src/backends/neon/NeonWorkloadFactory.cpp

//
// Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//

#include "NeonBackendId.hpp"
#include "NeonBackendModelContext.hpp"
#include "NeonTensorHandle.hpp"
#include "NeonWorkloadFactory.hpp"

#include <Layer.hpp>

#include <armnn/Utils.hpp>
#include <armnn/utility/IgnoreUnused.hpp>
#include <armnn/utility/NumericCast.hpp>
#include <armnn/utility/PolymorphicDowncast.hpp>

#include <backendsCommon/MakeWorkloadHelper.hpp>
#include <armnn/backends/MemCopyWorkload.hpp>
#include <backendsCommon/MemImportWorkload.hpp>
#include <armnn/backends/TensorHandle.hpp>

#include <neon/workloads/NeonWorkloadUtils.hpp>
#include <neon/workloads/NeonWorkloads.hpp>

namespace armnn
{

namespace
{
static const BackendId s_Id{NeonBackendId()};
}

bool NeonWorkloadFactory::IsLayerSupported(const Layer& layer,
                                           Optional<DataType> dataType,
                                           std::string& outReasonIfUnsupported)
{
    return IWorkloadFactory::IsLayerSupported(s_Id, layer, dataType, outReasonIfUnsupported);
}

bool NeonWorkloadFactory::IsLayerSupported(const IConnectableLayer& layer,
                                           Optional<DataType> dataType,
                                           std::string& outReasonIfUnsupported,
                                           const ModelOptions& modelOptions)
{
    return IWorkloadFactory::IsLayerSupported(s_Id, layer, dataType, outReasonIfUnsupported, modelOptions);
}

const BackendId& NeonWorkloadFactory::GetBackendId() const
{
    return s_Id;
}

void NeonWorkloadFactory::SetNumberOfThreads()
{
    if (m_ModelContextPtr)
    {
        const unsigned int MIN_THREADS = 1;
        const unsigned int MAX_THREADS = 64;

        // Set the number of threads to be used if the user has set NumberOfThreads param
        // Only set if within limit or valid input
        auto modelOptions = dynamic_cast<NeonBackendModelContext*>(m_ModelContextPtr.get());
        auto numberOfThreads = modelOptions->GetNumberOfThreads();

        if (numberOfThreads != 0 && numberOfThreads >= MIN_THREADS && numberOfThreads <= MAX_THREADS)
        {
            arm_compute::Scheduler::get().set_num_threads(numberOfThreads);
        }
    }
}

NeonWorkloadFactory::NeonWorkloadFactory(const std::shared_ptr<NeonMemoryManager>& memoryManager)
    : m_MemoryManager(memoryManager), m_ModelContextPtr(IBackendInternal::IBackendSpecificModelContextPtr{})
{
    SetNumberOfThreads();
}

NeonWorkloadFactory::NeonWorkloadFactory(const std::shared_ptr<NeonMemoryManager>& memoryManager,
                                         const IBackendInternal::IBackendSpecificModelContextPtr& modelContextPtr)
    : m_MemoryManager(memoryManager), m_ModelContextPtr(modelContextPtr)
{
    SetNumberOfThreads();
}

std::unique_ptr<ITensorHandle> NeonWorkloadFactory::CreateSubTensorHandle(ITensorHandle& parent,
    TensorShape const& subTensorShape,
    unsigned int const* subTensorOrigin) const
{
    const arm_compute::TensorShape shape = armcomputetensorutils::BuildArmComputeTensorShape(subTensorShape);

    arm_compute::Coordinates coords;
    coords.set_num_dimensions(subTensorShape.GetNumDimensions());
    for (unsigned int i = 0; i < subTensorShape.GetNumDimensions(); i++)
    {
        // Arm compute indexes tensor coords in reverse order.
        unsigned int revertedIndex = subTensorShape.GetNumDimensions() - i - 1;
        coords.set(i, armnn::numeric_cast<int>(subTensorOrigin[revertedIndex]));
    }

    const arm_compute::TensorShape parentShape = armcomputetensorutils::BuildArmComputeTensorShape(parent.GetShape());
    if (!::arm_compute::error_on_invalid_subtensor(__func__, __FILE__, __LINE__, parentShape, coords, shape))
    {
        return nullptr;
    }

    return std::make_unique<NeonSubTensorHandle>(
        PolymorphicDowncast<IAclTensorHandle*>(&parent), shape, coords);
}

std::unique_ptr<ITensorHandle> NeonWorkloadFactory::CreateTensorHandle(const TensorInfo& tensorInfo,
                                                                       const bool IsMemoryManaged) const
{
    auto tensorHandle = std::make_unique<NeonTensorHandle>(tensorInfo);
    if (IsMemoryManaged)
    {
        tensorHandle->SetMemoryGroup(m_MemoryManager->GetInterLayerMemoryGroup());
    }
    return tensorHandle;
}

