ClWorkloadFactory Class Reference

#include <ClWorkloadFactory.hpp>

Inheritance diagram for ClWorkloadFactory:

Collaboration diagram for ClWorkloadFactory:

Public Member Functions
	ClWorkloadFactory (const std::shared_ptr< ClMemoryManager > &memoryManager)
	ClWorkloadFactory (const std::shared_ptr< ClMemoryManager > &memoryManager, const IBackendInternal::IBackendSpecificModelContextPtr &modelContextPtr)
void	AfterWorkloadsCreated () override
const BackendId &	GetBackendId () const override
bool	SupportsSubTensors () const override
std::unique_ptr< ITensorHandle >	CreateSubTensorHandle (ITensorHandle &parent, TensorShape const &subTensorShape, unsigned int const *subTensorOrigin) const override
std::unique_ptr< ITensorHandle >	CreateTensorHandle (const TensorInfo &tensorInfo, const bool IsMemoryManaged=true) const override
std::unique_ptr< ITensorHandle >	CreateTensorHandle (const TensorInfo &tensorInfo, DataLayout dataLayout, const bool IsMemoryManaged=true) const override
std::unique_ptr< IWorkload >	CreateWorkload (LayerType type, const QueueDescriptor &descriptor, const WorkloadInfo &info) const override
	Backends should implement their own CreateWorkload function with a switch statement.
Public Member Functions inherited from WorkloadFactoryBase
bool	SupportsSubTensors () const override
std::unique_ptr< ITensorHandle >	CreateSubTensorHandle (ITensorHandle &, TensorShape const &, unsigned int const *) const override
std::unique_ptr< ITensorHandle >	CreateTensorHandle (const TensorInfo &, const bool) const override
std::unique_ptr< ITensorHandle >	CreateTensorHandle (const TensorInfo &, DataLayout, const bool) const override
std::unique_ptr< IWorkload >	CreateWorkload (LayerType, const QueueDescriptor &, const WorkloadInfo &) const override
	Backends should implement their own CreateWorkload function with a switch statement.
Public Member Functions inherited from IWorkloadFactory
virtual	~IWorkloadFactory ()

Static Public Member Functions
static bool	IsLayerSupported (const Layer &layer, Optional< DataType > dataType, std::string &outReasonIfUnsupported)
static bool	IsLayerSupported (const IConnectableLayer &layer, Optional< DataType > dataType, std::string &outReasonIfUnsupported, const ModelOptions &modelOptions)
Static Public Member Functions inherited from IWorkloadFactory
static bool	IsLayerSupported (const BackendId &backendId, const IConnectableLayer &layer, Optional< DataType > dataType, std::string &outReasonIfUnsupported)
static bool	IsLayerSupported (const IConnectableLayer &layer, Optional< DataType > dataType, std::string &outReasonIfUnsupported)
static bool	IsLayerSupported (const IConnectableLayer &layer, Optional< DataType > dataType, std::string &outReasonIfUnsupported, const ModelOptions &modelOptions)
static bool	IsLayerSupported (const BackendId &backendId, const IConnectableLayer &layer, Optional< DataType > dataType, std::string &outReasonIfUnsupported, const ModelOptions &modelOptions)

Detailed Description

Definition at line 21 of file ClWorkloadFactory.hpp.

Constructor & Destructor Documentation

ClWorkloadFactory() [1/2]

ClWorkloadFactory

(

const std::shared_ptr< ClMemoryManager > &

memoryManager

)

Definition at line 188 of file ClWorkloadFactory.cpp.

    : m_MemoryManager(memoryManager), m_ModelContextPtr(IBackendInternal::IBackendSpecificModelContextPtr{})
{
    InitializeCLCompileContext();
}

ClWorkloadFactory() [2/2]

ClWorkloadFactory	(	const std::shared_ptr< ClMemoryManager > &	memoryManager,
		const IBackendInternal::IBackendSpecificModelContextPtr &	modelContextPtr )

Definition at line 194 of file ClWorkloadFactory.cpp.

