Revert "MLCE-1093 Reshape and concat invalid results"

This reverts commit 4980e21193f0a14fef084a7f4b4197392f3c0845.

Reason for revert: Android Build for v82a failed due to schema not re-generating

Change-Id: Ic19cf471b487f321c97ff837d36526512fb12fa4

14 files changed, 292 insertions, 98 deletions

diff --git a/delegate/classic/src/Split.hpp b/delegate/classic/src/Split.hpp
index 57b7f8074e..aaa610259f 100644
--- a/delegate/classic/src/Split.hpp
+++ b/delegate/classic/src/Split.hpp
@@ -107,8 +107,6 @@ TfLiteStatus VisitSplitOperator(DelegateData& delegateData,
     splitterDimSizes[splitDim] /= numSplits;
 
     armnn::SplitterDescriptor splitDescriptor(numSplits, inputDimSize);
-    splitDescriptor.SetAxis(axis);
-
     for (unsigned int j = 0; j < numSplits; ++j)
     {
         // Set the size of the views.
@@ -303,8 +301,6 @@ TfLiteStatus VisitSplitVOperator(DelegateData& delegateData,
     }
 
     armnn::SplitterDescriptor splitDescriptor(numSplits, inputDimSize);
-    splitDescriptor.SetAxis(axis);
-
     unsigned int accumSplit = 0;
     for (unsigned int j = 0; j < numSplits; ++j)
     {
diff --git a/delegate/classic/src/Unpack.hpp b/delegate/classic/src/Unpack.hpp
index b3336ec990..2cd32564c1 100644
--- a/delegate/classic/src/Unpack.hpp
+++ b/delegate/classic/src/Unpack.hpp
@@ -96,8 +96,6 @@ TfLiteStatus VisitUnpackOperator(DelegateData& delegateData,
     unpackDimSizes[unpackAxis] /= unpackNum;
 
     armnn::SplitterDescriptor splitDesc(unpackNum, static_cast<unsigned int>(unpackDimSizes.size()));
-    splitDesc.SetAxis(unpackAxis);
-
     for (unsigned int j = 0; j < unpackNum; ++j)
     {
         // Set the size of the views.
diff --git a/delegate/opaque/src/Split.hpp b/delegate/opaque/src/Split.hpp
index 199f46b126..2dbfa602fb 100644
--- a/delegate/opaque/src/Split.hpp
+++ b/delegate/opaque/src/Split.hpp
@@ -134,8 +134,6 @@ TfLiteStatus VisitSplitOperator(DelegateData& delegateData,
     splitterDimSizes[splitDim] /= numSplits;
 
     armnn::SplitterDescriptor splitDescriptor(numSplits, inputDimSize);
-    splitDescriptor.SetAxis(axis);
-
     for (int j = 0; j < numSplits; ++j)
     {
         // Set the size of the views.
@@ -369,9 +367,7 @@ TfLiteStatus VisitSplitVOperator(DelegateData& delegateData,
     }
 
     armnn::SplitterDescriptor splitDescriptor(numSplits, inputDimSize);
-    splitDescriptor.SetAxis(axis);
     unsigned int accumSplit = 0;
-
     for (int j = 0; j < numSplits; ++j)
     {
         unsigned int splitSize = armnn::numeric_cast<unsigned int>(splitsTensorData[j]);
diff --git a/delegate/opaque/src/Unpack.hpp b/delegate/opaque/src/Unpack.hpp
index 525529ff7b..0956d1688e 100644
--- a/delegate/opaque/src/Unpack.hpp
+++ b/delegate/opaque/src/Unpack.hpp
@@ -96,8 +96,6 @@ TfLiteStatus VisitUnpackOperator(DelegateData& delegateData,
     unpackDimSizes[unpackAxis] /= unpackNum;
 
