src/backends/cl/test/ClFallbackTests.cpp

//
// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//

#include <backendsCommon/test/CommonTestUtils.hpp>

#include <test/GraphUtils.hpp>

#include <doctest/doctest.h>

TEST_SUITE("ClFallback")
{
TEST_CASE("ClImportEnabledFallbackToNeon")
{
    using namespace armnn;

    IRuntime::CreationOptions options;
    IRuntimePtr runtime(IRuntime::Create(options));

    // Builds up the structure of the network.
    INetworkPtr net(INetwork::Create());

    IConnectableLayer* input0 = net->AddInputLayer(0, "input0");
    IConnectableLayer* input1 = net->AddInputLayer(1, "input1");
    IConnectableLayer* input2 = net->AddInputLayer(2, "input2");
    IConnectableLayer* add = net->AddAdditionLayer("add");
    IConnectableLayer* sub = net->AddSubtractionLayer("sub");
    IConnectableLayer* output = net->AddOutputLayer(0, "output");

    input0->GetOutputSlot(0).Connect(add->GetInputSlot(0));
    input1->GetOutputSlot(0).Connect(add->GetInputSlot(1));
    input2->GetOutputSlot(0).Connect(sub->GetInputSlot(0));
    add->GetOutputSlot(0).Connect(sub->GetInputSlot(1));
    sub->GetOutputSlot(0).Connect(output->GetInputSlot(0));

    TensorInfo info = TensorInfo({ 1, 2, 4, 2 }, DataType::Float32);

    input0->GetOutputSlot(0).SetTensorInfo(info);
    input1->GetOutputSlot(0).SetTensorInfo(info);
    input2->GetOutputSlot(0).SetTensorInfo(info);
    add->GetOutputSlot(0).SetTensorInfo(info);
    sub->GetOutputSlot(0).SetTensorInfo(info);

    std::vector<BackendId> backends = { Compute::GpuAcc, Compute::CpuAcc };
    // Use BackendSelectionHint to specify CpuAcc for Subtraction layer
    sub->BackendSelectionHint(backends[1]);

    // optimize the network
    OptimizerOptions optOptions;
    optOptions.m_ImportEnabled = true;
    IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions);

    Graph& graph = GetGraphForTesting(optNet.get());

    armnn::Layer* const layer0 = GetFirstLayerWithName(graph, "input0");
    armnn::Layer* const layer1 = GetFirstLayerWithName(graph, "input1");
    armnn::Layer* const layer2 = GetFirstLayerWithName(graph, "input2");
    armnn::Layer* const layer3 = GetFirstLayerWithName(graph, "add");
    armnn::Layer* const layer4 = GetFirstLayerWithName(graph, "[ add (0) -> sub (1) ]");
    armnn::Layer* const layer5 = GetFirstLayerWithName(graph, "sub");
    armnn::Layer* const layer6 = GetFirstLayerWithName(graph, "output");

    // Checks order is valid.
    CHECK(CheckOrder(graph, layer0, layer1));
    CHECK(CheckOrder(graph, layer1, layer2));
    CHECK(CheckOrder(graph, layer2, layer3));
    CHECK(CheckOrder(graph, layer3, layer4));
    CHECK(CheckOrder(graph, layer4, layer5));
    CHECK(CheckOrder(graph, layer5, layer6));

    // Use memory import between backends
    CHECK((layer4->GetType() == LayerType::MemCopy));

    // Correctly use backend hint
    CHECK((layer5->GetBackendId() == Compute::CpuAcc ));

    // Load it into the runtime. It should pass.
    NetworkId netId;
    std::string ignoredErrorMessage;
    INetworkProperties networkProperties(false, MemorySource::Malloc, MemorySource::Malloc);
    runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties);

    // Creates structures for input & output
    std::vector<float> inputValue0
    {
        1.0f, 1.0f, 2.0f, 2.0f, 2.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f, 6.0f, 1.0f, 1.0f, 2.0f, 2.0f
    };
    std::vector<float> inputValue1
    {
        0.0f, 1.0f, 1.0f, 2.0f, 3.0f, 3.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f, 0.0f, 1.0f, 1.0f, 2.0f
    };
    std::vector<float> inputData2
    {
        12.0f, 11.0f, 10.0f, 9.0f, 8.0f, 7.0f, 6.0f, 5.0f, 4.0f, 3.0f, 2.0f, 1.0f, 12.0f, 11.0f, 10.0f, 9.0f
    };

    std::vector<float> outputData(16);

    std::vector<float> expectedOutput
    {
        11.0f, 9.0f, 7.0f, 5.0f, 3.0f, 1.0f, -1.0f, -3.0f, -5.0f, -7.0f, -9.0f, -11.0f, 11.0f, 9.0f, 7.0f, 5.0f
    };

