ArmNN
 24.08
Public Member Functions | List of all members
NeonQLstmWorkload Class Reference

#include <NeonQLstmWorkload.hpp>

Inheritance diagram for NeonQLstmWorkload:
[legend]
Collaboration diagram for NeonQLstmWorkload:
[legend]

Public Member Functions

 NeonQLstmWorkload (const QLstmQueueDescriptor &descriptor, const WorkloadInfo &info)
 
virtual void Execute () const override
 
- Public Member Functions inherited from NeonBaseWorkload< QLstmQueueDescriptor >
 NeonBaseWorkload (const QLstmQueueDescriptor &descriptor, const WorkloadInfo &info)
 
void ReplaceInputTensorHandle (ITensorHandle *tensorHandle, unsigned int slot) override
 
void ReplaceOutputTensorHandle (ITensorHandle *tensorHandle, unsigned int slot) override
 
- Public Member Functions inherited from BaseWorkload< QLstmQueueDescriptor >
 BaseWorkload (const QLstmQueueDescriptor &descriptor, const WorkloadInfo &info)
 
virtual const std::string & GetName () const override
 
void ExecuteAsync (ExecutionData &executionData) override
 
void PostAllocationConfigure () override
 
const QLstmQueueDescriptor & GetData () const
 
arm::pipe::ProfilingGuid GetGuid () const final
 
virtual bool SupportsTensorHandleReplacement () const override
 
- Public Member Functions inherited from IWorkload
virtual ~IWorkload ()
 
virtual arm::pipe::ProfilingGuid GetGuid () const =0
 
virtual bool SupportsTensorHandleReplacement () const =0
 
virtual const std::string & GetName () const =0
 
virtual void RegisterDebugCallback (const DebugCallbackFunction &)
 
virtual armnn::Optional< armnn::MemoryRequirements > GetMemoryRequirements ()
 

Additional Inherited Members

- Protected Member Functions inherited from NeonBaseWorkload< QLstmQueueDescriptor >
virtual void Reconfigure ()
 
- Protected Attributes inherited from BaseWorkload< QLstmQueueDescriptor >
QLstmQueueDescriptor m_Data
 
const arm::pipe::ProfilingGuid m_Guid
 
const std::string m_Name
 

Detailed Description

Definition at line 19 of file NeonQLstmWorkload.hpp.

Constructor & Destructor Documentation

◆ NeonQLstmWorkload()

NeonQLstmWorkload ( const QLstmQueueDescriptor &  descriptor,
const WorkloadInfo &  info 
)

Definition at line 17 of file NeonQLstmWorkload.cpp.

