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

use std::fmt::Debug;

use ndarray::{
    Array, ArrayBase, ArrayView, Data, Dim, Dimension, IntoDimension, Ix, RawData, RemoveAxis,
    SliceArg, SliceInfo, SliceInfoElem,
};
use num::traits::NumAssign;

use crate::{
    dilation::{IntoKernelWithDilation, KernelWithDilation},
    padding::PaddingExt,
    ConvMode, PaddingMode,
};

#[cfg(test)]
mod tests;

pub struct ExplicitConv<const N: usize> {
    pub padding: [[usize; 2]; N],
    pub strides: [usize; N],
}

impl<const N: usize> ConvMode<N> {
    pub(crate) fn unfold<S>(self, kernel: &KernelWithDilation<S, N>) -> ExplicitConv<N>
    where
        S: ndarray::RawData,
        Dim<[Ix; N]>: Dimension,
    {
        let kernel_dim = kernel.kernel.raw_dim();
        let kernel_dim: [usize; N] = std::array::from_fn(|i|
                // k + (k - 1) * (d - 1)
                kernel_dim[i] * kernel.dilation[i] - kernel.dilation[i] + 1);

        match self {
            ConvMode::Full => ExplicitConv {
                padding: std::array::from_fn(|i| [kernel_dim[i] - 1; 2]),
                strides: [1; N],
            },
            ConvMode::Same => ExplicitConv {
                padding: std::array::from_fn(|i| {
                    let k_size = kernel_dim[i];
                    if k_size % 2 == 0 {
                        [(k_size - 1) / 2 + 1, (k_size - 1) / 2]
                    } else {
                        [(k_size - 1) / 2; 2]
                    }
                }),
                strides: [1; N],
            },
            ConvMode::Valid => ExplicitConv {
                padding: [[0; 2]; N],
                strides: [1; N],
            },
            ConvMode::Custom { padding, strides } => ExplicitConv {
                padding: padding.map(|pad| [pad; 2]),
                strides,
            },
            ConvMode::Explicit { padding, strides } => ExplicitConv { padding, strides },
        }
    }
}

pub trait ConvExt<'a, T, S, SK, const N: usize>
where
    T: NumAssign + Copy,
    S: RawData,
    SK: RawData,
{
    fn conv(
        &self,
        kernel: impl IntoKernelWithDilation<'a, SK, N>,
        conv_mode: ConvMode<N>,
        padding_mode: PaddingMode<N, T>,
    ) -> Result<Array<T, Dim<[Ix; N]>>, crate::Error<N>>;
}

impl<'a, T, S, SK, const N: usize> ConvExt<'a, T, S, SK, N> for ArrayBase<S, Dim<[Ix; N]>>
where
    T: NumAssign + Copy + Debug,
    S: Data<Elem = T> + 'a,
    SK: Data<Elem = T> + 'a,
    Dim<[Ix; N]>: RemoveAxis,
    [Ix; N]: IntoDimension<Dim = Dim<[Ix; N]>>,
    SliceInfo<[SliceInfoElem; N], Dim<[Ix; N]>, Dim<[Ix; N]>>:
        SliceArg<Dim<[Ix; N]>, OutDim = Dim<[Ix; N]>>,
{
    fn conv(
        &self,
        kernel: impl IntoKernelWithDilation<'a, SK, N>,
        conv_mode: ConvMode<N>,
        padding_mode: PaddingMode<N, T>,
    ) -> Result<Array<T, Dim<[Ix; N]>>, crate::Error<N>> {
        let kwd = kernel.into_kernel_with_dilation();

        let self_raw_dim = self.raw_dim();
        if self.shape().iter().product::<usize>() == 0 {
            return Err(crate::Error::DataShape(self_raw_dim));
        }

        let kernel_raw_dim = kwd.kernel.raw_dim();
        if kwd.kernel.shape().iter().product::<usize>() == 0 {
            return Err(crate::Error::DataShape(kernel_raw_dim));
        }

        let kernel_raw_dim_with_dilation: [usize; N] =
            std::array::from_fn(|i| kernel_raw_dim[i] * kwd.dilation[i] - kwd.dilation[i] + 1);

        let cm = conv_mode.unfold(&kwd);
        let pds = self.padding(padding_mode, cm.padding);

        let pds_raw_dim = pds.raw_dim();
        if !(0..N).all(|i| kernel_raw_dim_with_dilation[i] <= pds_raw_dim[i]) {
            return Err(crate::Error::MismatchShape(
                conv_mode,
                kernel_raw_dim_with_dilation,
            ));
        }

        let offset_list = kwd.gen_offset_list(pds.strides());

        let output_shape: [usize; N] = std::array::from_fn(|i| {
            (cm.padding[i][0] + cm.padding[i][1] + self_raw_dim[i]
                - kernel_raw_dim_with_dilation[i])
                / cm.strides[i]
                + 1
        });
        let mut ret = Array::zeros(output_shape);

        let shape: [usize; N] = std::array::from_fn(|i| ret.raw_dim()[i]);
        let strides: [usize; N] =
            std::array::from_fn(|i| cm.strides[i] * pds.strides()[i] as usize);

        // dbg!(&offset_list);
        // dbg!(strides);

        unsafe {
            // use raw pointer to improve performance.
            let p: *mut T = ret.as_mut_ptr();

            // use ArrayView's iter without handle strides
            let view = ArrayView::from_shape(
                ndarray::ShapeBuilder::strides(shape, strides),
                pds.as_slice().unwrap(),
            )
            .unwrap();

            view.iter().enumerate().for_each(|(i, cur)| {
                let mut tmp_res = T::zero();

                offset_list.iter().for_each(|(tmp_offset, tmp_kernel)| {
                    tmp_res += *(cur as *const T).offset(*tmp_offset) * *tmp_kernel
                });

                *p.add(i) = tmp_res;
            });
        }

        Ok(ret)
    }
}