IVGCVSW-1988: Refactor backend-specific unit tests

Change-Id: I8eca81d2e0780390eaa837c186ffe1c7d41fdebe

6 files changed, 1088 insertions, 1 deletions

diff --git a/src/backends/neon/CMakeLists.txt b/src/backends/neon/CMakeLists.txt
index c6492bc076..93c7955a5f 100644
--- a/src/backends/neon/CMakeLists.txt
+++ b/src/backends/neon/CMakeLists.txt
@@ -14,7 +14,8 @@ if(ARMCOMPUTENEON)
         NeonTensorHandle.hpp
     )
 
-    add_subdirectory(workloads test)
+    add_subdirectory(workloads)
+    add_subdirectory(test)
 else()
     list(APPEND armnnNeonBackend_sources
         NeonLayerSupport.cpp
diff --git a/src/backends/neon/backend.cmake b/src/backends/neon/backend.cmake
index 5f02c845ed..0240d527b3 100644
--- a/src/backends/neon/backend.cmake
+++ b/src/backends/neon/backend.cmake
@@ -6,6 +6,7 @@
 if(ARMCOMPUTENEON)
     add_subdirectory(${PROJECT_SOURCE_DIR}/src/backends/neon)
     list(APPEND armnnLibraries armnnNeonBackend armnnNeonBackendWorkloads)
+    list(APPEND armnnUnitTestLibraries armnnNeonBackendUnitTests)
 else()
     message("NEON backend is disabled")
     add_subdirectory(${PROJECT_SOURCE_DIR}/src/backends/neon)
diff --git a/src/backends/neon/test/CMakeLists.txt b/src/backends/neon/test/CMakeLists.txt
index f41a074999..82156f380b 100644
--- a/src/backends/neon/test/CMakeLists.txt
+++ b/src/backends/neon/test/CMakeLists.txt
@@ -2,3 +2,14 @@
 # Copyright © 2017 Arm Ltd. All rights reserved.
 # SPDX-License-Identifier: MIT
 #
+
+list(APPEND armnnNeonBackendUnitTests_sources
+    NeonCreateWorkloadTests.cpp
+    NeonLayerSupportTests.cpp
+    NeonLayerTests.cpp
+)
+
+add_library(armnnNeonBackendUnitTests STATIC ${armnnNeonBackendUnitTests_sources})
+target_include_directories(armnnNeonBackendUnitTests PRIVATE ${PROJECT_SOURCE_DIR}/src)
+target_include_directories(armnnNeonBackendUnitTests PRIVATE ${PROJECT_SOURCE_DIR}/src/armnn)
+target_include_directories(armnnNeonBackendUnitTests PRIVATE ${PROJECT_SOURCE_DIR}/src/armnnUtils)
\ No newline at end of file
diff --git a/src/backends/neon/test/NeonCreateWorkloadTests.cpp b/src/backends/neon/test/NeonCreateWorkloadTests.cpp
new file mode 100644
index 0000000000..d1a5b2a5f2
--- /dev/null
+++ b/src/backends/neon/test/NeonCreateWorkloadTests.cpp
@@ -0,0 +1,531 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include <armnn/test/CreateWorkloadClNeon.hpp>
+
+#include <backends/MemCopyWorkload.hpp>
+#include <backends/neon/NeonWorkloadFactory.hpp>
+#include <backends/neon/NeonTensorHandle.hpp>
+#include <backends/neon/workloads/NeonWorkloadUtils.hpp>
+#include <backends/neon/workloads/NeonWorkloads.hpp>
+
+BOOST_AUTO_TEST_SUITE(CreateWorkloadNeon)
+
+namespace
+{
+
+bool TestNeonTensorHandleInfo(armnn::INeonTensorHandle* handle, const armnn::TensorInfo& expectedInfo)
+{
+    using namespace armnn::armcomputetensorutils;
+
+    const arm_compute::ITensorInfo* handleInfo = handle->GetTensor().info();
+    const arm_compute::TensorInfo expectedAclInfo = BuildArmComputeTensorInfo(expectedInfo);
+
+    if (handleInfo->data_type() != expectedAclInfo.data_type())
+    {
+        return false;
+    }
+
+    if (handleInfo->num_dimensions() != expectedAclInfo.num_dimensions())
+    {
+        return false;
+    }
+
+    if (handleInfo->quantization_info() != expectedAclInfo.quantization_info())
+    {
+        return false;
+    }
+
+    for (std::size_t d = 0; d < expectedAclInfo.num_dimensions(); ++d)
+    {
+        if (handleInfo->dimension(d) != expectedAclInfo.dimension(d))
+        {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+} // namespace
+
+template <typename ActivationWorkloadType, typename armnn::DataType DataType>
+static void NeonCreateActivationWorkloadTest()
+{
+    Graph graph;
+    NeonWorkloadFactory factory;
+    auto workload = CreateActivationWorkloadTest<ActivationWorkloadType, DataType>
+            (factory, graph);
+
+    // Checks that inputs/outputs are as we expect them (see definition of CreateActivationWorkloadTest).
+    ActivationQueueDescriptor queueDescriptor = workload->GetData();
+    auto inputHandle  = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
+    auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
+    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({1, 1}, DataType)));
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({1, 1}, DataType)));
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+BOOST_AUTO_TEST_CASE(CreateActivationFloat16Workload)
+{
+    NeonCreateActivationWorkloadTest<NeonActivationFloatWorkload, DataType::Float16>();
+}
+#endif
+
+BOOST_AUTO_TEST_CASE(CreateActivationFloatWorkload)
+{
+    NeonCreateActivationWorkloadTest<NeonActivationFloatWorkload, DataType::Float32>();
+}
+
+template <typename WorkloadType,
+          typename DescriptorType,
+          typename LayerType,
+          armnn::DataType DataType>
+static void NeonCreateArithmethicWorkloadTest()
+{
+    Graph graph;
+    NeonWorkloadFactory factory;
+    auto workload = CreateArithmeticWorkloadTest<WorkloadType, DescriptorType, LayerType, DataType>(factory, graph);
+
+    DescriptorType queueDescriptor = workload->GetData();
+    auto inputHandle1 = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
+    auto inputHandle2 = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[1]);
+    auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
+    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle1, TensorInfo({2, 3}, DataType)));
+    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle2, TensorInfo({2, 3}, DataType)));
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({2, 3}, DataType)));
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+BOOST_AUTO_TEST_CASE(CreateAdditionFloat16Workload)
+{
+    NeonCreateArithmethicWorkloadTest<NeonAdditionFloatWorkload,
+                                      AdditionQueueDescriptor,
+                                      AdditionLayer,
+                                      DataType::Float16>();
+}
+#endif
+
+BOOST_AUTO_TEST_CASE(CreateAdditionFloatWorkload)
+{
+    NeonCreateArithmethicWorkloadTest<NeonAdditionFloatWorkload,
+                                      AdditionQueueDescriptor,
+                                      AdditionLayer,
+                                      DataType::Float32>();
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+BOOST_AUTO_TEST_CASE(CreateSubtractionFloat16Workload)
+{
+    NeonCreateArithmethicWorkloadTest<NeonSubtractionFloatWorkload,
+                                      SubtractionQueueDescriptor,
+                                      SubtractionLayer,
+                                      DataType::Float16>();
+}
+#endif
+
+BOOST_AUTO_TEST_CASE(CreateSubtractionFloatWorkload)
+{
+    NeonCreateArithmethicWorkloadTest<NeonSubtractionFloatWorkload,
+                                      SubtractionQueueDescriptor,
+                                      SubtractionLayer,
+                                      DataType::Float32>();
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+BOOST_AUTO_TEST_CASE(CreateMultiplicationFloat16Workload)
+{
+    NeonCreateArithmethicWorkloadTest<NeonMultiplicationFloatWorkload,
+                                      MultiplicationQueueDescriptor,
+                                      MultiplicationLayer,
+                                      DataType::Float16>();
+}
+#endif
+
+BOOST_AUTO_TEST_CASE(CreateMultiplicationFloatWorkload)
+{
+    NeonCreateArithmethicWorkloadTest<NeonMultiplicationFloatWorkload,
+                                      MultiplicationQueueDescriptor,
+                                      MultiplicationLayer,
+                                      DataType::Float32>();
+}
+
+template <typename BatchNormalizationWorkloadType, typename armnn::DataType DataType>
+static void NeonCreateBatchNormalizationWorkloadTest()
+{
+    Graph                graph;
+    NeonWorkloadFactory  factory;
+    auto workload = CreateBatchNormalizationWorkloadTest<BatchNormalizationWorkloadType, DataType>(factory, graph);
+
+    // Checks that outputs and inputs are as we expect them (see definition of CreateBatchNormalizationWorkloadTest).
