diff --git a/src/armnn/BackendSettings.hpp b/src/armnn/BackendSettings.hpp
new file mode 100644
index 0000000..931a068
--- /dev/null
+++ b/src/armnn/BackendSettings.hpp
@@ -0,0 +1,87 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <armnn/BackendId.hpp>
+#include <vector>
+
+namespace armnn
+{
+
+struct BackendSettings
+{
+    BackendIdVector m_PreferredBackends;
+    BackendIdSet    m_SupportedBackends;
+    BackendIdSet    m_SelectedBackends;
+    BackendIdSet    m_IgnoredBackends;
+
+    BackendSettings() = default;
+
+    BackendSettings(const BackendIdVector& preferredBackends,
+                    const IDeviceSpec& deviceSpec)
+    {
+        Initialize(preferredBackends, deviceSpec);
+    }
+
+    bool IsBackendPreferred(const BackendId& backend) const
+    {
+        return IsBackendInCollection(backend, m_PreferredBackends);
+    }
+
+    bool IsBackendSupported(const BackendId& backend) const
+    {
+        return IsBackendInCollection(backend, m_SupportedBackends);
+    }
+
+    bool IsBackendSelected(const BackendId& backend) const
+    {
+        return IsBackendInCollection(backend, m_SelectedBackends);
+    }
+
+    bool IsBackendIgnored(const BackendId& backend) const
+    {
+        return IsBackendInCollection(backend, m_IgnoredBackends);
+    }
+
+    bool IsCpuRefUsed() const
+    {
+        BackendId cpuBackendId(Compute::CpuRef);
+        return IsBackendSupported(cpuBackendId) && IsBackendPreferred(cpuBackendId);
+    }
+
+    BackendIdVector GetAvailablePreferredBackends() const
+    {
+        BackendIdVector availablePreferredBackends;
+        for (const BackendId& backend : m_PreferredBackends)
+        {
+            if (IsBackendSupported(backend) && !IsBackendIgnored(backend))
+            {
+                availablePreferredBackends.push_back(backend);
+            }
+        }
+        return availablePreferredBackends;
+    }
+
+private:
+    void Initialize(const BackendIdVector& preferredBackends,
+                    const IDeviceSpec& deviceSpec)
+    {
+        // Copy preferred backends from input
+        m_PreferredBackends = preferredBackends;
+
+        // Obtain list of supported backends
+        const DeviceSpec& spec = *boost::polymorphic_downcast<const DeviceSpec*>(&deviceSpec);
+        m_SupportedBackends = spec.GetSupportedBackends();
+    }
+
+    template<typename Collection>
+    bool IsBackendInCollection(const BackendId& backend, const Collection& collection) const
+    {
+        return std::find(collection.begin(), collection.end(), backend) != collection.end();
+    }
+};
+
+} //namespace armnn
diff --git a/src/armnn/Graph.cpp b/src/armnn/Graph.cpp
index 83d82a5..831d85e 100644
--- a/src/armnn/Graph.cpp
+++ b/src/armnn/Graph.cpp
@@ -297,6 +297,65 @@
     }
 }
 
+void Graph::SubstituteSubGraph(std::unique_ptr<SubGraph> subGraph, IConnectableLayer* substituteLayer)
+{
+    BOOST_ASSERT(subGraph != nullptr);
+    BOOST_ASSERT(substituteLayer != nullptr);
+
+    ReplaceSubGraphConnections(*subGraph, substituteLayer);
+    EraseSubGraphLayers(*subGraph);
+}
+
+void Graph::ReplaceSubGraphConnections(const SubGraph& subGraph, IConnectableLayer* substituteLayer)
+{
+    BOOST_ASSERT(substituteLayer != nullptr);
+    BOOST_ASSERT_MSG(std::find(m_Layers.begin(), m_Layers.end(), substituteLayer) != m_Layers.end(),
+                     "Substitue layer is not a member of graph");
+
+    const SubGraph::InputSlots& subGraphInputSlots = subGraph.GetInputSlots();
+    const SubGraph::OutputSlots& subGraphOutputSlots = subGraph.GetOutputSlots();
+
+    const unsigned int numInputSlots = boost::numeric_cast<unsigned int>(subGraphInputSlots.size());
+    const unsigned int numOutputSlots = boost::numeric_cast<unsigned int>(subGraphOutputSlots.size());
+
+    BOOST_ASSERT(numInputSlots == substituteLayer->GetNumInputSlots());
+    BOOST_ASSERT(numOutputSlots == substituteLayer->GetNumOutputSlots());
+
+    // Disconnect the sub-graph and replace it with the substitute layer
+    // Step 1: process input slots
+    for(unsigned int inputSlotIdx = 0u; inputSlotIdx < numInputSlots; ++inputSlotIdx)
+    {
+        InputSlot* subGraphInputSlot = subGraphInputSlots.at(inputSlotIdx);
+        BOOST_ASSERT(subGraphInputSlot != nullptr);
+
+        IOutputSlot* connectedOutputSlot = subGraphInputSlot->GetConnection();
+        BOOST_ASSERT(connectedOutputSlot != nullptr);
+        connectedOutputSlot->Disconnect(*subGraphInputSlot);
+
+        IInputSlot& substituteInputSlot = substituteLayer->GetInputSlot(inputSlotIdx);
+        connectedOutputSlot->Connect(substituteInputSlot);
+    }
+
+    // Step 2: process output slots
+    for(unsigned int outputSlotIdx = 0u; outputSlotIdx < numOutputSlots; ++outputSlotIdx)
+    {
+        OutputSlot* subGraphOutputSlot = subGraphOutputSlots.at(outputSlotIdx);
+        BOOST_ASSERT(subGraphOutputSlot != nullptr);
+
+        OutputSlot* substituteOutputSlot = boost::polymorphic_downcast<OutputSlot*>(
+                &substituteLayer->GetOutputSlot(outputSlotIdx));
+        subGraphOutputSlot->MoveAllConnections(*substituteOutputSlot);
+    }
+}
+
+void Graph::EraseSubGraphLayers(const SubGraph &subGraph)
+{
+    for (auto layer : subGraph.GetLayers())
+    {
+        EraseLayer(layer);
+    }
+}
+
 void Graph::InferTensorInfos()
 {
     for (auto&& layer : TopologicalSort())
diff --git a/src/armnn/Graph.hpp b/src/armnn/Graph.hpp
index 7ace2e0..8f93f56 100644
--- a/src/armnn/Graph.hpp
+++ b/src/armnn/Graph.hpp
@@ -6,6 +6,7 @@
 
 #include "LayersFwd.hpp"
 #include "IGraphObservable.hpp"
+#include "SubGraph.hpp"
 
 #include <armnn/Types.hpp>
 #include <armnn/TensorFwd.hpp>
@@ -159,6 +160,8 @@
     /// and relinking them via an intermediary copy layers.
     void AddCopyLayers();
 
+    void SubstituteSubGraph(std::unique_ptr<SubGraph> subGraph, IConnectableLayer* substituteLayer);
+
     void InferTensorInfos();
 
     void AttachObservable(IGraphObservable* const observable, GraphEvent notifyOnEvent) {
@@ -210,6 +213,9 @@
     std::unordered_set<LayerBindingId> m_OutputIds;
     std::unordered_map<const Layer*, Iterator> m_PosInGraphMap;
 
+    void ReplaceSubGraphConnections(const SubGraph& subGraph, IConnectableLayer* substituteLayer);
+    void EraseSubGraphLayers(const SubGraph &subGraph);
+
     /// Mutable to allow sorting on const object.
     mutable LayersList m_Layers;
     mutable bool m_LayersInOrder;
diff --git a/src/armnn/ISubGraphConverter.hpp b/src/armnn/ISubGraphConverter.hpp
new file mode 100644
index 0000000..1d40c67
--- /dev/null
+++ b/src/armnn/ISubGraphConverter.hpp
@@ -0,0 +1,22 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <memory>
+
+namespace armnn
+{
+
+class ISubGraphConverter
+{
+public:
+    virtual ~ISubGraphConverter() {};
+
+    virtual std::shared_ptr<void> GetOutput() = 0;
+};
+
+}
+
diff --git a/src/armnn/InternalTypes.cpp b/src/armnn/InternalTypes.cpp
index 9ffd73a..16a1972 100644
--- a/src/armnn/InternalTypes.cpp
+++ b/src/armnn/InternalTypes.cpp
@@ -44,6 +44,7 @@
         case LayerType::Pad: return "Pad";
         case LayerType::Permute: return "Permute";
         case LayerType::Pooling2d: return "Pooling2d";
+        case LayerType::PreCompiled: return "PreCompiled";
         case LayerType::Reshape: return "Reshape";
         case LayerType::Rsqrt: return "Rsqrt";
         case LayerType::ResizeBilinear: return "ResizeBilinear";
diff --git a/src/armnn/InternalTypes.hpp b/src/armnn/InternalTypes.hpp
index f4996db..f05ea25 100644
--- a/src/armnn/InternalTypes.hpp
+++ b/src/armnn/InternalTypes.hpp
@@ -44,6 +44,7 @@
     Pad,
     Permute,
     Pooling2d,
+    PreCompiled,
     Reshape,
     ResizeBilinear,
     Rsqrt,
@@ -53,7 +54,7 @@
     StridedSlice,
     // Last layer goes here.
     LastLayer,
-    Subtraction = LastLayer,
+    Subtraction = LastLayer
 };
 
 const char* GetLayerTypeAsCString(LayerType type);
diff --git a/src/armnn/LayerSupportCommon.hpp b/src/armnn/LayerSupportCommon.hpp
index d6dda4f..c309f8c 100644
--- a/src/armnn/LayerSupportCommon.hpp
+++ b/src/armnn/LayerSupportCommon.hpp
@@ -7,6 +7,7 @@
 #include <armnn/DescriptorsFwd.hpp>
 #include <armnn/Types.hpp>
 #include <armnn/Tensor.hpp>
+#include <armnn/Optional.hpp>
 
 namespace armnn
 {
diff --git a/src/armnn/LayersFwd.hpp b/src/armnn/LayersFwd.hpp
index 9f55233..8b4ee08 100644
--- a/src/armnn/LayersFwd.hpp
+++ b/src/armnn/LayersFwd.hpp
@@ -36,6 +36,7 @@
 #include "layers/PadLayer.hpp"
 #include "layers/PermuteLayer.hpp"
 #include "layers/Pooling2dLayer.hpp"
+#include "layers/PreCompiledLayer.hpp"
 #include "layers/ReshapeLayer.hpp"
 #include "layers/ResizeBilinearLayer.hpp"
 #include "layers/RsqrtLayer.hpp"
@@ -102,6 +103,7 @@
 DECLARE_LAYER(Pad)
 DECLARE_LAYER(Permute)
 DECLARE_LAYER(Pooling2d)
+DECLARE_LAYER(PreCompiled)
 DECLARE_LAYER(Reshape)
 DECLARE_LAYER(ResizeBilinear)
 DECLARE_LAYER(Rsqrt)
diff --git a/src/armnn/Network.cpp b/src/armnn/Network.cpp
index 187d04e..7b9cb3d 100644
--- a/src/armnn/Network.cpp
+++ b/src/armnn/Network.cpp
@@ -2,11 +2,14 @@
 // Copyright © 2017 Arm Ltd. All rights reserved.
 // SPDX-License-Identifier: MIT
 //
+
 #include "Network.hpp"
 #include "Graph.hpp"
 #include "Layer.hpp"
 #include "DeviceSpec.hpp"
 #include "Optimizer.hpp"
+#include "SubGraphSelector.hpp"
+#include "BackendSettings.hpp"
 #include "optimizations/All.hpp"
 
 #include <backendsCommon/CpuTensorHandle.hpp>
@@ -71,6 +74,41 @@
     return m_Graph->SerializeToDot(stream);
 }
 
+struct OptimizationResult
+{
+    bool m_Warning;
+    bool m_Error;
+
+    OptimizationResult()
+        : m_Warning(false)
+        , m_Error(false)
+    {}
+};
+
+void ReportError(const std::string& errorMessage,
+                 Optional<std::vector<std::string>&> errorMessages)
+{
+    std::stringstream fullErrorMessage;
+    fullErrorMessage << "ERROR: " << errorMessage;
+    BOOST_LOG_TRIVIAL(warning) << fullErrorMessage.str();
+    if (errorMessages)
+    {
+        errorMessages.value().push_back(fullErrorMessage.str());
+    }
+}
+
+void ReportWarning(const std::string& warningMessage,
+                   Optional<std::vector<std::string>&> warningMessages)
+{
+    std::stringstream fullWarningMessage;
+    fullWarningMessage << "WARNING: " << warningMessage;
+    BOOST_LOG_TRIVIAL(warning) << fullWarningMessage.str();
+    if (warningMessages)
+    {
+        warningMessages.value().push_back(fullWarningMessage.str());
+    }
+}
+
 bool CheckScaleSetOnQuantizedType(Layer* layer, Optional<std::vector<std::string>&> errMessages)
 {
     bool noErrors = true;
@@ -82,108 +120,50 @@
             if (0.f == info.GetQuantizationScale()) {
                 noErrors = false;
                 std::stringstream ss;
-                ss << "ERROR: output " << i << " of layer " << GetLayerTypeAsCString(layer->GetType())
+                ss << "output " << i << " of layer " << GetLayerTypeAsCString(layer->GetType())
                    << " (" << layer->GetNameStr() << ") is of type"
                    << " Quantized 8 bit but its scale parameter has not been set";
-                BOOST_LOG_TRIVIAL(warning) << ss.str() ;
-                if (errMessages) {
-                    errMessages.value().push_back(ss.str());
-                }
+                ReportError(ss.str(), errMessages);
             }
         }
     }
     return noErrors;
 }
 
-IOptimizedNetworkPtr Optimize(const INetwork& inNetwork,
-                              const std::vector<BackendId>& backendPreferences,
-                              const IDeviceSpec& deviceSpec,
-                              const OptimizerOptions& options,
-                              Optional<std::vector<std::string>&> errMessages)
+OptimizationResult AssignBackends(OptimizedNetwork* optNetObjPtr,
+                                  BackendSettings& backendSettings,
+                                  Graph::Iterator& firstLayer,
+                                  Graph::Iterator& lastLayer,
+                                  Optional<std::vector<std::string>&> errMessages)
 {
-    if (backendPreferences.empty()) {
-        throw armnn::InvalidArgumentException("Invoked Optimize with no backends specified");
-    }
-    const Network& network = *boost::polymorphic_downcast<const Network*>(&inNetwork);
-    std::unique_ptr<Graph> graph = std::make_unique<Graph>(network.GetGraph());
+    OptimizationResult result;
 
