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

#[cfg(feature = "opencl")]
use crate::{
    cl_diagflat,
    ops::{cl_to_cpu_lr, cl_to_cpu_s},
};
use crate::{ColOp, DiagflatOp, FnsOps, Matrix, MaxOps, SumOps, TransposeOp};
use custos::{get_device, number::Float, range, GenericBlas, CPU};
#[cfg(feature = "opencl")]
use custos::{CDatatype, CLDevice};

#[cfg(feature = "cuda")]
use crate::{cu_to_cpu_lr, cu_to_cpu_s};
#[cfg(feature = "cuda")]
use custos::CudaDevice;

impl<'a, T: GenericBlas> Matrix<'a, T> {
    pub fn softmax(&self) -> Matrix<'a, T> {
        let device = get_device!(self.device(), SoftmaxOps<T>);
        device.softmax(self)
    }
    pub fn softmax_grad(&self, activated: &Matrix<T>) -> Matrix<'a, T> {
        let device = get_device!(self.device(), SoftmaxOps<T>);
        device.softmax_grad(activated, self)
    }
}

pub trait SoftmaxOps<T> {
    fn softmax(&self, inputs: &Matrix<T>) -> Matrix<T>;
    fn softmax_grad(&self, activated: &Matrix<T>, grads: &Matrix<T>) -> Matrix<T>;
}

impl<T: Float + GenericBlas> SoftmaxOps<T> for CPU {
    fn softmax(&self, inputs: &Matrix<T>) -> Matrix<T> {
        let exp = self.exp(&self.sub_col(inputs, &self.max_cols(inputs)));
        self.div_col(&exp, &self.sum_cols(&exp))
    }

    #[cfg(not(feature = "safe"))]
    fn softmax_grad(&self, activated: &Matrix<T>, grads: &Matrix<T>) -> Matrix<T> {
        use custos::Cache;

        use crate::{BaseOps, Gemm};

        let mut data: Matrix<T> = (Cache::get(self, grads.len, ()), grads.dims()).into();

        let rows = grads.rows();
        let cols = grads.cols();

        for idx in range(rows - 1) {
            let index = idx * cols;

            let single_out = Matrix::from((
                (&activated[index..index + cols]).as_ptr() as *mut T,
                (cols, 1),
            ));
            let single_grad =
                Matrix::from(((&grads[index..index + cols]).as_ptr() as *mut T, (cols, 1)));

            let diagflat = self.diagflat(&single_out);

            // cols 1 x 1 cols
            let jacobian_matrix = self.sub(
                &diagflat,
                &self.gemm(&single_out, &self.transpose(&single_out)),
            );

            let res = self.gemm(&jacobian_matrix, &single_grad);

            let data_row = &mut data[index..index + cols];
            data_row.copy_from_slice(&res);
        }
        data
    }

    #[cfg(feature = "safe")]
    fn softmax_grad(&self, activated: &Matrix<T>, grads: &Matrix<T>) -> Matrix<T> {
        use crate::{BaseOps, Gemm};

        let device = CPU::new();
        let mut data = cached(self, grads.dims());

        let rows = grads.rows();
        let cols = grads.cols();

        for idx in range(rows - 1) {
            let index = idx * cols;

            let single_out =
                Matrix::from((&device, (cols, 1), &activated[index..index + cols].to_vec()));

            let single_grad =
                Matrix::from((&device, (cols, 1), &grads[index..index + cols].to_vec()));

            let diagflat = self.diagflat(&single_out);

            let jacobian_matrix = self.sub(
                &diagflat,
                &self.gemm(&single_out, &self.transpose(&single_out)),
            );
            //cols cols x cols 1
            let res = self.gemm(&jacobian_matrix, &single_grad);

            let data_row = &mut data[index..index + cols];
            data_row.copy_from_slice(res.as_slice());
        }
        data
    }
}

#[cfg(feature = "cuda")]
impl<T: Default + Copy + GenericBlas> SoftmaxOps<T> for CudaDevice {
    fn softmax(&self, inputs: &Matrix<T>) -> Matrix<T> {
        cu_to_cpu_s(self, inputs, |cpu, x| cpu.softmax(&x))
    }

    fn softmax_grad(&self, activated: &Matrix<T>, grads: &Matrix<T>) -> Matrix<T> {
        cu_to_cpu_lr(self, activated, grads, |cpu, activated, grads| {
            cpu.softmax_grad(activated, grads)
        })
    }
}

#[cfg(feature = "opencl")]
// TODO: Softmax running on the opencl device
impl<T: GenericBlas + Float> SoftmaxOps<T> for CLDevice {
    fn softmax(&self, inputs: &Matrix<T>) -> Matrix<T> {
        cl_to_cpu_s(self, inputs, |device, inputs| device.softmax(inputs))
    }

    fn softmax_grad(&self, activated: &Matrix<T>, grads: &Matrix<T>) -> Matrix<T> {
        cl_to_cpu_lr(self, activated, grads, |device, activated, grads| {
            device.softmax_grad(activated, grads)
        })
    }
}

#[cfg(feature = "opencl")]
pub fn cl_softmax<'a, T: CDatatype>(
    device: &'a CLDevice,
    mut activated: Matrix<T>,
    grads: &Matrix<T>,
) -> custos::Result<Matrix<'a, T>> {
    use crate::{cl_tew, Gemm};

    let rows = grads.rows();
    let cols = grads.cols();

    let diag = cl_diagflat(device, &activated, activated.rows(), activated.cols())?;

    //println!("diag: {diag:?}");
    activated.reshape((cols, rows));

    //cols rows x rows cols

    let jacobian = cl_tew(
        device,
        &diag,
        &device.gemm(&activated, &device.transpose(&activated)),
        "-",
    )?;

    println!("jacobian: {jacobian:?}");

    let jacobian = (jacobian, rows, cols * cols).into();
    let mut jacobian = device.sum_rows(&jacobian);
    jacobian.reshape((cols, cols));

    // rows cols x cols cols
    let res = device.gemm(grads, &jacobian);
    Ok(res)
}