IVGCVSW-7109: Add Batch MatMul front end support - Reference

  * Descriptors added for BatchMatMul
  * Layer definition added
  * Input validation added (will likely change when opt. param support comes in)
  * Ref workload implementation for BatchMatMul added (will also change with opt. param support)
  * Ref layer tests made for BatchMatMul
  * CMake and other build files updated

Signed-off-by: Samuel Yap <samuel.yap@arm.com>
Change-Id: Ic885301da543ee0fbe7922b85e7f9658c4efc617

9 files changed, 1352 insertions, 0 deletions

diff --git a/src/backends/backendsCommon/LayerSupportRules.hpp b/src/backends/backendsCommon/LayerSupportRules.hpp
index e616ecf022..a83fd62867 100644
--- a/src/backends/backendsCommon/LayerSupportRules.hpp
+++ b/src/backends/backendsCommon/LayerSupportRules.hpp
@@ -186,4 +186,12 @@ struct TensorNumDimensionsAreCorrect : public Rule
     }
 };
 
+struct TensorNumDimensionsAreGreaterOrEqualTo : public Rule
+{
+    TensorNumDimensionsAreGreaterOrEqualTo(const TensorInfo& info, unsigned int numDimensionsToCompare)
+    {
+        m_Res = info.GetNumDimensions() >= numDimensionsToCompare;
+    }
+};
+
 } //namespace armnn
\ No newline at end of file
diff --git a/src/backends/backendsCommon/WorkloadData.cpp b/src/backends/backendsCommon/WorkloadData.cpp
index 606821b5e5..9a4c60f551 100644
--- a/src/backends/backendsCommon/WorkloadData.cpp
+++ b/src/backends/backendsCommon/WorkloadData.cpp
@@ -4143,5 +4143,232 @@ void UnidirectionalSequenceLstmQueueDescriptor::Validate(const WorkloadInfo& wor
     }
 }
 
+void BatchMatMulQueueDescriptor::Validate(const WorkloadInfo& workloadInfo) const
+{
+    const std::string descriptorName{"BatchMatMulDescriptor"};
+
+    ValidateNumInputs(workloadInfo,  descriptorName, 2);
+    ValidateNumOutputs(workloadInfo, descriptorName, 1);
+
+    // Inputs must be: both 2D+
+    // For inputs X and Y whose dimensions to be multiplied are (M,N) and (I,J) respectively,
+    // axes N and I must be the same size
+
+    const auto& inputTensorXInfo = workloadInfo.m_InputTensorInfos[0];
+    const auto& inputTensorYInfo = workloadInfo.m_InputTensorInfos[1];
+    const auto& outputTensorInfo = workloadInfo.m_OutputTensorInfos[0];
+
+    std::vector<DataType> supportedTypes =
+    {
+        DataType::BFloat16,
+        DataType::Float16,
+        DataType::Float32,
+        DataType::QAsymmS8,
+        DataType::QAsymmU8,
+        DataType::QSymmS16
+    };
+
+    ValidateDataTypes(inputTensorXInfo, supportedTypes, descriptorName);
+    ValidateDataTypes(inputTensorYInfo, supportedTypes, descriptorName);
+    ValidateDataTypes(outputTensorInfo, supportedTypes, descriptorName);
+
+    if ((inputTensorXInfo.GetNumDimensions() < 2) ||
+        (inputTensorYInfo.GetNumDimensions() < 2))
+    {
+        throw InvalidArgumentException(descriptorName + ": Input tensors are not 2D or greater.");
+    }
+
+    if(m_Parameters.m_DataLayoutX.has_value())
+    {
+        switch(m_Parameters.m_DataLayoutX.value())
+        {
+            case DataLayout::NCHW:
+            case DataLayout::NHWC:
+                if(inputTensorXInfo.GetNumDimensions() != 4)
+                {
+                    throw InvalidArgumentException(descriptorName +
+                        ": Input tensor X does not have the correct "
+                        "number of dimensions for the Data Layout that it has been assigned.");
+                }
+                break;
+            case DataLayout::NCDHW:
+            case DataLayout::NDHWC:
+                if(inputTensorXInfo.GetNumDimensions() != 5)
+                {
+                    throw InvalidArgumentException(descriptorName +
+                        ": Input tensor X does not have the correct "
+                        "number of dimensions for the Data Layout that it has been assigned.");
+                }
+                break;
+            default:
+                break;
+        }
+    }
+
+    if(m_Parameters.m_DataLayoutY.has_value())
+    {
+        switch(m_Parameters.m_DataLayoutY.value())
+        {
+            case DataLayout::NCHW:
+            case DataLayout::NHWC:
+                if(inputTensorYInfo.GetNumDimensions() != 4)
+                {
+                    throw InvalidArgumentException(descriptorName +
+                        ": Input tensor Y does not have the correct "
+                        "number of dimensions for the Data Layout that it has been assigned.");
+                }
+                break;
+            case DataLayout::NCDHW:
+            case DataLayout::NDHWC:
+                if(inputTensorYInfo.GetNumDimensions() != 5)
+                {
+                    throw InvalidArgumentException(descriptorName +
+                        ": Input tensor Y does not have the correct "
+                        "number of dimensions for the Data Layout that it has been assigned.");
+                }
+                break;
+            default:
+                break;
+        }
+    }
+
+    auto axesToMul = BatchMatMulDescriptor::GetAxesToMul(m_Parameters,
+                                                         inputTensorXInfo.GetShape(),
+                                                         inputTensorYInfo.GetShape());
+
+    if(inputTensorXInfo.GetShape()[axesToMul.first.second]
+       != inputTensorYInfo.GetShape()[axesToMul.second.first])
+    {
+        throw InvalidArgumentException(descriptorName +
+            ": The final axis of input tensor X must be the same size as "
+            "the second last axis of input tensor Y.");
+    }
+
