IVGCVSW-5322 Fix segfault between Neon and Cl layers

 * Fallback to memory copy if memory import is not supported
 * Remove direct compatibility between Neon and Cl Tensors
 * Unit tests fallback from Neon to Cl and Cl to Neon

Signed-off-by: Narumol Prangnawarat <narumol.prangnawarat@arm.com>
Change-Id: Iec00a77423fb23b37a6b1aefee1b2ec4d649efca

7 files changed, 1088 insertions, 23 deletions

diff --git a/src/armnn/LoadedNetwork.cpp b/src/armnn/LoadedNetwork.cpp
index 00ac90b121..b5a1b392b4 100644
--- a/src/armnn/LoadedNetwork.cpp
+++ b/src/armnn/LoadedNetwork.cpp
@@ -570,10 +570,12 @@ void LoadedNetwork::EnqueueInput(const BindableLayer& layer, ITensorHandle* tens
     info.m_OutputTensorInfos.push_back(outputTensorInfo);
 
     MemorySourceFlags importFlags = outputTensorHandle->GetImportFlags();
+    bool needMemCopy = true;
     if (m_IsImportEnabled)  // Try import the input tensor
     {
         if(CheckFlag(importFlags, MemorySource::Malloc) )
         {
+            needMemCopy = false;
             // This assumes a CPU Tensor handle
             void* mem = tensorHandle->Map(false);
             if (outputTensorHandle->Import(mem, MemorySource::Malloc))
@@ -584,12 +586,8 @@ void LoadedNetwork::EnqueueInput(const BindableLayer& layer, ITensorHandle* tens
             tensorHandle->Unmap();
             throw MemoryImportException("EnqueueInput: Memory Import failed");
         }
-        else
-        {
-            throw MemoryImportException("EnqueueInput: Memory Import failed, backend does not support Import");
-        }
     }
-    else
+    if (needMemCopy)
     {
         // Create a mem copy workload for input since we did not import
         std::unique_ptr<IWorkload> inputWorkload = std::make_unique<CopyMemGenericWorkload>(inputQueueDescriptor, info);
@@ -643,6 +641,7 @@ void LoadedNetwork::EnqueueOutput(const BindableLayer& layer, ITensorHandle* ten
     // c) There is only one connection to the OutputSlot and it is to an OutputLayer.
     // d) The output pointer is allocated via malloc. (Other types will be supported in a later release)
     // e) m_IsExportEnabled must be set to true
+    bool needMemCopy = true;
     if (m_IsExportEnabled && (layer.GetInputSlots()[0].GetConnectedOutputSlot()->GetNumConnections() == 1))
     {
         if(layer.GetInputSlots()[0].GetConnectedOutputSlot()->GetOwningLayer().GetType() != LayerType::Input)
@@ -650,6 +649,7 @@ void LoadedNetwork::EnqueueOutput(const BindableLayer& layer, ITensorHandle* ten
             MemorySourceFlags importFlags = inputTensorHandle->GetImportFlags();
             if (CheckFlag(importFlags, MemorySource::Malloc))
             {
+                needMemCopy = false;
                 void *mem = tensorHandle->Map(false);
                 bool importOk = inputTensorHandle->Import(mem, MemorySource::Malloc);
                 tensorHandle->Unmap();
@@ -669,17 +669,9 @@ void LoadedNetwork::EnqueueOutput(const BindableLayer& layer, ITensorHandle* ten
                     throw MemoryExportException("EnqueueOutput: Memory Export failed");
                 }
             }
-            else
-            {
-                throw MemoryExportException("EnqueueOutput: Memory Export failed, backend does not support Export");
-            }
-        }
-        else
-        {
-            throw MemoryExportException("EnqueueOutput: Memory Export failed, attempting to export Input Layer");
         }
     }
-    else
+    if (needMemCopy)
     {
         const Layer& connectedLayer = layer.GetInputSlots()[0].GetConnectedOutputSlot()->GetOwningLayer();
         // Do not add MemCopy Layer if OutputLayer is already connected the MemCopy Layer
diff --git a/src/backends/backendsCommon/test/CompatibilityTests.cpp b/src/backends/backendsCommon/test/CompatibilityTests.cpp
index 90aa76e3f3..b69e11253d 100644
--- a/src/backends/backendsCommon/test/CompatibilityTests.cpp
+++ b/src/backends/backendsCommon/test/CompatibilityTests.cpp
@@ -15,7 +15,7 @@
 
 using namespace armnn;
 