    // Prepare aligned data
    unsigned int numElements = info.GetNumElements();
    size_t totalBytes = numElements * sizeof(float);
    const size_t alignment = 64;
    size_t space = totalBytes + alignment + alignment;
    auto inputData0 = std::make_unique<uint8_t[]>(space);
    void* alignedInputPtr0 = inputData0.get();
    CHECK(std::align(alignment, totalBytes, alignedInputPtr0, space));

    auto* intputPtr0 = reinterpret_cast<float*>(alignedInputPtr0);
    std::copy(inputValue0.begin(), inputValue0.end(), intputPtr0);

    auto inputData1 = std::make_unique<uint8_t[]>(space);
    void* alignedInputPtr1 = inputData1.get();
    CHECK(std::align(alignment, totalBytes, alignedInputPtr1, space));

    auto* intputPtr1 = reinterpret_cast<float*>(alignedInputPtr1);
    std::copy(inputValue1.begin(), inputValue1.end(), intputPtr1);

    InputTensors inputTensors
    {
        { 0, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), alignedInputPtr0) },
        { 1, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 1), alignedInputPtr1) },
        { 2, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 2), inputData2.data()) }
    };
    OutputTensors outputTensors
    {
        { 0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data()) }
    };

    runtime->GetProfiler(netId)->EnableProfiling(true);

    // Do the inference
    runtime->EnqueueWorkload(netId, inputTensors, outputTensors);

    // Retrieve the Profiler.Print() output to get the workload execution
    ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance();
    std::stringstream ss;
    profilerManager.GetProfiler()->Print(ss);;
    std::string dump = ss.str();

    // Executed Subtraction using CpuAcc
    std::size_t found = dump.find("NeonSubtractionWorkload_Execute");
    CHECK(found != std::string::npos);

    // Contain CopyMemGeneric
    found = dump.find("CopyMemGeneric");
    CHECK(found != std::string::npos);

    // Check output is as expected
    CHECK(outputData == expectedOutput);

    runtime->UnloadNetwork(netId);
}

TEST_CASE("ClImportDisabledFallbackToNeon")
{
    using namespace armnn;

    IRuntime::CreationOptions options;
    IRuntimePtr runtime(IRuntime::Create(options));

    // Builds up the structure of the network.
    INetworkPtr net(INetwork::Create());

    IConnectableLayer* input0 = net->AddInputLayer(0, "input0");
    IConnectableLayer* input1 = net->AddInputLayer(1, "input1");
    IConnectableLayer* input2 = net->AddInputLayer(2, "input2");
    IConnectableLayer* add = net->AddAdditionLayer("add");
    IConnectableLayer* sub = net->AddSubtractionLayer("sub");
    IConnectableLayer* output = net->AddOutputLayer(0, "output");

    input0->GetOutputSlot(0).Connect(add->GetInputSlot(0));
    input1->GetOutputSlot(0).Connect(add->GetInputSlot(1));
    input2->GetOutputSlot(0).Connect(sub->GetInputSlot(0));
    add->GetOutputSlot(0).Connect(sub->GetInputSlot(1));
    sub->GetOutputSlot(0).Connect(output->GetInputSlot(0));

    TensorInfo info = TensorInfo({ 1, 2, 3, 2 }, DataType::Float32);

    input0->GetOutputSlot(0).SetTensorInfo(info);
    input1->GetOutputSlot(0).SetTensorInfo(info);
    input2->GetOutputSlot(0).SetTensorInfo(info);
    add->GetOutputSlot(0).SetTensorInfo(info);
    sub->GetOutputSlot(0).SetTensorInfo(info);

    std::vector<BackendId> backends = { Compute::GpuAcc, Compute::CpuAcc };
    // Use BackendSelectionHint to specify CpuAcc for Subtraction layer
    sub->BackendSelectionHint(backends[1]);

