burn-cubecl 0.22.0-pre.2

Generic backend that can be compiled just-in-time to any shader language target
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

use burn_backend::cubecl::dtype_to_storage_type;
use burn_backend::ops::ConvOptions;
use burn_std::Shape;
use cubek::{
    convolution::{
        AcceleratedTileKind, ConvAlgorithm, ConvolutionArgs, ConvolutionInputs, Strategy,
        components::ConvSetupError, launch_ref,
    },
    matmul::definition::{MatmulElems, MatmulGlobalElems},
    std::InputBinding,
};

use crate::{CubeRuntime, ops::numeric::empty_device_dtype, tensor::CubeTensor};

pub(crate) fn wgrad_gemm_simple_sync<R: CubeRuntime, const N: usize>(
    input: CubeTensor<R>,
    out_grad: CubeTensor<R>,
    weight_shape: Shape,
    options: ConvOptions<N>,
    tile_kind: AcceleratedTileKind,
) -> Result<CubeTensor<R>, ConvSetupError> {
    let algorithm = match tile_kind {
        AcceleratedTileKind::Cmma => ConvAlgorithm::SimpleSyncCyclic,
        AcceleratedTileKind::Mma => ConvAlgorithm::SimpleSyncStrided,
    };
    launch_backwards_weight::<R, N>(
        &Strategy::Inferred {
            algorithm,
            tile_kind,
        },
        input,
        out_grad,
        weight_shape,
        options,
    )
}

pub(crate) fn wgrad_gemm_simple_async<R: CubeRuntime, const N: usize>(
    input: CubeTensor<R>,
    out_grad: CubeTensor<R>,
    weight_shape: Shape,
    options: ConvOptions<N>,
    tile_kind: AcceleratedTileKind,
) -> Result<CubeTensor<R>, ConvSetupError> {
    let algorithm = match tile_kind {
        AcceleratedTileKind::Cmma => ConvAlgorithm::SimpleAsyncCyclic,
        AcceleratedTileKind::Mma => ConvAlgorithm::SimpleAsyncStrided,
    };
    launch_backwards_weight::<R, N>(
        &Strategy::Inferred {
            algorithm,
            tile_kind,
        },
        input,
        out_grad,
        weight_shape,
        options,
    )
}

pub(crate) fn wgrad_gemm_simple_tma<R: CubeRuntime, const N: usize>(
    input: CubeTensor<R>,
    out_grad: CubeTensor<R>,
    weight_shape: Shape,
    options: ConvOptions<N>,
    tile_kind: AcceleratedTileKind,
) -> Result<CubeTensor<R>, ConvSetupError> {
    launch_backwards_weight::<R, N>(
        &Strategy::Inferred {
            algorithm: ConvAlgorithm::SimpleAsyncTma,
            tile_kind,
        },
        input,
        out_grad,
        weight_shape,
        options,
    )
}

/// Perform a convolution backwards weight pass using the implicit GEMM (im2col) algorithm, using
/// cubecl tiling matmul components.
///
/// * `input` - The input feature map
/// * `out_grad` - The output gradients
/// * `weight_shape` - The shape of the weights/weight gradients
/// * `options` - The options to use for the convolution
pub fn launch_backwards_weight<R: CubeRuntime, const N: usize>(
    strategy: &Strategy,
    input: CubeTensor<R>,
    out_grad: CubeTensor<R>,
    weight_shape: Shape,
    options: ConvOptions<N>,
) -> Result<CubeTensor<R>, ConvSetupError> {
    if options.groups != 1 {
        return Err(ConvSetupError::Groups(options.groups));
    }

    let out_dtype = out_grad.dtype;

    let weight_grad = empty_device_dtype(
        input.client.clone(),
        input.device.clone(),
        weight_shape,
        out_dtype,
    );

    let client = input.client.clone();
    let dtypes = MatmulElems::from_globals(&MatmulGlobalElems {
        lhs: dtype_to_storage_type(input.dtype),
        rhs: dtype_to_storage_type(out_grad.dtype),
        out: dtype_to_storage_type(out_dtype),
    });
    let input_dtype = input.dtype;
    let out_grad_dtype = out_grad.dtype;
    let input = InputBinding::new(input.binding(), dtype_to_storage_type(input_dtype));
    let out_grad = InputBinding::new(out_grad.binding(), dtype_to_storage_type(out_grad_dtype));

    launch_ref::<R, N>(
        strategy,
        &client,
        ConvolutionInputs::BackwardWeight {
            input,
            out_grad,
            weight_grad: weight_grad.clone().binding(),
        },
        ConvolutionArgs {
            stride: options.stride,
            padding: options.padding,
            dilation: options.dilation,
        },
        dtypes,
    )?;

    Ok(weight_grad)
}