IVGCVSW-2325 Reduce duplication in LayerTests by reusing the ElementwiseTestHelper

* Refactored tests for element-wise operations to use ElementwiseTestHelper
* Moved the etasts for each operation in a separate file under backendsCommon/test/layerTests

Signed-off-by: Aron Virginas-Tar <Aron.Virginas-Tar@arm.com>
Change-Id: Icefb6b35df78b9619f69378229789505bf92670e

1 files changed, 617 insertions, 0 deletions

diff --git a/src/backends/backendsCommon/test/layerTests/AdditionTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/AdditionTestImpl.cpp
new file mode 100644
index 0000000000..c6d3982f92
--- /dev/null
+++ b/src/backends/backendsCommon/test/layerTests/AdditionTestImpl.cpp
@@ -0,0 +1,617 @@
+//
+// Copyright © 2017 Arm Ltd. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "AdditionTestImpl.hpp"
+
+#include "ElementwiseTestImpl.hpp"
+
+template<>
+std::unique_ptr<armnn::IWorkload> CreateWorkload<armnn::AdditionQueueDescriptor>(
+    const armnn::IWorkloadFactory& workloadFactory,
+    const armnn::WorkloadInfo& info,
+    const armnn::AdditionQueueDescriptor& descriptor)
+{
+    return workloadFactory.CreateAddition(descriptor, info);
+}
+
+LayerTestResult<float,4> AdditionTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    unsigned int batchSize = 2u;
+    unsigned int channels  = 2u;
+    unsigned int height    = 2u;
+    unsigned int width     = 3u;
+
+    unsigned int shape[] = { batchSize, channels, height, width };
+
+    std::vector<float> input1 =
+    {
+        0.0f, 2.0f, 1.0f,
+        0.2f, 1.0f, 2.0f,
+
+        1.0f, 2.0f, 1.0f,
+        0.2f, 1.0f, 2.0f,
+
+        0.0f, 2.0f, 1.0f,
+        4.2f, 1.0f, 2.0f,
+
+        0.0f, 0.0f, 1.0f,
+        0.2f, 1.0f, 2.0f,
+    };
+
+    std::vector<float> input2 =
+    {
+        1.0f, 2.0f,  1.0f,
+        0.0f, 1.0f,  2.0f,
+
+        1.0f, 2.0f, -2.0f,
+        0.2f, 1.0f,  2.0f,
+
+        0.0f, 2.0f,  1.0f,
+        4.2f, 0.0f, -3.0f,
+
+        0.0f, 0.0f,  1.0f,
+        0.7f, 1.0f,  5.0f,
+    };
+
+
+    std::vector<float> output
+    {
+        1.0f, 4.0f,  2.0f,
+        0.2f, 2.0f,  4.0f,
+
+        2.0f, 4.0f, -1.0f,
+        0.4f, 2.0f,  4.0f,
+
+        0.0f, 4.0f,  2.0f,
+        8.4f, 1.0f, -1.0f,
+
+        0.0f, 0.0f,  2.0f,
+        0.9f, 2.0f,  7.0f,
+    };
+
+    return ElementwiseTestHelper<4, armnn::AdditionQueueDescriptor, armnn::DataType::Float32>(
+        workloadFactory,
+        memoryManager,
+        shape,
+        input1,
+        shape,
+        input2,
+        shape,
+        output);
+}
+
+LayerTestResult<float, 5> Addition5dTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    unsigned int depth     = 2u;
+    unsigned int batchSize = 2u;
+    unsigned int channels  = 2u;
+    unsigned int height    = 2u;
+    unsigned int width     = 3u;
+
+    unsigned int shape[] = { depth, batchSize, channels, height, width };
+
+    std::vector<float> input1 =
+    {
+        2.6f, 4.0f, 4.4f,  2.7f, 4.6f, 2.8f,
+        2.3f, 1.9f, 3.4f,  2.9f, 2.2f, 4.5f,
+
+        2.8f, 1.9f, 2.3f,  2.6f, 4.7f, 3.5f,
+        0.4f, 1.5f, 2.1f,  0.7f, 5.0f, 1.1f,
+
+
+        1.0f, 2.7f, 0.0f,  0.6f, 0.8f, 0.9f,
+        1.0f, 2.6f, 0.4f,  3.8f, 0.4f, 0.8f,
+
+        0.5f, 4.3f, 3.1f,  4.4f, 0.7f, 1.4f,
+        0.4f, 4.4f, 0.7f,  0.6f, 4.7f, 1.2f,
+
+    };
+
+    std::vector<float> input2 =
+    {
+        4.4f, 3.0f, 1.0f,  0.0f, 3.9f, 3.1f,
+        1.7f, 2.9f, 1.3f,  0.4f, 0.4f, 4.3f,
+
+        4.5f, 0.2f, 2.2f,  4.1f, 3.9f, 3.0f,