    : m_MemoryManager(memoryManager), m_ModelContextPtr(modelContextPtr)
{
    InitializeCLCompileContext();
}

Member Function Documentation

AfterWorkloadsCreated()

void AfterWorkloadsCreated ( )

overridevirtual

Reimplemented from IWorkloadFactory.

Definition at line 66 of file ClWorkloadFactory.cpp.

{
    if(m_ModelContextPtr)
    {
        auto modelOptions = dynamic_cast<ClBackendModelContext*>(m_ModelContextPtr.get());
        if (modelOptions->SaveCachedNetwork())
        {
            ClContextSerializer serializer;
            serializer.Serialize(m_CLCompileContext);
            auto cachedFd = modelOptions->GetCachedFileDescriptor();
            if (cachedFd != -1)
            {
                std::vector<uint8_t> compiledContextData;
                std::stringstream stream;
                bool serialized = serializer.SaveSerializedToStream(stream);
                if (serialized)
                {
                    std::string const serializedString{stream.str()};
                    std::copy(serializedString.begin(),
                              serializedString.end(),
                              std::back_inserter(compiledContextData));
                    auto success = write(cachedFd, compiledContextData.data(), compiledContextData.size());
                    if (success == -1)
                    {
                        ARMNN_LOG(info) << "ClWorkloadFactory:: Could not cache the compiled context!";
                    }
                }
            }
 
            // Save map to a filepath provided in ModelOptions
            auto filePath = modelOptions->GetCachedNetworkFilePath();
            if (filePath != "" && fs::exists(filePath) && fs::is_regular_file(filePath))
            {
                // Serialize ClContext to the file specified
                std::ofstream file(filePath, std::ios::out | std::ios::binary);
                serializer.SaveSerializedToStream(file);
            }
        }
    }
}

References ARMNN_LOG, and armnn::info.

CreateSubTensorHandle()

std::unique_ptr< ITensorHandle > CreateSubTensorHandle ( ITensorHandle & parent, TensorShape const & subTensorShape, unsigned int const * subTensorOrigin ) const

overridevirtual

Implements IWorkloadFactory.

Definition at line 222 of file ClWorkloadFactory.cpp.

{
    arm_compute::Coordinates coords;
    arm_compute::TensorShape shape = armcomputetensorutils::BuildArmComputeTensorShape(subTensorShape);
 
    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<ClSubTensorHandle>(
        PolymorphicDowncast<IClTensorHandle*>(&parent), shape, coords);
}

References TensorShape::GetNumDimensions(), and ITensorHandle::GetShape().

CreateTensorHandle() [1/2]

std::unique_ptr< ITensorHandle > CreateTensorHandle ( const TensorInfo & tensorInfo, const bool IsMemoryManaged = true ) const

overridevirtual

Implements IWorkloadFactory.

Definition at line 201 of file ClWorkloadFactory.cpp.

{
    IgnoreUnused(IsMemoryManaged);
    std::unique_ptr<ClTensorHandle> tensorHandle = std::make_unique<ClTensorHandle>(tensorInfo);
    tensorHandle->SetMemoryGroup(m_MemoryManager->GetInterLayerMemoryGroup());
 
    return tensorHandle;
}

References armnn::IgnoreUnused().

CreateTensorHandle() [2/2]

std::unique_ptr< ITensorHandle > CreateTensorHandle ( const TensorInfo & tensorInfo, DataLayout dataLayout, const bool IsMemoryManaged = true ) const

overridevirtual

Implements IWorkloadFactory.

Definition at line 211 of file ClWorkloadFactory.cpp.

{
    IgnoreUnused(IsMemoryManaged);
    std::unique_ptr<ClTensorHandle> tensorHandle = std::make_unique<ClTensorHandle>(tensorInfo, dataLayout);
    tensorHandle->SetMemoryGroup(m_MemoryManager->GetInterLayerMemoryGroup());
 
    return tensorHandle;
}

References armnn::IgnoreUnused().