std::unique_ptr<ITensorHandle> NeonWorkloadFactory::CreateTensorHandle(const TensorInfo& tensorInfo,
                                                                       DataLayout dataLayout,
                                                                       const bool IsMemoryManaged) const
{
    auto tensorHandle = std::make_unique<NeonTensorHandle>(tensorInfo, dataLayout);
    if (IsMemoryManaged)
    {
        tensorHandle->SetMemoryGroup(m_MemoryManager->GetInterLayerMemoryGroup());
    }
    return tensorHandle;
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateWorkload(LayerType type,
                                                               const QueueDescriptor& descriptor,
                                                               const WorkloadInfo& info) const
{
    switch(type)
    {
        case LayerType::Activation :
        {
            auto activationQueueDescriptor = PolymorphicDowncast<const ActivationQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonActivationWorkload>(*activationQueueDescriptor, info);
        }
        case LayerType::Addition :
        {
            auto additionQueueDescriptor = PolymorphicDowncast<const AdditionQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonAdditionWorkload>(*additionQueueDescriptor, info);
        }
        case LayerType::ArgMinMax :
        {
            auto argMinMaxQueueDescriptor = PolymorphicDowncast<const ArgMinMaxQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonArgMinMaxWorkload>(*argMinMaxQueueDescriptor, info);
        }
        case LayerType::BatchNormalization :
        {
            auto batchNormalizationQueueDescriptor
                    = PolymorphicDowncast<const BatchNormalizationQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonBatchNormalizationWorkload>(*batchNormalizationQueueDescriptor, info);
        }
        case LayerType::BatchToSpaceNd :
        {
            auto batchToSpaceNdQueueDescriptor
                    = PolymorphicDowncast<const BatchToSpaceNdQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonBatchToSpaceNdWorkload>(*batchToSpaceNdQueueDescriptor, info);
        }
        case LayerType::Cast :
        {
            auto castQueueDescriptor = PolymorphicDowncast<const CastQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonCastWorkload>(*castQueueDescriptor, info);
        }
        case LayerType::ChannelShuffle :
        {
            auto channelShuffleQueueDescriptor = PolymorphicDowncast<const ChannelShuffleQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonChannelShuffleWorkload>(*channelShuffleQueueDescriptor, info);
        }
        case LayerType::Comparison :
        {
            auto comparisonQueueDescriptor = PolymorphicDowncast<const ComparisonQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonComparisonWorkload>(*comparisonQueueDescriptor, info);
        }
        case LayerType::Concat :
        {
            auto concatQueueDescriptor = PolymorphicDowncast<const ConcatQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonConcatWorkload>(*concatQueueDescriptor, info);
        }
        case LayerType::Constant :
        {
            auto constantQueueDescriptor = PolymorphicDowncast<const ConstantQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonConstantWorkload>(*constantQueueDescriptor, info);
        }
        case LayerType::ConvertBf16ToFp32 :
        {
            auto convertBf16ToFp32QueueDescriptor
                    = PolymorphicDowncast<const ConvertBf16ToFp32QueueDescriptor*>(&descriptor);
            return std::make_unique<NeonConvertBf16ToFp32Workload>(*convertBf16ToFp32QueueDescriptor, info);
        }
        case LayerType::ConvertFp16ToFp32 :
        {
            auto convertFp16ToFp32QueueDescriptor
                    = PolymorphicDowncast<const ConvertFp16ToFp32QueueDescriptor*>(&descriptor);
            return std::make_unique<NeonConvertFp16ToFp32Workload>(*convertFp16ToFp32QueueDescriptor, info);
        }
        case LayerType::ConvertFp32ToBf16 :
        {
            auto convertFp32ToBf16QueueDescriptor
                    = PolymorphicDowncast<const ConvertFp32ToBf16QueueDescriptor*>(&descriptor);
            return std::make_unique<NeonConvertFp32ToBf16Workload>(*convertFp32ToBf16QueueDescriptor, info);
        }
        case LayerType::ConvertFp32ToFp16 :
        {
            auto convertFp32ToFp16QueueDescriptor
                    = PolymorphicDowncast<const ConvertFp32ToFp16QueueDescriptor*>(&descriptor);
            return std::make_unique<NeonConvertFp32ToFp16Workload>(*convertFp32ToFp16QueueDescriptor, info);
        }
        case LayerType::Convolution2d :
        {
            auto convolution2dQueueDescriptor = PolymorphicDowncast<const Convolution2dQueueDescriptor*>(&descriptor);

            bool isFastMathEnabled = false;
            if (m_ModelContextPtr)
            {
                if (m_ModelContextPtr.get() != nullptr)
                {
                    auto modelOptions = dynamic_cast<NeonBackendModelContext*>(m_ModelContextPtr.get());
                    if (modelOptions)
                    {
                        isFastMathEnabled = modelOptions->IsFastMathEnabled();
                    }
                }
            }
            return std::make_unique<NeonConvolution2dWorkload>(*convolution2dQueueDescriptor,
                                                               info,
                                                               m_MemoryManager->GetIntraLayerManager(),
                                                               isFastMathEnabled);
        }
        case LayerType::Convolution3d :
        {
            auto convolution3dQueueDescriptor = PolymorphicDowncast<const Convolution3dQueueDescriptor*>(&descriptor);

