forked from NVIDIA/MinkowskiEngine
-
Notifications
You must be signed in to change notification settings - Fork 0
/
broadcast_kernel.hpp
163 lines (150 loc) · 6.64 KB
/
broadcast_kernel.hpp
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
/* Copyright (c) Chris Choy (chrischoy@ai.stanford.edu).
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
* FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
* IN THE SOFTWARE.
*
* Please cite "4D Spatio-Temporal ConvNets: Minkowski Convolutional Neural
* Networks", CVPR'19 (https://arxiv.org/abs/1904.08755) if you use any part
* of the code.
*/
#ifndef CPU_BROADCAST
#define CPU_BROADCAST
#include "math_functions.hpp"
#include "types.hpp"
#include "utils.hpp"
namespace minkowski {
template <typename Dtype, typename Itype>
void BroadcastForwardKernelCPU(const Dtype *p_in_feat, uint32_t in_nrows,
const Dtype *p_in_feat_global,
uint32_t in_nrows_global, Dtype *p_out_feat,
uint32_t nchannel, BroadcastMode::Type const op,
const cpu_in_maps &in_maps,
const cpu_out_maps &glob_maps) {
Dtype *p_curr_out_feat;
const Dtype *p_curr_in_feat_global;
// Compute the size
uint32_t num_map = 0;
for (const auto &in_map : in_maps)
num_map += in_map.size();
ASSERT(num_map == in_nrows, "The number of in-out map,", num_map,
" mismatches the number of features,", in_nrows);
// Copy all in_feat to out_feat
std::memcpy(p_out_feat, p_in_feat, sizeof(Dtype) * in_nrows * nchannel);
// To speed up, put switch outside for loops
switch (op) {
case BroadcastMode::ELEMENTWISE_ADDITON: // +
for (uint32_t k = 0; k < in_maps.size(); ++k) {
for (uint32_t row = 0; row < in_maps[k].size(); ++row) {
p_curr_out_feat = p_out_feat + in_maps[k][row] * nchannel;
p_curr_in_feat_global = p_in_feat_global + glob_maps[k][row] * nchannel;
cpu_add<Dtype>(nchannel, p_curr_in_feat_global, p_curr_out_feat,
p_curr_out_feat);
}
}
break;
case BroadcastMode::ELEMENTWISE_MULTIPLICATION: // *
for (uint32_t k = 0; k < in_maps.size(); ++k) {
for (uint32_t row = 0; row < in_maps[k].size(); ++row) {
p_curr_out_feat = p_out_feat + in_maps[k][row] * nchannel;
p_curr_in_feat_global = p_in_feat_global + glob_maps[k][row] * nchannel;
cpu_mul<Dtype>(nchannel, p_curr_in_feat_global, p_curr_out_feat,
p_curr_out_feat);
}
}
break;
/*
case 2: // division
for (int k = 0; k < in_maps.size(); ++k) {
for (int row = 0; row < in_maps[k].size(); ++row) {
p_curr_out_feat = p_out_feat + in_maps[k][row] * nchannel;
p_curr_in_feat_global = p_in_feat_global + glob_maps[k][row] * nchannel;
cpu_div<Dtype>(nchannel, p_curr_in_feat_global, p_curr_out_feat,
p_curr_out_feat);
}
}
break;
*/
default:
throw std::invalid_argument(Formatter() << "Operation not supported: "
<< std::to_string(op));
}
}
template <typename Dtype, typename Itype>
void BroadcastBackwardKernelCPU(const Dtype *p_in_feat, //
Dtype *p_grad_in_feat, uint32_t in_nrows, //
const Dtype *p_in_feat_global,
Dtype *p_grad_in_feat_global,
uint32_t in_nrows_global, //
const Dtype *p_grad_out_feat, //
uint32_t nchannel,
BroadcastMode::Type const op, //
const cpu_in_maps &in_maps,
const cpu_out_maps &glob_maps) {
Dtype *p_curr_grad_in_feat, *p_curr_grad_in_feat_global;
const Dtype *p_curr_in_feat_global, *p_curr_in_feat, *p_curr_grad_out_feat;
// Assume that the memory is cleared
/*
// Clear grad memory
std::memset(p_grad_in_feat_global, 0,
sizeof(Dtype) * in_nrows_global * nchannel);
*/
// Initialize the grad_in_feat as grad_out_feat
std::memcpy(p_grad_in_feat, p_grad_out_feat,
sizeof(Dtype) * in_nrows * nchannel);
// To speed up, put switch outside for loops
switch (op) {
case BroadcastMode::ELEMENTWISE_ADDITON: // +
// For p_grad_in_feat, copy all grad_out
for (uint32_t k = 0; k < in_maps.size(); ++k) {
for (uint32_t row = 0; row < in_maps[k].size(); ++row) {
p_curr_grad_out_feat = p_grad_out_feat + in_maps[k][row] * nchannel;
p_curr_grad_in_feat_global =
p_grad_in_feat_global + glob_maps[k][row] * nchannel;
cpu_add<Dtype>(nchannel, p_curr_grad_out_feat,
p_curr_grad_in_feat_global, p_curr_grad_in_feat_global);
}
}
break;
case BroadcastMode::ELEMENTWISE_MULTIPLICATION: // *
for (uint32_t k = 0; k < in_maps.size(); ++k) {
for (uint32_t row = 0; row < in_maps[k].size(); ++row) {
// In feat global
p_curr_in_feat = p_in_feat + in_maps[k][row] * nchannel;
p_curr_grad_in_feat = p_grad_in_feat + in_maps[k][row] * nchannel;
p_curr_grad_in_feat_global =
p_grad_in_feat_global + glob_maps[k][row] * nchannel;
p_curr_grad_out_feat = p_grad_out_feat + in_maps[k][row] * nchannel;
p_curr_in_feat_global = p_in_feat_global + glob_maps[k][row] * nchannel;
// In feat
cpu_mul<Dtype>(nchannel, p_curr_in_feat_global, p_curr_grad_out_feat,
p_curr_grad_in_feat);
// In feat glob
for (uint32_t j = 0; j < nchannel; j++) {
p_curr_grad_in_feat_global[j] +=
p_curr_grad_out_feat[j] * p_curr_in_feat[j];
}
}
}
break;
default:
throw std::invalid_argument(Formatter() << "Operation not supported: "
<< std::to_string(op));
}
}
} // namespace minkowski
#endif