src/armnn/test/CreateWorkload.hpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5 #pragma once
   6
   7 #include <boost/test/unit_test.hpp>
   8
   9 #include <boost/cast.hpp>
  10
  11 #include <backendsCommon/WorkloadData.hpp>
  12 #include <backendsCommon/WorkloadFactory.hpp>
  13 #include <backendsCommon/CpuTensorHandle.hpp>
  14
  15 #include <Graph.hpp>
  16
  17 #include <utility>
  18
  19
  20 using namespace armnn;
  21
  22 namespace
  23 {
  24
  25 using namespace std;
  26
  27 // Calls CreateWorkload for a layer, and checks the returned pointer is of the correct type.
  28 template<typename Workload>
  29 std::unique_ptr<Workload> MakeAndCheckWorkload(Layer& layer, Graph& graph, const IWorkloadFactory& factory)
  30 {
  31     std::unique_ptr<IWorkload> workload = layer.CreateWorkload(graph, factory);
  32     BOOST_TEST(workload.get() == boost::polymorphic_downcast<Workload*>(workload.get()),
  33                "Cannot convert to derived class");
  34     std::string reasonIfUnsupported;
  35     layer.SetBackendId(factory.GetBackendId());
  36     BOOST_TEST(factory.IsLayerSupported(layer, layer.GetDataType(), reasonIfUnsupported));
  37     return std::unique_ptr<Workload>(static_cast<Workload*>(workload.release()));
  38 }
  39
  40 // Connects two layers.
  41 void Connect(Layer* from, Layer* to, const TensorInfo& tensorInfo, unsigned int fromIndex = 0, unsigned int toIndex = 0)
  42 {
  43     from->GetOutputSlot(fromIndex).Connect(to->GetInputSlot(toIndex));
  44     from->GetOutputHandler(fromIndex).SetTensorInfo(tensorInfo);
  45 }
  46
  47 // Helper function to create tensor handlers for workloads, assuming they all use the same factory.
  48 void CreateTensorHandles(armnn::Graph& graph, armnn::IWorkloadFactory& factory)
  49 {
  50     for (auto&& layer : graph.TopologicalSort())
  51     {
  52         layer->CreateTensorHandles(graph, factory);
  53     }
  54 }
  55
  56 /////////////////////////////////////////////////////////////////////////////////////////////
  57 // The following functions are called by backendsCommon/test/CreateWorkload*.cpp
  58 // They build very simple graphs, and then create a workload.
  59 // Some checks are performed on the workload to ensure parameters have been passed correctly.
  60 // They return the created workloads so that backend-specific checks can be performed.
  61 /////////////////////////////////////////////////////////////////////////////////////////////
  62
  63 template <typename ActivationWorkload, armnn::DataType DataType>
  64 std::unique_ptr<ActivationWorkload> CreateActivationWorkloadTest(armnn::IWorkloadFactory& factory,
  65                                                                  armnn::Graph&            graph)
  66 {
  67     // Creates the layer we're testing.
  68     ActivationDescriptor layerDesc;
  69     layerDesc.m_Function = ActivationFunction::Abs;
  70     layerDesc.m_A        = 3.5f;
  71     layerDesc.m_B        = -10.0f;
  72
  73     ActivationLayer* const layer = graph.AddLayer<ActivationLayer>(layerDesc, "layer");
  74
  75     // Creates extra layers.
  76     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
  77     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
  78
  79     // Connects up.
  80     armnn::TensorInfo tensorInfo({1, 1}, DataType);
  81
  82     Connect(input, layer, tensorInfo);
  83     Connect(layer, output, tensorInfo);
  84
  85     CreateTensorHandles(graph, factory);
  86
  87     // Makes the workload and checks it.
  88     auto workload = MakeAndCheckWorkload<ActivationWorkload>(*layer, graph, factory);
  89
  90     ActivationQueueDescriptor queueDescriptor = workload->GetData();
  91     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
  92     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
  93     BOOST_TEST(queueDescriptor.m_Parameters.m_A == 3.5f);
  94     BOOST_TEST(queueDescriptor.m_Parameters.m_B == -10.0f);
  95     BOOST_TEST((queueDescriptor.m_Parameters.m_Function == ActivationFunction::Abs));
  96
  97     // Returns so we can do extra, backend-specific tests.
  98     return workload;
  99 }
 100
 101 template <typename WorkloadType,
 102           typename DescriptorType,
 103           typename LayerType,
 104           armnn::DataType DataType>
 105 std::unique_ptr<WorkloadType> CreateArithmeticWorkloadTest(armnn::IWorkloadFactory& factory,
 106                                                            armnn::Graph&            graph)
 107 {
 108     // Creates the layer we're testing.
 109     Layer* const layer = graph.AddLayer<LayerType>("layer");
 110
 111     // Creates extra layers.
 112     Layer* const input1 = graph.AddLayer<InputLayer>(1, "input1");
 113     Layer* const input2 = graph.AddLayer<InputLayer>(2, "input2");
 114     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
 115
 116     // Connects up.
 117     armnn::TensorInfo tensorInfo({2, 3}, DataType);
 118     Connect(input1, layer, tensorInfo, 0, 0);
 119     Connect(input2, layer, tensorInfo, 0, 1);
 120     Connect(layer, output, tensorInfo);
 121     CreateTensorHandles(graph, factory);
 122
 123     // Makes the workload and checks it.
 124     auto workload = MakeAndCheckWorkload<WorkloadType>(*layer, graph, factory);
 125
 126     DescriptorType queueDescriptor = workload->GetData();
 127     BOOST_TEST(queueDescriptor.m_Inputs.size() == 2);
 128     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
 129
 130     // Returns so we can do extra, backend-specific tests.
 131     return workload;
 132 }
 133
 134 template <typename BatchNormalizationFloat32Workload, armnn::DataType DataType>
 135 std::unique_ptr<BatchNormalizationFloat32Workload> CreateBatchNormalizationWorkloadTest(
 136     armnn::IWorkloadFactory& factory, armnn::Graph& graph, DataLayoutIndexed dataLayout = DataLayout::NCHW)
 137 {
 138
 139     TensorShape tensorShape;
 140     switch (dataLayout.GetDataLayout())
 141     {
 142         case DataLayout::NHWC:
 143             tensorShape = { 2, 4, 4, 3 };
 144             break;
 145         case DataLayout::NCHW:
 146         default:
 147             tensorShape = { 2, 3, 4, 4 };
 148     }
 149
 150     // Creates the layer we're testing.
 151     BatchNormalizationDescriptor layerDesc;
 152     layerDesc.m_Eps = 0.05f;
 153     layerDesc.m_DataLayout = dataLayout;
 154
 155     BatchNormalizationLayer* const layer = graph.AddLayer<BatchNormalizationLayer>(layerDesc, "layer");
 156
 157     armnn::TensorInfo weightInfo({3}, DataType);
 158     layer->m_Mean     = std::make_unique<ScopedCpuTensorHandle>(weightInfo);
 159     layer->m_Variance = std::make_unique<ScopedCpuTensorHandle>(weightInfo);
 160     layer->m_Beta     = std::make_unique<ScopedCpuTensorHandle>(weightInfo);
 161     layer->m_Gamma    = std::make_unique<ScopedCpuTensorHandle>(weightInfo);
 162     layer->m_Mean->Allocate();
 163     layer->m_Variance->Allocate();
 164     layer->m_Beta->Allocate();
 165     layer->m_Gamma->Allocate();
 166
 167     // Creates extra layers.
