src/backends/reference/test/RefOptimizedNetworkTests.cpp

   1 //
   2 // Copyright © 2017 Arm Ltd. All rights reserved.
   3 // SPDX-License-Identifier: MIT
   4 //
   5
   6 #include <armnn/ArmNN.hpp>
   7 #include <Graph.hpp>
   8 #include <Network.hpp>
   9
  10 #include <reference/RefWorkloadFactory.hpp>
  11
  12 #include <boost/test/unit_test.hpp>
  13
  14 BOOST_AUTO_TEST_SUITE(RefOptimizedNetwork)
  15
  16 BOOST_AUTO_TEST_CASE(OptimizeValidateCpuRefWorkloads)
  17 {
  18     const armnn::TensorInfo desc({3, 5}, armnn::DataType::Float32);
  19
  20     armnn::Network  net;
  21
  22     armnn::NormalizationDescriptor nmDesc;
  23     armnn::ActivationDescriptor acDesc;
  24
  25     //    in
  26     //     |
  27     //    nm
  28     //   /  |
  29     //  ac  |
  30     //   \  |
  31     //    ml
  32     //     |
  33     //    sm
  34     //     |
  35     //    ot
  36     armnn::IConnectableLayer* layer = net.AddInputLayer(0, "in");
  37     layer->GetOutputSlot(0).SetTensorInfo(desc);
  38
  39     armnn::IConnectableLayer* const normLayer = net.AddNormalizationLayer(nmDesc, "nm");
  40
  41     layer->GetOutputSlot(0).Connect(normLayer->GetInputSlot(0));
  42     normLayer->GetOutputSlot(0).SetTensorInfo(desc);
  43
  44     layer = net.AddActivationLayer(acDesc, "ac");
  45
  46     normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
  47     layer->GetOutputSlot(0).SetTensorInfo(desc);
  48
  49     armnn::IConnectableLayer* prevLayer = layer;
  50     layer = net.AddMultiplicationLayer("ml");
  51
  52     prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
  53     normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1));
  54     layer->GetOutputSlot(0).SetTensorInfo(desc);
  55
  56     prevLayer = layer;
  57     armnn::SoftmaxDescriptor softmaxDescriptor;
  58     layer = net.AddSoftmaxLayer(softmaxDescriptor, "sm");
  59
  60     prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
  61     layer->GetOutputSlot(0).SetTensorInfo(desc);
  62
  63     prevLayer = layer;
  64     layer = net.AddOutputLayer(0, "ot");
  65
  66     prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0));
  67
  68     armnn::IRuntime::CreationOptions options;
  69     armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
  70
  71     std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef };
  72     armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec());
  73     static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph().AllocateDynamicBuffers();
  74     BOOST_CHECK(optNet);
  75
  76     // Validates workloads.
  77     armnn::RefWorkloadFactory fact;
  78     for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
  79     {
  80         BOOST_CHECK_NO_THROW(
  81             layer->CreateWorkload(static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph(), fact));
  82     }
  83 }
  84
  85 BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsCpuRefPermuteLayer)
  86 {
  87     // Create runtime in which test will run
  88     armnn::IRuntime::CreationOptions options;
  89     armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
  90
  91     std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
  92
  93     // build up the structure of the network
  94     armnn::INetworkPtr net(armnn::INetwork::Create());
  95
  96     armnn::IConnectableLayer* input = net->AddInputLayer(0);
  97
  98     armnn::PermuteDescriptor descriptor({0, 2, 3, 1});
  99     armnn::IConnectableLayer* permute = net->AddPermuteLayer(descriptor);
 100
 101     armnn::IConnectableLayer* output = net->AddOutputLayer(0);
 102
 103     input->GetOutputSlot(0).Connect(permute->GetInputSlot(0));
 104     permute->GetOutputSlot(0).Connect(output->GetInputSlot(0));
 105
 106     input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32));
 107     permute->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 4, 1, 4 }, armnn::DataType::Float32));
 108
 109     // optimize the network
 110     armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
 111
 112     for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
 113     {