+    BatchNormalizationQueueDescriptor queueDescriptor = workload->GetData();
+    auto inputHandle  = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
+    auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
+    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({2, 3, 1, 1}, DataType)));
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({2, 3, 1, 1}, DataType)));
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+BOOST_AUTO_TEST_CASE(CreateBatchNormalizationFloat16Workload)
+{
+    NeonCreateBatchNormalizationWorkloadTest<NeonBatchNormalizationFloatWorkload, DataType::Float16>();
+}
+#endif
+
+BOOST_AUTO_TEST_CASE(CreateBatchNormalizationFloatWorkload)
+{
+    NeonCreateBatchNormalizationWorkloadTest<NeonBatchNormalizationFloatWorkload, DataType::Float32>();
+}
+
+template <typename Convolution2dWorkloadType, typename armnn::DataType DataType>
+static void NeonCreateConvolution2dWorkloadTest(DataLayout dataLayout = DataLayout::NCHW)
+{
+    Graph                graph;
+    NeonWorkloadFactory  factory;
+    auto                 workload = CreateConvolution2dWorkloadTest<Convolution2dWorkloadType,
+                                    DataType>(factory, graph, dataLayout);
+
+    TensorShape inputShape  = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 3, 8, 16} : TensorShape{2, 8, 16, 3};
+    TensorShape outputShape = (dataLayout == DataLayout::NCHW) ? TensorShape{2, 2, 2, 10} : TensorShape{2, 2, 10, 2};
+
+    // Checks that outputs and inputs are as we expect them (see definition of CreateConvolution2dWorkloadTest).
+    Convolution2dQueueDescriptor queueDescriptor = workload->GetData();
+    auto inputHandle  = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
+    auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
+    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo(inputShape, DataType)));
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle,  TensorInfo(outputShape, DataType)));
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+BOOST_AUTO_TEST_CASE(CreateConvolution2dFloat16NchwWorkload)
+{
+    NeonCreateConvolution2dWorkloadTest<NeonConvolution2dFloatWorkload, DataType::Float16>();
+}
+
+BOOST_AUTO_TEST_CASE(CreateConvolution2dFloat16NhwcWorkload)
+{
+    NeonCreateConvolution2dWorkloadTest<NeonConvolution2dFloatWorkload, DataType::Float16>(DataLayout::NHWC);
+}
+
+#endif
+BOOST_AUTO_TEST_CASE(CreateConvolution2dFloatNchwWorkload)
+{
+    NeonCreateConvolution2dWorkloadTest<NeonConvolution2dFloatWorkload, DataType::Float32>();
+}
+
+BOOST_AUTO_TEST_CASE(CreateConvolution2dFloatNhwcWorkload)
+{
+    NeonCreateConvolution2dWorkloadTest<NeonConvolution2dFloatWorkload, DataType::Float32>(DataLayout::NHWC);
+}
+
+template <typename FullyConnectedWorkloadType, typename armnn::DataType DataType>
+static void NeonCreateFullyConnectedWorkloadTest()
+{
+    Graph               graph;
+    NeonWorkloadFactory factory;
+    auto                workload = CreateFullyConnectedWorkloadTest<FullyConnectedWorkloadType,
+                                   DataType>(factory, graph);
+
+    // Checks that outputs and inputs are as we expect them (see definition of CreateFullyConnectedWorkloadTest).
+    FullyConnectedQueueDescriptor queueDescriptor = workload->GetData();
+    auto inputHandle  = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
+    auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
+    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({3, 1, 4, 5}, DataType)));
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({3, 7}, DataType)));
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+BOOST_AUTO_TEST_CASE(CreateFullyConnectedFloat16Workload)
+{
+    NeonCreateFullyConnectedWorkloadTest<NeonFullyConnectedWorkload, DataType::Float16>();
+}
+#endif
+
+BOOST_AUTO_TEST_CASE(CreateFullyConnectedFloatWorkload)
+{
+    NeonCreateFullyConnectedWorkloadTest<NeonFullyConnectedWorkload, DataType::Float32>();
+}
+
+template <typename NormalizationWorkloadType, typename armnn::DataType DataType>
+static void NeonCreateNormalizationWorkloadTest(DataLayout dataLayout)
+{
+    Graph graph;
+    NeonWorkloadFactory factory;
+    auto workload = CreateNormalizationWorkloadTest<NormalizationWorkloadType, DataType>(factory, graph, dataLayout);
+
+    // Checks that outputs and inputs are as we expect them (see definition of CreateNormalizationWorkloadTest).
+    NormalizationQueueDescriptor queueDescriptor = workload->GetData();
+    auto inputHandle  = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
+    auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
+    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({3, 5, 5, 1}, DataType)));
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({3, 5, 5, 1}, DataType)));
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+BOOST_AUTO_TEST_CASE(CreateNormalizationFloat16NchwWorkload)
+{
+    NeonCreateNormalizationWorkloadTest<NeonNormalizationFloatWorkload, DataType::Float16>(DataLayout::NCHW);
+}
+
+BOOST_AUTO_TEST_CASE(CreateNormalizationFloat16NhwcWorkload)
+{
+    NeonCreateNormalizationWorkloadTest<NeonNormalizationFloatWorkload, DataType::Float16>(DataLayout::NHWC);
+}
+#endif
+
+BOOST_AUTO_TEST_CASE(CreateNormalizationFloatNchwWorkload)
+{
+    NeonCreateNormalizationWorkloadTest<NeonNormalizationFloatWorkload, DataType::Float32>(DataLayout::NCHW);
+}
+
+BOOST_AUTO_TEST_CASE(CreateNormalizationFloatNhwcWorkload)
+{
+    NeonCreateNormalizationWorkloadTest<NeonNormalizationFloatWorkload, DataType::Float32>(DataLayout::NHWC);
+}
+
+
+template <typename Pooling2dWorkloadType, typename armnn::DataType DataType>
+static void NeonCreatePooling2dWorkloadTest(DataLayout dataLayout = DataLayout::NCHW)
+{
+    Graph               graph;
+    NeonWorkloadFactory factory;
+    auto                workload = CreatePooling2dWorkloadTest<Pooling2dWorkloadType, DataType>
+                                   (factory, graph, dataLayout);
+
+    TensorShape inputShape  = (dataLayout == DataLayout::NCHW) ? TensorShape{3, 2, 5, 5} : TensorShape{3, 5, 5, 2};
+    TensorShape outputShape = (dataLayout == DataLayout::NCHW) ? TensorShape{3, 2, 2, 4} : TensorShape{3, 2, 4, 2};
+
+    // Checks that outputs and inputs are as we expect them (see definition of CreatePooling2dWorkloadTest).