 168     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 169     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
 170
 171     //Connects up.
 172     armnn::TensorInfo tensorInfo(tensorShape, DataType);
 173     Connect(input, layer, tensorInfo);
 174     Connect(layer, output, tensorInfo);
 175     CreateTensorHandles(graph, factory);
 176
 177     // Makes the workload and checks it.
 178     auto workload = MakeAndCheckWorkload<BatchNormalizationFloat32Workload>(*layer, graph, factory);
 179     BatchNormalizationQueueDescriptor queueDescriptor = workload->GetData();
 180     BOOST_TEST(queueDescriptor.m_Parameters.m_Eps == 0.05f);
 181     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
 182     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
 183     BOOST_TEST((queueDescriptor.m_Mean->GetTensorInfo() == TensorInfo({3}, DataType)));
 184     BOOST_TEST((queueDescriptor.m_Variance->GetTensorInfo() == TensorInfo({3}, DataType)));
 185     BOOST_TEST((queueDescriptor.m_Gamma->GetTensorInfo() == TensorInfo({3}, DataType)));
 186     BOOST_TEST((queueDescriptor.m_Beta->GetTensorInfo() == TensorInfo({3}, DataType)));
 187     BOOST_TEST((queueDescriptor.m_Parameters.m_DataLayout.GetDataLayout() == dataLayout));
 188
 189     // Returns so we can do extra, backend-specific tests.
 190     return workload;
 191 }
 192
 193 template <typename Convolution2dWorkload, armnn::DataType DataType>
 194 std::unique_ptr<Convolution2dWorkload> CreateConvolution2dWorkloadTest(armnn::IWorkloadFactory& factory,
 195                                                                        armnn::Graph&            graph,
 196                                                                        DataLayout dataLayout = DataLayout::NCHW)
 197 {
 198     // Creates the layer we're testing.
 199     Convolution2dDescriptor layerDesc;
 200     layerDesc.m_PadLeft = 3;
 201     layerDesc.m_PadRight = 3;
 202     layerDesc.m_PadTop = 1;
 203     layerDesc.m_PadBottom = 1;
 204     layerDesc.m_StrideX = 2;
 205     layerDesc.m_StrideY = 4;
 206     layerDesc.m_BiasEnabled = true;
 207     layerDesc.m_DataLayout = dataLayout;
 208
 209     Convolution2dLayer* const layer = graph.AddLayer<Convolution2dLayer>(layerDesc, "layer");
 210
 211     TensorShape weightShape = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 3, 5, 3} : TensorShape{2, 5, 3, 3};
 212     TensorShape inputShape  = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 3, 8, 16} : TensorShape{2, 8, 16, 3};
 213     TensorShape outputShape = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 2, 2, 10} : TensorShape{2, 2, 10, 2};
 214
 215     layer->m_Weight = std::make_unique<ScopedCpuTensorHandle>(TensorInfo(weightShape, DataType));
 216     layer->m_Bias   = std::make_unique<ScopedCpuTensorHandle>(TensorInfo({2}, GetBiasDataType(DataType)));
 217
 218     layer->m_Weight->Allocate();
 219     layer->m_Bias->Allocate();
 220
 221     // Creates extra layers.
 222     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 223     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
 224
 225     // Connects up.
 226     Connect(input, layer, TensorInfo(inputShape, DataType));
 227     Connect(layer, output, TensorInfo(outputShape, DataType));
 228     CreateTensorHandles(graph, factory);
 229
 230     // Makes the workload and checks it.
 231     auto workload = MakeAndCheckWorkload<Convolution2dWorkload>(*layer, graph, factory);
 232
 233     Convolution2dQueueDescriptor queueDescriptor = workload->GetData();
 234     BOOST_TEST(queueDescriptor.m_Parameters.m_StrideX == 2);
 235     BOOST_TEST(queueDescriptor.m_Parameters.m_StrideY == 4);
 236     BOOST_TEST(queueDescriptor.m_Parameters.m_PadLeft == 3);
 237     BOOST_TEST(queueDescriptor.m_Parameters.m_PadRight == 3);
 238     BOOST_TEST(queueDescriptor.m_Parameters.m_PadTop == 1);
 239     BOOST_TEST(queueDescriptor.m_Parameters.m_PadBottom == 1);
 240     BOOST_TEST(queueDescriptor.m_Parameters.m_BiasEnabled);
 241     BOOST_TEST((queueDescriptor.m_Parameters.m_DataLayout == dataLayout));
 242
 243     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
 244     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
 245     BOOST_TEST((queueDescriptor.m_Weight->GetTensorInfo() == TensorInfo(weightShape, DataType)));
 246     BOOST_TEST((queueDescriptor.m_Bias->GetTensorInfo() ==
 247         TensorInfo({2}, GetBiasDataType(DataType))));
 248
 249     // Returns so we can do extra, backend-specific tests.
 250     return workload;
 251 }
 252
 253 template <typename LstmWorkload>
 254 std::unique_ptr<LstmWorkload> CreateLstmWorkloadTest(armnn::IWorkloadFactory& factory, armnn::Graph& graph)
 255 {
 256     // This parameter setting is for withCifgWithPeepholeNoProjection
 257     LstmDescriptor layerDesc;
 258     layerDesc.m_ActivationFunc = 4;
 259     layerDesc.m_ClippingThresCell = 0.0f;
 260     layerDesc.m_ClippingThresProj = 0.0f;
 261     layerDesc.m_CifgEnabled = true;
 262     layerDesc.m_PeepholeEnabled = true;
 263     layerDesc.m_ProjectionEnabled = false;
 264
 265     LstmLayer* const layer = graph.AddLayer<LstmLayer>(layerDesc, "layer");
 266     unsigned int batchSize = 2;
 267     unsigned int inputSize = 2;
 268     unsigned int numUnits = 4;
 269     unsigned int outputSize = 4;
 270
 271     layer->m_BasicParameters.m_InputToForgetWeights = std::make_unique<ScopedCpuTensorHandle>
 272             (TensorInfo({ numUnits, inputSize }, DataType::Float32));
 273     layer->m_BasicParameters.m_InputToCellWeights = std::make_unique<ScopedCpuTensorHandle>
 274             (TensorInfo({ numUnits, inputSize }, DataType::Float32));
 275     layer->m_BasicParameters.m_InputToOutputWeights = std::make_unique<ScopedCpuTensorHandle>
 276             (TensorInfo({ numUnits, inputSize }, DataType::Float32));
 277     layer->m_BasicParameters.m_RecurrentToForgetWeights = std::make_unique<ScopedCpuTensorHandle>
 278             (TensorInfo({ numUnits, outputSize }, DataType::Float32));
 279     layer->m_BasicParameters.m_RecurrentToCellWeights = std::make_unique<ScopedCpuTensorHandle>
 280             (TensorInfo({ numUnits, outputSize }, DataType::Float32));
 281     layer->m_BasicParameters.m_RecurrentToOutputWeights = std::make_unique<ScopedCpuTensorHandle>
