15 using namespace armcomputetensorutils;
22 m_InputToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
25 m_InputToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
28 m_InputToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
31 m_InputToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
34 m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::Tensor>();
37 m_RecurrentToForgetWeightsTensor = std::make_unique<arm_compute::Tensor>();
40 m_RecurrentToCellWeightsTensor = std::make_unique<arm_compute::Tensor>();
43 m_RecurrentToOutputWeightsTensor = std::make_unique<arm_compute::Tensor>();
46 m_InputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
49 m_ForgetGateBiasTensor = std::make_unique<arm_compute::Tensor>();
52 m_CellBiasTensor = std::make_unique<arm_compute::Tensor>();
55 m_OutputGateBiasTensor = std::make_unique<arm_compute::Tensor>();
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(),
121 m_QuantizedLstmLayer.prepare();
127 m_QuantizedLstmLayer.run();
138 const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input);
139 const arm_compute::TensorInfo aclCellStateInInfo = BuildArmComputeTensorInfo(cellStateIn);
140 const arm_compute::TensorInfo aclOutputStateInInfo = BuildArmComputeTensorInfo(outputStateIn);
141 const arm_compute::TensorInfo aclCellStateOutInfo = BuildArmComputeTensorInfo(cellStateOut);
142 const arm_compute::TensorInfo aclOutputStateOutInfo = BuildArmComputeTensorInfo(outputStateOut);
145 const arm_compute::TensorInfo aclInputToInputWeightsInfo
147 const arm_compute::TensorInfo aclInputToForgetWeightsInfo
149 const arm_compute::TensorInfo aclInputToCellWeightsInfo
151 const arm_compute::TensorInfo aclInputToOutputWeightsInfo
154 const arm_compute::TensorInfo aclRecurrentToInputWeightsInfo
156 const arm_compute::TensorInfo aclRecurrentToForgetWeightsInfo
158 const arm_compute::TensorInfo aclRecurrentToCellWeightsInfo
160 const arm_compute::TensorInfo aclRecurrentToOutputWeightsInfo
163 const arm_compute::TensorInfo aclInputGateBiasInfo
165 const arm_compute::TensorInfo aclForgetGateBiasInfo
167 const arm_compute::TensorInfo aclCellBiasInfo
168 = BuildArmComputeTensorInfo(paramsInfo.
GetCellBias());
169 const arm_compute::TensorInfo aclOutputGateBiasInfo
172 return arm_compute::NELSTMLayerQuantized::validate(&aclInputInfo,
173 &aclInputToInputWeightsInfo,
174 &aclInputToForgetWeightsInfo,
175 &aclInputToCellWeightsInfo,
176 &aclInputToOutputWeightsInfo,
177 &aclRecurrentToInputWeightsInfo,
178 &aclRecurrentToForgetWeightsInfo,
179 &aclRecurrentToCellWeightsInfo,
180 &aclRecurrentToOutputWeightsInfo,
181 &aclInputGateBiasInfo,
182 &aclForgetGateBiasInfo,
184 &aclOutputGateBiasInfo,
186 &aclOutputStateInInfo,
187 &aclCellStateOutInfo,
188 &aclOutputStateOutInfo);
191 void NeonQuantizedLstmWorkload::FreeUnusedTensors()
193 FreeTensorIfUnused(m_InputToInputWeightsTensor);
194 FreeTensorIfUnused(m_InputToForgetWeightsTensor);
195 FreeTensorIfUnused(m_InputToCellWeightsTensor);
196 FreeTensorIfUnused(m_InputToOutputWeightsTensor);
197 FreeTensorIfUnused(m_RecurrentToInputWeightsTensor);
198 FreeTensorIfUnused(m_RecurrentToForgetWeightsTensor);
199 FreeTensorIfUnused(m_RecurrentToCellWeightsTensor);
200 FreeTensorIfUnused(m_RecurrentToOutputWeightsTensor);
201 FreeTensorIfUnused(m_InputGateBiasTensor);
202 FreeTensorIfUnused(m_ForgetGateBiasTensor);
203 FreeTensorIfUnused(m_CellBiasTensor);
204 FreeTensorIfUnused(m_OutputGateBiasTensor);
205 FreeTensorIfUnused(m_CellStateInTensor);
206 FreeTensorIfUnused(m_OutputStateInTensor);
207 FreeTensorIfUnused(m_CellStateOutTensor);
const ConstCpuTensorHandle * m_CellBias
const ConstCpuTensorHandle * m_ForgetGateBias
const ConstCpuTensorHandle * m_InputToForgetWeights
const TensorInfo & GetTensorInfo() const
const ConstCpuTensorHandle * m_InputToInputWeights
const ConstCpuTensorHandle * m_RecurrentToInputWeights
const ConstCpuTensorHandle * m_OutputGateBias
const ConstCpuTensorHandle * m_InputGateBias
arm_compute::Status NeonQuantizedLstmWorkloadValidate(const TensorInfo &input, const TensorInfo &cellStateIn, const TensorInfo &outputStateIn, const TensorInfo &cellStateOut, const TensorInfo &outputStateOut, const QuantizedLstmInputParamsInfo ¶msInfo)
const QuantizedLstmQueueDescriptor m_Data
NeonQuantizedLstmWorkload(const QuantizedLstmQueueDescriptor &descriptor, const WorkloadInfo &info)
const ConstCpuTensorHandle * m_RecurrentToForgetWeights
const ConstCpuTensorHandle * m_RecurrentToOutputWeights
const ConstCpuTensorHandle * m_InputToOutputWeights
std::vector< ITensorHandle * > m_Outputs
std::vector< ITensorHandle * > m_Inputs
const ConstCpuTensorHandle * m_RecurrentToCellWeights
virtual void Execute() const override
void InitializeArmComputeTensorData(arm_compute::Tensor &tensor, const ConstCpuTensorHandle *handle)
const ConstCpuTensorHandle * m_InputToCellWeights