18  : NeonBaseWorkload<QLstmQueueDescriptor>(descriptor, info)
19 {
20  // Report Profiling Details
21  ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonQLstmWorkload_Construct",
22  descriptor.m_Parameters,
23  info,
24  this->GetGuid());
25 
26  arm_compute::LSTMParams<arm_compute::ITensor> qLstmParams;
27 
28  // Mandatory params
29  m_InputToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
30  BuildArmComputeTensor(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights->GetTensorInfo());
31 
32  m_InputToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
33  BuildArmComputeTensor(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights->GetTensorInfo());
34 
35  m_InputToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
36  BuildArmComputeTensor(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights->GetTensorInfo());
37 
38  m_RecurrentToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
39  BuildArmComputeTensor(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights->GetTensorInfo());
40 
41  m_RecurrentToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
42  BuildArmComputeTensor(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights->GetTensorInfo());
43 
44  m_RecurrentToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
45  BuildArmComputeTensor(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights->GetTensorInfo());
46 
47  m_ForgetGateBiasTensor = std::make_unique<arm_compute::Tensor>();
48  BuildArmComputeTensor(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias->GetTensorInfo());
49 
50  m_CellBiasTensor = std::make_unique<arm_compute::Tensor>();
51  BuildArmComputeTensor(*m_CellBiasTensor, m_Data.m_CellBias->GetTensorInfo());
52 
53  m_OutputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
54  BuildArmComputeTensor(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias->GetTensorInfo());
55 
56  // Create tensors for optional params if they are enabled
57  if (m_Data.m_Parameters.m_PeepholeEnabled)
58  {
59  m_CellToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
60 
61  if (!m_Data.m_Parameters.m_CifgEnabled)
62  {
63  // In ACL this is categorised as a CIFG param and not a Peephole param
64  BuildArmComputeTensor(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights->GetTensorInfo());
65  }
66 
67  m_CellToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
68  BuildArmComputeTensor(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights->GetTensorInfo());
69 
70  m_CellToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
71  BuildArmComputeTensor(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights->GetTensorInfo());
72 
73  // Set Peephole params
74  qLstmParams.set_peephole_params(m_CellToForgetWeightsTensor.get(),
75  m_CellToOutputWeightsTensor.get());
76  }
77 
78  if (m_Data.m_Parameters.m_ProjectionEnabled)
79  {
80  m_ProjectionWeightsTensor = std::make_unique<arm_compute::Tensor>();
81  BuildArmComputeTensor(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights->GetTensorInfo());
82 
83  m_ProjectionBiasTensor = std::make_unique<arm_compute::Tensor>();
84  if (m_Data.m_ProjectionBias != nullptr)
85  {
86  BuildArmComputeTensor(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias->GetTensorInfo());
87  }
88 
89  // Set projection params
90  qLstmParams.set_projection_params(
91  m_ProjectionWeightsTensor.get(),
92  m_Data.m_ProjectionBias != nullptr ? m_ProjectionBiasTensor.get() : nullptr);
93  }
94 
95  if (m_Data.m_Parameters.m_LayerNormEnabled)
96  {
97  m_InputLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
98 
99  if (!m_Data.m_Parameters.m_CifgEnabled)
100  {
101  BuildArmComputeTensor(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights->GetTensorInfo());
102  }
103 
104  m_ForgetLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
105  BuildArmComputeTensor(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights->GetTensorInfo());
106 
107  m_CellLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
108  BuildArmComputeTensor(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights->GetTensorInfo());
109 
110  m_OutputLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
111  BuildArmComputeTensor(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights->GetTensorInfo());
112 
113  // Set layer norm params
114  qLstmParams.set_layer_normalization_params(
115  m_Data.m_InputLayerNormWeights != nullptr ? m_InputLayerNormWeightsTensor.get() : nullptr,
116  m_ForgetLayerNormWeightsTensor.get(),
117  m_CellLayerNormWeightsTensor.get(),
118  m_OutputLayerNormWeightsTensor.get());
119  }
120 
121  if (!m_Data.m_Parameters.m_CifgEnabled)
122  {
123  m_InputToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
124  BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo());
125 
126  m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
127  BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo());
128 
129  m_InputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
130  BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo());
131 
132  // Set CIFG params
133  qLstmParams.set_cifg_params(
134  m_InputToInputWeightsTensor.get(),
135  m_RecurrentToInputWeightsTensor.get(),
136  m_Data.m_CellToInputWeights != nullptr ? m_CellToInputWeightsTensor.get() : nullptr,
137  m_InputGateBiasTensor.get());
138  }
139 
140  // Input/Output tensors
141  const arm_compute::ITensor& input = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
142  arm_compute::ITensor& outputStateIn = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
143  const arm_compute::ITensor& cellStateIn = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();
144 
145  arm_compute::ITensor& outputStateOut = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
146  arm_compute::ITensor& cellStateOut = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[1])->GetTensor();
147  arm_compute::ITensor& output = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[2])->GetTensor();
148 
149  // Set scalar descriptor params
150  qLstmParams.set_cell_clip_params(m_Data.m_Parameters.m_CellClip);
151  qLstmParams.set_projection_clip_params(m_Data.m_Parameters.m_ProjectionClip);
152  qLstmParams.set_hidden_state_params(m_Data.m_Parameters.m_HiddenStateZeroPoint,
153  m_Data.m_Parameters.m_HiddenStateScale);
154  qLstmParams.set_matmul_scale_params(m_Data.m_Parameters.m_InputIntermediateScale,
155  m_Data.m_Parameters.m_ForgetIntermediateScale,
156  m_Data.m_Parameters.m_CellIntermediateScale,
157  m_Data.m_Parameters.m_OutputIntermediateScale);
158 
159  // QLSTM NEON configure
160  m_QLstmLayer.configure(&input,
161  m_InputToForgetWeightsTensor.get(),
162  m_InputToCellWeightsTensor.get(),
163  m_InputToOutputWeightsTensor.get(),
164  m_RecurrentToForgetWeightsTensor.get(),
165  m_RecurrentToCellWeightsTensor.get(),
166  m_RecurrentToOutputWeightsTensor.get(),
167  m_ForgetGateBiasTensor.get(),
168  m_CellBiasTensor.get(),
169  m_OutputGateBiasTensor.get(),
170  &cellStateIn,
171  &outputStateIn,
172  &cellStateOut,
173  &outputStateOut,
174  &output,
175  qLstmParams);
176 
177  // Initialise ACL tensor data for mandatory params
178  InitializeArmComputeTensorData(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights);
179  InitializeArmComputeTensorData(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights);
180  InitializeArmComputeTensorData(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights);
181 
182  InitializeArmComputeTensorData(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights);
183  InitializeArmComputeTensorData(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights);
184  InitializeArmComputeTensorData(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights);
185 
186  InitializeArmComputeTensorData(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias);
187  InitializeArmComputeTensorData(*m_CellBiasTensor, m_Data.m_CellBias);
188  InitializeArmComputeTensorData(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias);
189 
190  // Initialise ACL tensor data for optional params
191  if (!m_Data.m_Parameters.m_CifgEnabled)
192  {
193  InitializeArmComputeTensorData(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights);
194  InitializeArmComputeTensorData(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights);
195  InitializeArmComputeTensorData(*m_InputGateBiasTensor, m_Data.m_InputGateBias);
196  }
197 
198  if (m_Data.m_Parameters.m_ProjectionEnabled)
199  {
200  InitializeArmComputeTensorData(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights);
201 
202  if (m_Data.m_ProjectionBias != nullptr)
203  {
204  InitializeArmComputeTensorData(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias);
205  }
206  }
207 
208  if (m_Data.m_Parameters.m_PeepholeEnabled)
209  {
210  if (!m_Data.m_Parameters.m_CifgEnabled)
211  {
212  InitializeArmComputeTensorData(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights);
213  }
214 
215  InitializeArmComputeTensorData(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights);
216  InitializeArmComputeTensorData(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights);
217  }
218 
219  if (m_Data.m_Parameters.m_LayerNormEnabled)
220  {
221  if (!m_Data.m_Parameters.m_CifgEnabled)
222  {
223  InitializeArmComputeTensorData(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights);
224  }
225 
226  InitializeArmComputeTensorData(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights);
227  InitializeArmComputeTensorData(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights);
228  InitializeArmComputeTensorData(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights);
229  }
230 
231  // QLSTM NEON prepare
232  m_QLstmLayer.prepare();
233 
234  FreeUnusedTensors();
235 }