-    auto optNet = IOptimizedNetworkPtr(new OptimizedNetwork(std::move(graph)), &IOptimizedNetwork::Destroy);
-
-    OptimizedNetwork* optNetObjPtr = boost::polymorphic_downcast<OptimizedNetwork*>(optNet.get());
-
-    // Perform optimisation passes
-    using namespace optimizations;
-    Optimizer::Pass(optNetObjPtr->GetGraph(), MakeOptimizations(SquashEqualPermuteSiblings(),
-                                                                SquashEqualReshapeSiblings(),
-                                                                OptimizeInversePermutes(),
-                                                                MovePermuteUp(),
-                                                                PermuteAsReshape(),
-                                                                OptimizeConsecutiveReshapes()));
-
-    // Infer the tensor infos for all output slots. Throws an exception on failure.
-    optNetObjPtr->GetGraph().InferTensorInfos();
-
-    // if Fp32 to Fp16 optimization is set convert Fp32 network to Fp16
-    if (options.m_ReduceFp32ToFp16)
-    {
-        Optimizer::Pass(optNetObjPtr->GetGraph(), MakeOptimizations(Fp32NetworkToFp16Converter()));
-    }
-
-    // if debug optimization is set, then print out data after each layer
-    if (options.m_Debug)
-    {
-        Optimizer::Pass(optNetObjPtr->GetGraph(), MakeOptimizations(InsertDebugLayer()));
-    }
-
-    // We know that DeviceSpec should be the only implementation of IDeviceSpec.
-    const DeviceSpec& spec = *boost::polymorphic_downcast<const DeviceSpec*>(&deviceSpec);
-    auto const& supportedBackends = spec.GetSupportedBackends();
-
-    // determine which of the preferred backends we have available for use
-    // and whether we have specified CpuRef as one of those backends.
-    bool cpuRefUsed = false;
-    std::vector<BackendId> availablePreferredBackends;
-    for (const auto& backend : backendPreferences)
-    {
-        // Check if the backend is in the available backend devices.
-        if (supportedBackends.count(backend) > 0)
-        {
-            availablePreferredBackends.push_back(backend);
-            if (backend == armnn::Compute::CpuRef) {
-                cpuRefUsed = true;
-            }
-        }
-    }
-    if (availablePreferredBackends.empty()) {
-        std::stringstream failureMsg;
-        failureMsg << "ERROR: None of the preferred backends " << backendPreferences
-                   << " are supported. Current platform provides " << supportedBackends;
-        BOOST_LOG_TRIVIAL(warning) << failureMsg.str();
-        if (errMessages) {
-            errMessages.value().push_back(failureMsg.str());
-        }
-        return IOptimizedNetworkPtr(nullptr, &IOptimizedNetwork::Destroy);
-    }
-
-    auto ReturnWithError = [&](Layer* layer)
+    // Helper lambda to compose meaningful error message before returning with error
+    auto ReturnWithError = [&](const Layer* layer)
     {
         std::stringstream failureMsg;
-        failureMsg << "ERROR: Layer of type " << GetLayerTypeAsCString(layer->GetType())
-                   << " is not supported on any preferred backend " << backendPreferences;
-        BOOST_LOG_TRIVIAL(warning) << failureMsg.str();
-        if (errMessages) {
-            errMessages.value().push_back(failureMsg.str());
-        }
-        return IOptimizedNetworkPtr(nullptr, &IOptimizedNetwork::Destroy);
+        failureMsg << "Layer of type " << GetLayerTypeAsCString(layer->GetType())
+                   << " is not supported on any preferred backend " << backendSettings.m_PreferredBackends;
+        ReportError(failureMsg.str(), errMessages);
+
+        result.m_Error = true;
+        return result;
     };
 
-    // The backends that we choose to run layers on
-    std::unordered_set<BackendId> chosenBackends;
-
-    // Assign a compute device for all nodes
-    bool bErrorFound = false;
-    for (auto&& layer : optNetObjPtr->GetGraph())
+    auto availablePreferredBackends = backendSettings.GetAvailablePreferredBackends();
+    if (availablePreferredBackends.empty())
     {
+        std::stringstream failureMsg;
+        failureMsg << "No preferred backends are available";
+        ReportError(failureMsg.str(), errMessages);
+
+        result.m_Error = true;
+        return result;
+    }
+
+    for (auto it = firstLayer; it != lastLayer; ++it)
+    {
+        auto layer = *it;
         DataType dataType = layer->GetDataType();
         std::string reasonIfUnsupported;
         bool found = false;
@@ -191,8 +171,9 @@
         {
             // don't bomb immediately, find all the quantized outputs
             // which haven't had a scale set and report them all back.
-            bErrorFound = true;
+            result.m_Error = true;
         }
+
         for (const auto& backend : availablePreferredBackends)
         {
             // need to set the compute device on the layer
@@ -273,38 +254,36 @@
                     }
                 }
                 std::stringstream warningMsg;
-                warningMsg << "WARNING: Layer of type " << GetLayerTypeAsCString(layer->GetType())
+                warningMsg << "Layer of type " << GetLayerTypeAsCString(layer->GetType())
                            << " is not supported on requested backend " << layer->GetBackendId().Get()
                            << " for data type " << GetDataTypeName(dataType)
                            << " (reason: " << reasonIfUnsupported
                            << "), falling back to the next backend.";
-                std::string wMsg = warningMsg.str();
-                BOOST_LOG_TRIVIAL(warning) << wMsg;
-                if (errMessages) {
-                    errMessages.value().push_back(wMsg);
-                }
+                ReportWarning(warningMsg.str(), errMessages);
             }
             else
             {
                 found = true;
-                chosenBackends.insert(backend);
+                backendSettings.m_SelectedBackends.insert(backend);
                 break;
             }
         }
 
         // If the layer is unsupported by any devices, log and return a null network.
-        if (!found) {
+        if (!found)
+        {
             // NOTE: if the layer is not an operation queue type AND we have not got CpuRef as a
             //       fallback we should set the compute device on the layer to CpuRef (these are not
             //       available as accelerated operations, or are only available under certain
             //       conditions, currently they comprise MemCopy, Constant, Permute)
             armnn::LayerType layerType = layer->GetType();
-            if (!cpuRefUsed && (layerType == armnn::LayerType::MemCopy ||
-                                layerType == armnn::LayerType::Constant ||
-                                layerType == armnn::LayerType::Permute))
+            if (!backendSettings.IsCpuRefUsed() && (layerType == armnn::LayerType::MemCopy ||
+                                                    layerType == armnn::LayerType::Constant ||
+                                                    layerType == armnn::LayerType::Permute))
             {
-                layer->SetBackendId(armnn::Compute::CpuRef);
-                chosenBackends.insert(armnn::Compute::CpuRef);
+                BackendId cpuBackendId(armnn::Compute::CpuRef);
+                layer->SetBackendId(cpuBackendId);
+                backendSettings.m_SelectedBackends.insert(cpuBackendId);
             }
             else
             {
@@ -312,14 +291,175 @@
             }
         }
     }
-    if (bErrorFound)
+
+    return result;
+}
+
+OptimizationResult InsertPreCompiledLayers(OptimizedNetwork* optNetObjPtr,
+                                           const IBackendInternalUniquePtr& backendObjPtr,
+                                           BackendSettings& backendSettings,
+                                           Optional<std::vector<std::string>&> errMessages)
+{
+    BOOST_ASSERT(backendObjPtr);
+
+    OptimizationResult result;
+
+    // Select sub-graphs based on backend
+    SubGraphSelector::SubGraphs subGraphs =
+            SubGraphSelector::SelectSubGraphs(optNetObjPtr->GetGraph(),
+                                              // select layers assigned to requested backend
+                                              [&](const Layer& layer)
+                                              {
+                                                  return layer.GetType() != LayerType::Input &&
+                                                         layer.GetType() != LayerType::Output &&
+                                                         layer.GetBackendId() == backendObjPtr->GetId();
+                                              });
+
+    if (subGraphs.empty())
     {
+        // No sub-graphs found -> return with no error
+        return result;
+    }
+
+    // Convert sub-graphs and substitute them with pre-compiled layers
+    unsigned int index = 0u;
+    for (auto& subGraph : subGraphs)
+    {
+        // Create a pre-compiled layer
+        PreCompiledLayer* preCompiledLayer = CreatePreCompiledLayer(optNetObjPtr->GetGraph(),
+                                                                    *subGraph,
+                                                                    index++,
+                                                                    backendObjPtr);
+        if (preCompiledLayer)
+        {
+            // Substitute sub-graph with pre-compiled layer in graph
+            optNetObjPtr->GetGraph().SubstituteSubGraph(std::move(subGraph), preCompiledLayer);
+        }
+        else
+        {
+            // Failed to create pre-compiled layer from sub-graph ->
+            // re-assign sub-graph layers to other available backends
+            std::stringstream warningMsg;
+            warningMsg << "Sub-graph #" << index << " failed to compile on "
+                       << backendObjPtr->GetId() << ". Re-assigning backends to "
+                       << subGraph->GetLayers().size() << " layers inside sub-graph";
+            ReportWarning(warningMsg.str(), errMessages);
+
+            backendSettings.m_IgnoredBackends = { backendObjPtr->GetId() };
+
+            Graph::Iterator firstLayer = subGraph->begin();
+            Graph::Iterator lastLayer  = subGraph->end();
+            OptimizationResult reassignmentResult = AssignBackends(optNetObjPtr,
+                                                                   backendSettings,
+                                                                   firstLayer,
+                                                                   lastLayer,
+                                                                   errMessages);
+
+            if (reassignmentResult.m_Error)
+            {
+                result.m_Error = true;
+                return result;
+            }
+        }
+    }
+
+    return result;
+}
+
+IOptimizedNetworkPtr Optimize(const INetwork& inNetwork,
+                              const std::vector<BackendId>& backendPreferences,
+                              const IDeviceSpec& deviceSpec,
+                              const OptimizerOptions& options,
+                              Optional<std::vector<std::string>&> errMessages)
+{
+    if (backendPreferences.empty())
+    {
+        throw armnn::InvalidArgumentException("Invoked Optimize with no backends specified");
+    }
+
+    const Network& network = *boost::polymorphic_downcast<const Network*>(&inNetwork);
+    std::unique_ptr<Graph> graph = std::make_unique<Graph>(network.GetGraph());
+
+    auto optNet = IOptimizedNetworkPtr(new OptimizedNetwork(std::move(graph)), &IOptimizedNetwork::Destroy);
+
+    OptimizedNetwork* optNetObjPtr = boost::polymorphic_downcast<OptimizedNetwork*>(optNet.get());
+
+    // Perform optimisation passes
+    using namespace optimizations;
+    Optimizer::Pass(optNetObjPtr->GetGraph(), MakeOptimizations(SquashEqualPermuteSiblings(),
+                                                                SquashEqualReshapeSiblings(),
+                                                                OptimizeInversePermutes(),
+                                                                MovePermuteUp(),
+                                                                PermuteAsReshape(),
+                                                                OptimizeConsecutiveReshapes()));
+
+    // Infer the tensor infos for all output slots. Throws an exception on failure.
+    optNetObjPtr->GetGraph().InferTensorInfos();
+
+    // If Fp32 to Fp16 optimization is set convert Fp32 network to Fp16
+    if (options.m_ReduceFp32ToFp16)
+    {
+        Optimizer::Pass(optNetObjPtr->GetGraph(), MakeOptimizations(Fp32NetworkToFp16Converter()));
+    }
+
+    // Initialize backend settings
+    BackendSettings backendSettings(backendPreferences, deviceSpec);
+    if (backendSettings.GetAvailablePreferredBackends().empty())
+    {
+        std::stringstream failureMsg;
+        failureMsg << "None of the preferred backends " << backendPreferences
+                   << " are supported. Current platform provides " << backendSettings.m_SupportedBackends;
+        ReportError(failureMsg.str(), errMessages);
+        return IOptimizedNetworkPtr(nullptr, &IOptimizedNetwork::Destroy);
+    }
+
+    // Assign an available backend to each layer
+    Graph::Iterator firstLayer = optNetObjPtr->GetGraph().begin();
+    Graph::Iterator lastLayer  = optNetObjPtr->GetGraph().end();
+    OptimizationResult assigBackendsResult = AssignBackends(optNetObjPtr,
+                                                            backendSettings,
+                                                            firstLayer,
+                                                            lastLayer,
+                                                            errMessages);
+    if (assigBackendsResult.m_Error)
+    {
+        // Failed to assign a backend to each layer
         return IOptimizedNetworkPtr(nullptr, &IOptimizedNetwork::Destroy);
     }
 
     Optimizer::Pass(optNetObjPtr->GetGraph(), MakeOptimizations(OptimizeInverseConversionsFp16(),
                                                                 OptimizeInverseConversionsFp32()));
 
+    // Insert pre-compiled layers where required by the backend
+    // TODO: This is a dummy/default backend id used for making the code build until
+    //       we've properly refactored the optimizer
+    const BackendId backendId(Compute::Undefined);
+    auto const& backendRegistry = BackendRegistryInstance();
+    if (backendRegistry.IsBackendRegistered(backendId))
+    {
+        // Obtain a backend object using the registered factory
+        auto backendFactory = backendRegistry.GetFactory(backendId);
+        auto backendObjPtr  = backendFactory();
+
+        OptimizationResult insertPreCompiledLayersResult = InsertPreCompiledLayers(optNetObjPtr,
+                                                                                   backendObjPtr,
+                                                                                   backendSettings,
+                                                                                   errMessages);
+        if (insertPreCompiledLayersResult.m_Error)
+        {
+            // Failed to insert pre-compiled layers
+            return IOptimizedNetworkPtr(nullptr, &IOptimizedNetwork::Destroy);
+        }
+    }
+
+    // If the debug flag is set, then insert a DebugLayer after each layer.
+    // NOTE: This optimization can only happen strictly after the PreCompiled layers have
+    //       already been inserted
+    if (options.m_Debug)
+    {
+        Optimizer::Pass(optNetObjPtr->GetGraph(), MakeOptimizations(InsertDebugLayer()));
+    }
+
     optNetObjPtr->GetGraph().AddCopyLayers();
 
     // Convert constants
@@ -327,7 +467,7 @@
     Optimizer::Pass(optNetObjPtr->GetGraph(), MakeOptimizations(ConvertConstantsHalfToFloat()));
 
     // Run backend specific optimizations
-    for (auto&& chosenBackend : chosenBackends)
+    for (auto&& chosenBackend : backendSettings.m_SelectedBackends)
     {
         auto factoryFun = BackendRegistryInstance().GetFactory(chosenBackend);
         auto backendPtr = factoryFun();
diff --git a/src/armnn/NetworkUtils.cpp b/src/armnn/NetworkUtils.cpp
index 9a4ce87..735a624 100644
--- a/src/armnn/NetworkUtils.cpp
+++ b/src/armnn/NetworkUtils.cpp
@@ -5,6 +5,12 @@
 
 #include "NetworkUtils.hpp"
 
+#include "SubGraphSelector.hpp"
+
+#include <armnn/Exceptions.hpp>
+
+#include <backendsCommon/BackendRegistry.hpp>
+
 namespace armnn
 {
 
@@ -74,7 +80,6 @@
     return convertLayers;
 }
 
-
 std::vector<DebugLayer*> InsertDebugLayerAfter(Graph& graph, Layer& layer)
 {
     std::vector<DebugLayer*> debugLayers;
@@ -97,10 +102,58 @@
 
         debugLayer->GetOutputSlot().SetTensorInfo(debugInfo);
 
+        // NOTE: It is OK to do this because DebugLayer is only supported on CpuRef
+        debugLayer->SetBackendId(Compute::CpuRef);
+
         debugLayers.emplace_back(debugLayer);
     }
 
     return debugLayers;
 }
 
+PreCompiledLayer* CreatePreCompiledLayer(Graph& graph,
+                                         const SubGraph& subGraph,
+                                         unsigned int subGraphIndex,
+                                         const IBackendInternalUniquePtr& backendObjPtr)
+{
+    BOOST_ASSERT(backendObjPtr);
+
+    IBackendInternal::ISubGraphConverterPtr converter =
+            backendObjPtr->CreateSubGraphConverter(std::make_shared<SubGraph>(subGraph));
+    if (!converter)
+    {
+        return nullptr;
+    }
+
+    try
+    {
+        // Attempt to convert and compile sub-graph
+        auto preCompiledObject = converter->GetOutput();
+    }
+    catch (std::exception&)
+    {
+        return nullptr;
+    }
+
+    // Create pre-compiled layer
+    std::string name = "pre-compiled" + std::to_string(subGraphIndex);
+    PreCompiledLayer* preCompiledLayer = graph.AddLayer<PreCompiledLayer>(
+        PreCompiledDescriptor(subGraph.GetNumInputSlots(), subGraph.GetNumOutputSlots()), name.c_str());
+
+    // Copy output tensor infos from sub-graph
+    for (unsigned int i = 0u; i < subGraph.GetNumOutputSlots(); i++)
+    {
+        preCompiledLayer->GetOutputSlot(i).SetTensorInfo(subGraph.GetOutputSlot(i)->GetTensorInfo());
+    }
+
+    // Assign pre-compiled object to layer
+    preCompiledLayer->SetPreCompiledObject(converter->GetOutput());
+
+    // Set the backend-id for the pre-compiled layer
+    BackendId backendId = backendObjPtr->GetId();
+    preCompiledLayer->SetBackendId(backendId);
+
+    return preCompiledLayer;
+}
+
 } // namespace armnn
diff --git a/src/armnn/NetworkUtils.hpp b/src/armnn/NetworkUtils.hpp
index b81d5cb..1a520b7 100644
--- a/src/armnn/NetworkUtils.hpp
+++ b/src/armnn/NetworkUtils.hpp
@@ -5,7 +5,11 @@
 