CreateWorkload()

std::unique_ptr< IWorkload > CreateWorkload ( LayerType type, const QueueDescriptor & descriptor, const WorkloadInfo & info ) const

overridevirtual

Backends should implement their own CreateWorkload function with a switch statement.

The case for the switch should be the LayerType and based on that they will call their specific workload creation functionality.

Implements IWorkloadFactory.

Definition at line 247 of file ClWorkloadFactory.cpp.

{
    switch(type)
    {
        case LayerType::Activation :
        {
            auto activationQueueDescriptor = PolymorphicDowncast<const ActivationQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClActivationWorkload>(*activationQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Addition :
        {
            auto additionQueueDescriptor = PolymorphicDowncast<const AdditionQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClAdditionWorkload>(*additionQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::ArgMinMax :
        {
            auto argMinMaxQueueDescriptor = PolymorphicDowncast<const ArgMinMaxQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClArgMinMaxWorkload>(*argMinMaxQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::BatchMatMul :
        {
            auto batchMatMulQueueDescriptor = PolymorphicDowncast<const BatchMatMulQueueDescriptor*>(&descriptor);
            return std::make_unique<ClBatchMatMulWorkload>(*batchMatMulQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::BatchNormalization :
        {
            auto batchNormalizationQueueDescriptor
                     = PolymorphicDowncast<const BatchNormalizationQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClBatchNormalizationFloatWorkload, NullWorkload>
                (*batchNormalizationQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::BatchToSpaceNd :
        {
            auto batchToSpaceNdQueueDescriptor
                     = PolymorphicDowncast<const BatchToSpaceNdQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClBatchToSpaceNdWorkload>(*batchToSpaceNdQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Cast :
        {
            auto castQueueDescriptor = PolymorphicDowncast<const CastQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClCastWorkload>(*castQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::ChannelShuffle :
        {
            auto channelShuffleQueueDescriptor
                     = PolymorphicDowncast<const ChannelShuffleQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClChannelShuffleWorkload>(*channelShuffleQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Comparison :
        {
            auto comparisonQueueDescriptor = PolymorphicDowncast<const ComparisonQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClComparisonWorkload>(*comparisonQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Concat :
        {
            auto concatQueueDescriptor = PolymorphicDowncast<const ConcatQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClConcatWorkload>(*concatQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Constant :
        {
            auto constantQueueDescriptor = PolymorphicDowncast<const ConstantQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClConstantWorkload>(*constantQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::ConvertFp16ToFp32 :
        {
            auto convertFp16ToFp32QueueDescriptor
                     = PolymorphicDowncast<const ConvertFp16ToFp32QueueDescriptor*>(&descriptor);
            return MakeWorkload<ClConvertFp16ToFp32Workload>(*convertFp16ToFp32QueueDescriptor,
                                                             info,
                                                             m_CLCompileContext);
        }
        case LayerType::ConvertFp32ToFp16 :
        {
            auto convertFp32ToFp16QueueDescriptor
                     = PolymorphicDowncast<const ConvertFp32ToFp16QueueDescriptor*>(&descriptor);
            return MakeWorkload<ClConvertFp32ToFp16Workload>(*convertFp32ToFp16QueueDescriptor,
                                                             info,
                                                             m_CLCompileContext);
        }
        case LayerType::Convolution2d :
        {
            auto convolution2dQueueDescriptor = PolymorphicDowncast<const Convolution2dQueueDescriptor*>(&descriptor);
            bool isFastMathEnabled = false;
            if (m_ModelContextPtr)
            {
                if (m_ModelContextPtr.get() != nullptr)
                {
                    auto modelOptions = dynamic_cast<ClBackendModelContext*>(m_ModelContextPtr.get());
                    if (modelOptions)
                    {
                        isFastMathEnabled = modelOptions->IsFastMathEnabled();