-BOOST_AUTO_TEST_SUITE(BackendsCompatibility)
+BOOST_AUTO_TEST_SUITE(BackendsCompatibility, * boost::unit_test::disabled())
 
 BOOST_AUTO_TEST_CASE(Neon_Cl_DirectCompatibility_Test)
 {
diff --git a/src/backends/cl/ClTensorHandleFactory.cpp b/src/backends/cl/ClTensorHandleFactory.cpp
index 33995f7b34..237f27a4ed 100644
--- a/src/backends/cl/ClTensorHandleFactory.cpp
+++ b/src/backends/cl/ClTensorHandleFactory.cpp
@@ -73,10 +73,11 @@ std::unique_ptr<ITensorHandle> ClTensorHandleFactory::CreateTensorHandle(const T
                                                                          const bool IsMemoryManaged) const
 {
     std::unique_ptr<ClTensorHandle> tensorHandle = std::make_unique<ClTensorHandle>(tensorInfo);
-    if (IsMemoryManaged)
+    if (!IsMemoryManaged)
     {
-        tensorHandle->SetMemoryGroup(m_MemoryManager->GetInterLayerMemoryGroup());
+        ARMNN_LOG(warning) << "ClTensorHandleFactory only has support for memory managed.";
     }
+    tensorHandle->SetMemoryGroup(m_MemoryManager->GetInterLayerMemoryGroup());
     return tensorHandle;
 }
 
@@ -85,10 +86,11 @@ std::unique_ptr<ITensorHandle> ClTensorHandleFactory::CreateTensorHandle(const T
                                                                          const bool IsMemoryManaged) const
 {
     std::unique_ptr<ClTensorHandle> tensorHandle = std::make_unique<ClTensorHandle>(tensorInfo, dataLayout);
-    if (IsMemoryManaged)
+    if (!IsMemoryManaged)
     {
-        tensorHandle->SetMemoryGroup(m_MemoryManager->GetInterLayerMemoryGroup());
+        ARMNN_LOG(warning) << "ClTensorHandleFactory only has support for memory managed.";
     }
+    tensorHandle->SetMemoryGroup(m_MemoryManager->GetInterLayerMemoryGroup());
     return tensorHandle;
 }
 
diff --git a/src/backends/cl/test/CMakeLists.txt b/src/backends/cl/test/CMakeLists.txt
index b0b330931d..2cf00d106b 100644
--- a/src/backends/cl/test/CMakeLists.txt
+++ b/src/backends/cl/test/CMakeLists.txt
@@ -23,6 +23,12 @@ if (ARMNNREF)
         )
 endif()
 