+    auto axesNotMul = BatchMatMulDescriptor::GetAxesNotMul(m_Parameters,
+                                                           inputTensorXInfo.GetShape(),
+                                                           inputTensorYInfo.GetShape());
+
+    {   // Separate scope so we don't pollute the rest of the scope with our temp variables
+        // e.g. NHWC isnt compatible with NCHW as of now
+        DataLayout xLayout;
+        DataLayout yLayout;
+
+        if(m_Parameters.m_DataLayoutX == EmptyOptional())
+        {
+            xLayout = DataLayout::NCHW; // Not equivalent - I'm just concerned with the last 2 axes
+        }
+        else
+        {
+            xLayout = m_Parameters.m_DataLayoutX.value();
+        }
+
+        if(m_Parameters.m_DataLayoutY == EmptyOptional())
+        {
+            yLayout = DataLayout::NCHW;
+        }
+        else
+        {
+            yLayout = m_Parameters.m_DataLayoutY.value();
+        }
+
+        if(xLayout == DataLayout::NCHW || xLayout == DataLayout::NCDHW)
+        {
+            if(yLayout == DataLayout::NHWC || yLayout == DataLayout::NDHWC)
+            {
+                throw InvalidArgumentException(descriptorName +
+                    ": Invalid input tensor data layout combination.");
+            }
+        }
+        if(yLayout == DataLayout::NCHW || yLayout == DataLayout::NCDHW)
+        {
+            if(xLayout == DataLayout::NHWC || xLayout == DataLayout::NDHWC)
+            {
+                throw InvalidArgumentException(descriptorName +
+                    ": Invalid input tensor data layout combination.");
+            }
+        }
+    }
+
+    // Simulate aligning the ends of the matrix dims and prepending 1's to the beginning of the shorter one
+    unsigned int outputTensorDimSize = std::max(inputTensorXInfo.GetNumDimensions(),
+                                                inputTensorYInfo.GetNumDimensions());
+    if(outputTensorDimSize-2 > 0)
+    {
+        TensorInfo tiXNotMul = TensorInfo(TensorShape(outputTensorDimSize-2),
+                                          DataType::Float32);
+        TensorInfo tiYNotMul = TensorInfo(TensorShape(outputTensorDimSize-2),
+                                          DataType::Float32);
+        TensorInfo tiOutNotMul = TensorInfo(TensorShape(outputTensorDimSize-2),
+                                            DataType::Float32);
+
+        auto doAxisExtension = [&](std::vector<unsigned int> axisIndices, TensorInfo& ti)
+        {
+            auto sizeDiff = (outputTensorDimSize-2) - axisIndices.size();
+
+            for(unsigned int i = 0; i < sizeDiff; i++)
+            {
+                axisIndices.insert(axisIndices.begin(), 1);
+            }
+
+            for(unsigned int i = 0; i < ti.GetNumDimensions(); i++)
+            {
+                ti.GetShape()[i] = inputTensorXInfo.GetShape()[i];
+            }
+        };
+
+        doAxisExtension(axesNotMul.first, tiXNotMul);
+        doAxisExtension(axesNotMul.second, tiYNotMul);
+
+        for(unsigned int i = 0; i < tiOutNotMul.GetNumDimensions(); i++)
+        {
+            tiOutNotMul.GetShape()[i] = std::max(tiXNotMul.GetShape()[i],
+                                                 tiYNotMul.GetShape()[i]);
+        }
+
+        ValidateBroadcastTensorShapesMatch(tiXNotMul,
+                                           tiYNotMul,
+                                           tiOutNotMul,
+                                           descriptorName,
+                                           "input_X",
+                                           "input_Y");
+    }
+
+    // Also check descriptor parameter validity
+    // This will eventually be moved to the start of the function as explained below
+    if ((!m_Parameters.m_TransposeX.empty() && !m_Parameters.m_AdjointX.empty()) ||
+        (!m_Parameters.m_TransposeY.empty() && !m_Parameters.m_AdjointY.empty()))
+    {
+        throw InvalidArgumentException(descriptorName +
+            ": Invalid descriptor parameters - Transpose and Adjoint "
+            "vectors cannot both be true for a given input tensor.");
+    }
+
+    if(m_Parameters.m_TransposeX.size() != 0 && m_Parameters.m_TransposeX.size() != inputTensorXInfo.GetNumDimensions())
+    {
+        throw InvalidArgumentException(descriptorName +
+            ": Invalid descriptor parameter - Transpose X vector must be "
+            "the same size as tensor input X's dimensionality.");
+    }
+    if(m_Parameters.m_AdjointX.size() != 0 && m_Parameters.m_AdjointX.size() != inputTensorXInfo.GetNumDimensions())
+    {
+        throw InvalidArgumentException(descriptorName +
+            ": Invalid descriptor parameter - Adjoint X vector must be "
+            "the same size as tensor input X's dimensionality.");
+    }
+    if(m_Parameters.m_TransposeY.size() != 0 && m_Parameters.m_TransposeY.size() != inputTensorYInfo.GetNumDimensions())
+    {
+        throw InvalidArgumentException(descriptorName +
+            ": Invalid descriptor parameter - Transpose Y vector must be "
+            "the same size as tensor input Y's dimensionality.");
+    }
+    if(m_Parameters.m_AdjointY.size() != 0 && m_Parameters.m_AdjointY.size() != inputTensorXInfo.GetNumDimensions())
+    {
+        throw InvalidArgumentException(descriptorName +
+            ": Invalid descriptor parameter - Adjoint Y vector must be "
+            "the same size as tensor input Y's dimensionality.");
+    }
+    // Note: for adjoint/transpose, you'll need to do the validation atop the resultant permutation.