                    }
                }
            }
            return MakeWorkload<ClConvolution2dWorkload>(*convolution2dQueueDescriptor,
                                                         info,
                                                         m_MemoryManager->GetIntraLayerManager(),
                                                         m_CLCompileContext,
                                                         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<ClBackendModelContext*>(m_ModelContextPtr.get());
                    if (modelOptions)
                    {
                        isFastMathEnabled = modelOptions->IsFastMathEnabled();
                    }
                }
            }
            return MakeWorkload<ClConvolution3dWorkload>(*convolution3dQueueDescriptor,
                                                         info,
                                                         m_MemoryManager->GetIntraLayerManager(),
                                                         m_CLCompileContext,
                                                         isFastMathEnabled);
        }
        case LayerType::Debug :
        {
            auto debugQueueDescriptor = PolymorphicDowncast<const DebugQueueDescriptor*>(&descriptor);
            return MakeWorkload<NullWorkload, NullWorkload>(*debugQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::DepthToSpace :
        {
            auto depthToSpaceQueueDescriptor = PolymorphicDowncast<const DepthToSpaceQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClDepthToSpaceWorkload>(*depthToSpaceQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::DepthwiseConvolution2d :
        {
            auto depthwiseConvolution2dQueueDescriptor
                     = PolymorphicDowncast<const DepthwiseConvolution2dQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClDepthwiseConvolutionWorkload>(*depthwiseConvolution2dQueueDescriptor,
                                                                info,
                                                                m_CLCompileContext);
        }
        case LayerType::Dequantize :
        {
            auto dequantizeQueueDescriptor = PolymorphicDowncast<const DequantizeQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClDequantizeWorkload>(*dequantizeQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::DetectionPostProcess :
        {
            auto detectionPostProcessQueueDescriptor
                     = PolymorphicDowncast<const DetectionPostProcessQueueDescriptor*>(&descriptor);
            return MakeWorkload<NullWorkload, NullWorkload>(*detectionPostProcessQueueDescriptor,
                                                            info,
                                                            m_CLCompileContext);
        }
        case LayerType::Division :
        {
            auto divisionQueueDescriptor = PolymorphicDowncast<const DivisionQueueDescriptor*>(&descriptor);
            return std::make_unique<ClDivisionWorkload>(*divisionQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::ElementwiseBinary :
        {
            auto elementwiseBinaryQueueDescriptor
                     = PolymorphicDowncast<const ElementwiseBinaryQueueDescriptor*>(&descriptor);
            switch (elementwiseBinaryQueueDescriptor->m_Parameters.m_Operation)
            {
                case BinaryOperation::Add:
                {
                    AdditionQueueDescriptor additionQueueDescriptor;
                    additionQueueDescriptor.m_Inputs  = descriptor.m_Inputs;
                    additionQueueDescriptor.m_Outputs = descriptor.m_Outputs;
                    additionQueueDescriptor.m_AdditionalInfoObject =
                        elementwiseBinaryQueueDescriptor->m_AdditionalInfoObject;
                    return std::make_unique<ClAdditionWorkload>(additionQueueDescriptor, info, m_CLCompileContext);
                }
                case BinaryOperation::Div:
                {
                    DivisionQueueDescriptor divisionQueueDescriptor;
                    divisionQueueDescriptor.m_Inputs  = descriptor.m_Inputs;
                    divisionQueueDescriptor.m_Outputs = descriptor.m_Outputs;
                    divisionQueueDescriptor.m_AdditionalInfoObject =
                        elementwiseBinaryQueueDescriptor->m_AdditionalInfoObject;
                    return std::make_unique<ClDivisionWorkload>(divisionQueueDescriptor, info, m_CLCompileContext);
                }
                case BinaryOperation::FloorDiv:
                {
                    DivisionQueueDescriptor divisionQueueDescriptor;
                    divisionQueueDescriptor.m_Inputs  = descriptor.m_Inputs;
                    divisionQueueDescriptor.m_Outputs = descriptor.m_Outputs;
