ArmNN
 20.05
Public Member Functions | List of all members
NeonLstmFloatWorkload Class Reference

#include <NeonLstmFloatWorkload.hpp>

Inheritance diagram for NeonLstmFloatWorkload:
TypedWorkload< QueueDescriptor, DataTypes > BaseWorkload< QueueDescriptor > IWorkload

Public Member Functions

 NeonLstmFloatWorkload (const LstmQueueDescriptor &descriptor, const WorkloadInfo &info)
 
virtual void Execute () const override
 
- Public Member Functions inherited from TypedWorkload< QueueDescriptor, DataTypes >
 TypedWorkload (const QueueDescriptor &descriptor, const WorkloadInfo &info)
 
- Public Member Functions inherited from BaseWorkload< QueueDescriptor >
 BaseWorkload (const QueueDescriptor &descriptor, const WorkloadInfo &info)
 
void PostAllocationConfigure () override
 
const QueueDescriptor & GetData () const
 
profiling::ProfilingGuid GetGuid () const final
 
- Public Member Functions inherited from IWorkload
virtual ~IWorkload ()
 
virtual void RegisterDebugCallback (const DebugCallbackFunction &)
 

Additional Inherited Members

- Protected Attributes inherited from BaseWorkload< QueueDescriptor >
const QueueDescriptor m_Data
 
const profiling::ProfilingGuid m_Guid
 

Detailed Description

Definition at line 19 of file NeonLstmFloatWorkload.hpp.

Constructor & Destructor Documentation

◆ NeonLstmFloatWorkload()

NeonLstmFloatWorkload ( const LstmQueueDescriptor &  descriptor,
const WorkloadInfo &  info 
)

Definition at line 17 of file NeonLstmFloatWorkload.cpp.

