cubecl-cpp 0.3.0

CPP transpiler for CubeCL
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

use std::fmt::Display;

use super::{Dialect, Elem, Variable};

#[derive(Debug, Clone, PartialEq, Eq, Copy)]
pub enum FragmentIdent {
    A,
    B,
    Accumulator,
}

#[derive(Debug, Clone, PartialEq, Eq, Copy)]
pub enum FragmentLayout {
    ColMajor,
    RowMajor,
}

#[derive(Debug, Clone, PartialEq, Eq, Copy)]
pub struct Fragment<D: Dialect> {
    pub ident: FragmentIdent,
    pub m: u8,
    pub n: u8,
    pub k: u8,
    pub elem: Elem<D>,
    pub layout: Option<FragmentLayout>,
}

/// Warp Matrix-Multiply and Accumulate Instruction.
#[derive(Debug, Clone, Copy)]
pub enum WmmaInstruction<D: Dialect> {
    /// Fill the fragment with the value.
    Fill {
        frag: Variable<D>,
        value: Variable<D>,
    },
    /// Load the value into the fragment given the stride.
    Load {
        frag: Variable<D>,
        value: Variable<D>,
        stride: Variable<D>,
        layout: Option<FragmentLayout>,
    },
    /// Executes D=A*B+C;
    ///
    /// For implementing a matmul, `D=C` : `C+=A*B`
    Execute {
        frag_a: Variable<D>,
        frag_b: Variable<D>,
        frag_c: Variable<D>,
        frag_d: Variable<D>,
    },
    /// Store the fragment in an output variable following the stride and the layout.
    Store {
        output: Variable<D>,
        frag: Variable<D>,
        stride: Variable<D>,
        layout: FragmentLayout,
    },
}

impl Display for FragmentLayout {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            FragmentLayout::ColMajor => f.write_str("nvcuda::wmma::col_major"),
            FragmentLayout::RowMajor => f.write_str("nvcuda::wmma::row_major"),
        }
    }
}

impl Display for FragmentIdent {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            FragmentIdent::A => f.write_str("nvcuda::wmma::matrix_a"),
            FragmentIdent::B => f.write_str("nvcuda::wmma::matrix_b"),
            FragmentIdent::Accumulator => f.write_str("nvcuda::wmma::accumulator"),
        }
    }
}

impl<D: Dialect> Display for Fragment<D> {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self.layout {
            Some(layout) => write!(
                f,
                "nvcuda::wmma::fragment<{}, {}, {}, {}, {}, {}>",
                self.ident, self.m, self.n, self.k, self.elem, layout
            ),
            None => write!(
                f,
                "nvcuda::wmma::fragment<{}, {}, {}, {}, {}>",
                self.ident, self.m, self.n, self.k, self.elem,
            ),
        }
    }
}

impl<D: Dialect> Display for WmmaInstruction<D> {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        match self {
            WmmaInstruction::Fill { frag, value } => {
                writeln!(f, "nvcuda::wmma::fill_fragment({frag}, {value});")
            }
            WmmaInstruction::Load {
                frag,
                value,
                stride,
                layout: None,
            } => writeln!(
                f,
                "nvcuda::wmma::load_matrix_sync({frag}, {value}, {stride});"
            ),
            WmmaInstruction::Load {
                frag,
                value,
                stride,
                layout: Some(layout),
            } => {
                let layout = match layout {
                    FragmentLayout::ColMajor => "nvcuda::wmma::mem_col_major",
                    FragmentLayout::RowMajor => "nvcuda::wmma::mem_row_major",
                };
                writeln!(
                    f,
                    "nvcuda::wmma::load_matrix_sync({frag}, {value}, {stride}, {layout});"
                )
            }
            WmmaInstruction::Execute {
                frag_a,
                frag_b,
                frag_c,
                frag_d,
            } => writeln!(
                f,
                "nvcuda::wmma::mma_sync({frag_d}, {frag_a}, {frag_b}, {frag_c});"
            ),
            WmmaInstruction::Store {
                output,
                frag,
                stride,
                layout,
            } => {
                let layout = match layout {
                    FragmentLayout::ColMajor => "nvcuda::wmma::mem_col_major",
                    FragmentLayout::RowMajor => "nvcuda::wmma::mem_row_major",
                };

                writeln!(
                    f,
                    "nvcuda::wmma::store_matrix_sync({output}, {frag}, {stride}, {layout});"
                )
            }
        }
    }
}