 #pragma once
 
+#include "DeviceSpec.hpp"
 #include "Graph.hpp"
+#include "SubGraph.hpp"
+
+#include <backendsCommon/IBackendInternal.hpp>
 
 namespace armnn
 {
@@ -16,4 +20,9 @@
 
 std::vector<DebugLayer*> InsertDebugLayerAfter(Graph& graph, Layer& layer);
 
+PreCompiledLayer* CreatePreCompiledLayer(Graph& graph,
+                                         const SubGraph& subGraph,
+                                         unsigned int subGraphIndex,
+                                         const IBackendInternalUniquePtr& backendObject);
+
 } // namespace armnn
diff --git a/src/armnn/SubGraph.cpp b/src/armnn/SubGraph.cpp
index 5d41f32..74a1838 100644
--- a/src/armnn/SubGraph.cpp
+++ b/src/armnn/SubGraph.cpp
@@ -69,4 +69,34 @@
     return m_Layers;
 }
 
+SubGraph::Layers::iterator SubGraph::begin()
+{
+    return m_Layers.begin();
+}
+
+SubGraph::Layers::iterator SubGraph::end()
+{
+    return m_Layers.end();
+}
+
+SubGraph::Layers::const_iterator SubGraph::begin() const
+{
+    return m_Layers.begin();
+}
+
+SubGraph::Layers::const_iterator SubGraph::end() const
+{
+    return m_Layers.end();
+}
+
+SubGraph::Layers::const_iterator SubGraph::cbegin() const
+{
+    return begin();
+}
+
+SubGraph::Layers::const_iterator SubGraph::cend() const
+{
+    return end();
+}
+
 } // namespace armnn
diff --git a/src/armnn/SubGraph.hpp b/src/armnn/SubGraph.hpp
index 312bb11..d22377d 100644
--- a/src/armnn/SubGraph.hpp
+++ b/src/armnn/SubGraph.hpp
@@ -8,7 +8,7 @@
 #include "Layer.hpp"
 
 #include <vector>
-#include <unordered_set>
+#include <list>
 
 namespace armnn
 {
@@ -24,7 +24,7 @@
 public:
     using InputSlots = std::vector<InputSlot *>;
     using OutputSlots = std::vector<OutputSlot *>;
-    using Layers = std::unordered_set<Layer *>;
+    using Layers = std::list<Layer*>;
 
     SubGraph();
     SubGraph(InputSlots && inputs,
@@ -44,6 +44,15 @@
     unsigned int GetNumInputSlots() const;
     unsigned int GetNumOutputSlots() const;
 
+    Layers::iterator begin();
+    Layers::iterator end();
+
+    Layers::const_iterator begin() const;
+    Layers::const_iterator end() const;
+
+    Layers::const_iterator cbegin() const;
+    Layers::const_iterator cend() const;
+
 private:
     InputSlots m_InputSlots;
     OutputSlots m_OutputSlots;
diff --git a/src/armnn/SubGraphSelector.cpp b/src/armnn/SubGraphSelector.cpp
index b87e2b7..d0542fd 100644
--- a/src/armnn/SubGraphSelector.cpp
+++ b/src/armnn/SubGraphSelector.cpp
@@ -166,7 +166,7 @@
             {
                 infoPtr->CollectNonSelectedOutputSlots(outputs, selector);
                 infoPtr->CollectNonSelectedInputs(inputs, selector);
-                layers.insert(infoPtr->m_Layer);
+                layers.push_back(infoPtr->m_Layer);
             }
             result.emplace_back(
                 std::make_unique<SubGraph>(
diff --git a/src/armnn/TypeUtils.hpp b/src/armnn/TypeUtils.hpp
index 01a0e64..5bb040f 100644
--- a/src/armnn/TypeUtils.hpp
+++ b/src/armnn/TypeUtils.hpp
@@ -11,16 +11,9 @@
 namespace armnn
 {
 
-
 template<DataType DT>
 struct ResolveTypeImpl;
 
-template<>
-struct ResolveTypeImpl<DataType::QuantisedAsymm8>
-{
-    using Type = uint8_t;
-};
-
 template <>
 struct ResolveTypeImpl<DataType::Float16>
 {
@@ -34,6 +27,18 @@
 };
 
 template<>
+struct ResolveTypeImpl<DataType::QuantisedAsymm8>
+{
+    using Type = uint8_t;
+};
+
+template<>
+struct ResolveTypeImpl<DataType::Signed32>
+{
+    using Type = int32_t;
+};
+
+template<>
 struct ResolveTypeImpl<DataType::Boolean>
 {
     using Type = bool;
@@ -42,5 +47,4 @@
 template<DataType DT>
 using ResolveType = typename ResolveTypeImpl<DT>::Type;
 
-
 } //namespace armnn
diff --git a/src/armnn/layers/MergerLayer.cpp b/src/armnn/layers/MergerLayer.cpp
index 85dc0e7..b4b5d3c 100644
--- a/src/armnn/layers/MergerLayer.cpp
+++ b/src/armnn/layers/MergerLayer.cpp
@@ -180,7 +180,7 @@
     VerifyLayerConnections(m_Param.GetNumViews(), CHECK_LOCATION());
 
     std::vector<TensorShape> inputShapes;
-    for (uint i = 0; i < GetNumInputSlots(); ++i)
+    for (unsigned int i = 0; i < GetNumInputSlots(); ++i)
     {
         inputShapes.push_back(GetInputSlot(i).GetConnection()->GetTensorInfo().GetShape());
     }
diff --git a/src/armnn/layers/PreCompiledLayer.cpp b/src/armnn/layers/PreCompiledLayer.cpp
new file mode 100644
index 0000000..c443f9a
--- /dev/null
+++ b/src/armnn/layers/PreCompiledLayer.cpp
@@ -0,0 +1,56 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "PreCompiledLayer.hpp"
+
+#include "LayerCloneBase.hpp"
+
+#include "backendsCommon/Workload.hpp"
+
+#include <armnn/TypesUtils.hpp>
+
+namespace armnn
+{
+
+PreCompiledLayer::PreCompiledLayer(const PreCompiledDescriptor& param, const char* name)
+    : LayerWithParameters(param.m_NumInputSlots, param.m_NumOutputSlots, LayerType::PreCompiled, param, name)
+    , m_PreCompiledObject(nullptr)
+{}
+
+PreCompiledLayer::~PreCompiledLayer()
+{}
+
+PreCompiledLayer* PreCompiledLayer::Clone(Graph& graph) const
+{
+    PreCompiledLayer* clone = CloneBase<PreCompiledLayer>(graph, m_Param, GetName());
+    clone->m_PreCompiledObject = this->m_PreCompiledObject;
+    return clone;
+}
+
+std::unique_ptr<IWorkload> PreCompiledLayer::CreateWorkload(const armnn::Graph& graph,
+                                                            const armnn::IWorkloadFactory& factory) const
+{
+    PreCompiledQueueDescriptor descriptor;
+    descriptor.m_PreCompiledObject = m_PreCompiledObject;
+    return factory.CreatePreCompiled(descriptor, PrepInfoAndDesc(descriptor, graph));
+}
+
+void PreCompiledLayer::ValidateTensorShapesFromInputs()
+{
+    // NOTE: since the PreCompiledLayer is an internal layer created from a valid SubGraph,
+    // we do not need to validate its input shapes
+}
+
+std::shared_ptr<void> PreCompiledLayer::GetPreCompiledObject() const
+{
+    return m_PreCompiledObject;
+}
+
+void PreCompiledLayer::SetPreCompiledObject(const std::shared_ptr<void>& preCompiledObject)
+{
+    m_PreCompiledObject = preCompiledObject;
+}
+
+} // namespace armnn
diff --git a/src/armnn/layers/PreCompiledLayer.hpp b/src/armnn/layers/PreCompiledLayer.hpp
new file mode 100644
index 0000000..fd28d0e
--- /dev/null
+++ b/src/armnn/layers/PreCompiledLayer.hpp
@@ -0,0 +1,42 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include "LayerWithParameters.hpp"
+#include <backendsCommon/WorkloadFactory.hpp>
+
+#include <armnn/Descriptors.hpp>
+
+#include <memory>
+
+namespace armnn
+{
+
+class PreCompiledLayer : public LayerWithParameters<PreCompiledDescriptor>
+{
+public:
+    PreCompiledLayer(const PreCompiledDescriptor& param, const char* name);
+    ~PreCompiledLayer();
+
+    virtual std::unique_ptr<IWorkload> CreateWorkload(const Graph& graph,
+                                                      const IWorkloadFactory& factory) const override;
+
+    PreCompiledLayer* Clone(Graph &graph) const override;
+
+    void ValidateTensorShapesFromInputs() override;
+
+    std::shared_ptr<void> GetPreCompiledObject() const;
+
+    void SetPreCompiledObject(const std::shared_ptr<void>& preCompiledObject);
+
+private:
+    PreCompiledLayer(const PreCompiledLayer& other) = delete;
+    PreCompiledLayer& operator=(const PreCompiledLayer& other) = delete;
+
+    std::shared_ptr<void> m_PreCompiledObject;
+};
+
+} // namespace armnn
diff --git a/src/armnn/test/CreateWorkload.hpp b/src/armnn/test/CreateWorkload.hpp
index f52f605..acc5cbd 100644
--- a/src/armnn/test/CreateWorkload.hpp
+++ b/src/armnn/test/CreateWorkload.hpp
@@ -14,6 +14,8 @@
 
 #include <Graph.hpp>
 #include <DataLayoutIndexed.hpp>
+#include <Network.hpp>
+#include <TypeUtils.hpp>
 
 #include <utility>
 
@@ -1093,4 +1095,130 @@
     return std::move(workloadMerger);
 }
 
+template <typename PreCompiledWorkload, armnn::DataType dataType>
+std::pair<armnn::IOptimizedNetworkPtr, std::unique_ptr<PreCompiledWorkload>> CreatePreCompiledWorkloadTest(
+    armnn::IWorkloadFactory& factory,
+    armnn::Graph& graph,
+    bool biasEnabled = false)
+{
+    // To create a PreCompiled layer, create a network and Optimize it.
+    armnn::Network net;
+
+    // Add an input layer
+    armnn::IConnectableLayer* const inputLayer = net.AddInputLayer(0, "input layer");
+    BOOST_TEST(inputLayer);
+
+    // ArmNN weights tensor shape is OIHW (out channels, in channels, height, width) for NCHW
+    // ArmNN weights tensor shape is OHWI (out channels, height, width, in channels) for NHWC
+    // this test is using NHWC, so the weights shape is OHWI
+    TensorInfo weightsTensorInfo(TensorShape({16, 1, 1, 16}), dataType, 0.9f, 0);
+    unsigned int weightsLength = weightsTensorInfo.GetNumElements();
+
+    using WeightType = armnn::ResolveType<dataType>;
+    std::vector<WeightType> convWeightsData(weightsLength);
+    for (unsigned int i = 0; i < weightsLength; ++i)
+    {
+        convWeightsData[i] = static_cast<WeightType>(i);
+    }
+
+    armnn::ConstTensor weights(weightsTensorInfo, convWeightsData);
+
+    // Add a layer that can be used in the PreCompiled layer
+    armnn::Convolution2dDescriptor convDesc2d;
+    convDesc2d.m_StrideX = 1;
+    convDesc2d.m_StrideY = 1;
+    convDesc2d.m_BiasEnabled = biasEnabled;
+    convDesc2d.m_DataLayout = armnn::DataLayout::NHWC;
+
+    armnn::IConnectableLayer* convLayer = nullptr;
+    const std::string convLayerName("conv layer");
+
+    if (biasEnabled)
+    {
+        constexpr armnn::DataType biasDataType = ( dataType == armnn::DataType::QuantisedAsymm8) ?
+            armnn::DataType::Signed32 : armnn::DataType::Float32;
+
+        TensorInfo biasTensorInfo(TensorShape({1, 1, 1, 16}), biasDataType, 0.9f * 0.9f, 0);
+        unsigned int biasLength = biasTensorInfo.GetNumElements();
+
+        using BiasType = armnn::ResolveType<biasDataType>;
+        std::vector<BiasType> biasData(biasLength);
+        std::fill(biasData.begin(), biasData.end(), static_cast<BiasType>(0));
+
+        armnn::ConstTensor biases(biasTensorInfo, biasData);
+
+        // Create convolution layer with biases
+        convLayer = net.AddConvolution2dLayer(convDesc2d, weights, biases, convLayerName.c_str());
+    }
+    else
+    {
+        // Create convolution layer without biases
+        convLayer = net.AddConvolution2dLayer(convDesc2d, weights, convLayerName.c_str());
+    }
+
+    BOOST_TEST(convLayer);
+
+    // Add an output layer
+    armnn::IConnectableLayer* const outputLayer = net.AddOutputLayer(0, "output layer");
+    BOOST_TEST(outputLayer);
+
+    // set the tensors in the network (NHWC format)
+    TensorInfo inputTensorInfo(TensorShape({ 1, 16, 16, 16 }), dataType);
+    if (dataType == armnn::DataType::QuantisedAsymm8)
+    {
+        inputTensorInfo.SetQuantizationOffset(0);
+        inputTensorInfo.SetQuantizationScale(0.9f);
+    }
+
+    TensorInfo outputTensorInfo(TensorShape({1, 16, 16, 16}), dataType);
+    if (dataType == armnn::DataType::QuantisedAsymm8)
+    {
+        outputTensorInfo.SetQuantizationOffset(0);
+        outputTensorInfo.SetQuantizationScale(0.9f);
+    }
+
+    // Connect the layers
+    inputLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(0));
+    inputLayer->GetOutputSlot(0).SetTensorInfo(inputTensorInfo);
+
+    convLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
+    convLayer->GetOutputSlot(0).SetTensorInfo(outputTensorInfo);
+
+    // Optimize the network for the backend supported by the factory
+    std::vector<armnn::BackendId> backends = {factory.GetBackendId()};
+    armnn::IRuntime::CreationOptions options;
+    armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
+    armnn::OptimizerOptions optimizerOptions;
+    armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(net, backends, runtime->GetDeviceSpec(),
+                                                               optimizerOptions);
+    BOOST_CHECK(optimizedNet != nullptr);
+
+    // Find the PreCompiled layer in the optimised graph
+    armnn::Graph& optimisedGraph = static_cast<armnn::OptimizedNetwork*>(optimizedNet.get())->GetGraph();
+    Layer* preCompiledLayer = nullptr;
+    for (auto& layer : optimisedGraph)
+    {
+        if (layer->GetType() == LayerType::PreCompiled)
+        {
+            preCompiledLayer = layer;
+        }
+    }
+    BOOST_TEST(preCompiledLayer);
+
+    // Create the TensorHandles.
+    CreateTensorHandles(optimisedGraph, factory);
+
+    // Make the workload and check it.
+    auto workload = MakeAndCheckWorkload<PreCompiledWorkload>(*preCompiledLayer, optimisedGraph, factory);
+
+    PreCompiledQueueDescriptor queueDescriptor = workload->GetData();
+    BOOST_TEST(queueDescriptor.m_Inputs.size()  == 1);
+    BOOST_TEST(queueDescriptor.m_Outputs.size() == 1);
+
+    // Returns the workload so we can do extra, backend-specific tests.
+    // NOTE: We need to return the optimised network as well, otherwise it gets
+    // out of scope and the tensor handles get destructed
+    return std::make_pair(std::move(optimizedNet), std::move(workload));
+}
+
 }
diff --git a/src/armnn/test/SubGraphTests.cpp b/src/armnn/test/SubGraphTests.cpp
index e516ac0..9e49197 100644
--- a/src/armnn/test/SubGraphTests.cpp
+++ b/src/armnn/test/SubGraphTests.cpp
@@ -17,6 +17,20 @@
 namespace
 {
 
+bool AreAnySubGraphLayersPresentInGraph(const SubGraph::Layers &subGraphLayers, const Graph &graph)
+{
+    for(auto&& layer : subGraphLayers)
+    {
+        auto posInGraph = std::find(graph.begin(), graph.end(), layer);
+        if(posInGraph != graph.end())
+        {
+            return true;
+        }
+    }
+
+    return false;
+}
+
 //
 // this helper only works if all layers where the inputs connect to are not selected
 //
@@ -112,6 +126,235 @@
 
