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

use crate::{shapes::*, tensor::*};

mod cpu_kernel;
#[cfg(feature = "cuda")]
mod cuda_kernel;

pub trait SliceKernel<E: Unit>: Storage<E> {
    fn forward<Src: Shape + SliceShape<Slice>, Slice>(
        &self,
        inp: &Tensor<Src, E, Self>,
        slice: &Slice,
    ) -> Result<Tensor<Src::Sliced, E, Self>, Self::Err>;

    fn backward<Src: Shape + SliceShape<Slice>, Slice>(
        &self,
        inp: &Tensor<Src, E, Self>,
        grad_inp: &mut Self::Vec,
        grad_out: &Self::Vec,
        slice: &Slice,
    ) -> Result<(), Self::Err>;
}

/// Slices all dimensions of a tensor, with the starting and ending indices of each dimension
/// determined by a tuple of ranges.
///
/// Slices are specified as tuples of ranges defined with the `..` and `..=` operators. All
/// sliced dimensions are changed to be of type usize except those sliced with `..`
/// ([std::ops::RangeFull]), whose types are not modified.
///
/// Example:
/// ```rust
/// # use dfdx::prelude::*;
/// # let dev = Cpu::default();
/// let a = dev.tensor([
///     [1., 2.],
///     [3., 4.],
/// ]);
///
/// // Slice the first row to get a 1x2 tensor
/// let b: Tensor<Rank2<1, 2>, _, _> = a.clone().slice((0..1, 0..2)).realize();
/// assert_eq!(b.array(), [[1., 2.]]);
///
/// // Slice the last column to get a 2x1 tensor
/// let c: Tensor<Rank2<2, 1>, _, _> = a.clone().slice((0..2, 1..)).realize();
/// assert_eq!(c.array(), [[2.], [4.]]);
/// ```
pub fn slice<S: SliceShape<Slice>, E: Unit, D: SliceKernel<E>, T: Tape<E, D>, Slice: 'static>(
    tensor: Tensor<S, E, D, T>,
    slice: Slice,
) -> Tensor<S::Sliced, E, D, T> {
    tensor.slice(slice)
}

impl<S: Shape, E: Unit, D: SliceKernel<E>, T: Tape<E, D>> Tensor<S, E, D, T> {
    /// Fallible version of [Tensor::slice]
    pub fn try_slice<Slice>(self, slice: Slice) -> Result<Tensor<S::Sliced, E, D, T>, D::Err>
    where
        S: SliceShape<Slice>,
        Slice: 'static,
    {
        let (inp, mut tape) = self.split_tape();
        let out = inp.device.forward(&inp, &slice)?;
        let inp_ghost = inp.ghost();
        let out_ghost = out.ghost();
        tape.add_backward_op(move |grads| {
            grads.try_alloc_for(&inp_ghost)?;
            grads.try_alloc_for(&out_ghost)?;
            let (grad_inp, grad_out) = grads.mut_and_ref(&inp_ghost, &out_ghost);
            inp.device.backward(&inp, grad_inp, grad_out, &slice)
        });
        Ok(out.put_tape(tape))
    }

    /// Calls [slice()].
    pub fn slice<Slice>(self, slice: Slice) -> Tensor<S::Sliced, E, D, T>
    where
        S: SliceShape<Slice>,
        Slice: 'static,
    {
        self.try_slice(slice).unwrap()
    }
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::{tensor_ops::*, tests::*};

    #[test]
    fn test_slice() {
        let dev = TestDevice::default();
        let a = dev
            .tensor([
                [1., 2., 3., 4.],
                [5., 6., 7., 8.],
                [9., 10., 11., 12.],
                [13., 14., 15., 16.],
            ])
            .to_dtype::<TestDtype>();

        let b: Tensor<Rank2<2, 2>, _, _> = a.clone().slice((2.., 2..)).realize();
        assert_close_to_literal!(b, [[11., 12.], [15., 16.]]);

        let b: Tensor<Rank2<2, 2>, _, _> = a.clone().slice((1..3, 1..3)).realize();
        assert_close_to_literal!(b, [[6., 7.], [10., 11.]]);

