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
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172

use std::marker::PhantomData;

use cubecl::prelude::*;
use cubecl_core::{self as cubecl, cmma::MmaDefinition, ir::MatrixIdent};
use cubecl_std::{CubeOption, CubeOptionExpand};

use crate::components::{
    MatrixLayout, as_cmma_layout,
    tile::{
        StridedTile,
        io::{Filled, Strided, TileKind},
    },
};

/// Generic CMMA reader over any tile type
#[cube]
pub(crate) trait MmaFragmentReader {
    type TileKind: TileKind;

    /// Fill a fragment with data, with the implementation depending on the tile kind.
    fn load_fragment<E: Numeric, V: Numeric, A: Numeric, B: Numeric, CD: Numeric>(
        tile: &<Self::TileKind as TileKind>::Tile<V>,
        fragment: &mut Sequence<Line<E>>,
        def: MmaDefinition<A, B, CD>,
        #[comptime] ident: MatrixIdent,
        #[comptime] layout: MatrixLayout,
    );
}

/// Reader to load the manual MMA registers. Tile kind determines implementation.
#[derive(CubeType)]
pub struct MmaStageReader<Kind: TileKind> {
    #[cube(comptime)]
    _ty: PhantomData<Kind>,
}

#[cube]
impl MmaFragmentReader for MmaStageReader<Strided> {
    type TileKind = Strided;

    fn load_fragment<E: Numeric, V: Numeric, A: Numeric, B: Numeric, CD: Numeric>(
        tile: &StridedTile<V>,
        fragment: &mut Sequence<Line<E>>,
        def: MmaDefinition<A, B, CD>,
        #[comptime] ident: MatrixIdent,
        #[comptime] layout: MatrixLayout,
    ) {
        let line_layout = def.line_layout(ident);

        let transposed = comptime![as_cmma_layout(layout) != line_layout];

        if transposed {
            load_transposed(tile, fragment, def, ident, layout);
        } else {
            load_plain(tile, fragment, def, ident, layout);
        }
    }
}

#[cube]
fn load_transposed<E: Numeric, V: Numeric, A: Numeric, B: Numeric, CD: Numeric>(
    tile: &StridedTile<V>,
    fragment: &mut Sequence<Line<E>>,
    def: MmaDefinition<A, B, CD>,
    #[comptime] ident: MatrixIdent,
    #[comptime] layout: MatrixLayout,
) {
    let num_lines = def.lines_per_lane(ident);
    let line_size = def.line_size(ident);
    let lane_id = UNIT_POS_PLANE;

    let (_, stride) = tile.as_unlined();
    let slice = tile.slice.with_line_size(1u32);

    let (stride_row, stride_col) = match layout {
        MatrixLayout::RowMajor => (stride, 1),
        MatrixLayout::ColMajor => (1, stride),
    };

    #[unroll]
    for i in 0..num_lines {
        let line = fragment.index_mut(i);
        #[unroll]
        for n in 0..line_size {
            let elem_idx = i * line_size + n;
            let (row, col) = def.position_of_nth(lane_id, elem_idx, ident);
            let offset = row * stride_row + col * stride_col;
            line[n] = E::cast_from(slice[offset]);
        }
    }
}

#[cube]
fn load_plain<E: Numeric, V: Numeric, A: Numeric, B: Numeric, CD: Numeric>(
    tile: &StridedTile<V>,
    fragment: &mut Sequence<Line<E>>,
    def: MmaDefinition<A, B, CD>,
    #[comptime] ident: MatrixIdent,
    #[comptime] layout: MatrixLayout,
) {
    let num_lines = def.lines_per_lane(ident);
    let line_size = def.line_size(ident);
    let lane_id = UNIT_POS_PLANE;
    let (_, stride) = tile.as_unlined();
    // Supported on all targets that support manual MMA
    let slice = tile.slice.with_line_size(line_size);

    let (stride_row, stride_col) = match layout {
        MatrixLayout::RowMajor => (stride, 1),
        MatrixLayout::ColMajor => (1, stride),
    };

    #[unroll]
    for i in 0..num_lines {
        let elem_idx = i * line_size;
        let (row, col) = def.position_of_nth(lane_id, elem_idx, ident);
        let offset = row * stride_row + col * stride_col;

        let value = slice[offset / line_size];
        *fragment.index_mut(i) = Line::cast_from(value);
    }
}

#[cube]
impl MmaFragmentReader for MmaStageReader<Filled> {
    type TileKind = Filled;

    fn load_fragment<E: Numeric, V: Numeric, A: Numeric, B: Numeric, CD: Numeric>(
        value: &V,
        fragment: &mut Sequence<Line<E>>,
        def: MmaDefinition<A, B, CD>,
        #[comptime] ident: MatrixIdent,
        #[comptime] _layout: MatrixLayout,
    ) {
        let num_lines = def.lines_per_lane(ident);
        let value = Line::<E>::cast_from(*value);

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

#[cube]
impl<Inner: TileKind> MmaFragmentReader for MmaStageReader<CubeOption<Inner>>
where
    MmaStageReader<Inner>: MmaFragmentReader<TileKind = Inner>,
{
    type TileKind = CubeOption<Inner>;

    fn load_fragment<E: Numeric, V: Numeric, A: Numeric, B: Numeric, CD: Numeric>(
        tile: &CubeOption<Inner::Tile<V>>,
        fragment: &mut Sequence<Line<E>>,
        def: MmaDefinition<A, B, CD>,
        #[comptime] ident: MatrixIdent,
        #[comptime] layout: MatrixLayout,
    ) {
        match tile {
            CubeOption::Some(tile) => {
                MmaStageReader::<Inner>::load_fragment(tile, fragment, def, ident, layout)
            }
            CubeOption::None => MmaStageReader::<Filled>::load_fragment::<E, V, A, B, CD>(
                &V::from_int(0),
                fragment,
                def,
                ident,
                layout,
            ),
        }
    }
}