+if (ARMCOMPUTENEON)
+    list(APPEND armnnClBackendUnitTests_sources
+        ClFallbackTests.cpp
+        )
+endif()
+
 add_library(armnnClBackendUnitTests OBJECT ${armnnClBackendUnitTests_sources})
 target_include_directories(armnnClBackendUnitTests PRIVATE ${PROJECT_SOURCE_DIR}/src/armnn)
 target_include_directories(armnnClBackendUnitTests PRIVATE ${PROJECT_SOURCE_DIR}/src/armnnUtils)
diff --git a/src/backends/cl/test/ClFallbackTests.cpp b/src/backends/cl/test/ClFallbackTests.cpp
new file mode 100644
index 0000000000..5885cbe8ef
--- /dev/null
+++ b/src/backends/cl/test/ClFallbackTests.cpp
@@ -0,0 +1,538 @@
+//
+// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include <backendsCommon/test/CommonTestUtils.hpp>
+
+#include <test/GraphUtils.hpp>
+
+#include <boost/test/unit_test.hpp>
+
+BOOST_AUTO_TEST_SUITE(ClFallback)
+
+BOOST_AUTO_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, 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;
+    optOptions.m_ImportEnabled = true;
+    IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions);
+
+    OptimizedNetwork* optNetObjPtr = PolymorphicDowncast<OptimizedNetwork*>(optNet.get());
+    Graph& graph = optNetObjPtr->GetGraph();
+
+    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.
+    BOOST_TEST(CheckOrder(graph, layer0, layer1));
+    BOOST_TEST(CheckOrder(graph, layer1, layer2));
+    BOOST_TEST(CheckOrder(graph, layer2, layer3));
+    BOOST_TEST(CheckOrder(graph, layer3, layer4));
+    BOOST_TEST(CheckOrder(graph, layer4, layer5));
+    BOOST_TEST(CheckOrder(graph, layer5, layer6));
+
+    // Use memory import between backends
+    BOOST_TEST((layer4->GetType() == LayerType::MemCopy));
+
+    // Correctly use backend hint
+    BOOST_TEST((layer5->GetBackendId() == Compute::CpuAcc ));
+
+    // Load it into the runtime. It should pass.
+    NetworkId netId;
+    std::string ignoredErrorMessage;
+    INetworkProperties networkProperties(true, true);
+
+    runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties);
+
+    // 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");
+    BOOST_TEST(found != std::string::npos);
+
+    // Contain CopyMemGeneric
+    found = dump.find("CopyMemGeneric");
+    BOOST_TEST(found != std::string::npos);
+
+    // Check output is as expected
+    BOOST_TEST(outputData == expectedOutput);
+}
+
+BOOST_AUTO_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);
+
+    OptimizedNetwork* optNetObjPtr = PolymorphicDowncast<OptimizedNetwork*>(optNet.get());
+    Graph& graph = optNetObjPtr->GetGraph();
+
+    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.
+    BOOST_TEST(CheckOrder(graph, layer0, layer1));
+    BOOST_TEST(CheckOrder(graph, layer1, layer2));
+    BOOST_TEST(CheckOrder(graph, layer2, layer3));
+    BOOST_TEST(CheckOrder(graph, layer3, layer4));
+    BOOST_TEST(CheckOrder(graph, layer4, layer5));
+    BOOST_TEST(CheckOrder(graph, layer5, layer6));
+
+    // Use memory import between backends
+    BOOST_TEST((layer4->GetType() == LayerType::MemCopy));
+
+    // Correctly use backend hint
+    BOOST_TEST((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");
+    BOOST_TEST(found != std::string::npos);
+
+    // Contain CopyMemGeneric
+    found = dump.find("CopyMemGeneric");
+    BOOST_TEST(found != std::string::npos);
+
+    // Check output is as expected
+    BOOST_TEST(outputData == expectedOutput);
+}
+
+BOOST_AUTO_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;
+
+    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;
+    optOptions.m_ImportEnabled = true;
+    IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions);
+
+    OptimizedNetwork* optNetObjPtr = PolymorphicDowncast<OptimizedNetwork*>(optNet.get());
+    Graph& graph = optNetObjPtr->GetGraph();
+
+    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.
+    BOOST_TEST(CheckOrder(graph, layer0, layer1));
+    BOOST_TEST(CheckOrder(graph, layer1, layer2));
+    BOOST_TEST(CheckOrder(graph, layer2, layer3));
+    BOOST_TEST(CheckOrder(graph, layer3, layer4));
+    BOOST_TEST(CheckOrder(graph, layer4, layer5));
+    BOOST_TEST(CheckOrder(graph, layer5, layer6));
+    BOOST_TEST(CheckOrder(graph, layer6, layer7));
+    BOOST_TEST(CheckOrder(graph, layer7, layer8));
+
+    // Use memory import between backends
+    BOOST_TEST((layer4->GetType() == LayerType::MemCopy));
+    BOOST_TEST((layer6->GetType() == LayerType::MemCopy));
+
+    // Correctly use backend hint
+    BOOST_TEST((layer5->GetBackendId() == Compute::CpuAcc ));
+
+    // Load it into the runtime. It should pass.
+    NetworkId netId;
+    std::string ignoredErrorMessage;
+    INetworkProperties networkProperties(true, true);
+
+    runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties);
+
+    // 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");
+    BOOST_TEST(found != std::string::npos);
+
+    // Correctly switch back to GpuAcc
+    found = dump.find("ClPooling2dWorkload_Execute");
+    BOOST_TEST(found != std::string::npos);
+
+    // Contain CopyMemGeneric
+    found = dump.find("CopyMemGeneric");
+    BOOST_TEST(found != std::string::npos);
+
+    // Check output is as expected
+    BOOST_TEST(outputData == expectedOutput);
+}
+
+BOOST_AUTO_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);
+
+    OptimizedNetwork* optNetObjPtr = PolymorphicDowncast<OptimizedNetwork*>(optNet.get());
+    Graph& graph = optNetObjPtr->GetGraph();
+
+    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.
+    BOOST_TEST(CheckOrder(graph, layer0, layer1));
+    BOOST_TEST(CheckOrder(graph, layer1, layer2));
+    BOOST_TEST(CheckOrder(graph, layer2, layer3));
+    BOOST_TEST(CheckOrder(graph, layer3, layer4));
+    BOOST_TEST(CheckOrder(graph, layer4, layer5));
+    BOOST_TEST(CheckOrder(graph, layer5, layer6));
+    BOOST_TEST(CheckOrder(graph, layer6, layer7));
+    BOOST_TEST(CheckOrder(graph, layer7, layer8));
+
+    // Use memory import between backends
+    BOOST_TEST((layer4->GetType() == LayerType::MemCopy));
+    BOOST_TEST((layer6->GetType() == LayerType::MemCopy));
+
+    // Correctly use backend hint
+    BOOST_TEST((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");
+    BOOST_TEST(found != std::string::npos);
+
+    // Correctly switch back to GpuAcc
+    found = dump.find("ClPooling2dWorkload_Execute");
+    BOOST_TEST(found != std::string::npos);
+
+    // Contain CopyMemGeneric
+    found = dump.find("CopyMemGeneric");
+    BOOST_TEST(found != std::string::npos);
+
+    // Check output is as expected
+    BOOST_TEST(outputData == expectedOutput);
+}
+
+BOOST_AUTO_TEST_SUITE_END()
diff --git a/src/backends/neon/NeonBackend.cpp b/src/backends/neon/NeonBackend.cpp
index 642c19dbe4..9862ddbd70 100644
--- a/src/backends/neon/NeonBackend.cpp
+++ b/src/backends/neon/NeonBackend.cpp
@@ -129,8 +129,7 @@ OptimizationViews NeonBackend::OptimizeSubgraphView(const SubgraphView& subgraph
 