 114         BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
 115     }
 116 }
 117
 118 BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsCpuRefMeanLayer)
 119 {
 120     // Create runtime in which test will run
 121     armnn::IRuntime::CreationOptions options;
 122     armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
 123
 124     std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
 125
 126     // build up the structure of the network
 127     armnn::INetworkPtr net(armnn::INetwork::Create());
 128
 129     armnn::IConnectableLayer* input = net->AddInputLayer(0);
 130
 131     armnn::MeanDescriptor descriptor({ 0, 1 }, false);
 132     armnn::IConnectableLayer* meanLayer = net->AddMeanLayer(descriptor);
 133
 134     armnn::IConnectableLayer* output = net->AddOutputLayer(0);
 135
 136     input->GetOutputSlot(0).Connect(meanLayer->GetInputSlot(0));
 137     meanLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0));
 138
 139     input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 4, 3, 2 }, armnn::DataType::Float32));
 140     meanLayer->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 2 }, armnn::DataType::Float32));
 141
 142     // optimize the network
 143     armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
 144
 145     for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph())
 146     {
 147         BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef);
 148     }
 149 }
 150
 151 BOOST_AUTO_TEST_CASE(FP16TurboModeTestOnCpuRef)
 152 {
 153     // Test to check when FP16 Turbo mode set
 154     // it converts the FP32 network to FP16 Network
 155     // add FP32ToFP16 conversion layer after the InputLayer
 156     // add FP16ToFP32 conversion layer after the OutputLayer
 157     // checks the other layers if they are supported in FP16
 158     // if they are not put the conversion layers before and after
 159     // if they are not supported in FP16 use FP32 instead
 160     // if there are inverse conversion layers remove them with optimization
 161     // at the moment FloorLayer is not supported in FP16 so it rolls back to FP32
 162     // and inverse conversion layers are removed by the optimizer
 163     armnn::Network net;
 164
 165     // Defines layers.
 166     auto input = net.AddInputLayer(0);
 167     auto floor = net.AddFloorLayer();
 168     auto output = net.AddOutputLayer(0);
 169
 170     // Connects layers.
 171     input->GetOutputSlot(0).Connect(floor->GetInputSlot(0));
 172     floor->GetOutputSlot(0).Connect(output->GetInputSlot(0));
 173
 174     armnn::TensorShape shape({4});
 175     armnn::TensorInfo info(shape, armnn::DataType::Float32);
 176     input->GetOutputSlot(0).SetTensorInfo(info);
 177     floor->GetOutputSlot(0).SetTensorInfo(info);
 178
 179     armnn::IRuntime::CreationOptions options;
 180     armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
 181
 182     std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef};
 183
 184     armnn::OptimizerOptions optimizerOptions;
 185     optimizerOptions.m_ReduceFp32ToFp16 = true;
 186
 187     armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec(),
 188                                                                optimizerOptions);
 189
 190     std::ostringstream ss;
 191     optimizedNet->SerializeToDot(ss);
 192
 193     auto inputId = input->GetGuid();
 194     auto floorId = floor->GetGuid();
 195     auto outputId = output->GetGuid();
 196
 197     std::stringstream expected;
 198     expected <<
 199              "digraph Optimized {\n"
 200              "    node [shape=\"record\"];\n"
 201              "    edge [fontsize=8 fontcolor=\"blue\" fontname=\"arial-bold\"];\n"
 202              "    " << inputId << " [label=\"{Input}\"];\n"
 203              "    " << floorId << " [label=\"{Floor}\"];\n"
 204              "    " << outputId << " [label=\"{Output}\"];\n"
 205              "    " << inputId << " -> " << floorId << " [label=< [4] >];\n"
 206              "    " << floorId << " -> " << outputId << " [label=< [4] >];\n"
 207              "}\n";
 208
 209     BOOST_TEST(ss.str() == expected.str());
 210 }
 211
 212 BOOST_AUTO_TEST_SUITE_END()