src/backends/backendsCommon/test/LstmTestImpl.hpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5 #pragma once
   6
   7 #include "QuantizeHelper.hpp"
   8 #include "WorkloadTestUtils.hpp"
   9
  10 #include <armnn/ArmNN.hpp>
  11 #include <armnn/Tensor.hpp>
  12 #include <armnn/TypesUtils.hpp>
  13
  14 #include <test/TensorHelpers.hpp>
  15
  16 #include <backendsCommon/CpuTensorHandle.hpp>
  17 #include <backendsCommon/WorkloadFactory.hpp>
  18
  19 LayerTestResult<float, 2> LstmNoCifgNoPeepholeNoProjectionTestImpl(
  20         armnn::IWorkloadFactory& workloadFactory,
  21         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
  22         const boost::multi_array<float, 2>& input,
  23         const boost::multi_array<float, 2>& outputExpected)
  24 {
  25     unsigned int batchSize = boost::numeric_cast<unsigned int>(input.shape()[0]);
  26     unsigned int inputSize = boost::numeric_cast<unsigned int>(input.shape()[1]);
  27     unsigned int outputSize = boost::numeric_cast<unsigned int>(outputExpected.shape()[1]);
  28     // cellSize and outputSize have the same size when there is no projection.
  29     unsigned numUnits = outputSize;
  30
  31
  32     armnn::TensorInfo inputTensorInfo({batchSize , inputSize}, armnn::DataType::Float32);
  33     armnn::TensorInfo cellStateInTensorInfo({batchSize , numUnits}, armnn::DataType::Float32);
  34     armnn::TensorInfo outputStateInTensorInfo({batchSize , outputSize}, armnn::DataType::Float32);
  35
  36
  37     armnn::TensorInfo scratchBufferTensorInfo({batchSize, numUnits * 4}, armnn::DataType::Float32);
  38     armnn::TensorInfo cellStateOutTensorInfo({batchSize, numUnits}, armnn::DataType::Float32);
  39     armnn::TensorInfo outputStateOutTensorInfo({batchSize, outputSize}, armnn::DataType::Float32);
  40     armnn::TensorInfo outputTensorInfo({batchSize, outputSize}, armnn::DataType::Float32);
  41
  42
  43     LayerTestResult<float, 2> ret(outputTensorInfo);
  44
  45     std::vector<float> inputVector;
  46     inputVector.assign(input.data(), input.data() + (batchSize * inputSize));
  47     auto inputTensor = MakeTensor<float,2>(inputTensorInfo, inputVector);
  48
  49     std::vector<float> cellStateInVector(batchSize * numUnits, 0.f);
  50     auto cellStateInTensor = MakeTensor<float,2>(cellStateInTensorInfo, cellStateInVector);
  51
  52     std::vector<float> outputStateInVector(batchSize * outputSize, 0.f);
  53     auto outputStateInTensor = MakeTensor<float,2>(outputStateInTensorInfo, outputStateInVector);
  54
  55     std::vector<float> scratchBufferVector(batchSize * numUnits * 4, 0.f);
  56     auto scratchBufferTensor = MakeTensor<float,2>(scratchBufferTensorInfo, scratchBufferVector);
  57
  58     std::vector<float> outputStateOutVector(batchSize * outputSize, 0.f);
  59     auto outputStateOutTensor = MakeTensor<float,2>(outputStateOutTensorInfo, outputStateOutVector);
  60
  61     std::vector<float> cellStateOutVector(batchSize * numUnits, 0.f);
  62     auto cellStateOutTensor = MakeTensor<float,2>(cellStateOutTensorInfo, cellStateOutVector);
  63
  64     std::vector<float> outputVector;
  65     outputVector.assign(outputExpected.data(), outputExpected.data() + (batchSize * outputSize));
  66     ret.outputExpected = MakeTensor<float, 2>(outputTensorInfo, outputVector);
  67
  68     std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo);
  69     std::unique_ptr<armnn::ITensorHandle> cellStateInHandle =
  70             workloadFactory.CreateTensorHandle(cellStateInTensorInfo);
  71     std::unique_ptr<armnn::ITensorHandle> outputStateInHandle =
  72             workloadFactory.CreateTensorHandle(outputStateInTensorInfo);
  73
  74     std::unique_ptr<armnn::ITensorHandle> scratchHandle = workloadFactory.CreateTensorHandle(scratchBufferTensorInfo);
  75     std::unique_ptr<armnn::ITensorHandle> outputStateOutHandle =
  76             workloadFactory.CreateTensorHandle(outputStateOutTensorInfo);
  77     std::unique_ptr<armnn::ITensorHandle> cellStateOutHandle =
  78             workloadFactory.CreateTensorHandle(cellStateOutTensorInfo);
  79     std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);
  80
  81
  82     armnn::LstmQueueDescriptor data;
  83     armnn::WorkloadInfo info;
  84
  85     AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
  86     AddInputToWorkload(data, info, outputStateInTensorInfo, outputStateInHandle.get());
  87     AddInputToWorkload(data, info, cellStateInTensorInfo, cellStateInHandle.get());
  88
  89     AddOutputToWorkload(data, info, scratchBufferTensorInfo, scratchHandle.get());
  90     AddOutputToWorkload(data, info, outputStateOutTensorInfo, outputStateOutHandle.get());
  91     AddOutputToWorkload(data, info, cellStateOutTensorInfo, cellStateOutHandle.get());
  92     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
  93
  94     armnn::TensorInfo tensorInfo4({numUnits}, armnn::DataType::Float32);
  95     armnn::TensorInfo tensorInfo8({numUnits, 2}, armnn::DataType::Float32);
  96     armnn::TensorInfo tensorInfo16({numUnits, 4}, armnn::DataType::Float32);
  97
  98     auto inputToInputWeights = MakeTensor<float, 2>(tensorInfo8, {-0.45018822f, -0.02338299f, -0.0870589f,
  99                                                                   -0.34550029f, 0.04266912f, -0.15680569f,
 100                                                                   -0.34856534f, 0.43890524f});
 101
 102     auto inputToForgetWeights = MakeTensor<float, 2>(tensorInfo8, {0.09701663f, 0.20334584f, -0.50592935f,
 103                                                                    -0.31343272f, -0.40032279f, 0.44781327f,
 104                                                                    0.01387155f, -0.35593212f});
 105
 106     auto inputToCellWeights = MakeTensor<float, 2>(tensorInfo8, {-0.50013041f, 0.1370284f, 0.11810488f, 0.2013163f,
 107                                                                  -0.20583314f, 0.44344562f, 0.22077113f,
 108                                                                  -0.29909778f});
 109
 110     auto inputToOutputWeights = MakeTensor<float, 2>(tensorInfo8, {-0.25065863f, -0.28290087f, 0.04613829f,
 111                                                                    0.40525138f, 0.44272184f, 0.03897077f,
 112                                                                    -0.1556896f, 0.19487578f});
 113
 114     auto recurrentToInputWeights = MakeTensor<float, 2>(tensorInfo16, {-0.0063535f, -0.2042388f, 0.31454784f,
 115                                                                        -0.35746509f, 0.28902304f, 0.08183324f,
 116                                                                        -0.16555229f, 0.02286911f, -0.13566875f,
 117                                                                        0.03034258f, 0.48091322f, -0.12528998f,
 118                                                                        0.24077177f, -0.51332325f, -0.33502164f,
 119                                                                        0.10629296f});
 120
 121     auto recurrentToForgetWeights = MakeTensor<float, 2>(tensorInfo16, {-0.48684245f, -0.06655136f, 0.42224967f,
 122                                                                         0.2112639f, 0.27654213f, 0.20864892f,
 123                                                                         -0.07646349f, 0.45877004f, 0.00141793f,
 124                                                                         -0.14609534f, 0.36447752f, 0.09196436f,
 125                                                                         0.28053468f, 0.01560611f, -0.20127171f,
 126                                                                         -0.01140004f});
 127
 128     auto recurrentToCellWeights = MakeTensor<float, 2>(tensorInfo16, {-0.3407414f, 0.24443203f, -0.2078532f,
 129                                                                       0.26320225f, 0.05695659f, -0.00123841f,
 130                                                                       -0.4744786f, -0.35869038f, -0.06418842f,
 131                                                                       -0.13502428f, -0.501764f, 0.22830659f,
 132                                                                       -0.46367589f, 0.26016325f, -0.03894562f,
 133                                                                       -0.16368064f});
 134
 135     auto recurrentToOutputWeights = MakeTensor<float, 2>(tensorInfo16, {0.43385774f, -0.17194885f, 0.2718237f,
 136                                                                         0.09215671f, 0.24107647f, -0.39835793f,
 137                                                                         0.18212086f, 0.01301402f, 0.48572797f,
 138                                                                         -0.50656658f, 0.20047462f, -0.20607421f,
 139                                                                         -0.51818722f, -0.15390486f, 0.0468148f,
 140                                                                         0.39922136f});
 141
 142     auto cellToInputWeights = MakeTensor<float, 1>(tensorInfo4, {0., 0., 0., 0.});
 143
 144     auto inputGateBias = MakeTensor<float, 1>(tensorInfo4, {0., 0., 0., 0.});
 145
 146     auto forgetGateBias = MakeTensor<float, 1>(tensorInfo4, {1., 1., 1., 1.});
 147
 148     auto cellBias = MakeTensor<float, 1>(tensorInfo4, {0., 0., 0., 0.});
 149
 150     auto outputGateBias = MakeTensor<float, 1>(tensorInfo4, {0., 0., 0., 0.});
 151
 152     armnn::ScopedCpuTensorHandle inputToInputWeightsTensor(tensorInfo8);
 153     armnn::ScopedCpuTensorHandle inputToForgetWeightsTensor(tensorInfo8);
 154     armnn::ScopedCpuTensorHandle inputToCellWeightsTensor(tensorInfo8);
 155     armnn::ScopedCpuTensorHandle inputToOutputWeightsTensor(tensorInfo8);
 156     armnn::ScopedCpuTensorHandle recurrentToInputWeightsTensor(tensorInfo16);
 157     armnn::ScopedCpuTensorHandle recurrentToForgetWeightsTensor(tensorInfo16);
 158     armnn::ScopedCpuTensorHandle recurrentToCellWeightsTensor(tensorInfo16);
 159     armnn::ScopedCpuTensorHandle recurrentToOutputWeightsTensor(tensorInfo16);
 160     armnn::ScopedCpuTensorHandle cellToInputWeightsTensor(tensorInfo4);
 161     armnn::ScopedCpuTensorHandle inputGateBiasTensor(tensorInfo4);
 162     armnn::ScopedCpuTensorHandle forgetGateBiasTensor(tensorInfo4);
 163     armnn::ScopedCpuTensorHandle cellBiasTensor(tensorInfo4);
 164     armnn::ScopedCpuTensorHandle outputGateBiasTensor(tensorInfo4);
 165
 166     AllocateAndCopyDataToITensorHandle(&inputToInputWeightsTensor, &inputToInputWeights[0][0]);
 167     AllocateAndCopyDataToITensorHandle(&inputToForgetWeightsTensor, &inputToForgetWeights[0][0]);
 168     AllocateAndCopyDataToITensorHandle(&inputToCellWeightsTensor, &inputToCellWeights[0][0]);
 169     AllocateAndCopyDataToITensorHandle(&inputToOutputWeightsTensor, &inputToOutputWeights[0][0]);
 170     AllocateAndCopyDataToITensorHandle(&recurrentToInputWeightsTensor, &recurrentToInputWeights[0][0]);
 171     AllocateAndCopyDataToITensorHandle(&recurrentToForgetWeightsTensor, &recurrentToForgetWeights[0][0]);
 172     AllocateAndCopyDataToITensorHandle(&recurrentToCellWeightsTensor, &recurrentToCellWeights[0][0]);
 173     AllocateAndCopyDataToITensorHandle(&recurrentToOutputWeightsTensor, &recurrentToOutputWeights[0][0]);
 174     AllocateAndCopyDataToITensorHandle(&cellToInputWeightsTensor, &cellToInputWeights[0]);
 175     AllocateAndCopyDataToITensorHandle(&inputGateBiasTensor, &inputGateBias[0]);
 176     AllocateAndCopyDataToITensorHandle(&forgetGateBiasTensor, &forgetGateBias[0]);
 177     AllocateAndCopyDataToITensorHandle(&cellBiasTensor, &cellBias[0]);
 178     AllocateAndCopyDataToITensorHandle(&outputGateBiasTensor, &outputGateBias[0]);
 179
 180     data.m_InputToInputWeights = &inputToInputWeightsTensor;
 181     data.m_InputToForgetWeights = &inputToForgetWeightsTensor;
 182     data.m_InputToCellWeights = &inputToCellWeightsTensor;
 183     data.m_InputToOutputWeights = &inputToOutputWeightsTensor;
