cubek-convolution 0.4.0-pre.1

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
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192

//! Force the split weight-gradient kernel and compare with the defining sum.

use cubecl::{prelude::*, std::tensor::TensorHandle};
use cubek_convolution::{
    ConvolutionArgs, DepthwiseWeightBackwardTensors, launch_depthwise_weight_backward,
};

fn upload(
    client: &Client,
    shape: [usize; 4],
    values: &[f32],
    channel_first: bool,
    gap: usize,
) -> TensorHandle {
    let order = if channel_first {
        [0, 3, 1, 2]
    } else {
        [0, 1, 2, 3]
    };
    let mut strides = [0; 4];
    let mut size = gap;
    for &axis in order.iter().rev() {
        strides[axis] = size;
        size *= shape[axis];
    }
    let mut data = vec![99.0f32; size];
    for (i, &value) in values.iter().enumerate() {
        let mut rem = i;
        let mut offset = 0;
        for axis in (0..4).rev() {
            offset += (rem % shape[axis]) * strides[axis];
            rem /= shape[axis];
        }
        data[offset] = value;
    }
    TensorHandle::new(
        client.create_from_slice(f32::as_bytes(&data)),
        shape,
        strides,
        f32::elem_type_native(),
    )
}

#[test]
fn depthwise_weight_split_matches_reference() {
    let client = cubecl::test_device().client();
    check_depthwise_weight_reference(&client);
}

pub fn check_depthwise_weight_reference(client: &Client) {
    if client.properties().hardware.max_units_per_cube < 128 {
        return;
    }
    // shape NHWC, kernel, stride, dilation, beginning/end padding, multiplier.
    for (shape, kernel, stride, dilation, padding, end_padding, multiplier) in [
        ([1, 1, 1, 1], [1, 1], [1, 1], [1, 1], [0, 0], [0, 0], 1),
        ([2, 7, 9, 5], [3, 3], [1, 1], [1, 1], [1, 1], [1, 1], 1),
        ([3, 9, 11, 3], [2, 3], [2, 1], [2, 1], [0, 2], [1, 0], 2),
        ([2, 5, 7, 4], [3, 2], [2, 2], [2, 2], [3, 0], [0, 2], 1),
        // Exactly one full split, followed below by a partial second split.
        ([2, 32, 64, 3], [3, 3], [1, 1], [1, 1], [1, 1], [1, 1], 1),
        // More than one 4096-sample split, with a partial final split.
        ([4, 33, 35, 7], [3, 3], [1, 1], [1, 1], [1, 1], [1, 1], 1),
        // NHWC channel tiles: partial channel and spatial tiles, then a
        // multiplier with stride and dilation. Other layouts take scalar tiles.
        ([2, 17, 19, 35], [3, 3], [1, 1], [1, 1], [1, 1], [1, 1], 1),
        ([2, 19, 21, 33], [2, 3], [2, 1], [2, 2], [1, 0], [0, 2], 2),
    ] {
        let [batch, height, width, channels] = shape;
        let [kh, kw] = kernel;
        let oh =
            (height + padding[0] + end_padding[0] - dilation[0] * (kh - 1) - 1) / stride[0] + 1;
        let ow = (width + padding[1] + end_padding[1] - dilation[1] * (kw - 1) - 1) / stride[1] + 1;
        let out_channels = channels * multiplier;
        let grad_shape = [batch, oh, ow, out_channels];
        let values = (0..shape.iter().product())
            .map(|i| ((i * 7 % 17) as f32 - 8.0) / 8.0)
            .collect::<Vec<_>>();
        let gradients = (0..grad_shape.iter().product())
            .map(|i| ((i * 3 % 13) as f32 - 6.0) / 16.0)
            .collect::<Vec<_>>();
        let mut expected = vec![0.0f32; out_channels * kh * kw];
        for n in 0..batch {
            for y in 0..oh {
                for x in 0..ow {
                    for oc in 0..out_channels {
                        let dy = gradients[((n * oh + y) * ow + x) * out_channels + oc];
                        for ky in 0..kh {
                            for kx in 0..kw {
                                let iy = (y * stride[0] + ky * dilation[0]) as isize
                                    - padding[0] as isize;
                                let ix = (x * stride[1] + kx * dilation[1]) as isize
                                    - padding[1] as isize;
                                if iy >= 0 && ix >= 0 && iy < height as isize && ix < width as isize
                                {
                                    let index = ((n * height + iy as usize) * width + ix as usize)
                                        * channels
                                        + oc / multiplier;
                                    expected[(oc * kh + ky) * kw + kx] += values[index] * dy;
                                }
                            }
                        }
                    }
                }
            }
        }
        for (input_nchw, grad_nchw, gap) in [(false, false, 1), (true, true, 1), (false, true, 2)] {
            let input = upload(client, shape, &values, input_nchw, gap);
            let out_grad = upload(client, grad_shape, &gradients, grad_nchw, gap);
            // The mixed-layout case also writes a permuted, gapped output.
            let weight_strides = if gap == 1 {
                [kh * kw, kw, 1, 1]
            } else {
                [gap, kw * out_channels * gap, out_channels * gap, 1]
            };
            let mut expected_storage = vec![99.0f32; expected.len() * gap];
            for oc in 0..out_channels {
                for ky in 0..kh {
                    for kx in 0..kw {
                        let offset = oc * weight_strides[0]
                            + ky * weight_strides[1]
                            + kx * weight_strides[2];
                        expected_storage[offset] = expected[(oc * kh + ky) * kw + kx];
                    }
                }
            }
            let weight_grad = TensorHandle::new(
                client.create_from_slice(f32::as_bytes(&vec![99.0; expected_storage.len()])),
                [out_channels, kh, kw, 1],
                weight_strides,
                f32::elem_type_native(),
            );
            launch_depthwise_weight_backward(
                client,
                DepthwiseWeightBackwardTensors {
                    input: input.binding(),
                    out_grad: out_grad.binding(),
                    weight_grad: weight_grad.clone().binding(),
                },
                ConvolutionArgs {
                    stride,
                    padding,
                    dilation,
                },
                channels,
                f32::elem_type_native(),
            )
            .unwrap();
            let bytes = client.read_one_unchecked(weight_grad.handle);
            assert_eq!(
                f32::from_bytes(&bytes),
                &expected_storage,
                "shape={shape:?}, multiplier={multiplier}, input_nchw={input_nchw}, grad_nchw={grad_nchw}, gap={gap}"
            );
        }
    }
}

#[test]
fn depthwise_weight_declines_dense_and_other_dtypes() {
    let client = cubecl::test_device().client();
    check_depthwise_weight_declines(&client);
}

pub fn check_depthwise_weight_declines(client: &Client) {
    for (groups, dtype) in [
        (1, f32::elem_type_native()),
        (4, half::f16::elem_type_native()),
    ] {
        let input = TensorHandle::empty(client, [1, 5, 5, 4], dtype);
        let out_grad = TensorHandle::empty(client, [1, 3, 3, 4], dtype);
        let weight_grad = TensorHandle::empty(client, [4, 3, 3, 1], dtype);
        assert!(
            launch_depthwise_weight_backward(
                client,
                DepthwiseWeightBackwardTensors {
                    input: input.binding(),
                    out_grad: out_grad.binding(),
                    weight_grad: weight_grad.binding(),
                },
                ConvolutionArgs {
                    stride: [1, 1],
                    padding: [0, 0],
                    dilation: [1, 1]
                },
                groups,
                dtype,
            )
            .is_err()
        );
    }
}