+    Pooling2dQueueDescriptor queueDescriptor = workload->GetData();
+    auto inputHandle  = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
+    auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
+    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo(inputShape, DataType)));
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo(outputShape, DataType)));
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+BOOST_AUTO_TEST_CASE(CreatePooling2dFloat16Workload)
+{
+    NeonCreatePooling2dWorkloadTest<NeonPooling2dFloatWorkload, DataType::Float16>();
+}
+#endif
+
+BOOST_AUTO_TEST_CASE(CreatePooling2dFloatNchwWorkload)
+{
+    NeonCreatePooling2dWorkloadTest<NeonPooling2dFloatWorkload, DataType::Float32>(DataLayout::NCHW);
+}
+
+BOOST_AUTO_TEST_CASE(CreatePooling2dFloatNhwcWorkload)
+{
+    NeonCreatePooling2dWorkloadTest<NeonPooling2dFloatWorkload, DataType::Float32>(DataLayout::NHWC);
+}
+
+BOOST_AUTO_TEST_CASE(CreatePooling2dUint8NchwWorkload)
+{
+    NeonCreatePooling2dWorkloadTest<NeonPooling2dUint8Workload, DataType::QuantisedAsymm8>(DataLayout::NCHW);
+}
+
+BOOST_AUTO_TEST_CASE(CreatePooling2dUint8NhwcWorkload)
+{
+    NeonCreatePooling2dWorkloadTest<NeonPooling2dUint8Workload, DataType::QuantisedAsymm8>(DataLayout::NHWC);
+}
+
+template <typename ReshapeWorkloadType, typename armnn::DataType DataType>
+static void NeonCreateReshapeWorkloadTest()
+{
+    Graph               graph;
+    NeonWorkloadFactory factory;
+    auto                workload = CreateReshapeWorkloadTest<ReshapeWorkloadType, DataType>(factory, graph);
+
+    // Checks that outputs and inputs are as we expect them (see definition of CreateReshapeWorkloadTest).
+    ReshapeQueueDescriptor queueDescriptor = workload->GetData();
+    auto inputHandle  = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
+    auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
+    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({4, 1}, DataType)));
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({1, 4}, DataType)));
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+BOOST_AUTO_TEST_CASE(CreateReshapeFloat16Workload)
+{
+    NeonCreateReshapeWorkloadTest<NeonReshapeFloatWorkload, DataType::Float16>();
+}
+#endif
+
+BOOST_AUTO_TEST_CASE(CreateReshapeFloatWorkload)
+{
+    NeonCreateReshapeWorkloadTest<NeonReshapeFloatWorkload, DataType::Float32>();
+}
+
+BOOST_AUTO_TEST_CASE(CreateReshapeUint8Workload)
+{
+    NeonCreateReshapeWorkloadTest<NeonReshapeUint8Workload, DataType::QuantisedAsymm8>();
+}
+
+template <typename SoftmaxWorkloadType, typename armnn::DataType DataType>
+static void NeonCreateSoftmaxWorkloadTest()
+{
+    Graph               graph;
+    NeonWorkloadFactory factory;
+    auto workload = CreateSoftmaxWorkloadTest<SoftmaxWorkloadType, DataType>(factory, graph);
+
+    // Checks that outputs and inputs are as we expect them (see definition of CreateSoftmaxWorkloadTest).
+    SoftmaxQueueDescriptor queueDescriptor = workload->GetData();
+    auto inputHandle  = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
+    auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
+    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({4, 1}, DataType)));
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({4, 1}, DataType)));
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+BOOST_AUTO_TEST_CASE(CreateSoftmaxFloat16Workload)
+{
+    NeonCreateSoftmaxWorkloadTest<NeonSoftmaxFloatWorkload, DataType::Float16>();
+}
+#endif
+
+BOOST_AUTO_TEST_CASE(CreateSoftmaxFloatWorkload)
+{
+    NeonCreateSoftmaxWorkloadTest<NeonSoftmaxFloatWorkload, DataType::Float32>();
+}
+
+BOOST_AUTO_TEST_CASE(CreateSplitterWorkload)
+{
+    Graph graph;
+    NeonWorkloadFactory factory;
+    auto workload = CreateSplitterWorkloadTest<NeonSplitterFloatWorkload, DataType::Float32>(factory, graph);
+
+    // Checks that outputs are as we expect them (see definition of CreateSplitterWorkloadTest).
+    SplitterQueueDescriptor queueDescriptor = workload->GetData();
+    auto inputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
+    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({5, 7, 7}, DataType::Float32)));
+
+    auto outputHandle0 = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle0, TensorInfo({1, 7, 7}, DataType::Float32)));
+
+    auto outputHandle1 = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[1]);
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle1, TensorInfo({2, 7, 7}, DataType::Float32)));
+
+    auto outputHandle2 = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[2]);
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle2, TensorInfo({2, 7, 7}, DataType::Float32)));
+}
+
+BOOST_AUTO_TEST_CASE(CreateSplitterMerger)
+{
+    // Tests that it is possible to decide which output of the splitter layer
+    // should be lined to which input of the merger layer.
+    // We tested that is is possible to specify 0th output
+    // of the splitter to be the 1st input to the merger, and the 1st output of the splitter to be 0th input
+    // of the merger.
+
+    Graph graph;
+    NeonWorkloadFactory factory;
+
+    auto workloads =
+        CreateSplitterMergerWorkloadTest<NeonSplitterFloatWorkload, NeonMergerFloatWorkload,
+            DataType::Float32>(factory, graph);
+
+    auto wlSplitter = std::move(workloads.first);
+    auto wlMerger = std::move(workloads.second);
+
+    //Checks that the index of inputs/outputs matches what we declared on InputDescriptor construction.
+    armnn::INeonTensorHandle* sOut0 = dynamic_cast<armnn::INeonTensorHandle*>(wlSplitter->GetData().m_Outputs[0]);
+    armnn::INeonTensorHandle* sOut1 = dynamic_cast<armnn::INeonTensorHandle*>(wlSplitter->GetData().m_Outputs[1]);
+    armnn::INeonTensorHandle* mIn0 = dynamic_cast<armnn::INeonTensorHandle*>(wlMerger->GetData().m_Inputs[0]);
+    armnn::INeonTensorHandle* mIn1 = dynamic_cast<armnn::INeonTensorHandle*>(wlMerger->GetData().m_Inputs[1]);
+
+    BOOST_TEST(sOut0);
+    BOOST_TEST(sOut1);
+    BOOST_TEST(mIn0);
+    BOOST_TEST(mIn1);
+
+    bool validDataPointers = (sOut0 == mIn1) && (sOut1 == mIn0);
+
+    BOOST_TEST(validDataPointers);
+}
+
+BOOST_AUTO_TEST_CASE(CreateSingleOutputMultipleInputs)
+{
+    // Tests that it is possible to assign multiple (two) different layers to each of the outputs of a splitter layer.
+    // We created a splitter with two outputs. That each of those outputs is used by two different activation layers
+
+    Graph graph;
+    NeonWorkloadFactory factory;
+    std::unique_ptr<NeonSplitterFloatWorkload> wlSplitter;
+    std::unique_ptr<NeonActivationFloatWorkload> wlActiv0_0;
+    std::unique_ptr<NeonActivationFloatWorkload> wlActiv0_1;
+    std::unique_ptr<NeonActivationFloatWorkload> wlActiv1_0;
+    std::unique_ptr<NeonActivationFloatWorkload> wlActiv1_1;
+
+    CreateSplitterMultipleInputsOneOutputWorkloadTest<NeonSplitterFloatWorkload,
+        NeonActivationFloatWorkload, DataType::Float32>(factory, graph, wlSplitter, wlActiv0_0, wlActiv0_1,
+                                                                 wlActiv1_0, wlActiv1_1);
+
+    armnn::INeonTensorHandle* sOut0 = dynamic_cast<armnn::INeonTensorHandle*>(wlSplitter->GetData().m_Outputs[0]);
+    armnn::INeonTensorHandle* sOut1 = dynamic_cast<armnn::INeonTensorHandle*>(wlSplitter->GetData().m_Outputs[1]);
+    armnn::INeonTensorHandle* activ0_0Im = dynamic_cast<armnn::INeonTensorHandle*>(wlActiv0_0->GetData().m_Inputs[0]);
+    armnn::INeonTensorHandle* activ0_1Im = dynamic_cast<armnn::INeonTensorHandle*>(wlActiv0_1->GetData().m_Inputs[0]);
+    armnn::INeonTensorHandle* activ1_0Im = dynamic_cast<armnn::INeonTensorHandle*>(wlActiv1_0->GetData().m_Inputs[0]);
+    armnn::INeonTensorHandle* activ1_1Im = dynamic_cast<armnn::INeonTensorHandle*>(wlActiv1_1->GetData().m_Inputs[0]);
+
+
+    BOOST_TEST(sOut0);
+    BOOST_TEST(sOut1);
+    BOOST_TEST(activ0_0Im);
+    BOOST_TEST(activ0_1Im);
+    BOOST_TEST(activ1_0Im);
+    BOOST_TEST(activ1_1Im);
+
+    bool validDataPointers = (sOut0 == activ0_0Im) && (sOut0 == activ0_1Im) &&
+                             (sOut1 == activ1_0Im) && (sOut1 == activ1_1Im);
+
+    BOOST_TEST(validDataPointers);
+}
+
+BOOST_AUTO_TEST_CASE(CreateMemCopyWorkloadsNeon)
+{
+    NeonWorkloadFactory    factory;
+    CreateMemCopyWorkloads<INeonTensorHandle>(factory);
+}
+
+template <typename L2NormalizationWorkloadType, typename armnn::DataType DataType>
+static void NeonCreateL2NormalizationWorkloadTest(DataLayout dataLayout)
+{
+    Graph                graph;
+    NeonWorkloadFactory  factory;
+    auto                 workload = CreateL2NormalizationWorkloadTest<L2NormalizationWorkloadType,
+                                    DataType>(factory, graph, dataLayout);
+
+    // Checks that inputs/outputs are as we expect them (see definition of CreateNormalizationWorkloadTest).