            bool isFastMathEnabled = false;
            if (m_ModelContextPtr)
            {
                if (m_ModelContextPtr.get() != nullptr)
                {
                    auto modelOptions = dynamic_cast<NeonBackendModelContext*>(m_ModelContextPtr.get());
                    if (modelOptions)
                    {
                        isFastMathEnabled = modelOptions->IsFastMathEnabled();
                    }
                }
            }
            return std::make_unique<NeonConvolution3dWorkload>(*convolution3dQueueDescriptor,
                                                               info,
                                                               m_MemoryManager->GetIntraLayerManager(),
                                                               isFastMathEnabled);
        }
        case LayerType::Debug :
        {
            auto debugQueueDescriptor = PolymorphicDowncast<const DebugQueueDescriptor*>(&descriptor);
            return MakeWorkloadHelper<NullWorkload, NullWorkload>(*debugQueueDescriptor, info);
        }
        case LayerType::DepthToSpace :
        {
            auto depthToSpaceQueueDescriptor = PolymorphicDowncast<const DepthToSpaceQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonDepthToSpaceWorkload>(*depthToSpaceQueueDescriptor, info);
        }
        case LayerType::DepthwiseConvolution2d :
        {
            auto depthwiseConvolution2dQueueDescriptor
                    = PolymorphicDowncast<const DepthwiseConvolution2dQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonDepthwiseConvolutionWorkload>(*depthwiseConvolution2dQueueDescriptor, info);
        }
        case LayerType::Dequantize :
        {
            auto dequantizeQueueDescriptor = PolymorphicDowncast<const DequantizeQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonDequantizeWorkload>(*dequantizeQueueDescriptor, info);
        }
        case LayerType::DetectionPostProcess :
        {
            auto detectionPostProcessQueueDescriptor
                    = PolymorphicDowncast<const DetectionPostProcessQueueDescriptor*>(&descriptor);
            return MakeWorkloadHelper<NullWorkload, NullWorkload>(*detectionPostProcessQueueDescriptor, info);
        }
        case LayerType::Division :
        {
            auto divisionQueueDescriptor = PolymorphicDowncast<const DivisionQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonDivisionWorkload>(*divisionQueueDescriptor, info);
        }
        case LayerType::ElementwiseUnary :
        {
            auto elementwiseUnaryQueueDescriptor
                    = PolymorphicDowncast<const ElementwiseUnaryQueueDescriptor*>(&descriptor);

            switch(elementwiseUnaryQueueDescriptor->m_Parameters.m_Operation)
            {
                case UnaryOperation::Abs:
                {
                    AbsQueueDescriptor absQueueDescriptor;
                    absQueueDescriptor.m_Inputs  = elementwiseUnaryQueueDescriptor->m_Inputs;
                    absQueueDescriptor.m_Outputs = elementwiseUnaryQueueDescriptor->m_Outputs;

                    return std::make_unique<NeonAbsWorkload>(absQueueDescriptor, info);
                }
                case UnaryOperation::Exp:
                    return std::make_unique<NeonExpWorkload>(*elementwiseUnaryQueueDescriptor, info);
                case UnaryOperation::LogicalNot:
                    return std::make_unique<NeonLogicalNotWorkload>(*elementwiseUnaryQueueDescriptor, info);
                case UnaryOperation::Log:
                    return std::make_unique<NeonLogWorkload>(*elementwiseUnaryQueueDescriptor, info);
                case UnaryOperation::Neg:
                    return std::make_unique<NeonNegWorkload>(*elementwiseUnaryQueueDescriptor, info);
                case UnaryOperation::Rsqrt:
                {
                    RsqrtQueueDescriptor rsqrtQueueDescriptor;
                    rsqrtQueueDescriptor.m_Inputs  = elementwiseUnaryQueueDescriptor->m_Inputs;
                    rsqrtQueueDescriptor.m_Outputs = elementwiseUnaryQueueDescriptor->m_Outputs;

