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

#[cfg(feature = "opencl")]
use crate::opencl::cl_str_op_mat;
use crate::{each_op, Matrix};
#[cfg(feature = "opencl")]
use custos::CLDevice;
use custos::{devices::cpu::CPU, get_device, number::Float, CDatatype};

#[cfg(feature = "cuda")]
use crate::cu_str_op;
#[cfg(feature = "cuda")]
use custos::CudaDevice;

impl<'a, T: CDatatype + Float> Matrix<'a, T> {
    #[inline]
    pub fn tanh(&self) -> Matrix<'a, T> {
        let device = get_device!(self.device(), ActivationOps<T>);
        device.tanh(self)
    }

    #[inline]
    pub fn tanh_grad(&self) -> Matrix<'a, T> {
        let device = get_device!(self.device(), ActivationOps<T>);
        device.tanh_grad(self)
    }

    #[inline]
    pub fn relu(&self) -> Matrix<'a, T> {
        let device = get_device!(self.device(), ActivationOps<T>);
        device.relu(self)
    }

    #[inline]
    pub fn relu_grad(&self) -> Matrix<'a, T> {
        let device = get_device!(self.device(), ActivationOps<T>);
        device.relu_grad(self)
    }
}

pub trait ActivationOps<T> {
    fn sigmoid(&self, x: &Matrix<T>) -> Matrix<T>;
    fn tanh(&self, x: &Matrix<T>) -> Matrix<T>;
    fn tanh_grad(&self, x: &Matrix<T>) -> Matrix<T>;
    fn relu(&self, x: &Matrix<T>) -> Matrix<T>;
    fn relu_grad(&self, x: &Matrix<T>) -> Matrix<T>;
}

#[cfg(feature = "opencl")]
impl<T: CDatatype + Float> ActivationOps<T> for CLDevice {
    #[inline]
    fn sigmoid(&self, x: &Matrix<T>) -> Matrix<T> {
        cl_str_op_mat(self, x, "1.0 / (1.0 + exp(-x))").unwrap()
    }

    #[inline]
    fn tanh(&self, x: &Matrix<T>) -> Matrix<T> {
        cl_str_op_mat(self, x, "tanh(x)").unwrap()
    }

    #[inline]
    fn tanh_grad(&self, x: &Matrix<T>) -> Matrix<T> {
        cl_str_op_mat(self, x, "1.0 - pow(tanh(x), 2)").unwrap()
    }

    #[inline]
    fn relu(&self, x: &Matrix<T>) -> Matrix<T> {
        cl_str_op_mat(self, x, "x * (x >= 0)").unwrap()
    }

    #[inline]
    fn relu_grad(&self, x: &Matrix<T>) -> Matrix<T> {
        cl_str_op_mat(self, x, "(x >= 0)").unwrap()
    }
}

impl<T: Float> ActivationOps<T> for CPU {
    #[inline]
    fn sigmoid(&self, x: &Matrix<T>) -> Matrix<T> {
        each_op(self, x, |x| T::one() / (T::one() + x.negate().exp()))
    }

    #[inline]
    fn tanh(&self, x: &Matrix<T>) -> Matrix<T> {
        each_op(self, x, |x| x.tanh())
    }

    #[inline]
    fn tanh_grad(&self, x: &Matrix<T>) -> Matrix<T> {
        each_op(self, x, |x| T::one() - x.tanh().powi(2))
    }

    #[inline]
    fn relu(&self, x: &Matrix<T>) -> Matrix<T> {
        each_op(self, x, |x| T::from_usize((x >= T::zero()) as usize) * x)
    }

    #[inline]
    fn relu_grad(&self, x: &Matrix<T>) -> Matrix<T> {
        each_op(self, x, |x| T::from_usize((x >= T::default()) as usize))
    }
}

#[cfg(feature = "cuda")]
impl<T: CDatatype> ActivationOps<T> for CudaDevice {
    #[inline]
    fn sigmoid(&self, x: &Matrix<T>) -> Matrix<T> {
        let out = cu_str_op(self, x, "1.0 / (1.0 + exp(-x))").unwrap();
        (out, x.dims()).into()
    }

    #[inline]
    fn tanh(&self, x: &Matrix<T>) -> Matrix<T> {
        let out = cu_str_op(self, x, "tanh(x)").unwrap();
        (out, x.dims()).into()
    }

    #[inline]
    fn tanh_grad(&self, x: &Matrix<T>) -> Matrix<T> {
        let out = cu_str_op(self, x, "1.0 - pow(tanh(x), 2)").unwrap();
        (out, x.dims()).into()
    }

    #[inline]
    fn relu(&self, x: &Matrix<T>) -> Matrix<T> {
        let out = cu_str_op(self, x, "x * (x >= 0)").unwrap();
        (out, x.dims()).into()
    }

    #[inline]
    fn relu_grad(&self, x: &Matrix<T>) -> Matrix<T> {
        let out = cu_str_op(self, x, "(x >= 0)").unwrap();
        (out, x.dims()).into()
    }
}