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
210
211
212
213
214
215
216
217
218
219
220
|
// Copyright (C) 2018 Intel Corporation
//
// SPDX-License-Identifier: Apache-2.0
//
#include "mkldnn_deconv_node.h"
#include "desc_iterator.hpp"
#include <ie_layers.h>
#include <mkldnn.hpp>
#include <string>
#include <vector>
#include <mkldnn_types.h>
#include <mkldnn_extension_utils.h>
#include "ie_parallel.hpp"
using namespace mkldnn;
using namespace MKLDNNPlugin;
using namespace InferenceEngine;
MKLDNNDeconvolutionNode::MKLDNNDeconvolutionNode(const InferenceEngine::CNNLayerPtr& layer, const mkldnn::engine& eng) : MKLDNNNode(layer, eng) {
internalBlobDesc.emplace_back([&](primitive_desc_iterator &primitive_desc_it, size_t idx) -> MKLDNNMemoryDesc {
return MKLDNNMemoryDesc(primitive_desc_it.weights_primitive_desc(0).desc());
});
}
void MKLDNNDeconvolutionNode::getSupportedDescriptors() {
if (!descs_fwd.empty() && !descs_bwd.empty())
return;
InferenceEngine::Precision precision = getCnnLayer()->insData[0].lock()->getPrecision();
if (precision != InferenceEngine::Precision::FP32)
precision = InferenceEngine::Precision::FP32;
auto inputDataType = MKLDNNExtensionUtils::IEPrecisionToDataType(precision);
precision = getCnnLayer()->outData[0]->getPrecision();
if (precision != InferenceEngine::Precision::FP32)
precision = InferenceEngine::Precision::FP32;
auto outputDataType = MKLDNNExtensionUtils::IEPrecisionToDataType(precision);
if (getParentEdges().size() != 1)
THROW_IE_EXCEPTION << "Incorrect number of input edges for layer " << getName();
if (getChildEdges().empty())
THROW_IE_EXCEPTION << "Incorrect number of output edges for layer " << getName();
auto * deconvLayer = dynamic_cast<DeconvolutionLayer*>(getCnnLayer().get());
if (deconvLayer == nullptr)
THROW_IE_EXCEPTION << "Cannot convert deconvolution layer.";
if (deconvLayer->_weights == nullptr) {
THROW_IE_EXCEPTION << "Weights are empty for layer: " << deconvLayer->name
<< " used in MKLDNN node: " << getName() << "\n"
<< "Use ReadWeights and SetWeights methods of InferenceEngine::CNNNetReader"
<< " to load them from .bin part of the IR";
}
withGroups = (deconvLayer->_group > 1);
isDW = withGroups && deconvLayer->_group == deconvLayer->_out_depth &&
deconvLayer->_group == deconvLayer->input()->getDims()[1];
withBiases = (deconvLayer->_biases != nullptr && deconvLayer->_biases->size() != 0);
if (withBiases)
biases = deconvLayer->_biases;
/* Original layout format for deconv weights is iohw (from Caffe).
* We specify oihw, but mean iohw, because there are no more
* suitable format in MKLDNN.
*/
SizeVector weightDims;
if (withGroups) {
weightDims = {
deconvLayer->_group,
deconvLayer->input()->getTensorDesc().getDims()[1] / deconvLayer->_group,
deconvLayer->_out_depth / deconvLayer->_group,
deconvLayer->_kernel[Y_AXIS],
deconvLayer->_kernel[X_AXIS]
};
groupNum = deconvLayer->_group;
} else {
weightDims = {
deconvLayer->input()->getTensorDesc().getDims()[1],
deconvLayer->_out_depth,
deconvLayer->_kernel[Y_AXIS],
deconvLayer->_kernel[X_AXIS]
};
}
internalBlobs.push_back(createInternalBlob(weightDims, true));
invertVectorCopyUtoI(deconvLayer->_stride, stride);
for (int i = 1; i <= deconvLayer->_dilation.size(); i++) {
dilation.push_back(static_cast<int>(deconvLayer->_dilation[deconvLayer->_dilation.size() - i]) - 1);
}
invertVectorCopyUtoI(deconvLayer->_padding, paddingL);
invertVectorCopyUtoI(deconvLayer->_pads_end, paddingR);
weightsDims = MKLDNNDims(weightDims);
for (int i = 0; i < 2; i++) {
int with_group = (withGroups) ? 1 : 0;
int krn = weightsDims[with_group + 2 + i];
int src = getChildEdgeAt(0)->getDims()[2 + i];
int dst = getParentEdgeAt(0)->getDims()[2 + i];
krn = (krn - 1)*(dilation[i] + 1) + 1;
int calc_dst = (src - krn + paddingL[i]) / stride[i] + 1;
paddingR[i] = (dst - calc_dst) * stride[i];
}
for (auto format : getAvailableFormatsForDims(getParentEdgeAt(0)->getDims())) {
MKLDNNMemoryDesc in_candidate(getParentEdgeAt(0)->getDims(), inputDataType, format);
MKLDNNMemoryDesc out_candidate(getChildEdgeAt(0)->getDims(), outputDataType, format);
createDescriptor({in_candidate}, {out_candidate});
}
}
void MKLDNNDeconvolutionNode::execute(mkldnn::stream strm) {
if (prim) {
strm.submit({*prim});
}
if (withBiases) {
const auto *bias = biases->buffer().as<const float*>();
auto dst = getChildEdgeAt(0)->getBlob();
float *output = dst->buffer().as<float *>() + dst->getTensorDesc().getBlockingDesc().getOffsetPadding();
const size_t N = dst->getTensorDesc().getDims()[0];
const size_t C = dst->getTensorDesc().getBlockingDesc().getBlockDims()[1] / groupNum;
const size_t H = dst->getTensorDesc().getDims()[2];
const size_t W = dst->getTensorDesc().getDims()[3];
const size_t blkC =
dst->getTensorDesc().getBlockingDesc().getBlockDims().size() > 4 ?
