ArmNN
 20.08
Public Member Functions | List of all members
NeonQLstmWorkload Class Reference

#include <NeonQLstmWorkload.hpp>

Inheritance diagram for NeonQLstmWorkload:
BaseWorkload< QLstmQueueDescriptor > IWorkload

Public Member Functions

 NeonQLstmWorkload (const QLstmQueueDescriptor &descriptor, const WorkloadInfo &info)
 
virtual void Execute () const override
 
- Public Member Functions inherited from BaseWorkload< QLstmQueueDescriptor >
 BaseWorkload (const QLstmQueueDescriptor &descriptor, const WorkloadInfo &info)
 
void PostAllocationConfigure () override
 
const QLstmQueueDescriptor & 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< QLstmQueueDescriptor >
const QLstmQueueDescriptor m_Data
 
const profiling::ProfilingGuid m_Guid
 

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  : BaseWorkload<QLstmQueueDescriptor>(descriptor, info)
19 {
20  arm_compute::LSTMParams<arm_compute::ITensor> qLstmParams;
21 
22  // Mandatory params
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  // Create tensors for optional params if they are enabled
51  if (m_Data.m_Parameters.m_PeepholeEnabled)
52  {
53  m_CellToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
54 
55  if (!m_Data.m_Parameters.m_CifgEnabled)
56  {
57  // In ACL this is categorised as a CIFG param and not a Peephole param
58  BuildArmComputeTensor(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights->GetTensorInfo());
59  }
60 
61  m_CellToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
62  BuildArmComputeTensor(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights->GetTensorInfo());
63 
64  m_CellToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
65  BuildArmComputeTensor(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights->GetTensorInfo());
66 
67  // Set Peephole params
68  qLstmParams.set_peephole_params(m_CellToForgetWeightsTensor.get(),
69  m_CellToOutputWeightsTensor.get());
70  }
71 
72  if (m_Data.m_Parameters.m_ProjectionEnabled)
73  {
74  m_ProjectionWeightsTensor = std::make_unique<arm_compute::Tensor>();
75  BuildArmComputeTensor(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights->GetTensorInfo());
76 
77  m_ProjectionBiasTensor = std::make_unique<arm_compute::Tensor>();
78  if (m_Data.m_ProjectionBias != nullptr)
79  {
80  BuildArmComputeTensor(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias->GetTensorInfo());
81  }
82 
83  // Set projection params
84  qLstmParams.set_projection_params(
85  m_ProjectionWeightsTensor.get(),
86  m_Data.m_ProjectionBias != nullptr ? m_ProjectionBiasTensor.get() : nullptr);
87  }
88 
89  if (m_Data.m_Parameters.m_LayerNormEnabled)
90  {
91  m_InputLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
92 
93  if (!m_Data.m_Parameters.m_CifgEnabled)
94  {
95  BuildArmComputeTensor(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights->GetTensorInfo());
96  }
97 
98  m_ForgetLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
99  BuildArmComputeTensor(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights->GetTensorInfo());
100 
101  m_CellLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
102  BuildArmComputeTensor(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights->GetTensorInfo());
103 
104  m_OutputLayerNormWeightsTensor = std::make_unique<arm_compute::Tensor>();
105  BuildArmComputeTensor(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights->GetTensorInfo());
106 
107  // Set layer norm params
108  qLstmParams.set_layer_normalization_params(
109  m_Data.m_InputLayerNormWeights != nullptr ? m_InputLayerNormWeightsTensor.get() : nullptr,
110  m_ForgetLayerNormWeightsTensor.get(),
111  m_CellLayerNormWeightsTensor.get(),
112  m_OutputLayerNormWeightsTensor.get());
113  }
114 
115  if (!m_Data.m_Parameters.m_CifgEnabled)
116  {
117  m_InputToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
118  BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo());
119 
120  m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
121  BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo());
122 
123  m_InputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
124  BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo());
125 
126  // Set CIFG params
127  qLstmParams.set_cifg_params(
128  m_InputToInputWeightsTensor.get(),
129  m_RecurrentToInputWeightsTensor.get(),
130  m_Data.m_CellToInputWeights != nullptr ? m_CellToInputWeightsTensor.get() : nullptr,
131  m_InputGateBiasTensor.get());
132  }
133 
134  // Input/Output tensors
135  const arm_compute::ITensor& input = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
136  const arm_compute::ITensor& outputStateIn = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
137  const arm_compute::ITensor& cellStateIn = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();
138 
139  arm_compute::ITensor& outputStateOut = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
140  arm_compute::ITensor& cellStateOut = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[1])->GetTensor();
141  arm_compute::ITensor& output = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[2])->GetTensor();
142 
143  // Set scalar descriptor params
144  qLstmParams.set_cell_clip_params(m_Data.m_Parameters.m_CellClip);
145  qLstmParams.set_projection_clip_params(m_Data.m_Parameters.m_ProjectionClip);
146  qLstmParams.set_hidden_state_params(m_Data.m_Parameters.m_HiddenStateZeroPoint,
147  m_Data.m_Parameters.m_HiddenStateScale);
148  qLstmParams.set_matmul_scale_params(m_Data.m_Parameters.m_InputIntermediateScale,
149  m_Data.m_Parameters.m_ForgetIntermediateScale,
150  m_Data.m_Parameters.m_CellIntermediateScale,
151  m_Data.m_Parameters.m_OutputIntermediateScale);
152 
153  // QLSTM NEON configure
154  m_QLstmLayer.configure(&input,
155  m_InputToForgetWeightsTensor.get(),
156  m_InputToCellWeightsTensor.get(),
157  m_InputToOutputWeightsTensor.get(),
158  m_RecurrentToForgetWeightsTensor.get(),
159  m_RecurrentToCellWeightsTensor.get(),
160  m_RecurrentToOutputWeightsTensor.get(),
161  m_ForgetGateBiasTensor.get(),
162  m_CellBiasTensor.get(),
163  m_OutputGateBiasTensor.get(),
164  &cellStateIn,
165  &outputStateIn,
166  &cellStateOut,
167  &outputStateOut,
168  &output,
169  qLstmParams);
170 
171  // Initialise ACL tensor data for mandatory params
172  InitializeArmComputeTensorData(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights);
173  InitializeArmComputeTensorData(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights);
174  InitializeArmComputeTensorData(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights);
175 
176  InitializeArmComputeTensorData(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights);
177  InitializeArmComputeTensorData(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights);
178  InitializeArmComputeTensorData(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights);
179 
180  InitializeArmComputeTensorData(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias);
181  InitializeArmComputeTensorData(*m_CellBiasTensor, m_Data.m_CellBias);
182  InitializeArmComputeTensorData(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias);
183 
184  // Initialise ACL tensor data for optional params
185  if (!m_Data.m_Parameters.m_CifgEnabled)
186  {
187  InitializeArmComputeTensorData(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights);
188  InitializeArmComputeTensorData(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights);
189  InitializeArmComputeTensorData(*m_InputGateBiasTensor, m_Data.m_InputGateBias);
190  }
191 
192  if (m_Data.m_Parameters.m_ProjectionEnabled)
193  {
194  InitializeArmComputeTensorData(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights);
195 
196  if (m_Data.m_ProjectionBias != nullptr)
197  {
198  InitializeArmComputeTensorData(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias);
199  }
200  }
201 
202  if (m_Data.m_Parameters.m_PeepholeEnabled)
203  {
204  if (!m_Data.m_Parameters.m_CifgEnabled)
205  {
206  InitializeArmComputeTensorData(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights);
207  }
208 
209  InitializeArmComputeTensorData(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights);
210  InitializeArmComputeTensorData(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights);
211  }
212 
213  if (m_Data.m_Parameters.m_LayerNormEnabled)
214  {
215  if (!m_Data.m_Parameters.m_CifgEnabled)
216  {
217  InitializeArmComputeTensorData(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights);
218  }
219 
220  InitializeArmComputeTensorData(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights);
221  InitializeArmComputeTensorData(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights);
222  InitializeArmComputeTensorData(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights);
223  }
224 
225  // QLSTM NEON prepare
226  m_QLstmLayer.prepare();
227 
228  FreeUnusedTensors();
229 }
armnn::QLstmQueueDescriptor::m_CellToForgetWeights
const ConstCpuTensorHandle * m_CellToForgetWeights
Definition: WorkloadData.hpp:570
armnn::QLstmQueueDescriptor::m_ProjectionWeights
const ConstCpuTensorHandle * m_ProjectionWeights
Definition: WorkloadData.hpp:576
armnn::QLstmDescriptor::m_PeepholeEnabled
bool m_PeepholeEnabled
Enable/disable peephole.
Definition: Descriptors.hpp:1176
armnn::QLstmQueueDescriptor::m_ProjectionBias
const ConstCpuTensorHandle * m_ProjectionBias
Definition: WorkloadData.hpp:577
armnn::QLstmDescriptor::m_HiddenStateScale
float m_HiddenStateScale
Hidden State quantization scale.
Definition: Descriptors.hpp:1192
armnn::QLstmQueueDescriptor::m_ForgetLayerNormWeights
const ConstCpuTensorHandle * m_ForgetLayerNormWeights
Definition: WorkloadData.hpp:579
armnn::BaseWorkload< QLstmQueueDescriptor >::m_Data
const QLstmQueueDescriptor m_Data
Definition: Workload.hpp:46
armnn::QLstmDescriptor::m_OutputIntermediateScale
float m_OutputIntermediateScale
Output intermediate quantization scale.
Definition: Descriptors.hpp:1188
armnn::QLstmQueueDescriptor::m_CellLayerNormWeights
const ConstCpuTensorHandle * m_CellLayerNormWeights
Definition: WorkloadData.hpp:580
armnn::QLstmQueueDescriptor::m_RecurrentToCellWeights
const ConstCpuTensorHandle * m_RecurrentToCellWeights
Definition: WorkloadData.hpp:567
armnn::QLstmQueueDescriptor::m_RecurrentToInputWeights
const ConstCpuTensorHandle * m_RecurrentToInputWeights
Definition: WorkloadData.hpp:565
armnn::QLstmQueueDescriptor::m_OutputGateBias
const ConstCpuTensorHandle * m_OutputGateBias
Definition: WorkloadData.hpp:575
armnn::QLstmQueueDescriptor::m_CellBias
const ConstCpuTensorHandle * m_CellBias
Definition: WorkloadData.hpp:574
armnn::QLstmQueueDescriptor::m_CellToOutputWeights
const ConstCpuTensorHandle * m_CellToOutputWeights
Definition: WorkloadData.hpp:571
armnn::QLstmQueueDescriptor::m_OutputLayerNormWeights
const ConstCpuTensorHandle * m_OutputLayerNormWeights
Definition: WorkloadData.hpp:581
armnn::QLstmDescriptor::m_LayerNormEnabled
bool m_LayerNormEnabled
Enable/disable layer normalization.
Definition: Descriptors.hpp:1180
armnn::QLstmQueueDescriptor::m_InputToForgetWeights
const ConstCpuTensorHandle * m_InputToForgetWeights
Definition: WorkloadData.hpp:562
armnn::QLstmDescriptor::m_ProjectionClip
float m_ProjectionClip
Clipping threshold value for the projection.
Definition: Descriptors.hpp:1172
armnn::QLstmDescriptor::m_InputIntermediateScale
float m_InputIntermediateScale
Input intermediate quantization scale.
Definition: Descriptors.hpp:1182
armnn::QLstmDescriptor::m_ForgetIntermediateScale
float m_ForgetIntermediateScale
Forget intermediate quantization scale.
Definition: Descriptors.hpp:1184
armnn::InitializeArmComputeTensorData
void InitializeArmComputeTensorData(arm_compute::Tensor &tensor, const ConstCpuTensorHandle *handle)
Definition: NeonWorkloadUtils.hpp:35
armnn::QLstmQueueDescriptor::m_CellToInputWeights
const ConstCpuTensorHandle * m_CellToInputWeights
Definition: WorkloadData.hpp:569
armnn::QLstmDescriptor::m_CellClip
float m_CellClip
Clipping threshold value for the cell state.
Definition: Descriptors.hpp:1170
armnn::QLstmQueueDescriptor::m_RecurrentToOutputWeights
const ConstCpuTensorHandle * m_RecurrentToOutputWeights
Definition: WorkloadData.hpp:568
armnn::QueueDescriptor::m_Outputs
std::vector< ITensorHandle * > m_Outputs
Definition: WorkloadData.hpp:31
armnn::QLstmDescriptor::m_ProjectionEnabled
bool m_ProjectionEnabled
Enable/disable the projection layer.
Definition: Descriptors.hpp:1178
armnn::QLstmQueueDescriptor::m_InputGateBias
const ConstCpuTensorHandle * m_InputGateBias
Definition: WorkloadData.hpp:572
armnn::QueueDescriptor::m_Inputs
std::vector< ITensorHandle * > m_Inputs
Definition: WorkloadData.hpp:30
armnn::QLstmQueueDescriptor::m_InputLayerNormWeights
const ConstCpuTensorHandle * m_InputLayerNormWeights
Definition: WorkloadData.hpp:578
armnn::QLstmQueueDescriptor::m_RecurrentToForgetWeights
const ConstCpuTensorHandle * m_RecurrentToForgetWeights
Definition: WorkloadData.hpp:566
armnn::QLstmQueueDescriptor::m_ForgetGateBias
const ConstCpuTensorHandle * m_ForgetGateBias
Definition: WorkloadData.hpp:573
armnn::QLstmDescriptor::m_CellIntermediateScale
float m_CellIntermediateScale
Cell intermediate quantization scale.
Definition: Descriptors.hpp:1186
armnn::QLstmQueueDescriptor::m_InputToOutputWeights
const ConstCpuTensorHandle * m_InputToOutputWeights
Definition: WorkloadData.hpp:564
armnn::QLstmDescriptor::m_CifgEnabled
bool m_CifgEnabled
Enable/disable CIFG (coupled input & forget gate).
Definition: Descriptors.hpp:1174
armnn::ConstCpuTensorHandle::GetTensorInfo
const TensorInfo & GetTensorInfo() const
Definition: CpuTensorHandle.hpp:37
armnn::QLstmQueueDescriptor::m_InputToInputWeights
const ConstCpuTensorHandle * m_InputToInputWeights
Definition: WorkloadData.hpp:561
armnn::QLstmDescriptor::m_HiddenStateZeroPoint
int32_t m_HiddenStateZeroPoint
Hidden State zero point.
Definition: Descriptors.hpp:1190
armnn::QLstmQueueDescriptor::m_InputToCellWeights
const ConstCpuTensorHandle * m_InputToCellWeights
Definition: WorkloadData.hpp:563

Member Function Documentation

◆ Execute()

void Execute ( ) const
overridevirtual

Implements IWorkload.

Definition at line 231 of file NeonQLstmWorkload.cpp.

232 {
233  m_QLstmLayer.run();
234 }
The documentation for this class was generated from the following files: