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 and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "SplitterLayer.hpp"
#include "LayerCloneBase.hpp"
#include <armnn/TypesUtils.hpp>
#include <backendsCommon/WorkloadData.hpp>
#include <backendsCommon/WorkloadFactory.hpp>
namespace armnn
{
SplitterLayer::SplitterLayer(const ViewsDescriptor& param, const char* name)
: LayerWithParameters(1, param.GetNumViews(), LayerType::Splitter, param, name)
{
}
std::unique_ptr<IWorkload> SplitterLayer::CreateWorkload(const IWorkloadFactory& factory) const
{
SplitterQueueDescriptor descriptor;
// Copies the window origins to the descriptor.
for (unsigned int i = 0; i < m_Param.GetNumViews(); ++i)
{
descriptor.m_ViewOrigins.emplace_back(
std::vector<unsigned int>(m_Param.GetViewOrigin(i), m_Param.GetViewOrigin(i) + m_Param.GetNumDimensions()));
}
return factory.CreateSplitter(descriptor, PrepInfoAndDesc(descriptor));
}
template<typename FactoryType>
void SplitterLayer::CreateTensors(const TensorHandleFactoryRegistry& registry, const FactoryType& factory)
{
//If sub tensors are supported than all the "splitter" need to do is to
//set the outputs to be appropriate sub tensors of the input.
bool useSubTensors = factory.SupportsSubTensors();
if (useSubTensors)
{
// Get outputHandler of previous layer
const OutputHandler& outputHandler = GetInputSlots()[0].GetConnectedOutputSlot()->GetOutputHandler();
const OutputSlot* slot = GetInputSlots()[0].GetConnectedOutputSlot();
const TensorInfo& parentInfo = outputHandler.GetTensorInfo();
ITensorHandle* inputData = outputHandler.GetData();
std::vector<std::unique_ptr<ITensorHandle>> subTensors;
// check if split is along the x or y (2 innermost dimensions)
auto numberOfDimensions = m_Param.GetNumDimensions();
// Compute split axis within class as aclCommon function causes header issues when included
auto ComputeSplitAxis = [&](const armnn::SplitterDescriptor& desc, const TensorShape& input)
{
unsigned int numSplit = desc.GetNumViews();
unsigned int numDimensions = desc.GetNumDimensions();
std::set<unsigned int> splitAxis;
for (unsigned int i = 0; i < numSplit; ++i)
{
for (unsigned int dimIdx = 0; dimIdx < numDimensions; ++dimIdx)
{
if (desc.GetViewSizes(i)[dimIdx] != input[dimIdx])
{
splitAxis.insert(dimIdx);
}
}
}
return splitAxis;
};
std::set<unsigned int> axis = ComputeSplitAxis(m_Param, parentInfo.GetShape());
std::set<unsigned int>::iterator axisIt = axis.begin();
bool isOnXorY = m_Param.GetNumDimensions() >= 3 &&
((*axisIt == numberOfDimensions - 1) ||
(*axisIt == numberOfDimensions - 2));
//Creates the outputs as subtensors of the input.
for (unsigned int i = 0; i < m_Param.GetNumViews(); ++i)
{
const TensorInfo& info = m_OutputHandlers[i].GetTensorInfo();
OutputSlot& outSlot = GetOutputSlot(i);
ITensorHandleFactory::FactoryId factoryId = outSlot.GetTensorHandleFactoryId();
const unsigned int numOutputSlots = GetNumOutputSlots();
// if split along x or y (2 innermost dimensions) and the next layers do not require padding
bool canUseSubTensorOnXorY = true;
bool isTensorHandleFactory = std::is_same<armnn::ITensorHandleFactory, FactoryType>::value;
if (isTensorHandleFactory)
{
for (unsigned int it = 0; it < numOutputSlots; ++it)
{
InputSlot* inputSlot = GetOutputSlot(it).GetConnection(0);
ITensorHandleFactory* handleFactory = registry.GetFactory(factoryId);
std::vector<Capability> capabilities =
handleFactory->GetCapabilities(&(inputSlot->GetOwningLayer()),
this,
CapabilityClass::PaddingRequired);
if (isOnXorY)
{
canUseSubTensorOnXorY = false;
if (capabilities.empty())
{
canUseSubTensorOnXorY = true;
}
}
if (!canUseSubTensorOnXorY)
{
break;
}
}
}
auto CreateSubTensor = [&]()
{
// Make sure:
// 1) quantization parameters are in the same space
// 2) the same TensorHandleFactory is used for input and split layer output
// 3) the output does not go to a Constant layer or input layer
// 4) if split along x or y (2 innermost dimensions) and the next layers do not require padding
if (parentInfo.IsTypeSpaceMatch(info) && //(1)
factoryId == slot->GetTensorHandleFactoryId() && //(2)
GetOutputSlot(i).GetConnection(0)->GetOwningLayer().GetType() != LayerType::Constant && //(3)
GetOutputSlot(i).GetConnection(0)->GetOwningLayer().GetType() != LayerType::Input && //(3)
canUseSubTensorOnXorY) //(4)
{
ARMNN_NO_DEPRECATE_WARN_BEGIN
return factory.CreateSubTensorHandle(*inputData,
info.GetShape(),
this->m_Param.GetViewOrigin(i));
ARMNN_NO_DEPRECATE_WARN_END
}
return std::unique_ptr<ITensorHandle>();
};
auto subTensor = CreateSubTensor();
if (!subTensor)
{
useSubTensors = false;
break; //Failed to create a valid sub-tensor, so stop trying with the rest of the views.
