ncnn

1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
4
//
5
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
6
// in compliance with the License. You may obtain a copy of the License at
7
//
8
// https://opensource.org/licenses/BSD-3-Clause
9
//
10
// Unless required by applicable law or agreed to in writing, software distributed
11
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
12
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
13
// specific language governing permissions and limitations under the License.
14

15
#include "fold.h"
16

17
namespace ncnn {
18

19
Fold::Fold()
20
{
21
    one_blob_only = true;
22
}
23

24
int Fold::load_param(const ParamDict& pd)
25
{
26
    kernel_w = pd.get(1, 0);
27
    kernel_h = pd.get(11, kernel_w);
28
    dilation_w = pd.get(2, 1);
29
    dilation_h = pd.get(12, dilation_w);
30
    stride_w = pd.get(3, 1);
31
    stride_h = pd.get(13, stride_w);
32
    pad_left = pd.get(4, 0);
33
    pad_right = pd.get(15, pad_left);
34
    pad_top = pd.get(14, pad_left);
35
    pad_bottom = pd.get(16, pad_top);
36
    output_w = pd.get(20, 0);
37
    output_h = pd.get(21, output_w);
38

39
    return 0;
40
}
41

42
int Fold::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
43
{
44
    const int max_channels = bottom_blob.h;
45
    size_t elemsize = bottom_blob.elemsize;
46

47
    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
48
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
49

50
    const int outw = output_w + pad_left + pad_right;
51
    const int outh = output_h + pad_top + pad_bottom;
52

53
    const int inw = (outw - kernel_extent_w) / stride_w + 1;
54
    const int inh = (outh - kernel_extent_h) / stride_h + 1;
55

56
    // assert inw * inh == size
57

58
    const int maxk = kernel_w * kernel_h;
59
    const int channels = max_channels / maxk;
60

61
    Mat top_blob_bordered;
62
    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
63
    {
64
        top_blob_bordered.create(outw, outh, channels, elemsize, opt.workspace_allocator);
65
    }
66
    else
67
    {
68
        top_blob_bordered = top_blob;
69
        top_blob_bordered.create(outw, outh, channels, elemsize, opt.blob_allocator);
70
    }
71
    if (top_blob_bordered.empty())
72
        return -100;
73

74
    // col2im
75
    const int gap = outw * stride_h - inw * stride_w;
76

77
    #pragma omp parallel for num_threads(opt.num_threads)
78
    for (int p = 0; p < channels; p++)
79
    {
80
        const float* sptr = bottom_blob.row(p * maxk);
81
        Mat outm = top_blob_bordered.channel(p);
82

83
        outm.fill(0.f);
84

85
        for (int u = 0; u < kernel_h; u++)
86
        {
87
            for (int v = 0; v < kernel_w; v++)
88
            {
89
                float* ptr = outm.row(dilation_h * u) + dilation_w * v;
90

91
                for (int i = 0; i < inh; i++)
92
                {
93
                    for (int j = 0; j < inw; j++)
94
                    {
95
                        ptr[0] += sptr[0];
96

97
                        ptr += stride_w;
98
                        sptr += 1;
99
                    }
100

101
                    ptr += gap;
102
                }
103
            }
104
        }
105
    }
106

107
    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
108
    {
109
        Option opt_b = opt;
110
        opt_b.use_packing_layout = false;
111
        copy_cut_border(top_blob_bordered, top_blob, pad_top, pad_bottom, pad_left, pad_right, opt_b);
112
        if (top_blob.empty())
113
            return -100;
114
    }
115
    else
116
    {
117
        top_blob = top_blob_bordered;
118
    }
119

120
    return 0;
121
}
122

123
} // namespace ncnn
124