     armnn::SplitterDescriptor splitDesc(unpackNum, static_cast<unsigned int>(unpackDimSizes.size()));
-    splitDesc.SetAxis(unpackAxis);
-
     for (unsigned int j = 0; j < unpackNum; ++j)
     {
         // Set the size of the views.
diff --git a/src/armnn/Descriptors.cpp b/src/armnn/Descriptors.cpp
index 5e4628bd77..e6374aea8f 100644
--- a/src/armnn/Descriptors.cpp
+++ b/src/armnn/Descriptors.cpp
@@ -216,16 +216,12 @@ void OriginsDescriptor::ReorderOrigins(unsigned int*  newOrdering, unsigned int
 ViewsDescriptor::ViewsDescriptor()
 : m_Origins()
 , m_ViewSizes(nullptr)
-, m_IsAxisSet(false)
-, m_Axis(0)
 {}
 
 ViewsDescriptor::ViewsDescriptor(uint32_t numViews, uint32_t numDimensions /*= 4*/)
     : m_Origins(numViews, numDimensions)
     , m_ViewSizes(numViews > 0 && numDimensions > 0 ?
                       new uint32_t *[numViews]() : nullptr)
-    , m_IsAxisSet(false)
-    , m_Axis(0)
 {
     if (m_ViewSizes)
     {
@@ -240,8 +236,6 @@ ViewsDescriptor::ViewsDescriptor(const ViewsDescriptor& other)
     : m_Origins(other.m_Origins)
     , m_ViewSizes(other.GetNumViews() > 0 && other.GetNumDimensions() > 0 ?
                       new uint32_t *[other.GetNumViews()]() : nullptr)
-    , m_IsAxisSet(other.m_IsAxisSet)
-    , m_Axis(other.m_Axis)
 {
     if (m_ViewSizes)
     {
@@ -367,8 +361,6 @@ void swap(ViewsDescriptor& first, ViewsDescriptor& second)
     using std::swap;
     swap(first.m_Origins, second.m_Origins);
     swap(first.m_ViewSizes, second.m_ViewSizes);
-    swap(first.m_IsAxisSet, second.m_IsAxisSet);
-    swap(first.m_Axis, second.m_Axis);
 }
 
 void ViewsDescriptor::SetAxis(int32_t axis)
diff --git a/src/armnnDeserializer/Deserializer.cpp b/src/armnnDeserializer/Deserializer.cpp
index 8ca1e83dfe..eb77f92842 100644
--- a/src/armnnDeserializer/Deserializer.cpp
+++ b/src/armnnDeserializer/Deserializer.cpp
@@ -3228,11 +3228,6 @@ void IDeserializer::DeserializerImpl::ParseSplitter(GraphPtr graph, unsigned int
         }
     }
 
-    if (flatBufferViewsDescriptor->hasAxis())
-    {
-        viewsDescriptor.SetAxis(flatBufferViewsDescriptor->axis());
-    }
-
     auto layerName = GetLayerName(graph, layerIndex);
     IConnectableLayer* layer = m_Network->AddSplitterLayer(viewsDescriptor, layerName.c_str());
 
diff --git a/src/armnnSerializer/Serializer.cpp b/src/armnnSerializer/Serializer.cpp
index 1b7e46e51a..6cadb598a2 100644
--- a/src/armnnSerializer/Serializer.cpp
+++ b/src/armnnSerializer/Serializer.cpp
@@ -1306,9 +1306,7 @@ void SerializerStrategy::SerializeSplitterLayer(const armnn::IConnectableLayer*
     // Create FlatBuffer ViewsDescriptor
     auto flatBufferViewsDescriptor = CreateViewsDescriptor(m_flatBufferBuilder,
                                                            flatBufferOriginDescriptor,
-                                                           m_flatBufferBuilder.CreateVector(flatBufferViewSizes),
-                                                           viewsDescriptor.HasAxis(),
-                                                           viewsDescriptor.GetAxis());
+                                                           m_flatBufferBuilder.CreateVector(flatBufferViewSizes));
 