 282             (TensorInfo({ numUnits, outputSize }, DataType::Float32));
 283     layer->m_BasicParameters.m_ForgetGateBias = std::make_unique<ScopedCpuTensorHandle>
 284             (TensorInfo({ numUnits }, DataType::Float32));
 285     layer->m_BasicParameters.m_CellBias = std::make_unique<ScopedCpuTensorHandle>
 286             (TensorInfo({ numUnits }, DataType::Float32));
 287     layer->m_BasicParameters.m_OutputGateBias = std::make_unique<ScopedCpuTensorHandle>
 288             (TensorInfo({ numUnits }, DataType::Float32));
 289
 290     layer->m_BasicParameters.m_InputToForgetWeights->Allocate();
 291     layer->m_BasicParameters.m_InputToCellWeights->Allocate();
 292     layer->m_BasicParameters.m_InputToOutputWeights->Allocate();
 293     layer->m_BasicParameters.m_RecurrentToForgetWeights->Allocate();
 294     layer->m_BasicParameters.m_RecurrentToCellWeights->Allocate();
 295     layer->m_BasicParameters.m_RecurrentToOutputWeights->Allocate();
 296     layer->m_BasicParameters.m_ForgetGateBias->Allocate();
 297     layer->m_BasicParameters.m_CellBias->Allocate();
 298     layer->m_BasicParameters.m_OutputGateBias->Allocate();
 299
 300
 301     if (layerDesc.m_PeepholeEnabled)
 302     {
 303         layer->m_PeepholeParameters.m_CellToForgetWeights = std::make_unique<ScopedCpuTensorHandle>
 304                 (TensorInfo({ numUnits }, DataType::Float32));
 305         layer->m_PeepholeParameters.m_CellToOutputWeights = std::make_unique<ScopedCpuTensorHandle>
 306                 (TensorInfo({ numUnits }, DataType::Float32));
 307         layer->m_PeepholeParameters.m_CellToForgetWeights->Allocate();
 308         layer->m_PeepholeParameters.m_CellToOutputWeights->Allocate();
 309     }
 310
 311     // create input and output layers
 312     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 313     Layer* const outputStateIn = graph.AddLayer<InputLayer>(1, "outputStateIn");
 314     Layer* const cellStateIn = graph.AddLayer<InputLayer>(2, "cellStateIn");
 315     Layer* const scratchBuffer = graph.AddLayer<OutputLayer>(0, "scratchBuffer");
 316     Layer* const outputStateOut = graph.AddLayer<OutputLayer>(1, "outputStateOut");
 317     Layer* const cellStateOut = graph.AddLayer<OutputLayer>(2, "cellStateOut");
 318     Layer* const output = graph.AddLayer<OutputLayer>(3, "output");
 319
 320     // connect up
 321     armnn::TensorInfo lstmTensorInfo1({ batchSize, inputSize }, DataType::Float32);
 322     armnn::TensorInfo lstmTensorInfo2({ batchSize, numUnits}, DataType::Float32);
 323     armnn::TensorInfo lstmTensorInfo3({ batchSize, outputSize }, DataType::Float32);
 324     armnn::TensorInfo lstmTensorInfoScratchBuff({ batchSize, numUnits*3 }, DataType::Float32);
 325     if (layerDesc.m_CifgEnabled)
 326     {
 327         lstmTensorInfoScratchBuff.SetShape({ batchSize, numUnits*4 });
 328     }
 329
 330     Connect(input, layer, lstmTensorInfo1, 0, 0);
 331     Connect(cellStateIn, layer, lstmTensorInfo2, 0, 1);
 332     Connect(outputStateIn, layer, lstmTensorInfo3, 0, 2);
 333     Connect(layer, scratchBuffer, lstmTensorInfoScratchBuff, 0, 0);
 334     Connect(layer, outputStateOut, lstmTensorInfo3, 1, 0);
 335     Connect(layer, cellStateOut, lstmTensorInfo2, 2, 0);
 336     Connect(layer, output, lstmTensorInfo3, 3, 0);
 337
 338     CreateTensorHandles(graph, factory);
 339
 340     // make the workload and check it
 341     auto workload = MakeAndCheckWorkload<LstmWorkload>(*layer, graph, factory);
 342     LstmQueueDescriptor queueDescriptor = workload->GetData();
 343     BOOST_TEST(queueDescriptor.m_Parameters.m_ActivationFunc == 4);
 344     BOOST_TEST(queueDescriptor.m_Parameters.m_ClippingThresCell == 0.0f);
 345     BOOST_TEST(queueDescriptor.m_Parameters.m_ClippingThresProj == 0.0f);
 346     BOOST_TEST(queueDescriptor.m_Inputs.size() == 3);
 347     BOOST_TEST(queueDescriptor.m_Outputs.size() == 4);
 348
 349     BOOST_TEST((queueDescriptor.m_InputToForgetWeights->GetTensorInfo() == TensorInfo({ numUnits, inputSize },
 350                                                                                      DataType::Float32)));
 351     BOOST_TEST((queueDescriptor.m_OutputGateBias->GetTensorInfo() == TensorInfo({ numUnits },
 352                                                                                      DataType::Float32)));
 353     BOOST_TEST((queueDescriptor.m_CellBias->GetTensorInfo() == TensorInfo({ numUnits }, DataType::Float32)));
 354     return workload;
 355 }
 356
 357 template <typename Convolution2dWorkload, armnn::DataType DataType>
 358 std::unique_ptr<Convolution2dWorkload> CreateDirectConvolution2dWorkloadTest(armnn::IWorkloadFactory& factory,
 359                                                                        armnn::Graph&            graph)
 360 {
 361     // Creates the layer we're testing.
 362     Convolution2dDescriptor layerDesc;
 363     layerDesc.m_PadLeft = 1;
 364     layerDesc.m_PadRight = 1;
 365     layerDesc.m_PadTop = 1;
 366     layerDesc.m_PadBottom = 1;
 367     layerDesc.m_StrideX = 1;
 368     layerDesc.m_StrideY = 1;
 369     layerDesc.m_BiasEnabled = true;
 370
 371     Convolution2dLayer* const layer = graph.AddLayer<Convolution2dLayer>(layerDesc, "layer");
 372
 373     float inputsQScale = DataType == armnn::DataType::QuantisedAsymm8 ? 1.0f : 0.0;
 374     float outputQScale = DataType == armnn::DataType::QuantisedAsymm8 ? 2.0f : 0.0;
 375
 376     layer->m_Weight = std::make_unique<ScopedCpuTensorHandle>(TensorInfo({ 2, 3, 3, 3 }, DataType, inputsQScale));
 377     layer->m_Bias   = std::make_unique<ScopedCpuTensorHandle>
 378         (TensorInfo({2},  GetBiasDataType(DataType), inputsQScale));
 379     layer->m_Weight->Allocate();
 380     layer->m_Bias->Allocate();
 381
 382     // Creates extra layers.
 383     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 384     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
 385
 386     // Connects up.
 387     Connect(input, layer, TensorInfo({2, 3, 6, 6}, DataType, inputsQScale));
 388     Connect(layer, output, TensorInfo({2, 2, 6, 6}, DataType, outputQScale));
 389     CreateTensorHandles(graph, factory);
 390
 391     // Makes the workload and checks it.