 184     data.m_RecurrentToInputWeights = &recurrentToInputWeightsTensor;
 185     data.m_RecurrentToForgetWeights = &recurrentToForgetWeightsTensor;
 186     data.m_RecurrentToCellWeights = &recurrentToCellWeightsTensor;
 187     data.m_RecurrentToOutputWeights = &recurrentToOutputWeightsTensor;
 188     data.m_CellToInputWeights = &cellToInputWeightsTensor;
 189     data.m_InputGateBias = &inputGateBiasTensor;
 190     data.m_ForgetGateBias = &forgetGateBiasTensor;
 191     data.m_CellBias = &cellBiasTensor;
 192     data.m_OutputGateBias = &outputGateBiasTensor;
 193
 194
 195     // Flags to set test configuration
 196     data.m_Parameters.m_ActivationFunc = 4;
 197     data.m_Parameters.m_CifgEnabled = false;
 198     data.m_Parameters.m_PeepholeEnabled = false;
 199     data.m_Parameters.m_ProjectionEnabled = false;
 200
 201
 202     std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateLstm(data, info);
 203     inputHandle->Allocate();
 204     outputStateInHandle->Allocate();
 205     cellStateInHandle->Allocate();
 206
 207     scratchHandle->Allocate();
 208     outputStateOutHandle->Allocate();
 209     cellStateOutHandle->Allocate();
 210     outputHandle->Allocate();
 211
 212     CopyDataToITensorHandle(inputHandle.get(), &inputTensor[0][0]);
 213     CopyDataToITensorHandle(outputStateInHandle.get(), &outputStateInTensor[0][0]);
 214     CopyDataToITensorHandle(cellStateInHandle.get(), &cellStateInTensor[0][0]);
 215
 216     workload->Execute();
 217
 218     CopyDataFromITensorHandle(&ret.output[0][0], outputHandle.get());
 219
 220     return ret;
 221 }
 222
 223
 224 LayerTestResult<float, 2>
 225 LstmLayerNoCifgWithPeepholeWithProjectionTestImpl(armnn::IWorkloadFactory& workloadFactory,
 226                                                   const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
 227                                                   const boost::multi_array<float, 2>& input,
 228                                                   const boost::multi_array<float, 2>& outputExpected)
 229 {
 230     unsigned int batchSize = 2;
 231     unsigned int outputSize = 16;
 232     unsigned int inputSize = 5;
 233     unsigned numUnits = 20;
 234
 235     armnn::TensorInfo inputTensorInfo({batchSize , inputSize}, armnn::DataType::Float32);
 236     armnn::TensorInfo cellStateInTensorInfo({batchSize , numUnits}, armnn::DataType::Float32);
 237     armnn::TensorInfo outputStateInTensorInfo({batchSize , outputSize}, armnn::DataType::Float32);
 238
 239     // Scratch buffer size without CIFG [batchSize, numUnits * 4]
 240     armnn::TensorInfo scratchBufferTensorInfo({batchSize, numUnits * 4}, armnn::DataType::Float32);
 241     armnn::TensorInfo cellStateOutTensorInfo({batchSize, numUnits}, armnn::DataType::Float32);
 242     armnn::TensorInfo outputStateOutTensorInfo({batchSize, outputSize}, armnn::DataType::Float32);
 243     armnn::TensorInfo outputTensorInfo({batchSize, outputSize}, armnn::DataType::Float32);
 244
 245     LayerTestResult<float, 2> ret(outputTensorInfo);
 246
 247     std::vector<float> inputVector;
 248     inputVector.assign(input.data(), input.data() + (batchSize * inputSize));
 249     auto inputTensor = MakeTensor<float,2>(inputTensorInfo, inputVector);
 250
 251     std::vector<float> cellStateInVector(batchSize * numUnits, 0.f);
 252     auto cellStateInTensor = MakeTensor<float,2>(cellStateInTensorInfo, cellStateInVector);
 253
 254     std::vector<float> outputStateInVector(batchSize * outputSize, 0.f);
 255     auto outputStateInTensor = MakeTensor<float,2>(outputStateInTensorInfo, outputStateInVector);
 256
 257     std::vector<float> scratchBufferVector(batchSize * numUnits * 4, 0.f);
 258     auto scratchBufferTensor = MakeTensor<float,2>(scratchBufferTensorInfo, scratchBufferVector);
 259
 260     std::vector<float> outputStateOutVector(batchSize * outputSize, 0.f);
 261     auto outputStateOutTensor = MakeTensor<float,2>(outputStateOutTensorInfo, outputStateOutVector);
 262
 263     std::vector<float> cellStateOutVector(batchSize * numUnits, 0.f);
 264     auto cellStateOutTensor = MakeTensor<float,2>(cellStateOutTensorInfo, cellStateOutVector);
 265
 266     std::vector<float> outputVector;
 267     outputVector.assign(outputExpected.data(), outputExpected.data() + (batchSize * outputSize));
 268     ret.outputExpected = MakeTensor<float, 2>(outputTensorInfo, outputVector);
 269
 270     std::unique_ptr<armnn::ITensorHandle> inputHandle = workloadFactory.CreateTensorHandle(inputTensorInfo);
 271     std::unique_ptr<armnn::ITensorHandle> cellStateInHandle =
 272             workloadFactory.CreateTensorHandle(cellStateInTensorInfo);
 273     std::unique_ptr<armnn::ITensorHandle> outputStateInHandle =
 274             workloadFactory.CreateTensorHandle(outputStateInTensorInfo);
 275
 276     std::unique_ptr<armnn::ITensorHandle> scratchHandle = workloadFactory.CreateTensorHandle(scratchBufferTensorInfo);
 277     std::unique_ptr<armnn::ITensorHandle> outputStateOutHandle =
 278             workloadFactory.CreateTensorHandle(outputStateOutTensorInfo);
 279     std::unique_ptr<armnn::ITensorHandle> cellStateOutHandle =
 280             workloadFactory.CreateTensorHandle(cellStateOutTensorInfo);
 281     std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);
 282
 283     armnn::LstmQueueDescriptor data;
 284     armnn::WorkloadInfo info;
 285
 286     AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
 287     AddInputToWorkload(data, info, outputStateInTensorInfo, outputStateInHandle.get());
 288     AddInputToWorkload(data, info, cellStateInTensorInfo, cellStateInHandle.get());
 289
 290     AddOutputToWorkload(data, info, scratchBufferTensorInfo, scratchHandle.get());
 291     AddOutputToWorkload(data, info, outputStateOutTensorInfo, outputStateOutHandle.get());
 292     AddOutputToWorkload(data, info, cellStateOutTensorInfo, cellStateOutHandle.get());
 293     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
 294
 295     armnn::TensorInfo tensorInfo16({outputSize}, armnn::DataType::Float32);
 296     armnn::TensorInfo tensorInfo20({numUnits}, armnn::DataType::Float32);
 297     armnn::TensorInfo tensorInfo20x5({numUnits, inputSize}, armnn::DataType::Float32);
 298     armnn::TensorInfo tensorInfo20x16({numUnits, outputSize}, armnn::DataType::Float32);
 299     armnn::TensorInfo tensorInfo16x20({outputSize, numUnits}, armnn::DataType::Float32);
 300
 301     auto inputToInputWeights =
 302             MakeTensor<float, 2>(tensorInfo20x5, {0.021393683f,0.06124551f,  0.046905167f,-0.014657677f,-0.03149463f,
 303                                                   0.09171803f, 0.14647801f,0.10797193f,   -0.0057968358f,0.0019193048f,
 304                                                   -0.2726754f, 0.10154029f, -0.018539885f, 0.080349885f, -0.10262385f,
 305                                                   -0.022599787f,-0.09121155f, -0.008675967f, -0.045206103f,-0.0821282f,
 306                                                   -0.008045952f,0.015478081f, 0.055217247f,  0.038719587f, 0.044153627f,
 307                                                   -0.06453243f,0.05031825f, -0.046935108f, -0.008164439f, 0.014574226f,
 308                                                   -0.1671009f,   -0.15519552f, -0.16819797f,-0.13971269f,-0.11953059f,
 309                                                   0.25005487f, -0.22790983f, 0.009855087f,  -0.028140958f, -0.11200698f,
 310                                                   0.11295408f, -0.0035217577f, 0.054485075f,  0.05184695f, 0.064711206f,
 311                                                   0.10989193f,   0.11674786f,  0.03490607f, 0.07727357f, 0.11390585f,
 312                                                   -0.1863375f,  -0.1034451f, -0.13945189f, -0.049401227f, -0.18767063f,
 313                                                   0.042483903f, 0.14233552f, 0.13832581f, 0.18350165f,    0.14545603f,
 314                                                   -0.028545704f,0.024939531f,0.050929718f,0.0076203286f,-0.0029723682f,
 315                                                   -0.042484224f, -0.11827596f, -0.09171104f,  -0.10808628f,-0.16327988f,
 316                                                   -0.2273378f,   -0.0993647f, -0.017155107f,0.0023917493f,0.049272764f,
 317                                                   0.0038534778f, 0.054764505f,   0.089753784f, 0.06947234f, 0.08014476f,
 318                                                   -0.04544234f, -0.0497073f,-0.07135631f,  -0.048929106f,-0.004042012f,
 319                                                   -0.009284026f, 0.018042054f, 0.0036860977f,-0.07427302f, -0.11434604f,
 320                                                   -0.018995456f, 0.031487543f, 0.012834908f,0.019977754f,0.044256654f,
 321                                                   -0.39292613f,  -0.18519334f, -0.11651281f,-0.06809892f, 0.011373677f
 322             });
 323
 324     auto inputToForgetWeights =
 325             MakeTensor<float, 2>(tensorInfo20x5, {-0.0018401089f, -0.004852237f,0.03698424f, 0.014181704f,0.028273236f,
 326                                                    -0.016726194f, -0.05249759f,-0.10204261f, 0.00861066f,-0.040979505f,
 327                                                    -0.009899187f,0.01923892f,-0.028177269f, -0.08535103f,-0.14585495f,
 328                                                    0.10662567f,-0.01909731f,-0.017883534f,-0.0047269356f,-0.045103323f,
 329                                                    0.0030784295f,0.076784775f,0.07463696f, 0.094531395f,0.0814421f,
 330                                                    -0.12257899f, -0.033945758f,-0.031303465f, 0.045630626f,0.06843887f,
 331                                                    -0.13492945f, -0.012480007f,-0.0811829f, -0.07224499f,-0.09628791f,
 332                                                    0.045100946f,0.0012300825f, 0.013964662f, 0.099372394f,0.02543059f,
 333                                                    0.06958324f,    0.034257296f, 0.0482646f, 0.06267997f,0.052625068f,
 334                                                    0.12784666f,    0.07077897f,  0.025725935f, 0.04165009f,0.07241905f,
 335                                                    0.018668644f, -0.037377294f,-0.06277783f,-0.08833636f,-0.040120605f,
 336                                                    -0.011405586f,-0.007808335f,-0.010301386f,-0.005102167f,0.027717464f,
 337                                                    0.05483423f, 0.11449111f, 0.11289652f,0.10939839f, 0.13396506f,
 338                                                    -0.08402166f,-0.01901462f,  -0.044678304f,-0.07720565f,0.014350063f,
 339                                                    -0.11757958f, -0.0652038f, -0.08185733f,-0.076754324f,-0.092614375f,
 340                                                    0.10405491f, 0.052960336f, 0.035755895f,0.035839386f,-0.012540553f,
 341                                                    0.036881298f,   0.02913376f,  0.03420159f,0.05448447f,-0.054523353f,
 342                                                    0.02582715f, 0.02327355f, -0.011857179f,-0.0011980024f,-0.034641717f,
 343                                                    -0.026125094f,-0.17582615f,-0.15923657f,-0.27486774f,-0.0006143371f,
 344                                                    0.0001771948f,  -8.470171e-05f, 0.02651807f,0.045790765f,0.06956496f
 345             });
 346
 347     auto inputToCellWeights =
 348             MakeTensor<float, 2>(tensorInfo20x5, {-0.04580283f,   -0.09549462f,   -0.032418985f,  -0.06454633f,
 349                                                   -0.043528453f,  0.043018587f,   -0.049152344f,  -0.12418144f,
 350                                                   -0.078985475f,  -0.07596889f,   0.019484362f,   -0.11434962f,
 351                                                   -0.0074034138f, -0.06314844f,   -0.092981495f,  0.0062155537f,
 352                                                   -0.025034338f,  -0.0028890965f, 0.048929527f,   0.06235075f,