        let b: Tensor<Rank2<1, 3>, _, _> = a.clone().slice((..1, 1..4)).realize();
        assert_close_to_literal!(b, [[2., 3., 4.]]);

        let b: Tensor<Rank2<2, 3>, _, _> = a.clone().slice((1..3, ..3)).realize();
        assert_close_to_literal!(b, [[5., 6., 7.], [9., 10., 11.]]);

        let b: Tensor<Rank2<2, 3>, _, _> = a.clone().slice((1..=2, 1..=3)).realize();
        assert_close_to_literal!(b, [[6., 7., 8.], [10., 11., 12.]]);

        let b: Tensor<Rank2<2, 2>, _, _> = a.clone().slice((0..=1, 2..=3)).realize();
        assert_close_to_literal!(b, [[3., 4.], [7., 8.]]);

        let b: Tensor<Rank2<3, 2>, _, _> = a.clone().slice((1.., ..2)).realize();
        assert_close_to_literal!(b, [[5., 6.], [9., 10.], [13., 14.]]);

        let b: Tensor<Rank2<2, 2>, _, _> = a.slice((..2, 2..)).realize();
        assert_close_to_literal!(b, [[3., 4.], [7., 8.]]);
    }

    #[test]
    fn test_slice_broadcast_top() {
        let dev = TestDevice::default();
        let a = dev
            .tensor([1., 2., 3., 4.])
            .to_dtype::<TestDtype>()
            .broadcast::<Rank2<5, 4>, _>();

        let b: Tensor<Rank2<3, 4>, _, _> = a.clone().slice((..3, ..)).realize();
        assert_close_to_literal!(b, [[1., 2., 3., 4.]; 3]);

        let b: Tensor<Rank2<5, 2>, _, _> = a.clone().slice((.., 1..3)).realize();
        assert_close_to_literal!(b, [[2., 3.]; 5]);

        let b: Tensor<Rank2<2, 2>, _, _> = a.clone().slice((1..3, 1..3)).realize();
        assert_close_to_literal!(b, [[2., 3.], [2., 3.]]);

        let b: Tensor<Rank2<3, 3>, _, _> = a.slice((1..4, 1..4)).realize();
        assert_close_to_literal!(b, [[2., 3., 4.]; 3]);
    }

    #[test]
    fn test_slice_broadcast_bottom() {
        let dev = TestDevice::default();
        let a: Tensor<Rank2<4, 5>, TestDtype, _> = dev
            .tensor([1., 2., 3., 4.])
            .to_dtype::<TestDtype>()
            .broadcast();

        let b: Tensor<Rank2<2, 5>, _, _> = a.clone().slice((1..3, ..)).realize();
        assert_close_to_literal!(b, [[2.; 5], [3.; 5]]);

        let b: Tensor<Rank2<4, 2>, _, _> = a.clone().slice((.., 1..3)).realize();
        assert_close_to_literal!(b, [[1., 1.], [2., 2.], [3., 3.], [4., 4.]]);

        let b: Tensor<Rank2<2, 2>, _, _> = a.clone().slice((1..3, 3..)).realize();
        assert_close_to_literal!(b, [[2., 2.], [3., 3.]]);

        let b: Tensor<Rank2<2, 2>, _, _> = a.slice((..2, 1..3)).realize();
        assert_close_to_literal!(b, [[1., 1.], [2., 2.]]);
    }

    #[test]
    fn test_slice_backward() {
        let dev = TestDevice::default();
        let a = dev
            .tensor([
                [1., 2., 3., 4.],
                [5., 6., 7., 8.],
                [9., 10., 11., 12.],
                [13., 14., 15., 16.],
            ])
            .to_dtype::<TestDtype>();

        let b: Tensor<Rank2<2, 2>, _, _, _> = a.leaky_trace().slice((2.., 2..)).realize();
        assert_close_to_literal!(b, [[11., 12.], [15., 16.]]);
        let g = b.square().sum().backward();
        assert_close_to_literal!(
            g.get(&a),
            [[0.; 4], [0.; 4], [0., 0., 22., 24.], [0., 0., 30., 32.]]
        );
    }
}