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

use cubecl_core::{
    self as cubecl,
    ir::{Instruction, ManagedVariable, Scope},
};
use cubecl_core::{
    cube,
    ir::{Allocator, CoopMma, MatrixIdent, Operation, Processor, ScopeProcessing},
};

#[derive(new, Debug)]
pub struct CudaMmaProcessor;

impl Processor for CudaMmaProcessor {
    fn transform(&self, mut processing: ScopeProcessing, allocator: Allocator) -> ScopeProcessing {
        let mut instructions = Vec::new();
        core::mem::swap(&mut processing.instructions, &mut instructions);

        for instruction in instructions {
            match instruction.operation {
                Operation::CoopMma(CoopMma::RowIndex { lane_id, i, matrix }) => {
                    let lane_id = ManagedVariable::Plain(lane_id);
                    let i = ManagedVariable::Plain(i);
                    let elems_per_reg = 32 / matrix.storage.elem_type().size_bits();
                    let mut scope = Scope::root(false)
                        .with_allocator(allocator.clone())
                        .with_types(processing.typemap.clone());
                    let row_idx: ManagedVariable = row_index::expand(
                        &mut scope,
                        lane_id.into(),
                        i.into(),
                        elems_per_reg as u32,
                        matrix.ident,
                    )
                    .into();
                    let tmp_processing = scope.process([]);
                    for inst in tmp_processing.instructions {
                        processing.instructions.push(inst);
                    }
                    for var in tmp_processing.variables {
                        processing.variables.push(var);
                    }

                    processing.instructions.push(Instruction::new(
                        Operation::Copy(*row_idx),
                        instruction.out(),
                    ));
                }
                Operation::CoopMma(CoopMma::ColIndex { lane_id, i, matrix }) => {
                    let lane_id = ManagedVariable::Plain(lane_id);
                    let i = ManagedVariable::Plain(i);
                    let elems_per_reg = 32 / matrix.storage.elem_type().size_bits();
                    let mut scope = Scope::root(false)
                        .with_allocator(allocator.clone())
                        .with_types(processing.typemap.clone());
                    let col_idx: ManagedVariable = col_index::expand(
                        &mut scope,
                        lane_id.into(),
                        i.into(),
                        elems_per_reg as u32,
                        matrix.ident,
                    )
                    .into();
                    let tmp_processing = scope.process([]);
                    for inst in tmp_processing.instructions {
                        processing.instructions.push(inst);
                    }
                    for var in tmp_processing.variables {
                        processing.variables.push(var);
                    }

                    processing.instructions.push(Instruction::new(
                        Operation::Copy(*col_idx),
                        instruction.out(),
                    ));
                }
                _ => {
                    processing.instructions.push(instruction);
                }
            }
        }

        processing
    }
}

/// Derived from PTX shape documentation
/// <https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-matrix-instructions-for-mma>
#[cube]
fn row_index(
    lane_id: u32,
    i: u32,
    #[comptime] elems_per_reg: u32,
    #[comptime] ident: MatrixIdent,
) -> u32 {
    match ident {
        MatrixIdent::A => {
            let group_id = lane_id / 4;
            let odd_register = (i / elems_per_reg) & 1;
            group_id + odd_register * 8
        }
        MatrixIdent::B => {
            let thread_id_in_group = lane_id % 4;
            let offset = thread_id_in_group * elems_per_reg + (i % elems_per_reg);
            let reg = i / elems_per_reg;
            offset + elems_per_reg * 4 * reg
        }
        MatrixIdent::Accumulator => {
            let group_id = lane_id / 4;
            let offset = (i << 2) & 8;
            group_id + offset
        }
    }
}

/// Derived from PTX shape documentation
/// <https://docs.nvidia.com/cuda/parallel-thread-execution/#warp-level-matrix-instructions-for-mma>
#[cube]
fn col_index(
    lane_id: u32,
    i: u32,
    #[comptime] elems_per_reg: u32,
    #[comptime] ident: MatrixIdent,
) -> u32 {
    match ident {
        MatrixIdent::A => {
            let thread_id_in_group = lane_id % 4;
            let offset = thread_id_in_group * elems_per_reg + (i % elems_per_reg);
            let group_2 = (i / (2 * elems_per_reg)) & 1;
            offset + 4 * elems_per_reg * group_2
        }
        MatrixIdent::B => lane_id >> 2,
        MatrixIdent::Accumulator => {
            let thread_id_in_group = lane_id % 4;
            (thread_id_in_group * 2) + (i % 2)
        }
    }
}