+}
+
 
 } // namespace armnn
\ No newline at end of file
diff --git a/src/backends/backendsCommon/WorkloadFactory.cpp b/src/backends/backendsCommon/WorkloadFactory.cpp
index 3660e6e721..70006e4f79 100644
--- a/src/backends/backendsCommon/WorkloadFactory.cpp
+++ b/src/backends/backendsCommon/WorkloadFactory.cpp
@@ -133,6 +133,22 @@ bool IWorkloadFactory::IsLayerConfigurationSupported(const BackendId& backendId,
                     reason);
             break;
         }
+        case LayerType::BatchMatMul:
+        {
+            auto cLayer = PolymorphicDowncast<const BatchMatMulLayer*>(&layer);
+            const BatchMatMulDescriptor& descriptor = cLayer->GetParameters();
+
+            const TensorInfo& input0 = layer.GetInputSlot(0).GetConnection()->GetTensorInfo();
+            const TensorInfo& input1 = layer.GetInputSlot(1).GetConnection()->GetTensorInfo();
+            const TensorInfo& output = layer.GetOutputSlot(0).GetTensorInfo();
+            result = layerSupportObject.IsBatchMatMulSupported(
+                            OverrideDataType(input0, dataType),
+                            OverrideDataType(input1, dataType),
+                            OverrideDataType(output, dataType),
+                            descriptor,
+                            reason);
+            break;
+        }
         case LayerType::BatchNormalization:
         {
             auto cLayer = PolymorphicDowncast<const BatchNormalizationLayer*>(&layer);
diff --git a/src/backends/backendsCommon/common.mk b/src/backends/backendsCommon/common.mk
index 86de7e331b..007cca57fa 100644
--- a/src/backends/backendsCommon/common.mk
+++ b/src/backends/backendsCommon/common.mk
@@ -46,6 +46,7 @@ COMMON_TEST_SOURCES := \
     test/layerTests/ActivationTestImpl.cpp \
     test/layerTests/AdditionTestImpl.cpp \
     test/layerTests/ArgMinMaxTestImpl.cpp \
+    test/layerTests/BatchMatMulTestImpl.cpp \
     test/layerTests/BatchNormalizationTestImpl.cpp \
     test/layerTests/CastTestImpl.cpp \
     test/layerTests/ChannelShuffleTestImpl.cpp \
diff --git a/src/backends/backendsCommon/test/CMakeLists.txt b/src/backends/backendsCommon/test/CMakeLists.txt
index 8beb7c4169..c5b97ebf4c 100644
--- a/src/backends/backendsCommon/test/CMakeLists.txt
+++ b/src/backends/backendsCommon/test/CMakeLists.txt
@@ -68,6 +68,8 @@ list(APPEND armnnBackendsCommonUnitTests_sources
     layerTests/AdditionTestImpl.hpp
     layerTests/ArgMinMaxTestImpl.cpp
     layerTests/ArgMinMaxTestImpl.hpp
+    layerTests/BatchMatMulTestImpl.cpp
+    layerTests/BatchMatMulTestImpl.hpp
     layerTests/BatchNormalizationTestImpl.cpp
     layerTests/BatchNormalizationTestImpl.hpp
     layerTests/BatchToSpaceNdTestImpl.hpp
diff --git a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
index ba8cfd5f68..5fdcd9c57a 100644
--- a/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
+++ b/src/backends/backendsCommon/test/IsLayerSupportedTestImpl.hpp
@@ -614,6 +614,8 @@ DECLARE_LAYER_POLICY_1_PARAM(Addition)
 
 DECLARE_LAYER_POLICY_2_PARAM(ArgMinMax)
 
+DECLARE_LAYER_POLICY_2_PARAM(BatchMatMul)
+
 DECLARE_LAYER_POLICY_2_PARAM(BatchNormalization)
 
 DECLARE_LAYER_POLICY_2_PARAM(BatchToSpaceNd)
diff --git a/src/backends/backendsCommon/test/LayerTests.hpp b/src/backends/backendsCommon/test/LayerTests.hpp
index 8d73027783..25435b24ec 100644
--- a/src/backends/backendsCommon/test/LayerTests.hpp
+++ b/src/backends/backendsCommon/test/LayerTests.hpp
@@ -9,6 +9,7 @@
 #include <backendsCommon/test/layerTests/ActivationTestImpl.hpp>
 #include <backendsCommon/test/layerTests/AdditionTestImpl.hpp>
 #include <backendsCommon/test/layerTests/ArgMinMaxTestImpl.hpp>
+#include <backendsCommon/test/layerTests/BatchMatMulTestImpl.hpp>
 #include <backendsCommon/test/layerTests/BatchNormalizationTestImpl.hpp>
 #include <backendsCommon/test/layerTests/BatchToSpaceNdTestImpl.hpp>
 #include <backendsCommon/test/layerTests/CastTestImpl.hpp>
diff --git a/src/backends/backendsCommon/test/layerTests/BatchMatMulTestImpl.cpp b/src/backends/backendsCommon/test/layerTests/BatchMatMulTestImpl.cpp
new file mode 100644
index 0000000000..41add6e6da
--- /dev/null
+++ b/src/backends/backendsCommon/test/layerTests/BatchMatMulTestImpl.cpp
@@ -0,0 +1,1010 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#include "BatchMatMulTestImpl.hpp"
+
+#include <armnn/backends/IBackendInternal.hpp>
+#include <armnn/backends/Workload.hpp>
+#include <armnn/backends/WorkloadData.hpp>
+#include <armnn/backends/WorkloadFactory.hpp>
+
+#include <armnnTestUtils/WorkloadTestUtils.hpp>
+#include <armnnUtils/QuantizeHelper.hpp>
+#include <armnnTestUtils/TensorCopyUtils.hpp>
+#include <armnn/Optional.hpp>
+
+
+template<armnn::DataType ArmnnType, typename T, std::size_t NumDims>
+LayerTestResult<T, NumDims> BatchMatMulTestImpl(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory,
+    armnn::BatchMatMulDescriptor descriptor,
+    const std::vector<T>& inputX,
+    const std::vector<T>& inputY,
+    const std::vector<T>& outputExpected,
+    const armnn::TensorInfo& inputXInfo,
+    const armnn::TensorInfo& inputYInfo,
+    const armnn::TensorInfo& outputInfo)
+{
+    std::vector<T> outputActual(outputInfo.GetNumElements());
+
+    std::unique_ptr<armnn::ITensorHandle> inputXHandle = tensorHandleFactory.CreateTensorHandle(inputXInfo);
+    std::unique_ptr<armnn::ITensorHandle> inputYHandle = tensorHandleFactory.CreateTensorHandle(inputYInfo);
+    std::unique_ptr<armnn::ITensorHandle> outputHandle = tensorHandleFactory.CreateTensorHandle(outputInfo);