 392     auto workload = MakeAndCheckWorkload<Convolution2dWorkload>(*layer, graph, factory);
 393
 394     Convolution2dQueueDescriptor queueDescriptor = workload->GetData();
 395     BOOST_TEST(queueDescriptor.m_Parameters.m_StrideX == 1);
 396     BOOST_TEST(queueDescriptor.m_Parameters.m_StrideY == 1);
 397     BOOST_TEST(queueDescriptor.m_Parameters.m_PadLeft == 1);
 398     BOOST_TEST(queueDescriptor.m_Parameters.m_PadRight == 1);
 399     BOOST_TEST(queueDescriptor.m_Parameters.m_PadTop == 1);
 400     BOOST_TEST(queueDescriptor.m_Parameters.m_PadBottom == 1);
 401     BOOST_TEST(queueDescriptor.m_Parameters.m_BiasEnabled == true);
 402
 403     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
 404     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
 405     BOOST_TEST((queueDescriptor.m_Weight->GetTensorInfo() == TensorInfo({2, 3, 3, 3},
 406         DataType, inputsQScale)));
 407     BOOST_TEST((queueDescriptor.m_Bias->GetTensorInfo()
 408                 == TensorInfo({2},  GetBiasDataType(DataType), inputsQScale)));
 409
 410     // Returns so we can do extra, backend-specific tests.
 411     return workload;
 412 }
 413
 414 template <typename DepthwiseConvolution2dFloat32Workload, armnn::DataType DataType>
 415 std::unique_ptr<DepthwiseConvolution2dFloat32Workload> CreateDepthwiseConvolution2dWorkloadTest(
 416     armnn::IWorkloadFactory& factory, armnn::Graph& graph, DataLayout dataLayout = DataLayout::NCHW)
 417 {
 418     // Creates the layer we're testing.
 419     DepthwiseConvolution2dDescriptor layerDesc;
 420     layerDesc.m_PadLeft         = 1;
 421     layerDesc.m_PadRight        = 2;
 422     layerDesc.m_PadTop          = 1;
 423     layerDesc.m_PadBottom       = 2;
 424     layerDesc.m_StrideX         = 1;
 425     layerDesc.m_StrideY         = 1;
 426     layerDesc.m_BiasEnabled     = false;
 427     layerDesc.m_DataLayout = dataLayout;
 428
 429     DepthwiseConvolution2dLayer* const layer = graph.AddLayer<DepthwiseConvolution2dLayer>(layerDesc, "layer");
 430
 431     layer->m_Weight = std::make_unique<ScopedCpuTensorHandle>(TensorInfo({1, 4, 4, 2}, DataType));
 432     layer->m_Weight->Allocate();
 433
 434     // Creates extra layers.
 435     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 436     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
 437
 438     TensorShape inputShape = (dataLayout == DataLayout::NCHW) ?
 439                 TensorShape{ 2, 2, 5, 5 } : TensorShape{ 2, 5, 5, 2 };
 440     TensorShape outputShape = (dataLayout == DataLayout::NCHW) ?
 441                 TensorShape{ 2, 2, 5, 5 } : TensorShape{ 2, 5, 5, 2 };
 442
 443     // Connects up.
 444     Connect(input, layer, TensorInfo(inputShape, DataType));
 445     Connect(layer, output, TensorInfo(outputShape, DataType));
 446     CreateTensorHandles(graph, factory);
 447
 448     // Makes the workload and checks it.
 449     auto workload = MakeAndCheckWorkload<DepthwiseConvolution2dFloat32Workload>(*layer, graph, factory);
 450
 451     DepthwiseConvolution2dQueueDescriptor queueDescriptor = workload->GetData();
 452     BOOST_TEST(queueDescriptor.m_Parameters.m_StrideX == 1);
 453     BOOST_TEST(queueDescriptor.m_Parameters.m_StrideY == 1);
 454     BOOST_TEST(queueDescriptor.m_Parameters.m_PadLeft == 1);
 455     BOOST_TEST(queueDescriptor.m_Parameters.m_PadRight == 2);
 456     BOOST_TEST(queueDescriptor.m_Parameters.m_PadTop == 1);
 457     BOOST_TEST(queueDescriptor.m_Parameters.m_PadBottom == 2);
 458     BOOST_TEST(queueDescriptor.m_Parameters.m_BiasEnabled == false);
 459     BOOST_TEST((queueDescriptor.m_Parameters.m_DataLayout == dataLayout));
 460
 461     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
 462     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
 463     BOOST_TEST((queueDescriptor.m_Weight->GetTensorInfo() == TensorInfo({1, 4, 4, 2}, DataType)));
 464
 465     // Returns so we can do extra, backend-specific tests.
 466     return workload;
 467 }
 468
 469 template <typename FullyConnectedWorkload, armnn::DataType DataType>
 470 std::unique_ptr<FullyConnectedWorkload> CreateFullyConnectedWorkloadTest(armnn::IWorkloadFactory& factory,
 471                                                                          armnn::Graph&            graph)
 472 {
 473     // Creates the layer we're testing.
 474     FullyConnectedDescriptor layerDesc;
 475     layerDesc.m_BiasEnabled = true;
 476     layerDesc.m_TransposeWeightMatrix = true;
 477
 478     FullyConnectedLayer* const layer = graph.AddLayer<FullyConnectedLayer>(layerDesc, "layer");
 479
 480     float inputsQScale = DataType == armnn::DataType::QuantisedAsymm8 ? 1.0f : 0.0;
 481     float outputQScale = DataType == armnn::DataType::QuantisedAsymm8 ? 2.0f : 0.0;
 482
 483     layer->m_Weight = std::make_unique<ScopedCpuTensorHandle>(TensorInfo({7, 20}, DataType, inputsQScale, 0));
 484     layer->m_Bias   = std::make_unique<ScopedCpuTensorHandle>(TensorInfo({7}, GetBiasDataType(DataType), inputsQScale));
 485     layer->m_Weight->Allocate();
 486     layer->m_Bias->Allocate();
 487
 488     // Creates extra layers.
 489     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 490     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
 491
 492     // Connects up.
 493     Connect(input, layer, TensorInfo({3, 1, 4, 5}, DataType, inputsQScale));
 494     Connect(layer, output, TensorInfo({3, 7}, DataType, outputQScale));
 495     CreateTensorHandles(graph, factory);
 496
 497     // Makes the workload and checks it.
 498     auto workload = MakeAndCheckWorkload<FullyConnectedWorkload>(*layer, graph, factory);
 499
 500     FullyConnectedQueueDescriptor queueDescriptor = workload->GetData();
 501     BOOST_TEST(queueDescriptor.m_Parameters.m_BiasEnabled == true);
 502     BOOST_TEST(queueDescriptor.m_Parameters.m_TransposeWeightMatrix == true);
 503
 504     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
 505     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
 506     BOOST_TEST((queueDescriptor.m_Weight->GetTensorInfo() == TensorInfo({7, 20}, DataType, inputsQScale)));
 507     BOOST_TEST((queueDescriptor.m_Bias->GetTensorInfo() == TensorInfo({7}, GetBiasDataType(DataType), inputsQScale)));
 508
 509     // Returns so we can do extra, backend-specific tests.
 510     return workload;
 511 }
 512
 513 template <typename NormalizationWorkload, armnn::DataType DataType>
 514 std::unique_ptr<NormalizationWorkload> CreateNormalizationWorkloadTest(armnn::IWorkloadFactory& factory,
 515                                                                        armnn::Graph& graph,
 516                                                                        DataLayout dataLayout = DataLayout::NCHW)
 517 {
 518     // Creates the layer we're testing.
 519     NormalizationDescriptor layerDesc;
 520     layerDesc.m_NormChannelType = NormalizationAlgorithmChannel::Across;
 521     layerDesc.m_NormMethodType = NormalizationAlgorithmMethod::LocalBrightness;
 522     layerDesc.m_NormSize = 3;
 523     layerDesc.m_Alpha = 0.5f;
 524     layerDesc.m_Beta = -1.0f;
 525     layerDesc.m_K = 0.2f;
 526     layerDesc.m_DataLayout = dataLayout;
 527
 528     NormalizationLayer* layer = graph.AddLayer<NormalizationLayer>(layerDesc, "layer");
 529
 530     // Creates extra layers.
 531     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 532     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
 533
 534     TensorShape inputShape = (dataLayout == DataLayout::NCHW) ?
 535                 TensorShape{ 3, 5, 5, 1 } : TensorShape{ 3, 1, 5, 5 };
 536     TensorShape outputShape = (dataLayout == DataLayout::NCHW) ?
 537                 TensorShape{ 3, 5, 5, 1 } : TensorShape{ 3, 1, 5, 5 };
 538
 539     // Connects up.
 540     armnn::TensorInfo inputTensorInfo(inputShape, DataType);
 541     armnn::TensorInfo outputTensorInfo(outputShape, DataType);
 542     Connect(input, layer, inputTensorInfo);
 543     Connect(layer, output, outputTensorInfo);
 544     CreateTensorHandles(graph, factory);
 545
 546     // Makes the workload and checks it.
 547     auto workload = MakeAndCheckWorkload<NormalizationWorkload>(*layer, graph, factory);
 548
 549     NormalizationQueueDescriptor queueDescriptor = workload->GetData();
 550     BOOST_TEST((queueDescriptor.m_Parameters.m_NormChannelType == NormalizationAlgorithmChannel::Across));
 551     BOOST_TEST((queueDescriptor.m_Parameters.m_NormMethodType == NormalizationAlgorithmMethod::LocalBrightness));
 552     BOOST_TEST(queueDescriptor.m_Parameters.m_NormSize == 3);
 553     BOOST_TEST(queueDescriptor.m_Parameters.m_Alpha == 0.5f);
 554     BOOST_TEST(queueDescriptor.m_Parameters.m_Beta == -1.0f);
 555     BOOST_TEST(queueDescriptor.m_Parameters.m_K == 0.2f);
 556     BOOST_TEST((queueDescriptor.m_Parameters.m_DataLayout == dataLayout));
 557
 558     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
 559     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
 560
 561     // Returns so we can do extra, backend-specific tests.
 562     return workload;
 563 }
 564
 565 template <typename Pooling2dWorkload, armnn::DataType DataType>
 566 std::unique_ptr<Pooling2dWorkload> CreatePooling2dWorkloadTest(armnn::IWorkloadFactory& factory,
 567                                                                armnn::Graph&            graph,
 568                                                                DataLayout dataLayout = DataLayout::NCHW)
 569 {
 570     // Creates the layer we're testing.
 571     Pooling2dDescriptor layerDesc;
 572     layerDesc.m_PoolType = PoolingAlgorithm::Average;
 573     layerDesc.m_PoolWidth = 3;
 574     layerDesc.m_PoolHeight = 3;
 575     layerDesc.m_PadLeft = 2;
 576     layerDesc.m_PadRight = 2;
 577     layerDesc.m_PadTop = 1;
 578     layerDesc.m_PadBottom = 1;
 579     layerDesc.m_StrideX = 2;
 580     layerDesc.m_StrideY = 3;
 581     layerDesc.m_OutputShapeRounding = OutputShapeRounding::Floor;
 582     layerDesc.m_DataLayout = dataLayout;
 583
 584     Pooling2dLayer* const layer = graph.AddLayer<Pooling2dLayer>(layerDesc, "layer");
 585
 586     // Create extra layers
 587     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 588     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
 589
 590     TensorShape inputShape  = (dataLayout == DataLayout::NCHW) ? TensorShape{3, 2, 5, 5} : TensorShape{3, 5, 5, 2};
 591     TensorShape outputShape = (dataLayout == DataLayout::NCHW) ? TensorShape{3, 2, 2, 4} : TensorShape{3, 2, 4, 2};
 592
 593     // Connect up
 594     Connect(input, layer, TensorInfo(inputShape, DataType));
 595     Connect(layer, output, TensorInfo(outputShape, DataType));
 596     CreateTensorHandles(graph, factory);
 597
 598     // Make the workload and checks it
 599     auto workload = MakeAndCheckWorkload<Pooling2dWorkload>(*layer, graph, factory);
 600
 601     Pooling2dQueueDescriptor queueDescriptor = workload->GetData();
 602     BOOST_TEST((queueDescriptor.m_Parameters.m_PoolType == PoolingAlgorithm::Average));
 603     BOOST_TEST((queueDescriptor.m_Parameters.m_OutputShapeRounding == OutputShapeRounding::Floor));
 604     BOOST_TEST(queueDescriptor.m_Parameters.m_PoolWidth == 3);
 605     BOOST_TEST(queueDescriptor.m_Parameters.m_PoolHeight == 3);
 606     BOOST_TEST(queueDescriptor.m_Parameters.m_StrideX == 2);
 607     BOOST_TEST(queueDescriptor.m_Parameters.m_StrideY == 3);
 608     BOOST_TEST(queueDescriptor.m_Parameters.m_PadLeft == 2);
 609     BOOST_TEST(queueDescriptor.m_Parameters.m_PadRight == 2);
 610     BOOST_TEST(queueDescriptor.m_Parameters.m_PadTop == 1);
 611     BOOST_TEST(queueDescriptor.m_Parameters.m_PadBottom == 1);
 612     BOOST_TEST((queueDescriptor.m_Parameters.m_DataLayout == dataLayout));
 613
 614     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
 615     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
 616
 617     // Return so we can do extra, backend-specific tests
 618     return workload;
 619 }
 620
 621 template <typename SoftmaxWorkload, armnn::DataType DataType>
 622 std::unique_ptr<SoftmaxWorkload> CreateSoftmaxWorkloadTest(armnn::IWorkloadFactory& factory,
 623                                                            armnn::Graph&            graph)
 624 {
 625     // Create the layer we're testing.
 626     SoftmaxDescriptor softmaxDescriptor;
 627     Layer* const layer = graph.AddLayer<SoftmaxLayer>(softmaxDescriptor, "layer");
 628
 629     // Create extra layers.
 630     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 631     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
 632
 633     // Connect up
 634     armnn::TensorInfo tensorInfo({4, 1}, DataType);
 635     Connect(input, layer, tensorInfo);
 636     Connect(layer, output, tensorInfo);
 637     CreateTensorHandles(graph, factory);
 638
 639     // Make the workload and checks it.
 640     auto workload = MakeAndCheckWorkload<SoftmaxWorkload>(*layer, graph, factory);
 641
 642     SoftmaxQueueDescriptor queueDescriptor = workload->GetData();
 643     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
 644     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
 645
 646     // Return so we can do extra, backend-specific tests.
 647     return workload;
 648 }
 649
 650 template<typename SplitterWorkload, armnn::DataType DataType>
 651 std::unique_ptr<SplitterWorkload>
 652     CreateSplitterWorkloadTest(armnn::IWorkloadFactory& factory, armnn::Graph& graph)
 653 {
 654     // Create the layer we're testing.
 655     // NOTE: need three dimensions channels, height/y, width/x because the Compute
 656     //       library restricts subtensors to have the same x and y dimensions as
 657     //       their parent tensors, and therefore the origin on the x and y dimension
 658     //       has to be zero for any view. So we need a third dimension to split...
 659     // NOTE: arguments are: number of views, number of dimensions.
 660     ViewsDescriptor layerDesc(3, 3);
 661     // NOTE: arguments are: view, dimension, value.
 662     layerDesc.SetViewOriginCoord(0, 0, 0);
 663     layerDesc.SetViewOriginCoord(1, 0, 1);
 664     layerDesc.SetViewOriginCoord(2, 0, 3);
 665
 666     Layer* const layer = graph.AddLayer<SplitterLayer>(layerDesc, "layer");
 667
 668     // Adds extra layers.
 669     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 670     Layer* const output0 = graph.AddLayer<OutputLayer>(0, "output0");
 671     Layer* const output1 = graph.AddLayer<OutputLayer>(1, "output1");
 672     Layer* const output2 = graph.AddLayer<OutputLayer>(2, "output2");
 673
 674     // Connects up.
 675     armnn::TensorInfo tensorInfo({5, 7, 7}, DataType);
 676     Connect(input, layer, tensorInfo);
 677
 678     armnn::TensorInfo output0Info({1, 7, 7}, DataType);
 679     armnn::TensorInfo output1Info({2, 7, 7}, DataType);
 680     armnn::TensorInfo output2Info({2, 7, 7}, DataType);
 681
 682     Connect(layer, output0, output0Info, 0, 0);
 683     Connect(layer, output1, output1Info, 1, 0);
 684     Connect(layer, output2, output2Info, 2, 0);
 685
 686     CreateTensorHandles(graph, factory);
 687
 688     // Makes the workload and checks it.
 689     auto workload = MakeAndCheckWorkload<SplitterWorkload>(*layer, graph, factory);
 690
 691     SplitterQueueDescriptor queueDescriptor = workload->GetData();
 692     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
 693     BOOST_TEST(queueDescriptor.m_Outputs.size() == 3);
 694     BOOST_TEST(queueDescriptor.m_ViewOrigins.size() == 3);
 695
 696     BOOST_TEST(queueDescriptor.m_ViewOrigins[0].m_Origin[0] == 0);
 697     BOOST_TEST(queueDescriptor.m_ViewOrigins[1].m_Origin[0] == 1);
 698     BOOST_TEST(queueDescriptor.m_ViewOrigins[2].m_Origin[0] == 3);
 699     BOOST_TEST(queueDescriptor.m_ViewOrigins[0].m_Origin[1] == 0);
 700     BOOST_TEST(queueDescriptor.m_ViewOrigins[1].m_Origin[1] == 0);
 701     BOOST_TEST(queueDescriptor.m_ViewOrigins[2].m_Origin[1] == 0);
 702     BOOST_TEST(queueDescriptor.m_ViewOrigins[0].m_Origin[2] == 0);
 703     BOOST_TEST(queueDescriptor.m_ViewOrigins[1].m_Origin[2] == 0);
 704     BOOST_TEST(queueDescriptor.m_ViewOrigins[2].m_Origin[2] == 0);
 705
 706     // Returns so we can do extra, backend-specific tests.
 707     return workload;
 708 }
 709
 710 /// This function constructs a graph with both a splitter and a merger, and returns a pair of the workloads.
 711 template<typename SplitterWorkload, typename MergerWorkload, armnn::DataType DataType>
 712 std::pair<std::unique_ptr<SplitterWorkload>, std::unique_ptr<MergerWorkload>>
 713     CreateSplitterMergerWorkloadTest(armnn::IWorkloadFactory& factory, armnn::Graph& graph)
 714 {
 715     armnn::TensorInfo inputTensorInfo({ 1, 2, 100, 10 }, DataType);
 716
 717     armnn::TensorInfo splitTensorInfo1({ 1, 1, 100, 10 }, DataType);
 718     armnn::TensorInfo splitTensorInfo2({ 1, 1, 100, 10 }, DataType);
 719
 720     //Constructs the graph.
 721     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 722
 723     armnn::ViewsDescriptor splitterViews(2);
 724     splitterViews.SetViewOriginCoord(0, 0, 0);
 725     splitterViews.SetViewOriginCoord(0, 1, 0);
 726     splitterViews.SetViewOriginCoord(0, 2, 0);
 727     splitterViews.SetViewOriginCoord(0, 3, 0);
 728
 729     splitterViews.SetViewOriginCoord(1, 0, 0);
 730     splitterViews.SetViewOriginCoord(1, 1, 1);
 731     splitterViews.SetViewOriginCoord(1, 2, 0);
 732     splitterViews.SetViewOriginCoord(1, 3, 0);
 733
 734     Layer* const splitter = graph.AddLayer<SplitterLayer>(splitterViews, "splitter");
 735     BOOST_TEST_CHECKPOINT("created splitter layer");
 736
 737     armnn::OriginsDescriptor mergerViews(2);
 738     mergerViews.SetViewOriginCoord(0, 0, 0);
 739     mergerViews.SetViewOriginCoord(0, 1, 1);
 740     mergerViews.SetViewOriginCoord(0, 2, 0);
 741     mergerViews.SetViewOriginCoord(0, 3, 0);
 742
 743     mergerViews.SetViewOriginCoord(1, 0, 0);
 744     mergerViews.SetViewOriginCoord(1, 1, 0);
 745     mergerViews.SetViewOriginCoord(1, 2, 0);
 746     mergerViews.SetViewOriginCoord(1, 3, 0);
 747
 748     Layer* const merger = graph.AddLayer<MergerLayer>(mergerViews, "merger");
 749     BOOST_TEST_CHECKPOINT("created merger layer");
 750
 751     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
 752
 753     // Adds connections.
 754     Connect(input, splitter, inputTensorInfo, 0, 0);
 755     BOOST_TEST_CHECKPOINT("connect input to splitter");
 756     Connect(splitter, merger, splitTensorInfo1, 0, 1); // The splitter & merger are connected up.
 757     BOOST_TEST_CHECKPOINT("connect splitter[0] to merger[1]");
 758     Connect(splitter, merger, splitTensorInfo2, 1, 0); // So that the outputs are flipped round.
 759     BOOST_TEST_CHECKPOINT("connect splitter[1] to merger[0]");
 760     Connect(merger, output, inputTensorInfo, 0, 0);
 761     BOOST_TEST_CHECKPOINT("connect merger to output");
 762
 763     CreateTensorHandles(graph, factory);
 764     BOOST_TEST_CHECKPOINT("created tensor handles");
 765
 766     auto workloadSplitter = MakeAndCheckWorkload<SplitterWorkload>(*splitter, graph, factory);
 767     BOOST_TEST_CHECKPOINT("created splitter workload");
 768     auto workloadMerger = MakeAndCheckWorkload<MergerWorkload>(*merger, graph, factory);
 769     BOOST_TEST_CHECKPOINT("created merger workload");
 770
 771     return {std::move(workloadSplitter), std::move(workloadMerger)};
 772 }
 773
 774
 775 /// This function constructs a graph with a splitter with two outputs. Each of the outputs is then
 776 /// connected to two different activation layers
 777 template<typename SplitterWorkload, typename ActivationWorkload, armnn::DataType DataType>
 778 void CreateSplitterMultipleInputsOneOutputWorkloadTest(armnn::IWorkloadFactory& factory, armnn::Graph& graph,
 779                                  std::unique_ptr<SplitterWorkload>& wlSplitter,
 780                                  std::unique_ptr<ActivationWorkload>& wlActiv0_0,
 781                                  std::unique_ptr<ActivationWorkload>& wlActiv0_1,
 782                                  std::unique_ptr<ActivationWorkload>& wlActiv1_0,
 783                                  std::unique_ptr<ActivationWorkload>& wlActiv1_1)
 784 {
 785     armnn::TensorInfo inputTensorInfo ({ 1, 3, 100, 50 }, DataType);
 786     armnn::TensorInfo splitTensorInfo1({ 1, 1, 100, 50 }, DataType);
 787     armnn::TensorInfo splitTensorInfo2({ 1, 2, 100, 50 }, DataType);
 788
 789     //Constructs the graph.