 std::vector<ITensorHandleFactory::FactoryId> NeonBackend::GetHandleFactoryPreferences() const
 {
-    return std::vector<ITensorHandleFactory::FactoryId>() = {"Arm/Neon/TensorHandleFactory",
-                                                             "Arm/Cl/TensorHandleFactory"};
+    return std::vector<ITensorHandleFactory::FactoryId>() = { NeonTensorHandleFactory::GetIdStatic() };
 }
 
 void NeonBackend::RegisterTensorHandleFactories(class TensorHandleFactoryRegistry& registry)
diff --git a/src/backends/neon/test/NeonFallbackTests.cpp b/src/backends/neon/test/NeonFallbackTests.cpp
index 9a07ed236f..fd7fbbc4d5 100644
--- a/src/backends/neon/test/NeonFallbackTests.cpp
+++ b/src/backends/neon/test/NeonFallbackTests.cpp
@@ -12,8 +12,6 @@
 
 BOOST_AUTO_TEST_SUITE(NeonFallback)
 
-std::vector<armnn::BackendId> defaultBackends = { armnn::Compute::CpuAcc };
-
 BOOST_AUTO_TEST_CASE(FallbackImportToCpuAcc)
 {
     using namespace armnn;
@@ -684,4 +682,534 @@ BOOST_AUTO_TEST_CASE(FallbackDisableImportFromCpuAcc)
     BOOST_TEST(outputData == expectedOutput);
 }
 
+#if defined(ARMCOMPUTECL_ENABLED)
+BOOST_AUTO_TEST_CASE(NeonImportEnabledFallbackToCl)
+{
+    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::CpuAcc, Compute::GpuAcc };
+    // Use BackendSelectionHint to specify GpuAcc for Subtraction layer
+    sub->BackendSelectionHint(backends[1]);
+
+    // optimize the network
+    OptimizerOptions optOptions;
+    optOptions.m_ImportEnabled = true;
+    IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions);
+
+    OptimizedNetwork* optNetObjPtr = PolymorphicDowncast<OptimizedNetwork*>(optNet.get());
+    Graph& graph = optNetObjPtr->GetGraph();
+
+    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.
+    BOOST_TEST(CheckOrder(graph, layer0, layer1));
+    BOOST_TEST(CheckOrder(graph, layer1, layer2));
+    BOOST_TEST(CheckOrder(graph, layer2, layer3));
+    BOOST_TEST(CheckOrder(graph, layer3, layer4));
+    BOOST_TEST(CheckOrder(graph, layer4, layer5));
+    BOOST_TEST(CheckOrder(graph, layer5, layer6));
+
+    // Use memory import between backends
+    BOOST_TEST((layer4->GetType() == LayerType::MemCopy));
+
+    // Correctly use backend hint
+    BOOST_TEST((layer5->GetBackendId() == Compute::GpuAcc ));
+
+    // Load it into the runtime. It should pass.
+    NetworkId netId;
+    std::string ignoredErrorMessage;
+    INetworkProperties networkProperties(true, true);
+
+    runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties);
+
+    // 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 GpuAcc
+    std::size_t found = dump.find("ClSubtractionWorkload_Execute");
+    BOOST_TEST(found != std::string::npos);
+
+    // Contain CopyMemGeneric
+    found = dump.find("CopyMemGeneric");
+    BOOST_TEST(found != std::string::npos);
+
+    // Check output is as expected
+    BOOST_TEST(outputData == expectedOutput);
+}
+
+BOOST_AUTO_TEST_CASE(NeonImportDisabledFallbackToCl)
+{
+    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::CpuAcc, Compute::GpuAcc };
+    // Use BackendSelectionHint to specify GpuAcc for Subtraction layer
+    sub->BackendSelectionHint(backends[1]);
+
+    // optimize the network
+    OptimizerOptions optOptions;
+    IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions);
+
+    OptimizedNetwork* optNetObjPtr = PolymorphicDowncast<OptimizedNetwork*>(optNet.get());
+    Graph& graph = optNetObjPtr->GetGraph();
+
+    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.
+    BOOST_TEST(CheckOrder(graph, layer0, layer1));
+    BOOST_TEST(CheckOrder(graph, layer1, layer2));
+    BOOST_TEST(CheckOrder(graph, layer2, layer3));
+    BOOST_TEST(CheckOrder(graph, layer3, layer4));
+    BOOST_TEST(CheckOrder(graph, layer4, layer5));
+    BOOST_TEST(CheckOrder(graph, layer5, layer6));
+
+    // Use memory import between backends
+    BOOST_TEST((layer4->GetType() == LayerType::MemCopy));
+
+    // Correctly use backend hint
+    BOOST_TEST((layer5->GetBackendId() == Compute::GpuAcc ));
+
+    // 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 GpuAcc
+    std::size_t found = dump.find("ClSubtractionWorkload_Execute");
+    BOOST_TEST(found != std::string::npos);
+
+    // Contain CopyMemGeneric
+    found = dump.find("CopyMemGeneric");
+    BOOST_TEST(found != std::string::npos);
+
+    // Check output is as expected
+    BOOST_TEST(outputData == expectedOutput);
+}
+
+BOOST_AUTO_TEST_CASE(NeonImportEnabledFallbackSubgraphToCl)
+{
+    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::CpuAcc, Compute::GpuAcc };
+    // Use BackendSelectionHint to specify GpuAcc for Subtraction layer
+    sub->BackendSelectionHint(backends[1]);
+
+    // optimize the network
+    OptimizerOptions optOptions;
+    optOptions.m_ImportEnabled = true;
+    IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions);
