src/backends/cl/test/Fp16SupportTest.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5
   6 #include <armnn/ArmNN.hpp>
   7 #include <armnn/Descriptors.hpp>
   8 #include <armnn/IRuntime.hpp>
   9 #include <armnn/INetwork.hpp>
  10 #include <Half.hpp>
  11
  12 #include <Graph.hpp>
  13 #include <Optimizer.hpp>
  14 #include <backendsCommon/CpuTensorHandle.hpp>
  15 #include <backendsCommon/test/QuantizeHelper.hpp>
  16
  17 #include <boost/core/ignore_unused.hpp>
  18 #include <boost/test/unit_test.hpp>
  19
  20 #include <set>
  21
  22 using namespace armnn;
  23
  24 BOOST_AUTO_TEST_SUITE(Fp16Support)
  25
  26 BOOST_AUTO_TEST_CASE(Fp16DataTypeSupport)
  27 {
  28     Graph graph;
  29
  30     Layer* const inputLayer1 = graph.AddLayer<InputLayer>(1, "input1");
  31     Layer* const inputLayer2 = graph.AddLayer<InputLayer>(2, "input2");
  32
  33     Layer* const additionLayer = graph.AddLayer<AdditionLayer>("addition");
  34     Layer* const outputLayer = graph.AddLayer<armnn::OutputLayer>(0, "output");
  35
  36     TensorInfo fp16TensorInfo({1, 2, 3, 5}, armnn::DataType::Float16);
  37     inputLayer1->GetOutputSlot(0).Connect(additionLayer->GetInputSlot(0));
  38     inputLayer2->GetOutputSlot(0).Connect(additionLayer->GetInputSlot(1));
  39     additionLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
  40
  41     inputLayer1->GetOutputSlot().SetTensorInfo(fp16TensorInfo);
  42     inputLayer2->GetOutputSlot().SetTensorInfo(fp16TensorInfo);
  43     additionLayer->GetOutputSlot().SetTensorInfo(fp16TensorInfo);
  44
  45     BOOST_CHECK(inputLayer1->GetOutputSlot(0).GetTensorInfo().GetDataType() == armnn::DataType::Float16);
  46     BOOST_CHECK(inputLayer2->GetOutputSlot(0).GetTensorInfo().GetDataType() == armnn::DataType::Float16);
  47     BOOST_CHECK(additionLayer->GetOutputSlot(0).GetTensorInfo().GetDataType() == armnn::DataType::Float16);
  48 }
  49
  50 BOOST_AUTO_TEST_CASE(Fp16AdditionTest)
  51 {
  52    using namespace half_float::literal;
  53    // Create runtime in which test will run
  54    IRuntime::CreationOptions options;
  55    IRuntimePtr runtime(IRuntime::Create(options));
  56
  57    // Builds up the structure of the network.
  58    INetworkPtr net(INetwork::Create());
  59
  60    IConnectableLayer* inputLayer1 = net->AddInputLayer(0);
  61    IConnectableLayer* inputLayer2 = net->AddInputLayer(1);
  62    IConnectableLayer* additionLayer = net->AddAdditionLayer();
  63    IConnectableLayer* outputLayer = net->AddOutputLayer(0);
  64
  65    inputLayer1->GetOutputSlot(0).Connect(additionLayer->GetInputSlot(0));
  66    inputLayer2->GetOutputSlot(0).Connect(additionLayer->GetInputSlot(1));
  67    additionLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
  68
  69    //change to float16
  70    TensorInfo fp16TensorInfo(TensorShape({4}), DataType::Float16);
  71    inputLayer1->GetOutputSlot(0).SetTensorInfo(fp16TensorInfo);
  72    inputLayer2->GetOutputSlot(0).SetTensorInfo(fp16TensorInfo);
  73    additionLayer->GetOutputSlot(0).SetTensorInfo(fp16TensorInfo);
  74
  75    // optimize the network
  76    std::vector<BackendId> backends = {Compute::GpuAcc};
  77    IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec());
  78
  79    // Loads it into the runtime.
  80    NetworkId netId;
  81    runtime->LoadNetwork(netId, std::move(optNet));
  82
  83    std::vector<Half> input1Data
  84    {
  85        1.0_h, 2.0_h, 3.0_h, 4.0_h
  86    };
  87
  88    std::vector<Half> input2Data
  89    {
  90        100.0_h, 200.0_h, 300.0_h, 400.0_h
  91    };
  92
  93    InputTensors inputTensors
  94    {
  95        {0,ConstTensor(runtime->GetInputTensorInfo(netId, 0), input1Data.data())},
  96        {1,ConstTensor(runtime->GetInputTensorInfo(netId, 0), input2Data.data())}
  97    };
  98
  99    std::vector<Half> outputData(input1Data.size());
 100    OutputTensors outputTensors
 101    {
 102        {0,Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data())}
 103    };
 104
 105    // Does the inference.
 106    runtime->EnqueueWorkload(netId, inputTensors, outputTensors);
 107
 108    // Checks the results.
 109    BOOST_TEST(outputData == std::vector<Half>({ 101.0_h, 202.0_h, 303.0_h, 404.0_h})); // Add
 110 }
 111
 112 BOOST_AUTO_TEST_SUITE_END()