 } // namespace <anonymous>
 
+BOOST_AUTO_TEST_SUITE(SubGraphSubstitution)
+
+BOOST_AUTO_TEST_CASE(SingleInputSingleOutput)
+{
+    // Construct graph
+    Graph graph;
+
+    Layer* const inputLayer = graph.AddLayer<InputLayer>(0, "input");
+
+    Convolution2dDescriptor convDescriptor;
+    Layer* const convLayer1 = graph.AddLayer<Convolution2dLayer>(convDescriptor, "conv1");
+    Layer* const convLayer2 = graph.AddLayer<Convolution2dLayer>(convDescriptor, "conv2");
+
+    Layer* const outputLayer = graph.AddLayer<OutputLayer>(0, "output");
+
+    inputLayer->GetOutputSlot(0).Connect(convLayer1->GetInputSlot(0));
+    convLayer1->GetOutputSlot(0).Connect(convLayer2->GetInputSlot(0));
+    convLayer2->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
+
+    // Construct sub-graph
+    SubGraphSelector::SubGraphPtr subGraph =
+            CreateSubGraphFrom(CreateInputsFrom({convLayer1}), CreateOutputsFrom({convLayer2}), {});
+
+    // Save sub-graph connections for comparison after substitution
+    IOutputSlot* subGraphInputConn = subGraph->GetInputSlot(0)->GetConnection();
+    IInputSlot* subGraphOutputConn = subGraph->GetOutputSlot(0)->GetConnection(0);
+
+    // Construct dummy pre-compiled layer
+    PreCompiledDescriptor preCompiledDescriptor(1, 1);
+    Layer* const preCompiledLayer = graph.AddLayer<PreCompiledLayer>(preCompiledDescriptor, "pre-compiled");
+
+    // Substitute sub-graph with pre-compiled layer
+    graph.SubstituteSubGraph(std::move(subGraph), preCompiledLayer);
+
+    // Check that connections are correct after substitution
+    BOOST_CHECK_EQUAL(preCompiledLayer->GetInputSlot(0).GetConnection(), subGraphInputConn);
+    BOOST_CHECK_EQUAL(preCompiledLayer->GetOutputSlot(0).GetConnection(0), subGraphOutputConn);
+}
+
+BOOST_AUTO_TEST_CASE(MultiInputSingleOutput)
+{
+    // Construct graph
+    Graph graph;
+
+    Layer* const inputLayer = graph.AddLayer<InputLayer>(0, "input");
+
+    ViewsDescriptor splitterDescriptor(2);
+    Layer* const splitterLayer = graph.AddLayer<SplitterLayer>(splitterDescriptor, "splitter");
+
+    Convolution2dDescriptor convDescriptor;
+    Layer* const convLayer1 = graph.AddLayer<Convolution2dLayer>(convDescriptor, "conv1");
+    Layer* const convLayer2 = graph.AddLayer<Convolution2dLayer>(convDescriptor, "conv2");
+
+    OriginsDescriptor mergerDescriptor(2);
+    Layer* const mergerLayer = graph.AddLayer<MergerLayer>(mergerDescriptor, "merger");
+
+    Layer* const outputLayer = graph.AddLayer<OutputLayer>(0, "output");
+
+    inputLayer->GetOutputSlot(0).Connect(splitterLayer->GetInputSlot(0));
+    splitterLayer->GetOutputSlot(0).Connect(convLayer1->GetInputSlot(0));
+    splitterLayer->GetOutputSlot(1).Connect(convLayer2->GetInputSlot(0));
+    convLayer1->GetOutputSlot(0).Connect(mergerLayer->GetInputSlot(0));
+    convLayer2->GetOutputSlot(0).Connect(mergerLayer->GetInputSlot(1));
+    mergerLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
+
+    // Construct sub-graph
+    SubGraphSelector::SubGraphPtr subGraph =
+            CreateSubGraphFrom(CreateInputsFrom({convLayer1, convLayer2}), CreateOutputsFrom({mergerLayer}), {});
+
+    // Save sub-graph connections for comparison after substitution
+    IOutputSlot* subGraphInputConn1 = subGraph->GetInputSlot(0)->GetConnection();
+    IOutputSlot* subGraphInputConn2 = subGraph->GetInputSlot(1)->GetConnection();
+
+    IInputSlot* subGraphOutputConn = subGraph->GetOutputSlot(0)->GetConnection(0);
+
+    // Construct dummy pre-compiled layer
+    PreCompiledDescriptor preCompiledDescriptor(2, 1);
+    Layer* const preCompiledLayer = graph.AddLayer<PreCompiledLayer>(preCompiledDescriptor, "pre-compiled");
+
+    // Substitute sub-graph with pre-compiled layer
+    graph.SubstituteSubGraph(std::move(subGraph), preCompiledLayer);
+
+    // Check that connections are correct after substitution
+    BOOST_CHECK_EQUAL(preCompiledLayer->GetInputSlot(0).GetConnection(), subGraphInputConn1);
+    BOOST_CHECK_EQUAL(preCompiledLayer->GetInputSlot(1).GetConnection(), subGraphInputConn2);
+
+    BOOST_CHECK_EQUAL(preCompiledLayer->GetOutputSlot(0).GetConnection(0), subGraphOutputConn);
+}
+
+BOOST_AUTO_TEST_CASE(SingleInputMultiOutput)
+{
+    // Construct graph
+    Graph graph;
+
+    Layer* const inputLayer = graph.AddLayer<InputLayer>(0, "input");
+
+    Convolution2dDescriptor convDescriptor;
+    Layer* const convLayer1 = graph.AddLayer<Convolution2dLayer>(convDescriptor, "conv1");
+    Layer* const convLayer2 = graph.AddLayer<Convolution2dLayer>(convDescriptor, "conv2");
+    OriginsDescriptor mergerDescriptor(2);
+    Layer* const mergerLayer = graph.AddLayer<MergerLayer>(mergerDescriptor, "merger");
+    Layer* const outputLayer = graph.AddLayer<OutputLayer>(0, "output");
+
+    ViewsDescriptor splitterDescriptor(2);
+    Layer* const splitterLayer = graph.AddLayer<SplitterLayer>(splitterDescriptor, "splitter");
+
+    inputLayer->GetOutputSlot(0).Connect(splitterLayer->GetInputSlot(0));
+    splitterLayer->GetOutputSlot(0).Connect(convLayer1->GetInputSlot(0));
+    splitterLayer->GetOutputSlot(1).Connect(convLayer2->GetInputSlot(0));
+    convLayer1->GetOutputSlot(0).Connect(mergerLayer->GetInputSlot(0));
+    convLayer2->GetOutputSlot(0).Connect(mergerLayer->GetInputSlot(1));
+    mergerLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
+
+    // Construct sub-graph
+    SubGraphSelector::SubGraphPtr subGraph =
+            CreateSubGraphFrom(CreateInputsFrom({splitterLayer}), CreateOutputsFrom({convLayer1, convLayer2}), {});
+
+    // Save sub-graph connections for comparison after substitution
+    IOutputSlot* subGraphInputConn1 = subGraph->GetInputSlot(0)->GetConnection();
+
+    IInputSlot* subGraphOutputConn1 = subGraph->GetOutputSlot(0)->GetConnection(0);
+    IInputSlot* subGraphOutputConn2 = subGraph->GetOutputSlot(1)->GetConnection(0);
+
+    // Construct dummy pre-compiled layer
+    PreCompiledDescriptor preCompiledDescriptor(1, 2);
+    Layer* const preCompiledLayer = graph.AddLayer<PreCompiledLayer>(preCompiledDescriptor, "pre-compiled");
+
+    // Substitute sub-graph with pre-compiled layer
+    graph.SubstituteSubGraph(std::move(subGraph), preCompiledLayer);
+
+    // Check that connections are correct after substitution
+    BOOST_CHECK_EQUAL(preCompiledLayer->GetInputSlot(0).GetConnection(), subGraphInputConn1);
+
+    BOOST_CHECK_EQUAL(preCompiledLayer->GetOutputSlot(0).GetConnection(0), subGraphOutputConn1);
+    BOOST_CHECK_EQUAL(preCompiledLayer->GetOutputSlot(1).GetConnection(0), subGraphOutputConn2);
+}
+
+BOOST_AUTO_TEST_CASE(MultiInputMultiOutput)
+{
+    // Construct graph
+    Graph graph;
+
+    Layer* const inputLayer = graph.AddLayer<InputLayer>(0, "input");
+
+    ViewsDescriptor splitterDescriptor(2);
+    Layer* const splitterLayer = graph.AddLayer<SplitterLayer>(splitterDescriptor, "splitter");
+
+    Convolution2dDescriptor convDescriptor;
+    Layer* const convLayer1 = graph.AddLayer<Convolution2dLayer>(convDescriptor, "conv1");
+    Layer* const convLayer2 = graph.AddLayer<Convolution2dLayer>(convDescriptor, "conv2");
+
+    OriginsDescriptor mergerDescriptor(2);
+    Layer* const mergerLayer = graph.AddLayer<MergerLayer>(mergerDescriptor, "merger");
+
+    Layer* const outputLayer = graph.AddLayer<OutputLayer>(0, "output");
+
+    inputLayer->GetOutputSlot(0).Connect(splitterLayer->GetInputSlot(0));
+    splitterLayer->GetOutputSlot(0).Connect(convLayer1->GetInputSlot(0));
+    splitterLayer->GetOutputSlot(1).Connect(convLayer2->GetInputSlot(0));
+    convLayer1->GetOutputSlot(0).Connect(mergerLayer->GetInputSlot(0));
+    convLayer2->GetOutputSlot(0).Connect(mergerLayer->GetInputSlot(1));
+    mergerLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
+
+    // Construct sub-graph
+    SubGraphSelector::SubGraphPtr subGraph = CreateSubGraphFrom(CreateInputsFrom({convLayer1, convLayer2}),
+                                                                CreateOutputsFrom({convLayer1, convLayer2}),
+                                                                {});
+
+    // Save sub-graph connections for comparison after substitution
+    IOutputSlot* subGraphInputConn1 = subGraph->GetInputSlot(0)->GetConnection();
+    IOutputSlot* subGraphInputConn2 = subGraph->GetInputSlot(1)->GetConnection();
+
+    IInputSlot* subGraphOutputConn1 = subGraph->GetOutputSlot(0)->GetConnection(0);
+    IInputSlot* subGraphOutputConn2 = subGraph->GetOutputSlot(1)->GetConnection(0);
+
+    // Construct dummy pre-compiled layer
+    PreCompiledDescriptor preCompiledDescriptor(2, 2);
+    Layer* const preCompiledLayer = graph.AddLayer<PreCompiledLayer>(preCompiledDescriptor, "pre-compiled");
+
+    // Substitute sub-graph with pre-compiled layer
+    graph.SubstituteSubGraph(std::move(subGraph), preCompiledLayer);
+
+    // Check that connections are correct after substitution
+    BOOST_CHECK_EQUAL(preCompiledLayer->GetInputSlot(0).GetConnection(), subGraphInputConn1);
+    BOOST_CHECK_EQUAL(preCompiledLayer->GetInputSlot(1).GetConnection(), subGraphInputConn2);
+
+    BOOST_CHECK_EQUAL(preCompiledLayer->GetOutputSlot(0).GetConnection(0), subGraphOutputConn1);
+    BOOST_CHECK_EQUAL(preCompiledLayer->GetOutputSlot(1).GetConnection(0), subGraphOutputConn2);
+}
+
+BOOST_AUTO_TEST_CASE(EraseReplacedLayers)
+{
+    // Construct graph
+    Graph graph;
+
+    graph.AddLayer<InputLayer>(0, "input");
+
+    ViewsDescriptor splitterDescriptor(2);
+    Layer* const splitterLayer = graph.AddLayer<SplitterLayer>(splitterDescriptor, "splitter");
+
+    Convolution2dDescriptor convDescriptor;
+    Layer* const convLayer1 = graph.AddLayer<Convolution2dLayer>(convDescriptor, "conv1");
+    Layer* const convLayer2 = graph.AddLayer<Convolution2dLayer>(convDescriptor, "conv2");
+
+    OriginsDescriptor mergerDescriptor(2);
+    Layer* const mergerLayer = graph.AddLayer<MergerLayer>(mergerDescriptor, "merger");
+
+    graph.AddLayer<OutputLayer>(0, "output");
+
+    // Construct sub-graph
+    SubGraphSelector::SubGraphPtr subGraph =
+            CreateSubGraphFrom({}, {}, {splitterLayer, convLayer1, convLayer2, mergerLayer});
+
+    // Construct dummy pre-compiled layer
+    PreCompiledDescriptor preCompiledDescriptor(0, 0);
+    Layer* const preCompiledLayer = graph.AddLayer<PreCompiledLayer>(preCompiledDescriptor, "pre-compiled");
+
+    // Save sub-graph layers for later verification
+    const SubGraph::Layers subGraphLayers = subGraph->GetLayers();
+
+    // Substitute sub-graph with pre-compiled layer
+    graph.SubstituteSubGraph(std::move(subGraph), preCompiledLayer);
+
+    // Check that the layers belonging to the sub-graph have been erased from the graph after substitution
+    BOOST_CHECK(!AreAnySubGraphLayersPresentInGraph(subGraphLayers, graph));
+}
+
+BOOST_AUTO_TEST_SUITE_END()
+
 BOOST_AUTO_TEST_SUITE(SubGraphSelection)
 