+
+    armnn::BatchMatMulQueueDescriptor queueDescriptor;
+    queueDescriptor.m_Parameters = descriptor;
+    armnn::WorkloadInfo workloadInfo;
+
+    AddInputToWorkload(queueDescriptor, workloadInfo, inputXInfo, inputXHandle.get());
+    AddInputToWorkload(queueDescriptor, workloadInfo, inputYInfo, inputYHandle.get());
+    AddOutputToWorkload(queueDescriptor, workloadInfo, outputInfo, outputHandle.get());
+
+    auto workload = workloadFactory.CreateWorkload(armnn::LayerType::BatchMatMul, queueDescriptor, workloadInfo);
+
+    inputXHandle->Allocate();
+    inputYHandle->Allocate();
+    outputHandle->Allocate();
+
+    CopyDataToITensorHandle(inputXHandle.get(), inputX.data());
+    CopyDataToITensorHandle(inputYHandle.get(), inputY.data());
+
+    workload->PostAllocationConfigure();
+    ExecuteWorkload(*workload, memoryManager);
+
+    CopyDataFromITensorHandle(outputActual.data(), outputHandle.get());
+
+    return LayerTestResult<T, NumDims>(outputActual,
+                                       outputExpected,
+                                       outputHandle->GetShape(),
+                                       outputInfo.GetShape());
+}
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<T, 2> BatchMatMul2DSimpleTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing
+
+    float qScale = 0.0f;
+    int32_t qOffset = 0;
+
+    switch(ArmnnType)
+    {
+        case armnn::DataType::QAsymmS8:
+        case armnn::DataType::QAsymmU8:
+        case armnn::DataType::QSymmS16:
+            qScale = 1.0f;
+            break;
+        default:
+            break;
+    }
+
+    armnn::TensorInfo inputXInfo({2,2}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo inputYInfo({2,2}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo outputInfo({2,2}, ArmnnType, qScale, qOffset);
+
+    std::vector<T> inputX = armnnUtils::QuantizedVector<T>({
+       1, 2,
+       3, 4
+    }, qScale, qOffset);
+
+    std::vector<T> inputY = armnnUtils::QuantizedVector<T>({
+        5, 6,
+        7, 8
+    }, qScale, qOffset);
+
+    std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({
+        19, 22,
+        43, 50
+    }, qScale, qOffset);
+
+    return BatchMatMulTestImpl<ArmnnType, T, 2>(workloadFactory,
+                                             memoryManager,
+                                             tensorHandleFactory,
+                                             descriptor,
+                                             inputX,
+                                             inputY,
+                                             outputExpected,
+                                             inputXInfo,
+                                             inputYInfo,
+                                             outputInfo);
+}
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 2>
+BatchMatMul2DSimpleTest<armnn::DataType::BFloat16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 2>
+BatchMatMul2DSimpleTest<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 2>
+BatchMatMul2DSimpleTest<armnn::DataType::Float16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 2>
+BatchMatMul2DSimpleTest<armnn::DataType::QAsymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 2>
+BatchMatMul2DSimpleTest<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 2>
+BatchMatMul2DSimpleTest<armnn::DataType::QSymmS16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<T, 3> BatchMatMul3DSimpleTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing
+
+    float qScale = 0.0f;
+    int32_t qOffset = 0;
+
+    switch(ArmnnType)
+    {
+        case armnn::DataType::QAsymmS8:
+        case armnn::DataType::QAsymmU8:
+        case armnn::DataType::QSymmS16:
+            qScale = 1.0f;
+            break;
+        default:
+            break;
+    }
+
+    armnn::TensorInfo inputXInfo({1,2,2}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo inputYInfo({1,2,2}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo outputInfo({1,2,2}, ArmnnType, qScale, qOffset);
+
+    std::vector<T> inputX = armnnUtils::QuantizedVector<T>({
+       1, 2,
+       3, 4
+    }, qScale, qOffset);
+
+    std::vector<T> inputY = armnnUtils::QuantizedVector<T>({
+        5, 6,
+        7, 8
+    }, qScale, qOffset);
+
+    std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({
+        19, 22,
+        43, 50
+    },qScale, qOffset);
+
+    return BatchMatMulTestImpl<ArmnnType, T, 3>(workloadFactory,
+                                                memoryManager,
+                                                tensorHandleFactory,
+                                                descriptor,
+                                                inputX,
+                                                inputY,
+                                                outputExpected,
+                                                inputXInfo,
+                                                inputYInfo,
+                                                outputInfo);
+}
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 3>
+BatchMatMul3DSimpleTest<armnn::DataType::BFloat16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 3>
+BatchMatMul3DSimpleTest<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 3>
+BatchMatMul3DSimpleTest<armnn::DataType::Float16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 3>
+BatchMatMul3DSimpleTest<armnn::DataType::QAsymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 3>