    // optimize the network
    OptimizerOptions optOptions;
    IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions);

    Graph& graph = GetGraphForTesting(optNet.get());

    armnn::Layer* const layer0 = GetFirstLayerWithName(graph, "input0");
    armnn::Layer* const layer1 = GetFirstLayerWithName(graph, "input1");
    armnn::Layer* const layer2 = GetFirstLayerWithName(graph, "input2");
    armnn::Layer* const layer3 = GetFirstLayerWithName(graph, "add");
    armnn::Layer* const layer4 = GetFirstLayerWithName(graph, "[ add (0) -> sub (1) ]");
    armnn::Layer* const layer5 = GetFirstLayerWithName(graph, "sub");
    armnn::Layer* const layer6 = GetFirstLayerWithName(graph, "output");

    // Checks order is valid.
    CHECK(CheckOrder(graph, layer0, layer1));
    CHECK(CheckOrder(graph, layer1, layer2));
    CHECK(CheckOrder(graph, layer2, layer3));
    CHECK(CheckOrder(graph, layer3, layer4));
    CHECK(CheckOrder(graph, layer4, layer5));
    CHECK(CheckOrder(graph, layer5, layer6));

    // Use memory import between backends
    CHECK((layer4->GetType() == LayerType::MemCopy));

    // Correctly use backend hint
    CHECK((layer5->GetBackendId() == Compute::CpuAcc ));

    // Load it into the runtime. It should pass.
    NetworkId netId;
    runtime->LoadNetwork(netId, std::move(optNet));

    // Creates structures for input & output
    std::vector<float> inputData0
    {
        1.0f, 1.0f, 2.0f, 2.0f, 2.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f, 6.0f
    };
    std::vector<float> inputData1
    {
        0.0f, 1.0f, 1.0f, 2.0f, 3.0f, 3.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f
    };
    std::vector<float> inputData2
    {
        12.0f, 11.0f, 10.0f, 9.0f, 8.0f, 7.0f, 6.0f, 5.0f, 4.0f, 3.0f, 2.0f, 1.0f
    };

    std::vector<float> outputData(12);

    std::vector<float> expectedOutput
    {
        11.0f, 9.0f, 7.0f, 5.0f, 3.0f, 1.0f, -1.0f, -3.0f, -5.0f, -7.0f, -9.0f, -11.0f
    };

    InputTensors inputTensors
    {
        { 0, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), inputData0.data()) },
        { 1, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 1), inputData1.data()) },
        { 2, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 2), inputData2.data()) }
    };
    OutputTensors outputTensors
    {
        { 0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data()) }
    };

    runtime->GetProfiler(netId)->EnableProfiling(true);

    // Do the inference
    runtime->EnqueueWorkload(netId, inputTensors, outputTensors);

    // Retrieve the Profiler.Print() output to get the workload execution
    ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance();
    std::stringstream ss;
    profilerManager.GetProfiler()->Print(ss);;
    std::string dump = ss.str();

    // Executed Subtraction using CpuAcc
    std::size_t found = dump.find("NeonSubtractionWorkload_Execute");
    CHECK(found != std::string::npos);

    // Contain CopyMemGeneric
    found = dump.find("CopyMemGeneric");
    CHECK(found != std::string::npos);

    // Check output is as expected
    CHECK(outputData == expectedOutput);
}

TEST_CASE("ClImportEnabledFallbackSubgraphToNeon")
{
    using namespace armnn;

    IRuntime::CreationOptions options;
    IRuntimePtr runtime(IRuntime::Create(options));

    // Builds up the structure of the network.
    INetworkPtr net(INetwork::Create());

    Pooling2dDescriptor desc;
    desc.m_PoolWidth = 2;
    desc.m_PoolHeight = 2;
    desc.m_StrideX = 2;
    desc.m_StrideY = 2;

    IConnectableLayer* input0 = net->AddInputLayer(0, "input0");
    IConnectableLayer* input1 = net->AddInputLayer(1, "input1");
    IConnectableLayer* input2 = net->AddInputLayer(2, "input2");
    IConnectableLayer* add = net->AddAdditionLayer("add");
    IConnectableLayer* sub = net->AddSubtractionLayer("sub");
    IConnectableLayer* pooling = net->AddPooling2dLayer(desc, "pooling");
    IConnectableLayer* output = net->AddOutputLayer(0, "output");

    input0->GetOutputSlot(0).Connect(add->GetInputSlot(0));
    input1->GetOutputSlot(0).Connect(add->GetInputSlot(1));
    input2->GetOutputSlot(0).Connect(sub->GetInputSlot(0));
    add->GetOutputSlot(0).Connect(sub->GetInputSlot(1));
    sub->GetOutputSlot(0).Connect(pooling->GetInputSlot(0));
    pooling->GetOutputSlot(0).Connect(output->GetInputSlot(0));

    TensorInfo info = TensorInfo({ 1, 2, 4, 2 }, DataType::Float32);
    TensorInfo poolingInfo = TensorInfo({ 1, 2, 2, 1 }, DataType::Float32);

    input0->GetOutputSlot(0).SetTensorInfo(info);
    input1->GetOutputSlot(0).SetTensorInfo(info);
    input2->GetOutputSlot(0).SetTensorInfo(info);
    add->GetOutputSlot(0).SetTensorInfo(info);
    sub->GetOutputSlot(0).SetTensorInfo(info);
    pooling->GetOutputSlot(0).SetTensorInfo(poolingInfo);

    std::vector<BackendId> backends = { Compute::GpuAcc, Compute::CpuAcc };
    // Use BackendSelectionHint to specify CpuAcc for Subtraction layer
    sub->BackendSelectionHint(backends[1]);

    // optimize the network
    OptimizerOptions optOptions;
    optOptions.m_ImportEnabled = true;
    IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions);

    Graph& graph = GetGraphForTesting(optNet.get());

    armnn::Layer* const layer0 = GetFirstLayerWithName(graph, "input0");
    armnn::Layer* const layer1 = GetFirstLayerWithName(graph, "input1");
    armnn::Layer* const layer2 = GetFirstLayerWithName(graph, "input2");
    armnn::Layer* const layer3 = GetFirstLayerWithName(graph, "add");
    armnn::Layer* const layer4 = GetFirstLayerWithName(graph, "[ add (0) -> sub (1) ]");
    armnn::Layer* const layer5 = GetFirstLayerWithName(graph, "sub");
    armnn::Layer* const layer6 = GetFirstLayerWithName(graph, "[ sub (0) -> pooling (0) ]");
    armnn::Layer* const layer7 = GetFirstLayerWithName(graph, "pooling");
    armnn::Layer* const layer8 = GetFirstLayerWithName(graph, "output");

    // Checks order is valid.
    CHECK(CheckOrder(graph, layer0, layer1));
    CHECK(CheckOrder(graph, layer1, layer2));
    CHECK(CheckOrder(graph, layer2, layer3));
    CHECK(CheckOrder(graph, layer3, layer4));
    CHECK(CheckOrder(graph, layer4, layer5));
    CHECK(CheckOrder(graph, layer5, layer6));
    CHECK(CheckOrder(graph, layer6, layer7));
    CHECK(CheckOrder(graph, layer7, layer8));

    // Use memory import between backends
    CHECK((layer4->GetType() == LayerType::MemCopy));
    CHECK((layer6->GetType() == LayerType::MemCopy));

    // Correctly use backend hint
    CHECK((layer5->GetBackendId() == Compute::CpuAcc ));

    // Load it into the runtime. It should pass.
    NetworkId netId;
    std::string ignoredErrorMessage;
    INetworkProperties networkProperties(false, MemorySource::Malloc, MemorySource::Malloc);
    runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties);

    // Creates structures for input & output
    std::vector<float> inputValue0
    {
        1.0f, 1.0f, 2.0f, 2.0f, 2.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f, 6.0f, 1.0f, 1.0f, 2.0f, 2.0f
    };
    std::vector<float> inputValue1
    {
        0.0f, 1.0f, 1.0f, 2.0f, 3.0f, 3.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f, 0.0f, 1.0f, 1.0f, 2.0f
    };
    std::vector<float> inputData2
    {
        12.0f, 11.0f, 10.0f, 9.0f, 8.0f, 7.0f, 6.0f, 5.0f, 4.0f, 3.0f, 2.0f, 1.0f, 12.0f, 11.0f, 10.0f, 9.0f
    };

    std::vector<float> outputData(4);

    std::vector<float> expectedOutput{ 11.0f, 3.0f, -5.0f, 11.0f };

    unsigned int numElements = info.GetNumElements();
    size_t totalBytes = numElements * sizeof(float);
    const size_t alignment = 64;
    size_t space = totalBytes + alignment + alignment;
    auto inputData0 = std::make_unique<uint8_t[]>(space);
    void* alignedInputPtr0 = inputData0.get();
    CHECK(std::align(alignment, totalBytes, alignedInputPtr0, space));

    auto* intputPtr0 = reinterpret_cast<float*>(alignedInputPtr0);
    std::copy(inputValue0.begin(), inputValue0.end(), intputPtr0);

    auto inputData1 = std::make_unique<uint8_t[]>(space);
    void* alignedInputPtr1 = inputData1.get();
    CHECK(std::align(alignment, totalBytes, alignedInputPtr1, space));

    auto* intputPtr1 = reinterpret_cast<float*>(alignedInputPtr1);
    std::copy(inputValue1.begin(), inputValue1.end(), intputPtr1);

    InputTensors inputTensors
    {
        { 0, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), alignedInputPtr0) },
        { 1, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 1), alignedInputPtr1) },
        { 2, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 2), inputData2.data()) }
    };
    OutputTensors outputTensors
    {
        { 0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data()) }
    };

    runtime->GetProfiler(netId)->EnableProfiling(true);

    // Do the inference
    runtime->EnqueueWorkload(netId, inputTensors, outputTensors);

    // Retrieve the Profiler.Print() output to get the workload execution
    ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance();
    std::stringstream ss;
    profilerManager.GetProfiler()->Print(ss);;
    std::string dump = ss.str();

    // Executed Subtraction using CpuAcc
    std::size_t found = dump.find("NeonSubtractionWorkload_Execute");
    CHECK(found != std::string::npos);

    // Correctly switch back to GpuAcc
    found = dump.find("ClPooling2dWorkload_Execute");
    CHECK(found != std::string::npos);

    // Contain CopyMemGeneric
    found = dump.find("CopyMemGeneric");
    CHECK(found != std::string::npos);

    // Check output is as expected
    CHECK(outputData == expectedOutput);

    runtime->UnloadNetwork(netId);
}

TEST_CASE("ClImportDisableFallbackSubgraphToNeon")
{
    using namespace armnn;

    IRuntime::CreationOptions options;
    IRuntimePtr runtime(IRuntime::Create(options));

    // Builds up the structure of the network.
    INetworkPtr net(INetwork::Create());

    Pooling2dDescriptor desc;

    IConnectableLayer* input0 = net->AddInputLayer(0, "input0");
    IConnectableLayer* input1 = net->AddInputLayer(1, "input1");
    IConnectableLayer* input2 = net->AddInputLayer(2, "input2");
    IConnectableLayer* add = net->AddAdditionLayer("add");
    IConnectableLayer* sub = net->AddSubtractionLayer("sub");
    IConnectableLayer* pooling = net->AddPooling2dLayer(desc, "pooling");
    IConnectableLayer* output = net->AddOutputLayer(0, "output");

    input0->GetOutputSlot(0).Connect(add->GetInputSlot(0));
    input1->GetOutputSlot(0).Connect(add->GetInputSlot(1));
    input2->GetOutputSlot(0).Connect(sub->GetInputSlot(0));
    add->GetOutputSlot(0).Connect(sub->GetInputSlot(1));
    sub->GetOutputSlot(0).Connect(pooling->GetInputSlot(0));
    pooling->GetOutputSlot(0).Connect(output->GetInputSlot(0));

    TensorInfo info = TensorInfo({ 1, 2, 3, 2 }, DataType::Float32);
    TensorInfo poolingInfo = TensorInfo({ 1, 2, 1, 1 }, DataType::Float32);

    input0->GetOutputSlot(0).SetTensorInfo(info);
    input1->GetOutputSlot(0).SetTensorInfo(info);
    input2->GetOutputSlot(0).SetTensorInfo(info);
    add->GetOutputSlot(0).SetTensorInfo(info);
    sub->GetOutputSlot(0).SetTensorInfo(info);
    pooling->GetOutputSlot(0).SetTensorInfo(poolingInfo);

    std::vector<BackendId> backends = { Compute::GpuAcc, Compute::CpuAcc };
    // Use BackendSelectionHint to specify CpuAcc for Subtraction layer
    sub->BackendSelectionHint(backends[1]);

