cubek-convolution 0.3.0

CubeK: Convolution Kernels
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

use cubecl::{ir::AddressType, prelude::*, zspace::shape};
use cubek_convolution::{
    ConvAlgorithm, ConvolutionArgs, ConvolutionInputs, Strategy,
    components::{ConvolutionOperation, ConvolutionProblem, Dimensionality},
    definition::{BackwardDataBlueprint, ConvBlueprint},
    launch_ref,
};
use cubek_matmul::definition::{AccumulatorOperand, MatmulElems};
use cubek_matmul::multi_level::{
    components::tile::TileMatmulKind, definition::BatchMatmulBlueprint,
};
use cubek_std::{InputBinding, MatrixLayout};
use cubek_test_utils::{
    ExecutionOutcome, HostData, HostDataType, TestInput, TestOutcome, launch_and_capture_outcome,
};

use super::common::{default_partition_buffering, default_swizzle, default_tiling_scheme};

#[test]
fn backward_data_masks_non_divisible_stride_coordinates() {
    run_backward_data_case(
        16,
        7,
        2,
        &[
            16., 16., 32., 16., 32., 16., 32., 16., 32., 16., 32., 16., 32., 16., 16., 0.,
        ],
    );
}

/// End padding is not present in `ConvolutionArgs`; it is encoded by the larger out-gradient
/// extent. For input length 8, a three-tap kernel, and no beginning padding, end padding 2 makes
/// the forward output (and therefore `out_grad`) length 8 rather than 6.
#[test]
fn backward_data_supports_end_only_padding() {
    run_backward_data_case(8, 8, 1, &[16., 32., 48., 48., 48., 48., 48., 48.]);
}

fn run_backward_data_case(
    in_len: usize,
    out_len: usize,
    stride: usize,
    expected_per_position: &[f32],
) {
    let client = cubecl::test_device().client();
    let dtypes = MatmulElems::from_single_dtype(half::f16::elem_type_native());

    let batches = 1;
    let kernel_len = 3;
    let channels = 16;
    let out_channels = 16;

    assert_eq!(expected_per_position.len(), in_len);

    let out_grad = TestInput::builder(client.clone(), shape![batches, out_len, out_channels])
        .dtype(dtypes.lhs_global)
        .custom(vec![1.; batches * out_len * out_channels])
        .generate_without_host_data();
    let weights = TestInput::builder(client.clone(), shape![out_channels, kernel_len, channels])
        .dtype(dtypes.rhs_global)
        .custom(vec![1.; out_channels * kernel_len * channels])
        .generate_without_host_data();
    let in_grad = TestInput::builder(client.clone(), shape![batches, in_len, channels])
        .dtype(dtypes.acc_global)
        .zeros()
        .generate_without_host_data();

    let problem = ConvolutionProblem {
        m: batches * in_len,
        n: channels,
        k: out_channels * kernel_len,
        lhs_strides: out_grad.strides().clone(),
        rhs_strides: weights.strides().clone(),
        lhs_layout: MatrixLayout::RowMajor,
        rhs_layout: MatrixLayout::RowMajor,
        kernel_size: vec![kernel_len as u32],
        stride: vec![stride as u32],
        padding: vec![0],
        dilation: vec![1],
        batches,
        channels,
        out_channels,
        in_shape: shape![in_len],
        out_shape: shape![out_len],
        padded_channels: out_channels,
        operation: ConvolutionOperation::BackwardData,
        dimensionality: Dimensionality::Dim1,
        global_dtypes: dtypes.as_global_elems(),
        address_type: AddressType::U32,
    };
    let matmul = BatchMatmulBlueprint::builder(
        TileMatmulKind::Cmma,
        default_tiling_scheme(),
        client.properties().hardware.plane_size_max,
        &problem.as_matmul_problem(AccumulatorOperand::Absent),
    )
    .shared_swizzle(default_swizzle())
    .partition_buffering(default_partition_buffering())
    .build();

    let inputs = ConvolutionInputs::BackwardData {
        out_grad: InputBinding::new(out_grad.binding(), dtypes.lhs_global),
        weights: InputBinding::new(weights.binding(), dtypes.rhs_global),
        in_grad: in_grad.clone().binding(),
    };
    let args = ConvolutionArgs::<1> {
        stride: [stride],
        padding: [0],
        dilation: [1],
    };
    let strategy = Strategy::Forced {
        algorithm: ConvAlgorithm::SimpleSyncCyclic,
        blueprint: ConvBlueprint::BackwardData(BackwardDataBlueprint {
            matmul,
            dimensionality: Dimensionality::Dim1,
        }),
    };

    let outcome =
        launch_and_capture_outcome(&client, &[&in_grad.handle], |client| {
            match launch_ref(&strategy, client, inputs, args, dtypes) {
                Ok(()) => ExecutionOutcome::Executed,
                Err(error) => ExecutionOutcome::CompileError(format!("{error:?}")),
            }
        });

    if let ExecutionOutcome::CompileError(error) = outcome {
        TestOutcome::CompileError(error).enforce();
        return;
    }

    let actual = HostData::from_tensor_handle(&client, in_grad, HostDataType::F32);

    for (position, &expected) in expected_per_position.iter().enumerate() {
        for channel in 0..channels {
            assert_eq!(actual.get_f32(&[0, position, channel]), expected);
        }
    }
}