dst->getTensorDesc().getBlockingDesc().getBlockDims()[4] :
1;
auto strides = dst->getTensorDesc().getBlockingDesc().getStrides();
parallel_for4d(N, C, H, W, [&](size_t n, size_t c, size_t h, size_t w) {
for (size_t g = 0; g < groupNum; g++) {
const size_t off = n * strides[0] + (g * C + c) * strides[1] + h * strides[2] + w * strides[3];
auto o = &output[off];
for (int bc = 0; bc < blkC; ++bc) {
o[bc] += bias[c * blkC + bc];
}
}
});
}
}
bool MKLDNNDeconvolutionNode::created() const {
return getType() == Deconvolution;
}
void MKLDNNDeconvolutionNode::createPrimitive() {
if (prim)
return;
auto prim_desc = createPrimitiveDescriptor<convolution_backward_data::primitive_desc,
convolution_backward_data::desc, convolution_forward::primitive_desc>();
prim.reset(new convolution_backward_data(prim_desc,
getParentEdgeAt(0)->getMemory().GetPrimitive(),
internalBlobMemory[0]->GetPrimitive(),
getChildEdgeAt(0)->getMemory().GetPrimitive()));
}
void MKLDNNDeconvolutionNode::createDescriptor(const std::vector<InferenceEngine::TensorDesc> &inputDesc,
const std::vector<InferenceEngine::TensorDesc> &outputDesc) {
MKLDNNMemoryDesc in_candidate(inputDesc[0]);
MKLDNNMemoryDesc out_candidate(outputDesc[0]);
auto in_fmt = in_candidate.getFormat();
auto out_fmt = out_candidate.getFormat();
int O_IND = withGroups ? 1 : 0;
int I_IND = withGroups ? 2 : 1;
// grouping and autoblicking is not compatible
if ((withGroups && !isDW) && (in_candidate.blocksExtended() || out_candidate.blocksExtended()))
return;
MKLDNNMemoryDesc wgh_candidate{weightsDims, in_candidate.getDataType(), memory::any};
for (auto alg : {algorithm::convolution_winograd, algorithm::convolution_direct}) {
try {
std::shared_ptr<mkldnn::convolution_forward::desc> conv_desc;
conv_desc.reset(new convolution_forward::desc(prop_kind::forward_inference, alg,
out_candidate, wgh_candidate, in_candidate, stride, dilation,
paddingL, paddingR, padding_kind::zero));
std::shared_ptr<mkldnn::convolution_backward_data::desc> deconv_desc;
deconv_desc.reset(new convolution_backward_data::desc(alg, out_candidate, wgh_candidate,
in_candidate, stride, dilation, paddingL, paddingR,
padding_kind::zero));
descs_fwd.push_back(conv_desc);
descs_bwd.push_back(deconv_desc);
descs.emplace_back(deconv_desc,
std::shared_ptr<convolution_forward::primitive_desc>(
new convolution_forward::primitive_desc(*conv_desc, getEngine())));
} catch(...) {}
}
}
MKLDNNMemoryDesc MKLDNNDeconvolutionNode::getSrcMemDesc(mkldnn::primitive_desc_iterator &primitive_desc_it, size_t idx) {
InferenceEngine::TensorDesc desc = MKLDNNMemoryDesc(primitive_desc_it.diff_dst_primitive_desc(idx).desc());
if (desc.getLayout() == InferenceEngine::Layout::ANY)
return MKLDNNMemoryDesc(InferenceEngine::TensorDesc(desc.getPrecision(),
getParentEdgeAt(idx)->getDims().ToSizeVector(),
desc.getLayout()));
else
return MKLDNNMemoryDesc(InferenceEngine::TensorDesc(desc.getPrecision(),
getParentEdgeAt(idx)->getDims().ToSizeVector(),
desc.getBlockingDesc()));
}
MKLDNNMemoryDesc MKLDNNDeconvolutionNode::getDstMemDesc(mkldnn::primitive_desc_iterator &primitive_desc_it, size_t idx) {
InferenceEngine::TensorDesc desc = MKLDNNMemoryDesc(primitive_desc_it.diff_src_primitive_desc(idx).desc());
if (desc.getLayout() == InferenceEngine::Layout::ANY)
return MKLDNNMemoryDesc(InferenceEngine::TensorDesc(desc.getPrecision(),
getChildEdgeAt(idx)->getDims().ToSizeVector(),
desc.getLayout()));
else
return MKLDNNMemoryDesc(InferenceEngine::TensorDesc(desc.getPrecision(),
getChildEdgeAt(idx)->getDims().ToSizeVector(),
desc.getBlockingDesc()));
}
|
