1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "ClQuantizedLstmWorkload.hpp"
7 #include "ClWorkloadUtils.hpp"
8
9 #include <armnn/backends/TensorHandle.hpp>
10 #include <aclCommon/ArmComputeTensorUtils.hpp>
11 #include <cl/ClTensorHandle.hpp>
12
13 namespace armnn
14 {
15
16 using namespace armcomputetensorutils;
17
ClQuantizedLstmWorkloadValidate(const TensorInfo & input,const TensorInfo & previousCellStateIn,const TensorInfo & previousOutputIn,const TensorInfo & cellStateOut,const TensorInfo & output,const QuantizedLstmInputParamsInfo & paramsInfo)18 arm_compute::Status ClQuantizedLstmWorkloadValidate(const TensorInfo& input, const TensorInfo& previousCellStateIn,
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
ClQuantizedLstmWorkload(const QuantizedLstmQueueDescriptor & descriptor,const WorkloadInfo & info,const arm_compute::CLCompileContext & clCompileContext)64 ClQuantizedLstmWorkload::ClQuantizedLstmWorkload(const QuantizedLstmQueueDescriptor &descriptor,
65 const WorkloadInfo &info,
66 const arm_compute::CLCompileContext& clCompileContext)
67 : ClBaseWorkload<QuantizedLstmQueueDescriptor>(descriptor, info)
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(Compute::Undefined, "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);
133 InitializeArmComputeClTensorData(*m_InputGateBiasTensor, m_Data.m_InputGateBias);
134 InitializeArmComputeClTensorData(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias);
135 InitializeArmComputeClTensorData(*m_CellBiasTensor, m_Data.m_CellBias);
136 InitializeArmComputeClTensorData(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias);
137
138 m_QuantizedLstmLayer.prepare();
139 FreeUnusedTensors();
140 }
141
Execute() const142 void ClQuantizedLstmWorkload::Execute() const
143 {
144 ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClQuantizedLstmWorkload_Execute", this->GetGuid());
145 RunClFunction(m_QuantizedLstmLayer, CHECK_LOCATION());
146 }
147
FreeUnusedTensors()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