-
Notifications
You must be signed in to change notification settings - Fork 4.1k
/
deformableconv2d_pack4to1.h
209 lines (199 loc) · 10.9 KB
/
deformableconv2d_pack4to1.h
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
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
// Tencent is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// Unless required by applicable law or agreed to in writing, software distributed
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
// specific language governing permissions and limitations under the License.
static void deformableconv2d_pack4to1_sse(const std::vector<Mat>& bottom_blobs, Mat& top_blob, const Mat& weight_data_packed, const Mat& bias_data, int kernel_w, int kernel_h, int dilation_w, int dilation_h, int stride_w, int stride_h, int pad_left, int pad_top, int activation_type, const Mat& activation_params, const Option& opt)
{
const Mat& bottom_blob = bottom_blobs[0];
const Mat& offset = bottom_blobs[1];
const bool has_mask = (bottom_blobs.size() == 3);
const bool offset_not_pack = offset.elempack == 1;
const bool mask_not_pack = has_mask ? bottom_blobs[2].elempack == 1 : true;
int w = bottom_blob.w;
int h = bottom_blob.h;
int inch = bottom_blob.c;
int outw = top_blob.w;
int outh = top_blob.h;
int outch = top_blob.c;
const float* bias_data_ptr = bias_data;
const int elempack = 4;
const int out_elempack = 1;
const int wstep = out_elempack * elempack;
#pragma omp parallel for num_threads(opt.num_threads)
for (int h_col = 0; h_col < outh; h_col++)
{
for (int w_col = 0; w_col < outw; w_col++)
{
int h_in = h_col * stride_h - pad_top;
int w_in = w_col * stride_w - pad_left;
for (int oc = 0; oc < outch; oc++)
{
const float* kptr = weight_data_packed.channel(oc);
float* outptr = top_blob.channel(oc);
float _sum = 0.f;
if (bias_data_ptr)
_sum = *(bias_data_ptr + oc);
for (int i = 0; i < kernel_h; i++)
{
for (int j = 0; j < kernel_w; j++)
{
float offset_h = 0.f;
float offset_w = 0.f;
float mask_ = 1.f;
if (offset_not_pack)
{
offset_h = offset.channel((i * kernel_w + j) * 2).row(h_col)[w_col];
offset_w = offset.channel((i * kernel_w + j) * 2 + 1).row(h_col)[w_col];
}
else
{
const int y_c = (i * kernel_w + j) * 2;
const int x_c = (i * kernel_w + j) * 2 + 1;
offset_h = offset.channel(y_c / offset.elempack).row(h_col)[w_col * offset.elempack + y_c % offset.elempack];
offset_w = offset.channel(x_c / offset.elempack).row(h_col)[w_col * offset.elempack + x_c % offset.elempack];
}
if (has_mask)
{
const Mat& mask = bottom_blobs[2];
if (mask_not_pack)
{
mask_ = mask.channel(i * kernel_w + j).row(h_col)[w_col];
}
else
{
const int m_c = i * kernel_w + j;
mask_ = mask.channel(m_c / mask.elempack).row(h_col)[w_col * mask.elempack + m_c % mask.elempack];
}
}
const float h_im = h_in + i * dilation_h + offset_h;
const float w_im = w_in + j * dilation_w + offset_w;
// Bilinear
const bool cond = h_im > -1 && w_im > -1 && h_im < h && w_im < w;
float w1 = 0.f;
float w2 = 0.f;
float w3 = 0.f;
float w4 = 0.f;
bool v1_cond = false;
bool v2_cond = false;
bool v3_cond = false;
bool v4_cond = false;
int v1_pos = 0;
int v2_pos = 0;
int v3_pos = 0;
int v4_pos = 0;
if (cond)
{
int h_low = (int)floorf(h_im);
int w_low = (int)floorf(w_im);