18  : FloatWorkload<LstmQueueDescriptor>(descriptor, info)
19 {
20  arm_compute::LSTMParams<arm_compute::ITensor> lstm_param;
21 
22  // Basic parameters
23  m_InputToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
24  BuildArmComputeTensor(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights->GetTensorInfo());
25 
26  m_InputToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
27  BuildArmComputeTensor(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights->GetTensorInfo());
28 
29  m_InputToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
30  BuildArmComputeTensor(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights->GetTensorInfo());
31 
32  m_RecurrentToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
33  BuildArmComputeTensor(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights->GetTensorInfo());
34 
35  m_RecurrentToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
36  BuildArmComputeTensor(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights->GetTensorInfo());
37 
38  m_RecurrentToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
39  BuildArmComputeTensor(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights->GetTensorInfo());
40 
41  m_ForgetGateBiasTensor = std::make_unique<arm_compute::Tensor>();
42  BuildArmComputeTensor(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias->GetTensorInfo());
43 
44  m_CellBiasTensor = std::make_unique<arm_compute::Tensor>();
45  BuildArmComputeTensor(*m_CellBiasTensor, m_Data.m_CellBias->GetTensorInfo());
46 
47  m_OutputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
48  BuildArmComputeTensor(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias->GetTensorInfo());
49 
50  // for future reference: check the AndroidNN API for the logic here
51  if (!m_Data.m_Parameters.m_CifgEnabled)
52  {
53  m_InputToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
54  BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo());
55 
56  m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
57  BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo());
58 
59  m_CellToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
60  if (m_Data.m_CellToInputWeights != nullptr)
61  {
62  BuildArmComputeTensor(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights->GetTensorInfo());
63  }
64 
65  m_InputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
66  BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo());
67 
68  lstm_param.set_cifg_params(m_InputToInputWeightsTensor.get(),
69  m_RecurrentToInputWeightsTensor.get(),
70  m_Data.m_CellToInputWeights != nullptr ? m_CellToInputWeightsTensor.get() : nullptr,
71  m_InputGateBiasTensor.get());
72  }
73 
74  if (m_Data.m_Parameters.m_ProjectionEnabled)
75  {
76  m_ProjectionWeightsTensor = std::make_unique<arm_compute::Tensor>();
77  BuildArmComputeTensor(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights->GetTensorInfo());
78 
79  m_ProjectionBiasTensor = std::make_unique<arm_compute::Tensor>();
80  if (m_Data.m_ProjectionBias != nullptr)
81  {
82  BuildArmComputeTensor(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias->GetTensorInfo());
83  }
84 
85  lstm_param.set_projection_params(m_ProjectionWeightsTensor.get(),
86  m_Data.m_ProjectionBias != nullptr ? m_ProjectionBiasTensor.get() : nullptr);
87  }
88 
89  if (m_Data.m_Parameters.m_PeepholeEnabled)
90  {
91  m_CellToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
92  BuildArmComputeTensor(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights->GetTensorInfo());
93 
94  m_CellToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
95  BuildArmComputeTensor(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights->GetTensorInfo());
96 
97  lstm_param.set_peephole_params(m_CellToForgetWeightsTensor.get(), m_CellToOutputWeightsTensor.get());
98  }
99 
100  if (m_Data.m_Parameters.m_LayerNormEnabled)
101  {
102  m_InputLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
103  if (!m_Data.m_Parameters.m_CifgEnabled)
104  {
105  BuildArmComputeTensor(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights->GetTensorInfo());
106  }
107 
108  m_ForgetLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
109  BuildArmComputeTensor(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights->GetTensorInfo());
110 
111  m_CellLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
112  BuildArmComputeTensor(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights->GetTensorInfo());
113 
114  m_OutputLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
115  BuildArmComputeTensor(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights->GetTensorInfo());
116 
117  lstm_param.set_layer_normalization_params(m_Data.m_Parameters.m_CifgEnabled ?
118  nullptr : m_InputLayerNormWeightsTensor.get(),
119  m_ForgetLayerNormWeightsTensor.get(),
120  m_CellLayerNormWeightsTensor.get(),
121  m_OutputLayerNormWeightsTensor.get());
122  }
123 
124  const arm_compute::ITensor& input = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
125  const arm_compute::ITensor& output_state_in = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
126  const arm_compute::ITensor& cell_state_in = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();
127 
128  arm_compute::ITensor& output_state_out = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[1])->GetTensor();
129  arm_compute::ITensor& cell_state_out = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[2])->GetTensor();
130  arm_compute::ITensor& output = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[3])->GetTensor();
131 
132  // Get the batch_size and the num_units from the cellStateIn dimensions
133  const TensorInfo& inputTensorInfo = info.m_InputTensorInfos[2];
134  const unsigned int batch_size = boost::numeric_cast<unsigned int>(inputTensorInfo.GetShape()[0]);
135  const unsigned int num_units = boost::numeric_cast<unsigned int>(inputTensorInfo.GetShape()[1]);
136 
137  m_ScratchBuffer = std::make_unique<arm_compute::Tensor>();
138  if (m_Data.m_Parameters.m_CifgEnabled)
139  {
140  // 2D tensor with dimensions [num_units * 3, batch_size] with CIFG
