Ryan OShea | 2323af4 | 2020-05-13 16:36:19 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2020 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "ClQLstmWorkload.hpp" |
| 7 | #include "ClWorkloadUtils.hpp" |
| 8 | |
| 9 | #include "aclCommon/ArmComputeTensorUtils.hpp" |
| 10 | |
| 11 | #include "cl/ClTensorHandle.hpp" |
| 12 | |
| 13 | namespace armnn |
| 14 | { |
| 15 | using namespace armcomputetensorutils; |
| 16 | |
| 17 | ClQLstmWorkload::ClQLstmWorkload(const QLstmQueueDescriptor &descriptor, const WorkloadInfo &info) |
| 18 | : BaseWorkload<QLstmQueueDescriptor>(descriptor, info) |
| 19 | { |
| 20 | arm_compute::LSTMParams<arm_compute::ICLTensor> qLstmParams; |
| 21 | |
| 22 | // Mandatory params |
| 23 | m_InputToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>(); |
| 24 | BuildArmComputeTensor(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights->GetTensorInfo()); |
| 25 | |
| 26 | m_InputToCellWeightsTensor = std::make_unique<arm_compute::CLTensor>(); |
| 27 | BuildArmComputeTensor(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights->GetTensorInfo()); |
| 28 | |
| 29 | m_InputToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); |
| 30 | BuildArmComputeTensor(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights->GetTensorInfo()); |
| 31 | |
| 32 | m_RecurrentToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>(); |
| 33 | BuildArmComputeTensor(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights->GetTensorInfo()); |
| 34 | |
| 35 | m_RecurrentToCellWeightsTensor = std::make_unique<arm_compute::CLTensor>(); |
| 36 | BuildArmComputeTensor(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights->GetTensorInfo()); |
| 37 | |
| 38 | m_RecurrentToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); |
| 39 | BuildArmComputeTensor(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights->GetTensorInfo()); |
| 40 | |
| 41 | m_ForgetGateBiasTensor = std::make_unique<arm_compute::CLTensor>(); |
| 42 | BuildArmComputeTensor(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias->GetTensorInfo()); |
| 43 | |
| 44 | m_CellBiasTensor = std::make_unique<arm_compute::CLTensor>(); |
| 45 | BuildArmComputeTensor(*m_CellBiasTensor, m_Data.m_CellBias->GetTensorInfo()); |
| 46 | |
| 47 | m_OutputGateBiasTensor = std::make_unique<arm_compute::CLTensor>(); |
| 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::CLTensor>(); |
| 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::CLTensor>(); |
| 62 | BuildArmComputeTensor(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights->GetTensorInfo()); |
| 63 | |
| 64 | m_CellToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); |
| 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::CLTensor>(); |
| 75 | BuildArmComputeTensor(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights->GetTensorInfo()); |
| 76 | |
| 77 | m_ProjectionBiasTensor = std::make_unique<arm_compute::CLTensor>(); |
| 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::CLTensor>(); |
| 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::CLTensor>(); |
| 99 | BuildArmComputeTensor(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights->GetTensorInfo()); |
| 100 | |
| 101 | m_CellLayerNormWeightsTensor = std::make_unique<arm_compute::CLTensor>(); |
| 102 | BuildArmComputeTensor(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights->GetTensorInfo()); |
| 103 | |
| 104 | m_OutputLayerNormWeightsTensor = std::make_unique<arm_compute::CLTensor>(); |
| 105 | BuildArmComputeTensor(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights->GetTensorInfo()); |
| 106 | |