                    return std::make_unique<NeonRsqrtWorkload>(rsqrtQueueDescriptor, info);
                }
                case UnaryOperation::Sin:
                    return std::make_unique<NeonSinWorkload>(*elementwiseUnaryQueueDescriptor, info);
                default:
                    return nullptr;
            }
        }
        case LayerType::Fill :
        {
            auto fillQueueDescriptor = PolymorphicDowncast<const FillQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonFillWorkload>(*fillQueueDescriptor, info);
        }
        case LayerType::Floor :
        {
            auto floorQueueDescriptor = PolymorphicDowncast<const FloorQueueDescriptor*>(&descriptor);
            return MakeWorkloadHelper<NeonFloorFloatWorkload, NullWorkload>(*floorQueueDescriptor, info);
        }
        case LayerType::FullyConnected :
        {
            auto fullyConnectedQueueDescriptor = PolymorphicDowncast<const FullyConnectedQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonFullyConnectedWorkload>(*fullyConnectedQueueDescriptor,
                                                                info,
                                                                m_MemoryManager->GetIntraLayerManager());
        }
        case LayerType::Gather :
        {
            auto gatherQueueDescriptor = PolymorphicDowncast<const GatherQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonGatherWorkload>(*gatherQueueDescriptor, info);
        }
        case LayerType::Input :
        {
            auto inputQueueDescriptor = PolymorphicDowncast<const InputQueueDescriptor*>(&descriptor);
            return std::make_unique<CopyMemGenericWorkload>(*inputQueueDescriptor, info);
        }
        case LayerType::InstanceNormalization :
        {
            auto instanceNormalizationQueueDescriptor
                    = PolymorphicDowncast<const InstanceNormalizationQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonInstanceNormalizationWorkload>(*instanceNormalizationQueueDescriptor, info);
        }
        case LayerType::L2Normalization :
        {
            auto l2NormalizationQueueDescriptor
                    = PolymorphicDowncast<const L2NormalizationQueueDescriptor*>(&descriptor);
            return MakeWorkloadHelper<NeonL2NormalizationFloatWorkload, NullWorkload>
                    (*l2NormalizationQueueDescriptor, info, m_MemoryManager->GetIntraLayerManager());
        }
        case LayerType::LogSoftmax :
        {
            auto logSoftmaxQueueDescriptor = PolymorphicDowncast<const LogSoftmaxQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonLogSoftmaxWorkload>(*logSoftmaxQueueDescriptor,
                                                            info,
                                                            m_MemoryManager->GetIntraLayerManager());
        }
        case LayerType::LogicalBinary :
        {
            auto logicalBinaryQueueDescriptor = PolymorphicDowncast<const LogicalBinaryQueueDescriptor*>(&descriptor);