    // optimize the network
    OptimizerOptions optOptions;
    IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions);

    Graph& graph = GetGraphForTesting(optNet.get());

    armnn::Layer* const layer0 = GetFirstLayerWithName(graph, "input0");
    armnn::Layer* const layer1 = GetFirstLayerWithName(graph, "input1");
    armnn::Layer* const layer2 = GetFirstLayerWithName(graph, "input2");
    armnn::Layer* const layer3 = GetFirstLayerWithName(graph, "add");
    armnn::Layer* const layer4 = GetFirstLayerWithName(graph, "[ add (0) -> sub (1) ]");
    armnn::Layer* const layer5 = GetFirstLayerWithName(graph, "sub");
    armnn::Layer* const layer6 = GetFirstLayerWithName(graph, "[ sub (0) -> pooling (0) ]");
    armnn::Layer* const layer7 = GetFirstLayerWithName(graph, "pooling");
    armnn::Layer* const layer8 = GetFirstLayerWithName(graph, "output");

    // Checks order is valid.
    CHECK(CheckOrder(graph, layer0, layer1));
    CHECK(CheckOrder(graph, layer1, layer2));
    CHECK(CheckOrder(graph, layer2, layer3));
    CHECK(CheckOrder(graph, layer3, layer4));
    CHECK(CheckOrder(graph, layer4, layer5));
    CHECK(CheckOrder(graph, layer5, layer6));
    CHECK(CheckOrder(graph, layer6, layer7));
    CHECK(CheckOrder(graph, layer7, layer8));

    // Use memory import between backends
    CHECK((layer4->GetType() == LayerType::MemCopy));
    CHECK((layer6->GetType() == LayerType::MemCopy));

    // Correctly use backend hint
    CHECK((layer5->GetBackendId() == Compute::CpuAcc ));

    // Load it into the runtime. It should pass.
    NetworkId netId;
    runtime->LoadNetwork(netId, std::move(optNet));

    // Creates structures for input & output
    std::vector<float> inputData0
    {
        1.0f, 1.0f, 2.0f, 2.0f, 2.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f, 6.0f
    };
    std::vector<float> inputData1
    {
        0.0f, 1.0f, 1.0f, 2.0f, 3.0f, 3.0f, 3.0f, 4.0f, 4.0f, 5.0f, 5.0f, 6.0f
    };
    std::vector<float> inputData2
    {
        12.0f, 11.0f, 10.0f, 9.0f, 8.0f, 7.0f, 6.0f, 5.0f, 4.0f, 3.0f, 2.0f, 1.0f
    };

    std::vector<float> outputData(2);

    std::vector<float> expectedOutput{ 11.0f, -1.0f };

    InputTensors inputTensors
    {
        { 0, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 0), inputData0.data()) },
        { 1, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 1), inputData1.data()) },
        { 2, armnn::ConstTensor(runtime->GetInputTensorInfo(netId, 2), inputData2.data()) }
    };
    OutputTensors outputTensors
    {
        { 0,armnn::Tensor(runtime->GetOutputTensorInfo(netId, 0), outputData.data()) }
    };

    runtime->GetProfiler(netId)->EnableProfiling(true);

    // Do the inference
    runtime->EnqueueWorkload(netId, inputTensors, outputTensors);

    // Retrieve the Profiler.Print() output to get the workload execution
    ProfilerManager& profilerManager = armnn::ProfilerManager::GetInstance();
    std::stringstream ss;
    profilerManager.GetProfiler()->Print(ss);;
    std::string dump = ss.str();

    // Executed Subtraction using CpuAcc
    std::size_t found = dump.find("NeonSubtractionWorkload_Execute");
    CHECK(found != std::string::npos);

    // Correctly switch back to GpuAcc
    found = dump.find("ClPooling2dWorkload_Execute");
    CHECK(found != std::string::npos);

    // Contain CopyMemGeneric
    found = dump.find("CopyMemGeneric");
    CHECK(found != std::string::npos);

    // Check output is as expected
    CHECK(outputData == expectedOutput);
}

}