diff options
Diffstat (limited to 'src/backends/reference')
| -rw-r--r-- | src/backends/reference/backend.mk | 3 | ||||
| -rw-r--r-- | src/backends/reference/test/CMakeLists.txt | 3 | ||||
| -rw-r--r-- | src/backends/reference/test/RefEndToEndTests.cpp | 251 | ||||
| -rw-r--r-- | src/backends/reference/test/RefJsonPrinterTests.cpp | 22 | ||||
| -rw-r--r-- | src/backends/reference/test/RefOptimizedNetworkTests.cpp | 212 |
5 files changed, 491 insertions, 0 deletions
diff --git a/src/backends/reference/backend.mk b/src/backends/reference/backend.mk index 455ab4618e..007efceb9b 100644 --- a/src/backends/reference/backend.mk +++ b/src/backends/reference/backend.mk @@ -65,6 +65,9 @@ BACKEND_SOURCES := \ BACKEND_TEST_SOURCES := \ test/RefCreateWorkloadTests.cpp \ + test/RefEndToEndTests.cpp \ + test/RefJsonPrinterTests.cpp \ test/RefLayerSupportTests.cpp \ test/RefLayerTests.cpp \ + test/RefOptimizedNetworkTests.cpp \ test/RefRuntimeTests.cpp diff --git a/src/backends/reference/test/CMakeLists.txt b/src/backends/reference/test/CMakeLists.txt index dea0ef6498..1eec594aa9 100644 --- a/src/backends/reference/test/CMakeLists.txt +++ b/src/backends/reference/test/CMakeLists.txt @@ -5,8 +5,11 @@ list(APPEND armnnRefBackendUnitTests_sources RefCreateWorkloadTests.cpp + RefEndToEndTests.cpp + RefJsonPrinterTests.cpp RefLayerSupportTests.cpp RefLayerTests.cpp + RefOptimizedNetworkTests.cpp RefRuntimeTests.cpp ) diff --git a/src/backends/reference/test/RefEndToEndTests.cpp b/src/backends/reference/test/RefEndToEndTests.cpp new file mode 100644 index 0000000000..8938d6f222 --- /dev/null +++ b/src/backends/reference/test/RefEndToEndTests.cpp @@ -0,0 +1,251 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include <backends/test/EndToEndTestImpl.hpp> + +#include <boost/test/unit_test.hpp> + +BOOST_AUTO_TEST_SUITE(RefEndToEnd) + +BOOST_AUTO_TEST_CASE(ConstantUsage_Ref_Float32) +{ + std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef}; + BOOST_TEST(ConstantUsageFloat32Test(backends)); +} + +BOOST_AUTO_TEST_CASE(ConstantUsage_Ref_Uint8) +{ + std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef}; + BOOST_TEST(ConstantUsageUint8Test(backends)); +} + +BOOST_AUTO_TEST_CASE(Unsigned8) +{ + using namespace armnn; + + // Create runtime in which test will run + armnn::IRuntime::CreationOptions options; + armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); + + // Builds up the structure of the network. + armnn::INetworkPtr net(INetwork::Create()); + + IConnectableLayer* input = net->AddInputLayer(0, "input"); + IConnectableLayer* softmax = net->AddSoftmaxLayer(SoftmaxDescriptor(), "softmax"); + IConnectableLayer* output = net->AddOutputLayer(0, "output"); + + input->GetOutputSlot(0).Connect(softmax->GetInputSlot(0)); + softmax->GetOutputSlot(0).Connect(output->GetInputSlot(0)); + + // Sets the tensors in the network. + TensorInfo inputTensorInfo(TensorShape({1, 5}), DataType::QuantisedAsymm8); + inputTensorInfo.SetQuantizationOffset(100); + inputTensorInfo.SetQuantizationScale(10000.0f); + input->GetOutputSlot(0).SetTensorInfo(inputTensorInfo); + + TensorInfo outputTensorInfo(TensorShape({1, 5}), DataType::QuantisedAsymm8); + outputTensorInfo.SetQuantizationOffset(0); + outputTensorInfo.SetQuantizationScale(1.0f/255.0f); + softmax->GetOutputSlot(0).SetTensorInfo(outputTensorInfo); + + // optimize the network + std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef}; + IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec()); + + // Loads it into the runtime. + NetworkId netId; + auto error = runtime->LoadNetwork(netId, std::move(optNet)); + BOOST_TEST(error == Status::Success); + + // Creates structures for input & output. + std::vector<uint8_t> inputData + { + 1, 10, 3, 200, 5 // Some inputs - one of which is sufficiently larger than the others to saturate softmax. + }; + std::vector<uint8_t> outputData(5); + + armnn::InputTensors inputTensors + { + {0, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), inputData.data())} + }; + armnn::OutputTensors outputTensors + { + {0, armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data())} + }; + + // Does the inference. + runtime->EnqueueWorkload(netId, inputTensors, outputTensors); + + // Checks the results. + BOOST_TEST(outputData[0] == 0); + BOOST_TEST(outputData[1] == 0); + BOOST_TEST(outputData[2] == 0); + BOOST_TEST(outputData[3] == 255); // softmax has been saturated. + BOOST_TEST(outputData[4] == 0); +} + +BOOST_AUTO_TEST_CASE(TrivialAdd) +{ + // This test was designed to match "AddTwo" in android nn/runtime/test/TestTrivialModel.cpp. + + using namespace armnn; + + // Create runtime in which test will run + armnn::IRuntime::CreationOptions options; + armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); + + // Builds up the structure of the network. + armnn::INetworkPtr net(INetwork::Create()); + + IConnectableLayer* input1 = net->AddInputLayer(0); + IConnectableLayer* input2 = net->AddInputLayer(1); + IConnectableLayer* add = net->AddAdditionLayer(); + IConnectableLayer* output = net->AddOutputLayer(0); + + input1->GetOutputSlot(0).Connect(add->GetInputSlot(0)); + input2->GetOutputSlot(0).Connect(add->GetInputSlot(1)); + add->GetOutputSlot(0).Connect(output->GetInputSlot(0)); + + // Sets the tensors in the network. + TensorInfo tensorInfo(TensorShape({3, 4}), DataType::Float32); + input1->GetOutputSlot(0).SetTensorInfo(tensorInfo); + input2->GetOutputSlot(0).SetTensorInfo(tensorInfo); + add->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + // optimize the network + std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef}; + IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec()); + + // Loads it into the runtime. + NetworkId netId; + runtime->LoadNetwork(netId, std::move(optNet)); + + // Creates structures for input & output - matching android nn test. + std::vector<float> input1Data + { + 1.f, 2.f, 3.f, 4.f, 5.f, 6.f, 7.f, 8.f, 9.f, 10.f, 11.f, 12.f + }; + std::vector<float> input2Data + { + 100.f, 200.f, 300.f, 400.f, 500.f, 600.f, 700.f, 800.f, 900.f, 1000.f, 1100.f, 1200.f + }; + std::vector<float> outputData(12); + + InputTensors inputTensors + { + {0,armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), input1Data.data())}, + {1,armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), input2Data.data())} + }; + OutputTensors outputTensors + { + {0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data())} + }; + + // Does the inference. + runtime->EnqueueWorkload(netId, inputTensors, outputTensors); + + // Checks the results + BOOST_TEST(outputData[0] == 101); + BOOST_TEST(outputData[1] == 202); + BOOST_TEST(outputData[2] == 303); + BOOST_TEST(outputData[3] == 404); + BOOST_TEST(outputData[4] == 505); + BOOST_TEST(outputData[5] == 606); + BOOST_TEST(outputData[6] == 707); + BOOST_TEST(outputData[7] == 808); + BOOST_TEST(outputData[8] == 909); + BOOST_TEST(outputData[9] == 1010); + BOOST_TEST(outputData[10] == 1111); + BOOST_TEST(outputData[11] == 1212); +} + +BOOST_AUTO_TEST_CASE(MultipleOutputs) +{ + using namespace armnn; + + // Create runtime in which test will run + armnn::IRuntime::CreationOptions options; + armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); + + // Builds up the structure of the network. + INetworkPtr net(INetwork::Create()); + + IConnectableLayer* input = net->AddInputLayer(0); + + // ReLu1 + ActivationDescriptor activation1Descriptor; + activation1Descriptor.m_Function = ActivationFunction::BoundedReLu; + activation1Descriptor.m_A = 1.f; + activation1Descriptor.m_B = -1.f; + IConnectableLayer* activation1 = net->AddActivationLayer(activation1Descriptor); + + // ReLu6 + ActivationDescriptor activation2Descriptor; + activation2Descriptor.m_Function = ActivationFunction::BoundedReLu; + activation2Descriptor.m_A = 6.0f; + IConnectableLayer* activation2 = net->AddActivationLayer(activation2Descriptor); + + // BoundedReLu(min=2, max=5) + ActivationDescriptor activation3Descriptor; + activation3Descriptor.m_Function = ActivationFunction::BoundedReLu; + activation3Descriptor.m_A = 5.0f; + activation3Descriptor.m_B = 2.0f; + IConnectableLayer* activation3 = net->AddActivationLayer(activation3Descriptor); + + IConnectableLayer* output1 = net->AddOutputLayer(0); + IConnectableLayer* output2 = net->AddOutputLayer(1); + IConnectableLayer* output3 = net->AddOutputLayer(2); + + input->GetOutputSlot(0).Connect(activation1->GetInputSlot(0)); + input->GetOutputSlot(0).Connect(activation2->GetInputSlot(0)); + input->GetOutputSlot(0).Connect(activation3->GetInputSlot(0)); + + activation1->GetOutputSlot(0).Connect(output1->GetInputSlot(0)); + activation2->GetOutputSlot(0).Connect(output2->GetInputSlot(0)); + activation3->GetOutputSlot(0).Connect(output3->GetInputSlot(0)); + + // Sets the tensors in the network. + TensorInfo tensorInfo(TensorShape({ 10 }), DataType::Float32); + input->GetOutputSlot(0).SetTensorInfo(tensorInfo); + activation1->GetOutputSlot(0).SetTensorInfo(tensorInfo); + activation2->GetOutputSlot(0).SetTensorInfo(tensorInfo); + activation3->GetOutputSlot(0).SetTensorInfo(tensorInfo); + + // optimize the network + std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef}; + IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec()); + + // Loads it into the runtime. + NetworkId netId; + runtime->LoadNetwork(netId, std::move(optNet)); + + // Creates structures for input & output. + const std::vector<float> inputData{ 3.f, 5.f, 2.f, 3.f, 7.f, 0.f, -2.f, -1.f, 3.f, 3.f }; + + std::vector<float> output1Data(inputData.size()); + std::vector<float> output2Data(inputData.size()); + std::vector<float> output3Data(inputData.size()); + + InputTensors inputTensors + { + {0,armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), inputData.data())} + }; + OutputTensors outputTensors + { + {0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), output1Data.data())}, + {1,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 1), output2Data.data())}, + {2,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 2), output3Data.data())} + }; + + // Does the inference. + runtime->EnqueueWorkload(netId, inputTensors, outputTensors); + + // Checks the results. + BOOST_TEST(output1Data == std::vector<float>({ 1.f, 1.f, 1.f, 1.f, 1.f, 0.f, -1.f, -1.f, 1.f, 1.f })); // ReLu1 + BOOST_TEST(output2Data == std::vector<float>({ 3.f, 5.f, 2.f, 3.f, 6.f, 0.f, 0.f, 0.f, 3.f, 3.f })); // ReLu6 + BOOST_TEST(output3Data == std::vector<float>({ 3.f, 5.f, 2.f, 3.f, 5.f, 2.f, 2.f, 2.f, 3.f, 3.f })); // [2, 5] +} + +BOOST_AUTO_TEST_SUITE_END()
\ No newline at end of file diff --git a/src/backends/reference/test/RefJsonPrinterTests.cpp b/src/backends/reference/test/RefJsonPrinterTests.cpp new file mode 100644 index 0000000000..ee668a2513 --- /dev/null +++ b/src/backends/reference/test/RefJsonPrinterTests.cpp @@ -0,0 +1,22 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include <armnn/BackendId.hpp> + +#include <backends/test/JsonPrinterTestImpl.hpp> + +#include <boost/test/unit_test.hpp> + +#include <vector> + +BOOST_AUTO_TEST_SUITE(RefJsonPrinter) + +BOOST_AUTO_TEST_CASE(SoftmaxProfilerJsonPrinterCpuRefTest) +{ + std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef}; + SetupSoftmaxProfilerWithSpecifiedBackendsAndValidateJsonPrinterResult(backends); +} + +BOOST_AUTO_TEST_SUITE_END()