References ARMNN_REPORT_PROFILING_WORKLOAD_DESC, armnn::info, and QueueDescriptorWithParameters< LayerDescriptor >::m_Parameters.

Member Function Documentation

◆ Execute()

void Execute ( ) const
overridevirtual

Implements IWorkload.

Definition at line 237 of file NeonQLstmWorkload.cpp.

238 {
239  ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID("NeonQLstmWorkload_Execute");
240  m_QLstmLayer.run();
241 }

References ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID.

The documentation for this class was generated from the following files:
armnn::QLstmDescriptor::m_ForgetIntermediateScale
float m_ForgetIntermediateScale
Forget intermediate quantization scale.
Definition: Descriptors.hpp:1428
armnn::QLstmDescriptor::m_ProjectionEnabled
bool m_ProjectionEnabled
Enable/disable the projection layer.
Definition: Descriptors.hpp:1422
armnn::QLstmQueueDescriptor::m_InputToForgetWeights
const ConstTensorHandle * m_InputToForgetWeights
Definition: WorkloadData.hpp:590
armnn::QLstmQueueDescriptor::m_InputLayerNormWeights
const ConstTensorHandle * m_InputLayerNormWeights
Definition: WorkloadData.hpp:606
armnn::QLstmDescriptor::m_InputIntermediateScale
float m_InputIntermediateScale
Input intermediate quantization scale.
Definition: Descriptors.hpp:1426
armnn::ConstTensorHandle::GetTensorInfo
const TensorInfo & GetTensorInfo() const
Definition: TensorHandle.hpp:40
armnn::QLstmQueueDescriptor::m_InputToInputWeights
const ConstTensorHandle * m_InputToInputWeights
Definition: WorkloadData.hpp:589
armnn::QLstmDescriptor::m_CellIntermediateScale
float m_CellIntermediateScale
Cell intermediate quantization scale.
Definition: Descriptors.hpp:1430
armnn::QLstmQueueDescriptor::m_ProjectionWeights
const ConstTensorHandle * m_ProjectionWeights
Definition: WorkloadData.hpp:604
armnn::QLstmQueueDescriptor::m_ForgetLayerNormWeights
const ConstTensorHandle * m_ForgetLayerNormWeights
Definition: WorkloadData.hpp:607
armnn::QLstmQueueDescriptor::m_RecurrentToOutputWeights
const ConstTensorHandle * m_RecurrentToOutputWeights
Definition: WorkloadData.hpp:596
armnn::InitializeArmComputeTensorData
void InitializeArmComputeTensorData(arm_compute::Tensor &tensor, TensorInfo tensorInfo, const ITensorHandle *handle)
Definition: NeonWorkloadUtils.hpp:69
armnn::QLstmQueueDescriptor::m_CellToOutputWeights
const ConstTensorHandle * m_CellToOutputWeights
Definition: WorkloadData.hpp:599
armnn::QueueDescriptorWithParameters::m_Parameters
LayerDescriptor m_Parameters
Definition: WorkloadData.hpp:66
armnn::QLstmQueueDescriptor::m_CellToForgetWeights
const ConstTensorHandle * m_CellToForgetWeights
Definition: WorkloadData.hpp:598
armnn::QLstmQueueDescriptor::m_OutputLayerNormWeights
const ConstTensorHandle * m_OutputLayerNormWeights
Definition: WorkloadData.hpp:609
armnn::QLstmQueueDescriptor::m_InputGateBias
const ConstTensorHandle * m_InputGateBias
Definition: WorkloadData.hpp:600
armnn::QLstmQueueDescriptor::m_OutputGateBias
const ConstTensorHandle * m_OutputGateBias
Definition: WorkloadData.hpp:603
armnn::QLstmQueueDescriptor::m_ForgetGateBias
const ConstTensorHandle * m_ForgetGateBias
Definition: WorkloadData.hpp:601
armnn::QLstmDescriptor::m_ProjectionClip
float m_ProjectionClip
Clipping threshold value for the projection.
Definition: Descriptors.hpp:1416
armnn::QLstmQueueDescriptor::m_RecurrentToInputWeights
const ConstTensorHandle * m_RecurrentToInputWeights
Definition: WorkloadData.hpp:593
armnn::QLstmDescriptor::m_OutputIntermediateScale
float m_OutputIntermediateScale
Output intermediate quantization scale.
Definition: Descriptors.hpp:1432
armnn::QueueDescriptor::m_Outputs
std::vector< ITensorHandle * > m_Outputs
Definition: WorkloadData.hpp:27
armnn::QLstmQueueDescriptor::m_RecurrentToForgetWeights
const ConstTensorHandle * m_RecurrentToForgetWeights
Definition: WorkloadData.hpp:594
ARMNN_REPORT_PROFILING_WORKLOAD_DESC
#define ARMNN_REPORT_PROFILING_WORKLOAD_DESC(name, desc, infos, guid)
Definition: Profiling.hpp:227
armnn::BaseWorkload< QLstmQueueDescriptor >::m_Data
QLstmQueueDescriptor m_Data
Definition: Workload.hpp:89
armnn::QLstmQueueDescriptor::m_InputToOutputWeights
const ConstTensorHandle * m_InputToOutputWeights
Definition: WorkloadData.hpp:592
armnn::QLstmDescriptor::m_HiddenStateZeroPoint
int32_t m_HiddenStateZeroPoint
Hidden State zero point.
Definition: Descriptors.hpp:1434
armnn::QLstmDescriptor::m_CifgEnabled
bool m_CifgEnabled
Enable/disable CIFG (coupled input & forget gate).
Definition: Descriptors.hpp:1418
armnn::QLstmDescriptor::m_HiddenStateScale
float m_HiddenStateScale
Hidden State quantization scale.
Definition: Descriptors.hpp:1436
armnn::QLstmQueueDescriptor::m_ProjectionBias
const ConstTensorHandle * m_ProjectionBias
Definition: WorkloadData.hpp:605
armnn::QLstmQueueDescriptor::m_CellToInputWeights
const ConstTensorHandle * m_CellToInputWeights
Definition: WorkloadData.hpp:597
armnn::QLstmDescriptor::m_CellClip
float m_CellClip
Clipping threshold value for the cell state.
Definition: Descriptors.hpp:1414
armnn::QLstmDescriptor::m_LayerNormEnabled
bool m_LayerNormEnabled
Enable/disable layer normalization.
Definition: Descriptors.hpp:1424
armnn::QLstmQueueDescriptor::m_CellBias
const ConstTensorHandle * m_CellBias
Definition: WorkloadData.hpp:602
ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID
#define ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID(label)
Creates a profiling event that uses GetGuid() and GetName() from the calling class.
Definition: NeonWorkloadUtils.hpp:33
armnn::QLstmQueueDescriptor::m_InputToCellWeights
const ConstTensorHandle * m_InputToCellWeights
Definition: WorkloadData.hpp:591
armnn::QLstmDescriptor::m_PeepholeEnabled
bool m_PeepholeEnabled
Enable/disable peephole.
Definition: Descriptors.hpp:1420
armnn::QLstmQueueDescriptor::m_CellLayerNormWeights
const ConstTensorHandle * m_CellLayerNormWeights
Definition: WorkloadData.hpp:608
armnn::QueueDescriptor::m_Inputs
std::vector< ITensorHandle * > m_Inputs
Definition: WorkloadData.hpp:26
armnn::QLstmQueueDescriptor::m_RecurrentToCellWeights
const ConstTensorHandle * m_RecurrentToCellWeights
Definition: WorkloadData.hpp:595