141  armnn::TensorInfo scratchBuffer1({ batch_size, num_units * 3 }, DataType::Float32);
142  BuildArmComputeTensor(*m_ScratchBuffer, scratchBuffer1);
143  }
144  else
145  {
146  // scratch_buffer [num_units * 4, batch_size] without CIFG
147  armnn::TensorInfo scratchBuffer2({ batch_size, num_units * 4 }, DataType::Float32);
148  BuildArmComputeTensor(*m_ScratchBuffer, scratchBuffer2);
149  }
150 
151  float cell_threshold = m_Data.m_Parameters.m_ClippingThresCell;
152  float projection_threshold = m_Data.m_Parameters.m_ClippingThresProj;
153 
154  // for preparing the object for the class ActivationLayerInfo, we need to consider 5 situations
155  arm_compute::ActivationLayerInfo activationLayerInfo;
156  if (m_Data.m_Parameters.m_ActivationFunc == 0)
157  {
158  // no activation, do nothing
159  }
160  else if (m_Data.m_Parameters.m_ActivationFunc == 1)
161  {
162  activationLayerInfo = arm_compute::ActivationLayerInfo(
163  arm_compute::ActivationLayerInfo::ActivationFunction::RELU);
164  }
165  else if (m_Data.m_Parameters.m_ActivationFunc == 3)
166  {
167  activationLayerInfo = arm_compute::ActivationLayerInfo(
168  arm_compute::ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.0);
169  }
170  else if (m_Data.m_Parameters.m_ActivationFunc == 4)
171  {
172  activationLayerInfo = arm_compute::ActivationLayerInfo(
173  arm_compute::ActivationLayerInfo::ActivationFunction::TANH, 1.0, 1.0);
174  }
175  else if (m_Data.m_Parameters.m_ActivationFunc == 6)
176  {
177  activationLayerInfo = arm_compute::ActivationLayerInfo(
178  arm_compute::ActivationLayerInfo::ActivationFunction::LOGISTIC);
179  }
180  else
181  {
182  throw armnn::Exception("Wrong Type of Activation Function!");
183  }
184 
185 
186  m_LstmLayer.configure(&input, m_InputToForgetWeightsTensor.get(), m_InputToCellWeightsTensor.get(),
187  m_InputToOutputWeightsTensor.get(), m_RecurrentToForgetWeightsTensor.get(),
188  m_RecurrentToCellWeightsTensor.get(), m_RecurrentToOutputWeightsTensor.get(),
189  m_ForgetGateBiasTensor.get(), m_CellBiasTensor.get(), m_OutputGateBiasTensor.get(),
190  &output_state_in, &cell_state_in, m_ScratchBuffer.get(), &output_state_out,
191  &cell_state_out, &output, lstm_param, activationLayerInfo,
192  cell_threshold, projection_threshold);
193 
194  armcomputetensorutils::InitialiseArmComputeTensorEmpty(*m_ScratchBuffer);
195 
196  InitializeArmComputeTensorData(*m_InputToForgetWeightsTensor,
197  m_Data.m_InputToForgetWeights);
198  InitializeArmComputeTensorData(*m_InputToCellWeightsTensor,
199  m_Data.m_InputToCellWeights);
200  InitializeArmComputeTensorData(*m_InputToOutputWeightsTensor,
201  m_Data.m_InputToOutputWeights);
202  InitializeArmComputeTensorData(*m_RecurrentToForgetWeightsTensor,
203  m_Data.m_RecurrentToForgetWeights);
204  InitializeArmComputeTensorData(*m_RecurrentToCellWeightsTensor,
205  m_Data.m_RecurrentToCellWeights);
206  InitializeArmComputeTensorData(*m_RecurrentToOutputWeightsTensor,
207  m_Data.m_RecurrentToOutputWeights);
208  InitializeArmComputeTensorData(*m_ForgetGateBiasTensor,
209  m_Data.m_ForgetGateBias);
210  InitializeArmComputeTensorData(*m_CellBiasTensor,
211  m_Data.m_CellBias);
212  InitializeArmComputeTensorData(*m_OutputGateBiasTensor,
213  m_Data.m_OutputGateBias);
214 
215  if (!m_Data.m_Parameters.m_CifgEnabled)
216  {
217  InitializeArmComputeTensorData(*m_InputToInputWeightsTensor,
218  m_Data.m_InputToInputWeights);
219  InitializeArmComputeTensorData(*m_RecurrentToInputWeightsTensor,
220  m_Data.m_RecurrentToInputWeights);
221  if (m_Data.m_CellToInputWeights != nullptr)
222  {
223  InitializeArmComputeTensorData(*m_CellToInputWeightsTensor,
224  m_Data.m_CellToInputWeights);
225  }
226  InitializeArmComputeTensorData(*m_InputGateBiasTensor,
227  m_Data.m_InputGateBias);
228  }
229 
230  if (m_Data.m_Parameters.m_ProjectionEnabled)
231  {
232  InitializeArmComputeTensorData(*m_ProjectionWeightsTensor,
233  m_Data.m_ProjectionWeights);
234  if (m_Data.m_ProjectionBias != nullptr)
235  {
236  InitializeArmComputeTensorData(*m_ProjectionBiasTensor,
237  m_Data.m_ProjectionBias);
238  }
239  }
240 
241  if (m_Data.m_Parameters.m_PeepholeEnabled)
242  {
243  InitializeArmComputeTensorData(*m_CellToForgetWeightsTensor,
244  m_Data.m_CellToForgetWeights);
245  InitializeArmComputeTensorData(*m_CellToOutputWeightsTensor,
246  m_Data.m_CellToOutputWeights);
247  }
248 
249  if (m_Data.m_Parameters.m_LayerNormEnabled)
250  {
251  if (!m_Data.m_Parameters.m_CifgEnabled)
252  {
253  InitializeArmComputeTensorData(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights);
254  }
255  InitializeArmComputeTensorData(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights);
256  InitializeArmComputeTensorData(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights);
257  InitializeArmComputeTensorData(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights);
258  }
259 
260  // Force Compute Library to perform the necessary copying and reshaping, after which
261  // delete all the input tensors that will no longer be needed
262  m_LstmLayer.prepare();
263  FreeUnusedTensors();
264 }
armnn::BaseWorkload::m_Data
const QueueDescriptor m_Data
Definition: Workload.hpp:46
armnn::numeric_cast
std::enable_if_t< std::is_unsigned< Source >::value &&std::is_unsigned< Dest >::value, Dest > numeric_cast(Source source)
Definition: NumericCast.hpp:33
armnn::InitializeArmComputeTensorData
void InitializeArmComputeTensorData(arm_compute::Tensor &tensor, const ConstCpuTensorHandle *handle)
Definition: NeonWorkloadUtils.hpp:35
armnn::QueueDescriptor::m_Outputs
std::vector< ITensorHandle * > m_Outputs
Definition: WorkloadData.hpp:31
armnn::Exception
Base class for all ArmNN exceptions so that users can filter to just those.
Definition: Exceptions.hpp:46
armnn::QueueDescriptor::m_Inputs
std::vector< ITensorHandle * > m_Inputs
Definition: WorkloadData.hpp:30

Member Function Documentation

◆ Execute()

void Execute ( ) const
overridevirtual

Implements IWorkload.

Definition at line 266 of file NeonLstmFloatWorkload.cpp.

267 {
268  m_LstmLayer.run();
269 }
The documentation for this class was generated from the following files: