ArmNN
 24.08
ClQuantizedLstmWorkload.cpp
Go to the documentation of this file.
1 //
2 // Copyright © 2019-2023 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 
7 #include "ClWorkloadUtils.hpp"
8 
11 #include <cl/ClTensorHandle.hpp>
12 
13 namespace armnn
14 {
15 
16 using namespace armcomputetensorutils;
17 
19  const TensorInfo& previousOutputIn, const TensorInfo& cellStateOut,
20  const TensorInfo& output,
21  const QuantizedLstmInputParamsInfo& paramsInfo)
22 {
23  // Inputs
24  const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input);
25  const arm_compute::TensorInfo aclPreviousCellStateInInfo = BuildArmComputeTensorInfo(previousCellStateIn);
26  const arm_compute::TensorInfo aclPreviousOutputInInfo = BuildArmComputeTensorInfo(previousOutputIn);
27 
28  // Outputs
29  const arm_compute::TensorInfo aclCellStateOutInfo = BuildArmComputeTensorInfo(cellStateOut);
30  const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output);
31 
32  // Basic parameters
33  const arm_compute::TensorInfo aclInputToInputWeightsInfo
34  = BuildArmComputeTensorInfo(paramsInfo.GetInputToInputWeights());
35  const arm_compute::TensorInfo aclInputToForgetWeightsInfo
36  = BuildArmComputeTensorInfo(paramsInfo.GetInputToForgetWeights());
37  const arm_compute::TensorInfo aclInputToCellWeightsInfo
38  = BuildArmComputeTensorInfo(paramsInfo.GetInputToCellWeights());
39  const arm_compute::TensorInfo aclInputToOutputWeightsInfo
40  = BuildArmComputeTensorInfo(paramsInfo.GetInputToOutputWeights());
41  const arm_compute::TensorInfo aclRecurrentToInputWeightsInfo
42  = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToInputWeights());
43  const arm_compute::TensorInfo aclRecurrentToForgetWeightsInfo
44  = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToForgetWeights());
45  const arm_compute::TensorInfo aclRecurrentToCellWeightsInfo
46  = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToCellWeights());
47  const arm_compute::TensorInfo aclRecurrentToOutputWeightsInfo
48  = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToOutputWeights());
49  const arm_compute::TensorInfo aclInputGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputGateBias());
50  const arm_compute::TensorInfo aclForgetGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetForgetGateBias());
51  const arm_compute::TensorInfo aclCellBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellBias());
52  const arm_compute::TensorInfo aclOutputGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetOutputGateBias());
53 
54  return arm_compute::CLLSTMLayerQuantized::validate(&aclInputInfo, &aclInputToInputWeightsInfo,
55  &aclInputToForgetWeightsInfo, &aclInputToCellWeightsInfo,
56  &aclInputToOutputWeightsInfo, &aclRecurrentToInputWeightsInfo,
57  &aclRecurrentToForgetWeightsInfo, &aclRecurrentToCellWeightsInfo,
58  &aclRecurrentToOutputWeightsInfo, &aclInputGateBiasInfo,
59  &aclForgetGateBiasInfo, &aclCellBiasInfo, &aclOutputGateBiasInfo,
60  &aclPreviousCellStateInInfo, &aclPreviousOutputInInfo,
61  &aclCellStateOutInfo, &aclOutputInfo);
62 }
63 
65  const WorkloadInfo &info,
66  const arm_compute::CLCompileContext& clCompileContext)
68 {
69  m_InputToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
70  BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo());
71 
72  m_InputToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>();
73  BuildArmComputeTensor(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights->GetTensorInfo());
74 
75  m_InputToCellWeightsTensor = std::make_unique<arm_compute::CLTensor>();
76  BuildArmComputeTensor(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights->GetTensorInfo());