 353                                                   0.10665918f,    -0.032036792f,  -0.08505916f,   -0.10843358f,
 354                                                   -0.13002433f,   -0.036816437f,  -0.02130134f,   -0.016518239f,
 355                                                   0.0047691227f,  -0.0025825808f, 0.066017866f,   0.029991534f,
 356                                                   -0.10652836f,   -0.1037554f,    -0.13056071f,   -0.03266643f,
 357                                                   -0.033702414f,  -0.006473424f,  -0.04611692f,   0.014419339f,
 358                                                   -0.025174323f,  0.0396852f,     0.081777506f,   0.06157468f,
 359                                                   0.10210095f,    -0.009658194f,  0.046511717f,   0.03603906f,
 360                                                   0.0069369148f,  0.015960095f,   -0.06507666f,   0.09551598f,
 361                                                   0.053568836f,   0.06408714f,    0.12835667f,    -0.008714329f,
 362                                                   -0.20211966f,   -0.12093674f,   0.029450472f,   0.2849013f,
 363                                                   -0.029227901f,  0.1164364f,     -0.08560263f,   0.09941786f,
 364                                                   -0.036999565f,  -0.028842626f,  -0.0033637602f, -0.017012902f,
 365                                                   -0.09720865f,   -0.11193351f,   -0.029155117f,  -0.017936034f,
 366                                                   -0.009768936f,  -0.04223324f,   -0.036159635f,  0.06505112f,
 367                                                   -0.021742892f,  -0.023377212f,  -0.07221364f,   -0.06430552f,
 368                                                   0.05453865f,    0.091149814f,   0.06387331f,    0.007518393f,
 369                                                   0.055960953f,   0.069779344f,   0.046411168f,   0.10509911f,
 370                                                   0.07463894f,    0.0075130584f,  0.012850982f,   0.04555431f,
 371                                                   0.056955688f,   0.06555285f,    0.050801456f,   -0.009862683f,
 372                                                   0.00826772f,    -0.026555609f,  -0.0073611983f, -0.0014897042f
 373             });
 374
 375     auto inputToOutputWeights =
 376             MakeTensor<float, 2>(tensorInfo20x5, {-0.0998932f,   -0.07201956f, -0.052803773f,-0.15629593f,-0.15001918f,
 377                                                   -0.07650751f,0.02359855f, -0.075155355f, -0.08037709f,  -0.15093534f,
 378                                                   0.029517552f, -0.04751393f, 0.010350531f,-0.02664851f, -0.016839722f,
 379                                                   -0.023121163f, 0.0077019283f, 0.012851257f, -0.05040649f,-0.0129761f,
 380                                                   -0.021737747f,-0.038305793f,-0.06870586f, -0.01481247f,-0.001285394f,
 381                                                   0.10124236f,  0.083122835f, 0.053313006f,-0.062235646f,-0.075637154f,
 382                                                   -0.027833903f, 0.029774971f,  0.1130802f, 0.09218906f, 0.09506135f,
 383                                                   -0.086665764f,-0.037162706f,-0.038880914f,-0.035832845f,-0.014481564f,
 384                                                   -0.09825003f,-0.12048569f,-0.097665586f,-0.05287633f, -0.0964047f,
 385                                                   -0.11366429f,  0.035777505f,  0.13568819f, 0.052451383f,0.050649304f,
 386                                                   0.05798951f, -0.021852335f,-0.099848844f,0.014740475f,-0.078897946f,
 387                                                   0.04974699f, 0.014160473f,  0.06973932f,    0.04964942f, 0.033364646f,
 388                                                   0.08190124f,   0.025535367f, 0.050893165f, 0.048514254f,0.06945813f,
 389                                                   -0.078907564f,-0.06707616f,  -0.11844508f, -0.09986688f,-0.07509403f,
 390                                                   0.06263226f,   0.14925587f,   0.20188436f, 0.12098451f,0.14639415f,
 391                                                   0.0015017595f, -0.014267382f, -0.03417257f,0.012711468f,0.0028300495f,
 392                                                   -0.024758482f, -0.05098548f,-0.0821182f, 0.014225672f,  0.021544158f,
 393                                                   0.08949725f,  0.07505268f, -0.0020780868f, 0.04908258f,0.06476295f,
 394                                                   -0.022907063f,0.027562456f,0.040185735f, 0.019567577f,-0.015598739f,
 395                                                   -0.049097303f, -0.017121866f, -0.083368234f,-0.02332002f,-0.0840956f
 396             });
 397
 398     auto inputGateBias =
 399             MakeTensor<float, 1>(tensorInfo20, {0.02234832f,  0.14757581f,   0.18176508f,  0.10380666f,  0.053110216f,
 400                                                 -0.06928846f, -0.13942584f,  -0.11816189f, 0.19483899f,  0.03652339f,
 401                                                 -0.10250295f, 0.036714908f,  -0.18426876f, 0.036065217f, 0.21810818f,
 402                                                 0.02383196f,  -0.043370757f, 0.08690144f,  -0.04444982f, 0.00030581196f
 403             });
 404
 405     auto forgetGateBias =
 406             MakeTensor<float, 1>(tensorInfo20, {0.035185695f, -0.042891346f, -0.03032477f, 0.23027696f,
 407                                                 0.11098921f,  0.15378423f,   0.09263801f,  0.09790885f,
 408                                                 0.09508917f,  0.061199076f,  0.07665568f,  -0.015443159f,
 409                                                 -0.03499149f, 0.046190713f,  0.08895977f,  0.10899629f,
 410                                                 0.40694186f,  0.06030037f,   0.012413437f, -0.06108739f
 411             });
 412
 413     auto cellBias =
 414             MakeTensor<float, 1>(tensorInfo20, {-0.024379363f, 0.0055531194f, 0.23377132f,   0.033463873f,
 415                                                 -0.1483596f,   -0.10639995f,  -0.091433935f, 0.058573797f,
 416                                                 -0.06809782f,  -0.07889636f,  -0.043246906f, -0.09829136f,
 417                                                 -0.4279842f,   0.034901652f,  0.18797937f,   0.0075234566f,
 418                                                 0.016178843f,  0.1749513f,    0.13975595f,   0.92058027f
 419             });
 420
 421     auto outputGateBias =
 422             MakeTensor<float, 1>(tensorInfo20, {0.046159424f,  -0.0012809046f, 0.03563469f, 0.12648113f, 0.027195795f,
 423                                                 0.35373217f,   -0.018957434f,  0.008907322f, -0.0762701f, 0.12018895f,
 424                                                 0.04216877f,   0.0022856654f,  0.040952638f,  0.3147856f,  0.08225149f,
 425                                                 -0.057416286f, -0.14995944f,   -0.008040261f, 0.13208859f, 0.029760877f
 426             });
 427
 428     auto recurrentToInputWeights =
 429             MakeTensor<float, 2>(tensorInfo20x16, {-0.001374326f,   -0.078856036f,   0.10672688f,    0.029162422f,
 430                                                    -0.11585556f,    0.02557986f,     -0.13446963f,   -0.035785314f,
 431                                                    -0.01244275f,    0.025961924f,    -0.02337298f,   -0.044228926f,
 432                                                    -0.055839065f,   -0.046598054f,   -0.010546039f,  -0.06900766f,
 433                                                    0.027239809f,    0.022582639f,    -0.013296484f,  -0.05459212f,
 434                                                    0.08981f,        -0.045407712f,   0.08682226f,    -0.06867011f,
 435                                                    -0.14390695f,    -0.02916037f,    0.000996957f,   0.091420636f,
 436                                                    0.14283475f,     -0.07390571f,    -0.06402044f,   0.062524505f,
 437                                                    -0.093129106f,   0.04860203f,     -0.08364217f,   -0.08119002f,
 438                                                    0.009352075f,    0.22920375f,     0.0016303885f,  0.11583097f,
 439                                                    -0.13732095f,    0.012405723f,    -0.07551853f,   0.06343048f,
 440                                                    0.12162708f,     -0.031923793f,   -0.014335606f,  0.01790974f,
 441                                                    -0.10650317f,    -0.0724401f,     0.08554849f,    -0.05727212f,
 442                                                    0.06556731f,     -0.042729504f,   -0.043227166f,  0.011683251f,
 443                                                    -0.013082158f,   -0.029302018f,   -0.010899579f,  -0.062036745f,
 444                                                    -0.022509435f,   -0.00964907f,    -0.01567329f,   0.04260106f,
 445                                                    -0.07787477f,    -0.11576462f,    0.017356863f,   0.048673786f,
 446                                                    -0.017577527f,   -0.05527947f,    -0.082487635f,  -0.040137455f,
 447                                                    -0.10820036f,    -0.04666372f,    0.022746278f,   -0.07851417f,
 448                                                    0.01068115f,     0.032956902f,    0.022433773f,   0.0026891115f,
 449                                                    0.08944216f,     -0.0685835f,     0.010513544f,   0.07228705f,
 450                                                    0.02032331f,     -0.059686817f,   -0.0005566496f, -0.086984694f,
 451                                                    0.040414046f,    -0.1380399f,     0.094208956f,   -0.05722982f,
 452                                                    0.012092817f,    -0.04989123f,    -0.086576f,     -0.003399834f,
 453                                                    -0.04696032f,    -0.045747425f,   0.10091314f,    0.048676282f,
 454                                                    -0.029037097f,   0.031399418f,    -0.0040285117f, 0.047237843f,
 455                                                    0.09504992f,     0.041799378f,    -0.049185462f,  -0.031518843f,
 456                                                    -0.10516937f,    0.026374253f,    0.10058866f,    -0.0033195973f,
 457                                                    -0.041975245f,   0.0073591834f,   0.0033782164f,  -0.004325073f,
 458                                                    -0.10167381f,    0.042500053f,    -0.01447153f,   0.06464186f,
 459                                                    -0.017142897f,   0.03312627f,     0.009205989f,   0.024138335f,
 460                                                    -0.011337001f,   0.035530265f,    -0.010912711f,  0.0706555f,
 461                                                    -0.005894094f,   0.051841937f,    -0.1401738f,    -0.02351249f,
 462                                                    0.0365468f,      0.07590991f,     0.08838724f,    0.021681072f,
 463                                                    -0.10086113f,    0.019608743f,    -0.06195883f,   0.077335775f,
 464                                                    0.023646897f,    -0.095322326f,   0.02233014f,    0.09756986f,
 465                                                    -0.048691444f,   -0.009579111f,   0.07595467f,    0.11480546f,
 466                                                    -0.09801813f,    0.019894179f,    0.08502348f,    0.004032281f,
 467                                                    0.037211012f,    0.068537936f,    -0.048005626f,  -0.091520436f,
 468                                                    -0.028379958f,   -0.01556313f,    0.06554592f,    -0.045599163f,
