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 "deconvolution3d.h"
16

17
#include "fused_activation.h"
18

19
namespace ncnn {
20

21
Deconvolution3D::Deconvolution3D()
22
{
23
    one_blob_only = true;
24
    support_inplace = false;
25
}
26

27
int Deconvolution3D::load_param(const ParamDict& pd)
28
{
29
    num_output = pd.get(0, 0);
30
    kernel_w = pd.get(1, 0);
31
    kernel_h = pd.get(11, kernel_w);
32
    kernel_d = pd.get(21, kernel_w);
33
    dilation_w = pd.get(2, 1);
34
    dilation_h = pd.get(12, dilation_w);
35
    dilation_d = pd.get(22, dilation_w);
36
    stride_w = pd.get(3, 1);
37
    stride_h = pd.get(13, stride_w);
38
    stride_d = pd.get(23, stride_w);
39
    pad_left = pd.get(4, 0);
40
    pad_right = pd.get(15, pad_left);
41
    pad_top = pd.get(14, pad_left);
42
    pad_bottom = pd.get(16, pad_top);
43
    pad_front = pd.get(24, pad_left);
44
    pad_behind = pd.get(17, pad_front);
45
    output_pad_right = pd.get(18, 0);
46
    output_pad_bottom = pd.get(19, output_pad_right);
47
    output_pad_behind = pd.get(20, output_pad_right);
48
    output_w = pd.get(25, 0);
49
    output_h = pd.get(26, output_w);
50
    output_d = pd.get(27, output_w);
51
    bias_term = pd.get(5, 0);
52
    weight_data_size = pd.get(6, 0);
53
    activation_type = pd.get(9, 0);
54
    activation_params = pd.get(10, Mat());
55

56
    return 0;
57
}
58

59
int Deconvolution3D::load_model(const ModelBin& mb)
60
{
61
    weight_data = mb.load(weight_data_size, 0);
62
    if (weight_data.empty())
63
        return -100;
64

65
    if (bias_term)
66
    {
67
        bias_data = mb.load(num_output, 1);
68
        if (bias_data.empty())
69
            return -100;
70
    }
71

72
    return 0;
73
}
74

75
static int deconvolution3d(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data, const Mat& bias_data, int kernel_w, int kernel_h, int kernel_d, int stride_w, int stride_h, int stride_d, int dilation_w, int dilation_h, int dilation_d, int activation_type, const Mat& activation_params, const Option& opt)
76
{
77
    const int outw = top_blob.w;
78
    const int outh = top_blob.h;
79
    const int outch = top_blob.c;
80

81
    const int maxk = kernel_w * kernel_h * kernel_d;
82

83
    // kernel offsets
84
    std::vector<int> _space_ofs(maxk);
85
    int* space_ofs = &_space_ofs[0];
86
    {
87
        int p1 = 0;
88
        int p2 = 0;
89
        int gap0 = outw * dilation_h - kernel_w * dilation_w;
90
        int gap1 = outh * outw * dilation_d - outw * kernel_h * dilation_h;
91
        for (int z = 0; z < kernel_d; z++)
92
        {
93
            for (int i = 0; i < kernel_h; i++)
94
            {
95
                for (int j = 0; j < kernel_w; j++)
96
                {
97
                    space_ofs[p1] = p2;
98
                    p1++;
99
                    p2 += dilation_w;
100
                }
101
                p2 += gap0;
102
            }
103
            p2 += gap1;
104
        }
105
    }
106

107
    #pragma omp parallel for num_threads(opt.num_threads)
108
    for (int p = 0; p < outch; p++)
109
    {
110
        Mat out = top_blob.channel(p);
111

112
        const float bias = bias_data.empty() ? 0.f : bias_data[p];
113

114
        out.fill(bias);
115

116
        // shadowed variable for less openmp task args
117
        const int w = bottom_blob.w;
118
        const int h = bottom_blob.h;
119
        const int d = bottom_blob.d;
120
        const int inch = bottom_blob.c;
121
        const int outw = top_blob.w;
122
        const int outh = top_blob.h;
123
        const int outd = top_blob.d;
124

125
        for (int z = 0; z < d; z++)
126
        {
127
            for (int i = 0; i < h; i++)
128
            {
129
                for (int j = 0; j < w; j++)
130
                {
131
                    float* outptr = out.depth(z * stride_d).row(i * stride_h) + j * stride_w;
132