+BatchMatMul3DSimpleTest<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 3>
+BatchMatMul3DSimpleTest<armnn::DataType::QSymmS16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<T, 4> BatchMatMulNCHWSimpleTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    auto descriptor = armnn::BatchMatMulDescriptor(
+        armnn::Optional<armnn::DataLayout>(armnn::DataLayout::NCHW),
+        armnn::Optional<armnn::DataLayout>(armnn::DataLayout::NCHW));
+
+    float qScale = 0.0f;
+    int32_t qOffset = 0;
+
+    switch(ArmnnType)
+    {
+        case armnn::DataType::QAsymmS8:
+        case armnn::DataType::QAsymmU8:
+        case armnn::DataType::QSymmS16:
+            qScale = 1.0f;
+            break;
+        default:
+            break;
+    }
+
+    armnn::TensorInfo inputXInfo({1,1,2,2}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo inputYInfo({1,1,2,2}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo outputInfo({1,1,2,2}, ArmnnType, qScale, qOffset);
+
+    std::vector<T> inputX = armnnUtils::QuantizedVector<T>({
+       1, 2,
+       3, 4
+    }, qScale, qOffset);
+
+    std::vector<T> inputY = armnnUtils::QuantizedVector<T>({
+        5, 6,
+        7, 8
+    }, qScale, qOffset);
+
+    std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({
+        19, 22,
+        43, 50
+    },qScale, qOffset);
+
+    return BatchMatMulTestImpl<ArmnnType, T, 4>(workloadFactory,
+                                                memoryManager,
+                                                tensorHandleFactory,
+                                                descriptor,
+                                                inputX,
+                                                inputY,
+                                                outputExpected,
+                                                inputXInfo,
+                                                inputYInfo,
+                                                outputInfo);
+}
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 4>
+BatchMatMulNCHWSimpleTest<armnn::DataType::BFloat16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 4>
+BatchMatMulNCHWSimpleTest<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 4>
+BatchMatMulNCHWSimpleTest<armnn::DataType::Float16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 4>
+BatchMatMulNCHWSimpleTest<armnn::DataType::QAsymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 4>
+BatchMatMulNCHWSimpleTest<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 4>
+BatchMatMulNCHWSimpleTest<armnn::DataType::QSymmS16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<T, 4> BatchMatMulNHWCSimpleTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    auto descriptor = armnn::BatchMatMulDescriptor(
+        armnn::Optional<armnn::DataLayout>(armnn::DataLayout::NHWC),
+        armnn::Optional<armnn::DataLayout>(armnn::DataLayout::NHWC));
+
+    float qScale = 0.0f;
+    int32_t qOffset = 0;
+
+    switch(ArmnnType)
+    {
+        case armnn::DataType::QAsymmS8:
+        case armnn::DataType::QAsymmU8:
+        case armnn::DataType::QSymmS16:
+            qScale = 1.0f;
+            break;
+        default:
+            break;
+    }
+
+    armnn::TensorInfo inputXInfo({1,2,2,1}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo inputYInfo({1,2,2,1}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo outputInfo({1,2,2,1}, ArmnnType, qScale, qOffset);
+
+    std::vector<T> inputX = armnnUtils::QuantizedVector<T>({
+       1, 2,
+       3, 4
+    }, qScale, qOffset);
+
+    std::vector<T> inputY = armnnUtils::QuantizedVector<T>({
+        5, 6,
+        7, 8
+    }, qScale, qOffset);
+
+    std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({
+        19, 22,
+        43, 50
+    },qScale, qOffset);
+
+    return BatchMatMulTestImpl<ArmnnType, T, 4>(workloadFactory,
+                                                memoryManager,
+                                                tensorHandleFactory,
+                                                descriptor,
+                                                inputX,
+                                                inputY,
+                                                outputExpected,
+                                                inputXInfo,
+                                                inputYInfo,
+                                                outputInfo);
+}
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 4>
+BatchMatMulNHWCSimpleTest<armnn::DataType::BFloat16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 4>
+BatchMatMulNHWCSimpleTest<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 4>
+BatchMatMulNHWCSimpleTest<armnn::DataType::Float16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 4>
+BatchMatMulNHWCSimpleTest<armnn::DataType::QAsymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 4>
+BatchMatMulNHWCSimpleTest<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 4>
+BatchMatMulNHWCSimpleTest<armnn::DataType::QSymmS16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<T, 3> BatchMatMul3DBatchTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing
+
+    float qScale = 0.0f;
+    int32_t qOffset = 0;
+
+    switch(ArmnnType)
+    {
+        case armnn::DataType::QAsymmS8:
+        case armnn::DataType::QAsymmU8:
+        case armnn::DataType::QSymmS16:
+            qScale = 1.0f;
+            break;
+        default:
+            break;
+    }
+