 469                                                    -0.01672207f,    -0.020169014f,   -0.011877351f,  -0.20212261f,
 470                                                    0.010889619f,    0.0047078193f,   0.038385306f,   0.08540671f,
 471                                                    -0.017140968f,   -0.0035865551f,  0.016678626f,   0.005633034f,
 472                                                    0.015963363f,    0.00871737f,     0.060130805f,   0.028611384f,
 473                                                    0.10109069f,     -0.015060172f,   -0.07894427f,   0.06401885f,
 474                                                    0.011584063f,    -0.024466386f,   0.0047652307f,  -0.09041358f,
 475                                                    0.030737216f,    -0.0046374933f,  0.14215417f,    -0.11823516f,
 476                                                    0.019899689f,    0.006106124f,    -0.027092824f,  0.0786356f,
 477                                                    0.05052217f,     -0.058925f,      -0.011402121f,  -0.024987547f,
 478                                                    -0.0013661642f,  -0.06832946f,    -0.015667673f,  -0.1083353f,
 479                                                    -0.00096863037f, -0.06988685f,    -0.053350925f,  -0.027275559f,
 480                                                    -0.033664223f,   -0.07978348f,    -0.025200296f,  -0.017207067f,
 481                                                    -0.058403496f,   -0.055697463f,   0.005798788f,   0.12965427f,
 482                                                    -0.062582195f,   0.0013350133f,   -0.10482091f,   0.0379771f,
 483                                                    0.072521195f,    -0.0029455067f,  -0.13797039f,   -0.03628521f,
 484                                                    0.013806405f,    -0.017858358f,   -0.01008298f,   -0.07700066f,
 485                                                    -0.017081132f,   0.019358726f,    0.0027079724f,  0.004635139f,
 486                                                    0.062634714f,    -0.02338735f,    -0.039547626f,  -0.02050681f,
 487                                                    0.03385117f,     -0.083611414f,   0.002862572f,   -0.09421313f,
 488                                                    0.058618143f,    -0.08598433f,    0.00972939f,    0.023867095f,
 489                                                    -0.053934585f,   -0.023203006f,   0.07452513f,    -0.048767887f,
 490                                                    -0.07314807f,    -0.056307215f,   -0.10433547f,   -0.06440842f,
 491                                                    0.04328182f,     0.04389765f,     -0.020006588f,  -0.09076438f,
 492                                                    -0.11652589f,    -0.021705797f,   0.03345259f,    -0.010329105f,
 493                                                    -0.025767034f,   0.013057034f,    -0.07316461f,   -0.10145612f,
 494                                                    0.06358255f,     0.18531723f,     0.07759293f,    0.12006465f,
 495                                                    0.1305557f,      0.058638252f,    -0.03393652f,   0.09622831f,
 496                                                    -0.16253184f,    -2.4580743e-06f, 0.079869635f,   -0.070196845f,
 497                                                    -0.005644518f,   0.06857898f,     -0.12598175f,   -0.035084512f,
 498                                                    0.03156317f,     -0.12794146f,    -0.031963028f,  0.04692781f,
 499                                                    0.030070418f,    0.0071660685f,   -0.095516115f,  -0.004643372f,
 500                                                    0.040170413f,    -0.062104587f,   -0.0037324072f, 0.0554317f,
 501                                                    0.08184801f,     -0.019164372f,   0.06791302f,    0.034257166f,
 502                                                    -0.10307039f,    0.021943003f,    0.046745934f,   0.0790918f,
 503                                                    -0.0265588f,     -0.007824208f,   0.042546265f,   -0.00977924f,
 504                                                    -0.0002440307f,  -0.017384544f,   -0.017990116f,  0.12252321f,
 505                                                    -0.014512694f,   -0.08251313f,    0.08861942f,    0.13589665f,
 506                                                    0.026351685f,    0.012641483f,    0.07466548f,    0.044301085f,
 507                                                    -0.045414884f,   -0.051112458f,   0.03444247f,    -0.08502782f,
 508                                                    -0.04106223f,    -0.028126027f,   0.028473156f,   0.10467447f
 509             });
 510
 511     auto recurrentToForgetWeights =
 512             MakeTensor<float, 2>(tensorInfo20x16, {-0.057784554f,  -0.026057621f,  -0.068447545f,   -0.022581743f,
 513                                                    0.14811787f,    0.10826372f,    0.09471067f,     0.03987225f,
 514                                                    -0.0039523416f, 0.00030638507f, 0.053185795f,    0.10572994f,
 515                                                    0.08414449f,    -0.022036452f,  -0.00066928595f, -0.09203576f,
 516                                                    0.032950465f,   -0.10985798f,   -0.023809856f,   0.0021431844f,
 517                                                    -0.02196096f,   -0.00326074f,   0.00058621005f,  -0.074678116f,
 518                                                    -0.06193199f,   0.055729095f,   0.03736828f,     0.020123724f,
 519                                                    0.061878487f,   -0.04729229f,   0.034919553f,    -0.07585433f,
 520                                                    -0.04421272f,   -0.044019096f,  0.085488975f,    0.04058006f,
 521                                                    -0.06890133f,   -0.030951202f,  -0.024628663f,   -0.07672815f,
 522                                                    0.034293607f,   0.08556707f,    -0.05293577f,    -0.033561368f,
 523                                                    -0.04899627f,   0.0241671f,     0.015736353f,    -0.095442444f,
 524                                                    -0.029564252f,  0.016493602f,   -0.035026584f,   0.022337519f,
 525                                                    -0.026871363f,  0.004780428f,   0.0077918363f,   -0.03601621f,
 526                                                    0.016435321f,   -0.03263031f,   -0.09543275f,    -0.047392778f,
 527                                                    0.013454138f,   0.028934088f,   0.01685226f,     -0.086110644f,
 528                                                    -0.046250615f,  -0.01847454f,   0.047608484f,    0.07339695f,
 529                                                    0.034546845f,   -0.04881143f,   0.009128804f,    -0.08802852f,
 530                                                    0.03761666f,    0.008096139f,   -0.014454086f,   0.014361001f,
 531                                                    -0.023502491f,  -0.0011840804f, -0.07607001f,    0.001856849f,
 532                                                    -0.06509276f,   -0.006021153f,  -0.08570962f,    -0.1451793f,
 533                                                    0.060212336f,   0.055259194f,   0.06974018f,     0.049454916f,
 534                                                    -0.027794661f,  -0.08077226f,   -0.016179763f,   0.1169753f,
 535                                                    0.17213494f,    -0.0056326236f, -0.053934924f,   -0.0124349f,
 536                                                    -0.11520337f,   0.05409887f,    0.088759385f,    0.0019655675f,
 537                                                    0.0042065294f,  0.03881498f,    0.019844765f,    0.041858196f,
 538                                                    -0.05695512f,   0.047233116f,   0.038937137f,    -0.06542224f,
 539                                                    0.014429736f,   -0.09719407f,   0.13908425f,     -0.05379757f,
 540                                                    0.012321099f,   0.082840554f,   -0.029899208f,   0.044217527f,
 541                                                    0.059855383f,   0.07711018f,    -0.045319796f,   0.0948846f,
 542                                                    -0.011724666f,  -0.0033288454f, -0.033542685f,   -0.04764985f,
 543                                                    -0.13873616f,   0.040668588f,   0.034832682f,    -0.015319203f,
 544                                                    -0.018715994f,  0.046002675f,   0.0599172f,      -0.043107376f,
 545                                                    0.0294216f,     -0.002314414f,  -0.022424703f,   0.0030315618f,
 546                                                    0.0014641669f,  0.0029166266f,  -0.11878115f,    0.013738511f,
 547                                                    0.12375372f,    -0.0006038222f, 0.029104086f,    0.087442465f,
 548                                                    0.052958444f,   0.07558703f,    0.04817258f,     0.044462286f,
 549                                                    -0.015213451f,  -0.08783778f,   -0.0561384f,     -0.003008196f,
 550                                                    0.047060397f,   -0.002058388f,  0.03429439f,     -0.018839769f,
 551                                                    0.024734668f,   0.024614193f,   -0.042046934f,   0.09597743f,
 552                                                    -0.0043254104f, 0.04320769f,    0.0064070094f,   -0.0019131786f,
 553                                                    -0.02558259f,   -0.022822596f,  -0.023273505f,   -0.02464396f,
 554                                                    -0.10991725f,   -0.006240552f,  0.0074488563f,   0.024044557f,
 555                                                    0.04383914f,    -0.046476185f,  0.028658995f,    0.060410924f,
 556                                                    0.050786525f,   0.009452605f,   -0.0073054377f,  -0.024810238f,
 557                                                    0.0052906186f,  0.0066939713f,  -0.0020913032f,  0.014515517f,
 558                                                    0.015898481f,   0.021362653f,   -0.030262267f,   0.016587038f,
 559                                                    -0.011442813f,  0.041154444f,   -0.007631438f,   -0.03423484f,
 560                                                    -0.010977775f,  0.036152758f,   0.0066366293f,   0.11915515f,
 561                                                    0.02318443f,    -0.041350313f,  0.021485701f,    -0.10906167f,
 562                                                    -0.028218046f,  -0.00954771f,   0.020531068f,    -0.11995105f,
 563                                                    -0.03672871f,   0.024019798f,   0.014255957f,    -0.05221243f,
 564                                                    -0.00661567f,   -0.04630967f,   0.033188973f,    0.10107534f,
 565                                                    -0.014027541f,  0.030796422f,   -0.10270911f,    -0.035999842f,
 566                                                    0.15443139f,    0.07684145f,    0.036571592f,    -0.035900835f,
 567                                                    -0.0034699554f, 0.06209149f,    0.015920248f,    -0.031122351f,
 568                                                    -0.03858649f,   0.01849943f,    0.13872518f,     0.01503974f,
 569                                                    0.069941424f,   -0.06948533f,   -0.0088794185f,  0.061282158f,
 570                                                    -0.047401894f,  0.03100163f,    -0.041533746f,   -0.10430945f,
 571                                                    0.044574402f,   -0.01425562f,   -0.024290353f,   0.034563623f,
 572                                                    0.05866852f,    0.023947537f,   -0.09445152f,    0.035450947f,
 573                                                    0.02247216f,    -0.0042998926f, 0.061146557f,    -0.10250651f,