                    divisionQueueDescriptor.m_AdditionalInfoObject =
                        elementwiseBinaryQueueDescriptor->m_AdditionalInfoObject;
                    return std::make_unique<ClFloorDivWorkload>(divisionQueueDescriptor, info, m_CLCompileContext);
                }
                case BinaryOperation::Maximum:
                {
                    MaximumQueueDescriptor maximumQueueDescriptor;
                    maximumQueueDescriptor.m_Inputs  = descriptor.m_Inputs;
                    maximumQueueDescriptor.m_Outputs = descriptor.m_Outputs;
                    maximumQueueDescriptor.m_AdditionalInfoObject =
                        elementwiseBinaryQueueDescriptor->m_AdditionalInfoObject;
                    return std::make_unique<ClMaximumWorkload>(maximumQueueDescriptor, info, m_CLCompileContext);
                }
                case BinaryOperation::Minimum:
                {
                    MinimumQueueDescriptor minimumQueueDescriptor;
                    minimumQueueDescriptor.m_Inputs  = descriptor.m_Inputs;
                    minimumQueueDescriptor.m_Outputs = descriptor.m_Outputs;
                    minimumQueueDescriptor.m_AdditionalInfoObject =
                        elementwiseBinaryQueueDescriptor->m_AdditionalInfoObject;
                    return std::make_unique<ClMinimumWorkload>(minimumQueueDescriptor, info, m_CLCompileContext);
                }
                case BinaryOperation::Mul:
                {
                    MultiplicationQueueDescriptor multiplicationQueueDescriptor;
                    multiplicationQueueDescriptor.m_Inputs  = descriptor.m_Inputs;
                    multiplicationQueueDescriptor.m_Outputs = descriptor.m_Outputs;
                    multiplicationQueueDescriptor.m_AdditionalInfoObject =
                        elementwiseBinaryQueueDescriptor->m_AdditionalInfoObject;
                    return std::make_unique<ClMultiplicationWorkload>(multiplicationQueueDescriptor,
                                                                      info,
                                                                      m_CLCompileContext);
                }
                case BinaryOperation::Power:
                case BinaryOperation::SqDiff:
                {
                    return std::make_unique<ClElementwiseBinaryWorkload>(*elementwiseBinaryQueueDescriptor,
                                                                         info,
                                                                         m_CLCompileContext);
                }
                case BinaryOperation::Sub:
                {
                    SubtractionQueueDescriptor subtractionQueueDescriptor;
                    subtractionQueueDescriptor.m_Inputs  = descriptor.m_Inputs;
                    subtractionQueueDescriptor.m_Outputs = descriptor.m_Outputs;
                    subtractionQueueDescriptor.m_AdditionalInfoObject =
                        elementwiseBinaryQueueDescriptor->m_AdditionalInfoObject;
                    return std::make_unique<ClSubtractionWorkload>(subtractionQueueDescriptor,
                                                                   info,
                                                                   m_CLCompileContext);
                }
                default:
                    return nullptr;
            }
        }
        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<ClAbsWorkload>(absQueueDescriptor, info, m_CLCompileContext);
                }
                case UnaryOperation::Exp:
                    return std::make_unique<ClExpWorkload>(*elementwiseUnaryQueueDescriptor, info, m_CLCompileContext);
                case UnaryOperation::Log:
                    return std::make_unique<ClLogWorkload>(*elementwiseUnaryQueueDescriptor, info, m_CLCompileContext);
                case UnaryOperation::LogicalNot:
                    return std::make_unique<ClLogicalNotWorkload>(*elementwiseUnaryQueueDescriptor,
                                                                  info,
                                                                  m_CLCompileContext);
                case UnaryOperation::Neg:
                    return std::make_unique<ClNegWorkload>(*elementwiseUnaryQueueDescriptor, info, m_CLCompileContext);
                case UnaryOperation::Rsqrt:
                {
                    RsqrtQueueDescriptor rsqrtQueueDescriptor;
                    rsqrtQueueDescriptor.m_Inputs  = elementwiseUnaryQueueDescriptor->m_Inputs;
                    rsqrtQueueDescriptor.m_Outputs = elementwiseUnaryQueueDescriptor->m_Outputs;
                    return std::make_unique<ClRsqrtWorkload>(rsqrtQueueDescriptor, info, m_CLCompileContext);
                }
