cubek-convolution 0.2.0-pre.3

CubeK: Convolution 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
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

use cubecl::{
    server::LaunchError,
    {Runtime, client::ComputeClient, ir::StorageType, prelude::TensorBinding},
};
use cubek_matmul::components::{global::read::FullLoadingStrategy, tile::TileMatmulFamily};
use cubek_matmul::components::{
    global::read::sync_full_cyclic::SyncFullCyclicLoading,
    stage::{ColMajorTilingOrder, RowMajorTilingOrder},
};
use cubek_matmul::{
    components::global::read::{
        async_full_tma::AsyncFullTmaLoading, sync_full_strided::SyncFullStridedLoading,
        sync_full_tilewise::SyncFullTilewiseLoading,
    },
    routines::simple::SimpleAlgorithm,
};
use cubek_matmul::{
    definition::AvailableVectorSizes,
    launch::{TensorArgs, TensorMapArgs},
};
use cubek_std::tile::Strided;
use std::marker::PhantomData;

use crate::{
    algorithm::{contiguous_pitched_layout, into_tensor_handle_tma},
    components::{
        ConvolutionOperation,
        global::{
            args::RuntimeArgs,
            read::strategy::{
                async_full_cyclic::AsyncFullCyclicLoading,
                async_full_strided::AsyncFullStridedLoading, sync_bias::SyncBiasLoading,
            },
        },
    },
};

use super::Algorithm;

/// Cmma convolution
pub struct SimpleConv<
    TMM: TileMatmulFamily,
    LL: FullLoadingStrategy<RuntimeArgs>,
    LR: FullLoadingStrategy<RuntimeArgs>,
> {
    _tmm: PhantomData<TMM>,
    _loader: PhantomData<(LL, LR)>,
}

pub type SimpleSyncCyclicConv<TMM> = SimpleConv<
    TMM,
    SyncFullCyclicLoading<RowMajorTilingOrder>,
    SyncFullCyclicLoading<ColMajorTilingOrder>,
>;
pub type SimpleSyncStridedConv<TMM> =
    SimpleConv<TMM, SyncFullStridedLoading, SyncFullStridedLoading>;
pub type SimpleSyncTilewiseConv<TMM> = SimpleConv<
    TMM,
    SyncFullTilewiseLoading<RowMajorTilingOrder>,
    SyncFullTilewiseLoading<ColMajorTilingOrder>,
>;
pub type SimpleAsyncCyclicConv<TMM> = SimpleConv<
    TMM,
    AsyncFullCyclicLoading<RowMajorTilingOrder>,
    AsyncFullCyclicLoading<ColMajorTilingOrder>,
>;
pub type SimpleAsyncStridedConv<TMM> =
    SimpleConv<TMM, AsyncFullStridedLoading, AsyncFullStridedLoading>;

pub struct SimpleAsyncTmaConv<TMM: TileMatmulFamily> {
    _tmm: PhantomData<TMM>,
}

impl<
    TMM: TileMatmulFamily<
            LhsTile = Strided,
            RhsTile = Strided,
            AccTile = Option<Strided>,
            OutTile = Strided,
        >,
    LL: FullLoadingStrategy<RuntimeArgs, TileKind = Strided>,
    LR: FullLoadingStrategy<RuntimeArgs, TileKind = Strided, SyncStrategy = LL::SyncStrategy>,
> Algorithm for SimpleConv<TMM, LL, LR>
{
    type Routine = SimpleAlgorithm<TMM, LL, LR, SyncBiasLoading>;
    type Args = TensorArgs<RuntimeArgs>;

    fn correct_layout<R: Runtime>(
        client: &ComputeClient<R>,
        handle: TensorBinding<R>,
        dtype: StorageType,
        _operation: ConvolutionOperation,
    ) -> Result<TensorBinding<R>, LaunchError> {
        contiguous_pitched_layout(client, handle, dtype)
    }
}

impl<
    TMM: TileMatmulFamily<
            LhsTile = Strided,
            RhsTile = Strided,
            AccTile = Option<Strided>,
            OutTile = Strided,
        >,
> Algorithm for SimpleAsyncTmaConv<TMM>
{
    type Routine = SimpleAlgorithm<TMM, AsyncFullTmaLoading, AsyncFullTmaLoading, SyncBiasLoading>;

    type Args = TensorMapArgs<RuntimeArgs>;

    fn correct_layout<R: Runtime>(
        client: &ComputeClient<R>,
        handle: TensorBinding<R>,
        dtype: StorageType,
        operation: ConvolutionOperation,
    ) -> Result<TensorBinding<R>, LaunchError> {
        into_tensor_handle_tma(client, handle, dtype, operation)
    }

    fn filter_vector_sizes(vector_sizes: AvailableVectorSizes) -> AvailableVectorSizes {
        AvailableVectorSizes {
            lhs: vec![1],
            rhs: vec![1],
            out: vector_sizes.out,
        }
    }
}