+    armnn::TensorInfo inputXInfo({2,2,2}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo inputYInfo({2,2,2}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo outputInfo({2,2,2}, ArmnnType, qScale, qOffset);
+
+    std::vector<T> inputX = armnnUtils::QuantizedVector<T>({
+       1, 2,
+       3, 4,
+
+       9, 10,
+       11, 12
+    }, qScale, qOffset);
+
+    std::vector<T> inputY = armnnUtils::QuantizedVector<T>({
+        5, 6,
+        7, 8,
+
+        13, 14,
+        15, 16
+    }, qScale, qOffset);
+
+    std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({
+        19, 22,
+        43, 50,
+
+        267, 286,
+        323, 346
+    },qScale, qOffset);
+
+    return BatchMatMulTestImpl<ArmnnType, T, 3>(workloadFactory,
+                                                memoryManager,
+                                                tensorHandleFactory,
+                                                descriptor,
+                                                inputX,
+                                                inputY,
+                                                outputExpected,
+                                                inputXInfo,
+                                                inputYInfo,
+                                                outputInfo);
+}
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 3>
+BatchMatMul3DBatchTest<armnn::DataType::BFloat16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 3>
+BatchMatMul3DBatchTest<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 3>
+BatchMatMul3DBatchTest<armnn::DataType::Float16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 3>
+BatchMatMul3DBatchTest<armnn::DataType::QAsymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 3>
+BatchMatMul3DBatchTest<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 3>
+BatchMatMul3DBatchTest<armnn::DataType::QSymmS16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<T, 3> BatchMatMul3DBroadcastTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing
+
+    float qScale = 0.0f;
+    int32_t qOffset = 0;
+
+    switch(ArmnnType)
+    {
+        case armnn::DataType::QAsymmS8:
+        case armnn::DataType::QAsymmU8:
+        case armnn::DataType::QSymmS16:
+            qScale = 1.0f;
+            break;
+        default:
+            break;
+    }
+
+    armnn::TensorInfo inputXInfo({2,2,2}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo inputYInfo({1,2,2}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo outputInfo({2,2,2}, ArmnnType, qScale, qOffset);
+
+    std::vector<T> inputX = armnnUtils::QuantizedVector<T>({
+       1, 2,
+       3, 4,
+
+       9, 10,
+       11, 12
+    }, qScale, qOffset);
+
+    std::vector<T> inputY = armnnUtils::QuantizedVector<T>({
+        13, 14,
+        15, 16
+    }, qScale, qOffset);
+
+    std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({
+        43, 46,
+        99, 106,
+
+        267, 286,
+        323, 346
+    },qScale, qOffset);
+
+    return BatchMatMulTestImpl<ArmnnType, T, 3>(workloadFactory,
+                                                memoryManager,
+                                                tensorHandleFactory,
+                                                descriptor,
+                                                inputX,
+                                                inputY,
+                                                outputExpected,
+                                                inputXInfo,
+                                                inputYInfo,
+                                                outputInfo);
+}
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 3>
+BatchMatMul3DBroadcastTest<armnn::DataType::BFloat16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 3>
+BatchMatMul3DBroadcastTest<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 3>
+BatchMatMul3DBroadcastTest<armnn::DataType::Float16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 3>
+BatchMatMul3DBroadcastTest<armnn::DataType::QAsymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 3>
+BatchMatMul3DBroadcastTest<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 3>
+BatchMatMul3DBroadcastTest<armnn::DataType::QSymmS16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<T, 3> BatchMatMul3D2DBroadcastTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing
+
+    float qScale = 0.0f;
+    int32_t qOffset = 0;
+
+    switch(ArmnnType)
+    {
+        case armnn::DataType::QAsymmS8:
+        case armnn::DataType::QAsymmU8:
+        case armnn::DataType::QSymmS16:
+            qScale = 1.0f;
+            break;
+        default:
+            break;
+    }
+
+    armnn::TensorInfo inputXInfo({2,2,2}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo inputYInfo({2,2}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo outputInfo({2,2,2}, ArmnnType, qScale, qOffset);
+
+    std::vector<T> inputX = armnnUtils::QuantizedVector<T>({
+       1, 2,
+       3, 4,
+
+       9, 10,
+       11, 12
+    }, qScale, qOffset);
+
+    std::vector<T> inputY = armnnUtils::QuantizedVector<T>({
+        13, 14,
+        15, 16
+    }, qScale, qOffset);
+
+    std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({
+        43, 46,
+        99, 106,
+
+        267, 286,
+        323, 346
+    },qScale, qOffset);
+
+    return BatchMatMulTestImpl<ArmnnType, T, 3>(workloadFactory,
+                                                memoryManager,
+                                                tensorHandleFactory,