\ No newline at end of file diff --git a/src/backends/reference/test/RefOptimizedNetworkTests.cpp b/src/backends/reference/test/RefOptimizedNetworkTests.cpp new file mode 100644 index 0000000000..63615e6859 --- /dev/null +++ b/src/backends/reference/test/RefOptimizedNetworkTests.cpp @@ -0,0 +1,212 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include <armnn/ArmNN.hpp> +#include <armnn/Graph.hpp> +#include <armnn/Network.hpp> + +#include <backends/reference/RefWorkloadFactory.hpp> + +#include <boost/test/unit_test.hpp> + +BOOST_AUTO_TEST_SUITE(RefOptimizedNetwork) + +BOOST_AUTO_TEST_CASE(OptimizeValidateCpuRefWorkloads) +{ + const armnn::TensorInfo desc({3, 5}, armnn::DataType::Float32); + + armnn::Network net; + + armnn::NormalizationDescriptor nmDesc; + armnn::ActivationDescriptor acDesc; + + // in + // | + // nm + // / | + // ac | + // \ | + // ml + // | + // sm + // | + // ot + armnn::IConnectableLayer* layer = net.AddInputLayer(0, "in"); + layer->GetOutputSlot(0).SetTensorInfo(desc); + + armnn::IConnectableLayer* const normLayer = net.AddNormalizationLayer(nmDesc, "nm"); + + layer->GetOutputSlot(0).Connect(normLayer->GetInputSlot(0)); + normLayer->GetOutputSlot(0).SetTensorInfo(desc); + + layer = net.AddActivationLayer(acDesc, "ac"); + + normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); + layer->GetOutputSlot(0).SetTensorInfo(desc); + + armnn::IConnectableLayer* prevLayer = layer; + layer = net.AddMultiplicationLayer("ml"); + + prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); + normLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(1)); + layer->GetOutputSlot(0).SetTensorInfo(desc); + + prevLayer = layer; + armnn::SoftmaxDescriptor softmaxDescriptor; + layer = net.AddSoftmaxLayer(softmaxDescriptor, "sm"); + + prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); + layer->GetOutputSlot(0).SetTensorInfo(desc); + + prevLayer = layer; + layer = net.AddOutputLayer(0, "ot"); + + prevLayer->GetOutputSlot(0).Connect(layer->GetInputSlot(0)); + + armnn::IRuntime::CreationOptions options; + armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); + + std::vector<armnn::BackendId> backends = { armnn::Compute::CpuRef }; + armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec()); + static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph().AllocateDynamicBuffers(); + BOOST_CHECK(optNet); + + // Validates workloads. + armnn::RefWorkloadFactory fact; + for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph()) + { + BOOST_CHECK_NO_THROW( + layer->CreateWorkload(static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph(), fact)); + } +} + +BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsCpuRefPermuteLayer) +{ + // Create runtime in which test will run + armnn::IRuntime::CreationOptions options; + armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); + + std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef}; + + // build up the structure of the network + armnn::INetworkPtr net(armnn::INetwork::Create()); + + armnn::IConnectableLayer* input = net->AddInputLayer(0); + + armnn::PermuteDescriptor descriptor({0, 2, 3, 1}); + armnn::IConnectableLayer* permute = net->AddPermuteLayer(descriptor); + + armnn::IConnectableLayer* output = net->AddOutputLayer(0); + + input->GetOutputSlot(0).Connect(permute->GetInputSlot(0)); + permute->GetOutputSlot(0).Connect(output->GetInputSlot(0)); + + input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 1, 4, 4 }, armnn::DataType::Float32)); + permute->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 1, 4, 1, 4 }, armnn::DataType::Float32)); + + // optimize the network + armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec()); + + for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph()) + { + BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef); + } +} + +BOOST_AUTO_TEST_CASE(OptimizeValidateWorkloadsCpuRefMeanLayer) +{ + // Create runtime in which test will run + armnn::IRuntime::CreationOptions options; + armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); + + std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef}; + + // build up the structure of the network + armnn::INetworkPtr net(armnn::INetwork::Create()); + + armnn::IConnectableLayer* input = net->AddInputLayer(0); + + armnn::MeanDescriptor descriptor({ 0, 1 }, false); + armnn::IConnectableLayer* meanLayer = net->AddMeanLayer(descriptor); + + armnn::IConnectableLayer* output = net->AddOutputLayer(0); + + input->GetOutputSlot(0).Connect(meanLayer->GetInputSlot(0)); + meanLayer->GetOutputSlot(0).Connect(output->GetInputSlot(0)); + + input->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 4, 3, 2 }, armnn::DataType::Float32)); + meanLayer->GetOutputSlot(0).SetTensorInfo(armnn::TensorInfo({ 2 }, armnn::DataType::Float32)); + + // optimize the network + armnn::IOptimizedNetworkPtr optNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec()); + + for (auto&& layer : static_cast<armnn::OptimizedNetwork*>(optNet.get())->GetGraph()) + { + BOOST_CHECK(layer->GetBackendId() == armnn::Compute::CpuRef); + } +} + +BOOST_AUTO_TEST_CASE(FP16TurboModeTestOnCpuRef) +{ + // Test to check when FP16 Turbo mode set + // it converts the FP32 network to FP16 Network + // add FP32ToFP16 conversion layer after the InputLayer + // add FP16ToFP32 conversion layer after the OutputLayer + // checks the other layers if they are supported in FP16 + // if they are not put the conversion layers before and after + // if they are not supported in FP16 use FP32 instead + // if there are inverse conversion layers remove them with optimization + // at the moment FloorLayer is not supported in FP16 so it rolls back to FP32 + // and inverse conversion layers are removed by the optimizer + armnn::Network net; + + // Defines layers. + auto input = net.AddInputLayer(0); + auto floor = net.AddFloorLayer(); + auto output = net.AddOutputLayer(0); + + // Connects layers. + input->GetOutputSlot(0).Connect(floor->GetInputSlot(0)); + floor->GetOutputSlot(0).Connect(output->GetInputSlot(0)); + + armnn::TensorShape shape({4}); + armnn::TensorInfo info(shape, armnn::DataType::Float32); + input->GetOutputSlot(0).SetTensorInfo(info); + floor->GetOutputSlot(0).SetTensorInfo(info); + + armnn::IRuntime::CreationOptions options; + armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options)); + + std::vector<armnn::BackendId> backends = {armnn::Compute::CpuRef}; + + armnn::OptimizerOptions optimizerOptions; + optimizerOptions.m_ReduceFp32ToFp16 = true; + + armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec(), + optimizerOptions); + + std::ostringstream ss; + optimizedNet->SerializeToDot(ss); + + auto inputId = input->GetGuid(); + auto floorId = floor->GetGuid(); + auto outputId = output->GetGuid(); + + std::stringstream expected; + expected << + "digraph Optimized {\n" + " node [shape=\"record\"];\n" + " edge [fontsize=8 fontcolor=\"blue\" fontname=\"arial-bold\"];\n" + " " << inputId << " [label=\"{Input}\"];\n" + " " << floorId << " [label=\"{Floor}\"];\n" + " " << outputId << " [label=\"{Output}\"];\n" + " " << inputId << " -> " << floorId << " [label=< [4] >];\n" + " " << floorId << " -> " << outputId << " [label=< [4] >];\n" + "}\n"; + + BOOST_TEST(ss.str() == expected.str()); +} + +BOOST_AUTO_TEST_SUITE_END() |