+        0.1f, 2.5f, 4.1f,  4.6f, 1.5f, 0.0f,
+
+
+        0.5f, 4.9f, 2.5f,  1.5f, 3.4f, 4.5f,
+        2.0f, 3.0f, 4.9f,  1.6f, 2.4f, 3.4f,
+
+        3.6f, 1.8f, 1.3f,  2.6f, 2.1f, 4.8f,
+        2.0f, 4.3f, 4.0f,  0.2f, 0.6f, 4.4f,
+    };
+
+    std::vector<float> output =
+    {
+        7.0f, 7.0f, 5.4f,  2.7f, 8.5f, 5.9f,
+        4.0f, 4.8f, 4.7f,  3.3f, 2.6f, 8.8f,
+
+        7.3f, 2.1f, 4.5f,  6.7f, 8.6f, 6.5f,
+        0.5f, 4.0f, 6.2f,  5.3f, 6.5f, 1.1f,
+
+
+        1.5f, 7.6f, 2.5f,  2.1f, 4.2f, 5.4f,
+        3.0f, 5.6f, 5.3f,  5.4f, 2.8f, 4.2f,
+
+        4.1f, 6.1f, 4.4f,  7.0f, 2.8f, 6.2f,
+        2.4f, 8.7f, 4.7f,  0.8f, 5.3f, 5.6f,
+    };
+
+    return ElementwiseTestHelper<5, armnn::AdditionQueueDescriptor, armnn::DataType::Float32>(
+        workloadFactory,
+        memoryManager,
+        shape,
+        input1,
+        shape,
+        input2,
+        shape,
+        output);
+}
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 4> AdditionBroadcastTestImpl(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    float qScale,
+    int32_t qOffset)
+{
+    armnn::TensorInfo inputTensorInfo1 = armnn::TensorInfo({1, 3, 2, 1}, ArmnnType);
+    armnn::TensorInfo inputTensorInfo2 = armnn::TensorInfo({1, 1, 2, 3}, ArmnnType);
+    armnn::TensorInfo outputTensorInfo = armnn::TensorInfo({1, 3, 2, 3}, ArmnnType);
+
+    if (armnn::IsQuantizedType<T>())
+    {
+        inputTensorInfo1.SetQuantizationScale(qScale);
+        inputTensorInfo1.SetQuantizationOffset(qOffset);
+        inputTensorInfo2.SetQuantizationScale(qScale);
+        inputTensorInfo2.SetQuantizationOffset(qOffset);
+        outputTensorInfo.SetQuantizationScale(qScale);
+        outputTensorInfo.SetQuantizationOffset(qOffset);
+    }
+
+    auto input1 = MakeTensor<T, 4>(inputTensorInfo1, QuantizedVector<T>(qScale, qOffset,
+    {
+        0.0f,
+        1.0f,
+
+        2.0f,
+        3.0f,
+
+        4.0f,
+        5.0f,
+    }));
+
+    auto input2 = MakeTensor<T, 4>(inputTensorInfo2, QuantizedVector<T>(qScale, qOffset,
+    {
+        0.5f, 1.5f, 2.5f,
+        3.5f, 4.5f, 5.5f,
+    }));
+
+    LayerTestResult<T,4> ret(outputTensorInfo);
+    ret.outputExpected = MakeTensor<T, 4>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset,
+    {
+        0.5f, 1.5f, 2.5f,
+        4.5f, 5.5f, 6.5f,
+
+        2.5f, 3.5f, 4.5f,
+        6.5f, 7.5f, 8.5f,
+
+        4.5f, 5.5f, 6.5f,
+        8.5f, 9.5f, 10.5f,
+    }));
+
+    std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1);
+    std::unique_ptr<armnn::ITensorHandle> inputHandle2 = workloadFactory.CreateTensorHandle(inputTensorInfo2);
+    std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);
+
+    armnn::AdditionQueueDescriptor data;
+    armnn::WorkloadInfo info;
+    AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get());
+    AddInputToWorkload(data, info, inputTensorInfo2, inputHandle2.get());
+    AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
+
+    std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateAddition(data, info);
+
+    inputHandle1->Allocate();
+    inputHandle2->Allocate();
+    outputHandle->Allocate();
+
+    CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]);
+    CopyDataToITensorHandle(inputHandle2.get(), &input2[0][0][0][0]);
+
+    workload->PostAllocationConfigure();
+    workload->Execute();
+
+    CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get());
+
+    return ret;
+}
