cubecl-cpp 0.10.0-pre.3

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

use core::fmt::Formatter;

use crate::{
    Dialect,
    cuda::ptx,
    shared::{Elem, FragmentIdent, MmaShape},
};

#[derive(Debug, Clone, Default, PartialEq)]
pub enum Extension<D: Dialect> {
    #[default]
    NoExtension,
    Mma(MmaExtension<D>),
}

#[derive(Debug, Clone, PartialEq)]
pub enum MmaExtension<D: Dialect> {
    Execute(MmaExecute<D>),
    ExecuteScaled(MmaExecuteScaled<D>),
    LdMatrix(LdMatrix<D>),
    StMatrix(StMatrix<D>),
}

#[derive(Debug, Clone, PartialEq, Copy)]
pub struct Fragment<D: Dialect>(pub Elem<D>);

impl<D: Dialect> Fragment<D> {
    pub fn elem(&self) -> Elem<D> {
        self.0
    }
}

#[derive(new, Debug, Clone, PartialEq)]
pub struct MmaExecute<D: Dialect> {
    pub shape: MmaShape<D>,
    pub frag_a: Fragment<D>,
    pub frag_b: Fragment<D>,
    pub frag_c: Fragment<D>,
    pub frag_d: Fragment<D>,
}

#[derive(new, Debug, Clone, PartialEq)]
pub struct MmaExecuteScaled<D: Dialect> {
    pub shape: MmaShape<D>,
    pub frag_a: Fragment<D>,
    pub frag_b: Fragment<D>,
    pub frag_c: Fragment<D>,
    pub frag_d: Fragment<D>,
    pub scales_elem: Elem<D>,
    pub scales_factor: u32,
}

#[derive(new, Debug, Clone, PartialEq)]
pub struct LdMatrix<D: Dialect> {
    pub elem: Elem<D>,
    pub factor: u32,
    pub transpose: bool,
}

#[derive(new, Debug, Clone, PartialEq)]
pub struct StMatrix<D: Dialect> {
    pub elem: Elem<D>,
    pub factor: u32,
    pub transpose: bool,
}

impl<D: Dialect> MmaExtension<D> {
    pub fn format_extension(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
        match self {
            MmaExtension::Execute(mma_execute) => mma_execute.format_extension(f),
            MmaExtension::ExecuteScaled(mma_execute) => mma_execute.format_extension(f),
            MmaExtension::LdMatrix(ld_matrix) => ld_matrix.format_extension(f),
            MmaExtension::StMatrix(st_matrix) => st_matrix.format_extension(f),
        }
    }
}

impl<D: Dialect> MmaExecute<D> {
    pub fn format_extension(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
        let a_elems = self.shape.num_elems(FragmentIdent::<D>::A) / 32;
        let b_elems = self.shape.num_elems(FragmentIdent::<D>::B) / 32;
        let c_elems = self.shape.num_elems(FragmentIdent::<D>::Accumulator) / 32;
        let d_elems = self.shape.num_elems(FragmentIdent::<D>::Accumulator) / 32;

        let a_regs = a_elems as usize / (32 / self.frag_a.elem().unpacked().size_bits());
        let b_regs = b_elems as usize / (32 / self.frag_b.elem().unpacked().size_bits());
        let c_regs = c_elems as usize / (32 / self.frag_c.elem().unpacked().size_bits());
        let d_regs = d_elems as usize / (32 / self.frag_d.elem().unpacked().size_bits());

        let ptx = ptx::mma_template(
            self.frag_a.elem().unpacked(),
            self.frag_b.elem().unpacked(),
            self.frag_c.elem().unpacked(),
            self.shape.k,
            a_regs,
            b_regs,
            c_regs,
            d_regs,
        );

        f.write_str(&ptx)
    }
}

impl<D: Dialect> MmaExecuteScaled<D> {
    pub fn format_extension(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
        let a_elems = self.shape.num_elems(FragmentIdent::<D>::A) / 32;
        let b_elems = self.shape.num_elems(FragmentIdent::<D>::B) / 32;
        let c_elems = self.shape.num_elems(FragmentIdent::<D>::Accumulator) / 32;
        let d_elems = self.shape.num_elems(FragmentIdent::<D>::Accumulator) / 32;

        let a_regs = a_elems as usize / (32 / self.frag_a.elem().unpacked().size_bits());
        let b_regs = b_elems as usize / (32 / self.frag_b.elem().unpacked().size_bits());
        let c_regs = c_elems as usize / (32 / self.frag_c.elem().unpacked().size_bits());
        let d_regs = d_elems as usize / (32 / self.frag_d.elem().unpacked().size_bits());

        let ptx = ptx::mma_scaled_template(
            self.frag_a.elem().unpacked(),
            self.frag_b.elem().unpacked(),
            self.frag_c.elem().unpacked(),
            self.shape.k,
            a_regs,
            b_regs,
            c_regs,
            d_regs,
            self.scales_elem,
            self.scales_factor,
        );

        f.write_str(&ptx)
    }
}

impl<D: Dialect> LdMatrix<D> {
    pub fn format_extension(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
        let ptx = ptx::ldmatrix_template(self.elem, self.factor, self.transpose);

        f.write_str(&ptx)
    }
}

impl<D: Dialect> StMatrix<D> {
    pub fn format_extension(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
        let ptx = ptx::stmatrix_template(self.elem, self.factor, self.transpose);

        f.write_str(&ptx)
    }
}