}
subTensors.push_back(std::move(subTensor));
}
if (useSubTensors)
{
unsigned int i = 0;
for (auto& subTensor : subTensors)
{
m_OutputHandlers[i].SetData(std::move(subTensor));
++i;
}
}
}
if (!useSubTensors)
{
for (unsigned int i = 0; i < m_Param.GetNumViews(); ++i)
{
m_OutputHandlers[i].CreateTensorHandles(factory);
}
}
}
void SplitterLayer::CreateTensorHandles(const TensorHandleFactoryRegistry& registry,
const IWorkloadFactory& workloadFactory,
const bool IsMemoryManaged)
{
IgnoreUnused(IsMemoryManaged);
OutputSlot& slot = GetOutputSlot(0);
ITensorHandleFactory::FactoryId factoryId = slot.GetTensorHandleFactoryId();
if (factoryId == ITensorHandleFactory::LegacyFactoryId)
{
CreateTensors(registry, workloadFactory);
}
else
{
ITensorHandleFactory* handleFactory = registry.GetFactory(factoryId);
ARMNN_ASSERT(handleFactory);
CreateTensors(registry, *handleFactory);
}
}
SplitterLayer* SplitterLayer::Clone(Graph& graph) const
{
return CloneBase<SplitterLayer>(graph, m_Param, GetName());
}
std::vector<TensorShape> SplitterLayer::InferOutputShapes(const std::vector<TensorShape>& inputShapes) const
{
IgnoreUnused(inputShapes);
ARMNN_ASSERT(inputShapes.size() == m_Param.GetNumViews());
std::vector<TensorShape> outShapes;
//Output shapes must match View shapes.
for (unsigned int viewIdx = 0; viewIdx < m_Param.GetNumViews(); viewIdx++)
{
const uint32_t* sizes = m_Param.GetViewSizes(viewIdx);
outShapes.push_back(TensorShape(m_Param.GetNumDimensions(), sizes));
}
return outShapes;
}
void SplitterLayer::ValidateTensorShapesFromInputs()
{
std::for_each(BeginOutputSlots(), EndOutputSlots(), [&](OutputSlot& outputSlot)
{
VerifyShapeInferenceType(outputSlot.GetTensorInfo().GetShape(), m_ShapeInferenceMethod);
});
std::vector<TensorShape> views;
for (unsigned int viewIdx = 0; viewIdx < m_Param.GetNumViews(); viewIdx++)
{
const uint32_t* sizes = m_Param.GetViewSizes(viewIdx);
views.push_back(TensorShape(m_Param.GetNumDimensions(), sizes));
}
auto inferredShapes = InferOutputShapes(views);
ARMNN_ASSERT(inferredShapes.size() == m_Param.GetNumViews());
for (unsigned int viewIdx = 0; viewIdx < m_Param.GetNumViews(); viewIdx++)
{
ValidateAndCopyShape(GetOutputSlot(viewIdx).GetTensorInfo().GetShape(),
inferredShapes[viewIdx],
m_ShapeInferenceMethod,
"SplitterLayer",
viewIdx);
}
}
void SplitterLayer::Accept(ILayerVisitor& visitor) const
{
visitor.VisitSplitterLayer(this, GetParameters(), GetName());
}
} // namespace armnn
|