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 4> AdditionBroadcast1ElementTestImpl(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    float qScale,
+    int32_t qOffset)
+{
+    armnn::TensorInfo inputTensorInfo1 = armnn::TensorInfo({1, 3, 2, 3}, ArmnnType);
+    armnn::TensorInfo inputTensorInfo2 = armnn::TensorInfo({1, 1, 1, 1}, ArmnnType);
+    armnn::TensorInfo outputTensorInfo = armnn::TensorInfo({1, 3, 2, 3}, ArmnnType);
+
+    if (armnn::IsQuantizedType<T>())
+    {
+        inputTensorInfo1.SetQuantizationScale(qScale);
+        inputTensorInfo1.SetQuantizationOffset(qOffset);
+        inputTensorInfo2.SetQuantizationScale(qScale);
+        inputTensorInfo2.SetQuantizationOffset(qOffset);
+        outputTensorInfo.SetQuantizationScale(qScale);
+        outputTensorInfo.SetQuantizationOffset(qOffset);
+    }
+
+    auto input1 = MakeTensor<T, 4>(inputTensorInfo1, QuantizedVector<T>(qScale, qOffset,
+    {
+            0.0f,  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,
+    }));
+
+    auto input2 = MakeTensor<T, 4>(inputTensorInfo2, QuantizedVector<T>(qScale, qOffset,
+    {
+        0.5f,
+    }));
+
+    LayerTestResult<T,4> ret(outputTensorInfo);
+    ret.outputExpected = MakeTensor<T, 4>(outputTensorInfo, QuantizedVector<T>(qScale, qOffset,
+    {
+            0.5f,  1.5f,  2.5f,
+            3.5f,  4.5f,  5.5f,
+            6.5f,  7.5f,  8.5f,
+            9.5f, 10.5f, 11.5f,
+        12.5f, 13.5f, 14.5f,
+        15.5f, 16.5f, 17.5f,
+    }));
+
+    std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1);
+    std::unique_ptr<armnn::ITensorHandle> inputHandle2 = workloadFactory.CreateTensorHandle(inputTensorInfo2);
+    std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);
+
+    armnn::AdditionQueueDescriptor data;
+    armnn::WorkloadInfo info;
+    AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get());
+    AddInputToWorkload(data, info, inputTensorInfo2, inputHandle2.get());
+    AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
+
+    std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateAddition(data, info);
+
+    inputHandle1->Allocate();
+    inputHandle2->Allocate();
+    outputHandle->Allocate();
+
+    CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]);
+    CopyDataToITensorHandle(inputHandle2.get(), &input2[0][0][0][0]);
+
+    workload->PostAllocationConfigure();
+    workload->Execute();
+
+    CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get());
+
+    return ret;
+}
+
+LayerTestResult<float, 4> AdditionBroadcastTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    return AdditionBroadcastTestImpl<armnn::DataType::Float32>(
+        workloadFactory, memoryManager, 0.0f, 0);
+}
+
+LayerTestResult<uint8_t, 4> AdditionBroadcastUint8Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    return AdditionBroadcastTestImpl<armnn::DataType::QuantisedAsymm8>(
+        workloadFactory, memoryManager, 2.f, 0);
+}
+
+LayerTestResult<int16_t, 4> AdditionBroadcastInt16Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    return AdditionBroadcastTestImpl<armnn::DataType::QuantisedSymm16>(
+        workloadFactory, memoryManager, 2.f, 0);
+}
+
+LayerTestResult<float, 4> AdditionBroadcast1ElementTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    return AdditionBroadcast1ElementTestImpl<armnn::DataType::Float32>(
+        workloadFactory, memoryManager, 0.0f, 0);
+}
+
+LayerTestResult<uint8_t, 4> AdditionBroadcast1ElementUint8Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    return AdditionBroadcast1ElementTestImpl<armnn::DataType::QuantisedAsymm8>(