            switch(logicalBinaryQueueDescriptor->m_Parameters.m_Operation)
            {
                case LogicalBinaryOperation::LogicalAnd:
                    return std::make_unique<NeonLogicalAndWorkload>(*logicalBinaryQueueDescriptor, info);
                case LogicalBinaryOperation::LogicalOr:
                    return std::make_unique<NeonLogicalOrWorkload>(*logicalBinaryQueueDescriptor, info);
                default:
                    return nullptr;
            }
        }
        case LayerType::Lstm :
        {
            auto lstmQueueDescriptor = PolymorphicDowncast<const LstmQueueDescriptor*>(&descriptor);
            return MakeWorkloadHelper<NeonLstmFloatWorkload, NullWorkload>(*lstmQueueDescriptor, info);
        }
        case LayerType::Maximum :
        {
            auto maximumQueueDescriptor = PolymorphicDowncast<const MaximumQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonMaximumWorkload>(*maximumQueueDescriptor, info);
        }
        case LayerType::Mean :
        {
            auto meanQueueDescriptor = PolymorphicDowncast<const MeanQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonMeanWorkload>(*meanQueueDescriptor, info);
        }
        case LayerType::MemCopy :
        {
            auto memCopyQueueDescriptor = PolymorphicDowncast<const MemCopyQueueDescriptor*>(&descriptor);
            if (memCopyQueueDescriptor->m_Inputs.empty() || !memCopyQueueDescriptor->m_Inputs[0])
            {
                throw InvalidArgumentException("NeonWorkloadFactory: Invalid null input for MemCopy workload");
            }
            return MakeWorkloadHelper<CopyMemGenericWorkload, CopyMemGenericWorkload>(*memCopyQueueDescriptor, info);
        }
        case LayerType::MemImport :
        {
            auto memImportQueueDescriptor = PolymorphicDowncast<const MemImportQueueDescriptor*>(&descriptor);
            if (memImportQueueDescriptor->m_Inputs.empty() || !memImportQueueDescriptor->m_Inputs[0])
            {
                throw InvalidArgumentException("NeonWorkloadFactory: Invalid null input for MemImport workload");
            }
            return std::make_unique<ImportMemGenericWorkload>(*memImportQueueDescriptor, info);
        }
        case LayerType::Minimum :
        {
            auto minimumQueueDescriptor = PolymorphicDowncast<const MinimumQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonMinimumWorkload>(*minimumQueueDescriptor, info);
        }
        case LayerType::Multiplication :
        {
            auto multiplicationQueueDescriptor = PolymorphicDowncast<const MultiplicationQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonMultiplicationWorkload>(*multiplicationQueueDescriptor, info);
        }
        case LayerType::Normalization :
        {
            auto normalizationQueueDescriptor = PolymorphicDowncast<const NormalizationQueueDescriptor*>(&descriptor);
            return MakeWorkloadHelper<NeonNormalizationFloatWorkload, NullWorkload>
                    (*normalizationQueueDescriptor, info, m_MemoryManager->GetIntraLayerManager());
        }
        case LayerType::Output :
        {
            auto outputQueueDescriptor = PolymorphicDowncast<const OutputQueueDescriptor*>(&descriptor);
            return std::make_unique<CopyMemGenericWorkload>(*outputQueueDescriptor, info);
        }
        case LayerType::Pad :
        {
            auto padQueueDescriptor = PolymorphicDowncast<const PadQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonPadWorkload>(*padQueueDescriptor, info);
        }
        case LayerType::Permute :
        {
            auto permuteQueueDescriptor = PolymorphicDowncast<const PermuteQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonPermuteWorkload>(*permuteQueueDescriptor, info);
        }
        case LayerType::Pooling2d :
        {
            auto pooling2dQueueDescriptor = PolymorphicDowncast<const Pooling2dQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonPooling2dWorkload>(*pooling2dQueueDescriptor, info);
        }
        case LayerType::PreCompiled :
        {
            auto preCompiledQueueDescriptor = PolymorphicDowncast<const PreCompiledQueueDescriptor*>(&descriptor);
            return MakeWorkloadHelper<NullWorkload, NullWorkload>(*preCompiledQueueDescriptor, info);
        }
        case LayerType::Prelu :
        {
            auto preluQueueDescriptor = PolymorphicDowncast<const PreluQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonPreluWorkload>(*preluQueueDescriptor, info);
        }
        case LayerType::QLstm :
        {
            auto qLstmQueueDescriptor = PolymorphicDowncast<const QLstmQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonQLstmWorkload>(*qLstmQueueDescriptor, info);
        }
        case LayerType::Quantize :
        {
            auto quantizeQueueDescriptor = PolymorphicDowncast<const QuantizeQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonQuantizeWorkload>(*quantizeQueueDescriptor, info);
        }
        case LayerType::QuantizedLstm :
        {
            auto quantizedLstmQueueDescriptor = PolymorphicDowncast<const QuantizedLstmQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonQuantizedLstmWorkload>(*quantizedLstmQueueDescriptor, info);
        }
        case LayerType::Rank :
        {
            auto rankQueueDescriptor = PolymorphicDowncast<const RankQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonRankWorkload>(*rankQueueDescriptor, info);
        }
        case LayerType::Reduce :
        {
            auto reduceQueueDescriptor = PolymorphicDowncast<const ReduceQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonReduceWorkload>(*reduceQueueDescriptor, info);
        }
        case LayerType::Reshape :
        {
            auto reshapeQueueDescriptor = PolymorphicDowncast<const ReshapeQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonReshapeWorkload>(*reshapeQueueDescriptor, info);
        }
        case LayerType::Resize :
        {
            auto resizeQueueDescriptor = PolymorphicDowncast<const ResizeQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonResizeWorkload>(*resizeQueueDescriptor, info);
        }
        case LayerType::Slice :
        {
            auto sliceQueueDescriptor = PolymorphicDowncast<const SliceQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonSliceWorkload>(*sliceQueueDescriptor, info);
        }
        case LayerType::Softmax :
        {
            auto softmaxQueueDescriptor = PolymorphicDowncast<const SoftmaxQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonSoftmaxWorkload>(*softmaxQueueDescriptor,
                                                         info,
                                                         m_MemoryManager->GetIntraLayerManager());
        }
        case LayerType::SpaceToBatchNd :
        {
            auto spaceToBatchNdQueueDescriptor
                    = PolymorphicDowncast<const SpaceToBatchNdQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonSpaceToBatchNdWorkload>(*spaceToBatchNdQueueDescriptor, info);
        }
        case LayerType::SpaceToDepth :
        {
            auto spaceToDepthQueueDescriptor = PolymorphicDowncast<const SpaceToDepthQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonSpaceToDepthWorkload>(*spaceToDepthQueueDescriptor, info);
        }
        case LayerType::Splitter :
        {
            auto splitterQueueDescriptor = PolymorphicDowncast<const SplitterQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonSplitterWorkload>(*splitterQueueDescriptor, info);
        }
        case LayerType::Stack :
        {
            auto stackQueueDescriptor = PolymorphicDowncast<const StackQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonStackWorkload>(*stackQueueDescriptor, info);
        }
        case LayerType::StridedSlice :
        {
            auto stridedSliceQueueDescriptor = PolymorphicDowncast<const StridedSliceQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonStridedSliceWorkload>(*stridedSliceQueueDescriptor, info);
        }
        case LayerType::Subtraction :
        {