+
+    OptimizedNetwork* optNetObjPtr = PolymorphicDowncast<OptimizedNetwork*>(optNet.get());
+    Graph& graph = optNetObjPtr->GetGraph();
+
+    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.
+    BOOST_TEST(CheckOrder(graph, layer0, layer1));
+    BOOST_TEST(CheckOrder(graph, layer1, layer2));
+    BOOST_TEST(CheckOrder(graph, layer2, layer3));
+    BOOST_TEST(CheckOrder(graph, layer3, layer4));
+    BOOST_TEST(CheckOrder(graph, layer4, layer5));
+    BOOST_TEST(CheckOrder(graph, layer5, layer6));
+    BOOST_TEST(CheckOrder(graph, layer6, layer7));
+    BOOST_TEST(CheckOrder(graph, layer7, layer8));
+
+    // Use memory import between backends
+    BOOST_TEST((layer4->GetType() == LayerType::MemCopy));
+    BOOST_TEST((layer6->GetType() == LayerType::MemCopy));
+
+    // Correctly use backend hint
+    BOOST_TEST((layer5->GetBackendId() == Compute::GpuAcc ));
+
+    // Load it into the runtime. It should pass.
+    NetworkId netId;
+    std::string ignoredErrorMessage;
+    INetworkProperties networkProperties(true, true);
+
+    runtime->LoadNetwork(netId, std::move(optNet), ignoredErrorMessage, networkProperties);
+
+    // 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 GpuAcc
+    std::size_t found = dump.find("ClSubtractionWorkload_Execute");
+    BOOST_TEST(found != std::string::npos);
+
+    // Correctly switch back to CpuAcc
+    found = dump.find("NeonPooling2dWorkload_Execute");
+    BOOST_TEST(found != std::string::npos);
+
+    // Contain CopyMemGeneric
+    found = dump.find("CopyMemGeneric");
+    BOOST_TEST(found != std::string::npos);
+
+    // Contains SyncMemGeneric for output
+    found = dump.find("SyncMemGeneric");
+    BOOST_TEST(found != std::string::npos);
+
+    // Check output is as expected
+    BOOST_TEST(outputData == expectedOutput);
+}
+
+BOOST_AUTO_TEST_CASE(NeonImportDisableFallbackSubgraphToCl)
+{
+    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::CpuAcc, Compute::GpuAcc };
+    // Use BackendSelectionHint to specify GpuAcc for Subtraction layer
+    sub->BackendSelectionHint(backends[1]);
+
+    // optimize the network
+    OptimizerOptions optOptions;
+    IOptimizedNetworkPtr optNet = Optimize(*net, backends, runtime->GetDeviceSpec(), optOptions);
+
+    OptimizedNetwork* optNetObjPtr = PolymorphicDowncast<OptimizedNetwork*>(optNet.get());
+    Graph& graph = optNetObjPtr->GetGraph();
+
+    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.
+    BOOST_TEST(CheckOrder(graph, layer0, layer1));
+    BOOST_TEST(CheckOrder(graph, layer1, layer2));
+    BOOST_TEST(CheckOrder(graph, layer2, layer3));
+    BOOST_TEST(CheckOrder(graph, layer3, layer4));
+    BOOST_TEST(CheckOrder(graph, layer4, layer5));
+    BOOST_TEST(CheckOrder(graph, layer5, layer6));
+    BOOST_TEST(CheckOrder(graph, layer6, layer7));
+    BOOST_TEST(CheckOrder(graph, layer7, layer8));
+
+    // Use memory import between backends
+    BOOST_TEST((layer4->GetType() == LayerType::MemCopy));
+    BOOST_TEST((layer6->GetType() == LayerType::MemCopy));
+
+    // Correctly use backend hint
+    BOOST_TEST((layer5->GetBackendId() == Compute::GpuAcc ));
+
+    // 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 GpuAcc
+    std::size_t found = dump.find("ClSubtractionWorkload_Execute");
+    BOOST_TEST(found != std::string::npos);
+
+    // Correctly switch back to CpuAcc
+    found = dump.find("NeonPooling2dWorkload_Execute");
+    BOOST_TEST(found != std::string::npos);
+
+    // Contain CopyMemGeneric
+    found = dump.find("CopyMemGeneric");
+    BOOST_TEST(found != std::string::npos);
+
+    // Check output is as expected
+    BOOST_TEST(outputData == expectedOutput);
+}
+#endif
+
 BOOST_AUTO_TEST_SUITE_END()