 574                                                    0.020881841f,   -0.06747029f,   0.10062043f,     -0.0023941975f,
 575                                                    0.03532124f,    -0.016341697f,  0.09685456f,     -0.016764693f,
 576                                                    0.051808182f,   0.05875331f,    -0.04536488f,    0.001626336f,
 577                                                    -0.028892258f,  -0.01048663f,   -0.009793449f,   -0.017093895f,
 578                                                    0.010987891f,   0.02357273f,    -0.00010856845f, 0.0099760275f,
 579                                                    -0.001845119f,  -0.03551521f,   0.0018358806f,   0.05763657f,
 580                                                    -0.01769146f,   0.040995963f,   0.02235177f,     -0.060430344f,
 581                                                    0.11475477f,    -0.023854522f,  0.10071741f,     0.0686208f,
 582                                                    -0.014250481f,  0.034261297f,   0.047418304f,    0.08562733f,
 583                                                    -0.030519066f,  0.0060542435f,  0.014653856f,    -0.038836084f,
 584                                                    0.04096551f,    0.032249358f,   -0.08355519f,    -0.026823482f,
 585                                                    0.056386515f,   -0.010401743f,  -0.028396193f,   0.08507674f,
 586                                                    0.014410365f,   0.020995233f,   0.17040324f,     0.11511526f,
 587                                                    0.02459721f,    0.0066619175f,  0.025853224f,    -0.023133837f,
 588                                                    -0.081302024f,  0.017264642f,   -0.009585969f,   0.09491168f,
 589                                                    -0.051313367f,  0.054532815f,   -0.014298593f,   0.10657464f,
 590                                                    0.007076659f,   0.10964551f,    0.0409152f,      0.008275321f,
 591                                                    -0.07283536f,   0.07937492f,    0.04192024f,     -0.1075027f
 592             });
 593
 594     auto recurrentToCellWeights =
 595             MakeTensor<float, 2>(tensorInfo20x16, {-0.037322544f,   0.018592842f,   0.0056175636f,  -0.06253426f,
 596                                                    0.055647098f,    -0.05713207f,   -0.05626563f,   0.005559383f,
 597                                                    0.03375411f,     -0.025757805f,  -0.088049285f,  0.06017052f,
 598                                                    -0.06570978f,    0.007384076f,   0.035123326f,   -0.07920549f,
 599                                                    0.053676967f,    0.044480428f,   -0.07663568f,   0.0071805613f,
 600                                                    0.08089997f,     0.05143358f,    0.038261272f,   0.03339287f,
 601                                                    -0.027673481f,   0.044746667f,   0.028349208f,   0.020090483f,
 602                                                    -0.019443132f,   -0.030755889f,  -0.0040000007f, 0.04465846f,
 603                                                    -0.021585021f,   0.0031670958f,  0.0053199246f,  -0.056117613f,
 604                                                    -0.10893326f,    0.076739706f,   -0.08509834f,   -0.027997585f,
 605                                                    0.037871376f,    0.01449768f,    -0.09002357f,   -0.06111149f,
 606                                                    -0.046195522f,   0.0422062f,     -0.005683705f,  -0.1253618f,
 607                                                    -0.012925729f,   -0.04890792f,   0.06985068f,    0.037654128f,
 608                                                    0.03398274f,     -0.004781977f,  0.007032333f,   -0.031787455f,
 609                                                    0.010868644f,    -0.031489216f,  0.09525667f,    0.013939797f,
 610                                                    0.0058680447f,   0.0167067f,     0.02668468f,    -0.04797466f,
 611                                                    -0.048885044f,   -0.12722108f,   0.035304096f,   0.06554885f,
 612                                                    0.00972396f,     -0.039238118f,  -0.05159735f,   -0.11329045f,
 613                                                    0.1613692f,      -0.03750952f,   0.06529313f,    -0.071974665f,
 614                                                    -0.11769596f,    0.015524369f,   -0.0013754242f, -0.12446318f,
 615                                                    0.02786344f,     -0.014179351f,  0.005264273f,   0.14376344f,
 616                                                    0.015983658f,    0.03406988f,    -0.06939408f,   0.040699873f,
 617                                                    0.02111075f,     0.09669095f,    0.041345075f,   -0.08316494f,
 618                                                    -0.07684199f,    -0.045768797f,  0.032298047f,   -0.041805092f,
 619                                                    0.0119405f,      0.0061010392f,  0.12652606f,    0.0064572375f,
 620                                                    -0.024950314f,   0.11574242f,    0.04508852f,    -0.04335324f,
 621                                                    0.06760663f,     -0.027437469f,  0.07216407f,    0.06977076f,
 622                                                    -0.05438599f,    0.034033038f,   -0.028602652f,  0.05346137f,
 623                                                    0.043184172f,    -0.037189785f,  0.10420091f,    0.00882477f,
 624                                                    -0.054019816f,   -0.074273005f,  -0.030617684f,  -0.0028467078f,
 625                                                    0.024302477f,    -0.0038869337f, 0.005332455f,   0.0013399826f,
 626                                                    0.04361412f,     -0.007001822f,  0.09631092f,    -0.06702025f,
 627                                                    -0.042049985f,   -0.035070654f,  -0.04103342f,   -0.10273396f,
 628                                                    0.0544271f,      0.037184782f,   -0.13150354f,   -0.0058036847f,
 629                                                    -0.008264958f,   0.042035464f,   0.05891794f,    0.029673764f,
 630                                                    0.0063542654f,   0.044788733f,   0.054816857f,   0.062257513f,
 631                                                    -0.00093483756f, 0.048938446f,   -0.004952862f,  -0.007730018f,
 632                                                    -0.04043371f,    -0.017094059f,  0.07229206f,    -0.023670016f,
 633                                                    -0.052195564f,   -0.025616996f,  -0.01520939f,   0.045104615f,
 634                                                    -0.007376126f,   0.003533447f,   0.006570588f,   0.056037236f,
 635                                                    0.12436656f,     0.051817212f,   0.028532185f,   -0.08686856f,
 636                                                    0.11868599f,     0.07663395f,    -0.07323171f,   0.03463402f,
 637                                                    -0.050708205f,   -0.04458982f,   -0.11590894f,   0.021273347f,
 638                                                    0.1251325f,      -0.15313013f,   -0.12224372f,   0.17228661f,
 639                                                    0.023029093f,    0.086124025f,   0.006445803f,   -0.03496501f,
 640                                                    0.028332196f,    0.04449512f,    -0.042436164f,  -0.026587414f,
 641                                                    -0.006041347f,   -0.09292539f,   -0.05678812f,   0.03897832f,
 642                                                    0.09465633f,     0.008115513f,   -0.02171956f,   0.08304309f,
 643                                                    0.071401566f,    0.019622514f,   0.032163795f,   -0.004167056f,
 644                                                    0.02295182f,     0.030739572f,   0.056506045f,   0.004612461f,
 645                                                    0.06524936f,     0.059999723f,   0.046395954f,   -0.0045512207f,
 646                                                    -0.1335546f,     -0.030136576f,  0.11584653f,    -0.014678886f,
 647                                                    0.0020118146f,   -0.09688814f,   -0.0790206f,    0.039770417f,
 648                                                    -0.0329582f,     0.07922767f,    0.029322514f,   0.026405897f,
 649                                                    0.04207835f,     -0.07073373f,   0.063781224f,   0.0859677f,
 650                                                    -0.10925287f,    -0.07011058f,   0.048005477f,   0.03438226f,
 651                                                    -0.09606514f,    -0.006669445f,  -0.043381985f,  0.04240257f,
 652                                                    -0.06955775f,    -0.06769346f,   0.043903265f,   -0.026784198f,
 653                                                    -0.017840602f,   0.024307009f,   -0.040079936f,  -0.019946516f,
 654                                                    0.045318738f,    -0.12233574f,   0.026170589f,   0.0074471775f,
 655                                                    0.15978073f,     0.10185836f,    0.10298046f,    -0.015476589f,
 656                                                    -0.039390966f,   -0.072174534f,  0.0739445f,     -0.1211869f,
 657                                                    -0.0347889f,     -0.07943156f,   0.014809798f,   -0.12412325f,
 658                                                    -0.0030663363f,  0.039695457f,   0.0647603f,     -0.08291318f,
 659                                                    -0.018529687f,   -0.004423833f,  0.0037507233f,  0.084633216f,
 660                                                    -0.01514876f,    -0.056505352f,  -0.012800942f,  -0.06994386f,
 661                                                    0.012962922f,    -0.031234352f,  0.07029052f,    0.016418684f,
 662                                                    0.03618972f,     0.055686004f,   -0.08663945f,   -0.017404709f,
 663                                                    -0.054761406f,   0.029065743f,   0.052404847f,   0.020238016f,
 664                                                    0.0048197987f,   -0.0214882f,    0.07078733f,    0.013016777f,
 665                                                    0.06262858f,     0.009184685f,   0.020785125f,   -0.043904778f,
 666                                                    -0.0270329f,     -0.03299152f,   -0.060088247f,  -0.015162964f,
 667                                                    -0.001828936f,   0.12642565f,    -0.056757294f,  0.013586685f,
 668                                                    0.09232601f,     -0.035886683f,  0.06000002f,    0.05229691f,
 669                                                    -0.052580316f,   -0.082029596f,  -0.010794592f,  0.012947712f,
 670                                                    -0.036429964f,   -0.085508935f,  -0.13127148f,   -0.017744139f,
 671                                                    0.031502828f,    0.036232427f,   -0.031581745f,  0.023051167f,
 672                                                    -0.05325106f,    -0.03421577f,   0.028793324f,   -0.034633752f,
 673                                                    -0.009881397f,   -0.043551125f,  -0.018609839f,  0.0019097115f,
 674                                                    -0.008799762f,   0.056595087f,   0.0022273948f,  0.055752404f
 675             });
 676
 677     auto recurrentToOutputWeights =
 678             MakeTensor<float, 2>(tensorInfo20x16, {0.025825322f, -0.05813119f, 0.09495884f,-0.045984812f, -0.01255415f,
 679                                                     -0.0026479573f,-0.08196161f,-0.054914974f,-0.0046604523f,
 680                                                    -0.029587349f, -0.044576716f,  -0.07480124f,  -0.082868785f,
