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
247
248
249
250
251
252
253
254
255
256
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// See LICENSE file in the project root for full license information.
//
#include "WorkloadFactory.hpp"
#include "RefWorkloadFactory.hpp"
#include "NeonWorkloadFactory.hpp"
#include "ClWorkloadFactory.hpp"
#include "armnn/Types.hpp"
#include "armnn/LayerSupport.hpp"
#include "Layer.hpp"
#include "LayersFwd.hpp"
#include "CpuTensorHandle.hpp"
#include <boost/cast.hpp>
#include <cstring>
#include <boost/iterator/transform_iterator.hpp>
namespace armnn
{
bool IWorkloadFactory::IsLayerSupported(Compute compute, const Layer& layer, DataType dataType,
std::string& outReasonIfUnsupported)
{
constexpr size_t reasonCapacity = 1024;
char reason[reasonCapacity];
bool result;
switch(layer.GetType())
{
case LayerType::Activation:
{
auto cLayer = boost::polymorphic_downcast<const ActivationLayer*>(&layer);
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
result = IsActivationSupported(compute, input, cLayer->GetParameters(), reason, reasonCapacity);
break;
}
case LayerType::Addition:
{
const TensorInfo& input0 = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
const TensorInfo& input1 = layer.GetInputSlot(1).GetConnection()->GetTensorInfo();
const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo();
result = IsAdditionSupported(compute, input0, input1, output, reason, reasonCapacity);
break;
}
case LayerType::BatchNormalization:
{
auto cLayer = boost::polymorphic_downcast<const BatchNormalizationLayer*>(&layer);
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
result = IsBatchNormalizationSupported(compute, input, cLayer->GetParameters(), reason, reasonCapacity);
break;
}
case LayerType::Constant:
{
const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo();
result = IsConstantSupported(compute, output, reason, reasonCapacity);
break;
}
case LayerType::Convolution2d:
{
auto cLayer = boost::polymorphic_downcast<const Convolution2dLayer*>(&layer);
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo();
BOOST_ASSERT(cLayer->m_Weight.get() != nullptr);
const TensorInfo * biasInfo = nullptr;
static const TensorInfo dummyFloat32Bias(TensorShape({1,1,1,1}), DataType::Float32);
static const TensorInfo dummyQA8Bias(TensorShape({1,1,1,1}), DataType::Signed32);
const Convolution2dDescriptor& descriptor = cLayer->GetParameters();
if (descriptor.m_BiasEnabled)
{
BOOST_ASSERT(cLayer->m_Bias.get() != nullptr);
biasInfo = &(cLayer->m_Bias->GetTensorInfo());
}
else
{
// If biases are not enabled I pass a dummy tensorinfo for the validation
switch(input.GetDataType())
{
case DataType::Float32:
{
biasInfo = &dummyFloat32Bias;
break;
}
case DataType::QuantisedAsymm8:
{
biasInfo = &dummyQA8Bias;
break;
}
default:
{
BOOST_ASSERT_MSG(false, "Unexpected input type");
}
}
}
result = IsConvolution2dSupported(compute,
input,
output,
descriptor,
cLayer->m_Weight->GetTensorInfo(),
*biasInfo,
reason,
reasonCapacity);
break;
}
case LayerType::MemCopy:
{
// MemCopy supported for CpuRef, CpuAcc and GpuAcc backends
// (also treat Undefined as CpuRef to avoid breaking lots of Unit tests)
result = compute == Compute::CpuRef || compute == Compute::Undefined
|| compute == Compute::CpuAcc || compute == Compute::GpuAcc;
strcpy(reason, "Unsupported backend type");
break;
}
case LayerType::DepthwiseConvolution2d:
{
auto cLayer = boost::polymorphic_downcast<const DepthwiseConvolution2dLayer*>(&layer);
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
result = IsDepthwiseConvolutionSupported(compute, input, cLayer->GetParameters(),
cLayer->m_Weight->GetTensorInfo(), reason, reasonCapacity);
break;
}
case LayerType::FakeQuantization:
{
auto cLayer = boost::polymorphic_downcast<const FakeQuantizationLayer*>(&layer);
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
result = IsFakeQuantizationSupported(compute, input, cLayer->GetParameters(), reason, reasonCapacity);