+    L2NormalizationQueueDescriptor queueDescriptor = workload->GetData();
+    auto inputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Inputs[0]);
+    auto outputHandle = boost::polymorphic_downcast<INeonTensorHandle*>(queueDescriptor.m_Outputs[0]);
+    BOOST_TEST(TestNeonTensorHandleInfo(inputHandle, TensorInfo({ 5, 20, 50, 67 }, DataType)));
+    BOOST_TEST(TestNeonTensorHandleInfo(outputHandle, TensorInfo({ 5, 20, 50, 67 }, DataType)));
+}
+
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+BOOST_AUTO_TEST_CASE(CreateL2NormalizationFloat16NchwWorkload)
+{
+    NeonCreateL2NormalizationWorkloadTest<NeonL2NormalizationFloatWorkload, DataType::Float16>(DataLayout::NCHW);
+}
+
+BOOST_AUTO_TEST_CASE(CreateL2NormalizationFloat16NhwcWorkload)
+{
+    NeonCreateL2NormalizationWorkloadTest<NeonL2NormalizationFloatWorkload, DataType::Float16>(DataLayout::NHWC);
+}
+#endif
+
+BOOST_AUTO_TEST_CASE(CreateL2NormalizationNchwWorkload)
+{
+    NeonCreateL2NormalizationWorkloadTest<NeonL2NormalizationFloatWorkload, DataType::Float32>(DataLayout::NCHW);
+}
+
+BOOST_AUTO_TEST_CASE(CreateL2NormalizationNhwcWorkload)
+{
+    NeonCreateL2NormalizationWorkloadTest<NeonL2NormalizationFloatWorkload, DataType::Float32>(DataLayout::NHWC);
+}
+
+BOOST_AUTO_TEST_SUITE_END()
diff --git a/src/backends/neon/test/NeonLayerSupportTests.cpp b/src/backends/neon/test/NeonLayerSupportTests.cpp
new file mode 100644
index 0000000000..db7897fc28
--- /dev/null
+++ b/src/backends/neon/test/NeonLayerSupportTests.cpp
@@ -0,0 +1,59 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include <armnn/layers/ConvertFp16ToFp32Layer.hpp>
+#include <armnn/layers/ConvertFp32ToFp16Layer.hpp>
+#include <armnn/test/TensorHelpers.hpp>
+
+#include <backends/CpuTensorHandle.hpp>
+#include <backends/neon/NeonWorkloadFactory.hpp>
+#include <backends/test/IsLayerSupportedTestImpl.hpp>
+#include <backends/test/LayerTests.hpp>
+
+#include <boost/test/unit_test.hpp>
+
+#include <string>
+
+BOOST_AUTO_TEST_SUITE(NeonLayerSupport)
+
+BOOST_AUTO_TEST_CASE(IsLayerSupportedFloat16Neon)
+{
+    armnn::NeonWorkloadFactory factory;
+    IsLayerSupportedTests<armnn::NeonWorkloadFactory, armnn::DataType::Float16>(&factory);
+}
+
+BOOST_AUTO_TEST_CASE(IsLayerSupportedFloat32Neon)
+{
+    armnn::NeonWorkloadFactory factory;
+    IsLayerSupportedTests<armnn::NeonWorkloadFactory, armnn::DataType::Float32>(&factory);
+}
+
+BOOST_AUTO_TEST_CASE(IsLayerSupportedUint8Neon)
+{
+    armnn::NeonWorkloadFactory factory;
+    IsLayerSupportedTests<armnn::NeonWorkloadFactory, armnn::DataType::QuantisedAsymm8>(&factory);
+}
+
+BOOST_AUTO_TEST_CASE(IsConvertFp16ToFp32SupportedNeon)
+{
+    std::string reasonIfUnsupported;
+
+    bool result = IsConvertLayerSupportedTests<armnn::NeonWorkloadFactory, armnn::ConvertFp16ToFp32Layer,
+      armnn::DataType::Float16, armnn::DataType::Float32>(reasonIfUnsupported);
+
+    BOOST_CHECK(result);
+}
+
+BOOST_AUTO_TEST_CASE(IsConvertFp32ToFp16SupportedNeon)
+{
+    std::string reasonIfUnsupported;
+
+    bool result = IsConvertLayerSupportedTests<armnn::NeonWorkloadFactory, armnn::ConvertFp32ToFp16Layer,
+      armnn::DataType::Float32, armnn::DataType::Float16>(reasonIfUnsupported);
+
+    BOOST_CHECK(result);
+}
+
+BOOST_AUTO_TEST_SUITE_END()
diff --git a/src/backends/neon/test/NeonLayerTests.cpp b/src/backends/neon/test/NeonLayerTests.cpp
new file mode 100644
index 0000000000..2d4ee996a4
--- /dev/null
+++ b/src/backends/neon/test/NeonLayerTests.cpp
@@ -0,0 +1,484 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include <armnn/test/TensorHelpers.hpp>
+#include <armnn/test/UnitTests.hpp>
+
+#include <backends/CpuTensorHandle.hpp>
+#include <backends/neon/NeonLayerSupport.hpp>
+#include <backends/neon/NeonWorkloadFactory.hpp>
+#include <backends/reference/RefWorkloadFactory.hpp>
+#include <backends/test/ActivationFixture.hpp>
+#include <backends/test/LayerTests.hpp>
+#include <backends/test/TensorCopyUtils.hpp>
+#include <backends/test/WorkloadTestUtils.hpp>
+
+#include <boost/test/unit_test.hpp>
+
+BOOST_AUTO_TEST_SUITE(Compute_ArmComputeNeon)
+using FactoryType = armnn::NeonWorkloadFactory;
+
+// ============================================================================
+// UNIT tests
+
+// Convolution
+ARMNN_AUTO_TEST_CASE(SimpleConvolution1d, Convolution1dTest, true)
+
+ARMNN_AUTO_TEST_CASE(SimpleConvolution2d, SimpleConvolution2d3x5Test, true)
+ARMNN_AUTO_TEST_CASE(SimpleConvolution2dSquare, SimpleConvolution2d3x3Test, true)
+ARMNN_AUTO_TEST_CASE(UnbiasedConvolution2d, SimpleConvolution2d3x5Test, false)
+ARMNN_AUTO_TEST_CASE(UnbiasedConvolution2dSquare, SimpleConvolution2d3x3Test, false)
+ARMNN_AUTO_TEST_CASE(SimpleConvolution2dAsymmetricPadding, Convolution2dAsymmetricPaddingTest)
+
+ARMNN_AUTO_TEST_CASE(SimpleConvolution2dSquareNhwc, SimpleConvolution2d3x3NhwcTest, false)
+namespace
+{
+
+armnn::Convolution2dDescriptor MakeConv2dDesc(uint32_t strideX, uint32_t strideY,
+    uint32_t padLeft = 0, uint32_t padRight = 0, uint32_t padTop = 0, uint32_t padBottom = 0)
+{
+    armnn::Convolution2dDescriptor result;
+    result.m_StrideX = strideX;
+    result.m_StrideY = strideY;
+    result.m_PadLeft = padLeft;
+    result.m_PadRight = padRight;
+    result.m_PadTop = padTop;
+    result.m_PadBottom = padBottom;
+    result.m_BiasEnabled = true;
+    return result;
+}
+
+}
+
+BOOST_AUTO_TEST_CASE(Conv2dUtils)
+{
+    // The only preferred Neon convolution is 1x1 with padding=0 and stride size {1,2,3}.