            auto subtractionQueueDescriptor = PolymorphicDowncast<const SubtractionQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonSubtractionWorkload>(*subtractionQueueDescriptor, info);
        }
        case LayerType::Transpose :
        {
            auto transposeQueueDescriptor = PolymorphicDowncast<const TransposeQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonTransposeWorkload>(*transposeQueueDescriptor, info);
        }
        case LayerType::TransposeConvolution2d :
        {
            auto transposeConvolution2dQueueDescriptor
                    = PolymorphicDowncast<const TransposeConvolution2dQueueDescriptor*>(&descriptor);
            return std::make_unique<NeonTransposeConvolution2dWorkload>(*transposeConvolution2dQueueDescriptor,
                                                                        info,
                                                                        m_MemoryManager->GetIntraLayerManager());
        }
        case LayerType::UnidirectionalSequenceLstm :
        {
            auto desc = PolymorphicDowncast<const UnidirectionalSequenceLstmQueueDescriptor*>(&descriptor);
            return MakeWorkloadHelper<NeonUnidirectionalSequenceLstmFloatWorkload, NullWorkload>(*desc, info);
        }
        default:
            return nullptr;
    }
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateActivation(const ActivationQueueDescriptor& descriptor,
                                                                 const WorkloadInfo&              info) const
{
    return std::make_unique<NeonActivationWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateAddition(const AdditionQueueDescriptor& descriptor,
                                                                      const WorkloadInfo&            info) const
{
    return std::make_unique<NeonAdditionWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateArgMinMax(const ArgMinMaxQueueDescriptor& descriptor,
                                                                const WorkloadInfo& info) const
{
    return std::make_unique<NeonArgMinMaxWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateBatchNormalization(
    const BatchNormalizationQueueDescriptor& descriptor, const WorkloadInfo& info) const
{
    return std::make_unique<NeonBatchNormalizationWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateBatchToSpaceNd(const BatchToSpaceNdQueueDescriptor& descriptor,
                                                                     const WorkloadInfo& info) const
{
    return std::make_unique<NeonBatchToSpaceNdWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateCast(const CastQueueDescriptor& descriptor,
                                                           const WorkloadInfo& info) const
{
    return std::make_unique<NeonCastWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateChannelShuffle(const ChannelShuffleQueueDescriptor& descriptor,
                                                                     const WorkloadInfo& info) const
{
    return std::make_unique<NeonChannelShuffleWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateComparison(const ComparisonQueueDescriptor& descriptor,
                                                                 const WorkloadInfo& info) const
{
    return std::make_unique<NeonComparisonWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateConcat(const ConcatQueueDescriptor& descriptor,
                                                                    const WorkloadInfo&          info) const
{
    return std::make_unique<NeonConcatWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConstant(const ConstantQueueDescriptor& descriptor,
                                                               const WorkloadInfo& info) const
{
    return std::make_unique<NeonConstantWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConvertBf16ToFp32(
    const ConvertBf16ToFp32QueueDescriptor& descriptor,
    const WorkloadInfo& info) const
{
    return std::make_unique<NeonConvertBf16ToFp32Workload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConvertFp16ToFp32(
    const ConvertFp16ToFp32QueueDescriptor& descriptor,
    const WorkloadInfo& info) const
{
    return std::make_unique<NeonConvertFp16ToFp32Workload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConvertFp32ToBf16(
    const ConvertFp32ToBf16QueueDescriptor& descriptor,
    const WorkloadInfo& info) const
{
    return std::make_unique<NeonConvertFp32ToBf16Workload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateConvertFp32ToFp16(
    const ConvertFp32ToFp16QueueDescriptor& descriptor,
    const WorkloadInfo& info) const
{
    return std::make_unique<NeonConvertFp32ToFp16Workload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateConvolution2d(
    const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info) const
{
    bool isFastMathEnabled = false;
    if (m_ModelContextPtr)
    {
        if (m_ModelContextPtr.get() != nullptr)
        {
            auto modelOptions = dynamic_cast<NeonBackendModelContext*>(m_ModelContextPtr.get());
            if (modelOptions)
            {
                isFastMathEnabled = modelOptions->IsFastMathEnabled();
            }
        }
    }
    return std::make_unique<NeonConvolution2dWorkload>(descriptor,
                                                       info,
                                                       m_MemoryManager->GetIntraLayerManager(),
                                                       isFastMathEnabled);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateConvolution3d(
        const Convolution3dQueueDescriptor& descriptor, const WorkloadInfo& info) const
{
    bool isFastMathEnabled = false;
    if (m_ModelContextPtr)
    {
        if (m_ModelContextPtr.get() != nullptr)
        {
            auto modelOptions = dynamic_cast<NeonBackendModelContext*>(m_ModelContextPtr.get());
            if (modelOptions)
            {
                isFastMathEnabled = modelOptions->IsFastMathEnabled();
            }
        }
    }
    return std::make_unique<NeonConvolution3dWorkload>(descriptor,
                                                       info,
                                                       m_MemoryManager->GetIntraLayerManager(),
                                                       isFastMathEnabled);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDebug(const DebugQueueDescriptor& descriptor,
                                                            const WorkloadInfo& info) const
{
    return MakeWorkloadHelper<NullWorkload, NullWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDepthToSpace(const DepthToSpaceQueueDescriptor& descriptor,
                                                                   const WorkloadInfo& info) const
{
    return std::make_unique<NeonDepthToSpaceWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDepthwiseConvolution2d(
    const DepthwiseConvolution2dQueueDescriptor& descriptor, const WorkloadInfo& info) const
{
    return std::make_unique<NeonDepthwiseConvolutionWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDequantize(const DequantizeQueueDescriptor& descriptor,
                                                                 const WorkloadInfo& info) const
{
    return std::make_unique<NeonDequantizeWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateDetectionPostProcess(
    const armnn::DetectionPostProcessQueueDescriptor& descriptor, const armnn::WorkloadInfo& info) const
{
    return MakeWorkloadHelper<NullWorkload, NullWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateDivision(
    const DivisionQueueDescriptor& descriptor, const WorkloadInfo& info) const
{
    return std::make_unique<NeonDivisionWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateElementwiseUnary(
    const ElementwiseUnaryQueueDescriptor& descriptor, const WorkloadInfo& info) const
{
    switch(descriptor.m_Parameters.m_Operation)
    {
        case UnaryOperation::Abs:
        {
            AbsQueueDescriptor absQueueDescriptor;
            absQueueDescriptor.m_Inputs  = descriptor.m_Inputs;
            absQueueDescriptor.m_Outputs = descriptor.m_Outputs;

