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
193
194

use crate::{
    compute::StaticKernel,
    element::WgpuElement,
    kernel::{
        self, elemwise_workgroup,
        pool::{build_output_and_info_pool2d, build_pool2d_info},
        KernelSettings, WORKGROUP_DEFAULT,
    },
    kernel_wgsl,
    ops::numeric::empty_device,
    tensor::WgpuTensor,
};

kernel_wgsl!(MaxPool2d, "../../template/pool/max_pool2d.wgsl");
kernel_wgsl!(
    MaxPool2dWithIndicesBackward,
    "../../template/pool/max_pool2d_with_indices_backward.wgsl"
);
kernel_wgsl!(
    MaxPool2dWithIndices,
    "../../template/pool/max_pool2d_with_indices.wgsl"
);

pub(crate) fn max_pool2d<E: WgpuElement>(
    x: WgpuTensor<E, 4>,
    kernel_size: [usize; 2],
    stride: [usize; 2],
    padding: [usize; 2],
    dilation: [usize; 2],
) -> WgpuTensor<E, 4> {
    let (info_handle, output) =
        build_output_and_info_pool2d(&x, kernel_size, stride, padding, dilation);
    let kernel = StaticKernel::<
        KernelSettings<MaxPool2d, E, i32, WORKGROUP_DEFAULT, WORKGROUP_DEFAULT, 1>,
    >::new(elemwise_workgroup(
        output.shape.num_elements(),
        WORKGROUP_DEFAULT,
    ));

    x.client
        .execute(Box::new(kernel), &[&x.handle, &output.handle, &info_handle]);

    output
}

pub(crate) fn max_pool2d_with_indices<E: WgpuElement, I: WgpuElement>(
    x: WgpuTensor<E, 4>,
    kernel_size: [usize; 2],
    stride: [usize; 2],
    padding: [usize; 2],
    dilation: [usize; 2],
) -> (WgpuTensor<E, 4>, WgpuTensor<I, 4>) {
    let (info_handle, output) =
        build_output_and_info_pool2d(&x, kernel_size, stride, padding, dilation);
    let indices = empty_device(x.client.clone(), x.device, output.shape.clone());

    let kernel = StaticKernel::<
        KernelSettings<MaxPool2dWithIndices, E, i32, WORKGROUP_DEFAULT, WORKGROUP_DEFAULT, 1>,
    >::new(elemwise_workgroup(
        output.shape.num_elements(),
        WORKGROUP_DEFAULT,
    ));

    x.client.execute(
        Box::new(kernel),
        &[&x.handle, &output.handle, &indices.handle, &info_handle],
    );

    (output, indices)
}

pub(crate) fn max_pool2d_with_indices_backward<E: WgpuElement, I: WgpuElement>(
    x: WgpuTensor<E, 4>,
    grad: WgpuTensor<E, 4>,
    indices: WgpuTensor<I, 4>,
    kernel_size: [usize; 2],
    stride: [usize; 2],
    padding: [usize; 2],
    dilation: [usize; 2],
) -> WgpuTensor<E, 4> {
    let grad = kernel::into_contiguous(grad);
    let indices = kernel::into_contiguous(indices);

    let num_elems = x.shape.num_elements();
    let buffer = x.client.empty(num_elems * core::mem::size_of::<E>());
    let output = WgpuTensor::new(x.client.clone(), x.device.clone(), x.shape.clone(), buffer);

    let info_handle = build_pool2d_info(&x, &grad, kernel_size, stride, padding, dilation);

    let kernel = StaticKernel::<
        KernelSettings<MaxPool2dWithIndicesBackward, E, I, WORKGROUP_DEFAULT, WORKGROUP_DEFAULT, 1>,
    >::new(elemwise_workgroup(
        output.shape.num_elements(),
        WORKGROUP_DEFAULT,
    ));

    x.client.execute(
        Box::new(kernel),
        &[&indices.handle, &grad.handle, &output.handle, &info_handle],
    );
    output
}

#[cfg(test)]
mod tests {
    use crate::tests::{ReferenceBackend, TestBackend};
    use burn_tensor::{module, ops::ModuleOps, Distribution, Tensor};

    #[test]
    pub fn max_pool2d_should_work_with_multiple_invocations() {
        let tensor = Tensor::<TestBackend, 4>::random([32, 32, 32, 32], Distribution::Default);
        let tensor_ref = Tensor::<ReferenceBackend, 4>::from_data(tensor.to_data());
        let kernel_size = [3, 3];
        let stride = [2, 2];
        let padding = [1, 1];
        let dilation = [1, 1];

        let pooled = module::max_pool2d(tensor, kernel_size, stride, padding, dilation);
        let pooled_ref = module::max_pool2d(tensor_ref, kernel_size, stride, padding, dilation);

        pooled
            .into_data()
            .assert_approx_eq(&pooled_ref.into_data(), 3);
    }

    #[test]
    pub fn max_pool2d_with_indices_should_work_with_multiple_invocations() {
        let tensor = Tensor::<TestBackend, 4>::random([32, 32, 32, 32], Distribution::Default);
        let tensor_ref = Tensor::<ReferenceBackend, 4>::from_data(tensor.to_data());
        let kernel_size = [3, 3];
        let stride = [2, 2];
        let padding = [1, 1];
        let dilation = [1, 1];

        let (pooled, indices) =
            module::max_pool2d_with_indices(tensor, kernel_size, stride, padding, dilation);
        let (pooled_ref, indices_ref) =
            module::max_pool2d_with_indices(tensor_ref, kernel_size, stride, padding, dilation);

        pooled
            .into_data()
            .assert_approx_eq(&pooled_ref.into_data(), 3);
        assert_eq!(indices.into_data(), indices_ref.into_data().convert());
    }

    #[test]
    pub fn max_pool2d_with_indices_backward_should_work_with_multiple_invocations() {
        let tensor = Tensor::<TestBackend, 4>::random([32, 32, 32, 32], Distribution::Default);
        let grad_output = Tensor::<TestBackend, 4>::random([32, 32, 16, 16], Distribution::Default);
        let tensor_ref = Tensor::<ReferenceBackend, 4>::from_data(tensor.to_data());
        let grad_output_ref = Tensor::<ReferenceBackend, 4>::from_data(grad_output.to_data());
        let kernel_size = [3, 3];
        let stride = [2, 2];
        let padding = [1, 1];
        let dilation = [1, 1];

        let (_, indices) =
            module::max_pool2d_with_indices(tensor.clone(), kernel_size, stride, padding, dilation);
        let (_, indices_ref) = module::max_pool2d_with_indices(
            tensor_ref.clone(),
            kernel_size,
            stride,
            padding,
            dilation,
        );
        let grad = TestBackend::max_pool2d_with_indices_backward(
            tensor.into_primitive(),
            kernel_size,
            stride,
            padding,
            dilation,
            grad_output.into_primitive(),
            indices.into_primitive(),
        )
        .x_grad;
        let grad_ref = ReferenceBackend::max_pool2d_with_indices_backward(
            tensor_ref.into_primitive(),
            kernel_size,
            stride,
            padding,
            dilation,
            grad_output_ref.into_primitive(),
            indices_ref.into_primitive(),
        )
        .x_grad;

        Tensor::<TestBackend, 4>::from_primitive(grad)
            .into_data()
            .assert_approx_eq(
                &Tensor::<ReferenceBackend, 4>::from_primitive(grad_ref).into_data(),
                3,
            );
    }
}