+        workloadFactory, memoryManager, 0.1333333f, 128);
+}
+
+LayerTestResult<int16_t, 4> AdditionBroadcast1ElementInt16Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    return AdditionBroadcast1ElementTestImpl<armnn::DataType::QuantisedSymm16>(
+        workloadFactory, memoryManager, 0.1333333f, 0);
+}
+
+LayerTestResult<uint8_t, 4> AdditionUint8Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    const unsigned int shape0[] = { 1, 2, 2, 3 };
+    const unsigned int shape1[] = { 1, 2, 2, 3 };
+
+    std::vector<uint8_t> input0(
+    {
+        63,  35,  77,  70,  56, 112, //  420, 224,  518,  469,  371, 763
+        203,  28, 252, 168, 245,  91  // 1400, 175, 1743, 1155, 1694, 616
+    });
+
+    std::vector<uint8_t> input1(
+    {
+        21,   7, 175, 231, 175, 210, // 126,   28, 1204, 1596, 1204, 1449
+        126, 161,  63,  21, 105, 126  // 861, 1106,  420,  126,  714,  861
+    });
+
+    std::vector<uint8_t> output(
+    {
+        81,  39, 249, 255, 228, 255, //  546,  252, 1722, 2065(clamped), 1575, 2212(clamped)
+        255, 186, 255, 186, 255, 214, // 2261(clamped), 1281, 2163(clamped), 1281, 2408(clamped), 1477
+    });
+
+    return ElementwiseTestHelper<4, armnn::AdditionQueueDescriptor, armnn::DataType::QuantisedAsymm8>(
+        workloadFactory,
+        memoryManager,
+        shape0,
+        input0,
+        7.0f,
+        3,
+        shape1,
+        input1,
+        7.0f,
+        3,
+        shape0,
+        output,
+        7.0f,
+        3);
+}
+
+LayerTestResult<int16_t, 4> AdditionInt16Test(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    const unsigned int shape0[] = { 1, 2, 2, 3 };
+    const unsigned int shape1[] = { 1, 2, 2, 3 };
+
+    std::vector<int16_t> input0 =
+    {
+        63,  35,  77,  70,  56, 112, //  441, 245,  539,  490,  392, 184
+        203,  28, 252, 168, 245,  91  // 1421, 196, 1764, 1176, 1715, 637
+    };
+
+    std::vector<int16_t> input1 =
+    {
+        21,   7, 175, 231, 175, 210, // 126,   28, 1204, 1596, 1204, 1449
+        126, 161,  63,  21, 105, 126  // 861, 1106,  420,  126,  714,  861
+    };
+
+    std::vector<int16_t> output =
+    {
+        84,  42, 252, 301, 231, 322, //  588,  294, 1764, 2107(clamped), 1617, 2254(clamped)
+        329, 189, 315, 189, 350, 217, // 2303(clamped), 1323, 2205(clamped), 1323, 2450(clamped), 1519
+    };
+
+    return ElementwiseTestHelper<4, armnn::AdditionQueueDescriptor, armnn::DataType::QuantisedSymm16>(
+        workloadFactory,
+        memoryManager,
+        shape0,
+        input0,
+        7.0f,
+        0,
+        shape1,
+        input1,
+        7.0f,
+        0,
+        shape0,
+        output,
+        7.0f,
+        0);
+}
+
+LayerTestResult<float, 4> AdditionAfterMaxPoolTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager)
+{
+    // Create Initial Tensor
+    // 1, 2, 3
+    // 4, 5, 6
+    // 7, 8, 9
+
+    armnn::TensorInfo poolingInputTensorInfo({ 1, 1, 3, 3}, armnn::DataType::Float32);
+    armnn::TensorInfo poolingOutputTensorInfo({ 1, 1, 2, 2}, armnn::DataType::Float32);
+
+    boost::multi_array<float, 4> poolingInput = MakeTensor<float,4>(poolingInputTensorInfo,
+                                                            {1, 2, 3,
+                                                             4, 5, 6,
+                                                             7, 8, 9
+                                                            });
