1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "NeonQuantizedLstmWorkload.hpp"
7 #include "NeonWorkloadUtils.hpp"
8
9 #include <aclCommon/ArmComputeTensorUtils.hpp>
10 #include <armnn/backends/TensorHandle.hpp>
11 #include <neon/NeonTensorHandle.hpp>
12
13 namespace armnn
14 {
15 using namespace armcomputetensorutils;
16
NeonQuantizedLstmWorkload(const QuantizedLstmQueueDescriptor & descriptor,const WorkloadInfo & info)17 NeonQuantizedLstmWorkload::NeonQuantizedLstmWorkload(const QuantizedLstmQueueDescriptor &descriptor,
18 const WorkloadInfo &info)
19 : NeonBaseWorkload<QuantizedLstmQueueDescriptor>(descriptor, info)
20 {
21 // Basic parameters
22 m_InputToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
23 BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo());
24
25 m_InputToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
26 BuildArmComputeTensor(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights->GetTensorInfo());
27
28 m_InputToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
29 BuildArmComputeTensor(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights->GetTensorInfo());
30
31 m_InputToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
32 BuildArmComputeTensor(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights->GetTensorInfo());
33
34 m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
35 BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo());
36
37 m_RecurrentToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
38 BuildArmComputeTensor(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights->GetTensorInfo());
39
40 m_RecurrentToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
41 BuildArmComputeTensor(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights->GetTensorInfo());
42
43 m_RecurrentToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
44 BuildArmComputeTensor(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights->GetTensorInfo());
45
46 m_InputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
47 BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo());
48
49 m_ForgetGateBiasTensor = std::make_unique<arm_compute::Tensor>();
50 BuildArmComputeTensor(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias->GetTensorInfo());
51
52 m_CellBiasTensor = std::make_unique<arm_compute::Tensor>();
53 BuildArmComputeTensor(*m_CellBiasTensor, m_Data.m_CellBias->GetTensorInfo());
54
55 m_OutputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
56 BuildArmComputeTensor(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias->GetTensorInfo());
57
58 const arm_compute::ITensor& input = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
59 arm_compute::ITensor& cell_state_in = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
60 const arm_compute::ITensor& output_state_in = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();
61
62 arm_compute::ITensor& cell_state_out = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
63 arm_compute::ITensor& output_state_out = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[1])->GetTensor();
64
65 m_QuantizedLstmLayer.configure(&input,
66 m_InputToInputWeightsTensor.get(),
67 m_InputToForgetWeightsTensor.get(),
68 m_InputToCellWeightsTensor.get(),
69 m_InputToOutputWeightsTensor.get(),
70 m_RecurrentToInputWeightsTensor.get(),
71 m_RecurrentToForgetWeightsTensor.get(),
72 m_RecurrentToCellWeightsTensor.get(),
73 m_RecurrentToOutputWeightsTensor.get(),
74 m_InputGateBiasTensor.get(),
75 m_ForgetGateBiasTensor.get(),
76 m_CellBiasTensor.get(),
77 m_OutputGateBiasTensor.get(),
78 &cell_state_in,
79 &output_state_in,
80 &cell_state_out,
81 &output_state_out);
82
83 InitializeArmComputeTensorData(*m_InputToInputWeightsTensor,
84 m_Data.m_InputToInputWeights);
85
86 InitializeArmComputeTensorData(*m_InputToForgetWeightsTensor,
87 m_Data.m_InputToForgetWeights);
88
89 InitializeArmComputeTensorData(*m_InputToCellWeightsTensor,
90 m_Data.m_InputToCellWeights);
91
92 InitializeArmComputeTensorData(*m_InputToOutputWeightsTensor,
93 m_Data.m_InputToOutputWeights);
94
95 InitializeArmComputeTensorData(*m_RecurrentToInputWeightsTensor,
96 m_Data.m_RecurrentToInputWeights);
97
98 InitializeArmComputeTensorData(*m_RecurrentToForgetWeightsTensor,
99 m_Data.m_RecurrentToForgetWeights);
100
101 InitializeArmComputeTensorData(*m_RecurrentToCellWeightsTensor,
102 m_Data.m_RecurrentToCellWeights);
103
104 InitializeArmComputeTensorData(*m_RecurrentToOutputWeightsTensor,
105 m_Data.m_RecurrentToOutputWeights);
106
107 InitializeArmComputeTensorData(*m_InputGateBiasTensor,
108 m_Data.m_InputGateBias);
109
110 InitializeArmComputeTensorData(*m_ForgetGateBiasTensor,
111 m_Data.m_ForgetGateBias);
112
113 InitializeArmComputeTensorData(*m_CellBiasTensor,
114 m_Data.m_CellBias);
115
116 InitializeArmComputeTensorData(*m_OutputGateBiasTensor,
117 m_Data.m_OutputGateBias);
118
119 // Force Compute Library to perform the necessary copying and reshaping, after which
120 // delete all the input tensors that will no longer be needed
121 m_QuantizedLstmLayer.prepare();
122 FreeUnusedTensors();
123 }
124
Execute() const125 void NeonQuantizedLstmWorkload::Execute() const
126 {
127 ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonQuantizedLstmWorkload_Execute", this->GetGuid());
128 m_QuantizedLstmLayer.run();
129 }
130
NeonQuantizedLstmWorkloadValidate(const TensorInfo & input,const TensorInfo & cellStateIn,const TensorInfo & outputStateIn,const TensorInfo & cellStateOut,const TensorInfo & outputStateOut,const QuantizedLstmInputParamsInfo & paramsInfo)131 arm_compute::Status NeonQuantizedLstmWorkloadValidate(const TensorInfo& input,
132 const TensorInfo& cellStateIn,
133 const TensorInfo& outputStateIn,
134 const TensorInfo& cellStateOut,
135 const TensorInfo& outputStateOut,
136 const QuantizedLstmInputParamsInfo& paramsInfo)
137 {
138 // The inputs and outputs
139 const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input);
140 const arm_compute::TensorInfo aclCellStateInInfo = BuildArmComputeTensorInfo(cellStateIn);
141 const arm_compute::TensorInfo aclOutputStateInInfo = BuildArmComputeTensorInfo(outputStateIn);
142 const arm_compute::TensorInfo aclCellStateOutInfo = BuildArmComputeTensorInfo(cellStateOut);
143 const arm_compute::TensorInfo aclOutputStateOutInfo = BuildArmComputeTensorInfo(outputStateOut);
144
145 // Basic parameters
146 const arm_compute::TensorInfo aclInputToInputWeightsInfo
147 = BuildArmComputeTensorInfo(paramsInfo.GetInputToInputWeights());
148 const arm_compute::TensorInfo aclInputToForgetWeightsInfo
149 = BuildArmComputeTensorInfo(paramsInfo.GetInputToForgetWeights());
150 const arm_compute::TensorInfo aclInputToCellWeightsInfo
151 = BuildArmComputeTensorInfo(paramsInfo.GetInputToCellWeights());
152 const arm_compute::TensorInfo aclInputToOutputWeightsInfo
153 = BuildArmComputeTensorInfo(paramsInfo.GetInputToOutputWeights());
154
155 const arm_compute::TensorInfo aclRecurrentToInputWeightsInfo
156 = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToInputWeights());
157 const arm_compute::TensorInfo aclRecurrentToForgetWeightsInfo
158 = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToForgetWeights());
159 const arm_compute::TensorInfo aclRecurrentToCellWeightsInfo
160 = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToCellWeights());
161 const arm_compute::TensorInfo aclRecurrentToOutputWeightsInfo
162 = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToOutputWeights());
163
164 const arm_compute::TensorInfo aclInputGateBiasInfo
165 = BuildArmComputeTensorInfo(paramsInfo.GetInputGateBias());
166 const arm_compute::TensorInfo aclForgetGateBiasInfo
167 = BuildArmComputeTensorInfo(paramsInfo.GetForgetGateBias());
168 const arm_compute::TensorInfo aclCellBiasInfo
169 = BuildArmComputeTensorInfo(paramsInfo.GetCellBias());
170 const arm_compute::TensorInfo aclOutputGateBiasInfo
171 = BuildArmComputeTensorInfo(paramsInfo.GetOutputGateBias());
172
173 return arm_compute::NELSTMLayerQuantized::validate(&aclInputInfo,
174 &aclInputToInputWeightsInfo,
175 &aclInputToForgetWeightsInfo,
176 &aclInputToCellWeightsInfo,
177 &aclInputToOutputWeightsInfo,
178 &aclRecurrentToInputWeightsInfo,
179 &aclRecurrentToForgetWeightsInfo,
180 &aclRecurrentToCellWeightsInfo,
181 &aclRecurrentToOutputWeightsInfo,
182 &aclInputGateBiasInfo,
183 &aclForgetGateBiasInfo,
184 &aclCellBiasInfo,
185 &aclOutputGateBiasInfo,
186 &aclCellStateInInfo,
187 &aclOutputStateInInfo,
188 &aclCellStateOutInfo,
189 &aclOutputStateOutInfo);
190 }
191
FreeUnusedTensors()192 void NeonQuantizedLstmWorkload::FreeUnusedTensors()
193 {
194 FreeTensorIfUnused(m_InputToInputWeightsTensor);
195 FreeTensorIfUnused(m_InputToForgetWeightsTensor);
196 FreeTensorIfUnused(m_InputToCellWeightsTensor);
197 FreeTensorIfUnused(m_InputToOutputWeightsTensor);
198 FreeTensorIfUnused(m_RecurrentToInputWeightsTensor);
199 FreeTensorIfUnused(m_RecurrentToForgetWeightsTensor);
200 FreeTensorIfUnused(m_RecurrentToCellWeightsTensor);
201 FreeTensorIfUnused(m_RecurrentToOutputWeightsTensor);
202 FreeTensorIfUnused(m_InputGateBiasTensor);
203 FreeTensorIfUnused(m_ForgetGateBiasTensor);
204 FreeTensorIfUnused(m_CellBiasTensor);
205 FreeTensorIfUnused(m_OutputGateBiasTensor);
206 FreeTensorIfUnused(m_CellStateInTensor);
207 FreeTensorIfUnused(m_OutputStateInTensor);
208 FreeTensorIfUnused(m_CellStateOutTensor);
209 }
210
211 } //namespace armnn
212