+                                                descriptor,
+                                                inputX,
+                                                inputY,
+                                                outputExpected,
+                                                inputXInfo,
+                                                inputYInfo,
+                                                outputInfo);
+}
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 3>
+BatchMatMul3D2DBroadcastTest<armnn::DataType::BFloat16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 3>
+BatchMatMul3D2DBroadcastTest<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 3>
+BatchMatMul3D2DBroadcastTest<armnn::DataType::Float16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 3>
+BatchMatMul3D2DBroadcastTest<armnn::DataType::QAsymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 3>
+BatchMatMul3D2DBroadcastTest<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 3>
+BatchMatMul3D2DBroadcastTest<armnn::DataType::QSymmS16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<T, 5> BatchMatMulNDHWCNHWCTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    auto descriptor = armnn::BatchMatMulDescriptor(
+        armnn::Optional<armnn::DataLayout>(armnn::DataLayout::NDHWC),
+        armnn::Optional<armnn::DataLayout>(armnn::DataLayout::NHWC));
+
+    float qScale = 0.0f;
+    int32_t qOffset = 0;
+
+    switch(ArmnnType)
+    {
+        case armnn::DataType::QAsymmS8:
+        case armnn::DataType::QAsymmU8:
+        case armnn::DataType::QSymmS16:
+            qScale = 1.0f;
+            break;
+        default:
+            break;
+    }
+
+    armnn::TensorInfo inputXInfo({1,1,2,2,2}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo inputYInfo({1,2,2,2}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo outputInfo({1,1,2,2,2}, ArmnnType, qScale, qOffset);
+
+    std::vector<T> inputX = armnnUtils::QuantizedVector<T>({
+        1, 20,
+        3, 22,
+
+        2, 21,
+        4, 23
+    }, qScale, qOffset);
+
+    std::vector<T> inputY = armnnUtils::QuantizedVector<T>({
+        5, 24,
+        7, 26,
+
+        6, 25,
+        8, 27
+    }, qScale, qOffset);
+
+    std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({
+       23, 1030,
+       31, 1114,
+
+       34, 1079,
+       46, 1167
+    },qScale, qOffset);
+
+    return BatchMatMulTestImpl<ArmnnType, T, 5>(workloadFactory,
+                                                memoryManager,
+                                                tensorHandleFactory,
+                                                descriptor,
+                                                inputX,
+                                                inputY,
+                                                outputExpected,
+                                                inputXInfo,
+                                                inputYInfo,
+                                                outputInfo);
+}
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 5>
+BatchMatMulNDHWCNHWCTest<armnn::DataType::BFloat16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 5>
+BatchMatMulNDHWCNHWCTest<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 5>
+BatchMatMulNDHWCNHWCTest<armnn::DataType::Float16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 5>
+BatchMatMulNDHWCNHWCTest<armnn::DataType::QAsymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 5>
+BatchMatMulNDHWCNHWCTest<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 5>
+BatchMatMulNDHWCNHWCTest<armnn::DataType::QSymmS16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<T, 2> BatchMatMul2DTinyTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing
+
+    float qScale = 0.0f;
+    int32_t qOffset = 0;
+
+    switch(ArmnnType)
+    {
+        case armnn::DataType::QAsymmS8:
+        case armnn::DataType::QAsymmU8:
+        case armnn::DataType::QSymmS16:
+            qScale = 1.0f;
+            break;
+        default:
+            break;
+    }
+
+    armnn::TensorInfo inputXInfo({1,1}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo inputYInfo({1,1}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo outputInfo({1,1}, ArmnnType, qScale, qOffset);
+
+    std::vector<T> inputX = armnnUtils::QuantizedVector<T>({
+       3
+    }, qScale, qOffset);
+
+    std::vector<T> inputY = armnnUtils::QuantizedVector<T>({
+        5
+    }, qScale, qOffset);
+
+    std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({
+        15
+    }, qScale, qOffset);
+
+    return BatchMatMulTestImpl<ArmnnType, T, 2>(workloadFactory,
+                                             memoryManager,
+                                             tensorHandleFactory,
+                                             descriptor,
+                                             inputX,
+                                             inputY,
+                                             outputExpected,
+                                             inputXInfo,
+                                             inputYInfo,
+                                             outputInfo);
+}
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 2>
+BatchMatMul2DTinyTest<armnn::DataType::BFloat16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 2>
+BatchMatMul2DTinyTest<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 2>
+BatchMatMul2DTinyTest<armnn::DataType::Float16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 2>