 681                                                    0.023254942f,    0.027502948f, -0.0039728214f, -0.08683098f,
 682                                                    -0.08116779f,  -0.014675607f,   -0.037924774f, -0.023314456f,
 683                                                    -0.007401714f, -0.09255757f,  0.029460307f,    -0.08829125f,
 684                                                     -0.005139627f,  -0.08989442f,  -0.0555066f,   0.13596267f,
 685                                                    -0.025062224f, -0.048351806f,  -0.03850004f,  0.07266485f,
 686                                                    -0.022414139f,   0.05940088f, 0.075114764f,   0.09597592f,
 687                                                    -0.010211725f, -0.0049794707f,  -0.011523867f, -0.025980417f,
 688                                                    0.072999895f,  0.11091378f,   -0.081685916f,   0.014416728f,
 689                                                     0.043229222f,   0.034178585f,  -0.07530371f,  0.035837382f,
 690                                                    -0.085607f, -0.007721233f,  -0.03287832f,  -0.043848954f,
 691                                                    -0.06404588f,    -0.06632928f, -0.073643476f,  0.008214239f,
 692                                                    -0.045984086f, 0.039764922f,    0.03474462f, 0.060612556f,
 693                                                    -0.080590084f, 0.049127717f,  0.04151091f,     -0.030063879f,
 694                                                     0.008801774f,   -0.023021035f, -0.019558564f, 0.05158114f,
 695                                                    -0.010947698f, -0.011825728f,  0.0075720972f, 0.0699727f,
 696                                                    -0.0039981045f,  0.069350146f, 0.08799282f,    0.016156472f,
 697                                                    0.035502106f,  0.11695009f,     0.006217345f, 0.13392477f,
 698                                                    -0.037875112f, 0.025745004f,  0.08940699f,     -0.00924166f,
 699                                                     0.0046702605f,  -0.036598757f, -0.08811812f,  0.10522024f,
 700                                                    -0.032441203f, 0.008176899f,   -0.04454919f,  0.07058152f,
 701                                                    0.0067963637f,   0.039206743f, 0.03259838f,    0.03725492f,
 702                                                    -0.09515802f,  0.013326398f,    -0.052055415f, -0.025676316f,
 703                                                    0.03198509f,   -0.015951829f, -0.058556724f,   0.036879618f,
 704                                                     0.043357447f,   0.028362012f,  -0.05908629f,  0.0059240665f,
 705                                                    -0.04995891f, -0.019187413f,0.0276265f, -0.01628143f, 0.0025863599f,
 706                                                    0.08800015f, 0.035250366f,   -0.022165963f, -0.07328642f,
 707                                                    -0.009415526f,   -0.07455109f, 0.11690406f,    0.0363299f,
 708                                                    0.07411125f,   0.042103454f,    -0.009660886f, 0.019076364f,
 709                                                    0.018299393f, -0.046004917f, 0.08891175f,0.0431396f, -0.026327137f,
 710                                                    -0.051502608f, 0.08979574f,   -0.051670972f,   0.04940282f,
 711                                                     -0.07491107f,   -0.021240504f, 0.022596184f,  -0.034280192f,
 712                                                    0.060163025f, -0.058211457f,  -0.051837247f, -0.01349775f,
 713                                                    -0.04639988f,    -0.035936575f, -0.011681591f,  0.064818054f,
 714                                                    0.0073146066f, -0.021745546f,   -0.043124277f, -0.06471268f,
 715                                                    -0.07053354f,  -0.029321948f, -0.05330136f,    0.016933719f,
 716                                                     -0.053782392f,  0.13747959f,   -0.1361751f,   -0.11569455f,
 717                                                    0.0033329215f, 0.05693899f,    -0.053219706f, 0.063698f,
 718                                                    0.07977434f,     -0.07924483f, 0.06936997f,    0.0034815092f,
 719                                                    -0.007305279f, -0.037325785f,   -0.07251102f, -0.033633437f,
 720                                                    -0.08677009f,  0.091591336f,  -0.14165086f,    0.021752775f,
 721                                                     0.019683983f,   0.0011612234f, -0.058154266f, 0.049996935f,
 722                                                    0.0288841f, -0.0024567875f, -0.14345716f, 0.010955264f,-0.10234828f,
 723                                                    0.1183656f, -0.0010731248f, -0.023590032f,-0.072285876f,-0.0724771f,
 724                                                    -0.026382286f, -0.0014920527f, 0.042667855f,  0.0018776858f,
 725                                                    0.02986552f,     0.009814309f, 0.0733756f,     0.12289186f,
 726                                                    0.018043943f,  -0.0458958f,     0.049412545f, 0.033632483f,
 727                                                    0.05495232f,   0.036686596f,  -0.013781798f,   -0.010036754f,
 728                                                     0.02576849f,    -0.08307328f,  0.010112348f,  0.042521734f,
 729                                                    -0.05869831f, -0.071689695f, 0.03876447f, -0.13275425f, -0.0352966f,
 730                                                    -0.023077697f, 0.10285965f,    0.084736146f,  0.15568255f,
 731                                                    -0.00040734606f, 0.027835453f, -0.10292561f,   -0.032401145f,
 732                                                    0.10053256f,   -0.026142767f,   -0.08271222f, -0.0030240538f,
 733                                                    -0.016368777f, 0.1070414f,    0.042672627f,    0.013456989f,
 734                                                     -0.0437609f,    -0.022309763f, 0.11576483f,   0.04108048f,
 735                                                    0.061026827f, -0.0190714f,  -0.0869359f, 0.037901703f,  0.0610107f,
 736                                                    0.07202949f, 0.01675338f,    0.086139716f,  -0.08795751f,
 737                                                    -0.014898893f,   -0.023771819f, -0.01965048f,   0.007955471f,
 738                                                    -0.043740474f, 0.03346837f,     -0.10549954f, 0.090567775f,
 739                                                    0.042013682f,  -0.03176985f,  0.12569028f,     -0.02421228f,
 740                                                     -0.029526481f,  0.023851605f,  0.031539805f,  0.05292009f,
 741                                                    -0.02344001f, -0.07811758f,   -0.08834428f,  0.10094801f,
 742                                                    0.16594367f,     -0.06861939f, -0.021256343f,  -0.041093912f,
 743                                                    -0.06669611f,  0.035498552f,    0.021757556f, -0.09302526f,
 744                                                    -0.015403468f, -0.06614931f,  -0.051798206f,   -0.013874718f,
 745                                                     0.03630673f,    0.010412845f,  -0.08077351f,  0.046185967f,
 746                                                    0.0035662893f, 0.03541868f,    -0.094149634f, -0.034814864f,
 747                                                    0.003128424f,    -0.020674974f, -0.03944324f,   -0.008110165f,
 748                                                    -0.11113267f,  0.08484226f,     0.043586485f, 0.040582247f,
 749                                                    0.0968012f,    -0.065249965f, -0.028036479f,   0.0050708856f,
 750                                                     0.0017462453f,  0.0326779f,    0.041296225f,  0.09164146f,
 751                                                    -0.047743853f, -0.015952192f,  -0.034451712f, 0.084197424f,
 752                                                    -0.05347844f,    -0.11768019f, 0.085926116f,   -0.08251791f,
 753                                                    -0.045081906f, 0.0948852f,      0.068401024f, 0.024856757f,
 754                                                    0.06978981f,   -0.057309967f, -0.012775832f,   -0.0032452994f,
 755                                                     0.01977615f, -0.041040014f, -0.024264973f,0.063464895f, 0.05431621f
 756             });
 757
 758     auto cellToInputWeights =
 759             MakeTensor<float, 1>(tensorInfo20, {0.040369894f, 0.030746894f,  0.24704495f,  0.018586371f, -0.037586458f,
 760                                                 -0.15312155f, -0.11812848f,  -0.11465643f, 0.20259799f,   0.11418174f,
 761                                                 -0.10116027f, -0.011334949f, 0.12411352f, -0.076769054f,-0.052169047f,
 762                                                 0.21198851f,  -0.38871562f,  -0.09061183f, -0.09683246f,  -0.21929175f
 763             });
 764
 765
 766     auto cellToForgetWeights =
 767             MakeTensor<float, 1>(tensorInfo20, {-0.01998659f,-0.15568835f,-0.24248174f,   -0.012770197f, 0.041331276f,
 768                                                 -0.072311886f, -0.052123554f,-0.0066330447f,-0.043891653f,0.036225766f,
 769                                                 -0.047248036f, 0.021479502f,0.033189066f, 0.11952997f,   -0.020432774f,
 770                                                 0.64658105f,   -0.06650122f,  -0.03467612f,  0.095340036f, 0.23647355f
 771             });
 772
 773     auto cellToOutputWeights =
 774             MakeTensor<float, 1>(tensorInfo20, {0.08286371f,  -0.08261836f, -0.51210177f, 0.002913762f, 0.17764764f,
 775                                                 -0.5495371f,  -0.08460716f, -0.24552552f, 0.030037103f, 0.04123544f,
 776                                                 -0.11940523f, 0.007358328f, 0.1890978f,   0.4833202f,   -0.34441817f,
 777                                                 0.36312827f,  -0.26375428f, 0.1457655f,   -0.19724406f, 0.15548733f
 778             });
 779
 780     auto projectionWeights =
 781             MakeTensor<float, 2>(tensorInfo16x20,
 782                                  {-0.009802181f,  0.09401916f,    0.0717386f,     -0.13895074f,  0.09641832f,
 783                                   0.060420845f,   0.08539281f,    0.054285463f,   0.061395317f,  0.034448683f,
 784                                   -0.042991187f,  0.019801661f,   -0.16840284f,   -0.015726732f, -0.23041931f,
 785                                   -0.024478018f,  -0.10959692f,   -0.013875541f,  0.18600968f,   -0.061274476f,
 786                                   0.0138165f,     -0.08160894f,   -0.07661644f,   0.032372914f,  0.16169067f,
 787                                   0.22465782f,    -0.03993472f,   -0.004017731f,  0.08633481f,   -0.28869787f,
 788                                   0.08682067f,    0.17240396f,    0.014975425f,   0.056431185f,  0.031037588f,
 789                                   0.16702051f,    0.0077946745f,  0.15140012f,    0.29405436f,   0.120285f,
 790                                   -0.188994f,     -0.027265169f,  0.043389652f,   -0.022061434f, 0.014777949f,
 791                                   -0.20203483f,   0.094781205f,   0.19100232f,    0.13987629f,   -0.036132768f,
 792                                   -0.06426278f,   -0.05108664f,   0.13221376f,    0.009441198f,  -0.16715929f,
 793                                   0.15859416f,    -0.040437475f,  0.050779544f,   -0.022187516f, 0.012166504f,
 794                                   0.027685808f,   -0.07675938f,   -0.0055694645f, -0.09444123f,  0.0046453946f,
 795                                   0.050794356f,   0.10770313f,    -0.20790008f,   -0.07149004f,  -0.11425117f,
 796                                   0.008225835f,   -0.035802525f,  0.14374903f,    0.15262283f,   0.048710253f,
 797                                   0.1847461f,     -0.007487823f,  0.11000021f,    -0.09542012f,  0.22619456f,
 798                                   -0.029149994f,  0.08527916f,    0.009043713f,   0.0042746216f, 0.016261552f,
 799                                   0.022461696f,   0.12689082f,    -0.043589946f,  -0.12035478f,  -0.08361797f,
 800                                   -0.050666027f,  -0.1248618f,    -0.1275799f,    -0.071875185f, 0.07377272f,
 801                                   0.09944291f,    -0.18897448f,   -0.1593054f,    -0.06526116f,  -0.040107165f,
 802                                   -0.004618631f,  -0.067624845f,  -0.007576253f,  0.10727444f,   0.041546922f,
 803                                   -0.20424393f,   0.06907816f,    0.050412357f,   0.00724631f,   0.039827548f,
 804                                   0.12449835f,    0.10747581f,    0.13708383f,    0.09134148f,   -0.12617786f,
 805                                   -0.06428341f,   0.09956831f,    0.1208086f,     -0.14676677f,  -0.0727722f,
 806                                   0.1126304f,     0.010139365f,   0.015571211f,   -0.038128063f, 0.022913318f,
 807                                   -0.042050496f,  0.16842307f,    -0.060597885f,  0.10531834f,   -0.06411776f,
 808                                   -0.07451711f,   -0.03410368f,   -0.13393489f,   0.06534304f,   0.003620307f,
 809                                   0.04490757f,    0.05970546f,    0.05197996f,    0.02839995f,   0.10434969f,
 810                                   -0.013699693f,  -0.028353551f,  -0.07260381f,   0.047201227f,  -0.024575593f,
 811                                   -0.036445823f,  0.07155557f,    0.009672501f,   -0.02328883f,  0.009533515f,
 812                                   -0.03606021f,   -0.07421458f,   -0.028082801f,  -0.2678904f,   -0.13221288f,
 813                                   0.18419984f,    -0.13012612f,   -0.014588381f,  -0.035059117f, -0.04824723f,
 814                                   0.07830115f,    -0.056184657f,  0.03277091f,    0.025466874f,  0.14494097f,
 815                                   -0.12522776f,   -0.098633975f,  -0.10766018f,   -0.08317623f,  0.08594209f,
 816                                   0.07749552f,    0.039474737f,   0.1776665f,     -0.07409566f,  -0.0477268f,
 817                                   0.29323658f,    0.10801441f,    0.1154011f,     0.013952499f,  0.10739139f,
 818                                   0.10708251f,    -0.051456142f,  0.0074137426f,  -0.10430189f,  0.10034707f,
 819                                   0.045594677f,   0.0635285f,     -0.0715442f,    -0.089667566f, -0.10811871f,
 820                                   0.00026344223f, 0.08298446f,    -0.009525053f,  0.006585689f,  -0.24567553f,
 821                                   -0.09450807f,   0.09648481f,    0.026996298f,   -0.06419476f,  -0.04752702f,
 822                                   -0.11063944f,   -0.23441927f,   -0.17608605f,   -0.052156363f, 0.067035615f,
 823                                   0.19271925f,    -0.0032889997f, -0.043264326f,  0.09663576f,   -0.057112187f,
 824                                   -0.10100678f,   0.0628376f,     0.04447668f,    0.017961001f,  -0.10094388f,
 825                                   -0.10190601f,   0.18335468f,    0.10494553f,    -0.052095775f, -0.0026118709f,
 826                                   0.10539724f,    -0.04383912f,   -0.042349473f,  0.08438151f,   -0.1947263f,
 827                                   0.02251204f,    0.11216432f,    -0.10307853f,   0.17351969f,   -0.039091777f,
 828                                   0.08066188f,    -0.00561982f,   0.12633002f,    0.11335965f,   -0.0088127935f,
 829                                   -0.019777594f,  0.06864014f,    -0.059751723f,  0.016233567f,  -0.06894641f,
 830                                   -0.28651384f,   -0.004228674f,  0.019708522f,   -0.16305895f,  -0.07468996f,
 831                                   -0.0855457f,    0.099339016f,   -0.07580735f,   -0.13775392f,  0.08434318f,
 832                                   0.08330512f,    -0.12131499f,   0.031935584f,   0.09180414f,   -0.08876437f,
 833                                   -0.08049874f,   0.008753825f,   0.03498998f,    0.030215185f,  0.03907079f,
 834                                   0.089751154f,   0.029194152f,   -0.03337423f,   -0.019092513f, 0.04331237f,
 835                                   0.04299654f,    -0.036394123f,  -0.12915532f,   0.09793732f,   0.07512415f,
 836                                   -0.11319543f,   -0.032502122f,  0.15661901f,    0.07671967f,   -0.005491124f,
 837                                   -0.19379048f,   -0.218606f,     0.21448623f,    0.017840758f,  0.1416943f,
 838                                   -0.07051762f,   0.19488361f,    0.02664691f,    -0.18104725f,  -0.09334311f,
 839                                   0.15026465f,    -0.15493552f,   -0.057762887f,  -0.11604192f,  -0.262013f,
 840                                   -0.01391798f,   0.012185008f,   0.11156489f,    -0.07483202f,  0.06693364f,
 841                                   -0.26151478f,   0.046425626f,   0.036540434f,   -0.16435726f,  0.17338543f,
 842                                   -0.21401681f,   -0.11385144f,   -0.08283257f,   -0.069031075f, 0.030635102f,
 843                                   0.010969227f,   0.11109743f,    0.010919218f,   0.027526086f,  0.13519906f,
 844                                   0.01891392f,    -0.046839405f,  -0.040167913f,  0.017953383f,  -0.09700955f,
 845                                   0.0061885654f,  -0.07000971f,   0.026893595f,   -0.038844477f, 0.14543656f
 846                                  });
 847
 848     std::vector<float> projectionBiasVector(outputSize, 0.f);
 849     auto projectionBias = MakeTensor<float,1>(tensorInfo16, projectionBiasVector);
 850
 851     armnn::ScopedCpuTensorHandle inputToInputWeightsTensor(tensorInfo20x5);
 852     armnn::ScopedCpuTensorHandle inputToForgetWeightsTensor(tensorInfo20x5);
 853     armnn::ScopedCpuTensorHandle inputToCellWeightsTensor(tensorInfo20x5);
 854     armnn::ScopedCpuTensorHandle inputToOutputWeightsTensor(tensorInfo20x5);
 855     armnn::ScopedCpuTensorHandle recurrentToForgetWeightsTensor(tensorInfo20x16);
 856     armnn::ScopedCpuTensorHandle recurrentToInputWeightsTensor(tensorInfo20x16);
 857     armnn::ScopedCpuTensorHandle recurrentToCellWeightsTensor(tensorInfo20x16);
 858     armnn::ScopedCpuTensorHandle recurrentToOutputWeightsTensor(tensorInfo20x16);
 859     armnn::ScopedCpuTensorHandle cellToInputWeightsTensor(tensorInfo20);
 860     armnn::ScopedCpuTensorHandle inputGateBiasTensor(tensorInfo20);
 861     armnn::ScopedCpuTensorHandle forgetGateBiasTensor(tensorInfo20);
 862     armnn::ScopedCpuTensorHandle cellBiasTensor(tensorInfo20);
 863     armnn::ScopedCpuTensorHandle outputGateBiasTensor(tensorInfo20);
 864     armnn::ScopedCpuTensorHandle cellToForgetWeightsTensor(tensorInfo20);
 865     armnn::ScopedCpuTensorHandle cellToOutputWeightsTensor(tensorInfo20);
 866     armnn::ScopedCpuTensorHandle projectionWeightsTensor(tensorInfo16x20);
 867     armnn::ScopedCpuTensorHandle projectionBiasTensor(tensorInfo16);
 868
 869     AllocateAndCopyDataToITensorHandle(&inputToInputWeightsTensor, &inputToInputWeights[0][0]);
 870     AllocateAndCopyDataToITensorHandle(&inputToForgetWeightsTensor, &inputToForgetWeights[0][0]);
 871     AllocateAndCopyDataToITensorHandle(&inputToCellWeightsTensor, &inputToCellWeights[0][0]);
 872     AllocateAndCopyDataToITensorHandle(&inputToOutputWeightsTensor, &inputToOutputWeights[0][0]);
 873     AllocateAndCopyDataToITensorHandle(&recurrentToInputWeightsTensor, &recurrentToInputWeights[0][0]);
 874     AllocateAndCopyDataToITensorHandle(&recurrentToForgetWeightsTensor, &recurrentToForgetWeights[0][0]);
 875     AllocateAndCopyDataToITensorHandle(&recurrentToCellWeightsTensor, &recurrentToCellWeights[0][0]);
 876     AllocateAndCopyDataToITensorHandle(&recurrentToOutputWeightsTensor, &recurrentToOutputWeights[0][0]);
 877     AllocateAndCopyDataToITensorHandle(&cellToInputWeightsTensor, &cellToInputWeights[0]);
 878     AllocateAndCopyDataToITensorHandle(&inputGateBiasTensor, &inputGateBias[0]);
 879     AllocateAndCopyDataToITensorHandle(&forgetGateBiasTensor, &forgetGateBias[0]);
 880     AllocateAndCopyDataToITensorHandle(&cellBiasTensor, &cellBias[0]);
 881     AllocateAndCopyDataToITensorHandle(&outputGateBiasTensor, &outputGateBias[0]);
 882     AllocateAndCopyDataToITensorHandle(&cellToForgetWeightsTensor, &cellToForgetWeights[0]);
 883     AllocateAndCopyDataToITensorHandle(&cellToOutputWeightsTensor, &cellToOutputWeights[0]);
 884     AllocateAndCopyDataToITensorHandle(&projectionWeightsTensor, &projectionWeights[0][0]);
 885     AllocateAndCopyDataToITensorHandle(&projectionBiasTensor, &projectionBias[0]);
 886
 887     data.m_InputToInputWeights = &inputToInputWeightsTensor;
 888     data.m_InputToForgetWeights = &inputToForgetWeightsTensor;
 889     data.m_InputToCellWeights = &inputToCellWeightsTensor;
 890     data.m_InputToOutputWeights = &inputToOutputWeightsTensor;
 891     data.m_RecurrentToInputWeights = &recurrentToInputWeightsTensor;
 892     data.m_RecurrentToForgetWeights = &recurrentToForgetWeightsTensor;
 893     data.m_RecurrentToCellWeights = &recurrentToCellWeightsTensor;
 894     data.m_RecurrentToOutputWeights = &recurrentToOutputWeightsTensor;
 895     data.m_CellToInputWeights = &cellToInputWeightsTensor;
 896     data.m_InputGateBias = &inputGateBiasTensor;
 897     data.m_ForgetGateBias = &forgetGateBiasTensor;
 898     data.m_CellBias = &cellBiasTensor;
 899     data.m_OutputGateBias = &outputGateBiasTensor;
 900     data.m_CellToForgetWeights = &cellToForgetWeightsTensor;
 901     data.m_CellToOutputWeights = &cellToOutputWeightsTensor;
 902     data.m_ProjectionWeights = &projectionWeightsTensor;
 903     data.m_ProjectionBias = &projectionBiasTensor;
 904
 905     // Flags to set test configuration
 906     data.m_Parameters.m_ActivationFunc = 4;
 907     data.m_Parameters.m_CifgEnabled = false;
 908     data.m_Parameters.m_PeepholeEnabled = true;
 909     data.m_Parameters.m_ProjectionEnabled = true;
 910
 911
 912     std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateLstm(data, info);
 913     inputHandle->Allocate();
 914     outputStateInHandle->Allocate();
 915     cellStateInHandle->Allocate();
 916
 917     scratchHandle->Allocate();
 918     outputStateOutHandle->Allocate();
 919     cellStateOutHandle->Allocate();
 920     outputHandle->Allocate();
 921
 922     CopyDataToITensorHandle(inputHandle.get(), &inputTensor[0][0]);
 923     CopyDataToITensorHandle(outputStateInHandle.get(), &outputStateInTensor[0][0]);
 924     CopyDataToITensorHandle(cellStateInHandle.get(), &cellStateInTensor[0][0]);
 925
 926     workload->Execute();
 927
 928     CopyDataFromITensorHandle(&ret.output[0][0], outputHandle.get());
 929
 930     return ret;
 931
 932 }
 933
 934
 935 LayerTestResult<float, 2> LstmLayerWithCifgWithPeepholeNoProjectionTestImpl(
 936         armnn::IWorkloadFactory& workloadFactory,
 937         const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
 938         const boost::multi_array<float, 2>& input,
 939         const boost::multi_array<float, 2>& outputExpected)
 940 {
 941     bool cifgEnabled = true;
 942     bool peepholeEnabled = true;
 943     bool projectionEnabled = false;
 944     // These are not the input and the output of Lstm yet
 945     unsigned int batchSize = boost::numeric_cast<unsigned int>(input.shape()[0]);
 946     unsigned int inputSize = boost::numeric_cast<unsigned int>(input.shape()[1]);
 947
 948     unsigned int outputSize = boost::numeric_cast<unsigned int>(outputExpected.shape()[1]);
 949
 950     const unsigned int cellSize = outputSize;
 951
 952     // Decide the shape of all input tensors