| 107 | qLstmParams.set_layer_normalization_params( |
| 108 | m_Data.m_InputLayerNormWeights != nullptr ? m_InputLayerNormWeightsTensor.get() : nullptr, |
| 109 | m_ForgetLayerNormWeightsTensor.get(), |
| 110 | m_CellLayerNormWeightsTensor.get(), |
| 111 | m_OutputLayerNormWeightsTensor.get()); |
| 112 | } |
| 113 | |
| 114 | if (!m_Data.m_Parameters.m_CifgEnabled) |
| 115 | { |
| 116 | m_InputToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); |
| 117 | BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo()); |
| 118 | |
| 119 | m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); |
| 120 | BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo()); |
| 121 | |
| 122 | m_InputGateBiasTensor = std::make_unique<arm_compute::CLTensor>(); |
| 123 | BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo()); |
| 124 | |
| 125 | qLstmParams.set_cifg_params( |
| 126 | m_InputToInputWeightsTensor.get(), |
| 127 | m_RecurrentToInputWeightsTensor.get(), |
| 128 | m_Data.m_CellToInputWeights != nullptr ? m_CellToInputWeightsTensor.get() : nullptr, |
| 129 | m_InputGateBiasTensor.get()); |
| 130 | } |
| 131 | |
| 132 | // Input/Output tensors |
| 133 | const arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); |
| 134 | const arm_compute::ICLTensor& outputStateIn = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor(); |
Teresa Charlin | 57512af | 2020-07-29 12:06:40 +0100 | [diff] [blame] | 135 | arm_compute::ICLTensor& cellStateIn = static_cast<IClTensorHandle*>(m_Data.m_Inputs[2])->GetTensor(); |
Ryan OShea | 2323af4 | 2020-05-13 16:36:19 +0100 | [diff] [blame] | 136 | |
| 137 | arm_compute::ICLTensor& outputStateOut = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); |
| 138 | arm_compute::ICLTensor& cellStateOut = static_cast<IClTensorHandle*>(m_Data.m_Outputs[1])->GetTensor(); |
| 139 | arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[2])->GetTensor(); |
| 140 | |
| 141 | // Set scalar descriptor params |
| 142 | qLstmParams.set_cell_clip_params(m_Data.m_Parameters.m_CellClip); |
| 143 | qLstmParams.set_projection_clip_params(m_Data.m_Parameters.m_ProjectionClip); |
| 144 | qLstmParams.set_hidden_state_params(m_Data.m_Parameters.m_HiddenStateZeroPoint, |
| 145 | m_Data.m_Parameters.m_HiddenStateScale); |
| 146 | qLstmParams.set_matmul_scale_params(m_Data.m_Parameters.m_InputIntermediateScale, |
| 147 | m_Data.m_Parameters.m_ForgetIntermediateScale, |
| 148 | m_Data.m_Parameters.m_CellIntermediateScale, |
| 149 | m_Data.m_Parameters.m_OutputIntermediateScale); |
| 150 | |
| 151 | m_QLstmLayer.configure(&input, |
| 152 | m_InputToForgetWeightsTensor.get(), |
| 153 | m_InputToCellWeightsTensor.get(), |
| 154 | m_InputToOutputWeightsTensor.get(), |
| 155 | m_RecurrentToForgetWeightsTensor.get(), |
| 156 | m_RecurrentToCellWeightsTensor.get(), |
| 157 | m_RecurrentToOutputWeightsTensor.get(), |
| 158 | m_ForgetGateBiasTensor.get(), |
| 159 | m_CellBiasTensor.get(), |
| 160 | m_OutputGateBiasTensor.get(), |
| 161 | &cellStateIn, |
| 162 | &outputStateIn, |
| 163 | &cellStateOut, |
| 164 | &outputStateOut, |
| 165 | &output, |
| 166 | qLstmParams); |
| 167 | |
| 168 | // InitializeArmComputeTensorData for mandatory params |
| 169 | InitializeArmComputeClTensorData(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights); |
| 170 | InitializeArmComputeClTensorData(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights); |
| 171 | InitializeArmComputeClTensorData(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights); |
| 172 | |
| 173 | InitializeArmComputeClTensorData(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights); |
