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
|
// Copyright (C) 2018 Intel Corporation
//
// SPDX-License-Identifier: Apache-2.0
//
#pragma once
#include <vector>
#include <map>
#include <set>
#include <memory>
#include <string>
#include "ie_blob.h"
#include "ie_plugin.hpp"
#include "cpp/ie_cnn_network.h"
#include "debug_options.h"
#include "inference_engine.hpp"
#include <CPP/network.hpp>
#include <CPP/memory.hpp>
#include <CPP/primitive.hpp>
#include <CPP/topology.hpp>
#include <CPP/pooling.hpp>
#include <CPP/eltwise.hpp>
#include <CPP/concatenation.hpp>
#include <CPP/detection_output.hpp>
#include <CPP/softmax.hpp>
#include <cpp_interfaces/impl/ie_executable_network_thread_safe_default.hpp>
#include <CPP/upsampling.hpp>
#include "cldnn_custom_layer.h"
namespace CLDNNPlugin {
struct InferenceEnv {
std::shared_ptr<const cldnn::engine> engine;
std::shared_ptr<cldnn::network> network;
std::map<std::string, cldnn::primitive_id> primitiveIDs;
std::map<std::string, std::vector<cldnn::primitive_id>> prevPrimitiveIDs;
std::map<std::string, InferenceEngine::InferenceEngineProfileInfo> perfMap;
std::set<cldnn::primitive_id> profilingIDs;
DebugOptions debugOptions;
std::map<std::string, InferenceEngine::SizeVector> outputDims;
std::map<std::string, cldnn::layout> inputLayouts;
std::vector<std::shared_ptr<cldnn::network>> batchNetworks;
int m_max_batch;
int m_bv_sz;
};
class CLDNNGraph : public InferenceEngine::ExecutableNetworkThreadSafeDefault {
public:
typedef std::shared_ptr<CLDNNGraph> Ptr;
struct Config {
Config() : useProfiling(false), dumpCustomKernels(false), exclusiveAsyncRequests(false),
memory_pool_on(true),
enableDynamicBatch(false),
queuePriority(cldnn::priority_mode_types::disabled),
queueThrottle(cldnn::throttle_mode_types::disabled) {}
void LoadFromMap(const std::map<std::string, std::string>& configMap);
bool enableDynamicBatch;
bool useProfiling;
bool dumpCustomKernels;
bool exclusiveAsyncRequests;
bool memory_pool_on;
cldnn::priority_mode_types queuePriority;
cldnn::throttle_mode_types queueThrottle;
CLDNNCustomLayerMap customLayers;
cldnn::tuning_config_options tuningConfig;
std::string graph_dumps_dir;
std::string sources_dumps_dir;
};
explicit CLDNNGraph(InferenceEngine::ICNNNetwork &network, const Config& config = {}, int max_batch = -1);
InferenceEngine::InferRequestInternal::Ptr
CreateInferRequestImpl(InferenceEngine::InputsDataMap networkInputs, InferenceEngine::OutputsDataMap networkOutputs) override;
static bool IsLayerSupported(const std::string &type) {
return LayerTypeFromStr(type) != NO_TYPE;
}
protected:
// graph members
std::shared_ptr<cldnn::topology> m_topology;
InferenceEnv m_env;
Config m_config;
InferenceEngine::InputsDataMap* p_currentInputs;
InferenceEngine::OutputsDataMap* p_currentOutputs;
int m_curBatch;
static const cldnn::primitive_id m_preProcessTag;
static const cldnn::primitive_id m_weightsTag;
static const cldnn::primitive_id m_biasesTag;
static const cldnn::primitive_id m_meanValuesTag;
static const cldnn::primitive_id m_postProcessTag;
static const cldnn::primitive_id m_scalesTag;
static const cldnn::primitive_id m_workaroundTag;
static const cldnn::primitive_id m_preCustomLayerTag;
static const cldnn::primitive_id m_postCustomLayerTag;
// internal types
enum LayerType {
Convolution,
ReLU,
ReLU6,
Sigmoid,
TanH,
ELU,
Activation,
LRN,
Pooling,
FullyConnected,
SoftMax,
Power,
Split,
Concatenate,
Eltwise,
SimplerNMS,
ROIPooling,
Crop,
Deconvolution,
PriorBox,
DetectionOutput,
Normalize,
Reshape,
Permute,
Flatten,
BatchNormalization,
PReLU,
ScaleShift,
Proposal,
PSROIPooling,
Clamp,
Copy,
Upsampling,
Resample,
RegionYolo,
ReorgYolo,
ConstantBlob,
ArgMax,
MVN,
Unpooling,
Tile,
RNN,
NO_TYPE
};
enum WeightRearrangeType {
BroadcastFeatures,
FlipDeconvDims,
NO_REARRANGE
};
cldnn::format m_defaultFormat;
cldnn::data_types m_networkPrecision;
void InitFormat(InferenceEngine::ICNNNetwork &network);
static cldnn::data_types DataTypeFromPrecision(InferenceEngine::Precision p);
static cldnn::format FormatFromLayout(InferenceEngine::Layout l);
static cldnn::upsampling_sample_type UpsamplingTypeFromString(const std::string& str);
void Load(InferenceEngine::ICNNNetwork &network);
static LayerType LayerTypeFromStr(const std::string& str);
static cldnn::pooling_mode PoolingModeFromIEPooling(InferenceEngine::PoolingLayer::PoolType pt, bool excludePadding = false);
static cldnn::eltwise_mode EltwiseModeFromIEEltwise(InferenceEngine::EltwiseLayer::eOperation op);
static cldnn::concatenation::concatenation_axis ConcatAxisFromIEAxis(unsigned axis);
static cldnn::prior_box_code_type PriorBoxCodeFromString(const std::string& str);
static cldnn::softmax::dimension_t SoftmaxDimensionFromIEAxis(const InferenceEngine::SoftMaxLayer* softmaxLayer, bool isPrevFC = false);
void CreatePrimitiveFromBlob(cldnn::primitive_id primID,
const InferenceEngine::Blob::Ptr pBlob,
cldnn::layout blobLayout,