77 
78  m_InputToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
79  BuildArmComputeTensor(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights->GetTensorInfo());
80 
81  m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
82  BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo());
83 
84  m_RecurrentToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>();
85  BuildArmComputeTensor(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights->GetTensorInfo());
86 
87  m_RecurrentToCellWeightsTensor = std::make_unique<arm_compute::CLTensor>();
88  BuildArmComputeTensor(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights->GetTensorInfo());
89 
90  m_RecurrentToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
91  BuildArmComputeTensor(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights->GetTensorInfo());
92 
93  m_InputGateBiasTensor = std::make_unique<arm_compute::CLTensor>();
94  BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo());
95 
96  m_ForgetGateBiasTensor = std::make_unique<arm_compute::CLTensor>();
97  BuildArmComputeTensor(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias->GetTensorInfo());
98 
99  m_CellBiasTensor = std::make_unique<arm_compute::CLTensor>();
100  BuildArmComputeTensor(*m_CellBiasTensor, m_Data.m_CellBias->GetTensorInfo());
101 
102  m_OutputGateBiasTensor = std::make_unique<arm_compute::CLTensor>();
103  BuildArmComputeTensor(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias->GetTensorInfo());
104 
105  const arm_compute::ICLTensor& inputTensor = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
106  arm_compute::ICLTensor& cellStateInTensor = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
107  const arm_compute::ICLTensor& outputStateInTensor = static_cast<IClTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();
108 
109  arm_compute::ICLTensor& cellStateOutTensor = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
110  arm_compute::ICLTensor& outputStateOutTensor = static_cast<IClTensorHandle*>(m_Data.m_Outputs[1])->GetTensor();
111 
112  {
113  ARMNN_SCOPED_PROFILING_EVENT_CL_NAME_GUID("ClQuantizedLstmWorkload_configure");
114  m_QuantizedLstmLayer.configure(clCompileContext, &inputTensor, m_InputToInputWeightsTensor.get(),
115  m_InputToForgetWeightsTensor.get(),
116  m_InputToCellWeightsTensor.get(), m_InputToOutputWeightsTensor.get(),
117  m_RecurrentToInputWeightsTensor.get(), m_RecurrentToForgetWeightsTensor.get(),
118  m_RecurrentToCellWeightsTensor.get(), m_RecurrentToOutputWeightsTensor.get(),
119  m_InputGateBiasTensor.get(), m_ForgetGateBiasTensor.get(),
120  m_CellBiasTensor.get(),
121  m_OutputGateBiasTensor.get(), &cellStateInTensor, &outputStateInTensor,
122  &cellStateOutTensor, &outputStateOutTensor);
123  }
124 
125  InitializeArmComputeClTensorData(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights);
126  InitializeArmComputeClTensorData(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights);
127  InitializeArmComputeClTensorData(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights);
128  InitializeArmComputeClTensorData(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights);
129  InitializeArmComputeClTensorData(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights);
130  InitializeArmComputeClTensorData(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights);
131  InitializeArmComputeClTensorData(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights);
132  InitializeArmComputeClTensorData(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights);
134  InitializeArmComputeClTensorData(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias);
136  InitializeArmComputeClTensorData(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias);
137 
138  m_QuantizedLstmLayer.prepare();
139  FreeUnusedTensors();
140 }
141 
143 {
144  ARMNN_SCOPED_PROFILING_EVENT_CL_NAME_GUID("ClQuantizedLstmWorkload_Execute");
145  RunClFunction(m_QuantizedLstmLayer, CHECK_LOCATION());
146 }
147 
148 void ClQuantizedLstmWorkload::FreeUnusedTensors()
149 {
150  FreeTensorIfUnused(m_InputToInputWeightsTensor);
151  FreeTensorIfUnused(m_InputToForgetWeightsTensor);
152  FreeTensorIfUnused(m_InputToCellWeightsTensor);
153  FreeTensorIfUnused(m_InputToOutputWeightsTensor);
154  FreeTensorIfUnused(m_RecurrentToInputWeightsTensor);
155  FreeTensorIfUnused(m_RecurrentToForgetWeightsTensor);
156  FreeTensorIfUnused(m_RecurrentToCellWeightsTensor);
157  FreeTensorIfUnused(m_RecurrentToOutputWeightsTensor);
158  FreeTensorIfUnused(m_InputGateBiasTensor);
159  FreeTensorIfUnused(m_ForgetGateBiasTensor);
160  FreeTensorIfUnused(m_CellBiasTensor);
161  FreeTensorIfUnused(m_OutputGateBiasTensor);
162 }
163 
164 } // namespace armnn
armnn::QuantizedLstmQueueDescriptor::m_CellBias
const ConstTensorHandle * m_CellBias
Definition: WorkloadData.hpp:645
armnn::ClQuantizedLstmWorkload::ClQuantizedLstmWorkload
ClQuantizedLstmWorkload(const QuantizedLstmQueueDescriptor &descriptor, const WorkloadInfo &info, const arm_compute::CLCompileContext &clCompileContext)
Definition: ClQuantizedLstmWorkload.cpp:64
armnn::QuantizedLstmInputParamsInfo::GetForgetGateBias
const TensorInfo & GetForgetGateBias() const
Definition: QuantizedLstmParams.hpp:203
armnn::QuantizedLstmQueueDescriptor::m_RecurrentToInputWeights
const ConstTensorHandle * m_RecurrentToInputWeights
Definition: WorkloadData.hpp:638
armnn::InitializeArmComputeClTensorData
void InitializeArmComputeClTensorData(arm_compute::CLTensor &clTensor, const ConstTensorHandle *handle)
Definition: ClWorkloadUtils.hpp:124
armnn::QuantizedLstmInputParamsInfo::GetInputToForgetWeights
const TensorInfo & GetInputToForgetWeights() const
Definition: QuantizedLstmParams.hpp:169
armnn::QuantizedLstmQueueDescriptor::m_InputToForgetWeights
const ConstTensorHandle * m_InputToForgetWeights
Definition: WorkloadData.hpp:634
armnn::QuantizedLstmInputParamsInfo::GetRecurrentToForgetWeights
const TensorInfo & GetRecurrentToForgetWeights() const
Definition: QuantizedLstmParams.hpp:186
ClQuantizedLstmWorkload.hpp
armnn::TensorInfo
Definition: Tensor.hpp:152
CHECK_LOCATION
#define CHECK_LOCATION()
Definition: Exceptions.hpp:203
armnn::ClBaseWorkload
Definition: ClBaseWorkload.hpp:13
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::ConstTensorHandle::GetTensorInfo
const TensorInfo & GetTensorInfo() const
Definition: TensorHandle.hpp:40
armnn::QuantizedLstmInputParamsInfo::GetInputGateBias
const TensorInfo & GetInputGateBias() const
Definition: QuantizedLstmParams.hpp:199
armnn::QuantizedLstmInputParamsInfo
Definition: QuantizedLstmParams.hpp:119
armnn::QuantizedLstmQueueDescriptor::m_RecurrentToCellWeights
const ConstTensorHandle * m_RecurrentToCellWeights
Definition: WorkloadData.hpp:640
armnn::QuantizedLstmInputParamsInfo::GetRecurrentToOutputWeights
const TensorInfo & GetRecurrentToOutputWeights() const
Definition: QuantizedLstmParams.hpp:194
armnn::QuantizedLstmInputParamsInfo::GetInputToInputWeights
const TensorInfo & GetInputToInputWeights() const
Definition: QuantizedLstmParams.hpp:165
ARMNN_SCOPED_PROFILING_EVENT_CL_NAME_GUID
#define ARMNN_SCOPED_PROFILING_EVENT_CL_NAME_GUID(label)
Creates a profiling event that uses GetGuid() and GetName() from the calling class.
