cubek-matmul 0.2.0

CubeK: Matrix Multiplication Kernels
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

use cubecl::{prelude::barrier::Barrier, prelude::*};

use crate::{
    components::{
        global::{SharedGlobalMatmulConfig, read::SyncStrategy},
        stage::StageConfig,
    },
    definition::MatmulTypes,
};

/// Asynchronous barrier for `async_memcpy`
pub struct AsyncBarrier {}

#[cube]
impl SyncStrategy for AsyncBarrier {
    type Barrier = Shared<Barrier>;

    fn create_barrier() -> Self::Barrier {
        Barrier::shared(CUBE_DIM, UNIT_POS == 0)
    }

    fn sync<MP: MatmulTypes, S: StageConfig>(
        barrier: &mut Self::Barrier,
        #[comptime] _config: SharedGlobalMatmulConfig<S>,
    ) {
        barrier.arrive_and_wait();
    }
}

/// Asynchronous barrier for `async_copy`
pub struct AsyncCopy {}

#[cube]
impl SyncStrategy for AsyncCopy {
    type Barrier = Shared<Barrier>;

    fn create_barrier() -> Self::Barrier {
        Barrier::shared(CUBE_DIM, UNIT_POS == 0)
    }

    fn sync<MP: MatmulTypes, S: StageConfig>(
        barrier: &mut Self::Barrier,
        #[comptime] _config: SharedGlobalMatmulConfig<S>,
    ) {
        barrier.commit_copy_async();
        barrier.arrive_and_wait();
    }
}