+    armnn::TensorShape shape1x1({ 1,1,1,1 });
+    armnn::TensorInfo info1x1(shape1x1, armnn::DataType::Float32);
+    BOOST_TEST(armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(1, 1)));
+    BOOST_TEST(armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(1, 2)));
+    BOOST_TEST(armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(1, 3)));
+    BOOST_TEST(armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(2, 1)));
+    BOOST_TEST(armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(2, 2)));
+    BOOST_TEST(armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(2, 3)));
+    BOOST_TEST(armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(3, 1)));
+    BOOST_TEST(armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(3, 2)));
+    BOOST_TEST(armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(3, 3)));
+
+    BOOST_TEST(!armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(4, 1)));
+    BOOST_TEST(!armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(4, 5)));
+    BOOST_TEST(!armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(3, 6)));
+
+    // non zero padding is not preferred for direct convolution
+    BOOST_TEST(!armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(1, 1, 1, 0)));
+    BOOST_TEST(!armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(1, 1, 0, 1)));
+    BOOST_TEST(!armnn::IsNeonDirectConvolutionPreferred(info1x1, MakeConv2dDesc(1, 1, 1, 1)));
+
+    // 2x2 filter not preferred for direct convolution
+    armnn::TensorShape shape2x2({ 1,1,2,2 });
+    armnn::TensorInfo info2x2(shape2x2, armnn::DataType::Float32);
+    BOOST_TEST(!armnn::IsNeonDirectConvolutionPreferred(info2x2, MakeConv2dDesc(1, 1)));
+}
+
+// Depthwise Convolution
+ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dDepthMul1, DepthwiseConvolution2dDepthMul1Test, true)
+ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dDepthMul1, DepthwiseConvolution2dDepthMul1Test, false)
+ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dDepthMul1Uint8, DepthwiseConvolution2dDepthMul1Uint8Test, true)
+ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dDepthMul1Uint8, DepthwiseConvolution2dDepthMul1Uint8Test, false)
+
+ARMNN_AUTO_TEST_CASE(DepthwiseConvolution2dAsymmetric, DepthwiseConvolution2dAsymmetricTest, true)
+ARMNN_AUTO_TEST_CASE(UnbiasedDepthwiseConvolution2dAsymmetric, DepthwiseConvolution2dAsymmetricTest, false)
+
+namespace
+{
+
+armnn::DepthwiseConvolution2dDescriptor MakeDepthwiseConv2dDesc(uint32_t strideX, uint32_t strideY,
+    uint32_t depthMultiplier = 1, uint32_t padLeft = 0, uint32_t padRight = 0,
+    uint32_t padTop = 0, uint32_t padBottom = 0)
+{
+    boost::ignore_unused(depthMultiplier);
+
+    armnn::DepthwiseConvolution2dDescriptor desc;
+
+    desc.m_PadLeft = padLeft;
+    desc.m_PadRight = padRight;
+
+    desc.m_PadTop = padTop;
+    desc.m_PadBottom = padBottom;
+    desc.m_StrideX = strideX;
+    desc.m_StrideY = strideY;
+    desc.m_BiasEnabled = false;
+
+    return desc;
+}
+
+armnn::TensorInfo CreateOutputTensorInfo(const armnn::TensorInfo& inputInfo,
+                                         const armnn::TensorInfo& weightsInfo,
+                                         const armnn::DepthwiseConvolution2dDescriptor& descriptor,
+                                         armnn::DataType dataType)
+{
+    const armnn::TensorShape& inputShape  = inputInfo.GetShape();
+    const armnn::TensorShape& filterShape = weightsInfo.GetShape();
+
+    unsigned int inWidth = inputShape[3];
+    unsigned int inHeight = inputShape[2];
+    unsigned int inBatchSize = inputShape[0];
+
+    unsigned int filterWidth = filterShape[3];
+    unsigned int readWidth = (inWidth + descriptor.m_PadLeft + descriptor.m_PadRight) - (filterWidth);
+    unsigned int outWidth =  1u + (readWidth / descriptor.m_StrideX);
+
+    unsigned int filterHeight = filterShape[2];
+    unsigned int readHeight = (inHeight + descriptor.m_PadTop + descriptor.m_PadBottom) - (filterHeight);
+    unsigned int outHeight = 1u + (readHeight / descriptor.m_StrideY);
+    unsigned int depthMultiplier = filterShape[0];
+
+    unsigned int outChannels = filterShape[1] * depthMultiplier;
+    unsigned int outBatchSize = inBatchSize;
+
+    armnn::TensorShape outputShape({outBatchSize, outChannels, outHeight, outWidth});
+    return armnn::TensorInfo(outputShape, dataType);
+}
+}
+
+BOOST_AUTO_TEST_CASE(DepthwiseConv2dUtils)
+{
+    const armnn::DataType dataType = armnn::DataType::Float32;
+
+    armnn::TensorInfo inputInfo({1, 1, 10, 10 }, dataType);
+    armnn::TensorInfo outputInfo;
+    armnn::TensorInfo weightsInfo3x3({ 1, 1, 3, 3 }, dataType);
+    armnn::TensorInfo biasesInfo;
+
+    armnn::DepthwiseConvolution2dDescriptor descriptor;
+
+    // Strides supported: 1,2,3
+    descriptor = MakeDepthwiseConv2dDesc(1, 1);
+    outputInfo = CreateOutputTensorInfo(inputInfo, weightsInfo3x3, descriptor, dataType);
+    BOOST_TEST(armnn::IsDepthwiseConvolutionSupportedNeon(inputInfo, outputInfo, descriptor,
+                                                          weightsInfo3x3, biasesInfo));
+
+    descriptor = MakeDepthwiseConv2dDesc(1, 2);
+    outputInfo = CreateOutputTensorInfo(inputInfo, weightsInfo3x3, descriptor, dataType);
+    BOOST_TEST(armnn::IsDepthwiseConvolutionSupportedNeon(inputInfo, outputInfo, descriptor,
+                                                          weightsInfo3x3, biasesInfo));
+
+    descriptor = MakeDepthwiseConv2dDesc(1, 3);
+    outputInfo = CreateOutputTensorInfo(inputInfo, weightsInfo3x3, descriptor, dataType);
+    BOOST_TEST(armnn::IsDepthwiseConvolutionSupportedNeon(inputInfo, outputInfo, descriptor,
+                                                          weightsInfo3x3, biasesInfo));
+
+    descriptor = MakeDepthwiseConv2dDesc(2, 1);
+    outputInfo = CreateOutputTensorInfo(inputInfo, weightsInfo3x3, descriptor, dataType);
+    BOOST_TEST(armnn::IsDepthwiseConvolutionSupportedNeon(inputInfo, outputInfo, descriptor,
+                                                          weightsInfo3x3, biasesInfo));
+
+    descriptor = MakeDepthwiseConv2dDesc(2, 2);