size_t blobByteOffset = 0,
WeightRearrangeType rearrange = NO_REARRANGE);
void CreateWeightAndBiasPrimitives(const InferenceEngine::CNNLayerPtr& layer,
std::vector<cldnn::primitive_id>& weightsPrimID,
std::vector<cldnn::primitive_id>& biasesPrimID);
void CreateScaleWeightsAndBiasesFromBN(const InferenceEngine::BatchNormalizationLayer* bnLayer,
cldnn::primitive_id weightsPrimID,
cldnn::primitive_id biasesPrimID);
void AddPreProcessPrimitive(InferenceEngine::InputInfo::Ptr inputInfo);
void AddInputPrimitive(InferenceEngine::InputInfo::Ptr inputInfo);
void AddOutputPrimitive(std::string outputName, const InferenceEngine::DataPtr outputData,
InferenceEngine::Precision outputPrecision = InferenceEngine::Precision::UNSPECIFIED);
void CreateSingleLayerPrimitive(InferenceEngine::CNNLayerPtr& layer);
bool IsValidSplitConvMerge(const InferenceEngine::SplitLayer* splitLayer) const;
bool CanProcessDynBatch(InferenceEngine::ICNNNetwork &network) const;
static std::vector<InferenceEngine::CNNLayerPtr> GetNextLayers(const InferenceEngine::DataPtr data);
static std::vector<InferenceEngine::CNNLayerPtr> GetNextLayers(const InferenceEngine::CNNLayerPtr layer);
static InferenceEngine::CNNLayerPtr GetNextSingleLayer(const InferenceEngine::DataPtr data);
static InferenceEngine::CNNLayerPtr GetNextSingleLayer(const InferenceEngine::CNNLayerPtr layer);
std::vector<cldnn::primitive_id> GetPrevLayersPrimitives(const InferenceEngine::CNNLayerPtr layer) const;
void AddSingleValuePrimitive(cldnn::primitive_id valPrimID, cldnn::data_types dataType, float value);
void CreateGenericLayerBlobPrimitives(const InferenceEngine::GenericLayer* layer);
static void ValidateGenericLayerBlobs(const InferenceEngine::GenericLayer* layer, const std::vector<std::string>& blobNames);
static cldnn::tensor CldnnTensorFromIEDims(const InferenceEngine::SizeVector& dims);
static bool HasParam(const std::map<std::string, std::string>& layerParams, std::string paramName) {
auto p = layerParams.find(paramName);
return p != layerParams.end();
}
void InitProfileInfo(const std::string& layerName,
const std::string& layerType,
const std::string& execType,
InferenceEngine::InferenceEngineProfileInfo::LayerStatus status);
void changeInputBatch(size_t batch);
void CompileNetwork();
// Layer Primitive Creators
void CreatePReLUPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateBatchNormalizationPrimitive(InferenceEngine::CNNLayerPtr & layer);
void CreateFlattenPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreatePermutePrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateReshapePrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateNormalizePrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateDetectionOutputPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreatePriorBoxPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateDeconvolutionPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateCropPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateROIPoolingPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateSimplerNMSPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateEltwisePrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateConcatenatePrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateSplitPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateFusedSplitConvMergePrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreatePowerPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateSoftMaxPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateFullyConnectedPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreatePoolingPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateLRNPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateActivationPrimitive(InferenceEngine::CNNLayerPtr &layer, const LayerType type);
void CreateConvolutionPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateScaleShiftPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateProposalPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreatePSROIPoolingPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateCopyPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateUpsamplingPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateResamplePrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateYOLO2RegionPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateYOLO2ReorgPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateArgMaxPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateMaxUnpoolingPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateMVNPrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateTilePrimitive(InferenceEngine::CNNLayerPtr &layer);
void CreateRNNPrimitive(InferenceEngine::CNNLayerPtr &layer);
void AddConstantBlobInput(InferenceEngine::CNNLayerPtr &layer);
void CreateCustomLayerPrimitive(InferenceEngine::CNNLayerPtr &layer, CLDNNCustomLayerPtr customLayer);
};
}; // namespace CLDNNPlugin
|