+
+    std::unique_ptr<armnn::ITensorHandle> poolingInputHandle =
+            workloadFactory.CreateTensorHandle(poolingInputTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> poolingOutputHandle =
+            workloadFactory.CreateTensorHandle(poolingOutputTensorInfo);
+
+    // Apply MaxPool poolSize = 1x1, stride=2x2
+    // Result =
+    // 1, 3
+    // 7, 9
+    armnn::Pooling2dDescriptor descriptor;
+    descriptor.m_PoolHeight = 1;
+    descriptor.m_PoolWidth = 1;
+    descriptor.m_StrideX = 2;
+    descriptor.m_StrideY = 2;
+    descriptor.m_PoolType = armnn::PoolingAlgorithm::Max;
+
+    armnn::Pooling2dQueueDescriptor queueDescriptor;
+    queueDescriptor.m_Parameters = descriptor;
+    armnn::WorkloadInfo workloadInfo;
+    AddInputToWorkload(queueDescriptor, workloadInfo, poolingInputTensorInfo, poolingInputHandle.get());
+    AddOutputToWorkload(queueDescriptor, workloadInfo, poolingOutputTensorInfo, poolingOutputHandle.get());
+
+    // Create the MaxPool
+    std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreatePooling2d(queueDescriptor, workloadInfo);
+
+    //LayerTestResult<float, 4> result(poolingOutputTensorInfo);
+    auto shape( GetTensorShapeAsArray<4>(poolingOutputTensorInfo));
+    boost::multi_array<float, 4> resultMaxPool;
+    resultMaxPool.resize(shape);
+
+
+    // Create addition with another tensor the same size
+    // This would be the result to apply a Conv2d with kernel ones(2) and stride 1x1
+    // with the initial tensor.
+    // 12, 16
+    // 24, 28
+
+    armnn::TensorInfo addInputTensorInfo({ 1,1,2,2}, armnn::DataType::Float32);
+    armnn::TensorInfo addOutputTensorInfo({ 1,1,2,2}, armnn::DataType::Float32);
+
+    boost::multi_array<float, 4> addInput = MakeTensor<float,4>(addInputTensorInfo,
+                                                                    {12, 16,
+                                                                     24, 28,
+                                                                    });
+
+    // Expected output tensor after MaxPool and Addition.
+    LayerTestResult<float,4> addRet(addOutputTensorInfo);
+    addRet.outputExpected = MakeTensor<float, 4>(addOutputTensorInfo, std::vector<float>(
+    {
+        13, 19,
+        31, 37
+    }));
+
+    std::unique_ptr<armnn::ITensorHandle> addInputHandle = workloadFactory.CreateTensorHandle(addInputTensorInfo);
+    std::unique_ptr<armnn::ITensorHandle> addOutputHandle = workloadFactory.CreateTensorHandle(addOutputTensorInfo);
+
+    armnn::AdditionQueueDescriptor data;
+    armnn::WorkloadInfo info;
+
+    // Add the output of the MaxPool and the new tensor
+    AddInputToWorkload(data, info, poolingOutputTensorInfo, poolingOutputHandle.get());
+    AddInputToWorkload(data, info, addInputTensorInfo, addInputHandle.get());
+    AddOutputToWorkload(data, info, addOutputTensorInfo, addOutputHandle.get());
+
+    std::unique_ptr<armnn::IWorkload> addWorkload = workloadFactory.CreateAddition(data, info);
+
+    poolingInputHandle->Allocate();
+    poolingOutputHandle->Allocate();
+    addInputHandle->Allocate();
+    addOutputHandle->Allocate();
+
+    CopyDataToITensorHandle(poolingInputHandle.get(), &poolingInput[0][0][0][0]);
+    CopyDataFromITensorHandle(&resultMaxPool[0][0][0][0], poolingOutputHandle.get());
+
+    CopyDataToITensorHandle(poolingOutputHandle.get(), &resultMaxPool[0][0][0][0]);
+    CopyDataToITensorHandle(addInputHandle.get(), &addInput[0][0][0][0]);
+
