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
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157

use std::marker::PhantomData;

use crate::components::tile::io::{Strided, TileKind};
use crate::components::tile::{
    TileConfig, TileMatmul,
    mma::{reader::MmaStageReader, writer::MmaStageWriter},
};
use crate::components::tile::{mma::reader::MmaFragmentReader, tile_data::StridedTile};
use crate::components::{StageIdent, tile::mma::config::MmaMatmulConfig};
use cubecl_core::prelude::*;
use cubecl_core::{self as cubecl, cmma::MmaDefinition, ir::MatrixIdent};

/// Uses one plane to perform a small matmul using accelerated instructions, with manual register
/// management.
/// Currently requires matrix layout to match the platform's preferred layout.
pub struct MmaMatmul<Acc: TileKind> {
    _ty: PhantomData<Acc>,
}

#[cube]
impl<L: Numeric, R: Numeric, A: Numeric, AccTile: TileKind> TileMatmul<L, R, A>
    for MmaMatmul<AccTile>
where
    MmaStageReader<AccTile>: MmaFragmentReader<TileKind = AccTile>,
{
    type Config = MmaMatmulConfig;
    type LhsFragment = Sequence<Line<L>>;
    type RhsFragment = Sequence<Line<R>>;
    type AccFragment = Sequence<Line<A>>;

    type LhsTile = Strided;
    type RhsTile = Strided;
    type AccTile = AccTile;
    type OutTile = Strided;

    fn execute(
        lhs: &Self::LhsFragment,
        rhs: &Self::RhsFragment,
        out: &mut Self::AccFragment,
        #[comptime] config: Self::Config,
    ) {
        let def = mma_definition(config);
        let out_arr = def.execute(lhs, rhs, out);
        let num_lines = def.lines_per_lane(MatrixIdent::Accumulator);

        #[unroll]
        for i in 0..num_lines {
            *out.index_mut(i) = out_arr[i];
        }
    }

    fn allocate_lhs(#[comptime] config: Self::Config) -> Self::LhsFragment {
        let def = mma_definition::<L, R, A>(config);
        let line_size = def.line_size(MatrixIdent::A);
        let mut frag = Sequence::new();
        #[unroll]
        for _ in 0..def.lines_per_lane(MatrixIdent::A) {
            // Needs to be mut because sequence is dodgy
            #[allow(unused_mut)]
            let mut reg = Line::empty(line_size);
            frag.push(reg);
        }
        frag
    }

    fn allocate_rhs(#[comptime] config: Self::Config) -> Self::RhsFragment {
        let def = mma_definition::<L, R, A>(config);
        let line_size = def.line_size(MatrixIdent::B);
        let mut frag = Sequence::new();
        #[unroll]
        for _ in 0..def.lines_per_lane(MatrixIdent::B) {
            // Needs to be mut because sequence is dodgy
            #[allow(unused_mut)]
            let mut reg = Line::empty(line_size);
            frag.push(reg);
        }
        frag
    }

    fn allocate_acc(#[comptime] config: Self::Config) -> Self::AccFragment {
        let def = mma_definition::<L, R, A>(config);
        let line_size = def.line_size(MatrixIdent::Accumulator);
        let mut frag = Sequence::new();
        #[unroll]
        for _ in 0..def.lines_per_lane(MatrixIdent::Accumulator) {
            // Needs to be mut because sequence is dodgy
            #[allow(unused_mut)]
            let mut reg = Line::empty(line_size);
            frag.push(reg);
        }
        frag
    }

    fn load_lhs<E: Numeric>(
        tile: &StridedTile<E>,
        lhs: &mut Self::LhsFragment,
        #[comptime] config: Self::Config,
    ) {
        MmaStageReader::<Self::LhsTile>::load_fragment(
            tile,
            lhs,
            mma_definition::<L, R, A>(config),
            MatrixIdent::A,
            config.matrix_layout(StageIdent::Lhs),
        );
    }

    fn load_rhs<E: Numeric>(
        tile: &StridedTile<E>,
        rhs: &mut Self::RhsFragment,
        #[comptime] config: Self::Config,
    ) {
        MmaStageReader::<Self::LhsTile>::load_fragment(
            tile,
            rhs,
            mma_definition::<L, R, A>(config),
            MatrixIdent::B,
            config.matrix_layout(StageIdent::Rhs),
        );
    }

    fn load_acc<E: Numeric>(
        tile: &AccTile::Tile<E>,
        acc: &mut Self::AccFragment,
        #[comptime] config: Self::Config,
    ) {
        MmaStageReader::<Self::AccTile>::load_fragment(
            tile,
            acc,
            mma_definition::<L, R, A>(config),
            MatrixIdent::Accumulator,
            config.matrix_layout(StageIdent::Acc),
        );
    }

    fn write_results<E: Numeric>(
        tile: &mut StridedTile<E, ReadWrite>,
        out: &Self::AccFragment,
        #[comptime] config: Self::Config,
    ) {
        MmaStageWriter::store_fragment(
            tile,
            out,
            mma_definition::<L, R, A>(config),
            MatrixIdent::Accumulator,
            config.matrix_layout(StageIdent::Out),
        );
    }
}

#[cube]
pub(super) fn mma_definition<L: Numeric, R: Numeric, A: Numeric>(
    #[comptime] config: MmaMatmulConfig,
) -> MmaDefinition<L, R, A> {
    let size = config.tile_size();
    MmaDefinition::new(size.m(), size.n(), size.k())
}