| 174 | InitializeArmComputeClTensorData(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights); |
| 175 | InitializeArmComputeClTensorData(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights); |
| 176 | |
| 177 | InitializeArmComputeClTensorData(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias); |
| 178 | InitializeArmComputeClTensorData(*m_CellBiasTensor, m_Data.m_CellBias); |
| 179 | InitializeArmComputeClTensorData(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias); |
| 180 | |
| 181 | if (!m_Data.m_Parameters.m_CifgEnabled) |
| 182 | { |
| 183 | InitializeArmComputeClTensorData(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights); |
| 184 | InitializeArmComputeClTensorData(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights); |
| 185 | InitializeArmComputeClTensorData(*m_InputGateBiasTensor, m_Data.m_InputGateBias); |
| 186 | } |
| 187 | |
| 188 | if (m_Data.m_Parameters.m_ProjectionEnabled) |
| 189 | { |
| 190 | InitializeArmComputeClTensorData(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights); |
| 191 | |
| 192 | if (m_Data.m_ProjectionBias != nullptr) |
| 193 | { |
| 194 | InitializeArmComputeClTensorData(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias); |
| 195 | } |
| 196 | } |
| 197 | |
| 198 | if (m_Data.m_Parameters.m_PeepholeEnabled) |
| 199 | { |
| 200 | if (!m_Data.m_Parameters.m_CifgEnabled) |
| 201 | { |
| 202 | InitializeArmComputeClTensorData(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights); |
| 203 | } |
| 204 | |
| 205 | InitializeArmComputeClTensorData(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights); |
| 206 | InitializeArmComputeClTensorData(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights); |
| 207 | } |
| 208 | |
| 209 | if (m_Data.m_Parameters.m_LayerNormEnabled) |
| 210 | { |
| 211 | if (!m_Data.m_Parameters.m_CifgEnabled) |
| 212 | { |
| 213 | InitializeArmComputeClTensorData(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights); |
| 214 | } |
| 215 | InitializeArmComputeClTensorData(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights); |
| 216 | InitializeArmComputeClTensorData(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights); |
| 217 | InitializeArmComputeClTensorData(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights); |
| 218 | } |
| 219 | |
| 220 | m_QLstmLayer.prepare(); |
| 221 | |
| 222 | FreeUnusedTensors(); |
| 223 | } |
| 224 | |
| 225 | void ClQLstmWorkload::Execute() const |
| 226 | { |
| 227 | m_QLstmLayer.run(); |
| 228 | } |
| 229 | |
| 230 | arm_compute::Status ClQLstmWorkloadValidate(const TensorInfo& input, |
| 231 | const TensorInfo& cellStateIn, |
| 232 | const TensorInfo& outputStateIn, |
| 233 | const TensorInfo& cellStateOut, |
| 234 | const TensorInfo& outputStateOut, |
| 235 | const TensorInfo& output, |
| 236 | const QLstmDescriptor& descriptor, |
| 237 | const LstmInputParamsInfo& paramsInfo) |
| 238 | { |
| 239 | arm_compute::LSTMParams<arm_compute::ITensorInfo> aclParamsInfo; |
| 240 | |
| 241 | // The inputs and outputs |
| 242 | const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input); |
| 243 | const arm_compute::TensorInfo aclOutputStateInInfo = BuildArmComputeTensorInfo(outputStateIn); |
| 244 | const arm_compute::TensorInfo aclCellStateInInfo = BuildArmComputeTensorInfo(cellStateIn); |
| 245 | |
| 246 | const arm_compute::TensorInfo aclOutputStateOutInfo = BuildArmComputeTensorInfo(outputStateOut); |
| 247 | const arm_compute::TensorInfo aclCellStateOutInfo = BuildArmComputeTensorInfo(cellStateOut); |
| 248 | const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); |