+BatchMatMul2DTinyTest<armnn::DataType::QAsymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 2>
+BatchMatMul2DTinyTest<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 2>
+BatchMatMul2DTinyTest<armnn::DataType::QSymmS16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T>
+LayerTestResult<T, 3> BatchMatMul3DNonSquareTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory)
+{
+    auto descriptor = armnn::BatchMatMulDescriptor(); // Arbitrary layout with no transpose/adjointing
+
+    float qScale = 0.0f;
+    int32_t qOffset = 0;
+
+    switch(ArmnnType)
+    {
+        case armnn::DataType::QAsymmS8:
+        case armnn::DataType::QAsymmU8:
+        case armnn::DataType::QSymmS16:
+            qScale = 1.0f;
+            break;
+        default:
+            break;
+    }
+
+    armnn::TensorInfo inputXInfo({2,5,3}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo inputYInfo({2,3,4}, ArmnnType, qScale, qOffset);
+    armnn::TensorInfo outputInfo({2,5,4}, ArmnnType, qScale, qOffset);
+
+    std::vector<T> inputX = armnnUtils::QuantizedVector<T>({
+       8, 8, 4,
+       6, 1, 3,
+       8, 8, 3,
+       8, 9, 8,
+       5, 4, 4,
+
+       1, 8, 5,
+       7, 1, 1,
+       8, 7, 9,
+       3, 2, 7,
+       8, 5, 3
+    }, qScale, qOffset);
+
+    std::vector<T> inputY = armnnUtils::QuantizedVector<T>({
+        6, 2, 3, 2,
+        6, 2, 2, 8,
+        3, 7, 8, 1,
+
+        7, 2, 9, 5,
+        2, 3, 1, 3,
+        2, 7, 7, 5
+    }, qScale, qOffset);
+
+    std::vector<T> outputExpected = armnnUtils::QuantizedVector<T>({
+        108, 60, 72, 84,
+        51, 35, 44, 23,
+        105, 53, 64, 83,
+        126, 90, 106, 96,
+        66, 46, 55, 46,
+
+        33, 61, 52, 54,
+        53, 24, 71, 43,
+        88, 100, 142, 106,
+        39, 61, 78, 56,
+        72, 52, 98, 70
+    },qScale, qOffset);
+
+    return BatchMatMulTestImpl<ArmnnType, T, 3>(workloadFactory,
+                                                memoryManager,
+                                                tensorHandleFactory,
+                                                descriptor,
+                                                inputX,
+                                                inputY,
+                                                outputExpected,
+                                                inputXInfo,
+                                                inputYInfo,
+                                                outputInfo);
+}
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::BFloat16>, 3>
+BatchMatMul3DNonSquareTest<armnn::DataType::BFloat16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float32>, 3>
+BatchMatMul3DNonSquareTest<armnn::DataType::Float32>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::Float16>, 3>
+BatchMatMul3DNonSquareTest<armnn::DataType::Float16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmS8>, 3>
+BatchMatMul3DNonSquareTest<armnn::DataType::QAsymmS8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QAsymmU8>, 3>
+BatchMatMul3DNonSquareTest<armnn::DataType::QAsymmU8>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template LayerTestResult<armnn::ResolveType<armnn::DataType::QSymmS16>, 3>
+BatchMatMul3DNonSquareTest<armnn::DataType::QSymmS16>(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
\ No newline at end of file
diff --git a/src/backends/backendsCommon/test/layerTests/BatchMatMulTestImpl.hpp b/src/backends/backendsCommon/test/layerTests/BatchMatMulTestImpl.hpp
new file mode 100644
index 0000000000..9e2139667b
--- /dev/null
+++ b/src/backends/backendsCommon/test/layerTests/BatchMatMulTestImpl.hpp
@@ -0,0 +1,85 @@
+//
+// Copyright © 2022 Arm Ltd and Contributors. All rights reserved.
+// SPDX-License-Identifier: MIT
+//
+
+#pragma once
+
+#include <armnnTestUtils/LayerTestResult.hpp>
+
+#include <ResolveType.hpp>
+
+#include <armnn/backends/IBackendInternal.hpp>
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>, std::size_t NumDims>
+LayerTestResult<T, NumDims> BatchMatMulTestImpl(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory,
+    armnn::BatchMatMulDescriptor descriptor,
+    const std::vector<T>& inputX,
+    const std::vector<T>& inputY,
+    const std::vector<T>& outputExpected,
+    const armnn::TensorInfo& inputXInfo,
+    const armnn::TensorInfo& inputYInfo,
+    const armnn::TensorInfo& outputInfo);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 2> BatchMatMul2DSimpleTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 3> BatchMatMul3DSimpleTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 4> BatchMatMulNCHWSimpleTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 4> BatchMatMulNHWCSimpleTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 3> BatchMatMul3DBatchTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 3> BatchMatMul3DBroadcastTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 3> BatchMatMul3D2DBroadcastTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 5> BatchMatMulNDHWCNHWCTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 2> BatchMatMul2DTinyTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
+
+template<armnn::DataType ArmnnType, typename T = armnn::ResolveType<ArmnnType>>
+LayerTestResult<T, 3> BatchMatMul3DNonSquareTest(
+    armnn::IWorkloadFactory& workloadFactory,
+    const armnn::IBackendInternal::IMemoryManagerSharedPtr& memoryManager,
+    const armnn::ITensorHandleFactory& tensorHandleFactory);
\ No newline at end of file