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
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// See LICENSE file in the project root for full license information.
//
#include "Graph.hpp"
#include "Layers.hpp"
#include <armnn/Utils.hpp>
#include <armnn/TypesUtils.hpp>
#include <boost/polymorphic_cast.hpp>
#include <boost/log/trivial.hpp>
#include <boost/assert.hpp>
#include <boost/format.hpp>
#include <unordered_map>
#include <DotSerializer.hpp>
#include <sstream>
namespace armnn
{
Graph::Graph(const Graph& other)
: m_LayersInOrder(other.m_LayersInOrder)
{
std::unordered_map<const Layer*, Layer*> otherToClonedMap;
for (auto&& otherLayer : other.m_Layers)
{
Layer* const layer = otherLayer->Clone(*this);
otherToClonedMap.emplace(otherLayer, layer);
}
// Copy slot connections
for (auto&& otherLayer : other.m_Layers)
{
Layer* const thisLayer = otherToClonedMap[otherLayer];
auto outputSlot = thisLayer->BeginOutputSlots();
for (auto&& otherOutputSlot : otherLayer->GetOutputSlots())
{
for (auto&& otherInputSlot : otherOutputSlot.GetConnections())
{
const Layer& otherTgtLayer = otherInputSlot->GetOwningLayer();
Layer* const thisTgtLayer = otherToClonedMap[&otherTgtLayer];
InputSlot& inputSlot = thisTgtLayer->GetInputSlot(otherInputSlot->GetSlotIndex());
outputSlot->Connect(inputSlot);
}
outputSlot->SetTensorInfo(otherOutputSlot.GetTensorInfo());
++outputSlot;
}
}
}
Status Graph::Print() const
{
if (m_Layers.empty())
{
BOOST_LOG_TRIVIAL(info) << "\n Graph is empty.\n";
return Status::Success;
}
BOOST_LOG_TRIVIAL(info) << "\n";
BOOST_LOG_TRIVIAL(info) << "Walking Pattern: \n";
for (auto&& it : TopologicalSort())
{
BOOST_LOG_TRIVIAL(info) << it->GetName() << ":" << GetLayerTypeAsCString(it->GetType())
<< ":" << GetComputeDeviceAsCString(it->GetComputeDevice());
}
BOOST_LOG_TRIVIAL(info) << "\n\n";
return Status::Success;
}
Status Graph::SerializeToDot(std::ostream& stream)
{
{
DotGraph graph(stream, "Optimized");
{
// Default node attributes:
DotDefaults nodes(stream, "node");
nodes.GetAttributeSet()
.AddAttribute("shape", "record");
}
{
// Default edge attributes:
DotDefaults edges(stream, "edge");
edges.GetAttributeSet()
.AddAttribute("fontsize", 8)
.AddAttribute("fontcolor", "blue")
.AddAttribute("fontname", "arial-bold");
}
// First declare the nodes
for (auto&& layer : m_Layers)
{
DotNode node(stream, layer->GetGuid(), GetLayerTypeAsCString(layer->GetType()));
// Extract the layer parameters
ParameterStringifyFunction extractParams = [&node](const std::string & name, const std::string & value){
node.GetContents().AddContent(name + " : " + value);
};
layer->SerializeLayerParameters(extractParams);
}
// Second declare the edges
for (auto&& layer : m_Layers)
{
LayerGuid toId = layer->GetGuid();
for (unsigned int i=0;i<layer->GetNumInputSlots(); i++)
{
OutputSlot* outputSlot = static_cast<OutputSlot*>(layer->GetInputSlot(i).GetConnection());
LayerGuid fromId = outputSlot->GetOwningLayer().GetGuid();
DotEdge edge(stream, fromId, toId);
// Now Print the tensor shape on the edge
{
// Construct the label attribute with HTML markup
std::stringstream ss;
{
ss << "< [";
const TensorShape& shape = outputSlot->GetTensorInfo().GetShape();
for (unsigned int i = 0; i < shape.GetNumDimensions(); i++)
{
if (i != 0)
{
ss << ",";
}
ss << shape[i];
}
ss << "] >";
}
edge.GetAttributeSet().AddAttribute("label", ss);
}
}
}
}
if (stream.bad())
{
return Status::Failure;
}
return Status::Success;
}
Status Graph::AllocateDynamicBuffers()
{
for (auto&& layer : m_Layers)
{
for (auto slot = layer->BeginOutputSlots(); slot != layer->EndOutputSlots(); ++slot)
{
slot->GetOutputHandler().AllocateTensors();
}
}
return Status::Success;
}
const Graph& Graph::TopologicalSort() const
{
if (!m_LayersInOrder)
{
//Reset layer order
for (auto&& it : m_Layers)
{
it->ResetPriority();
}
auto compareLayerPriority = [](const LayersList::value_type& layerA, const LayersList::value_type& layerB)
{
return layerA->GetPriority() < layerB->GetPriority();
};
m_Layers.sort(compareLayerPriority);
m_LayersInOrder = true;
}
return *this;
}
void Graph::AddCopyLayers()
{
// Returns true if the given layer could potentially need an intermediate copy layer (depending on its
// connections to other layers). At the time of writing, copy layers will be inserted in the following situations:
// CPU -> CL (and viceversa)
// CPU -> Neon (and viceversa)
auto MayNeedCopyLayer = [](const Layer& layer)
{
// All layers should have been associated with a valid compute device at this point
BOOST_ASSERT(layer.GetComputeDevice() != Compute::Undefined);
// Do not need another copy layer if copy layer is already present
return layer.GetType() != LayerType::MemCopy;
};
for (auto&& srcLayer : m_Layers)
{
if (MayNeedCopyLayer(*srcLayer))
{
unsigned int srcOutputIndex = 0;
for (auto&& srcOutput : srcLayer->GetOutputSlots())
{
for (auto&& dstInput : srcOutput.GetConnections())
{
Layer& dstLayer = dstInput->GetOwningLayer();
if (MayNeedCopyLayer(dstLayer) && (dstLayer.GetComputeDevice() != srcLayer->GetComputeDevice()))
{
// A copy layer is needed in between the source and destination layers
// Record the operation rather than attempting to modify the graph as we go
// (invalidating iterators)
const std::string copyLayerName = boost::str(boost::format("[ %1% (%2%) -> %3% (%4%) ]")
% srcLayer->GetName()
% srcOutputIndex
% dstLayer.GetName()
% dstInput->GetSlotIndex());
MemCopyLayer* const copyLayer = InsertNewLayer<MemCopyLayer>(*dstInput, copyLayerName.c_str());
copyLayer->SetComputeDevice(dstLayer.GetComputeDevice());
}
}
++srcOutputIndex;
}
}
}
}
void Graph::InferTensorInfos()
{
for (auto&& layer : TopologicalSort())
{
for (auto&& input : layer->GetInputSlots())
{
boost::ignore_unused(input);
BOOST_ASSERT_MSG(input.GetConnectedOutputSlot()->IsTensorInfoSet(),
"All inputs must have the TensorInfo set at this point.");
}
layer->ValidateTensorShapesFromInputs();
}
}
} // namespace armnn
|