            return std::make_unique<NeonAbsWorkload>(absQueueDescriptor, info);
        }
        case UnaryOperation::Exp:
            return std::make_unique<NeonExpWorkload>(descriptor, info);
        case UnaryOperation::LogicalNot:
            return std::make_unique<NeonLogicalNotWorkload>(descriptor, info);
        case UnaryOperation::Log:
            return std::make_unique<NeonLogWorkload>(descriptor, info);
        case UnaryOperation::Neg:
            return std::make_unique<NeonNegWorkload>(descriptor, info);
        case UnaryOperation::Rsqrt:
        {
            RsqrtQueueDescriptor rsqrtQueueDescriptor;
            rsqrtQueueDescriptor.m_Inputs  = descriptor.m_Inputs;
            rsqrtQueueDescriptor.m_Outputs = descriptor.m_Outputs;

            return std::make_unique<NeonRsqrtWorkload>(rsqrtQueueDescriptor, info);
        }
        case UnaryOperation::Sin:
            return std::make_unique<NeonSinWorkload>(descriptor, info);
        default:
            return nullptr;
    }
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateFill(const FillQueueDescriptor& descriptor,
                                                           const WorkloadInfo& info) const
{
    return std::make_unique<NeonFillWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateFloor(const FloorQueueDescriptor& descriptor,
                                                            const WorkloadInfo& info) const
{
    return MakeWorkloadHelper<NeonFloorFloatWorkload, NullWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateFullyConnected(
    const FullyConnectedQueueDescriptor& descriptor, const WorkloadInfo& info) const
{
    return std::make_unique<NeonFullyConnectedWorkload>(descriptor, info, m_MemoryManager->GetIntraLayerManager());
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateGather(const armnn::GatherQueueDescriptor& descriptor,
                                                             const armnn::WorkloadInfo& info) const
{
    return std::make_unique<NeonGatherWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateInput(const InputQueueDescriptor& descriptor,
                                                            const WorkloadInfo&        info) const
{
    return std::make_unique<CopyMemGenericWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateInstanceNormalization(
    const InstanceNormalizationQueueDescriptor& descriptor,
    const WorkloadInfo& info) const
{
    return std::make_unique<NeonInstanceNormalizationWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateL2Normalization(const L2NormalizationQueueDescriptor& descriptor,
                                                                      const WorkloadInfo& info) const
{
    return MakeWorkloadHelper<NeonL2NormalizationFloatWorkload, NullWorkload>(descriptor, info,
                                                                              m_MemoryManager->GetIntraLayerManager());
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateLogSoftmax(const LogSoftmaxQueueDescriptor& descriptor,
                                                                 const WorkloadInfo& info) const
{
    return std::make_unique<NeonLogSoftmaxWorkload>(descriptor, info, m_MemoryManager->GetIntraLayerManager());
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateLogicalBinary(const LogicalBinaryQueueDescriptor& descriptor,
                                                                    const WorkloadInfo& info) const
{
    switch(descriptor.m_Parameters.m_Operation)
    {
        case LogicalBinaryOperation::LogicalAnd:
            return std::make_unique<NeonLogicalAndWorkload>(descriptor, info);
        case LogicalBinaryOperation::LogicalOr:
            return std::make_unique<NeonLogicalOrWorkload>(descriptor, info);
        default:
            return nullptr;
    }
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateLstm(const LstmQueueDescriptor& descriptor,
                                                           const WorkloadInfo& info) const
{
    return MakeWorkloadHelper<NeonLstmFloatWorkload, NullWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMaximum(const MaximumQueueDescriptor& descriptor,
                                                              const WorkloadInfo& info) const
{
    return std::make_unique<NeonMaximumWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMean(const MeanQueueDescriptor& descriptor,
                                                           const WorkloadInfo& info) const
{
    return std::make_unique<NeonMeanWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateMemCopy(const MemCopyQueueDescriptor& descriptor,
                                                                     const WorkloadInfo&        info) const
{
    if (descriptor.m_Inputs.empty() || !descriptor.m_Inputs[0])
    {
        throw InvalidArgumentException("NeonWorkloadFactory: Invalid null input for MemCopy workload");
    }

    return MakeWorkloadHelper<CopyMemGenericWorkload, CopyMemGenericWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateMemImport(const MemImportQueueDescriptor& descriptor,
                                                                       const WorkloadInfo&        info) const
{
    if (descriptor.m_Inputs.empty() || !descriptor.m_Inputs[0])
    {
        throw InvalidArgumentException("NeonWorkloadFactory: Invalid null input for MemImport workload");
    }