 BOOST_AUTO_TEST_CASE(NoSubGraphsForNoMatch)
@@ -585,3 +828,167 @@
 }
 
 BOOST_AUTO_TEST_SUITE_END()
+
+BOOST_AUTO_TEST_SUITE(IntegrationTests)
+
+BOOST_AUTO_TEST_CASE(SingleSubGraph)
+{
+    // This test case represents the scenario when we have one subgraph
+    // in which two layers have GpuAcc backend assigned
+
+    //Construct graph
+    Graph graph;
+
+    Layer* const inputLayer = graph.AddLayer<InputLayer>(0, "input");
+
+    Convolution2dDescriptor convDescriptor;
+    Layer* const convLayer1 = graph.AddLayer<Convolution2dLayer>(convDescriptor, "conv1");
+    convLayer1->SetBackendId(Compute::GpuAcc);
+
+    Layer* const convLayer2 = graph.AddLayer<Convolution2dLayer>(convDescriptor, "conv2");
+    convLayer2->SetBackendId(Compute::GpuAcc);
+
+    Layer* const outputLayer = graph.AddLayer<OutputLayer>(0, "output");
+
+    inputLayer->GetOutputSlot(0).Connect(convLayer1->GetInputSlot(0));
+    convLayer1->GetOutputSlot(0).Connect(convLayer2->GetInputSlot(0));
+    convLayer2->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
+
+    // GpuAcc sub graph selector
+    SubGraphSelector::SubGraphs subGraphs =
+        SubGraphSelector::SelectSubGraphs(
+            graph,
+            // select the GpuAcc layers only
+            [](const Layer & l){
+                bool toSelect = (l.GetBackendId() == Compute::GpuAcc);
+                return toSelect;
+            });
+
+    BOOST_TEST(subGraphs.size() == 1);
+    if(subGraphs.size() == 1)
+    {
+        BOOST_TEST((subGraphs[0] != nullptr));
+
+        if (subGraphs[0].get() != nullptr)
+        {
+            unsigned int numInputSlots = boost::numeric_cast<unsigned int>(subGraphs[0]->GetInputSlots().size());
+            unsigned int numOutputSlots = boost::numeric_cast<unsigned int>(subGraphs[0]->GetOutputSlots().size());
+
+            BOOST_TEST((numInputSlots == 1));
+            BOOST_TEST((numOutputSlots == 1));
+
+            // Save sub-graph connections for comparison after substitution
+            IOutputSlot* subGraphInputConn1 = subGraphs[0]->GetInputSlot(0)->GetConnection();
+            IInputSlot* subGraphOutputConn1 = subGraphs[0]->GetOutputSlot(0)->GetConnection(0);
+
+            // Construct dummy pre-compiled layer
+            PreCompiledDescriptor preCompiledDescriptor(numInputSlots, numOutputSlots);
+            Layer* const preCompiledLayer = graph.AddLayer<PreCompiledLayer>(preCompiledDescriptor, "pre-compiled");
+
+            // Substitute sub-graph with pre-compiled layer
+            graph.SubstituteSubGraph((std::move(subGraphs[0])), preCompiledLayer);
+
+            // Check that connections are correct after substitution
+            BOOST_CHECK_EQUAL(preCompiledLayer->GetInputSlot(0).GetConnection(), subGraphInputConn1);
+
+            BOOST_CHECK_EQUAL(preCompiledLayer->GetOutputSlot(0).GetConnection(0), subGraphOutputConn1);
+        }
+    }
+}
+
+BOOST_AUTO_TEST_CASE(MultipleSubGraphs)
+{
+    // This test case represents the scenario when we have two subgraphs
+    // in which two layers have CpuAcc backend assigned
+
+    //Construct graph
+    Graph graph;
+
+    Layer* const inputLayer = graph.AddLayer<InputLayer>(0, "input");
+
+    ViewsDescriptor splitterDescriptor(2);
+    Layer* const splitterLayer = graph.AddLayer<SplitterLayer>(splitterDescriptor, "splitter");
+    splitterLayer->SetBackendId(Compute::CpuAcc);
+
+    Convolution2dDescriptor convDescriptor;
+    Layer* const convLayer1 = graph.AddLayer<Convolution2dLayer>(convDescriptor, "conv1");
+    Layer* const convLayer2 = graph.AddLayer<Convolution2dLayer>(convDescriptor, "conv2");
+
+    OriginsDescriptor mergerDescriptor(2);
+    Layer* const mergerLayer = graph.AddLayer<MergerLayer>(mergerDescriptor, "merger");
+    mergerLayer->SetBackendId(Compute::CpuAcc);
+
+    Layer* const outputLayer = graph.AddLayer<OutputLayer>(0, "output");
+
+    inputLayer->GetOutputSlot(0).Connect(splitterLayer->GetInputSlot(0));
+    splitterLayer->GetOutputSlot(0).Connect(convLayer1->GetInputSlot(0));
+    splitterLayer->GetOutputSlot(1).Connect(convLayer2->GetInputSlot(0));
+    convLayer1->GetOutputSlot(0).Connect(mergerLayer->GetInputSlot(0));
+    convLayer2->GetOutputSlot(0).Connect(mergerLayer->GetInputSlot(1));
+    mergerLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
+
+    // CpuAcc sub graph selector
+    SubGraphSelector::SubGraphs subGraphs =
+        SubGraphSelector::SelectSubGraphs(
+            graph,
+            // select the CpuAcc layers only
+            [](const Layer & l){
+                bool toSelect = (l.GetBackendId() == Compute::CpuAcc);
+                return toSelect;
+            });
+
+    BOOST_TEST(subGraphs.size() == 2);
+    if(subGraphs.size() == 2)
+    {
+        BOOST_TEST((subGraphs[0] != nullptr));
+        BOOST_TEST((subGraphs[1] != nullptr));
+
+        if (subGraphs[0].get() != nullptr && subGraphs[1].get() != nullptr)
+        {
+            //Sort subGraphs by their inputSlot size.
+            std::sort(subGraphs.begin(), subGraphs.end(),
+                      [](SubGraphSelector::SubGraphPtr & lhs, SubGraphSelector::SubGraphPtr & rhs)
+                      {
+                          return (lhs->GetInputSlots().size() < rhs->GetInputSlots().size());
+                      }
+            );
+
+            unsigned int numInputSlots1  = boost::numeric_cast<unsigned int>(subGraphs[0]->GetInputSlots().size());
+            unsigned int numOutputSlots1 = boost::numeric_cast<unsigned int>(subGraphs[0]->GetOutputSlots().size());
+
+            unsigned int numInputSlots2  = boost::numeric_cast<unsigned int>(subGraphs[1]->GetInputSlots().size());
+            unsigned int numOutputSlots2 = boost::numeric_cast<unsigned int>(subGraphs[1]->GetOutputSlots().size());
+
+            // Save sub-graph connections for comparison after substitution
+            IOutputSlot* subGraph1InputConn  = subGraphs[0]->GetInputSlot(0)->GetConnection();
+            IInputSlot* subGraph1OutputConn1 = subGraphs[0]->GetOutputSlot(0)->GetConnection(0);
+            IInputSlot* subGraph1OutputConn2 = subGraphs[0]->GetOutputSlot(1)->GetConnection(0);
+
+            // Save sub-graph connections for comparison after substitution
+            IOutputSlot* subGraph2InputConn1 = subGraphs[1]->GetInputSlot(0)->GetConnection();
+            IOutputSlot* subGraph2InputConn2 = subGraphs[1]->GetInputSlot(1)->GetConnection();
+            IInputSlot* subGraph2OutputConn  = subGraphs[1]->GetOutputSlot(0)->GetConnection(0);
+
+            PreCompiledDescriptor preCompiledDescriptor1(numInputSlots1, numOutputSlots1);
+            Layer* const preCompiledLayer1 = graph.AddLayer<PreCompiledLayer>(preCompiledDescriptor1, "pre-compiled1");
+
+            PreCompiledDescriptor preCompiledDescriptor2(numInputSlots2, numOutputSlots2);
+            Layer* const preCompiledLayer2 = graph.AddLayer<PreCompiledLayer>(preCompiledDescriptor2, "pre-compiled2");
+
+            // Substitute sub-graph with pre-compiled layer
+            graph.SubstituteSubGraph((std::move(subGraphs[0])), preCompiledLayer1);
+            graph.SubstituteSubGraph((std::move(subGraphs[1])), preCompiledLayer2);
+
+            // Check that connections are correct after substitution
+            BOOST_CHECK_EQUAL(preCompiledLayer1->GetInputSlot(0).GetConnection(), subGraph1InputConn);
+            BOOST_CHECK_EQUAL(preCompiledLayer1->GetOutputSlot(0).GetConnection(0), subGraph1OutputConn1);
+            BOOST_CHECK_EQUAL(preCompiledLayer1->GetOutputSlot(1).GetConnection(0), subGraph1OutputConn2);
+
+            BOOST_CHECK_EQUAL(preCompiledLayer2->GetInputSlot(0).GetConnection(), subGraph2InputConn1);
+            BOOST_CHECK_EQUAL(preCompiledLayer2->GetInputSlot(1).GetConnection(), subGraph2InputConn2);
+            BOOST_CHECK_EQUAL(preCompiledLayer2->GetOutputSlot(0).GetConnection(0), subGraph2OutputConn);
+        }
+    }
+}
+
+BOOST_AUTO_TEST_SUITE_END()
diff --git a/src/backends/aclCommon/test/MemCopyTestImpl.hpp b/src/backends/aclCommon/test/MemCopyTestImpl.hpp
index 4247cc5..4e0bfa8 100644
--- a/src/backends/aclCommon/test/MemCopyTestImpl.hpp
+++ b/src/backends/aclCommon/test/MemCopyTestImpl.hpp
@@ -4,6 +4,8 @@
 //
 #pragma once
 
+#include <TypeUtils.hpp>
+
 #include <backendsCommon/IBackendInternal.hpp>
 
 #include <backendsCommon/test/LayerTests.hpp>
@@ -18,33 +20,29 @@
 namespace
 {
 
-LayerTestResult<float, 4> MemCopyTest(armnn::IWorkloadFactory& srcWorkloadFactory,
-                                      armnn::IWorkloadFactory& dstWorkloadFactory,
-                                      bool withSubtensors)
+template<armnn::DataType dataType, typename T = armnn::ResolveType<dataType>>
+LayerTestResult<T, 4> MemCopyTest(armnn::IWorkloadFactory& srcWorkloadFactory,
+                                  armnn::IWorkloadFactory& dstWorkloadFactory,
+                                  bool withSubtensors)
 {
     const std::array<unsigned int, 4> shapeData = { { 1u, 1u, 6u, 5u } };
     const armnn::TensorShape tensorShape(4, shapeData.data());
-    const armnn::TensorInfo tensorInfo(tensorShape, armnn::DataType::Float32);
-    boost::multi_array<float, 4> inputData = MakeTensor<float, 4>(tensorInfo, std::vector<float>(
+    const armnn::TensorInfo tensorInfo(tensorShape, dataType);
+    boost::multi_array<T, 4> inputData = MakeTensor<T, 4>(tensorInfo, std::vector<T>(
         {
-            1.0f, 2.0f, 3.0f, 4.0f, 5.0f,
-
-            6.0f, 7.0f, 8.0f, 9.0f, 10.0f,
-
-            11.0f, 12.0f, 13.0f, 14.0f, 15.0f,
-
-            16.0f, 17.0f, 18.0f, 19.0f, 20.0f,
-
-            21.0f, 22.0f, 23.0f, 24.0f, 25.0f,
-
-            26.0f, 27.0f, 28.0f, 29.0f, 30.0f,
+             1,  2,  3,  4,  5,
+             6,  7,  8,  9, 10,
+            11, 12, 13, 14, 15,
+            16, 17, 18, 19, 20,
+            21, 22, 23, 24, 25,
+            26, 27, 28, 29, 30,
         })
     );
 
-    LayerTestResult<float, 4> ret(tensorInfo);
+    LayerTestResult<T, 4> ret(tensorInfo);
     ret.outputExpected = inputData;
 
-    boost::multi_array<float, 4> outputData(shapeData);
+    boost::multi_array<T, 4> outputData(shapeData);
 
     auto inputTensorHandle = srcWorkloadFactory.CreateTensorHandle(tensorInfo);
     auto outputTensorHandle = dstWorkloadFactory.CreateTensorHandle(tensorInfo);
@@ -75,8 +73,11 @@
     return ret;
 }
 
-template<typename SrcWorkloadFactory, typename DstWorkloadFactory>
-LayerTestResult<float, 4> MemCopyTest(bool withSubtensors)
+template<typename SrcWorkloadFactory,
+         typename DstWorkloadFactory,
+         armnn::DataType dataType,
+         typename T = armnn::ResolveType<dataType>>
+LayerTestResult<T, 4> MemCopyTest(bool withSubtensors)
 {
     armnn::IBackendInternal::IMemoryManagerSharedPtr srcMemoryManager =
         WorkloadFactoryHelper<SrcWorkloadFactory>::GetMemoryManager();
@@ -87,7 +88,7 @@
     SrcWorkloadFactory srcWorkloadFactory = WorkloadFactoryHelper<SrcWorkloadFactory>::GetFactory(srcMemoryManager);
     DstWorkloadFactory dstWorkloadFactory = WorkloadFactoryHelper<DstWorkloadFactory>::GetFactory(dstMemoryManager);
 
-    return MemCopyTest(srcWorkloadFactory, dstWorkloadFactory, withSubtensors);
+    return MemCopyTest<dataType>(srcWorkloadFactory, dstWorkloadFactory, withSubtensors);
 }
 
 } // anonymous namespace
diff --git a/src/backends/aclCommon/test/MemCopyTests.cpp b/src/backends/aclCommon/test/MemCopyTests.cpp
index 7099a70..78cd95b 100644
--- a/src/backends/aclCommon/test/MemCopyTests.cpp
+++ b/src/backends/aclCommon/test/MemCopyTests.cpp
@@ -47,25 +47,29 @@
 
 BOOST_AUTO_TEST_CASE(CopyBetweenNeonAndGpu)
 {
-    LayerTestResult<float, 4> result = MemCopyTest<armnn::NeonWorkloadFactory, armnn::ClWorkloadFactory>(false);
+    LayerTestResult<float, 4> result =
+        MemCopyTest<armnn::NeonWorkloadFactory, armnn::ClWorkloadFactory, armnn::DataType::Float32>(false);
     BOOST_TEST(CompareTensors(result.output, result.outputExpected));
 }
 
 BOOST_AUTO_TEST_CASE(CopyBetweenGpuAndNeon)
 {
-    LayerTestResult<float, 4> result = MemCopyTest<armnn::ClWorkloadFactory, armnn::NeonWorkloadFactory>(false);
+    LayerTestResult<float, 4> result =
+        MemCopyTest<armnn::ClWorkloadFactory, armnn::NeonWorkloadFactory, armnn::DataType::Float32>(false);
     BOOST_TEST(CompareTensors(result.output, result.outputExpected));
 }
 
 BOOST_AUTO_TEST_CASE(CopyBetweenNeonAndGpuWithSubtensors)
 {
-    LayerTestResult<float, 4> result = MemCopyTest<armnn::NeonWorkloadFactory, armnn::ClWorkloadFactory>(true);
+    LayerTestResult<float, 4> result =
+        MemCopyTest<armnn::NeonWorkloadFactory, armnn::ClWorkloadFactory, armnn::DataType::Float32>(true);
     BOOST_TEST(CompareTensors(result.output, result.outputExpected));
 }
 
 BOOST_AUTO_TEST_CASE(CopyBetweenGpuAndNeonWithSubtensors)
 {
-    LayerTestResult<float, 4> result = MemCopyTest<armnn::ClWorkloadFactory, armnn::NeonWorkloadFactory>(true);
+    LayerTestResult<float, 4> result =
+        MemCopyTest<armnn::ClWorkloadFactory, armnn::NeonWorkloadFactory, armnn::DataType::Float32>(true);
     BOOST_TEST(CompareTensors(result.output, result.outputExpected));
 }
 
diff --git a/src/backends/backends.cmake b/src/backends/backends.cmake
index c82de8d..438fda3 100644
--- a/src/backends/backends.cmake
+++ b/src/backends/backends.cmake
@@ -5,8 +5,8 @@
 
 # single place to use wildcards, so we can include
 # yet unknown backend modules and corresponding common libraries
-FILE(GLOB commonIncludes ${PROJECT_SOURCE_DIR}/src/backends/*/common.cmake)
-FILE(GLOB backendIncludes ${PROJECT_SOURCE_DIR}/src/backends/*/backend.cmake)
+file(GLOB commonIncludes ${PROJECT_SOURCE_DIR}/src/backends/*/common.cmake)
+file(GLOB backendIncludes ${PROJECT_SOURCE_DIR}/src/backends/*/backend.cmake)
 
 # prefer to include common code first
 foreach(includeFile ${commonIncludes})
diff --git a/src/backends/backendsCommon/IBackendInternal.hpp b/src/backends/backendsCommon/IBackendInternal.hpp
index b102d1a..2e6b056 100644
--- a/src/backends/backendsCommon/IBackendInternal.hpp
+++ b/src/backends/backendsCommon/IBackendInternal.hpp
@@ -6,6 +6,10 @@
 
 #include <armnn/Types.hpp>
 #include <armnn/IRuntime.hpp>
+
+#include <ISubGraphConverter.hpp>
+#include <SubGraph.hpp>
+
 #include <vector>
 
 namespace armnn
@@ -37,6 +41,8 @@
     using IMemoryManagerUniquePtr = std::unique_ptr<IMemoryManager>;
     using IMemoryManagerSharedPtr = std::shared_ptr<IMemoryManager>;
 
+    using ISubGraphConverterPtr = std::unique_ptr<ISubGraphConverter>;
+
     virtual IMemoryManagerUniquePtr CreateMemoryManager() const = 0;
 
     virtual IWorkloadFactoryPtr CreateWorkloadFactory(
@@ -44,6 +50,8 @@
 
     virtual IBackendContextPtr CreateBackendContext(const IRuntime::CreationOptions&) const = 0;
 
+    virtual ISubGraphConverterPtr CreateSubGraphConverter(const std::shared_ptr<SubGraph>& subGraph) const = 0;
+
     virtual Optimizations GetOptimizations() const = 0;
     virtual ILayerSupportSharedPtr GetLayerSupport() const = 0;
 };
diff --git a/src/backends/backendsCommon/LayerSupportBase.cpp b/src/backends/backendsCommon/LayerSupportBase.cpp
index 2987e5d..187d2f7 100644
--- a/src/backends/backendsCommon/LayerSupportBase.cpp
+++ b/src/backends/backendsCommon/LayerSupportBase.cpp
@@ -285,6 +285,13 @@
     return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
 }
 
+bool LayerSupportBase::IsPreCompiledSupported(const TensorInfo& input,
+                                              const PreCompiledDescriptor& descriptor,
+                                              Optional<std::string&> reasonIfUnsupported) const
+{
+    return DefaultLayerSupport(__func__, __FILE__, __LINE__, reasonIfUnsupported);
+}
+
 bool LayerSupportBase::IsReshapeSupported(const TensorInfo& input,
                                           const ReshapeDescriptor& descriptor,
                                           Optional<std::string&> reasonIfUnsupported) const
diff --git a/src/backends/backendsCommon/LayerSupportBase.hpp b/src/backends/backendsCommon/LayerSupportBase.hpp
index 8c7aa98..c6f943c 100644
--- a/src/backends/backendsCommon/LayerSupportBase.hpp
+++ b/src/backends/backendsCommon/LayerSupportBase.hpp
@@ -184,6 +184,10 @@
                               const Pooling2dDescriptor& descriptor,
                               Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
 
+    bool IsPreCompiledSupported(const TensorInfo& input,
+                                const PreCompiledDescriptor& descriptor,
+                                Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
+
     bool IsReshapeSupported(const TensorInfo& input,
                             const ReshapeDescriptor& descriptor,
                             Optional<std::string&> reasonIfUnsupported = EmptyOptional()) const override;
diff --git a/src/backends/backendsCommon/WorkloadData.cpp b/src/backends/backendsCommon/WorkloadData.cpp
index a5db088..97981e2 100644
--- a/src/backends/backendsCommon/WorkloadData.cpp
+++ b/src/backends/backendsCommon/WorkloadData.cpp
@@ -1051,4 +1051,9 @@
                               "output");
 }
 
+void PreCompiledQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const
+{
+    // This is internally generated so it should not need validation.
+}
+
 } //namespace armnn
diff --git a/src/backends/backendsCommon/WorkloadData.hpp b/src/backends/backendsCommon/WorkloadData.hpp
index 59e3dfb..453896b 100644
--- a/src/backends/backendsCommon/WorkloadData.hpp
+++ b/src/backends/backendsCommon/WorkloadData.hpp
@@ -378,4 +378,16 @@
     void Validate(const WorkloadInfo& workloadInfo) const;
 };
 
+struct PreCompiledQueueDescriptor : QueueDescriptorWithParameters<PreCompiledDescriptor>
+{
+    PreCompiledQueueDescriptor()
+        : m_PreCompiledObject(nullptr)
+    {
+    }
+
+    std::shared_ptr<void> m_PreCompiledObject;
+
+    void Validate(const WorkloadInfo& workloadInfo) const;
+};
+
 } //namespace armnn
diff --git a/src/backends/backendsCommon/WorkloadDataFwd.hpp b/src/backends/backendsCommon/WorkloadDataFwd.hpp
index 9ae20e0..9fbd81b 100644
--- a/src/backends/backendsCommon/WorkloadDataFwd.hpp
+++ b/src/backends/backendsCommon/WorkloadDataFwd.hpp
@@ -23,5 +23,6 @@
 struct BatchNormalizationQueueDescriptor;
 struct FakeQuantizationQueueDescriptor;
 struct ReshapeQueueDescriptor;
+struct PreCompiledQueueDescriptor;
 
-} // namespace armnn
\ No newline at end of file
+} // namespace armnn
diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp
index 209ba6a..0f015bd 100644
--- a/src/backends/backendsCommon/WorkloadFactory.cpp
+++ b/src/backends/backendsCommon/WorkloadFactory.cpp
@@ -193,14 +193,13 @@
         case LayerType::Debug:
         {
             auto cLayer = boost::polymorphic_downcast<const DebugLayer*>(&layer);
-            const DebugDescriptor& descriptor = cLayer->GetParameters();
 
             const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
             const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo();
 
             result = layerSupportObject->IsDebugSupported(OverrideDataType(input, dataType),
                                                           OverrideDataType(output, dataType),
-                                                          descriptor,
+                                                          cLayer->GetParameters(),
                                                           reason);
             break;
         }
@@ -577,6 +576,15 @@
                                                               reason);
             break;
         }
+        case LayerType::PreCompiled:
+        {
+            auto cLayer = boost::polymorphic_downcast<const PreCompiledLayer*>(&layer);
+            const TensorInfo& input = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
+            result = layerSupportObject->IsPreCompiledSupported(OverrideDataType(input, dataType),
+                                                                cLayer->GetParameters(),
+                                                                reason);
+            break;
+        }
         case LayerType::Division:
         {
             const TensorInfo& input0 = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
diff --git a/src/backends/backendsCommon/WorkloadFactory.hpp b/src/backends/backendsCommon/WorkloadFactory.hpp
index aee9f91..d516698 100644
--- a/src/backends/backendsCommon/WorkloadFactory.hpp
+++ b/src/backends/backendsCommon/WorkloadFactory.hpp
@@ -159,6 +159,9 @@
 
     virtual std::unique_ptr<IWorkload> CreateRsqrt(const RsqrtQueueDescriptor& descriptor,
                                                    const WorkloadInfo& info) const = 0;
+
+    virtual std::unique_ptr<IWorkload> CreatePreCompiled(const PreCompiledQueueDescriptor& descriptor,
+                                                         const WorkloadInfo& info) const = 0;
 };
 
 } //namespace armnn
diff --git a/src/backends/backendsCommon/test/CMakeLists.txt b/src/backends/backendsCommon/test/CMakeLists.txt
index 8107176..7edd93e 100644
--- a/src/backends/backendsCommon/test/CMakeLists.txt
+++ b/src/backends/backendsCommon/test/CMakeLists.txt
@@ -28,6 +28,8 @@
     OptimizedNetworkTests.cpp
     PermuteTestImpl.hpp
     Pooling2dTestImpl.hpp
+    PreCompiledTestImpl.cpp
+    PreCompiledTestImpl.hpp
     QuantizeHelper.hpp
     ReshapeTestImpl.hpp
     RuntimeTestImpl.hpp
diff --git a/src/backends/backendsCommon/test/DebugTestImpl.hpp b/src/backends/backendsCommon/test/DebugTestImpl.hpp
index e0f8a35..d112054 100644
--- a/src/backends/backendsCommon/test/DebugTestImpl.hpp
+++ b/src/backends/backendsCommon/test/DebugTestImpl.hpp
@@ -64,15 +64,15 @@
     outputHandle->Allocate();
 
     CopyDataToITensorHandle(inputHandle.get(), input.data());
-  
+
     std::ostringstream oss;
     std::streambuf* coutStreambuf = std::cout.rdbuf();
     std::cout.rdbuf(oss.rdbuf());
 
     ExecuteWorkload(*workload, memoryManager);
-  
+
     std::cout.rdbuf(coutStreambuf);
-    
+
     BOOST_TEST(oss.str() == expectedStringOutput);
 
     CopyDataFromITensorHandle(ret.output.data(), outputHandle.get());
diff --git a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
index 78716ef..edc58cf 100644
--- a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
+++ b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
@@ -372,6 +372,8 @@
 
 DECLARE_LAYER_POLICY_2_PARAM(Pooling2d)
 
+DECLARE_LAYER_POLICY_2_PARAM(PreCompiled)
+
 DECLARE_LAYER_POLICY_1_PARAM(Division)
 