 790     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 791
 792     armnn::ViewsDescriptor splitterViews(2);
 793
 794     splitterViews.SetViewOriginCoord(0, 0, 0);
 795     splitterViews.SetViewOriginCoord(0, 1, 0);
 796     splitterViews.SetViewOriginCoord(0, 2, 0);
 797     splitterViews.SetViewOriginCoord(0, 3, 0);
 798
 799     splitterViews.SetViewOriginCoord(1, 0, 0);
 800     splitterViews.SetViewOriginCoord(1, 1, 1);
 801     splitterViews.SetViewOriginCoord(1, 2, 0);
 802     splitterViews.SetViewOriginCoord(1, 3, 0);
 803
 804     Layer* const splitter = graph.AddLayer<SplitterLayer>(splitterViews, "splitter");
 805
 806     armnn::ActivationDescriptor activationDesc;
 807
 808     Layer* const activ0_0 = graph.AddLayer<ActivationLayer>(activationDesc, "activ0_0");
 809     Layer* const activ0_1 = graph.AddLayer<ActivationLayer>(activationDesc, "activ0_1");
 810     Layer* const activ1_0 = graph.AddLayer<ActivationLayer>(activationDesc, "activ1_0");
 811     Layer* const activ1_1 = graph.AddLayer<ActivationLayer>(activationDesc, "activ1_1");
 812
 813     Layer* const output1 = graph.AddLayer<OutputLayer>(1, "output1");
 814     Layer* const output2 = graph.AddLayer<OutputLayer>(2, "output2");
 815     Layer* const output3 = graph.AddLayer<OutputLayer>(3, "output3");
 816     Layer* const output4 = graph.AddLayer<OutputLayer>(4, "output4");
 817
 818     // Adds connections.
 819     Connect(input, splitter, inputTensorInfo, 0, 0);
 820     Connect(splitter, activ0_0, splitTensorInfo1, 0, 0);
 821     Connect(splitter, activ0_1, splitTensorInfo1, 0, 0);
 822
 823     Connect(splitter, activ1_0, splitTensorInfo2, 1, 0);
 824     Connect(splitter, activ1_1, splitTensorInfo2, 1, 0);
 825
 826     Connect(activ0_0, output1, splitTensorInfo1, 0, 0);
 827     Connect(activ0_1, output2, splitTensorInfo1, 0, 0);
 828     Connect(activ1_0, output3, splitTensorInfo2, 0, 0);
 829     Connect(activ1_1, output4, splitTensorInfo2, 0, 0);
 830
 831     CreateTensorHandles(graph, factory);
 832
 833     auto workloadSplitter = MakeAndCheckWorkload<SplitterWorkload>(*splitter, graph, factory);
 834     auto workloadActiv0_0 = MakeAndCheckWorkload<ActivationWorkload>(*activ0_0, graph, factory);
 835     auto workloadActiv0_1 = MakeAndCheckWorkload<ActivationWorkload>(*activ0_1, graph, factory);
 836     auto workloadActiv1_0 = MakeAndCheckWorkload<ActivationWorkload>(*activ1_0, graph, factory);
 837     auto workloadActiv1_1 = MakeAndCheckWorkload<ActivationWorkload>(*activ1_1, graph, factory);
 838
 839     wlSplitter = std::move(workloadSplitter);
 840     wlActiv0_0 = std::move(workloadActiv0_0);
 841     wlActiv0_1 = std::move(workloadActiv0_1);
 842     wlActiv1_0 = std::move(workloadActiv1_0);
 843     wlActiv1_1 = std::move(workloadActiv1_1);
 844 }
 845
 846 template <typename ResizeBilinearWorkload, armnn::DataType DataType>
 847 std::unique_ptr<ResizeBilinearWorkload> CreateResizeBilinearWorkloadTest(armnn::IWorkloadFactory& factory,
 848                                                                          armnn::Graph& graph,
 849                                                                          DataLayoutIndexed dataLayout =
 850                                                                              DataLayout::NCHW)
 851 {
 852     TensorShape inputShape;
 853     TensorShape outputShape;
 854
 855     switch (dataLayout.GetDataLayout()) {
 856         case DataLayout::NHWC:
 857             inputShape =  { 2, 4, 4, 3 };
 858             outputShape = { 2, 2, 2, 3 };
 859             break;
 860         case DataLayout::NCHW:
 861         default:
 862             inputShape =  { 2, 3, 4, 4 };
 863             outputShape = { 2, 3, 2, 2 };
 864     }
 865
 866     // Creates the layer we're testing.
 867     ResizeBilinearDescriptor resizeDesc;
 868     resizeDesc.m_TargetWidth = outputShape[dataLayout.GetWidthIndex()];
 869     resizeDesc.m_TargetHeight = outputShape[dataLayout.GetHeightIndex()];
 870     resizeDesc.m_DataLayout = dataLayout;
 871     Layer* const layer = graph.AddLayer<ResizeBilinearLayer>(resizeDesc, "layer");
 872
 873     // Creates extra layers.
 874     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 875     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
 876
 877     // Connects up.
 878     armnn::TensorInfo inputTensorInfo(inputShape, DataType);
 879     armnn::TensorInfo outputTensorInfo(outputShape, DataType);
 880     Connect(input, layer, inputTensorInfo);
 881     Connect(layer, output, outputTensorInfo);
 882     CreateTensorHandles(graph, factory);
 883
 884     // Makes the workload and checks it.
 885     auto workload = MakeAndCheckWorkload<ResizeBilinearWorkload>(*layer, graph, factory);
 886
 887     ResizeBilinearQueueDescriptor queueDescriptor = workload->GetData();
 888     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
 889     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
 890     BOOST_TEST((queueDescriptor.m_Parameters.m_DataLayout.GetDataLayout() == dataLayout));
 891
 892     // Returns so we can do extra, backend-specific tests.
 893     return workload;
 894 }
 895
 896 template <typename L2NormalizationWorkload, armnn::DataType DataType>
 897 std::unique_ptr<L2NormalizationWorkload> CreateL2NormalizationWorkloadTest(armnn::IWorkloadFactory& factory,
 898     armnn::Graph& graph, DataLayout dataLayout = DataLayout::NCHW)
 899 {
 900     // Creates the layer we're testing.