    return std::make_unique<ImportMemGenericWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateMinimum(const MinimumQueueDescriptor& descriptor,
                                                              const WorkloadInfo& info) const
{
    return std::make_unique<NeonMinimumWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateMultiplication(
    const MultiplicationQueueDescriptor& descriptor, const WorkloadInfo& info) const
{
    return std::make_unique<NeonMultiplicationWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateNormalization(
    const NormalizationQueueDescriptor& descriptor, const WorkloadInfo& info) const
{
    return MakeWorkloadHelper<NeonNormalizationFloatWorkload, NullWorkload>(descriptor, info,
                                                                            m_MemoryManager->GetIntraLayerManager());
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateOutput(const OutputQueueDescriptor& descriptor,
                                                             const WorkloadInfo& info) const
{
    return std::make_unique<CopyMemGenericWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreatePad(const PadQueueDescriptor& descriptor,
                                                          const WorkloadInfo& info) const
{
    return std::make_unique<NeonPadWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreatePermute(const PermuteQueueDescriptor& descriptor,
                                                                     const WorkloadInfo& info) const
{
    return std::make_unique<NeonPermuteWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreatePooling2d(const Pooling2dQueueDescriptor& descriptor,
                                                                       const WorkloadInfo& info) const
{
    return std::make_unique<NeonPooling2dWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreatePreCompiled(const PreCompiledQueueDescriptor& descriptor,
                                                                  const WorkloadInfo& info) const
{
    return MakeWorkloadHelper<NullWorkload, NullWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreatePrelu(const armnn::PreluQueueDescriptor &descriptor,
                                                                   const armnn::WorkloadInfo &info) const
{
    return std::make_unique<NeonPreluWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateQLstm(const QLstmQueueDescriptor& descriptor,
                                                            const WorkloadInfo& info) const
{
    return std::make_unique<NeonQLstmWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateQuantize(const QuantizeQueueDescriptor& descriptor,
                                                                      const WorkloadInfo& info) const
{
    return std::make_unique<NeonQuantizeWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateQuantizedLstm(const QuantizedLstmQueueDescriptor& descriptor,
                                                                    const WorkloadInfo& info) const
{
    return std::make_unique<NeonQuantizedLstmWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateRank(const RankQueueDescriptor& descriptor,
                                                           const WorkloadInfo& info) const
{
    return std::make_unique<NeonRankWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateReduce(const ReduceQueueDescriptor& descriptor,
                                                             const WorkloadInfo& info) const
{
    return std::make_unique<NeonReduceWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateReshape(const ReshapeQueueDescriptor& descriptor,
                                                              const WorkloadInfo& info) const
{
    return std::make_unique<NeonReshapeWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateResize(const ResizeQueueDescriptor& descriptor,
                                                             const WorkloadInfo& info) const
{
    return std::make_unique<NeonResizeWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSlice(const SliceQueueDescriptor& descriptor,
                                                            const WorkloadInfo& info) const
{
    return std::make_unique<NeonSliceWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSoftmax(const SoftmaxQueueDescriptor& descriptor,
                                                              const WorkloadInfo& info) const
{
    return std::make_unique<NeonSoftmaxWorkload>(descriptor, info, m_MemoryManager->GetIntraLayerManager());
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSpaceToBatchNd(const SpaceToBatchNdQueueDescriptor& descriptor,
                                                                     const WorkloadInfo& info) const
{
    return std::make_unique<NeonSpaceToBatchNdWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSpaceToDepth(const SpaceToDepthQueueDescriptor& descriptor,
                                                                   const WorkloadInfo& info) const
{
    return std::make_unique<NeonSpaceToDepthWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateSplitter(const SplitterQueueDescriptor& descriptor,
                                                               const WorkloadInfo&            info) const
{
    return std::make_unique<NeonSplitterWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateStack(const StackQueueDescriptor& descriptor,
                                                            const WorkloadInfo& info) const
{
    return std::make_unique<NeonStackWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateStridedSlice(const StridedSliceQueueDescriptor& descriptor,
                                                                   const WorkloadInfo& info) const
{
    return std::make_unique<NeonStridedSliceWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateSubtraction(
    const SubtractionQueueDescriptor& descriptor, const WorkloadInfo& info) const
{
    return std::make_unique<NeonSubtractionWorkload>(descriptor, info);
}

std::unique_ptr<armnn::IWorkload> NeonWorkloadFactory::CreateTranspose(const TransposeQueueDescriptor& descriptor,
                                                                       const WorkloadInfo& info) const
{
    return std::make_unique<NeonTransposeWorkload>(descriptor, info);
}

std::unique_ptr<IWorkload> NeonWorkloadFactory::CreateTransposeConvolution2d(
    const TransposeConvolution2dQueueDescriptor &descriptor,
    const WorkloadInfo &info) const
{
    return std::make_unique<NeonTransposeConvolution2dWorkload>(descriptor, info,
                                                                m_MemoryManager->GetIntraLayerManager());
}

} // namespace armnn