| 249 | |
| 250 | // Mandatory tensor info |
| 251 | const arm_compute::TensorInfo aclInputToForgetWeightsInfo |
| 252 | = BuildArmComputeTensorInfo(paramsInfo.GetInputToForgetWeights()); |
| 253 | const arm_compute::TensorInfo aclInputToCellWeightsInfo |
| 254 | = BuildArmComputeTensorInfo(paramsInfo.GetInputToCellWeights()); |
| 255 | const arm_compute::TensorInfo aclInputToOutputWeightsInfo |
| 256 | = BuildArmComputeTensorInfo(paramsInfo.GetInputToOutputWeights()); |
| 257 | const arm_compute::TensorInfo aclRecurrentToForgetWeightsInfo |
| 258 | = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToForgetWeights()); |
| 259 | const arm_compute::TensorInfo aclRecurrentToCellWeightsInfo |
| 260 | = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToCellWeights()); |
| 261 | const arm_compute::TensorInfo aclRecurrentToOutputWeightsInfo |
| 262 | = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToOutputWeights()); |
| 263 | const arm_compute::TensorInfo aclForgetGateBiasInfo |
| 264 | = BuildArmComputeTensorInfo(paramsInfo.GetForgetGateBias()); |
| 265 | const arm_compute::TensorInfo aclCellBiasInfo |
| 266 | = BuildArmComputeTensorInfo(paramsInfo.GetCellBias()); |
| 267 | const arm_compute::TensorInfo aclOutputGateBiasInfo |
| 268 | = BuildArmComputeTensorInfo(paramsInfo.GetOutputGateBias()); |
| 269 | |
| 270 | // Optional tensor info |
| 271 | arm_compute::TensorInfo aclInputToInputWeightsInfo; |
| 272 | arm_compute::TensorInfo aclRecurrentToInputWeightsInfo; |
| 273 | arm_compute::TensorInfo aclCellToInputWeightsInfo; |
| 274 | arm_compute::TensorInfo aclCellToForgetWeightsInfo; |
| 275 | arm_compute::TensorInfo aclCellToOutputWeightsInfo; |
| 276 | arm_compute::TensorInfo aclInputGateBiasInfo; |
| 277 | arm_compute::TensorInfo aclProjectionWeightsInfo; |
| 278 | arm_compute::TensorInfo aclProjectionBiasInfo; |
| 279 | arm_compute::TensorInfo aclInputLayerNormWeightsInfo; |
| 280 | arm_compute::TensorInfo aclForgetLayerNormWeightsInfo; |
| 281 | arm_compute::TensorInfo aclCellLayerNormWeightsInfo; |
| 282 | arm_compute::TensorInfo aclOutputLayerNormWeightsInfo; |
| 283 | |
| 284 | |
| 285 | if (descriptor.m_PeepholeEnabled) |
| 286 | { |
| 287 | if (!descriptor.m_CifgEnabled) |
| 288 | { |
| 289 | aclCellToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToInputWeights()); |
| 290 | } |
| 291 | |
| 292 | aclCellToForgetWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToForgetWeights()); |
| 293 | aclCellToOutputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToOutputWeights()); |
| 294 | |
| 295 | aclParamsInfo.set_peephole_params(&aclCellToForgetWeightsInfo, |
| 296 | &aclCellToOutputWeightsInfo); |
| 297 | } |
| 298 | |
| 299 | if (descriptor.m_ProjectionEnabled) |
| 300 | { |
| 301 | aclProjectionWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetProjectionWeights()); |
| 302 | |
| 303 | if (paramsInfo.m_ProjectionBias != nullptr) |
| 304 | { |
| 305 | aclProjectionBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetProjectionBias()); |
| 306 | } |
| 307 | |
| 308 | aclParamsInfo.set_projection_params( |
| 309 | &aclProjectionWeightsInfo, |
| 310 | paramsInfo.m_ProjectionBias != nullptr ? &aclProjectionBiasInfo : nullptr); |
| 311 | } |
| 312 | |
| 313 | if (descriptor.m_LayerNormEnabled) |
| 314 | { |
| 315 | if (!descriptor.m_CifgEnabled) |
| 316 | { |
| 317 | aclInputLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputLayerNormWeights()); |
| 318 | } |
| 319 | |
| 320 | aclForgetLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetForgetLayerNormWeights()); |