 DECLARE_LAYER_POLICY_2_PARAM(ResizeBilinear)
diff --git a/src/backends/backendsCommon/test/LayerTests.cpp b/src/backends/backendsCommon/test/LayerTests.cpp
old mode 100755
new mode 100644
index 8e4596b..0bf56e2
--- a/src/backends/backendsCommon/test/LayerTests.cpp
+++ b/src/backends/backendsCommon/test/LayerTests.cpp
@@ -37,6 +37,7 @@
 #include "StridedSliceTestImpl.hpp"
 #include "NormTestImpl.hpp"
 #include "PermuteTestImpl.hpp"
+#include "PreCompiledTestImpl.hpp"
 #include "LstmTestImpl.hpp"
 #include "ConvertFp16ToFp32TestImpl.hpp"
 #include "ConvertFp32ToFp16TestImpl.hpp"
@@ -8567,3 +8568,38 @@
 {
     return Debug1DTest<uint8_t>(workloadFactory, memoryManager);
 }
+
+LayerTestResult<uint8_t, 4> PreCompiledConvolution2dTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    return PreCompiledConvolution2dTestImpl(workloadFactory, memoryManager);
+}
+
+LayerTestResult<uint8_t, 4> PreCompiledConvolution2dStride2x2Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    return PreCompiledConvolution2dStride2x2TestImpl(workloadFactory, memoryManager);
+}
+
+LayerTestResult<uint8_t, 4> PreCompiledDepthwiseConvolution2dTest(
+    armnn::IWorkloadFactory & workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr & memoryManager)
+{
+    return PreCompiledDepthwiseConvolution2dTestImpl(workloadFactory, memoryManager);
+}
+
+LayerTestResult<uint8_t, 4> PreCompiledDepthwiseConvolution2dStride2x2Test(
+    armnn::IWorkloadFactory & workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr & memoryManager)
+{
+    return PreCompiledDepthwiseConvolution2dStride2x2TestImpl(workloadFactory, memoryManager);
+}
+
+LayerTestResult<uint8_t, 4> PreCompiledMaxPooling2dTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    return PreCompiledMaxPooling2dTestImpl(workloadFactory, memoryManager);
+}
diff --git a/src/backends/backendsCommon/test/LayerTests.hpp b/src/backends/backendsCommon/test/LayerTests.hpp
index 98c0806..744470d 100644
--- a/src/backends/backendsCommon/test/LayerTests.hpp
+++ b/src/backends/backendsCommon/test/LayerTests.hpp
@@ -1317,3 +1317,55 @@
 LayerTestResult<uint8_t, 1> Debug1DUint8Test(
     armnn::IWorkloadFactory& workloadFactory,
     const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<uint8_t, 4> PreCompiledConvolution2dTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<uint8_t, 4> PreCompiledConvolution2dStride2x2Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<uint8_t, 4> PreCompiledDepthwiseConvolution2dTest(
+    armnn::IWorkloadFactory & workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr & memoryManager);
+
+LayerTestResult<uint8_t, 4> PreCompiledDepthwiseConvolution2dStride2x2Test(
+    armnn::IWorkloadFactory & workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr & memoryManager);
+
+LayerTestResult<uint8_t, 4> PreCompiledMaxPooling2dTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<float, 4> Debug4DFloat32Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<float, 3> Debug3DFloat32Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<float, 2> Debug2DFloat32Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<float, 1> Debug1DFloat32Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<uint8_t, 4> Debug4DUint8Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<uint8_t, 3> Debug3DUint8Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<uint8_t, 2> Debug2DUint8Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<uint8_t, 1> Debug1DUint8Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
diff --git a/src/backends/backendsCommon/test/PreCompiledTestImpl.cpp b/src/backends/backendsCommon/test/PreCompiledTestImpl.cpp
new file mode 100644
index 0000000..5a2bba1
--- /dev/null
+++ b/src/backends/backendsCommon/test/PreCompiledTestImpl.cpp
@@ -0,0 +1,491 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "PreCompiledTestImpl.hpp"
+
+#include "TensorCopyUtils.hpp"
+
+#include <Graph.hpp>
+#include <Network.hpp>
+#include <Runtime.hpp>
+
+#include <armnn/Exceptions.hpp>
+#include <armnn/INetwork.hpp>
+
+#include <test/TensorHelpers.hpp>
+
+#include <backendsCommon/WorkloadFactory.hpp>
+
+#include <boost/polymorphic_pointer_cast.hpp>
+
+using namespace armnn;
+
+namespace
+{
+
+template<typename ConvolutionDescriptor>
+struct PreCompiledConvolutionHelper
+{
+};
+
+template<>
+struct PreCompiledConvolutionHelper<Convolution2dDescriptor>
+{
+    static IConnectableLayer* AddConvolutionLayerToNetwork(
+        Network& network,
+        Convolution2dDescriptor descriptor,
+        const ConstTensor& weights,
+        const ConstTensor& biases)
+    {
+        return network.AddConvolution2dLayer(descriptor, weights, biases, "convolution");
+    }
+};
+
+template<>
+struct PreCompiledConvolutionHelper<DepthwiseConvolution2dDescriptor>
+{
+    static IConnectableLayer* AddConvolutionLayerToNetwork(
+        Network& network,
+        DepthwiseConvolution2dDescriptor descriptor,
+        const ConstTensor& weights,
+        const ConstTensor& biases)
+    {
+        return network.AddDepthwiseConvolution2dLayer(descriptor, weights, biases, "depthwiseConvolution");
+    }
+};
+
+template<typename ConvolutionDescriptor>
+ConvolutionDescriptor CreateConvolutionDescriptor(unsigned int stride, unsigned int padding)
+{
+    ConvolutionDescriptor descriptor;
+
+    descriptor.m_StrideX     = stride;
+    descriptor.m_StrideY     = stride;
+    descriptor.m_PadLeft     = padding;
+    descriptor.m_PadRight    = padding;
+    descriptor.m_PadTop      = padding;
+    descriptor.m_PadBottom   = padding;
+    descriptor.m_BiasEnabled = true;
+    descriptor.m_DataLayout  = DataLayout::NHWC;
+
+    return descriptor;
+}
+
+static std::vector<uint8_t> CreateIdentityConvolutionKernel(
+    unsigned int kernelSize, unsigned int channels)
+{
+    BOOST_ASSERT(kernelSize % 2 == 1); // kernelSize need to be an odd number
+
+    const unsigned int numElements = channels * (kernelSize * kernelSize);
+    std::vector<uint8_t> kernel(numElements, 0u);
+
+    unsigned int centerIndex = kernelSize / 2;
+    for(unsigned int y = 0u; y < kernelSize; y++)
+    {
+        for(unsigned int x = 0u; x < kernelSize; x++)
+        {
+            for(unsigned int channel = 0u; channel < channels; channel++)
+            {
+                if (x == centerIndex && y == centerIndex)
+                {
+                    const unsigned int flatIndex =
+                        (y * kernelSize * channels) + (x * channels) + channel;
+
+                    kernel[flatIndex] = 1u;
+                }
+            }
+        }
+    }
+
+    return kernel;
+}
+
+template<typename ConvolutionDescriptor>
+std::vector<uint8_t> GetIdentityConvolutionExpectedOutputData(
+    const TensorInfo& inputInfo,
+    const TensorInfo& outputInfo,
+    const ConvolutionDescriptor& descriptor,
+    const std::vector<uint8_t>& inputData)
+{
+    const unsigned int outputDataSize = outputInfo.GetNumElements();
+    std::vector<uint8_t> expectedOutputData(outputDataSize);
+
+    const unsigned int channels = outputInfo.GetShape()[3];
+    BOOST_ASSERT(channels == inputInfo.GetShape()[3]);
+
+    const unsigned int inputW  = inputInfo.GetShape()[2];
+
+    const unsigned int outputH = outputInfo.GetShape()[1];
+    const unsigned int outputW = outputInfo.GetShape()[2];
+
+    // Pick values from the input buffer, but after each iteration skip a number of
+    // rows and columns equal to the stride in the respective dimension
+    for (unsigned int inputY = 0, outputY = 0; outputY < outputH; inputY += descriptor.m_StrideY, outputY++)
+    {
+        for (unsigned int inputX = 0, outputX = 0; outputX < outputW; inputX += descriptor.m_StrideX, outputX++)
+        {
+            for (unsigned int channel = 0u; channel < channels; channel++)
+            {
+                const unsigned int inputIndex  =
+                    (inputY * inputW * channels) + (inputX * channels) + channel;
+                const unsigned int outputIndex =
+                    (outputY * outputW * channels) + (outputX * channels) + channel;
+
+                expectedOutputData[outputIndex] = inputData[inputIndex];
+            }
+        }
+    }
+
+    return expectedOutputData;
+}
+
+armnn::PreCompiledLayer* FindPreCompiledLayer(armnn::Graph& optimisedGraph)
+{
+    for (auto& layer : optimisedGraph)
+    {
+        if (layer->GetType() == armnn::LayerType::PreCompiled)
+        {
+            return boost::polymorphic_pointer_downcast<armnn::PreCompiledLayer>(layer);
+        }
+    }
+
+    // No pre-compiled layer found
+    return nullptr;
+}
+
+// NOTE: This only supports a single input and a single output
+LayerTestResult<uint8_t, 4> OptimiseAndRunNetwork(armnn::IWorkloadFactory& workloadFactory,
+                                                  Network& net,
+                                                  TensorInfo inputInfo,
+                                                  std::vector<uint8_t> inputData,
+                                                  TensorInfo outputInfo,
+                                                  std::vector<uint8_t> expectedOutputData)
+{
+    // Optimize the network for the backend supported by the factory
+    std::vector<BackendId> backends = {workloadFactory.GetBackendId()};
+    IRuntimePtr runtime(IRuntime::Create(IRuntime::CreationOptions()));
+    IOptimizedNetworkPtr optimizedNet = Optimize(net, backends, runtime->GetDeviceSpec(), OptimizerOptions());
+    if (!optimizedNet)
+    {
+        throw RuntimeException(std::string("Failed to optimize network for ") + std::string(backends[0]),
+                               CHECK_LOCATION());
+    }
+
+    // Find the pre-compiled layer in the optimised graph
+    Graph& optimisedGraph = static_cast<OptimizedNetwork*>(optimizedNet.get())->GetGraph();
+    PreCompiledLayer* preCompiledLayer = FindPreCompiledLayer(optimisedGraph);
+    if (!preCompiledLayer)
+    {
+        throw RuntimeException("Could not find pre-compiled layer in optimised graph", CHECK_LOCATION());
+    }
+
+    // Create the tensor handles
+    for (auto&& layer : optimisedGraph.TopologicalSort())
+    {
+        layer->CreateTensorHandles(optimisedGraph, workloadFactory);
+    }
+
+    // Create the pre-compiled workload
+    auto workload = preCompiledLayer->CreateWorkload(optimisedGraph, workloadFactory);
+
+    // Set the input data
+    boost::multi_array<uint8_t, 4> input = MakeTensor<uint8_t, 4>(inputInfo, inputData);
+    const QueueDescriptor& workloadData =
+        static_cast<BaseWorkload<PreCompiledQueueDescriptor>*>(workload.get())->GetData();
+    CopyDataToITensorHandle(workloadData.m_Inputs[0], &input[0][0][0][0]);
+
+    // Execute the workload
+    workload->Execute();
+
+    // Set the expected and actual outputs
+    LayerTestResult<uint8_t, 4> result(outputInfo);
+    result.outputExpected = MakeTensor<uint8_t, 4>(outputInfo, expectedOutputData);
+    CopyDataFromITensorHandle(&result.output[0][0][0][0], workloadData.m_Outputs[0]);
+    return result;
+}
+
+} // anonymous namespace
+
+template<typename ConvolutionDescriptor>
+LayerTestResult<uint8_t, 4> PreCompiledConvolution2dTestImpl(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    unsigned int inputSize,
+    unsigned int outputSize,
+    unsigned int channels,
+    unsigned int kernelSize,
+    const ConvolutionDescriptor& descriptor,
+    bool isDepthwiseConvolution = false)
+{
+    BOOST_ASSERT(descriptor.m_BiasEnabled == true);
+    BOOST_ASSERT(descriptor.m_DataLayout  == DataLayout::NHWC);
+
+    // Set up tensor shapes and infos
+    const TensorShape inputShape ({1, inputSize,  inputSize,  channels});
+    const TensorShape outputShape({1, outputSize, outputSize, channels});
+    const TensorShape kernelShape = isDepthwiseConvolution
+                                    // The format for the depthwise convolution is MIHW
+                                    ? TensorShape({1, channels, kernelSize, kernelSize})
+                                    // The format for the regular convolution depends on the layout of the inputs,
+                                    // in this case is NHWC
+                                    : TensorShape({1, kernelSize, kernelSize, channels});
+    const TensorShape biasesShape({1, 1, 1, channels});
+
+    // NOTE: inputScale * weightsScale / outputScale must be >= 0.0 and < 1.0
+    TensorInfo inputInfo(inputShape, DataType::QuantisedAsymm8, 1.0f, 0);
+    TensorInfo outputInfo(outputShape, DataType::QuantisedAsymm8, 2.0f, 0);
+    TensorInfo weightsInfo(kernelShape, DataType::QuantisedAsymm8, 1.0f, 0);
+    TensorInfo biasesInfo(biasesShape, DataType::Signed32, 1.0f, 0);
+
+    // Populate weight and bias data
+    std::vector<uint8_t> weightsData = CreateIdentityConvolutionKernel(kernelSize, channels);
+
+    // NOTE: We need to multiply the elements of the identity kernel by 2
+    // to compensate for the scaling factor
+    std::transform(weightsData.begin(), weightsData.end(), weightsData.begin(),
+        [](uint8_t w) -> uint8_t { return static_cast<uint8_t>(w * 2); });
+
+    const unsigned int biasDataSize = biasesInfo.GetNumElements();
+    std::vector<int32_t> biasesData(biasDataSize, 0);
+
+    // Construct network
+    Network network;
+    ConstTensor weights(weightsInfo, weightsData);
+    ConstTensor biases(biasesInfo, biasesData);
+
+    IConnectableLayer* const inputLayer = network.AddInputLayer(0, "input");
+
+    IConnectableLayer* const convolutionLayer =
+        PreCompiledConvolutionHelper<ConvolutionDescriptor>
+            ::AddConvolutionLayerToNetwork(network, descriptor, weights, biases);
+
+    IConnectableLayer* const outputLayer = network.AddOutputLayer(0, "output");
+
+    inputLayer->GetOutputSlot(0).Connect(convolutionLayer->GetInputSlot(0));
+    inputLayer->GetOutputSlot(0).SetTensorInfo(inputInfo);
+
+    convolutionLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
+    convolutionLayer->GetOutputSlot(0).SetTensorInfo(outputInfo);
+
+    // Generate input data: sequence [0, 1 .. 255]
+    const unsigned int inputDataSize = inputInfo.GetNumElements();
+    std::vector<uint8_t> inputData(inputDataSize);
+    std::iota(inputData.begin(), inputData.end(), 0);
+
+    // Set expected output
+    std::vector<uint8_t> expectedOutputData =
+        GetIdentityConvolutionExpectedOutputData(inputInfo,
+                                                 outputInfo,
+                                                 descriptor,
+                                                 inputData);
+
+    return OptimiseAndRunNetwork(workloadFactory,
+                                 network,
+                                 inputInfo,
+                                 inputData,
+                                 outputInfo,
+                                 expectedOutputData);
+}
+
+LayerTestResult<uint8_t, 4> PreCompiledConvolution2dTestImpl(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    const unsigned int inputSize  = 16;
+    const unsigned int outputSize = 16;
+    const unsigned int channels   = 1;
+    const unsigned int kernelSize = 3;
+    const unsigned int stride     = 1;
+    const unsigned int padding    = 1;
+
+    Convolution2dDescriptor descriptor =
+        CreateConvolutionDescriptor<Convolution2dDescriptor>(stride, padding);
+
+    return PreCompiledConvolution2dTestImpl(workloadFactory,
+                                            memoryManager,
+                                            inputSize,
+                                            outputSize,
+                                            channels,
+                                            kernelSize,
+                                            descriptor);
+}
+
+LayerTestResult<uint8_t, 4> PreCompiledConvolution2dStride2x2TestImpl(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    const unsigned int inputSize  = 16;
+    const unsigned int outputSize = 8;
+    const unsigned int channels   = 1;
+    const unsigned int kernelSize = 3;
+    const unsigned int stride     = 2;
+    const unsigned int padding    = 1;
+
+    Convolution2dDescriptor descriptor =
+        CreateConvolutionDescriptor<Convolution2dDescriptor>(stride, padding);
+
+    return PreCompiledConvolution2dTestImpl(workloadFactory,
+                                            memoryManager,
+                                            inputSize,
+                                            outputSize,
+                                            channels,
+                                            kernelSize,
+                                            descriptor);
+}
+
+LayerTestResult<uint8_t, 4> PreCompiledDepthwiseConvolution2dTestImpl(
+    armnn::IWorkloadFactory & workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr & memoryManager)
+{
+    const unsigned int inputSize  = 16;
+    const unsigned int outputSize = 16;
+    const unsigned int channels   = 3;
+    const unsigned int kernelSize = 1;
+    const unsigned int stride     = 1;
+    const unsigned int padding    = 0;
+
+    DepthwiseConvolution2dDescriptor descriptor =
+        CreateConvolutionDescriptor<DepthwiseConvolution2dDescriptor>(stride, padding);
+
+    return PreCompiledConvolution2dTestImpl(workloadFactory,
+                                            memoryManager,
+                                            inputSize,
+                                            outputSize,
+                                            channels,
+                                            kernelSize,
+                                            descriptor,
+                                            true);
+}
+
+LayerTestResult<uint8_t, 4> PreCompiledDepthwiseConvolution2dStride2x2TestImpl(
+    armnn::IWorkloadFactory & workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr & memoryManager)
+{
+    const unsigned int inputSize  = 16;
+    const unsigned int outputSize = 8;
+    const unsigned int channels   = 3;
+    const unsigned int kernelSize = 3;
+    const unsigned int stride     = 2;
+    const unsigned int padding    = 1;
+
+    DepthwiseConvolution2dDescriptor descriptor =
+        CreateConvolutionDescriptor<DepthwiseConvolution2dDescriptor>(stride, padding);
+
+    return PreCompiledConvolution2dTestImpl(workloadFactory,
+                                            memoryManager,
+                                            inputSize,
+                                            outputSize,
+                                            channels,
+                                            kernelSize,
+                                            descriptor);
+}
+
+LayerTestResult<uint8_t, 4> PreCompiledMaxPooling2dTestImpl(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    // Pooling cannot be run in isolation, it must be fused with the previous layer, e.g. Convolution2d.
+
+    // Set up the Convolution descriptor
+    Convolution2dDescriptor convDescriptor;
+    convDescriptor.m_StrideX = 1;
+    convDescriptor.m_StrideY = 1;
+    convDescriptor.m_BiasEnabled = true;
+    convDescriptor.m_DataLayout  = DataLayout::NHWC;
+
+    // Set up the Convolution weights
+    TensorInfo weightsInfo(TensorShape({16, 1, 1, 16}), DataType::QuantisedAsymm8, 2.0f, 0);
+    const unsigned int weightsDataSize = weightsInfo.GetNumElements();
+    std::vector<uint8_t> weightsData(weightsDataSize);
+    for (unsigned int i = 0; i < 16; ++i)
+    {
+        for (unsigned int j = 0; j < 16; ++j)
+        {
+            weightsData[(i * 16) + j] = i == j ? 1.0f : 0.0f;
+        }
+    }
+    ConstTensor weights(weightsInfo, weightsData);
+
+    // Set up the Convolution biases
+    TensorInfo biasInfo(TensorShape({1, 1, 1, 16}), DataType::Signed32, 1.0f * 2.0f, 0);
+    const unsigned int biasDataSize = biasInfo.GetNumElements();
+    std::vector<int32_t> biasData(biasDataSize, 0);
+    ConstTensor biases(biasInfo, biasData);
+
+    // Set up the Convolution input
+    TensorInfo inputInfo(TensorShape({1, 16, 16, 16 }), DataType::QuantisedAsymm8, 1.0f, 0);
+    const unsigned int inputDataSize = inputInfo.GetNumElements();
+    std::vector<uint8_t> inputData(inputDataSize);
+    for (unsigned int i = 0; i < inputDataSize; ++i)
+    {
+        inputData[i] = boost::numeric_cast<uint8_t>((i * 4) % 250);
+    }
+
+    // Set up the Convolution output / Pooling input info
+    TensorInfo convOutputInfo(TensorShape({1, 16, 16, 16 }), DataType::QuantisedAsymm8, 4.0f, 0);
+
+    // Set up the Pooling descriptor
+    Pooling2dDescriptor poolDescriptor;
+    poolDescriptor.m_PoolType = PoolingAlgorithm::Max;
+    poolDescriptor.m_PoolWidth = 2;
+    poolDescriptor.m_PoolHeight = 2;
+    poolDescriptor.m_StrideX = 2;
+    poolDescriptor.m_StrideY = 2;
+    poolDescriptor.m_PaddingMethod = PaddingMethod::Exclude;
+    poolDescriptor.m_DataLayout = DataLayout::NHWC;
+
+    // Set the expected output from the Pooling layer
+    TensorInfo outputInfo(TensorShape({1, 8, 8, 16 }), DataType::QuantisedAsymm8, 4.0f, 0);
+    const unsigned int outputDataSize = outputInfo.GetNumElements();
+    std::vector<uint8_t> expectedOutputData(outputDataSize);
+    // The Maxpooling inputs are the Convolution outputs, i.e. (Convolution inputs / 2) after scale adjustments
+    // Maxpooling selects the max value in each pool from its inputs and our pool size is 2x2
+    for (unsigned int channel = 0; channel < 16; ++channel)
+    {
+        for (unsigned int row = 0; row < 8; ++row)
+        {
+            for (unsigned int column = 0; column < 8; ++column)
+            {
+                // The input and output data indexes are calculated for NHWC data layout.
+                // Output index: (row * columns * channels) + (column * channels) + channel
+                auto outIndex = (row * 8 * 16) + (column * 16) + channel;
+                // Input index: (row * strideY * columns * channels) + (column * strideX * channels) + channel
+                //      and we take 4 entries for the 2x2 pool
+                auto in0Index = ((row * 2) * 16 * 16) + ((column * 2) * 16) + channel;
+                auto in1Index = ((row * 2) * 16 * 16) + (((column * 2) + 1) * 16) + channel;
+                auto in2Index = (((row * 2) + 1) * 16 * 16) + ((column * 2) * 16) + channel;
+                auto in3Index = (((row * 2) + 1) * 16 * 16) + (((column * 2) + 1) * 16) + channel;
+                // output value is the maximum of the input pool values, adjusted for the quantization scale change
+                auto maxIn = std::max<uint8_t>({inputData[in0Index],
+                                                inputData[in1Index],
+                                                inputData[in2Index],
+                                                inputData[in3Index]});
+                expectedOutputData[outIndex] = maxIn / 2;
+            }
+        }
+    }
+
+    // Construct the network
+    Network net;
+    IConnectableLayer* const inputLayer   = net.AddInputLayer(0, "input");
+    IConnectableLayer* const convLayer    = net.AddConvolution2dLayer(convDescriptor, weights, biases, "conv");
+    IConnectableLayer* const poolingLayer = net.AddPooling2dLayer(poolDescriptor, "pooling2d");
+    IConnectableLayer* const outputLayer  = net.AddOutputLayer(0, "output");
+
+    // Connect the layers
+    inputLayer->GetOutputSlot(0).Connect(convLayer->GetInputSlot(0));
+    inputLayer->GetOutputSlot(0).SetTensorInfo(inputInfo);
+    convLayer->GetOutputSlot(0).Connect(poolingLayer->GetInputSlot(0));
+    convLayer->GetOutputSlot(0).SetTensorInfo(convOutputInfo);
+    poolingLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
+    poolingLayer->GetOutputSlot(0).SetTensorInfo(outputInfo);
+
+    return OptimiseAndRunNetwork(workloadFactory,
+                                 net,
+                                 inputInfo,
+                                 inputData,
+                                 outputInfo,
+                                 expectedOutputData);
+}
diff --git a/src/backends/backendsCommon/test/PreCompiledTestImpl.hpp b/src/backends/backendsCommon/test/PreCompiledTestImpl.hpp
new file mode 100644
index 0000000..f4e78b6
--- /dev/null
+++ b/src/backends/backendsCommon/test/PreCompiledTestImpl.hpp
@@ -0,0 +1,27 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+#pragma once
+
+#include "LayerTests.hpp"
+
+LayerTestResult<uint8_t, 4> PreCompiledConvolution2dTestImpl(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<uint8_t, 4> PreCompiledConvolution2dStride2x2TestImpl(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
+
+LayerTestResult<uint8_t, 4> PreCompiledDepthwiseConvolution2dTestImpl(
+    armnn::IWorkloadFactory & workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr & memoryManager);
+
+LayerTestResult<uint8_t, 4> PreCompiledDepthwiseConvolution2dStride2x2TestImpl(
+    armnn::IWorkloadFactory & workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr & memoryManager);
+
+LayerTestResult<uint8_t, 4> PreCompiledMaxPooling2dTestImpl(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager);
diff --git a/src/backends/cl/ClBackend.cpp b/src/backends/cl/ClBackend.cpp
index 2b82c18..2f9dfa9 100644
--- a/src/backends/cl/ClBackend.cpp
+++ b/src/backends/cl/ClBackend.cpp
@@ -63,6 +63,12 @@
     return IBackendContextPtr{new ClBackendContext{options}};
 }
 