Definition: ClWorkloadUtils.hpp:36
armnn::QuantizedLstmInputParamsInfo::GetCellBias
const TensorInfo & GetCellBias() const
Definition: QuantizedLstmParams.hpp:207
armnn::QuantizedLstmQueueDescriptor::m_InputToInputWeights
const ConstTensorHandle * m_InputToInputWeights
Definition: WorkloadData.hpp:633
armnn::WorkloadInfo
Contains information about TensorInfos of a layer.
Definition: WorkloadInfo.hpp:16
ClWorkloadUtils.hpp
armnn::QuantizedLstmQueueDescriptor::m_RecurrentToForgetWeights
const ConstTensorHandle * m_RecurrentToForgetWeights
Definition: WorkloadData.hpp:639
armnn::QuantizedLstmInputParamsInfo::GetRecurrentToCellWeights
const TensorInfo & GetRecurrentToCellWeights() const
Definition: QuantizedLstmParams.hpp:190
armnn::QuantizedLstmQueueDescriptor::m_InputToCellWeights
const ConstTensorHandle * m_InputToCellWeights
Definition: WorkloadData.hpp:635
armnn::QuantizedLstmInputParamsInfo::GetInputToOutputWeights
const TensorInfo & GetInputToOutputWeights() const
Definition: QuantizedLstmParams.hpp:177
armnn::BoostLogSeverityMapping::info
@ info
armnn::QueueDescriptor::m_Outputs
std::vector< ITensorHandle * > m_Outputs
Definition: WorkloadData.hpp:27
armnn::QuantizedLstmQueueDescriptor::m_InputToOutputWeights
const ConstTensorHandle * m_InputToOutputWeights
Definition: WorkloadData.hpp:636
armnn::ClQuantizedLstmWorkload::Execute
void Execute() const override
Definition: ClQuantizedLstmWorkload.cpp:142
armnn::QuantizedLstmQueueDescriptor
Definition: WorkloadData.hpp:614
armnn::IClTensorHandle
Definition: IClTensorHandle.hpp:13
TensorHandle.hpp
armnn::Status
Status
Definition: Types.hpp:42
ClTensorHandle.hpp
armnn::BaseWorkload< QuantizedLstmQueueDescriptor >::m_Data
QuantizedLstmQueueDescriptor m_Data
Definition: Workload.hpp:89
armnn::RunClFunction
void RunClFunction(arm_compute::IFunction &function, const CheckLocation &location)
Definition: ClWorkloadUtils.hpp:167
armnn::QuantizedLstmInputParamsInfo::GetOutputGateBias
const TensorInfo & GetOutputGateBias() const
Definition: QuantizedLstmParams.hpp:211
armnn::ClQuantizedLstmWorkloadValidate
arm_compute::Status ClQuantizedLstmWorkloadValidate(const TensorInfo &input, const TensorInfo &previousCellStateIn, const TensorInfo &previousOutputIn, const TensorInfo &cellStateOut, const TensorInfo &output, const QuantizedLstmInputParamsInfo &paramsInfo)
Definition: ClQuantizedLstmWorkload.cpp:18
armnn
Copyright (c) 2021 ARM Limited and Contributors.
Definition: 01_00_quick_start.dox:6
ArmComputeTensorUtils.hpp
armnn::QuantizedLstmInputParamsInfo::GetInputToCellWeights
const TensorInfo & GetInputToCellWeights() const
Definition: QuantizedLstmParams.hpp:173
armnn::QuantizedLstmQueueDescriptor::m_OutputGateBias
const ConstTensorHandle * m_OutputGateBias
Definition: WorkloadData.hpp:646
armnn::QuantizedLstmInputParamsInfo::GetRecurrentToInputWeights
const TensorInfo & GetRecurrentToInputWeights() const
Definition: QuantizedLstmParams.hpp:182
armnn::QueueDescriptor::m_Inputs
std::vector< ITensorHandle * > m_Inputs
Definition: WorkloadData.hpp:26
armnn::QuantizedLstmQueueDescriptor::m_ForgetGateBias
const ConstTensorHandle * m_ForgetGateBias
Definition: WorkloadData.hpp:644