ArmNN
 24.08
QuantizedLstmLayer.cpp
Go to the documentation of this file.
1 //
2 // Copyright © 2017,2019-2024 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 #include "QuantizedLstmLayer.hpp"
6 
7 #include "LayerCloneBase.hpp"
8 
10 #include <armnn/TypesUtils.hpp>
13 
14 namespace armnn
15 {
16 
18  : Layer(3, 2, LayerType::QuantizedLstm, name)
19 {
20 }
21 
22 std::unique_ptr<IWorkload> QuantizedLstmLayer::CreateWorkload(const IWorkloadFactory& factory) const
23 {
25 
26  // QuantizedLstmLayer parameters - there are no optional params
31 
36 
41 
42  SetAdditionalInfo(descriptor);
43 
44  return factory.CreateWorkload(LayerType::QuantizedLstm, descriptor, PrepInfoAndDesc(descriptor));
45 }
46 
48 {
49  auto layer = CloneBase<QuantizedLstmLayer>(graph, GetName());
50 
51  layer->m_QuantizedLstmParameters.m_InputToInputWeights = m_QuantizedLstmParameters.m_InputToInputWeights ?
53  layer->m_QuantizedLstmParameters.m_InputToForgetWeights = m_QuantizedLstmParameters.m_InputToForgetWeights ?
55  layer->m_QuantizedLstmParameters.m_InputToCellWeights = m_QuantizedLstmParameters.m_InputToCellWeights ?
57  layer->m_QuantizedLstmParameters.m_InputToOutputWeights = m_QuantizedLstmParameters.m_InputToOutputWeights ?
59 
60  layer->m_QuantizedLstmParameters.m_RecurrentToInputWeights = m_QuantizedLstmParameters.m_RecurrentToInputWeights ?
62  layer->m_QuantizedLstmParameters.m_RecurrentToForgetWeights = m_QuantizedLstmParameters.m_RecurrentToForgetWeights
64  layer->m_QuantizedLstmParameters.m_RecurrentToCellWeights = m_QuantizedLstmParameters.m_RecurrentToCellWeights ?
66  layer->m_QuantizedLstmParameters.m_RecurrentToOutputWeights = m_QuantizedLstmParameters.m_RecurrentToOutputWeights
68 
69  layer->m_QuantizedLstmParameters.m_InputGateBias = m_QuantizedLstmParameters.m_InputGateBias ?
71  layer->m_QuantizedLstmParameters.m_ForgetGateBias = m_QuantizedLstmParameters.m_ForgetGateBias ?
73  layer->m_QuantizedLstmParameters.m_CellBias = m_QuantizedLstmParameters.m_CellBias ?
75  layer->m_QuantizedLstmParameters.m_OutputGateBias = m_QuantizedLstmParameters.m_OutputGateBias ?
77 
78  return std::move(layer);
79 }
80 
81 std::vector<TensorShape> QuantizedLstmLayer::InferOutputShapes(const std::vector<TensorShape>& inputShapes) const
82 {
83  if (inputShapes.size() != 3)
84  {
85  throw armnn::Exception("inputShapes' size is \"" + std::to_string(inputShapes.size()) +
86  "\" - should be \"3\".");
87  }
88 
89  // Get input values for validation
90  unsigned int numBatches = inputShapes[0][0];
91  unsigned int outputSize = inputShapes[1][1];
92 
93  std::vector<TensorShape> outShapes;
94  outShapes.push_back(TensorShape({numBatches, outputSize})); // cellStateOut
95  outShapes.push_back(TensorShape({numBatches, outputSize})); // output
96 
97  return outShapes;
98 }
99 
101 {
103 
104  const TensorShape& outputShape = GetOutputSlot(0).GetTensorInfo().GetShape();
105 
107 
108  auto inferredShapes = InferOutputShapes(
109  {
110  GetInputSlot(0).GetTensorInfo().GetShape(), // input
111  GetInputSlot(1).GetTensorInfo().GetShape(), // previousCellStateIn
112  GetInputSlot(2).GetTensorInfo().GetShape() // previousOutputIn
113  });
114 
115  if (inferredShapes.size() != 2)
116  {
117  throw armnn::LayerValidationException("inferredShapes has "
118  + std::to_string(inferredShapes.size()) +
119  " element(s) - should only have 2.");
120  }
121 
122  // Check weights and bias for nullptr
124  {
125  throw armnn::LayerValidationException("QuantizedLstmLayer: "
126  "m_QuantizedLstmParameters.m_InputToInputWeights "
127  "should not be null.");
128  }
129 
131  {
132  throw armnn::LayerValidationException("QuantizedLstmLayer: "
133  "m_QuantizedLstmParameters.m_InputToForgetWeights "
134  "should not be null.");
135  }
136 
138  {
139  throw armnn::LayerValidationException("QuantizedLstmLayer: "
140  "m_QuantizedLstmParameters.m_InputToCellWeights "
141  "should not be null.");
142  }
143 
145  {
146  throw armnn::LayerValidationException("QuantizedLstmLayer: "
147  "m_QuantizedLstmParameters.m_InputToOutputWeights "
148  "should not be null.");
149  }
150 
152  {