                case UnaryOperation::Sin:
                    return std::make_unique<ClSinWorkload>(*elementwiseUnaryQueueDescriptor, info, m_CLCompileContext);
                case UnaryOperation::Sqrt:
                    return std::make_unique<ClSqrtWorkload>(*elementwiseUnaryQueueDescriptor, info, m_CLCompileContext);
                default:
                    return nullptr;
            }
        }
        case LayerType::Fill :
        {
            auto fillQueueDescriptor = PolymorphicDowncast<const FillQueueDescriptor*>(&descriptor);
            return std::make_unique<ClFillWorkload>(*fillQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Floor :
        {
            auto floorQueueDescriptor = PolymorphicDowncast<const FloorQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClFloorFloatWorkload, NullWorkload>(*floorQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::FullyConnected :
        {
            auto fullyConnectedQueueDescriptor
                     = PolymorphicDowncast<const FullyConnectedQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClFullyConnectedWorkload>(*fullyConnectedQueueDescriptor,
                                                          info,
                                                          m_MemoryManager->GetIntraLayerManager(),
                                                          m_CLCompileContext);
        }
        case LayerType::Gather :
        {
            auto gatherQueueDescriptor = PolymorphicDowncast<const GatherQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClGatherWorkload>(*gatherQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::GatherNd :
        {
            auto gatherNdQueueDescriptor = PolymorphicDowncast<const GatherNdQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClGatherNdWorkload>(*gatherNdQueueDescriptor, info, m_CLCompileContext);
        }
        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 MakeWorkload<ClInstanceNormalizationWorkload>(*instanceNormalizationQueueDescriptor,
                                                                 info,
                                                                 m_CLCompileContext);
        }
        case LayerType::L2Normalization :
        {
            auto l2NormalizationQueueDescriptor
                     = PolymorphicDowncast<const L2NormalizationQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClL2NormalizationFloatWorkload, NullWorkload>(*l2NormalizationQueueDescriptor,
                                                                              info,
                                                                              m_CLCompileContext);
        }
        case LayerType::LogicalBinary :
        {
            auto logicalBinaryQueueDescriptor = PolymorphicDowncast<const LogicalBinaryQueueDescriptor*>(&descriptor);
            switch(logicalBinaryQueueDescriptor->m_Parameters.m_Operation)
            {
                case LogicalBinaryOperation::LogicalAnd:
                    return std::make_unique<ClLogicalAndWorkload>(*logicalBinaryQueueDescriptor,
                                                                  info,
                                                                  m_CLCompileContext);
                case LogicalBinaryOperation::LogicalOr:
                    return std::make_unique<ClLogicalOrWorkload>(*logicalBinaryQueueDescriptor,
                                                                 info,
                                                                 m_CLCompileContext);
                default:
                    return nullptr;
            }
        }
        case LayerType::LogSoftmax :
        {
            auto logSoftmaxQueueDescriptor = PolymorphicDowncast<const LogSoftmaxQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClLogSoftmaxWorkload>(*logSoftmaxQueueDescriptor,
                                                      info,
                                                      m_MemoryManager->GetIntraLayerManager(),
                                                      m_CLCompileContext);
        }
        case LayerType::Lstm :
        {
            auto lstmQueueDescriptor = PolymorphicDowncast<const LstmQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClLstmFloatWorkload, NullWorkload>(*lstmQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Maximum :
        {
            auto maximumQueueDescriptor = PolymorphicDowncast<const MaximumQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClMaximumWorkload>(*maximumQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Mean :
        {
            auto meanQueueDescriptor = PolymorphicDowncast<const MeanQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClMeanWorkload>(*meanQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::MemCopy :
