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 "unfold.h"
16

17
namespace ncnn {
18

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

24
int Unfold::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
    pad_value = pd.get(18, 0.f);
37

38
    return 0;
39
}
40

41
int Unfold::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
42
{
43
    Mat bottom_blob_bordered;
44
    {
45
        Option opt_b = opt;
46
        opt_b.blob_allocator = opt.workspace_allocator;
47
        opt_b.use_packing_layout = false;
48
        make_padding(bottom_blob, bottom_blob_bordered, opt_b);
49
        if (bottom_blob_bordered.empty())
50
            return -100;
51
    }
52

53
    const int w = bottom_blob_bordered.w;
54
    const int h = bottom_blob_bordered.h;
55
    const int channels = bottom_blob_bordered.c;
56
    const size_t elemsize = bottom_blob_bordered.elemsize;
57

58
    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
59
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
60

61
    const int outw = (w - kernel_extent_w) / stride_w + 1;
62
    const int outh = (h - kernel_extent_h) / stride_h + 1;
63

64
    const int size = outw * outh;
65
    const int maxk = kernel_w * kernel_h;
66

67
    top_blob.create(size, maxk * channels, elemsize, opt.blob_allocator);
68
    if (top_blob.empty())
69
        return -100;
70

71
    // im2col
72
    const int gap = w * stride_h - outw * stride_w;
73

74
    #pragma omp parallel for num_threads(opt.num_threads)
75
    for (int p = 0; p < channels; p++)
76
    {
77
        const Mat img = bottom_blob_bordered.channel(p);
78
        float* ptr = top_blob.row(p * maxk);
79

80
        for (int u = 0; u < kernel_h; u++)
81
        {
82
            for (int v = 0; v < kernel_w; v++)
83
            {
84
                const float* sptr = img.row(dilation_h * u) + dilation_w * v;
85

86
                for (int i = 0; i < outh; i++)
87
                {
88
                    for (int j = 0; j < outw; j++)
89
                    {
90
                        ptr[0] = sptr[0];
91

92
                        sptr += stride_w;
93
                        ptr += 1;
94
                    }
95

96
                    sptr += gap;
97
                }
98
            }
99
        }
100
    }
101

102
    return 0;
103
}
104

105
void Unfold::make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, const Option& opt) const
106
{
107
    int w = bottom_blob.w;
108
    int h = bottom_blob.h;
109

110
    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
111
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
112

113
    bottom_blob_bordered = bottom_blob;
114
    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
115
    {
116
        Option opt_b = opt;
117
        opt_b.blob_allocator = opt.workspace_allocator;
118
        copy_make_border(bottom_blob, bottom_blob_bordered, pad_top, pad_bottom, pad_left, pad_right, BORDER_CONSTANT, pad_value, opt_b);
119
    }
120
    else if (pad_left == -233 && pad_right == -233 && pad_top == -233 && pad_bottom == -233)
121
    {
122
        // tensorflow padding=SAME or onnx padding=SAME_UPPER
123
        int wpad = kernel_extent_w + (w - 1) / stride_w * stride_w - w;
124
        int hpad = kernel_extent_h + (h - 1) / stride_h * stride_h - h;
125
        if (wpad > 0 || hpad > 0)
126
        {
127
            Option opt_b = opt;
128
            opt_b.blob_allocator = opt.workspace_allocator;
129
            copy_make_border(bottom_blob, bottom_blob_bordered, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, BORDER_CONSTANT, pad_value, opt_b);
130
        }
131
    }
132
    else if (pad_left == -234 && pad_right == -234 && pad_top == -234 && pad_bottom == -234)
133
    {
134
        // onnx padding=SAME_LOWER
135
        int wpad = kernel_extent_w + (w - 1) / stride_w * stride_w - w;
136
        int hpad = kernel_extent_h + (h - 1) / stride_h * stride_h - h;
137
        if (wpad > 0 || hpad > 0)
138
        {
139
            Option opt_b = opt;
140
            opt_b.blob_allocator = opt.workspace_allocator;
141
            copy_make_border(bottom_blob, bottom_blob_bordered, hpad - hpad / 2, hpad / 2, wpad - wpad / 2, wpad / 2, BORDER_CONSTANT, pad_value, opt_b);
142
        }
143
    }
144
}
145

146
} // namespace ncnn
147