break;
}
case LayerType::Floor:
{
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo();
result = IsFloorSupported(compute, input, output, reason, reasonCapacity);
break;
}
case LayerType::FullyConnected:
{
auto cLayer = boost::polymorphic_downcast<const FullyConnectedLayer*>(&layer);
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
result = IsFullyConnectedSupported(compute, input, cLayer->GetParameters(), reason, reasonCapacity);
break;
}
case LayerType::Input:
{
const TensorInfo& input = layer.GetOutputSlot(0).GetTensorInfo();
result = IsInputSupported(compute, input, reason, reasonCapacity);
break;
}
case LayerType::L2Normalization:
{
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
result = IsL2NormalizationSupported(compute, input, reason, reasonCapacity);
break;
}
case LayerType::Merger:
{
auto cLayer = boost::polymorphic_downcast<const MergerLayer*>(&layer);
// Get vector of all inputs
auto getTensorInfo = [](const InputSlot& slot)
{
return &slot.GetConnectedOutputSlot()->GetTensorInfo();
};
auto begin = boost::make_transform_iterator(layer.GetInputSlots().begin(), getTensorInfo);
auto end = boost::make_transform_iterator(layer.GetInputSlots().end(), getTensorInfo);
std::vector<const TensorInfo*> inputs(begin, end);
result = IsMergerSupported(compute, inputs, cLayer->GetParameters(), reason, reasonCapacity);
break;
}
case LayerType::Multiplication:
{
const TensorInfo& input0 = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
const TensorInfo& input1 = layer.GetInputSlot(1).GetConnection()->GetTensorInfo();
result = IsMultiplicationSupported(compute, input0, input1, reason, reasonCapacity);
break;
}
case LayerType::Normalization:
{
auto cLayer = boost::polymorphic_downcast<const NormalizationLayer*>(&layer);
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo();
result = IsNormalizationSupported(compute, input, output, cLayer->GetParameters(), reason, reasonCapacity);
break;
}
case LayerType::Output:
{
const TensorInfo& output = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
result = IsOutputSupported(compute, output, reason, reasonCapacity);
break;
}
case LayerType::Permute:
{
auto cLayer = boost::polymorphic_downcast<const PermuteLayer*>(&layer);
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo();
result = IsPermuteSupported(compute, input, output, cLayer->GetParameters(), reason, reasonCapacity);
break;
}
case LayerType::Pooling2d:
{
auto cLayer = boost::polymorphic_downcast<const Pooling2dLayer*>(&layer);
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo();
result = IsPooling2dSupported(compute, input, output, cLayer->GetParameters(), reason, reasonCapacity);
break;
}
case LayerType::Reshape:
{
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
result = IsReshapeSupported(compute, input, reason, reasonCapacity);
break;
}
case LayerType::ResizeBilinear:
{
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
result = IsResizeBilinearSupported(compute, input, reason, reasonCapacity);
break;
}
case LayerType::Softmax:
{
auto cLayer = boost::polymorphic_downcast<const SoftmaxLayer*>(&layer);
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
result = IsSoftmaxSupported(compute, input, cLayer->GetParameters(), reason, reasonCapacity);
break;
}
case LayerType::Splitter:
{
auto cLayer = boost::polymorphic_downcast<const SplitterLayer*>(&layer);
const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
result = IsSplitterSupported(compute, input, cLayer->GetParameters(), reason, reasonCapacity);
break;
}
default:
{
BOOST_ASSERT_MSG(false, "WorkloadFactory did not recognise type of layer.");
strcpy(reason, "Unrecognised layer type");
result = false;
break;
}
}
outReasonIfUnsupported = reason;
return result;
}
bool IWorkloadFactory::IsLayerSupported(const Layer& layer, DataType dataType, std::string& outReasonIfUnsupported)
{
return IsLayerSupported(layer.GetComputeDevice(), layer, dataType, outReasonIfUnsupported);
}
}
|