153  throw armnn::LayerValidationException("QuantizedLstmLayer: "
154  "m_QuantizedLstmParameters.m_RecurrentToInputWeights "
155  "should not be null.");
156  }
157 
159  {
160  throw armnn::LayerValidationException("QuantizedLstmLayer: "
161  "m_QuantizedLstmParameters.m_RecurrentToForgetWeights "
162  "should not be null.");
163  }
164 
166  {
167  throw armnn::LayerValidationException("QuantizedLstmLayer: "
168  "m_QuantizedLstmParameters.m_RecurrentToCellWeights "
169  "should not be null.");
170  }
171 
173  {
174  throw armnn::LayerValidationException("QuantizedLstmLayer: "
175  "m_QuantizedLstmParameters.m_RecurrentToOutputWeights "
176  "should not be null.");
177  }
178 
180  {
181  throw armnn::LayerValidationException("QuantizedLstmLayer: "
182  "m_QuantizedLstmParameters.m_InputGateBias "
183  "should not be null.");
184  }
185 
187  {
188  throw armnn::LayerValidationException("QuantizedLstmLayer: "
189  "m_QuantizedLstmParameters.m_ForgetGateBias "
190  "should not be null.");
191  }
192 
194  {
195  throw armnn::LayerValidationException("QuantizedLstmLayer: "
196  "m_QuantizedLstmParameters.m_CellBias "
197  "should not be null.");
198  }
199 
201  {
202  throw armnn::LayerValidationException("QuantizedLstmLayer: "
203  "m_QuantizedLstmParameters.m_OutputGateBias "
204  "should not be null.");
205  }
206 
207  // Check output TensorShape(s) match inferred shape
208  ValidateAndCopyShape(outputShape, inferredShapes[0], m_ShapeInferenceMethod, "QuantizedLstmLayer");
209 
211  inferredShapes[1],
213  "QuantizedLstmLayer",
214  1);
215 }
216 
218 {
219  // For API stability DO NOT ALTER order and add new members to the end of vector
220  return
221  {
226 
231 
236  };
237 }
238 
240 {
241  std::vector<ConstTensor> constTensors;
242 
247 
252 
257 
258  // InputToX weight tensors
260  {
261  constTensors.emplace_back(ConstTensor(managedInputToInputWeights.GetTensorInfo(),
262  managedInputToInputWeights.Map()));
263  }
264 
266  {
267  constTensors.emplace_back(ConstTensor(managedInputToForgetWeights.GetTensorInfo(),
268  managedInputToForgetWeights.Map()));
269  }
270 
272  {
273  constTensors.emplace_back(ConstTensor(managedInputToCellWeights.GetTensorInfo(),
274  managedInputToCellWeights.Map()));
275  }
276 
278  {
279  constTensors.emplace_back(ConstTensor(managedInputToOutputWeights.GetTensorInfo(),
280  managedInputToOutputWeights.Map()));
281  }
282 
283  // RecurrentToX weight tensors
285  {
286  constTensors.emplace_back(ConstTensor(
287  managedRecurrentToInputWeights.GetTensorInfo(),
288  managedRecurrentToInputWeights.Map()));
289  }
290 
292  {
293  constTensors.emplace_back(ConstTensor(
294  managedRecurrentToForgetWeights.GetTensorInfo(),
295  managedRecurrentToForgetWeights.Map()));
296  }
297 
299  {
300  constTensors.emplace_back(ConstTensor(
301  managedRecurrentToCellWeights.GetTensorInfo(),
302  managedRecurrentToCellWeights.Map()));
303  }
304 
306  {
307  constTensors.emplace_back(ConstTensor(
308  managedRecurrentToOutputWeights.GetTensorInfo(),
309  managedRecurrentToOutputWeights.Map()));
310  }
311 
312  // Bias tensors
314  {
315  constTensors.emplace_back(ConstTensor(managedInputGateBias.GetTensorInfo(),
316  managedInputGateBias.Map()));
317  }
318 
320  {
321  constTensors.emplace_back(ConstTensor(managedForgetGateBias.GetTensorInfo(),
322  managedForgetGateBias.Map()));
323  }
324 
325  if (m_QuantizedLstmParameters.m_CellBias != nullptr)
326  {
327  constTensors.emplace_back(ConstTensor(managedCellBias.GetTensorInfo(),
328  managedCellBias.Map()));
329  }
330 
332  {
333  constTensors.emplace_back(ConstTensor(managedOutputGateBias.GetTensorInfo(),
334  managedOutputGateBias.Map()));
335  }
336 
337 
338  strategy.ExecuteStrategy(this, BaseDescriptor(), constTensors, GetName());
339 }
340 
341 } // namespace armnn
armnn::QuantizedLstmQueueDescriptor::m_CellBias
const ConstTensorHandle * m_CellBias
Definition: WorkloadData.hpp:645
armnn::QuantizedLstmParameters::m_InputToInputWeights
std::shared_ptr< ConstTensorHandle > m_InputToInputWeights
A unique pointer to represent 2D weights tensor with dimensions [outputSize, inputSize] (QAsymm8).
Definition: QuantizedLstmLayer.hpp:17
armnn::QuantizedLstmQueueDescriptor::m_RecurrentToInputWeights
const ConstTensorHandle * m_RecurrentToInputWeights
Definition: WorkloadData.hpp:638
armnn::QuantizedLstmParameters::m_CellBias
std::shared_ptr< ConstTensorHandle > m_CellBias
A unique pointer to represent 1D bias tensor with dimensions [outputSize] (int32).
Definition: QuantizedLstmLayer.hpp:39
armnn::QuantizedLstmQueueDescriptor::m_InputToForgetWeights
const ConstTensorHandle * m_InputToForgetWeights
Definition: WorkloadData.hpp:634
armnn::QuantizedLstmLayer::ValidateTensorShapesFromInputs
void ValidateTensorShapesFromInputs() override
Check if the input tensor shape(s) will lead to a valid configuration of QuantizedLstmLayer.
Definition: QuantizedLstmLayer.cpp:100
armnn::OutputSlot::GetTensorInfo
const TensorInfo & GetTensorInfo() const override
Definition: Layer.cpp:100
armnn::QuantizedLstmLayer::GetConstantTensorsByRef
Layer::ImmutableConstantTensors GetConstantTensorsByRef() const override
Retrieve the handles to the constant values stored by the layer.
Definition: QuantizedLstmLayer.cpp:217
TypesUtils.hpp
armnn::QuantizedLstmLayer::Clone
QuantizedLstmLayer * Clone(Graph &graph) const override
Creates a dynamically-allocated copy of this layer.
Definition: QuantizedLstmLayer.cpp:47
armnn::QuantizedLstmParameters::m_OutputGateBias
std::shared_ptr< ConstTensorHandle > m_OutputGateBias
A unique pointer to represent 1D bias tensor with dimensions [outputSize] (int32).
Definition: QuantizedLstmLayer.hpp:41
CHECK_LOCATION
#define CHECK_LOCATION()
Definition: Exceptions.hpp:203
armnn::Layer::ValidateAndCopyShape
void ValidateAndCopyShape(const TensorShape &outputShape, const TensorShape &inferredShape, const ShapeInferenceMethod shapeInferenceMethod, const std::string &layerName, const unsigned int outputSlotIndex=0)
Definition: Layer.cpp:457
armnn::Layer::GetOutputSlot
const OutputSlot & GetOutputSlot(unsigned int index=0) const override
Get the const output slot handle by slot index.
Definition: Layer.hpp:339
armnn::QuantizedLstmQueueDescriptor::m_RecurrentToOutputWeights
const ConstTensorHandle * m_RecurrentToOutputWeights
Definition: WorkloadData.hpp:641
armnn::QuantizedLstmQueueDescriptor::m_InputGateBias
const ConstTensorHandle * m_InputGateBias
Definition: WorkloadData.hpp:643
armnn::LayerType::QuantizedLstm
@ QuantizedLstm
armnn::IStrategy
Definition: IStrategy.hpp:16
armnn::ManagedConstTensorHandle
Definition: TensorHandle.hpp:187
armnn::Layer::GetInputSlot
const InputSlot & GetInputSlot(unsigned int index) const override
Get a const input slot handle by slot index.
Definition: Layer.hpp:337
WorkloadFactory.hpp
armnn::Layer::GetName
const char * GetName() const override
Returns the name of the layer.
Definition: Layer.hpp:332
armnn::QuantizedLstmParameters::m_RecurrentToOutputWeights
std::shared_ptr< ConstTensorHandle > m_RecurrentToOutputWeights
A unique pointer to represent 2D weights tensor with dimensions [outputSize, outputSize] (QAsymm8).
Definition: QuantizedLstmLayer.hpp:32
armnn::QuantizedLstmParameters::m_InputToForgetWeights
std::shared_ptr< ConstTensorHandle > m_InputToForgetWeights
A unique pointer to represent 2D weights tensor with dimensions [outputSize, inputSize] (QAsymm8).
Definition: QuantizedLstmLayer.hpp:19
armnn::Layer
Definition: Layer.hpp:230
armnn::IConnectableLayer::ImmutableConstantTensors
std::vector< std::reference_wrapper< const std::shared_ptr< ConstTensorHandle > >> ImmutableConstantTensors
Definition: INetwork.hpp:141
armnn::InputSlot::GetTensorInfo
const TensorInfo & GetTensorInfo() const override
Gets the TensorInfo for this InputSlot.
Definition: Layer.cpp:614
armnn::QuantizedLstmParameters::m_InputToCellWeights
std::shared_ptr< ConstTensorHandle > m_InputToCellWeights
A unique pointer to represent 2D weights tensor with dimensions [outputSize, inputSize] (QAsymm8).
Definition: QuantizedLstmLayer.hpp:21
armnn::TensorShape
Definition: Tensor.hpp:20
armnn::QuantizedLstmQueueDescriptor::m_RecurrentToCellWeights
const ConstTensorHandle * m_RecurrentToCellWeights
Definition: WorkloadData.hpp:640
armnn::QuantizedLstmQueueDescriptor::m_InputToInputWeights
const ConstTensorHandle * m_InputToInputWeights
Definition: WorkloadData.hpp:633
armnn::QuantizedLstmParameters::m_ForgetGateBias
std::shared_ptr< ConstTensorHandle > m_ForgetGateBias
A unique pointer to represent 1D bias tensor with dimensions [outputSize] (int32).
Definition: QuantizedLstmLayer.hpp:37
armnn::Layer::PrepInfoAndDesc
WorkloadInfo PrepInfoAndDesc(QueueDescriptor &descriptor) const
Helper function to reduce duplication in *Layer::CreateWorkload.
Definition: Layer.hpp:409
QuantizedLstmLayer.hpp
armnn::QuantizedLstmParameters::m_InputToOutputWeights
std::shared_ptr< ConstTensorHandle > m_InputToOutputWeights
A unique pointer to represent 2D weights tensor with dimensions [outputSize, inputSize] (QAsymm8).
Definition: QuantizedLstmLayer.hpp:23
armnn::LayerValidationException
Definition: Exceptions.hpp:105
armnn::IWorkloadFactory
Definition: WorkloadFactory.hpp:22
armnn::QuantizedLstmQueueDescriptor::m_RecurrentToForgetWeights
const ConstTensorHandle * m_RecurrentToForgetWeights
Definition: WorkloadData.hpp:639
armnn::Layer::VerifyShapeInferenceType
void VerifyShapeInferenceType(const TensorShape &outputShape, ShapeInferenceMethod shapeInferenceMethod)
Definition: Layer.cpp:526
armnn::GetTensorInfo
const TensorInfo & GetTensorInfo(const ITensorHandle *tensorHandle)
float32 helpers
Definition: RefWorkloadUtils.hpp:33
armnn::ManagedConstTensorHandle::Map
const void * Map(bool blocking=true)
RAII Managed resource Unmaps MemoryArea once out of scope.
Definition: TensorHandle.hpp:196
armnn::Layer::SetAdditionalInfo
void SetAdditionalInfo(QueueDescriptor &descriptor) const
Definition: Layer.cpp:303
armnn::QuantizedLstmQueueDescriptor::m_InputToCellWeights
const ConstTensorHandle * m_InputToCellWeights
Definition: WorkloadData.hpp:635
armnn::Exception
Base class for all ArmNN exceptions so that users can filter to just those.
Definition: Exceptions.hpp:46
armnn::QuantizedLstmLayer
This layer represents a QuantizedLstm operation.
Definition: QuantizedLstmLayer.hpp:45
armnn::BaseDescriptor
Base class for all descriptors.
Definition: Descriptors.hpp:22
armnn::QuantizedLstmQueueDescriptor::m_InputToOutputWeights
const ConstTensorHandle * m_InputToOutputWeights
Definition: WorkloadData.hpp:636
armnn::QuantizedLstmQueueDescriptor
Definition: WorkloadData.hpp:614
armnn::QuantizedLstmLayer::m_QuantizedLstmParameters
QuantizedLstmParameters m_QuantizedLstmParameters
Definition: QuantizedLstmLayer.hpp:49
TensorHandle.hpp
armnn::QuantizedLstmLayer::CreateWorkload
virtual std::unique_ptr< IWorkload > CreateWorkload(const IWorkloadFactory &factory) const override
Makes a workload for the QuantizedLstm type.
Definition: QuantizedLstmLayer.cpp:22
armnn::QuantizedLstmLayer::InferOutputShapes
std::vector< TensorShape > InferOutputShapes(const std::vector< TensorShape > &inputShapes) const override
By default returns inputShapes if the number of inputs are equal to number of outputs,...
Definition: QuantizedLstmLayer.cpp:81
armnn::TensorInfo::GetShape
const TensorShape & GetShape() const
Definition: Tensor.hpp:193
armnn::QuantizedLstmLayer::ExecuteStrategy
void ExecuteStrategy(IStrategy &strategy) const override
Apply a visitor to this layer.
Definition: QuantizedLstmLayer.cpp:239
armnn
Copyright (c) 2021 ARM Limited and Contributors.
Definition: 01_00_quick_start.dox:6
QuantizedLstmParams.hpp
armnn::QuantizedLstmParameters::m_RecurrentToCellWeights
std::shared_ptr< ConstTensorHandle > m_RecurrentToCellWeights
A unique pointer to represent 2D weights tensor with dimensions [outputSize, outputSize] (QAsymm8).
Definition: QuantizedLstmLayer.hpp:30
armnn::ConstTensor
A tensor defined by a TensorInfo (shape and data type) and an immutable backing store.
Definition: Tensor.hpp:329
armnn::Layer::VerifyLayerConnections
void VerifyLayerConnections(unsigned int expectedConnections, const CheckLocation &location) const
Definition: Layer.cpp:410
armnn::QuantizedLstmParameters::m_InputGateBias
std::shared_ptr< ConstTensorHandle > m_InputGateBias
A unique pointer to represent 1D bias tensor with dimensions [outputSize] (int32).
Definition: QuantizedLstmLayer.hpp:35
armnn::Layer::m_ShapeInferenceMethod
ShapeInferenceMethod m_ShapeInferenceMethod
Definition: Layer.hpp:441
armnn::QuantizedLstmLayer::QuantizedLstmLayer
QuantizedLstmLayer(const char *name)
Constructor to create a QuantizedLstmLayer.
Definition: QuantizedLstmLayer.cpp:17
armnn::LayerType
LayerType
When adding a new layer, adapt also the LastLayer enum value in the enum class LayerType below.
Definition: Types.hpp:491
armnn::QuantizedLstmQueueDescriptor::m_OutputGateBias
const ConstTensorHandle * m_OutputGateBias
Definition: WorkloadData.hpp:646
armnn::Graph
Definition: Graph.hpp:30
armnn::IWorkloadFactory::CreateWorkload
virtual std::unique_ptr< IWorkload > CreateWorkload(LayerType type, const QueueDescriptor &descriptor, const WorkloadInfo &info) const =0
Backends should implement their own CreateWorkload function with a switch statement.
armnn::IStrategy::ExecuteStrategy
virtual void ExecuteStrategy(const IConnectableLayer *layer, const armnn::BaseDescriptor &descriptor, const std::vector< armnn::ConstTensor > &constants, const char *name, const armnn::LayerBindingId id=0)=0
armnn::QuantizedLstmParameters::m_RecurrentToInputWeights
std::shared_ptr< ConstTensorHandle > m_RecurrentToInputWeights
A unique pointer to represent 2D weights tensor with dimensions [outputSize, outputSize] (QAsymm8).
Definition: QuantizedLstmLayer.hpp:26
armnn::QuantizedLstmParameters::m_RecurrentToForgetWeights
std::shared_ptr< ConstTensorHandle > m_RecurrentToForgetWeights
A unique pointer to represent 2D weights tensor with dimensions [outputSize, outputSize] (QAsymm8).
Definition: QuantizedLstmLayer.hpp:28
LayerCloneBase.hpp
armnn::ManagedConstTensorHandle::GetTensorInfo
const TensorInfo & GetTensorInfo() const
Definition: TensorHandle.hpp:239
armnn::QuantizedLstmQueueDescriptor::m_ForgetGateBias
const ConstTensorHandle * m_ForgetGateBias
Definition: WorkloadData.hpp:644