        {
            auto memCopyQueueDescriptor = PolymorphicDowncast<const MemCopyQueueDescriptor*>(&descriptor);
            if (memCopyQueueDescriptor->m_Inputs.empty() || !memCopyQueueDescriptor->m_Inputs[0])
            {
                throw InvalidArgumentException("ClWorkloadFactory: Invalid null input for MemCopy workload");
            }
            return MakeWorkload<CopyMemGenericWorkload>(*memCopyQueueDescriptor, info);
        }
        case LayerType::MemImport :
        {
            auto memImportQueueDescriptor = PolymorphicDowncast<const MemImportQueueDescriptor*>(&descriptor);
            if (memImportQueueDescriptor->m_Inputs.empty() || !memImportQueueDescriptor->m_Inputs[0])
            {
                throw InvalidArgumentException("ClWorkloadFactory: Invalid null input for MemImport workload");
            }
            return std::make_unique<ImportMemGenericWorkload>(*memImportQueueDescriptor, info);
        }
        case LayerType::Minimum :
        {
            auto minimumQueueDescriptor = PolymorphicDowncast<const MinimumQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClMinimumWorkload>(*minimumQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Multiplication :
        {
            auto multiplicationQueueDescriptor = PolymorphicDowncast<const MultiplicationQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClMultiplicationWorkload>(*multiplicationQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Normalization :
        {
            auto normalizationQueueDescriptor = PolymorphicDowncast<const NormalizationQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClNormalizationFloatWorkload, NullWorkload>(*normalizationQueueDescriptor,
                                                                            info,
                                                                            m_CLCompileContext);
        }
        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 MakeWorkload<ClPadWorkload>(*padQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Permute :
        {
            auto permuteQueueDescriptor = PolymorphicDowncast<const PermuteQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClPermuteWorkload>(*permuteQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Pooling2d :
        {
            auto pooling2dQueueDescriptor = PolymorphicDowncast<const Pooling2dQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClPooling2dWorkload>(*pooling2dQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Pooling3d :
        {
            auto pooling3dQueueDescriptor = PolymorphicDowncast<const Pooling3dQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClPooling3dWorkload>(*pooling3dQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::PreCompiled :
        {
            auto preCompiledQueueDescriptor = PolymorphicDowncast<const PreCompiledQueueDescriptor*>(&descriptor);
            return MakeWorkload<NullWorkload, NullWorkload>(*preCompiledQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Prelu :
        {
            auto preluQueueDescriptor = PolymorphicDowncast<const PreluQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClPreluWorkload>(*preluQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::QLstm :
        {
            auto qLstmQueueDescriptor = PolymorphicDowncast<const QLstmQueueDescriptor*>(&descriptor);
            return std::make_unique<ClQLstmWorkload>(*qLstmQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Quantize :
        {
            auto quantizeQueueDescriptor = PolymorphicDowncast<const QuantizeQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClQuantizeWorkload>(*quantizeQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::QuantizedLstm :
        {
            auto quantizedLstmQueueDescriptor = PolymorphicDowncast<const QuantizedLstmQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClQuantizedLstmWorkload>(*quantizedLstmQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Rank :
        {
            auto rankQueueDescriptor = PolymorphicDowncast<const RankQueueDescriptor*>(&descriptor);
            return std::make_unique<ClRankWorkload>(*rankQueueDescriptor, info);
        }
        case LayerType::Reduce :
        {
            auto reduceQueueDescriptor = PolymorphicDowncast<const ReduceQueueDescriptor*>(&descriptor);
            return std::make_unique<ClReduceWorkload>(*reduceQueueDescriptor, info);
        }
        case LayerType::Reshape :
        {
            auto reshapeQueueDescriptor = PolymorphicDowncast<const ReshapeQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClReshapeWorkload>(*reshapeQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Resize :
        {