+    outputInfo = CreateOutputTensorInfo(inputInfo, weightsInfo3x3, descriptor, dataType);
+    BOOST_TEST(armnn::IsDepthwiseConvolutionSupportedNeon(inputInfo, outputInfo, descriptor,
+                                                          weightsInfo3x3, biasesInfo));
+
+    descriptor = MakeDepthwiseConv2dDesc(2, 3);
+    outputInfo = CreateOutputTensorInfo(inputInfo, weightsInfo3x3, descriptor, dataType);
+    BOOST_TEST(armnn::IsDepthwiseConvolutionSupportedNeon(inputInfo, outputInfo, descriptor,
+                                                          weightsInfo3x3, biasesInfo));
+
+    descriptor = MakeDepthwiseConv2dDesc(3, 1);
+    outputInfo = CreateOutputTensorInfo(inputInfo, weightsInfo3x3, descriptor, dataType);
+    BOOST_TEST(armnn::IsDepthwiseConvolutionSupportedNeon(inputInfo, outputInfo, descriptor,
+                                                          weightsInfo3x3, biasesInfo));
+
+    descriptor = MakeDepthwiseConv2dDesc(3, 2);
+    outputInfo = CreateOutputTensorInfo(inputInfo, weightsInfo3x3, descriptor, dataType);
+    BOOST_TEST(armnn::IsDepthwiseConvolutionSupportedNeon(inputInfo, outputInfo, descriptor,
+                                                          weightsInfo3x3, biasesInfo));
+
+    descriptor = MakeDepthwiseConv2dDesc(3, 3);
+    outputInfo = CreateOutputTensorInfo(inputInfo, weightsInfo3x3, descriptor, dataType);
+    BOOST_TEST(armnn::IsDepthwiseConvolutionSupportedNeon(inputInfo, outputInfo, descriptor,
+                                                          weightsInfo3x3, biasesInfo));
+
+    // Supported stride 4
+    descriptor = MakeDepthwiseConv2dDesc(4, 1);
+    outputInfo = CreateOutputTensorInfo(inputInfo, weightsInfo3x3, descriptor, dataType);
+    BOOST_TEST(armnn::IsDepthwiseConvolutionSupportedNeon(inputInfo, outputInfo, descriptor,
+                                                          weightsInfo3x3, biasesInfo));
+
+    // Supported weights shape 1x1
+    armnn::TensorInfo weightsInfo1x1({ 1, 1, 1, 1 }, armnn::DataType::Float32);
+    descriptor = MakeDepthwiseConv2dDesc(1, 1);
+    outputInfo = CreateOutputTensorInfo(inputInfo, weightsInfo1x1, descriptor, dataType);
+    BOOST_TEST(armnn::IsDepthwiseConvolutionSupportedNeon(inputInfo, outputInfo, descriptor,
+                                                          weightsInfo1x1, biasesInfo));
+
+    // Supported shape 2x2
+    armnn::TensorInfo weightsInfo2x2({ 1, 1, 2, 2 }, armnn::DataType::Float32);
+    descriptor = MakeDepthwiseConv2dDesc(1, 1);
+    outputInfo = CreateOutputTensorInfo(inputInfo, weightsInfo2x2, descriptor, dataType);
+    BOOST_TEST(armnn::IsDepthwiseConvolutionSupportedNeon(inputInfo, outputInfo, descriptor,
+                                                          weightsInfo2x2, biasesInfo));
+
+    // Asymmetric padding
+    descriptor = MakeDepthwiseConv2dDesc(1, 1, 1, 1, 2, 1, 2);
+    outputInfo = CreateOutputTensorInfo(inputInfo, weightsInfo3x3, descriptor, dataType);
+    BOOST_TEST(armnn::IsDepthwiseConvolutionSupportedNeon(inputInfo, outputInfo, descriptor,
+                                                          weightsInfo3x3, biasesInfo));
+}
+
+// Pooling
+ARMNN_AUTO_TEST_CASE(SimpleMaxPooling2dSize3x3Stride2x4, SimpleMaxPooling2dSize3x3Stride2x4Test, true)
+ARMNN_AUTO_TEST_CASE(SimpleMaxPooling2dSize3x3Stride2x4Uint8, SimpleMaxPooling2dSize3x3Stride2x4Uint8Test, true)
+ARMNN_AUTO_TEST_CASE(SimpleAveragePooling2d, SimpleAveragePooling2dTest)
+ARMNN_AUTO_TEST_CASE(SimpleAveragePooling2dNhwc, SimpleAveragePooling2dNhwcTest)
+ARMNN_AUTO_TEST_CASE(SimpleAveragePooling2dUint8, SimpleAveragePooling2dUint8Test)
+
+ARMNN_AUTO_TEST_CASE(LargeTensorsAveragePooling2d, LargeTensorsAveragePooling2dTest)
+ARMNN_AUTO_TEST_CASE(LargeTensorsAveragePooling2dUint8, LargeTensorsAveragePooling2dUint8Test)
+
+ARMNN_AUTO_TEST_CASE(SimpleL2Pooling2d, SimpleL2Pooling2dTest)
+ARMNN_AUTO_TEST_CASE(UNSUPPORTED_SimpleL2Pooling2dUint8, SimpleL2Pooling2dUint8Test)
+ARMNN_AUTO_TEST_CASE(L2Pooling2dSize3Stride1, L2Pooling2dSize3Stride1Test)
+ARMNN_AUTO_TEST_CASE(UNSUPPORTED_L2Pooling2dSize3Stride1Uint8, L2Pooling2dSize3Stride1Uint8Test)
+ARMNN_AUTO_TEST_CASE(L2Pooling2dSize3Stride3, L2Pooling2dSize3Stride3Test)
+ARMNN_AUTO_TEST_CASE(UNSUPPORTED_L2Pooling2dSize3Stride3Uint8, L2Pooling2dSize3Stride3Uint8Test)
+ARMNN_AUTO_TEST_CASE(L2Pooling2dSize3Stride4, L2Pooling2dSize3Stride4Test)
+ARMNN_AUTO_TEST_CASE(UNSUPPORTED_L2Pooling2dSize3Stride4Uint8, L2Pooling2dSize3Stride4Uint8Test)
+ARMNN_AUTO_TEST_CASE(L2Pooling2dSize7, L2Pooling2dSize7Test)
+ARMNN_AUTO_TEST_CASE(UNSUPPORTED_L2Pooling2dSize7Uint8, L2Pooling2dSize7Uint8Test)
+ARMNN_AUTO_TEST_CASE(L2Pooling2dSize9, L2Pooling2dSize9Test)
+ARMNN_AUTO_TEST_CASE(UNSUPPORTED_L2Pooling2dSize9Uint8, L2Pooling2dSize9Uint8Test)
+
+// Ignore padding values for pooling but count padding fields into the divisor
+ARMNN_AUTO_TEST_CASE(IgnorePaddingSimpleMaxPooling2d, IgnorePaddingSimpleMaxPooling2dTest)
+ARMNN_AUTO_TEST_CASE(IgnorePaddingSimpleMaxPooling2dUint8, IgnorePaddingSimpleMaxPooling2dUint8Test)
+ARMNN_AUTO_TEST_CASE(IgnorePaddingMaxPooling2dSize3, IgnorePaddingMaxPooling2dSize3Test)
+ARMNN_AUTO_TEST_CASE(IgnorePaddingMaxPooling2dSize3Uint8, IgnorePaddingMaxPooling2dSize3Uint8Test)
+
+ARMNN_AUTO_TEST_CASE(IgnorePaddingSimpleAveragePooling2d, IgnorePaddingSimpleAveragePooling2dTest)
+ARMNN_AUTO_TEST_CASE(IgnorePaddingSimpleAveragePooling2dUint8, IgnorePaddingSimpleAveragePooling2dUint8Test)
+ARMNN_AUTO_TEST_CASE(IgnorePaddingSimpleAveragePooling2dNoPadding, IgnorePaddingSimpleAveragePooling2dNoPaddingTest)
+ARMNN_AUTO_TEST_CASE(IgnorePaddingSimpleAveragePooling2dNoPaddingUint8,
+    IgnorePaddingSimpleAveragePooling2dNoPaddingUint8Test)
+ARMNN_AUTO_TEST_CASE(IgnorePaddingAveragePooling2dSize3, IgnorePaddingAveragePooling2dSize3Test)
+ARMNN_AUTO_TEST_CASE(IgnorePaddingAveragePooling2dSize3Uint8, IgnorePaddingAveragePooling2dSize3Uint8Test)
+ARMNN_AUTO_TEST_CASE(IgnorePaddingAveragePooling2dSize3x2Stride2x2,
+                             IgnorePaddingAveragePooling2dSize3x2Stride2x2Test, false)
+ARMNN_AUTO_TEST_CASE(IgnorePaddingAveragePooling2dSize3x2Stride2x2NoPadding,
