSizes, MatmulProblem, MatmulSelection, MatmulSetupError,
        PartitionSize, TileSize, TilingScheme,
        batch::{
            CubeCountPlanSelection, GlobalOrderSelection, HypercubeSelection,
            PartitionedBatchMatmulFamily, RowMajorGlobalPartitionMatmul, SmAllocation,
        },
        global::{
            PlaneWriterFamily,
            multi_stage::double_buffering::DoubleBufferingMatmulFamily,
            read::{
                sync_full_cyclic::SyncFullCyclicLoading,
                sync_partial_cyclic::SyncPartialCyclicLoading,
            },
            single_stage::simple::SimpleMatmulFamily,
        },
        stage::{
            ColMajorTilingOrder, FilledStageFamily, PartitionBuffering, PlaneMatmulFamily,
            RowMajorTilingOrder, StridedStageFamily,
        },
        tile::{io::Filled, plane_vec_mat_inner_product::PlaneVecMatInnerProduct},
    },
    kernels::layered::Algorithm,
};

pub struct SimpleVecMatAlgorithm {}

impl Algorithm for SimpleVecMatAlgorithm {
    type SelectionArgs = ();
    type TileMatmul = PlaneVecMatInnerProduct<Filled>;
    type StageMatmul = PlaneMatmulFamily<
        Self::TileMatmul,
        StridedStageFamily,
        StridedStageFamily,
        FilledStageFamily,
    >;
    type GlobalMatmul = SimpleMatmulFamily<
        Self::StageMatmul,
        SyncFullCyclicLoading<RowMajorTilingOrder>,
        SyncFullCyclicLoading<ColMajorTilingOrder>,
        PlaneWriterFamily,
    >;
    type BatchMatmul =
        PartitionedBatchMatmulFamily<Self::GlobalMatmul, RowMajorGlobalPartitionMatmul>;

    fn selection<R: Runtime>(
        client: &ComputeClient<R::Server>,
        problem: &MatmulProblem,
        plane_dim: u32,
        line_sizes: &MatmulLineSizes,
        _elems: MatmulElems,
        _args: &Self::SelectionArgs,
    ) -> Result<MatmulSelection, MatmulSetupError> {
        Ok(selection_vecmat::<R>(
            client,
            problem,
            (1, line_sizes.out as u32, plane_dim * line_sizes.lhs as u32).into(),
            plane_dim,
        ))
    }
}

pub struct DoubleVecMatAlgorithm {}

impl Algorithm for DoubleVecMatAlgorithm {
    type SelectionArgs = ();
    type TileMatmul = PlaneVecMatInnerProduct<Filled>;
    type StageMatmul = PlaneMatmulFamily<
        Self::TileMatmul,
        StridedStageFamily,
        StridedStageFamily,
        FilledStageFamily,
    >;
    type GlobalMatmul = DoubleBufferingMatmulFamily<
        Self::StageMatmul,
        SyncPartialCyclicLoading<RowMajorTilingOrder>,
        SyncPartialCyclicLoading<ColMajorTilingOrder>,
        PlaneWriterFamily,
    >;
    type BatchMatmul =
        PartitionedBatchMatmulFamily<Self::GlobalMatmul, RowMajorGlobalPartitionMatmul>;

    fn selection<R: Runtime>(
        client: &ComputeClient<R::Server>,
        problem: &MatmulProblem,
        plane_dim: u32,
        line_sizes: &MatmulLineSizes,
        _elems: MatmulElems,
        _args: &Self::SelectionArgs,
    ) -> Result<MatmulSelection, MatmulSetupError> {
        Ok(selection_vecmat::<R>(
            client,
            problem,
            (1, line_sizes.out as u32, plane_dim * line_sizes.lhs as u32).into(),
            plane_dim,
        ))
    }
}

fn selection_vecmat<R: Runtime>(
    client: &ComputeClient<R::Server>,
    problem: &MatmulProblem,
    tile_size: TileSize,
    plane_dim: u32,
) -> MatmulSelection {
    let tiling_scheme = TilingScheme::builder()
        .with_tile_size(tile_size)
        .with_partition_size(PartitionSize::new(1, 1, 1))
        .with_stage_size((1, 1, 1).into())
        .build()
        .unwrap();
    let cube_count_plan = match client.properties().hardware.num_streaming_multiprocessors {
        Some(num_sms) => CubeCountPlanSelection::Sm {
            num_sms,
            sm_usage: SmAllocation::Exact,
            cubes_first: true,
        },
        None => CubeCountPlanSelection::FromProblem,
    };

    let hypercube = HypercubeSelection::builder(&tiling_scheme)
        .global_order(GlobalOrderSelection::SwizzleRow {
            m: problem.m as u32,
            w: 2,
        })
        .cube_count_plan(cube_count_plan)
        .build();

    MatmulSelection::builder(tiling_scheme, plane_dim)
        .partition_buffering(PartitionBuffering::Single)
        .hypercube_config(hypercube)
        .build()
}