+    workload->PostAllocationConfigure();
+    workload->Execute();
+    addWorkload->PostAllocationConfigure();
+    addWorkload->Execute();
+
+    CopyDataFromITensorHandle(&addRet.output[0][0][0][0], addOutputHandle.get());
+
+    return addRet;
+}
+
+LayerTestResult<float,4> CompareAdditionTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    armnn::IWorkloadFactory& refWorkloadFactory)
+{
+    unsigned int batchSize = 4;
+    unsigned int channels  = 1;
+    unsigned int height    = 2;
+    unsigned int width     = 3;
+
+    armnn::TensorInfo inputTensorInfo1, inputTensorInfo2;
+    armnn::TensorInfo outputTensorInfo;
+
+    unsigned int shape[] = {batchSize, channels, height, width};
+
+    inputTensorInfo1 = armnn::TensorInfo(4, shape, armnn::DataType::Float32);
+    inputTensorInfo2 = armnn::TensorInfo(4, shape, armnn::DataType::Float32);
+    outputTensorInfo = armnn::TensorInfo(4, shape, armnn::DataType::Float32);
+
+    auto input1 = MakeRandomTensor<float, 4>(inputTensorInfo1, 1232);
+    auto input2 = MakeRandomTensor<float, 4>(inputTensorInfo2, 456);
+
+    LayerTestResult<float,4> ret(outputTensorInfo);
+
+    std::unique_ptr<armnn::ITensorHandle> inputHandle1 = workloadFactory.CreateTensorHandle(inputTensorInfo1);
+    std::unique_ptr<armnn::ITensorHandle> inputHandle2 = workloadFactory.CreateTensorHandle(inputTensorInfo2);
+    std::unique_ptr<armnn::ITensorHandle> outputHandle = workloadFactory.CreateTensorHandle(outputTensorInfo);
+
+    std::unique_ptr<armnn::ITensorHandle> inputHandle1Ref = refWorkloadFactory.CreateTensorHandle(inputTensorInfo1);
+    std::unique_ptr<armnn::ITensorHandle> inputHandle2Ref = refWorkloadFactory.CreateTensorHandle(inputTensorInfo2);
+    std::unique_ptr<armnn::ITensorHandle> outputHandleRef = refWorkloadFactory.CreateTensorHandle(outputTensorInfo);
+
+    armnn::AdditionQueueDescriptor data;
+    armnn::WorkloadInfo info;
+    AddInputToWorkload(data, info, inputTensorInfo1, inputHandle1.get());
+    AddInputToWorkload(data, info, inputTensorInfo2, inputHandle2.get());
+    AddOutputToWorkload(data, info, outputTensorInfo, outputHandle.get());
+
+    armnn::AdditionQueueDescriptor refData = data;
+    armnn::WorkloadInfo refInfo = info;
+    SetWorkloadInput(refData, refInfo, 0, inputTensorInfo1, inputHandle1Ref.get());
+    SetWorkloadInput(refData, refInfo, 1, inputTensorInfo2, inputHandle2Ref.get());
+    SetWorkloadOutput(refData, refInfo, 0, outputTensorInfo, outputHandleRef.get());
+
+    std::unique_ptr<armnn::IWorkload> workload = workloadFactory.CreateAddition(data, info);
+    std::unique_ptr<armnn::IWorkload> workloadRef = refWorkloadFactory.CreateAddition(refData, refInfo);
+
+    inputHandle1->Allocate();
+    inputHandle2->Allocate();
+    outputHandle->Allocate();
+    inputHandle1Ref->Allocate();
+    inputHandle2Ref->Allocate();
+    outputHandleRef->Allocate();
+
+    CopyDataToITensorHandle(inputHandle1.get(), &input1[0][0][0][0]);
+    CopyDataToITensorHandle(inputHandle2.get(), &input2[0][0][0][0]);
+    CopyDataToITensorHandle(inputHandle1Ref.get(), &input1[0][0][0][0]);
+    CopyDataToITensorHandle(inputHandle2Ref.get(), &input2[0][0][0][0]);
+
+    workload->PostAllocationConfigure();
+    workload->Execute();
+    workloadRef->PostAllocationConfigure();
+    workloadRef->Execute();
+
+    CopyDataFromITensorHandle(&ret.output[0][0][0][0], outputHandle.get());
+    CopyDataFromITensorHandle(&ret.outputExpected[0][0][0][0], outputHandleRef.get());
+
+    return ret;
+}