     // Create FlatBuffer BaseLayer
     auto flatBufferBaseLayer = CreateLayerBase(layer, serializer::LayerType::LayerType_Splitter);
diff --git a/src/armnnSerializer/test/SerializerTests.cpp b/src/armnnSerializer/test/SerializerTests.cpp
index bfe3fc6467..b2590eaa42 100644
--- a/src/armnnSerializer/test/SerializerTests.cpp
+++ b/src/armnnSerializer/test/SerializerTests.cpp
@@ -2644,7 +2644,6 @@ TEST_CASE("SerializeSplitter")
             desc.SetViewSize(g, dimIdx, splitterDimSizes[dimIdx]);
         }
     }
-    desc.SetAxis(1);
 
     const std::string layerName("splitter");
     const armnn::TensorInfo inputInfo(numDimensions, inputShape, armnn::DataType::Float32);
diff --git a/src/armnnTfLiteParser/TfLiteParser.cpp b/src/armnnTfLiteParser/TfLiteParser.cpp
index f40f7ffe0e..301989ebb7 100644
--- a/src/armnnTfLiteParser/TfLiteParser.cpp
+++ b/src/armnnTfLiteParser/TfLiteParser.cpp
@@ -4213,7 +4213,7 @@ void TfLiteParserImpl::ParseUnpack(size_t subgraphIndex, size_t operatorIndex)
         }
         splitDesc.SetViewOriginCoord(j, unpackAxis, unpackDimSizes[unpackAxis] * j);
     }
-    splitDesc.SetAxis(unpackAxis);
+
     auto layerName = fmt::format("Unpack:{}:{}", subgraphIndex, operatorIndex);
     IConnectableLayer* layer = m_Network->AddSplitterLayer(splitDesc, layerName.c_str());
 
@@ -4352,10 +4352,7 @@ void TfLiteParserImpl::ParseSplit(size_t subgraphIndex, size_t operatorIndex)
         }
         splitDesc.SetViewOriginCoord(j, splitDim, splitterDimSizes[splitDim] * j);
     }
-    if (axisTensorInfo.GetNumElements() == 1)
-    {
-        splitDesc.SetAxis(axis);
-    }
+
     auto layerName = fmt::format("Split:{}:{}", subgraphIndex, operatorIndex);
     IConnectableLayer* layer = m_Network->AddSplitterLayer(splitDesc, layerName.c_str());
 
@@ -4534,7 +4531,6 @@ void TfLiteParserImpl::ParseSplitV(size_t subgraphIndex, size_t operatorIndex)
         splitDesc.SetViewOriginCoord(j, splitDim, accumSplit);
         accumSplit += splitSize;
     }
-    splitDesc.SetAxis(axis);
 
     auto layerName = fmt::format("SplitV:{}:{}", subgraphIndex, operatorIndex);
     IConnectableLayer* layer = m_Network->AddSplitterLayer(splitDesc, layerName.c_str());
diff --git a/src/backends/aclCommon/ArmComputeUtils.hpp b/src/backends/aclCommon/ArmComputeUtils.hpp
index 9552b7620a..fc59b281b5 100644
--- a/src/backends/aclCommon/ArmComputeUtils.hpp
+++ b/src/backends/aclCommon/ArmComputeUtils.hpp
@@ -9,7 +9,6 @@
 #include <armnn/utility/Assert.hpp>
 #include <armnn/utility/NumericCast.hpp>
 #include <armnn/backends/WorkloadData.hpp>
-#include <armnnUtils/TensorUtils.hpp>
 
 #include <arm_compute/runtime/FunctionDescriptors.h>
 #include <arm_compute/function_info/FullyConnectedLayerInfo.h>
@@ -248,20 +247,13 @@ inline std::set<unsigned int> ComputeSplitAxis(const armnn::SplitterDescriptor&
     unsigned int numDimensions = desc.GetNumDimensions();
     std::set<unsigned int> splitAxis;
 
-    if (desc.HasAxis())
+    for (unsigned int i = 0; i < numSplit; ++i)
     {
-        splitAxis.insert(armnnUtils::GetUnsignedAxis(desc.GetNumDimensions(), desc.GetAxis()));
-    }
-    else
-    {
-        for (unsigned int i = 0; i < numSplit; ++i)
+        for (unsigned int dimIdx = 0; dimIdx < numDimensions; ++dimIdx)
         {
-            for (unsigned int dimIdx = 0; dimIdx < numDimensions; ++dimIdx)
+            if (desc.GetViewSizes(i)[dimIdx] != input[dimIdx])
             {
-                if (desc.GetViewSizes(i)[dimIdx] != input[dimIdx])
-                {
-                    splitAxis.insert(dimIdx);
-                }
+                splitAxis.insert(dimIdx);
             }
         }
     }
diff --git a/src/backends/backendsCommon/test/layerTests/SplitterTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/SplitterTestImpl.cpp
index 13483e5ebd..9e3d83c0f4 100644
--- a/src/backends/backendsCommon/test/layerTests/SplitterTestImpl.cpp
+++ b/src/backends/backendsCommon/test/layerTests/SplitterTestImpl.cpp
@@ -1,5 +1,5 @@
 //
-// Copyright © 2019-2020,2023 Arm Ltd and Contributors. All rights reserved.
+// Copyright © 2017, 2023 Arm Ltd and Contributors. All rights reserved.
 // SPDX-License-Identifier: MIT
 //
 
@@ -199,29 +199,11 @@ std::vector<LayerTestResult<T,3>> SplitterTestCommon(
 
     // Do the first split
     armnn::SplitterQueueDescriptor data;
-    data.m_Parameters = armnn::SplitterDescriptor(2, 3);
-
     armnn::WorkloadInfo info;
     AddInputToWorkload(data, info, inputTensorInfo, inputHandle.get());
     AddOutputToWorkload(data, info, outputTensorInfo1, outputHandle1.get());
     AddOutputToWorkload(data, info, outputTensorInfo2, outputHandle2.get());
 
-    data.m_Parameters.SetViewSize(0, 0, outputChannels1);
-    data.m_Parameters.SetViewSize(0, 1, outputHeight1);
-    data.m_Parameters.SetViewSize(0, 2, outputWidth1);
-
-    data.m_Parameters.SetViewSize(1, 0, outputChannels2);
-    data.m_Parameters.SetViewSize(1, 1, outputHeight2);
-    data.m_Parameters.SetViewSize(1, 2, outputWidth2);
-
-    data.m_Parameters.SetViewOriginCoord(0, 0, 0);
-    data.m_Parameters.SetViewOriginCoord(0, 1, 0);
-    data.m_Parameters.SetViewOriginCoord(0, 2, 0);
-
-    data.m_Parameters.SetViewOriginCoord(1, 0, 1);
-    data.m_Parameters.SetViewOriginCoord(1, 1, 0);
-    data.m_Parameters.SetViewOriginCoord(1, 2, 0);
-
     data.m_ViewOrigins.push_back(window1);
     data.m_ViewOrigins.push_back(window2);
 
@@ -242,29 +224,11 @@ std::vector<LayerTestResult<T,3>> SplitterTestCommon(
 
     // Do the second split.
     armnn::SplitterQueueDescriptor data2;
-    data2.m_Parameters = armnn::SplitterDescriptor(2, 3);
-
     armnn::WorkloadInfo info2;
     AddInputToWorkload(data2, info2, outputTensorInfo2, outputHandle2.get());
     AddOutputToWorkload(data2, info2, outputTensorInfo3, outputHandle3.get());
     AddOutputToWorkload(data2, info2, outputTensorInfo4, outputHandle4.get());
 
-    data2.m_Parameters.SetViewSize(0, 0, outputChannels1);
-    data2.m_Parameters.SetViewSize(0, 1, outputHeight1);
-    data2.m_Parameters.SetViewSize(0, 2, outputWidth1);
-
-    data2.m_Parameters.SetViewSize(1, 0, outputChannels2);
-    data2.m_Parameters.SetViewSize(1, 1, outputHeight2);
-    data2.m_Parameters.SetViewSize(1, 2, outputWidth1);
-
-    data2.m_Parameters.SetViewOriginCoord(0, 0, 0);
-    data2.m_Parameters.SetViewOriginCoord(0, 1, 0);
-    data2.m_Parameters.SetViewOriginCoord(0, 2, 0);
-
-    data2.m_Parameters.SetViewOriginCoord(1, 0, 1);
-    data2.m_Parameters.SetViewOriginCoord(1, 1, 0);
-    data2.m_Parameters.SetViewOriginCoord(1, 2, 0);
-
     data2.m_ViewOrigins.push_back(window3);
     data2.m_ViewOrigins.push_back(window4);
 
@@ -343,17 +307,6 @@ LayerTestResult<T, 3> CopyViaSplitterTestImpl(
     AddInputToWorkload(data, info, tensorInfo, inputHandle.get());
     AddOutputToWorkload(data, info, tensorInfo, outputHandle.get());
 
-    data.m_Parameters = armnn::SplitterDescriptor(1, 3);
-    data.m_Parameters.SetAxis(0);
-
-    data.m_Parameters.SetViewSize(0, 0, 3);
-    data.m_Parameters.SetViewSize(0, 1, 6);
-    data.m_Parameters.SetViewSize(0, 2, 5);
-
-    data.m_Parameters.SetViewOriginCoord(0, 0, 0);
-    data.m_Parameters.SetViewOriginCoord(0, 1, 0);
-    data.m_Parameters.SetViewOriginCoord(0, 2, 0);
-
     data.m_ViewOrigins.push_back(window);
 
     std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateWorkload(armnn::LayerType::Splitter,
diff --git a/src/backends/cl/ClTensorHandleFactory.cpp b/src/backends/cl/ClTensorHandleFactory.cpp
index be3ca5e05a..82e41d3ff6 100644
--- a/src/backends/cl/ClTensorHandleFactory.cpp
+++ b/src/backends/cl/ClTensorHandleFactory.cpp
@@ -103,7 +103,7 @@ const FactoryId& ClTensorHandleFactory::GetId() const
 
 bool ClTensorHandleFactory::SupportsSubTensors() const
 {
-    return false;
+    return true;
 }
 
 MemorySourceFlags ClTensorHandleFactory::GetExportFlags() const
diff --git a/src/backends/neon/NeonTensorHandleFactory.cpp b/src/backends/neon/NeonTensorHandleFactory.cpp
index 2597b5f28b..ce3ce5c0d7 100644
--- a/src/backends/neon/NeonTensorHandleFactory.cpp
+++ b/src/backends/neon/NeonTensorHandleFactory.cpp
@@ -104,7 +104,7 @@ bool NeonTensorHandleFactory::SupportsInPlaceComputation() const
 
 bool NeonTensorHandleFactory::SupportsSubTensors() const
 {
-    return false;
+    return true;
 }
 
 MemorySourceFlags NeonTensorHandleFactory::GetExportFlags() const
diff --git a/src/backends/neon/test/NeonTensorHandleTests.cpp b/src/backends/neon/test/NeonTensorHandleTests.cpp
index a94e4dd187..c8e781b71d 100644
--- a/src/backends/neon/test/NeonTensorHandleTests.cpp
+++ b/src/backends/neon/test/NeonTensorHandleTests.cpp
@@ -1,5 +1,5 @@
 //
-// Copyright © 2020-2021,2023 Arm Ltd and Contributors. All rights reserved.
+// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
 // SPDX-License-Identifier: MIT
 //
 #include <Graph.hpp>
@@ -89,6 +89,81 @@ TEST_CASE("NeonTensorHandleGetCapabilitiesPadding")
     CHECK(capabilities[0].m_Value);
 }
 
+TEST_CASE("ConcatOnXorYSubTensorsNoPaddingRequiredTest")
+{
+    armnn::INetworkPtr net(armnn::INetwork::Create());
+
+    // Set up tensor infos
+    const armnn::TensorInfo inputInfo = armnn::TensorInfo({2, 3, 2, 2}, armnn::DataType::Float32);
+    const armnn::TensorInfo intermediateInfo = armnn::TensorInfo({2, 3, 2, 2}, armnn::DataType::Float32);
+    const armnn::TensorInfo outputInfo = armnn::TensorInfo({2, 3, 4, 2}, armnn::DataType::Float32);
+
+    armnn::ElementwiseUnaryDescriptor descriptor(armnn::UnaryOperation::Abs);
+
+    // Create the network
+    armnn::IConnectableLayer* const input0Layer = net->AddInputLayer(0, "input_0");
+    input0Layer->GetOutputSlot(0).SetTensorInfo(inputInfo);
+    armnn::IConnectableLayer* elementwiseUnaryLayer0 = net->AddElementwiseUnaryLayer(descriptor, "elementwiseUnary_0");
+    elementwiseUnaryLayer0->GetOutputSlot(0).SetTensorInfo(intermediateInfo);
+    input0Layer->GetOutputSlot(0).Connect(elementwiseUnaryLayer0->GetInputSlot(0));
+
+    armnn::IConnectableLayer* const input1Layer = net->AddInputLayer(1, "input_1");
+    input1Layer->GetOutputSlot(0).SetTensorInfo(inputInfo);
+    armnn::IConnectableLayer* elementwiseUnaryLayer1 = net->AddElementwiseUnaryLayer(descriptor, "elementwiseUnary_1");
+    elementwiseUnaryLayer1->GetOutputSlot(0).SetTensorInfo(intermediateInfo);
+    input1Layer->GetOutputSlot(0).Connect(elementwiseUnaryLayer1->GetInputSlot(0));
+
+    std::array<armnn::TensorShape, 2> concatInputShapes = { intermediateInfo.GetShape(), intermediateInfo.GetShape() };
+    armnn::IConnectableLayer* const concatLayer = net->AddConcatLayer(armnn::CreateDescriptorForConcatenation(
+        concatInputShapes.begin(), concatInputShapes.end(), 2), "concatenation");
+    concatLayer->GetOutputSlot(0).SetTensorInfo(outputInfo);
+    elementwiseUnaryLayer0->GetOutputSlot(0).Connect(concatLayer->GetInputSlot(0));
+    elementwiseUnaryLayer1->GetOutputSlot(0).Connect(concatLayer->GetInputSlot(1));
+
+    armnn::IConnectableLayer* const outputLayer = net->AddOutputLayer(0, "output");
+    concatLayer->GetOutputSlot(0).Connect(outputLayer->GetInputSlot(0));
+
+    armnn::IRuntime::CreationOptions options;
+    armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
+
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
+
+    const armnn::Graph& theGraph = GetGraphForTesting(optimizedNet.get());
+
+    // Load graph into runtime
+    armnn::NetworkId networkIdentifier;
+    runtime->LoadNetwork(networkIdentifier, std::move(optimizedNet));
+
+    // now check the concat how many sub-tensors it is using..
+    auto TraceSubTensorHandleAncestry = [](armnn::ITensorHandle* const subTensorHandle)
+    {
+        if (subTensorHandle && subTensorHandle->GetParent())
+        {
+            return true;
+        }
+        return false;
+    };
+
+    for (auto&& layer : theGraph)
+    {
+        if(layer->GetType() == armnn::LayerType::Concat)
+        {
+            unsigned int numberOfSubTensors = 0;
+            for (unsigned int i = 0; i < layer->GetNumInputSlots(); ++i)
+            {
+                const armnn::OutputSlot* slot = layer->GetInputSlot(i).GetConnectedOutputSlot();
+                if (TraceSubTensorHandleAncestry(slot->GetOutputHandler().GetData()))
+                {
+                    ++numberOfSubTensors;
+                }
+            }
+            // sub-tensors should be supported in this configuration
+            ARMNN_ASSERT(numberOfSubTensors > 0);
+        }
+    }
+}
+
 TEST_CASE("ConcatonXorYPaddingRequiredTest")
 {
     armnn::INetworkPtr net(armnn::INetwork::Create());
@@ -172,6 +247,212 @@ TEST_CASE("ConcatonXorYPaddingRequiredTest")
     ARMNN_ASSERT(numberOfSubTensors == 0);
 }
 
+TEST_CASE("SplitteronXorYNoPaddingRequiredTest")
+{
+    using namespace armnn;
+
+    unsigned int splitAxis = 2;
+    unsigned int numSplit = 2;
+
+    const TensorShape& inputShape = { 2, 3, 4, 2 };
+    const armnn::TensorInfo intermediateInfo = armnn::TensorInfo({ 2, 3, 2, 2 }, armnn::DataType::Float32);
+    const std::vector<TensorShape> outputShapes{{ 2, 3, 2, 2 },
+                                                { 2, 3, 2, 2 }};
+    const float qScale = 1.0f;
+    const int32_t qOffset = 0;
+
+    // Creates structures for input & output.
+    std::vector<float> inputData{
+            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,
+            31, 32,
+            33, 34,
+            35, 36,
+            37, 38,
+            39, 40,
+            41, 42,
+            43, 44,
+            45, 46,
+            47, 48
+    };
+
+    std::vector<float> expectedOutput0{
+            1, 2,
+            3, 4,
+            9, 10,
+            11, 12,
+            17, 18,
+            19, 20,
+            25, 26,
+            27, 28,
+            33, 34,
+            35, 36,
+            41, 42,
+            43, 44
+    };
+
+    std::vector<float> expectedOutput1{
+            5, 6,
+            7, 8,
+            13, 14,
+            15, 16,
+            21, 22,
+            23, 24,
+            29, 30,
+            31, 32,
+            37, 38,
+            39, 40,
+            45, 46,
+            47, 48
+    };
+
+    // Builds up the structure of the network.
+    INetworkPtr net(INetwork::Create());
+
+    TensorInfo inputTensorInfo(inputShape, armnn::DataType::Float32, qScale, qOffset);
+
+    armnn::ElementwiseUnaryDescriptor descriptor(armnn::UnaryOperation::Abs);
+
+    // Splitter
+    std::vector<unsigned int> splitterDimSizes(inputShape.GetNumDimensions());
+
+    // Add current input shape to splitterDimSizes
+    for (unsigned int i = 0; i < inputShape.GetNumDimensions(); ++i)
+    {
+        splitterDimSizes[i] = inputTensorInfo.GetShape()[i];
+    }
+
+    if (splitterDimSizes[splitAxis] % numSplit != 0)
+    {
+        throw ParseException("Number of splits must evenly divide the dimension");
+    }
+
+    splitterDimSizes[splitAxis] /= numSplit;
+
+    SplitterDescriptor splitDesc(numSplit, inputShape.GetNumDimensions());
+
+    for (unsigned int g = 0; g < numSplit; ++g)
+    {
+        // Set the size of the views.
+        for (unsigned int dimIdx = 0; dimIdx < splitterDimSizes.size(); ++dimIdx)
+        {
+            splitDesc.SetViewSize(g, dimIdx, splitterDimSizes[dimIdx]);
+        }
+        splitDesc.SetViewOriginCoord(g, splitAxis, splitterDimSizes[splitAxis] * g);
+    }
+    IConnectableLayer* input = net->AddInputLayer(0, "input");
+    IConnectableLayer* elementWiseUnary0 = net->AddElementwiseUnaryLayer(descriptor, "elementwiseunary_0");
+    IConnectableLayer* elementWiseUnary1 = net->AddElementwiseUnaryLayer(descriptor, "elementwiseunary_0");
+    IConnectableLayer* splitter = net->AddSplitterLayer(splitDesc, "splitter");
+
+    // Connections
+    Connect(input, splitter, inputTensorInfo, 0, 0);
+    Connect(splitter, elementWiseUnary0, intermediateInfo, 0, 0);
+    Connect(splitter, elementWiseUnary1, intermediateInfo, 1, 0);
+
+    std::vector<IConnectableLayer*> pooling2dLayers{elementWiseUnary0, elementWiseUnary1};
+
+    for (unsigned int i = 0; i < outputShapes.size(); ++i)
+    {
+        TensorInfo outputTensorInfo(outputShapes[i], armnn::DataType::Float32, qScale, qOffset);
