cubek-std 0.2.0

CubeK: Standard Library
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
173
174
175

use cubecl::{cmma::MmaDefinition, ir::MatrixIdent, prelude::*};

use crate::{
    MatrixLayout, TileSize,
    tile::{
        compute::matmul::mma::{MmaFragmentReader as _, MmaStageReader, MmaStageWriter},
        data::{Filled, MmaMatmul, NA, NL, NR, Strided, StridedTile},
    },
};

#[cube]
fn make_mma_definition<L: Numeric, R: Numeric, A: Numeric>(
    #[comptime] config: MmaMatmul,
) -> MmaDefinition<L, R, A> {
    MmaDefinition::new(
        config.tile_size.m() as usize,
        config.tile_size.n() as usize,
        config.tile_size.k() as usize,
    )
}

#[cube]
pub fn mma_execute<L: Numeric, R: Numeric, A: Numeric>(
    lhs: &Array<Vector<L, NL>>,
    rhs: &Array<Vector<R, NR>>,
    acc: &mut Array<Vector<A, NA>>,
    #[comptime] _matrix_layout: MatrixLayout,
    #[comptime] config: MmaMatmul,
) {
    let def = MmaDefinition::<L, R, A>::new(
        config.tile_size.m() as usize,
        config.tile_size.n() as usize,
        config.tile_size.k() as usize,
    );
    let out_arr = def.execute(lhs, rhs, acc);
    let num_vectors = def.vectors_per_lane(MatrixIdent::Accumulator);
    #[unroll]
    for i in 0..num_vectors {
        acc[i] = out_arr[i];
    }
}

#[cube]
pub fn mma_load_lhs_from_shared<
    E: Numeric,
    ES: Size,
    L: Numeric,
    R: Numeric,
    A: Numeric,
    IO: SliceVisibility,
>(
    shared: &StridedTile<E, ES, IO>,
    fragment: &mut Array<Vector<L, NL>>,
    #[comptime] matrix_layout: MatrixLayout,
    #[comptime] config: MmaMatmul,
) {
    let shared = shared.to_read_only();
    let def = make_mma_definition::<L, R, A>(config);
    MmaStageReader::<Strided>::load_fragment(
        &shared,
        fragment,
        def,
        MatrixIdent::A,
        matrix_layout,
        comptime!(TileSize::new(
            config.tile_size.m(),
            config.tile_size.n(),
            config.tile_size.k(),
        )),
        config.mma_io_config,
    );
}

#[cube]
pub fn mma_load_rhs_from_shared<
    E: Numeric,
    ES: Size,
    R: Numeric,
    L: Numeric,
    A: Numeric,
    IO: SliceVisibility,
>(
    shared: &StridedTile<E, ES, IO>,
    fragment: &mut Array<Vector<R, NR>>,
    #[comptime] matrix_layout: MatrixLayout,
    #[comptime] config: MmaMatmul,
) {
    let shared = shared.to_read_only();
    let def = make_mma_definition::<L, R, A>(config);
    MmaStageReader::<Strided>::load_fragment(
        &shared,
        fragment,
        def,
        MatrixIdent::B,
        matrix_layout,
        comptime!(TileSize::new(
            config.tile_size.m(),
            config.tile_size.n(),
            config.tile_size.k(),
        )),
        config.mma_io_config,
    );
}

#[cube]
pub fn mma_load_acc_from_shared<
    E: Numeric,
    ES: Size,
    A: Numeric,
    L: Numeric,
    R: Numeric,
    IO: SliceVisibility,
>(
    shared: &StridedTile<E, ES, IO>,
    fragment: &mut Array<Vector<A, NA>>,
    #[comptime] matrix_layout: MatrixLayout,
    #[comptime] config: MmaMatmul,
) {
    let shared = shared.to_read_only();
    let def = make_mma_definition::<L, R, A>(config);
    MmaStageReader::<Strided>::load_fragment(
        &shared,
        fragment,
        def,
        MatrixIdent::Accumulator,
        matrix_layout,
        comptime!(TileSize::new(
            config.tile_size.m(),
            config.tile_size.n(),
            config.tile_size.k(),
        )),
        config.mma_io_config,
    );
}

#[cube]
pub fn mma_load_acc_zeros<E: Numeric, ES: Size, A: Numeric, L: Numeric, R: Numeric>(
    fragment: &mut Array<Vector<A, NA>>,
    #[comptime] matrix_layout: MatrixLayout,
    #[comptime] config: MmaMatmul,
) {
    let def = make_mma_definition::<L, R, A>(config);
    MmaStageReader::<Filled>::load_fragment::<A, NA, E, ES, L, R, A>(
        &E::from_int(0),
        fragment,
        def,
        MatrixIdent::Accumulator,
        matrix_layout,
        comptime!(TileSize::new(
            config.tile_size.m(),
            config.tile_size.n(),
            config.tile_size.k(),
        )),
        config.mma_io_config,
    );
}

#[cube]
pub fn mma_write_to_shared<E: Numeric, ES: Size, A: Numeric, L: Numeric, R: Numeric>(
    shared: &mut StridedTile<E, ES, ReadWrite>,
    fragment: &Array<Vector<A, NA>>,
    #[comptime] config: MmaMatmul,
) {
    let def = make_mma_definition::<L, R, A>(config);
    let out_layout = comptime!(shared.layout);
    MmaStageWriter::store_fragment(
        shared,
        fragment,
        def,
        MatrixIdent::Accumulator,
        out_layout,
        config.tile_size.m(),
        config.mma_io_config,
    );
}