 953     armnn::TensorInfo inputTensorInfo({batchSize , inputSize}, armnn::DataType::Float32);
 954     armnn::TensorInfo outputStateInTensorInfo({batchSize, outputSize}, armnn::DataType::Float32);
 955     armnn::TensorInfo cellStateInTensorInfo({batchSize, cellSize}, armnn::DataType::Float32);
 956
 957     unsigned int scratchBufferSize = cifgEnabled ? cellSize * 3 : cellSize * 4;
 958     armnn::TensorInfo scratchBufferTensorInfo({batchSize, scratchBufferSize}, armnn::DataType::Float32);
 959     armnn::TensorInfo outputStateOutTensorInfo({batchSize, outputSize}, armnn::DataType::Float32);
 960     armnn::TensorInfo cellStateOutTensorInfo({batchSize, cellSize}, armnn::DataType::Float32);
 961     armnn::TensorInfo outputTensorInfo({batchSize, outputSize}, armnn::DataType::Float32);
 962
 963     // List of inputs
 964     std::vector<float> inputData;
 965     inputData.assign(input.data(), input.data() + batchSize*inputSize);
 966     auto inputTensor = MakeTensor<float,2>(inputTensorInfo, inputData);
 967
 968     std::vector<float> outputStateInVector(batchSize * outputSize, 0.f);
 969     auto outputStateInTensor = MakeTensor<float, 2>(outputStateInTensorInfo, outputStateInVector);
 970
 971     std::vector<float> cellStateInVector(batchSize * cellSize, 0.f);
 972     auto cellStateInTensor = MakeTensor<float, 2>(cellStateInTensorInfo, cellStateInVector);
 973
 974
 975     // Prepare all the weights in the descriptor for LSTM
 976     armnn::LstmQueueDescriptor data;
 977     armnn::TensorInfo tensorInfoInput({cellSize, inputSize}, armnn::DataType::Float32);
 978     armnn::TensorInfo tensorInfoOutput({cellSize, outputSize}, armnn::DataType::Float32);
 979     armnn::TensorInfo tensorInfoNumUnits({cellSize}, armnn::DataType::Float32);
 980
 981     auto inputToCellWeights = MakeTensor<float, 2>(tensorInfoInput,
 982                                                      {-0.49770179f, -0.27711356f, -0.09624726f, 0.05100781f,
 983                                                      0.04717243f, 0.48944736f, -0.38535351f,
 984                                                      -0.17212132f});
 985     auto inputToForgetWeights = MakeTensor<float, 2>(tensorInfoInput,
 986                                                      {-0.55291498f, -0.42866567f, 0.13056988f,
 987                                                        -0.3633365f, -0.22755712f, 0.28253698f, 0.24407166f,
 988                                                        0.33826375f});
 989     auto inputToOutputWeights = MakeTensor<float, 2>(tensorInfoInput,
 990                                                      {0.10725588f, -0.02335852f, -0.55932593f,
 991                                                        -0.09426838f, -0.44257352f, 0.54939759f,
 992                                                        0.01533556f, 0.42751634f});
 993     auto cellBias = MakeTensor<float, 1>(tensorInfoNumUnits, {0.f, 0.f, 0.f, 0.f});
 994     auto forgetGateBias = MakeTensor<float, 1>(tensorInfoNumUnits, {1.f, 1.f, 1.f, 1.f});
 995     auto outputGateBias = MakeTensor<float, 1>(tensorInfoNumUnits, {0.f, 0.f, 0.f, 0.f});
 996
 997     auto recurrentToCellWeights = MakeTensor<float, 2>(tensorInfoOutput,
 998                 {0.54066205f, -0.32668582f, -0.43562764f, -0.56094903f, 0.42957711f,
 999                  0.01841056f, -0.32764608f, -0.33027974f, -0.10826075f, 0.20675004f,
1000                  0.19069612f, -0.03026325f, -0.54532051f, 0.33003211f, 0.44901288f,
1001                  0.21193194f});
1002     auto recurrentToForgetWeights = MakeTensor<float, 2>(tensorInfoOutput,
1003                  {-0.13832897f, -0.0515101f, -0.2359007f, -0.16661474f, -0.14340827f,
1004                   0.36986142f, 0.23414481f, 0.55899f, 0.10798943f, -0.41174671f, 0.17751795f,
1005                   -0.34484994f, -0.35874045f, -0.11352962f, 0.27268326f, 0.54058349f});
1006
1007     auto recurrentToOutputWeights = MakeTensor<float, 2>(tensorInfoOutput,
1008                 {0.41613156f, 0.42610586f, -0.16495961f, -0.5663873f, 0.30579174f, -0.05115908f,
1009                  -0.33941799f, 0.23364776f, 0.11178309f, 0.09481031f, -0.26424935f, 0.46261835f,
1010                  0.50248802f, 0.26114327f, -0.43736315f, 0.33149987f});
1011
1012     auto cellToForgetWeights = MakeTensor<float, 1>(tensorInfoNumUnits,
1013                 {0.47485286f, -0.51955009f, -0.24458408f, 0.31544167f});
1014     auto cellToOutputWeights = MakeTensor<float, 1>(tensorInfoNumUnits,
1015                 {-0.17135078f, 0.82760304f, 0.85573703f, -0.77109635f});
1016
1017     armnn::ScopedCpuTensorHandle inputToCellWeightsTensor(tensorInfoInput);
1018     armnn::ScopedCpuTensorHandle inputToForgetWeightsTensor(tensorInfoInput);
1019     armnn::ScopedCpuTensorHandle inputToOutputWeightsTensor(tensorInfoInput);
1020
1021     armnn::ScopedCpuTensorHandle cellBiasTensor(tensorInfoNumUnits);
1022     armnn::ScopedCpuTensorHandle forgetGateBiasTensor(tensorInfoNumUnits);
1023     armnn::ScopedCpuTensorHandle outputGateBiasTensor(tensorInfoNumUnits);
1024
1025     armnn::ScopedCpuTensorHandle recurrentToCellWeightsTensor(tensorInfoOutput);
1026     armnn::ScopedCpuTensorHandle recurrentToForgetWeightsTensor(tensorInfoOutput);
1027     armnn::ScopedCpuTensorHandle recurrentToOutputWeightsTensor(tensorInfoOutput);
1028
1029
1030     armnn::ScopedCpuTensorHandle cellToForgetWeightsTensor(tensorInfoNumUnits);
1031     armnn::ScopedCpuTensorHandle cellToOutputWeightsTensor(tensorInfoNumUnits);
1032
1033     AllocateAndCopyDataToITensorHandle(&inputToCellWeightsTensor, &inputToCellWeights[0][0]);
1034     AllocateAndCopyDataToITensorHandle(&inputToForgetWeightsTensor, &inputToForgetWeights[0][0]);
1035     AllocateAndCopyDataToITensorHandle(&inputToOutputWeightsTensor, &inputToOutputWeights[0][0]);
1036
1037     AllocateAndCopyDataToITensorHandle(&cellBiasTensor, &cellBias[0]);
1038     AllocateAndCopyDataToITensorHandle(&forgetGateBiasTensor, &forgetGateBias[0]);
1039     AllocateAndCopyDataToITensorHandle(&outputGateBiasTensor, &outputGateBias[0]);
1040
1041     AllocateAndCopyDataToITensorHandle(&recurrentToCellWeightsTensor, &recurrentToCellWeights[0][0]);
1042     AllocateAndCopyDataToITensorHandle(&recurrentToForgetWeightsTensor, &recurrentToForgetWeights[0][0]);
1043     AllocateAndCopyDataToITensorHandle(&recurrentToOutputWeightsTensor, &recurrentToOutputWeights[0][0]);
1044
1045     AllocateAndCopyDataToITensorHandle(&cellToForgetWeightsTensor, &cellToForgetWeights[0]);
1046     AllocateAndCopyDataToITensorHandle(&cellToOutputWeightsTensor, &cellToOutputWeights[0]);
1047
1048
1049     data.m_InputToCellWeights = &inputToCellWeightsTensor;
1050     data.m_InputToForgetWeights = &inputToForgetWeightsTensor;
1051     data.m_InputToOutputWeights = &inputToOutputWeightsTensor;
1052
1053     data.m_CellBias = &cellBiasTensor;
1054     data.m_ForgetGateBias = &forgetGateBiasTensor;
1055     data.m_OutputGateBias = &outputGateBiasTensor;
1056
1057     data.m_RecurrentToCellWeights = &recurrentToCellWeightsTensor;
1058     data.m_RecurrentToForgetWeights = &recurrentToForgetWeightsTensor;
1059     data.m_RecurrentToOutputWeights = &recurrentToOutputWeightsTensor;
1060
1061     data.m_CellToForgetWeights = &cellToForgetWeightsTensor;
1062     data.m_CellToOutputWeights = &cellToOutputWeightsTensor;
1063
1064     // other parameters for the descriptor
1065     data.m_Parameters.m_CifgEnabled = cifgEnabled;
1066     data.m_Parameters.m_ProjectionEnabled = projectionEnabled;
1067     data.m_Parameters.m_PeepholeEnabled = peepholeEnabled;
1068
1069     data.m_Parameters.m_ActivationFunc = 4;
1070     data.m_Parameters.m_ClippingThresProj = 0.0;
1071     data.m_Parameters.m_ClippingThresCell = 0.0;
1072
1073
1074     // List of outputs
1075     std::vector<float> scratchBufferVector(batchSize * scratchBufferSize, 0.f);
1076     auto scratchBufferTensor = MakeTensor<float,2>(scratchBufferTensorInfo, scratchBufferVector);
1077     LayerTestResult<float, 2> ret0(scratchBufferTensorInfo);
1078
1079     // Output state for a certain time step
1080     std::vector<float> outputStateOutVector(batchSize * outputSize, 0.f);
1081     auto outputStateOutTensor = MakeTensor<float,2>(outputStateOutTensorInfo, outputStateOutVector);
1082     LayerTestResult<float, 2> ret1(outputStateOutTensorInfo);
1083
1084     // Cell state for a certain time step
1085     std::vector<float> cellStateOutVector(batchSize * cellSize, 0.f);
1086     auto cellStateOutTensor = MakeTensor<float,2>(cellStateOutTensorInfo, cellStateOutVector);
1087     LayerTestResult<float, 2> ret2(cellStateOutTensorInfo);
1088
1089     // Output for a certain time step
1090     std::vector<float> outputVector(batchSize * outputSize, 0.f);
1091     auto outputTensor = MakeTensor<float, 2>(outputTensorInfo, outputVector);
1092     std::vector<float> outputData;
1093     outputData.assign(outputExpected.data(), outputExpected.data() + batchSize*outputSize);
1094     LayerTestResult<float, 2> ret3(outputTensorInfo);
1095     ret3.outputExpected = MakeTensor<float, 2>(outputTensorInfo, outputData);
1096
1097     // Prepare the inputs and outputs for the workload
1098     std::unique_ptr<armnn::ITensorHandle> inputHandle =
1099             workloadFactory.CreateTensorHandle(inputTensorInfo);
1100     std::unique_ptr<armnn::ITensorHandle> outputStateInHandle =
1101             workloadFactory.CreateTensorHandle(outputStateInTensorInfo);
1102     std::unique_ptr<armnn::ITensorHandle> cellStateInHandle =
1103             workloadFactory.CreateTensorHandle(cellStateInTensorInfo);
1104
1105     std::unique_ptr<armnn::ITensorHandle> scratchBufferHandle =
1106             workloadFactory.CreateTensorHandle(scratchBufferTensorInfo);
1107     std::unique_ptr<armnn::ITensorHandle> outputStateOutHandle =
1108             workloadFactory.CreateTensorHandle(outputStateOutTensorInfo);
1109     std::unique_ptr<armnn::ITensorHandle> cellStateOutHandle =
1110             workloadFactory.CreateTensorHandle(cellStateOutTensorInfo);
1111     std::unique_ptr<armnn::ITensorHandle> outputHandle =
1112             workloadFactory.CreateTensorHandle(outputTensorInfo);
1113
1114     armnn::WorkloadInfo info;
1115     AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
1116     AddInputToWorkload(data, info, outputStateInTensorInfo, outputStateInHandle.get());
1117     AddInputToWorkload(data, info, cellStateInTensorInfo, cellStateInHandle.get());
1118
1119     AddOutputToWorkload(data, info, scratchBufferTensorInfo, scratchBufferHandle.get());
1120     AddOutputToWorkload(data, info, outputStateOutTensorInfo, outputStateOutHandle.get());
1121     AddOutputToWorkload(data, info, cellStateOutTensorInfo, cellStateOutHandle.get());
1122     AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
1123
1124     std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateLstm(data, info);
1125
1126
1127     inputHandle->Allocate();
1128     outputStateInHandle->Allocate();
1129     cellStateInHandle->Allocate();
1130
1131     scratchBufferHandle->Allocate();
1132     outputStateOutHandle->Allocate();
1133     cellStateOutHandle->Allocate();
1134     outputHandle->Allocate();
1135
1136
1137     CopyDataToITensorHandle(inputHandle.get(), &inputTensor[0][0]);
1138     CopyDataToITensorHandle(outputStateInHandle.get(), &outputStateInTensor[0][0]);
1139     CopyDataToITensorHandle(cellStateInHandle.get(), &cellStateInTensor[0][0]);
1140
1141     CopyDataToITensorHandle(scratchBufferHandle.get(), &scratchBufferTensor[0][0]);
1142     CopyDataToITensorHandle(outputStateOutHandle.get(), &outputStateOutTensor[0][0]);
1143     CopyDataToITensorHandle(cellStateOutHandle.get(), &cellStateOutTensor[0][0]);
1144
1145     workload->Execute();
1146
1147     CopyDataFromITensorHandle(&ret0.output[0][0], scratchBufferHandle.get());
1148     CopyDataFromITensorHandle(&ret1.output[0][0], outputStateOutHandle.get());
1149     CopyDataFromITensorHandle(&ret2.output[0][0], cellStateOutHandle.get());
1150     CopyDataFromITensorHandle(&ret3.output[0][0], outputHandle.get());
1151
1152     return ret3;
1153 }