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

use std::any::Any;

use tracing::Level;

use crate::{Primitive, Shape, Tensor};

#[derive(Clone, Debug, PartialEq, Eq)]
pub struct ReduceSum {
    pub dims: Vec<usize>,
    pub keep_dim: bool,
}

impl ReduceSum {
    pub fn new(dims: impl Into<Vec<usize>>, keep_dim: bool) -> Self {
        Self {
            dims: dims.into(),
            keep_dim,
        }
    }

    pub fn dims(&self) -> &[usize] {
        &self.dims
    }
}

impl Primitive for ReduceSum {
    fn clone_boxed(&self) -> Box<dyn Primitive> {
        Box::new(self.clone())
    }

    fn dot_label(&self) -> String {
        format!("ReduceSum({:?}, {})", &self.dims, &self.keep_dim)
    }

    fn as_any(&self) -> &dyn Any {
        self
    }

    #[tracing::instrument(ret(level = Level::TRACE))]
    fn jvp(&self, _output: &Tensor, _primals: &[Tensor], tangents: &[Tensor]) -> Tensor {
        let tangent_x = &tangents[0];
        tangent_x.sum((self.dims(), false))
    }

    #[tracing::instrument(ret(level = Level::TRACE))]
    fn vjp(&self, output: &Tensor, primals: &[Tensor], cotangent: &Tensor) -> Vec<Tensor> {
        let x = &primals[0];
        let mut shape = x.shape().to_vec();
        for dim in self.dims() {
            shape[*dim] = 1;
        }
        let cotangent_x = cotangent.reshape(&shape).broadcast_to(x);
        vec![cotangent_x]
    }
}

#[derive(Clone, Debug, PartialEq, Eq)]
pub struct ReduceMax {
    pub dims: Vec<usize>,
    pub keep_dim: bool,
}

impl ReduceMax {
    pub fn new(dims: impl Into<Vec<usize>>, keep_dim: bool) -> Self {
        Self {
            dims: dims.into(),
            keep_dim,
        }
    }

    pub fn dims(&self) -> &[usize] {
        &self.dims
    }
}

impl Primitive for ReduceMax {
    fn clone_boxed(&self) -> Box<dyn Primitive> {
        Box::new(self.clone())
    }

    fn dot_label(&self) -> String {
        format!("ReduceMax({:?}, {})", &self.dims, &self.keep_dim)
    }

    fn as_any(&self) -> &dyn Any {
        self
    }

    #[tracing::instrument(ret(level = Level::TRACE))]
    fn jvp(&self, output: &Tensor, primals: &[Tensor], tangents: &[Tensor]) -> Tensor {
        let x = &primals[0];
        let tangent_x = &tangents[0];
        reduce_chooser_jvp_rule(tangent_x, output, x, self.dims())
    }

    #[tracing::instrument(ret(level = Level::TRACE))]
    fn vjp(&self, output: &Tensor, primals: &[Tensor], cotangent: &Tensor) -> Vec<Tensor> {
        let x = &primals[0];
        let mut shape = x.shape().to_vec();
        for dim in self.dims() {
            shape[*dim] = 1;
        }
        let mask = x.eq(output).as_type(x.dtype());
        let normalizer = mask.sum((self.dims(), true));
        let cotangent_x = (cotangent.reshape(shape) / normalizer) * mask;
        vec![cotangent_x]
    }
}

#[derive(Clone, Debug, PartialEq, Eq)]
pub struct ReduceMin {
    pub dims: Vec<usize>,
    pub keep_dim: bool,
}

impl ReduceMin {
    pub fn new(dims: impl Into<Vec<usize>>, keep_dim: bool) -> Self {
        Self {
            dims: dims.into(),
            keep_dim,
        }
    }

    pub fn dims(&self) -> &[usize] {
        &self.dims
    }
}

impl Primitive for ReduceMin {
    fn clone_boxed(&self) -> Box<dyn Primitive> {
        Box::new(self.clone())
    }

    fn dot_label(&self) -> String {
        format!("ReduceMax({:?}, {})", &self.dims, &self.keep_dim)
    }

    fn as_any(&self) -> &dyn Any {
        self
    }

    #[tracing::instrument(ret(level = Level::TRACE))]
    fn jvp(&self, output: &Tensor, primals: &[Tensor], tangents: &[Tensor]) -> Tensor {
        let x = &primals[0];
        let tangent_x = &tangents[0];
        reduce_chooser_jvp_rule(tangent_x, output, x, self.dims())
    }

    #[tracing::instrument(ret(level = Level::TRACE))]
    fn vjp(&self, output: &Tensor, primals: &[Tensor], cotangent: &Tensor) -> Vec<Tensor> {
        let x = &primals[0];
        let mut shape = x.shape().to_vec();
        for dim in self.dims() {
            shape[*dim] = 1;
        }
        let mask = x.eq(output).as_type(x.dtype());
        let normalizer = mask.sum((self.dims(), true));
        let cotangent_x = (cotangent.reshape(shape) / normalizer) * mask;
        vec![cotangent_x]
    }
}

fn reduce_chooser_jvp_rule(g: &Tensor, ans: &Tensor, operand: &Tensor, dims: &[usize]) -> Tensor {
    let mut shape = operand.shape().to_vec();
    for dim in dims {
        shape[*dim] = 1;
    }
    let location_indicators = operand.eq(ans.reshape(shape)).as_type(g.dtype());
    let counts = location_indicators.sum(dims);
    (g * location_indicators).sum(dims) / counts
}