+                             IgnorePaddingAveragePooling2dSize3x2Stride2x2Test,
+                                          true)
+
+ARMNN_AUTO_TEST_CASE(IgnorePaddingSimpleL2Pooling2d, IgnorePaddingSimpleL2Pooling2dTest)
+ARMNN_AUTO_TEST_CASE(UNSUPPORTED_IgnorePaddingSimpleL2Pooling2dUint8, IgnorePaddingSimpleL2Pooling2dUint8Test)
+ARMNN_AUTO_TEST_CASE(IgnorePaddingL2Pooling2dSize3, IgnorePaddingL2Pooling2dSize3Test)
+ARMNN_AUTO_TEST_CASE(UNSUPPORTED_IgnorePaddingL2Pooling2dSize3Uint8, IgnorePaddingL2Pooling2dSize3Uint8Test)
+
+// Activation
+ARMNN_AUTO_TEST_CASE(ConstantLinearActivation, ConstantLinearActivationTest)
+
+ARMNN_AUTO_TEST_CASE(SimpleSoftmaxBeta1, SimpleSoftmaxTest, 1.0f)
+ARMNN_AUTO_TEST_CASE(SimpleSoftmaxBeta2, SimpleSoftmaxTest, 2.0f)
+
+ARMNN_AUTO_TEST_CASE(SimpleSoftmaxBeta1Uint8, SimpleSoftmaxUint8Test, 1.0f)
+ARMNN_AUTO_TEST_CASE(SimpleSoftmaxBeta2Uint8, SimpleSoftmaxUint8Test, 2.0f)
+
+ARMNN_AUTO_TEST_CASE(ReLu1Uint8, BoundedReLuUint8UpperAndLowerBoundTest)
+ARMNN_AUTO_TEST_CASE(ReLu6Uint8, BoundedReLuUint8UpperBoundOnlyTest)
+
+// Softmax
+BOOST_AUTO_TEST_CASE(Softmax4dSupport)
+{
+    const unsigned int numDimensions = 4u;
+    std::array<unsigned int, numDimensions> dimensionSizes;
+    dimensionSizes.fill(1u);
+
+    const armnn::TensorInfo inputInfo(numDimensions, &dimensionSizes.front(), armnn::DataType::Float32);
+    const armnn::TensorInfo outputInfo(numDimensions, &dimensionSizes.front(), armnn::DataType::Float32);
+
+    // 4D Softmax should be reported as unsupported on the NEON backend
+    BOOST_TEST(!armnn::IsSoftmaxSupportedNeon(inputInfo, outputInfo, armnn::SoftmaxDescriptor()));
+}
+
+// Splitter
+ARMNN_AUTO_TEST_CASE(SimpleSplitter, SplitterTest)
+ARMNN_AUTO_TEST_CASE(SimpleSplitterUint8, SplitterUint8Test)
+
+ARMNN_AUTO_TEST_CASE(CopyViaSplitter, CopyViaSplitterTest)
+ARMNN_AUTO_TEST_CASE(CopyViaSplitterUint8, CopyViaSplitterUint8Test)
+
+// Merger
+ARMNN_AUTO_TEST_CASE(SimpleMerger, MergerTest)
+ARMNN_AUTO_TEST_CASE(MergerUint8, MergerUint8Test)
+
+// Fully Connected
+ARMNN_AUTO_TEST_CASE(SimpleFullyConnected, FullyConnectedFloat32Test, false, false)
+ARMNN_AUTO_TEST_CASE(SimpleFullyConnectedWithBias, FullyConnectedFloat32Test, true, false)
+ARMNN_AUTO_TEST_CASE(SimpleFullyConnectedWithTranspose, FullyConnectedFloat32Test, false, true)
+ARMNN_AUTO_TEST_CASE(FullyConnectedLarge, FullyConnectedLargeTest, false)
+ARMNN_AUTO_TEST_CASE(FullyConnectedLargeTransposed, FullyConnectedLargeTest, true)
+ARMNN_AUTO_TEST_CASE(FullyConnectedUint8, FullyConnectedUint8Test, false)
+ARMNN_AUTO_TEST_CASE(FullyConnectedBiasedUint8, FullyConnectedUint8Test, true)
+
+// Add
+ARMNN_AUTO_TEST_CASE(SimpleAdd, AdditionTest)
+ARMNN_AUTO_TEST_CASE(AddBroadcast, AdditionBroadcastTest)
+ARMNN_AUTO_TEST_CASE(AddBroadcast1Element, AdditionBroadcast1ElementTest)
+
+// Sub
+ARMNN_AUTO_TEST_CASE(SimpleSub, SubtractionTest)
+
+// Mul
+ARMNN_AUTO_TEST_CASE(SimpleMultiplication, MultiplicationTest)
+ARMNN_AUTO_TEST_CASE(MultiplicationBroadcast1Element, MultiplicationBroadcast1ElementTest)
+ARMNN_AUTO_TEST_CASE(MultiplicationBroadcast1DVector, MultiplicationBroadcast1DVectorTest)
+
+// Batch Norm
+ARMNN_AUTO_TEST_CASE(BatchNorm, BatchNormTest)
+
+// Constant
+ARMNN_AUTO_TEST_CASE(Constant, ConstantTest)
+ARMNN_AUTO_TEST_CASE(ConstantUint8, ConstantTestUint8)
+
+// Concatenation
+ARMNN_AUTO_TEST_CASE(Concatenation1d, Concatenation1dTest)
+ARMNN_AUTO_TEST_CASE(Concatenation1dUint8, Concatenation1dUint8Test)
+
+ARMNN_AUTO_TEST_CASE(Concatenation2dDim0, Concatenation2dDim0Test)
+ARMNN_AUTO_TEST_CASE(Concatenation2dDim0Uint8, Concatenation2dDim0Uint8Test)
+ARMNN_AUTO_TEST_CASE(Concatenation2dDim1, Concatenation2dDim1Test)
+ARMNN_AUTO_TEST_CASE(Concatenation2dDim1Uint8, Concatenation2dDim1Uint8Test)
+
+ARMNN_AUTO_TEST_CASE(Concatenation2dDim0DiffInputDims, Concatenation2dDim0DiffInputDimsTest)
+ARMNN_AUTO_TEST_CASE(Concatenation2dDim0DiffInputDimsUint8, Concatenation2dDim0DiffInputDimsUint8Test)
+ARMNN_AUTO_TEST_CASE(Concatenation2dDim1DiffInputDims, Concatenation2dDim1DiffInputDimsTest)
+ARMNN_AUTO_TEST_CASE(Concatenation2dDim1DiffInputDimsUint8, Concatenation2dDim1DiffInputDimsUint8Test)
+
+ARMNN_AUTO_TEST_CASE(Concatenation3dDim0, Concatenation3dDim0Test)
+ARMNN_AUTO_TEST_CASE(Concatenation3dDim0Uint8, Concatenation3dDim0Uint8Test)
+ARMNN_AUTO_TEST_CASE(Concatenation3dDim1, Concatenation3dDim1Test)
+ARMNN_AUTO_TEST_CASE(Concatenation3dDim1Uint8, Concatenation3dDim1Uint8Test)
+ARMNN_AUTO_TEST_CASE(Concatenation3dDim2, Concatenation3dDim2Test)
+ARMNN_AUTO_TEST_CASE(Concatenation3dDim2Uint8, Concatenation3dDim2Uint8Test)
+
+ARMNN_AUTO_TEST_CASE(Concatenation3dDim0DiffInputDims, Concatenation3dDim0DiffInputDimsTest)
+ARMNN_AUTO_TEST_CASE(Concatenation3dDim0DiffInputDimsUint8, Concatenation3dDim0DiffInputDimsUint8Test)
+ARMNN_AUTO_TEST_CASE(Concatenation3dDim1DiffInputDims, Concatenation3dDim1DiffInputDimsTest)
+ARMNN_AUTO_TEST_CASE(Concatenation3dDim1DiffInputDimsUint8, Concatenation3dDim1DiffInputDimsUint8Test)
+ARMNN_AUTO_TEST_CASE(Concatenation3dDim2DiffInputDims, Concatenation3dDim2DiffInputDimsTest)
+ARMNN_AUTO_TEST_CASE(Concatenation3dDim2DiffInputDimsUint8, Concatenation3dDim2DiffInputDimsUint8Test)
+
+// L2 Normalization
+ARMNN_AUTO_TEST_CASE(L2Normalization1d, L2Normalization1dTest)
+ARMNN_AUTO_TEST_CASE(L2Normalization2d, L2Normalization2dTest)
+ARMNN_AUTO_TEST_CASE(L2Normalization3d, L2Normalization3dTest)
+ARMNN_AUTO_TEST_CASE(L2Normalization4d, L2Normalization4dTest)
+
+ARMNN_AUTO_TEST_CASE(L2Normalization1dNhwc, L2Normalization1dNhwcTest)
+ARMNN_AUTO_TEST_CASE(L2Normalization2dNhwc, L2Normalization2dNhwcTest)
+ARMNN_AUTO_TEST_CASE(L2Normalization3dNhwc, L2Normalization3dNhwcTest)
+ARMNN_AUTO_TEST_CASE(L2Normalization4dNhwc, L2Normalization4dNhwcTest)
+
+// Floor
+ARMNN_AUTO_TEST_CASE(SimpleFloor, SimpleFloorTest)
+
+// Reshape
+ARMNN_AUTO_TEST_CASE(SimpleReshapeFloat32, SimpleReshapeFloat32Test)
+ARMNN_AUTO_TEST_CASE(SimpleReshapeUint8, SimpleReshapeUint8Test)
+
+// Permute
+ARMNN_AUTO_TEST_CASE(SimplePermuteFloat32, SimplePermuteFloat32Test)
+ARMNN_AUTO_TEST_CASE(SimplePermuteUint8, SimplePermuteUint8Test)
+ARMNN_AUTO_TEST_CASE(PermuteFloat32ValueSet1, PermuteFloat32ValueSet1Test)