133
                    const float* kptr = (const float*)weight_data + maxk * inch * p;
134

135
                    for (int q = 0; q < inch; q++)
136
                    {
137
                        const float val = bottom_blob.channel(q).depth(z).row(i)[j];
138

139
                        for (int k = 0; k < maxk; k++)
140
                        {
141
                            float w = kptr[k];
142
                            outptr[space_ofs[k]] += val * w;
143
                        }
144

145
                        kptr += maxk;
146
                    }
147
                }
148
            }
149
        }
150

151
        {
152
            float* outptr = out;
153
            int size = outw * outh * outd;
154

155
            for (int i = 0; i < size; i++)
156
            {
157
                outptr[i] = activation_ss(outptr[i], activation_type, activation_params);
158
            }
159
        }
160
    }
161

162
    return 0;
163
}
164

165
int Deconvolution3D::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
166
{
167
    int w = bottom_blob.w;
168
    int h = bottom_blob.h;
169
    int d = bottom_blob.d;
170
    size_t elemsize = bottom_blob.elemsize;
171

172
    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
173
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
174
    const int kernel_extent_d = dilation_d * (kernel_d - 1) + 1;
175

176
    int outw = (w - 1) * stride_w + kernel_extent_w + output_pad_right;
177
    int outh = (h - 1) * stride_h + kernel_extent_h + output_pad_bottom;
178
    int outd = (d - 1) * stride_d + kernel_extent_d + output_pad_behind;
179

180
    Mat top_blob_bordered;
181
    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0 || pad_front > 0 || pad_behind > 0 || (output_w > 0 && output_h > 0 && output_d > 0))
182
    {
183
        top_blob_bordered.create(outw, outh, outd, num_output, elemsize, opt.workspace_allocator);
184
    }
185
    else
186
    {
187
        top_blob_bordered = top_blob;
188
        top_blob_bordered.create(outw, outh, outd, num_output, elemsize, opt.blob_allocator);
189
    }
190
    if (top_blob_bordered.empty())
191
        return -100;
192

193
    int ret = deconvolution3d(bottom_blob, top_blob_bordered, weight_data, bias_data, kernel_w, kernel_h, kernel_d, stride_w, stride_h, stride_d, dilation_w, dilation_h, dilation_d, activation_type, activation_params, opt);
194
    if (ret != 0)
195
        return ret;
196

197
    cut_padding(top_blob_bordered, top_blob, opt);
198
    if (top_blob.empty())
199
        return -100;
200

201
    return 0;
202
}
203

204
void Deconvolution3D::cut_padding(const Mat& top_blob_bordered, Mat& top_blob, const Option& opt) const
205
{
206
    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0 || pad_front > 0 || pad_behind > 0)
207
    {
208
        copy_cut_border_3d(top_blob_bordered, top_blob, pad_top, pad_bottom, pad_left, pad_right, pad_front, pad_behind, opt);
209
    }
210
    else if (output_w > 0 && output_h > 0 && output_d > 0)
211
    {
212
        int wcut = top_blob_bordered.w - output_w;
213
        int hcut = top_blob_bordered.h - output_h;
214
        int dcut = top_blob_bordered.d - output_d;
215

216
        if (pad_left == -233 || pad_right == -233 || pad_top == -233 || pad_bottom == -233 || pad_front == -233 || pad_behind == -233)
217
        {
218
            // onnx padding=SAME_UPPER
219
            copy_cut_border_3d(top_blob_bordered, top_blob, hcut / 2, hcut - hcut / 2, wcut / 2, wcut - wcut / 2, dcut / 2, dcut - dcut / 2, opt);
220
        }
221
        else if (pad_left == -234 || pad_right == -234 || pad_top == -234 || pad_bottom == -234 || pad_front == -234 || pad_behind == -234)
222
        {
223
            // onnx padding=SAME_LOWER
224
            copy_cut_border_3d(top_blob_bordered, top_blob, hcut - hcut / 2, hcut / 2, wcut - wcut / 2, wcut / 2, dcut - dcut / 2, dcut / 2, opt);
225
        }
226
    }
227
    else
228
    {
229
        top_blob = top_blob_bordered;
230
    }
231
}
232

233
} // namespace ncnn
234