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
|
/*
* Copyright (c) Facebook, Inc. and its affiliates.
* All rights reserved.
* This source code is licensed under the BSD-style license found in the
* LICENSE file in the root directory of this source tree.
*/
#include <algorithm>
#include <memory>
#include "fbgemm/Fbgemm.h"
namespace fbgemm {
template <int SPATIAL_DIM, typename T, typename accT>
PackWeightsForConv<SPATIAL_DIM, T, accT>::PackWeightsForConv(
const conv_param_t<SPATIAL_DIM>& conv_p,
const T* sdata,
const BlockingFactors* blocking_params)
: conv_param_(conv_p) {
static_assert(
SPATIAL_DIM == 2 || SPATIAL_DIM == 3,
"Only 2D and 3D convolutions are supported");
// Note: The following logic should *exactly* match with what we have in
// FbgemmConv.cc
switch (ConvFastPath<SPATIAL_DIM, accT>(conv_p)) {
case optimized_conv_t::depthwise: {
W_dw_packed_ = std::make_shared<PackedDepthWiseConvMatrix>(
conv_p.G, SPATIAL_DIM == 3 ? 3 * 3 * 3 : 3 * 3, sdata);
break;
}
case optimized_conv_t::groupwise: {
W_gconv_packed_ =
std::make_shared<PackWeightMatrixForGConv<T, accT, SPATIAL_DIM>>(
matrix_op_t::Transpose, conv_p, sdata, nullptr);
break;
}
case optimized_conv_t::pointwise: {
int NDim = conv_p.OC / conv_p.G;
int KDim = conv_p.K[0] * conv_p.K[1] * conv_p.IC;
W_pointwise_packed_ = std::make_shared<PackBMatrix<T, accT>>(
matrix_op_t::Transpose,
KDim,
NDim,
sdata,
KDim / conv_p.G,
nullptr,
conv_p.G,
blocking_params);
break;
}
case optimized_conv_t::im2col: {
int NDim = conv_p.OC / conv_p.G;
int KDim = conv_p.K[0] * conv_p.K[1] * conv_p.IC;
W_im2col_packed_ = std::make_shared<PackBMatrix<T, accT>>(
matrix_op_t::Transpose,
KDim,
NDim,
sdata,
KDim / conv_p.G,
nullptr,
conv_p.G,
blocking_params);
break;
}
} // switch
}
template <int SPATIAL_DIM, typename T, typename accT>
void PackWeightsForConv<SPATIAL_DIM, T, accT>::unpack(T* origin_buf) {
if (W_dw_packed_) {
W_dw_packed_->unpack(origin_buf);
} else if (W_gconv_packed_) {
W_gconv_packed_->unpack(origin_buf);
} else if (W_im2col_packed_) {
W_im2col_packed_->unpack(origin_buf);
} else if (W_pointwise_packed_) {
W_pointwise_packed_->unpack(origin_buf);
} else {
assert(false && "At least one packed weights object should exist");
}
}
template <int SPATIAL_DIM, typename T, typename accT>
bool PackWeightsForConv<SPATIAL_DIM, T, accT>::isPackingCompliant(
const conv_param_t<SPATIAL_DIM>& test_conv_p) {
return conv_param_.IC == test_conv_p.IC && conv_param_.OC == test_conv_p.OC &&
conv_param_.G == test_conv_p.G &&
std::equal(
conv_param_.K.begin(),
conv_param_.K.end(),
test_conv_p.K.begin()) &&
std::equal(
conv_param_.stride.begin(),
conv_param_.stride.end(),
test_conv_p.stride.begin()) &&
std::equal(
conv_param_.pad.begin(),
conv_param_.pad.end(),
test_conv_p.pad.begin()) &&
std::equal(
conv_param_.dilation.begin(),
conv_param_.dilation.end(),
test_conv_p.dilation.begin());
}
template <int SPATIAL_DIM, typename T, typename accT>
std::string PackWeightsForConv<SPATIAL_DIM, T, accT>::mismatchingParams(
const conv_param_t<SPATIAL_DIM>& test_conv_p) {
std::string msg = "";
auto combineStr = [](std::string id, std::string str1, std::string str2) {
std::string out = id + std::string(" ");
out += str1;
out += std::string(" vs ") + str2;
out += std::string(";");
return out;
};
auto combineInt = [&combineStr](std::string id, int int1, int int2) {
return combineStr(id, std::to_string(int1), std::to_string(int2));
};
if (conv_param_.IC != test_conv_p.IC) {
msg += combineInt("input_channels", conv_param_.IC, test_conv_p.IC);
}
if (conv_param_.OC != test_conv_p.OC) {
msg += combineInt("output_channels", conv_param_.IC, test_conv_p.IC);
}
if (conv_param_.G != test_conv_p.G) {
msg += combineInt("groups", conv_param_.G, test_conv_p.G);
}
if (!std::equal(
conv_param_.K.begin(), conv_param_.K.end(), test_conv_p.K.begin())) {
msg += combineStr(
"kernel",
arrayToString<SPATIAL_DIM>(conv_param_.K),
arrayToString<SPATIAL_DIM>(test_conv_p.K));
}
if (!std::equal(
conv_param_.stride.begin(),
conv_param_.stride.end(),
test_conv_p.stride.begin())) {
msg += combineStr(
"stride",
arrayToString<SPATIAL_DIM>(conv_param_.stride),
arrayToString<SPATIAL_DIM>(test_conv_p.stride));
}
if (!std::equal(
conv_param_.pad.begin(),
conv_param_.pad.end(),
test_conv_p.pad.begin())) {
msg += combineStr(
"pad",
arrayToString<2 * SPATIAL_DIM>(conv_param_.pad),
arrayToString<2 * SPATIAL_DIM>(test_conv_p.pad));
}
if (!std::equal(
conv_param_.dilation.begin(),
conv_param_.dilation.end(),
test_conv_p.dilation.begin())) {
msg += combineStr(
"dilation",
arrayToString<SPATIAL_DIM>(conv_param_.dilation),
arrayToString<SPATIAL_DIM>(test_conv_p.dilation));
}
return msg;
}
template class PackWeightsForConv<2, int8_t, int32_t>;
template class PackWeightsForConv<3, int8_t, int32_t>;
} // namespace fbgemm
|