+        IConnectableLayer* output = net->AddOutputLayer(armnn::numeric_cast<LayerBindingId>(i));
+        Connect(pooling2dLayers[i], output, outputTensorInfo, 0, 0);
+    }
+
+    std::map<int, std::vector<float>> inputTensorData = {{ 0,inputData }};
+    std::map<int, std::vector<float>> expectedOutputData = {{ 0, expectedOutput0 }, { 1, expectedOutput1 }};
+
+    armnn::IRuntime::CreationOptions options;
+    armnn::IRuntimePtr runtime(armnn::IRuntime::Create(options));
+
+    std::vector<armnn::BackendId> backends = { armnn::Compute::CpuAcc };
+    armnn::IOptimizedNetworkPtr optimizedNet = armnn::Optimize(*net, backends, runtime->GetDeviceSpec());
+
+    const armnn::Graph& theGraph = GetGraphForTesting(optimizedNet.get());
+
+    // Load graph into runtime
+    armnn::NetworkId networkIdentifier;
+    runtime->LoadNetwork(networkIdentifier, std::move(optimizedNet));
+
+    // now check the concat how many sub-tensors it is using..
+    auto TraceSubTensorHandleAncestry = [](armnn::ITensorHandle* const subTensorHandle)
+    {
+        if (subTensorHandle && subTensorHandle->GetParent())
+        {
+            return true;
+        }
+        return false;
+    };
+
+    for (auto&& layer : theGraph)
+    {
+        if(layer->GetType() == armnn::LayerType::ElementwiseUnary)
+        {
+            unsigned int numberOfSubTensors = 0;
+            for (unsigned int i = 0; i < layer->GetNumInputSlots(); ++i)
+            {
+                const armnn::OutputSlot* slot = layer->GetInputSlot(i).GetConnectedOutputSlot();
+                if (TraceSubTensorHandleAncestry(slot->GetOutputHandler().GetData()))
+                {
+                    ++numberOfSubTensors;
+                }
+            }
+            // sub-tensors should be supported in this configuration
+            ARMNN_ASSERT(numberOfSubTensors > 0);
+        }
+    }
+
+    InputTensors inputTensors;
+    inputTensors.reserve(inputTensorData.size());
+    for (auto&& it : inputTensorData)
+    {
+        TensorInfo inputTensorInfo = runtime->GetInputTensorInfo(networkIdentifier, it.first);
+        inputTensorInfo.SetConstant(true);
+        inputTensors.push_back({it.first,
+                                ConstTensor(inputTensorInfo, it.second.data())});
+    }
+    OutputTensors outputTensors;
+    outputTensors.reserve(expectedOutputData.size());
+    std::map<int, std::vector<float>> outputStorage;
+    for (auto&& it : expectedOutputData)
+    {
+        std::vector<float> out(it.second.size());
+        outputStorage.emplace(it.first, out);
+        outputTensors.push_back({it.first,
+                                 Tensor(runtime->GetOutputTensorInfo(networkIdentifier, it.first),
+                                               outputStorage.at(it.first).data())});
+    }
+
+    // Does the inference.
+    runtime->EnqueueWorkload(networkIdentifier, inputTensors, outputTensors);
+
+    // Checks the results.
+    float tolerance = 0.000001f;
+    for (auto&& it : expectedOutputData)
+    {
+        std::vector<float> out = outputStorage.at(it.first);
+        for (unsigned int i = 0; i < out.size(); ++i)
+        {
+            CHECK_MESSAGE(Compare<armnn::DataType::Float32>(it.second[i], out[i], tolerance) == true,
+                    "Actual output: " << out[i] << ". Expected output:" << it.second[i]);
+
+        }
+    }
+}
+
 TEST_CASE("SplitteronXorYPaddingRequiredTest")
 {
     using namespace armnn;