 901     L2NormalizationDescriptor layerDesc;
 902     layerDesc.m_DataLayout = dataLayout;
 903
 904     Layer* const layer = graph.AddLayer<L2NormalizationLayer>(layerDesc, "l2norm");
 905
 906     // Creates extra layers.
 907     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 908     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
 909
 910     TensorShape inputShape = (dataLayout == DataLayout::NCHW) ?
 911                 TensorShape{ 5, 20, 50, 67 } : TensorShape{ 5, 50, 67, 20 };
 912     TensorShape outputShape = (dataLayout == DataLayout::NCHW) ?
 913                 TensorShape{ 5, 20, 50, 67 } : TensorShape{ 5, 50, 67, 20 };
 914
 915     // Connects up.
 916     armnn::TensorInfo inputTensorInfo(inputShape, DataType);
 917     armnn::TensorInfo outputTensorInfo(outputShape, DataType);
 918     Connect(input, layer, inputTensorInfo);
 919     Connect(layer, output, outputTensorInfo);
 920     CreateTensorHandles(graph, factory);
 921
 922     // Makes the workload and checks it.
 923     auto workload = MakeAndCheckWorkload<L2NormalizationWorkload>(*layer, graph, factory);
 924
 925     L2NormalizationQueueDescriptor queueDescriptor = workload->GetData();
 926     BOOST_TEST((queueDescriptor.m_Parameters.m_DataLayout == dataLayout));
 927     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
 928     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
 929
 930     // Returns so we can do extra, backend-specific tests.
 931     return workload;
 932 }
 933
 934 template <typename ReshapeWorkload, armnn::DataType DataType>
 935 std::unique_ptr<ReshapeWorkload> CreateReshapeWorkloadTest(armnn::IWorkloadFactory& factory,
 936     armnn::Graph& graph)
 937 {
 938     // Creates the layer we're testing.
 939     TensorShape outputShape({ 1, 4 });
 940     ReshapeDescriptor reshapeDesc;
 941     reshapeDesc.m_TargetShape = outputShape;
 942     Layer* const layer = graph.AddLayer<ReshapeLayer>(reshapeDesc, "layer");
 943
 944     // Creates extra layers.
 945     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 946     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
 947
 948     // Connects up.
 949     armnn::TensorInfo inputTensorInfo({ 4, 1 }, DataType);
 950     armnn::TensorInfo outputTensorInfo(outputShape, DataType);
 951     Connect(input, layer, inputTensorInfo);
 952     Connect(layer, output, outputTensorInfo);
 953     CreateTensorHandles(graph, factory);
 954
 955     // Makes the workload and checks it.
 956     auto workload = MakeAndCheckWorkload<ReshapeWorkload>(*layer, graph, factory);
 957
 958     ReshapeQueueDescriptor queueDescriptor = workload->GetData();
 959     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
 960     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
 961
 962     // Returns so we can do extra, backend-specific tests.
 963     return workload;
 964 }
 965
 966 template <typename ConvertFp16ToFp32Float32Workload>
 967 std::unique_ptr<ConvertFp16ToFp32Float32Workload> CreateConvertFp16ToFp32WorkloadTest(
 968     armnn::IWorkloadFactory& factory, armnn::Graph& graph)
 969 {
 970     // Creates the layer we're testing.
 971     ConvertFp16ToFp32Layer* const layer = graph.AddLayer<ConvertFp16ToFp32Layer>("Fp16ToFp32Converter");
 972
 973     // Creates extra layers.
 974     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
 975     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
 976
 977     // Connects up.
 978     armnn::TensorInfo inputTensorInfo({1, 3, 2, 3}, armnn::DataType::Float16);
 979     armnn::TensorInfo outputTensorInfo({1, 3, 2, 3}, armnn::DataType::Float32);
 980     Connect(input, layer, inputTensorInfo);
 981     Connect(layer, output, outputTensorInfo);
 982     CreateTensorHandles(graph, factory);
 983
 984     // Makes the workload and checks it.
 985     auto workload = MakeAndCheckWorkload<ConvertFp16ToFp32Float32Workload>(*layer, graph, factory);
 986
 987     ConvertFp16ToFp32QueueDescriptor queueDescriptor = workload->GetData();
 988     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
 989     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
 990
 991     // Returns so we can do extra, backend-specific tests.
 992     return workload;
 993 }
 994
 995 template <typename ConvertFp32ToFp16Float16Workload>
 996 std::unique_ptr<ConvertFp32ToFp16Float16Workload> CreateConvertFp32ToFp16WorkloadTest(
 997     armnn::IWorkloadFactory& factory, armnn::Graph& graph)
 998 {
 999     // Creates the layer we're testing.
1000     ConvertFp32ToFp16Layer* const layer = graph.AddLayer<ConvertFp32ToFp16Layer>("Fp32ToFp16Converter");
1001
1002     // Creates extra layers.
1003     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
1004     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
1005
1006     // Connects up.
1007     armnn::TensorInfo inputTensorInfo({1, 3, 2, 3}, armnn::DataType::Float32);
1008     armnn::TensorInfo outputTensorInfo({1, 3, 2, 3}, armnn::DataType::Float16);
1009     Connect(input, layer, inputTensorInfo);
1010     Connect(layer, output, outputTensorInfo);
1011     CreateTensorHandles(graph, factory);
1012
1013     // Makes the workload and checks it.
1014     auto workload = MakeAndCheckWorkload<ConvertFp32ToFp16Float16Workload>(*layer, graph, factory);
1015
1016     ConvertFp32ToFp16QueueDescriptor queueDescriptor = workload->GetData();
1017     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
1018     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
1019
1020     // Returns so we can do extra, backend-specific tests.
1021     return workload;
1022 }
1023
1024 template <typename MeanWorkload, armnn::DataType DataType>
1025 std::unique_ptr<MeanWorkload> CreateMeanWorkloadTest(armnn::IWorkloadFactory& factory, armnn::Graph& graph)
1026 {
1027     // Reduce along the first and second dimensions, and do not keep the reduced dimensions.
1028     MeanDescriptor descriptor({ 1, 2 }, false);
1029
1030     // Creates the layer we're testing.
1031     Layer* const layer = graph.AddLayer<MeanLayer>(descriptor, "mean");
1032
1033     // Creates extra layers.
1034     Layer* const input = graph.AddLayer<InputLayer>(0, "input");
1035     Layer* const output = graph.AddLayer<OutputLayer>(0, "output");
1036
1037     // Connects up.
1038     armnn::TensorInfo inputTensorInfo({ 1, 3, 7, 4 }, DataType);
1039     armnn::TensorInfo outputTensorInfo({ 1, 4 }, DataType);
1040     Connect(input, layer, inputTensorInfo);
1041     Connect(layer, output, outputTensorInfo);
1042     CreateTensorHandles(graph, factory);
1043
1044     // Makes the workload and checks it.
1045     auto workload = MakeAndCheckWorkload<MeanWorkload>(*layer, graph, factory);
1046
1047     MeanQueueDescriptor queueDescriptor = workload->GetData();
1048     BOOST_TEST(queueDescriptor.m_Parameters.m_Axis == descriptor.m_Axis);
1049     BOOST_TEST(queueDescriptor.m_Parameters.m_KeepDims == descriptor.m_KeepDims);
1050     BOOST_TEST(queueDescriptor.m_Inputs.size() == 1);
1051     BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
1052
1053     // Returns so we can do extra, backend-specific tests.
1054     return workload;
1055 }
1056
1057 }