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

use crate::{
    dim::{DimDyn, DimTrait},
    index::Index0D,
    matrix::{AsMutPtr, AsPtr, IndexAxisDyn, IndexAxisMutDyn, MatrixBase, ToViewMutMatrix},
    matrix_impl::Matrix,
    memory_impl::{ViewMem, ViewMutMem},
    num::Num,
};

pub trait Relu<T: Num> {
    fn relu(&mut self, source: Matrix<ViewMem<T>, DimDyn>);
    fn relu_backward_mask(&mut self, source: Matrix<ViewMem<T>, DimDyn>);
}

impl<'a, T: Num> Relu<T> for Matrix<ViewMutMem<'a, T>, DimDyn> {
    fn relu(&mut self, source: Matrix<ViewMem<T>, DimDyn>) {
        if self.shape() != source.shape() {
            panic!("shape mismatch");
        }

        let len = self.shape().len();
        if len == 0 {
            unsafe {
                *self.as_mut_ptr() = if *source.as_ptr() > T::zero() {
                    *source.as_ptr()
                } else {
                    T::zero()
                };
            };
        } else if len == 1 {
            let num_elm = self.shape().num_elm();
            let stride_self = self.stride()[0];
            let stride_source = source.stride()[0];
            let self_ptr = self.as_mut_slice();
            let source_ptr = source.as_slice();
            relu_kernel_cpu(source_ptr, self_ptr, num_elm, stride_source, stride_self);
        } else {
            for i in 0..self.shape()[0] {
                self.to_view_mut()
                    .index_axis_mut_dyn(Index0D::new(i))
                    .relu(source.index_axis_dyn(Index0D::new(i)).clone());
            }
        }
    }

    fn relu_backward_mask(&mut self, source: Matrix<ViewMem<T>, DimDyn>) {
        if self.shape() != source.shape() {
            panic!("shape mismatch");
        }

        let len = self.shape().len();
        if len == 0 {
            unsafe {
                *self.as_mut_ptr() = if *source.as_ptr() > T::zero() {
                    T::one()
                } else {
                    T::zero()
                };
            };
        } else if len <= 1 {
            let num_elm = self.shape().num_elm();
            let stride_self = self.stride()[0];
            let stride_source = source.stride()[0];
            let self_ptr = self.as_mut_slice();
            let source_ptr = source.as_slice();
            relu_backward_mask_kernel_cpu(
                source_ptr,
                self_ptr,
                num_elm,
                stride_source,
                stride_self,
            );
        } else {
            for i in 0..self.shape()[0] {
                self.to_view_mut()
                    .index_axis_mut_dyn(Index0D::new(i))
                    .relu_backward_mask(source.index_axis_dyn(Index0D::new(i)).clone());
            }
        }
    }
}

fn relu_kernel_cpu<T: Num>(x: &[T], y: &mut [T], len: usize, incx: usize, incy: usize) {
    if incx == 1 && incy == 1 {
        for i in 0..len {
            y[i] = if x[i] > T::zero() { x[i] } else { T::zero() };
        }
    } else {
        for i in 0..len {
            y[i * incy] = if x[i * incx] > T::zero() {
                x[i * incx]
            } else {
                T::zero()
            };
        }
    }
}

fn relu_backward_mask_kernel_cpu<T: Num>(
    x: &[T],
    mask: &mut [T],
    len: usize,
    incx: usize,
    incmask: usize,
) {
    if incx == 1 && incmask == 1 {
        for i in 0..len {
            mask[i] = if x[i] > T::zero() {
                T::one()
            } else {
                T::zero()
            };
        }
    } else {
        for i in 0..len {
            mask[i * incmask] = if x[i * incx] > T::zero() {
                T::one()
            } else {
                T::zero()
            };
        }
    }
}

#[cfg(test)]
mod relu {
    use crate::{
        matrix::{OwnedMatrix, ToViewMatrix, ToViewMutMatrix},
        matrix_impl::OwnedMatrixDyn,
        operation::{asum::Asum, zeros::Zeros},
    };

    use super::Relu;

    #[test]
    fn relu() {
        let x = OwnedMatrixDyn::from_vec(vec![1.0, -1.0, 0.0, 2.0], [2, 2]);
        let mut y = OwnedMatrixDyn::zeros([2, 2]);
        y.to_view_mut().relu(x.to_view());
        let ans = OwnedMatrixDyn::from_vec(vec![1.0, 0.0, 0.0, 2.0], [2, 2]);
        let diff = y.to_view() - ans.to_view();
        let diff_asum = diff.asum();
        assert!(diff_asum < 1.0e-6);
    }

    #[test]
    fn relu_backward_mask() {
        let x = OwnedMatrixDyn::from_vec(vec![1.0, -1.0, 0.0, 2.0], [2, 2]);
        let mut y = OwnedMatrixDyn::zeros([2, 2]);
        y.to_view_mut().relu_backward_mask(x.to_view());
        let ans = OwnedMatrixDyn::from_vec(vec![1.0, 0.0, 0.0, 1.0], [2, 2]);
        let diff = y.to_view() - ans.to_view();
        let diff_asum = diff.asum();
        assert!(diff_asum < 1.0e-6);
    }

    #[test]
    fn relu_0d() {
        let x = OwnedMatrixDyn::from_vec(vec![1.0], []);
        let mut y = OwnedMatrixDyn::zeros([]);
        y.to_view_mut().relu(x.to_view());
        let ans = OwnedMatrixDyn::from_vec(vec![1.0], []);
        let diff = y.to_view() - ans.to_view();
        let diff_asum = diff.asum();
        assert!(diff_asum < 1.0e-6);
    }

    #[test]
    fn relu_backward_mask_0d() {
        let x = OwnedMatrixDyn::from_vec(vec![1.0], []);
        let mut y = OwnedMatrixDyn::zeros([]);
        y.to_view_mut().relu_backward_mask(x.to_view());
        let ans = OwnedMatrixDyn::from_vec(vec![1.0], []);
        let diff = y.to_view() - ans.to_view();
        let diff_asum = diff.asum();
        assert!(diff_asum < 1.0e-6);
    }
}