| 321 | aclCellLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellLayerNormWeights()); |
| 322 | aclOutputLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetOutputLayerNormWeights()); |
| 323 | |
| 324 | aclParamsInfo.set_layer_normalization_params( |
| 325 | paramsInfo.m_InputLayerNormWeights != nullptr ? &aclInputLayerNormWeightsInfo : nullptr, |
| 326 | &aclForgetLayerNormWeightsInfo, |
| 327 | &aclCellLayerNormWeightsInfo, |
| 328 | &aclOutputLayerNormWeightsInfo); |
| 329 | } |
| 330 | |
| 331 | if (!descriptor.m_CifgEnabled) |
| 332 | { |
| 333 | aclInputToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputToInputWeights()); |
| 334 | aclRecurrentToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToInputWeights()); |
| 335 | aclInputGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputGateBias()); |
| 336 | |
| 337 | |
| 338 | aclParamsInfo.set_cifg_params( |
| 339 | &aclInputToInputWeightsInfo, |
| 340 | &aclRecurrentToInputWeightsInfo, |
| 341 | paramsInfo.m_CellToInputWeights != nullptr ? &aclCellToInputWeightsInfo : nullptr, |
| 342 | &aclInputGateBiasInfo); |
| 343 | } |
| 344 | |
| 345 | aclParamsInfo.set_cell_clip_params(descriptor.m_CellClip); |
| 346 | aclParamsInfo.set_projection_clip_params(descriptor.m_ProjectionClip); |
| 347 | aclParamsInfo.set_hidden_state_params(descriptor.m_HiddenStateZeroPoint, descriptor.m_HiddenStateScale); |
| 348 | aclParamsInfo.set_matmul_scale_params(descriptor.m_InputIntermediateScale, |
| 349 | descriptor.m_ForgetIntermediateScale, |
| 350 | descriptor.m_CellIntermediateScale, |
| 351 | descriptor.m_OutputIntermediateScale); |
| 352 | |
| 353 | return arm_compute::CLQLSTMLayer::validate(&aclInputInfo, |
| 354 | &aclInputToForgetWeightsInfo, |
| 355 | &aclInputToCellWeightsInfo, |
| 356 | &aclInputToOutputWeightsInfo, |
| 357 | &aclRecurrentToForgetWeightsInfo, |
| 358 | &aclRecurrentToCellWeightsInfo, |
| 359 | &aclRecurrentToOutputWeightsInfo, |
| 360 | &aclForgetGateBiasInfo, |
| 361 | &aclCellBiasInfo, |
| 362 | &aclOutputGateBiasInfo, |
| 363 | &aclCellStateInInfo, |
| 364 | &aclOutputStateInInfo, |
| 365 | &aclCellStateOutInfo, |
| 366 | &aclOutputStateOutInfo, |
| 367 | &aclOutputInfo, |
| 368 | aclParamsInfo); |
| 369 | } |
| 370 | |
| 371 | void ClQLstmWorkload::FreeUnusedTensors() |
| 372 | { |
| 373 | FreeTensorIfUnused(m_InputToInputWeightsTensor); |
| 374 | FreeTensorIfUnused(m_InputToForgetWeightsTensor); |
| 375 | FreeTensorIfUnused(m_InputToCellWeightsTensor); |
| 376 | FreeTensorIfUnused(m_InputToOutputWeightsTensor); |
| 377 | |
| 378 | FreeTensorIfUnused(m_RecurrentToInputWeightsTensor); |
| 379 | FreeTensorIfUnused(m_RecurrentToForgetWeightsTensor); |
| 380 | FreeTensorIfUnused(m_RecurrentToCellWeightsTensor); |
| 381 | FreeTensorIfUnused(m_RecurrentToOutputWeightsTensor); |
| 382 | |
| 383 | FreeTensorIfUnused(m_CellToInputWeightsTensor); |
| 384 | FreeTensorIfUnused(m_CellToForgetWeightsTensor); |
| 385 | FreeTensorIfUnused(m_CellToOutputWeightsTensor); |
| 386 | |
| 387 | FreeTensorIfUnused(m_InputGateBiasTensor); |
| 388 | FreeTensorIfUnused(m_ForgetGateBiasTensor); |
| 389 | FreeTensorIfUnused(m_CellBiasTensor); |
| 390 | FreeTensorIfUnused(m_OutputGateBiasTensor); |
| 391 | |
| 392 | FreeTensorIfUnused(m_ProjectionWeightsTensor); |
| 393 | FreeTensorIfUnused(m_ProjectionBiasTensor); |
| 394 | |
| 395 | FreeTensorIfUnused(m_InputLayerNormWeightsTensor); |
| 396 | FreeTensorIfUnused(m_ForgetLayerNormWeightsTensor); |
| 397 | FreeTensorIfUnused(m_CellLayerNormWeightsTensor); |
| 398 | FreeTensorIfUnused(m_OutputLayerNormWeightsTensor); |
| 399 | } |
| 400 | |
| 401 | } //namespace armnn |