int h_high = h_low + 1;
int w_high = w_low + 1;
float lh = h_im - h_low;
float lw = w_im - w_low;
float hh = 1 - lh;
float hw = 1 - lw;
v1_cond = (h_low >= 0 && w_low >= 0);
v2_cond = (h_low >= 0 && w_high <= w - 1);
v3_cond = (h_high <= h - 1 && w_low >= 0);
v4_cond = (h_high <= h - 1 && w_high <= w - 1);
if (v1_cond)
v1_pos = h_low * w + w_low;
if (v2_cond)
v2_pos = h_low * w + w_high;
if (v3_cond)
v3_pos = h_high * w + w_low;
if (v4_cond)
v4_pos = h_high * w + w_high;
w1 = hh * hw;
w2 = hh * lw;
w3 = lh * hw;
w4 = lh * lw;
}
for (int ic = 0; ic < inch; ic++)
{
const float* data_im_ptr = bottom_blob.channel(ic);
float _val_channel0 = 0.f;
float _val_channel1 = _val_channel0;
float _val_channel2 = _val_channel0;
float _val_channel3 = _val_channel0;
if (cond)
{
float _v1_channel0 = _val_channel0;
float _v1_channel1 = _val_channel0;
float _v1_channel2 = _val_channel0;
float _v1_channel3 = _val_channel0;
float _v2_channel0 = _val_channel0;
float _v2_channel1 = _val_channel0;
float _v2_channel2 = _val_channel0;
float _v2_channel3 = _val_channel0;
float _v3_channel0 = _val_channel0;
float _v3_channel1 = _val_channel0;
float _v3_channel2 = _val_channel0;
float _v3_channel3 = _val_channel0;
float _v4_channel0 = _val_channel0;
float _v4_channel1 = _val_channel0;
float _v4_channel2 = _val_channel0;
float _v4_channel3 = _val_channel0;
if (v1_cond)
{
_v1_channel0 = *(data_im_ptr + v1_pos * elempack);
_v1_channel1 = *(data_im_ptr + v1_pos * elempack + 1);
_v1_channel2 = *(data_im_ptr + v1_pos * elempack + 2);
_v1_channel3 = *(data_im_ptr + v1_pos * elempack + 3);
}
if (v2_cond)
{
_v2_channel0 = *(data_im_ptr + v2_pos * elempack);
_v2_channel1 = *(data_im_ptr + v2_pos * elempack + 1);
_v2_channel2 = *(data_im_ptr + v2_pos * elempack + 2);
_v2_channel3 = *(data_im_ptr + v2_pos * elempack + 3);
}
if (v3_cond)
{
_v3_channel0 = *(data_im_ptr + v3_pos * elempack);
_v3_channel1 = *(data_im_ptr + v3_pos * elempack + 1);
_v3_channel2 = *(data_im_ptr + v3_pos * elempack + 2);
_v3_channel3 = *(data_im_ptr + v3_pos * elempack + 3);
}
if (v4_cond)
{
_v4_channel0 = *(data_im_ptr + v4_pos * elempack);
_v4_channel1 = *(data_im_ptr + v4_pos * elempack + 1);
_v4_channel2 = *(data_im_ptr + v4_pos * elempack + 2);
_v4_channel3 = *(data_im_ptr + v4_pos * elempack + 3);
}
_val_channel0 = w1 * _v1_channel0 + w2 * _v2_channel0 + w3 * _v3_channel0 + w4 * _v4_channel0;
_val_channel1 = w1 * _v1_channel1 + w2 * _v2_channel1 + w3 * _v3_channel1 + w4 * _v4_channel1;
_val_channel2 = w1 * _v1_channel2 + w2 * _v2_channel2 + w3 * _v3_channel2 + w4 * _v4_channel2;
_val_channel3 = w1 * _v1_channel3 + w2 * _v2_channel3 + w3 * _v3_channel3 + w4 * _v4_channel3;
}
if (has_mask)
{
_val_channel0 *= mask_;
_val_channel1 *= mask_;
_val_channel2 *= mask_;
_val_channel3 *= mask_;
}
float _conv_w0 = *(kptr);
float _conv_w1 = *(kptr + out_elempack); // 1 * out_elempack
float _conv_w2 = *(kptr + 2); // 2 * out_elempack
float _conv_w3 = *(kptr + 3); // 3 * out_elempack
_sum += (_val_channel0 * _conv_w0 + _val_channel1 * _conv_w1 + _val_channel2 * _conv_w2 + _val_channel3 * _conv_w3);
kptr += wstep;
}
}
}
_sum = activation_ss(_sum, activation_type, activation_params);
*(outptr + h_col * outw + w_col) = _sum;
}
}
}
}