+IBackendInternal::ISubGraphConverterPtr ClBackend::CreateSubGraphConverter(
+    const std::shared_ptr<SubGraph>& subGraph) const
+{
+    return ISubGraphConverterPtr{};
+}
+
 IBackendInternal::Optimizations ClBackend::GetOptimizations() const
 {
     return Optimizations{};
diff --git a/src/backends/cl/ClBackend.hpp b/src/backends/cl/ClBackend.hpp
index ef98da0..84b5b9a 100644
--- a/src/backends/cl/ClBackend.hpp
+++ b/src/backends/cl/ClBackend.hpp
@@ -25,6 +25,9 @@
 
     IBackendInternal::IBackendContextPtr CreateBackendContext(const IRuntime::CreationOptions&) const override;
 
+    IBackendInternal::ISubGraphConverterPtr CreateSubGraphConverter(
+        const std::shared_ptr<SubGraph>& subGraph) const override;
+
     IBackendInternal::Optimizations GetOptimizations() const override;
     IBackendInternal::ILayerSupportSharedPtr GetLayerSupport() const override;
 };
diff --git a/src/backends/cl/ClWorkloadFactory.cpp b/src/backends/cl/ClWorkloadFactory.cpp
index af47f65..28011cf 100644
--- a/src/backends/cl/ClWorkloadFactory.cpp
+++ b/src/backends/cl/ClWorkloadFactory.cpp
@@ -356,4 +356,10 @@
     return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info);
 }
 
+std::unique_ptr<IWorkload> ClWorkloadFactory::CreatePreCompiled(const PreCompiledQueueDescriptor& descriptor,
+                                                                const WorkloadInfo& info) const
+{
+    return MakeWorkload<NullWorkload, NullWorkload>(descriptor, info);
+}
+
 } // namespace armnn
diff --git a/src/backends/cl/ClWorkloadFactory.hpp b/src/backends/cl/ClWorkloadFactory.hpp
index 85cbd91..286e897 100644
--- a/src/backends/cl/ClWorkloadFactory.hpp
+++ b/src/backends/cl/ClWorkloadFactory.hpp
@@ -150,6 +150,9 @@
     virtual std::unique_ptr<IWorkload> CreateRsqrt(const RsqrtQueueDescriptor& descriptor,
                                                    const WorkloadInfo& info) const override;
 
+    virtual std::unique_ptr<IWorkload> CreatePreCompiled(const PreCompiledQueueDescriptor& descriptor,
+                                                         const WorkloadInfo& info) const override;
+
 private:
     template<typename FloatWorkload, typename Uint8Workload, typename QueueDescriptorType, typename... Args>
     static std::unique_ptr<IWorkload> MakeWorkload(const QueueDescriptorType& descriptor,
diff --git a/src/backends/cl/test/ClMemCopyTests.cpp b/src/backends/cl/test/ClMemCopyTests.cpp
index 93d8dd5..3cd9af7 100644
--- a/src/backends/cl/test/ClMemCopyTests.cpp
+++ b/src/backends/cl/test/ClMemCopyTests.cpp
@@ -17,25 +17,29 @@
 
 BOOST_AUTO_TEST_CASE(CopyBetweenCpuAndGpu)
 {
-    LayerTestResult<float, 4> result = MemCopyTest<armnn::RefWorkloadFactory, armnn::ClWorkloadFactory>(false);
+    LayerTestResult<float, 4> result =
+        MemCopyTest<armnn::RefWorkloadFactory, armnn::ClWorkloadFactory, armnn::DataType::Float32>(false);
     BOOST_TEST(CompareTensors(result.output, result.outputExpected));
 }
 
 BOOST_AUTO_TEST_CASE(CopyBetweenGpuAndCpu)
 {
-    LayerTestResult<float, 4> result = MemCopyTest<armnn::ClWorkloadFactory, armnn::RefWorkloadFactory>(false);
+    LayerTestResult<float, 4> result =
+        MemCopyTest<armnn::ClWorkloadFactory, armnn::RefWorkloadFactory, armnn::DataType::Float32>(false);
     BOOST_TEST(CompareTensors(result.output, result.outputExpected));
 }
 
 BOOST_AUTO_TEST_CASE(CopyBetweenCpuAndGpuWithSubtensors)
 {
-    LayerTestResult<float, 4> result = MemCopyTest<armnn::RefWorkloadFactory, armnn::ClWorkloadFactory>(true);
+    LayerTestResult<float, 4> result =
+        MemCopyTest<armnn::RefWorkloadFactory, armnn::ClWorkloadFactory, armnn::DataType::Float32>(true);
     BOOST_TEST(CompareTensors(result.output, result.outputExpected));
 }
 
 BOOST_AUTO_TEST_CASE(CopyBetweenGpuAndCpuWithSubtensors)
 {
-    LayerTestResult<float, 4> result = MemCopyTest<armnn::ClWorkloadFactory, armnn::RefWorkloadFactory>(true);
+    LayerTestResult<float, 4> result =
+        MemCopyTest<armnn::ClWorkloadFactory, armnn::RefWorkloadFactory, armnn::DataType::Float32>(true);
     BOOST_TEST(CompareTensors(result.output, result.outputExpected));
 }
 
diff --git a/src/backends/neon/NeonBackend.cpp b/src/backends/neon/NeonBackend.cpp
index fd2b766..ce97a1d 100644
--- a/src/backends/neon/NeonBackend.cpp
+++ b/src/backends/neon/NeonBackend.cpp
@@ -63,6 +63,12 @@
     return IBackendContextPtr{};
 }
 
+IBackendInternal::ISubGraphConverterPtr NeonBackend::CreateSubGraphConverter(
+    const std::shared_ptr<SubGraph>& subGraph) const
+{
+    return ISubGraphConverterPtr{};
+}
+
 IBackendInternal::Optimizations NeonBackend::GetOptimizations() const
 {
     return Optimizations{};
diff --git a/src/backends/neon/NeonBackend.hpp b/src/backends/neon/NeonBackend.hpp
index 127a5a4..3b1d186 100644
--- a/src/backends/neon/NeonBackend.hpp
+++ b/src/backends/neon/NeonBackend.hpp
@@ -25,6 +25,9 @@
 
     IBackendInternal::IBackendContextPtr CreateBackendContext(const IRuntime::CreationOptions&) const override;
 
+    IBackendInternal::ISubGraphConverterPtr CreateSubGraphConverter(
+        const std::shared_ptr<SubGraph>& subGraph) const override;
+
     IBackendInternal::Optimizations GetOptimizations() const override;
     IBackendInternal::ILayerSupportSharedPtr GetLayerSupport() const override;
 };
diff --git a/src/backends/neon/NeonWorkloadFactory.cpp b/src/backends/neon/NeonWorkloadFactory.cpp
index 65093fb..3728c86 100644
--- a/src/backends/neon/NeonWorkloadFactory.cpp
+++ b/src/backends/neon/NeonWorkloadFactory.cpp
@@ -324,4 +324,10 @@
     return MakeWorkloadHelper<NullWorkload, NullWorkload>(descriptor, info);
 }
 
+std::unique_ptr<IWorkload> NeonWorkloadFactory::CreatePreCompiled(const PreCompiledQueueDescriptor& descriptor,
+                                                                  const WorkloadInfo& info) const
+{
+    return MakeWorkloadHelper<NullWorkload, NullWorkload>(descriptor, info);
+}
+
 } // namespace armnn
diff --git a/src/backends/neon/NeonWorkloadFactory.hpp b/src/backends/neon/NeonWorkloadFactory.hpp
index 9a63088..68317ed 100644
--- a/src/backends/neon/NeonWorkloadFactory.hpp
+++ b/src/backends/neon/NeonWorkloadFactory.hpp
@@ -151,6 +151,9 @@
     virtual std::unique_ptr<IWorkload> CreateRsqrt(const RsqrtQueueDescriptor& descriptor,
                                                    const WorkloadInfo& info) const override;
 
+    virtual std::unique_ptr<IWorkload> CreatePreCompiled(const PreCompiledQueueDescriptor& descriptor,
+                                                         const WorkloadInfo& info) const override;
+
 private:
     mutable std::shared_ptr<NeonMemoryManager> m_MemoryManager;
 };
diff --git a/src/backends/neon/test/NeonMemCopyTests.cpp b/src/backends/neon/test/NeonMemCopyTests.cpp
index f6699a6..dbe1f8d 100644
--- a/src/backends/neon/test/NeonMemCopyTests.cpp
+++ b/src/backends/neon/test/NeonMemCopyTests.cpp
@@ -18,25 +18,29 @@
 
 BOOST_AUTO_TEST_CASE(CopyBetweenCpuAndNeon)
 {
-    LayerTestResult<float, 4> result = MemCopyTest<armnn::RefWorkloadFactory, armnn::NeonWorkloadFactory>(false);
+    LayerTestResult<float, 4> result =
+        MemCopyTest<armnn::RefWorkloadFactory, armnn::NeonWorkloadFactory, armnn::DataType::Float32>(false);
     BOOST_TEST(CompareTensors(result.output, result.outputExpected));
 }
 
 BOOST_AUTO_TEST_CASE(CopyBetweenNeonAndCpu)
 {
-    LayerTestResult<float, 4> result = MemCopyTest<armnn::NeonWorkloadFactory, armnn::RefWorkloadFactory>(false);
+    LayerTestResult<float, 4> result =
+        MemCopyTest<armnn::NeonWorkloadFactory, armnn::RefWorkloadFactory, armnn::DataType::Float32>(false);
     BOOST_TEST(CompareTensors(result.output, result.outputExpected));
 }
 
 BOOST_AUTO_TEST_CASE(CopyBetweenCpuAndNeonWithSubtensors)
 {
-    LayerTestResult<float, 4> result = MemCopyTest<armnn::RefWorkloadFactory, armnn::NeonWorkloadFactory>(true);
+    LayerTestResult<float, 4> result =
+        MemCopyTest<armnn::RefWorkloadFactory, armnn::NeonWorkloadFactory, armnn::DataType::Float32>(true);
     BOOST_TEST(CompareTensors(result.output, result.outputExpected));
 }
 
 BOOST_AUTO_TEST_CASE(CopyBetweenNeonAndCpuWithSubtensors)
 {
-    LayerTestResult<float, 4> result = MemCopyTest<armnn::NeonWorkloadFactory, armnn::RefWorkloadFactory>(true);
+    LayerTestResult<float, 4> result =
+        MemCopyTest<armnn::NeonWorkloadFactory, armnn::RefWorkloadFactory, armnn::DataType::Float32>(true);
     BOOST_TEST(CompareTensors(result.output, result.outputExpected));
 }
 
diff --git a/src/backends/reference/RefBackend.cpp b/src/backends/reference/RefBackend.cpp
index 8f5e9c4..e4f468c 100644
--- a/src/backends/reference/RefBackend.cpp
+++ b/src/backends/reference/RefBackend.cpp
@@ -56,6 +56,12 @@
     return IMemoryManagerUniquePtr{};
 }
 
+IBackendInternal::ISubGraphConverterPtr RefBackend::CreateSubGraphConverter(
+    const std::shared_ptr<SubGraph>& subGraph) const
+{
+    return ISubGraphConverterPtr{};
+}
+
 IBackendInternal::Optimizations RefBackend::GetOptimizations() const
 {
     return Optimizations{};
diff --git a/src/backends/reference/RefBackend.hpp b/src/backends/reference/RefBackend.hpp
index 1a0aef5..5136622 100644
--- a/src/backends/reference/RefBackend.hpp
+++ b/src/backends/reference/RefBackend.hpp
@@ -25,6 +25,9 @@
 
     IBackendInternal::IBackendContextPtr CreateBackendContext(const IRuntime::CreationOptions&) const override;
 
+    IBackendInternal::ISubGraphConverterPtr CreateSubGraphConverter(
+        const std::shared_ptr<SubGraph>& subGraph) const override;
+
     IBackendInternal::Optimizations GetOptimizations() const override;
     IBackendInternal::ILayerSupportSharedPtr GetLayerSupport() const override;
 };
diff --git a/src/backends/reference/RefWorkloadFactory.cpp b/src/backends/reference/RefWorkloadFactory.cpp
index 7929363..361a3f1 100644
--- a/src/backends/reference/RefWorkloadFactory.cpp
+++ b/src/backends/reference/RefWorkloadFactory.cpp
@@ -318,4 +318,10 @@
     return MakeWorkload<RefRsqrtFloat32Workload, NullWorkload>(descriptor, info);
 }
 
-} // namespace armnn
\ No newline at end of file
+std::unique_ptr<IWorkload> RefWorkloadFactory::CreatePreCompiled(const PreCompiledQueueDescriptor& descriptor,
+                                                                 const WorkloadInfo& info) const
+{
+    return nullptr;
+}
+
+} // namespace armnn
diff --git a/src/backends/reference/RefWorkloadFactory.hpp b/src/backends/reference/RefWorkloadFactory.hpp
index f6707f5..432ac72 100644
--- a/src/backends/reference/RefWorkloadFactory.hpp
+++ b/src/backends/reference/RefWorkloadFactory.hpp
@@ -167,6 +167,10 @@
 
     virtual std::unique_ptr<IWorkload> CreateRsqrt(const RsqrtQueueDescriptor& descriptor,
                                                    const WorkloadInfo& info) const override;
+
+    virtual std::unique_ptr<IWorkload> CreatePreCompiled(const PreCompiledQueueDescriptor& descriptor,
+                                                         const WorkloadInfo& info) const override;
+
 private:
 
     template <typename F32Workload, typename U8Workload, typename QueueDescriptorType>
diff --git a/src/backends/reference/workloads/Debug.cpp b/src/backends/reference/workloads/Debug.cpp
index dfcbbd8..cc83c7b 100644
--- a/src/backends/reference/workloads/Debug.cpp
+++ b/src/backends/reference/workloads/Debug.cpp
@@ -98,4 +98,4 @@
                              const DebugDescriptor& descriptor,
                              const uint8_t* inputData,
                              uint8_t* outputData);
-} //namespace armnn
+} // namespace armnn
diff --git a/src/backends/reference/workloads/RefDebugWorkload.cpp b/src/backends/reference/workloads/RefDebugWorkload.cpp
index 17eb8fc..d9a47c0 100644
--- a/src/backends/reference/workloads/RefDebugWorkload.cpp
+++ b/src/backends/reference/workloads/RefDebugWorkload.cpp
@@ -2,11 +2,12 @@
 // Copyright © 2017 Arm Ltd. All rights reserved.
 // SPDX-License-Identifier: MIT
 //
+
 #include "RefDebugWorkload.hpp"
 #include "Debug.hpp"
-
 #include "RefWorkloadUtils.hpp"
-#include "TypeUtils.hpp"
+
+#include <TypeUtils.hpp>
 
 namespace armnn
 {
@@ -30,4 +31,4 @@
 template class RefDebugWorkload<DataType::Float32>;
 template class RefDebugWorkload<DataType::QuantisedAsymm8>;
 
-} //namespace armnn
+} // namespace armnn
diff --git a/src/backends/reference/workloads/RefDebugWorkload.hpp b/src/backends/reference/workloads/RefDebugWorkload.hpp
index a1231f9..c1a3e26 100644
--- a/src/backends/reference/workloads/RefDebugWorkload.hpp
+++ b/src/backends/reference/workloads/RefDebugWorkload.hpp
@@ -2,12 +2,13 @@
 // Copyright © 2017 Arm Ltd. All rights reserved.
 // SPDX-License-Identifier: MIT
 //
+
 #pragma once
 
-#include <backendsCommon/Workload.hpp>
-
 #include <armnn/TypesUtils.hpp>
 
+#include <backendsCommon/Workload.hpp>
+
 namespace armnn
 {
 
@@ -30,4 +31,4 @@
 using RefDebugFloat32Workload = RefDebugWorkload<DataType::Float32>;
 using RefDebugUint8Workload = RefDebugWorkload<DataType::QuantisedAsymm8>;
 
-} //namespace armnn
+} // namespace armnn