+ARMNN_AUTO_TEST_CASE(PermuteFloat32ValueSet2, PermuteFloat32ValueSet2Test)
+ARMNN_AUTO_TEST_CASE(PermuteFloat32ValueSet3, PermuteFloat32ValueSet3Test)
+
+// Lstm
+ARMNN_AUTO_TEST_CASE(LstmLayerFloat32WithCifgWithPeepholeNoProjection,
+                     LstmLayerFloat32WithCifgWithPeepholeNoProjectionTest)
+ARMNN_AUTO_TEST_CASE(LstmLayerFloat32NoCifgNoPeepholeNoProjection,
+                     LstmLayerFloat32NoCifgNoPeepholeNoProjectionTest)
+ARMNN_AUTO_TEST_CASE(LstmLayerFloat32NoCifgWithPeepholeWithProjection,
+                     LstmLayerFloat32NoCifgWithPeepholeWithProjectionTest)
+
+// Normalization
+ARMNN_AUTO_TEST_CASE(SimpleNormalizationAcross, SimpleNormalizationAcrossTest)
+ARMNN_AUTO_TEST_CASE(SimpleNormalizationWithin, SimpleNormalizationWithinTest)
+ARMNN_AUTO_TEST_CASE(SimpleNormalizationAcrossNhwc, SimpleNormalizationAcrossNhwcTest)
+
+// ============================================================================
+// COMPARE tests
+
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareConv2dWithReference, CompareConvolution2dTest)
+
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareDepthwiseConv2dWithReferenceFloat32, CompareDepthwiseConvolution2dTest<float>)
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareDepthwiseConv2dWithReferenceUint8, CompareDepthwiseConvolution2dTest<uint8_t>)
+
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareNormalizationWithinWithReference, CompareNormalizationTest,
+                                 armnn::NormalizationAlgorithmChannel::Within,
+                                 armnn::NormalizationAlgorithmMethod::LocalBrightness)
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareNormalizationAcrossWithReference, CompareNormalizationTest,
+                                 armnn::NormalizationAlgorithmChannel::Across,
+                                 armnn::NormalizationAlgorithmMethod::LocalBrightness)
+
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareMaxPooling2dWithReference, ComparePooling2dTest, armnn::PoolingAlgorithm::Max)
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareMaxPooling2dWithReferenceUint8, ComparePooling2dUint8Test,
+                                 armnn::PoolingAlgorithm::Max)
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareAveragePooling2dWithReference, ComparePooling2dTest,
+                                 armnn::PoolingAlgorithm::Average)
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareAveragePooling2dWithReferenceUint8, ComparePooling2dUint8Test,
+                                 armnn::PoolingAlgorithm::Average)
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareL2Pooling2dWithReference, ComparePooling2dTest, armnn::PoolingAlgorithm::L2)
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(UNSUPPORTED_CompareL2Pooling2dWithReferenceUint8, ComparePooling2dUint8Test,
+                                 armnn::PoolingAlgorithm::L2)
+
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareSoftmaxBeta1WithReference, CompareSoftmaxTest, 1.0f)
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareSoftmaxBeta2WithReference, CompareSoftmaxTest, 2.0f)
+
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareSoftmaxUint8Beta1WithReference, CompareSoftmaxUint8Test, 1.0f)
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareSoftmaxUint8Beta2WithReference, CompareSoftmaxUint8Test, 2.0f)
+
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareAddition, CompareAdditionTest)
+
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareMultiplicationWithReference, CompareMultiplicationTest)
+
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(CompareBatchNorm, CompareBatchNormTest)
+
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(ReLu1, CompareBoundedReLuTest, 1.0f, -1.0f)
+ARMNN_COMPARE_REF_AUTO_TEST_CASE(ReLu6, CompareBoundedReLuTest, 6.0f, 0.0f)
+
+// ============================================================================
+// FIXTURE tests
+
+ARMNN_COMPARE_REF_FIXTURE_TEST_CASE(CompareSigmoidActivationWithReference, ActivationFixture,
+                                    CompareActivationTest, armnn::ActivationFunction::Sigmoid, 5u)
+
+ARMNN_COMPARE_REF_FIXTURE_TEST_CASE(CompareTanhActivationWithReference, ActivationFixture,
+                                    CompareActivationTest, armnn::ActivationFunction::TanH, 5u)
+
+ARMNN_COMPARE_REF_FIXTURE_TEST_CASE(CompareLinearActivationWithReference, ActivationFixture,
+                                    CompareActivationTest, armnn::ActivationFunction::Linear, 5u)
+
+ARMNN_COMPARE_REF_FIXTURE_TEST_CASE(CompareReLuActivationWithReference, ActivationFixture,
+                                    CompareActivationTest, armnn::ActivationFunction::ReLu, 5u)
+
+ARMNN_COMPARE_REF_FIXTURE_TEST_CASE(CompareBoundedReLuActivationWithReference, ActivationFixture,
+                                    CompareActivationTest, armnn::ActivationFunction::BoundedReLu, 5u)
+ARMNN_COMPARE_REF_FIXTURE_TEST_CASE(CompareBoundedReLuActivationWithReferenceUint8, ActivationFixture,
+                                    CompareActivationUint8Test, armnn::ActivationFunction::BoundedReLu)
+
+ARMNN_COMPARE_REF_FIXTURE_TEST_CASE(CompareSoftReLuActivationWithReference, ActivationFixture,
+                                    CompareActivationTest, armnn::ActivationFunction::SoftReLu, 1u)
+
+ARMNN_COMPARE_REF_FIXTURE_TEST_CASE(CompareLeakyReLuActivationWithReference, ActivationFixture,
+                                    CompareActivationTest, armnn::ActivationFunction::LeakyReLu, 5u)
+
+ARMNN_COMPARE_REF_FIXTURE_TEST_CASE(CompareAbsActivationWithReference, ActivationFixture,
+                                    CompareActivationTest, armnn::ActivationFunction::Abs, 5u)
+
+ARMNN_COMPARE_REF_FIXTURE_TEST_CASE(CompareSqrtActivationWithReference, PositiveActivationFixture,
+                                    CompareActivationTest, armnn::ActivationFunction::Sqrt, 5u)
+
+ARMNN_COMPARE_REF_FIXTURE_TEST_CASE(CompareSquareActivationWithReference, ActivationFixture,
+                                    CompareActivationTest, armnn::ActivationFunction::Square, 5u)
+BOOST_AUTO_TEST_SUITE_END()