            auto resizeQueueDescriptor = PolymorphicDowncast<const ResizeQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClResizeWorkload>(*resizeQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::ReverseV2:
        {
            auto reverseV2QueueDescriptor = PolymorphicDowncast<const ReverseV2QueueDescriptor*>(&descriptor);
            return MakeWorkload<ClReverseV2Workload>(*reverseV2QueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::ScatterNd :
        {
            auto scatterNdQueueDescriptor = PolymorphicDowncast<const ScatterNdQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClScatterNdWorkload>(*scatterNdQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Slice :
        {
            auto sliceQueueDescriptor = PolymorphicDowncast<const SliceQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClSliceWorkload>(*sliceQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Softmax :
        {
            auto softmaxQueueDescriptor = PolymorphicDowncast<const SoftmaxQueueDescriptor*>(&descriptor);
            return std::make_unique<ClSoftmaxWorkload>(*softmaxQueueDescriptor,
                                                       info,
                                                       m_MemoryManager->GetIntraLayerManager(),
                                                       m_CLCompileContext);
        }
        case LayerType::SpaceToBatchNd :
        {
            auto spaceToBatchNdQueueDescriptor
                     = PolymorphicDowncast<const SpaceToBatchNdQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClSpaceToBatchNdWorkload>(*spaceToBatchNdQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::SpaceToDepth :
        {
            auto spaceToDepthQueueDescriptor = PolymorphicDowncast<const SpaceToDepthQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClSpaceToDepthWorkload>(*spaceToDepthQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Splitter :
        {
            auto splitterQueueDescriptor = PolymorphicDowncast<const SplitterQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClSplitterWorkload>(*splitterQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Stack :
        {
            auto stackQueueDescriptor = PolymorphicDowncast<const StackQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClStackWorkload>(*stackQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::StridedSlice :
        {
            auto stridedSliceQueueDescriptor = PolymorphicDowncast<const StridedSliceQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClStridedSliceWorkload>(*stridedSliceQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Subtraction :
        {
            auto subtractionQueueDescriptor = PolymorphicDowncast<const SubtractionQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClSubtractionWorkload>(*subtractionQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Tile:
        {
            auto tileQueueDescriptor = PolymorphicDowncast<const TileQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClTileWorkload>(*tileQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::Transpose :
        {
            auto transposeQueueDescriptor = PolymorphicDowncast<const TransposeQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClTransposeWorkload>(*transposeQueueDescriptor, info, m_CLCompileContext);
        }
        case LayerType::TransposeConvolution2d :
        {
            auto transposeConvolution2dQueueDescriptor
                     = PolymorphicDowncast<const TransposeConvolution2dQueueDescriptor*>(&descriptor);
            return MakeWorkload<ClTransposeConvolution2dWorkload>(*transposeConvolution2dQueueDescriptor,
                                                                  info,
                                                                  m_MemoryManager->GetIntraLayerManager(),
                                                                  m_CLCompileContext);
        }
        case LayerType::UnidirectionalSequenceLstm :
        {
            auto desc = PolymorphicDowncast<const UnidirectionalSequenceLstmQueueDescriptor*>(&descriptor);
            return MakeWorkloadHelper<ClUnidirectionalSequenceLstmFloatWorkload, NullWorkload>(*desc,
                                                                                               info,
                                                                                               m_CLCompileContext);
        }
        default:
            return nullptr;
    }
}

References armnn::Abs, armnn::Activation, armnn::Add, armnn::Addition, armnn::ArgMinMax, armnn::BatchMatMul, armnn::BatchNormalization, armnn::BatchToSpaceNd, armnn::Cast, armnn::ChannelShuffle, armnn::Comparison, armnn::Concat, armnn::Constant, armnn::ConvertFp16ToFp32, armnn::ConvertFp32ToFp16, armnn::Convolution2d, armnn::Convolution3d, armnn::Debug, armnn::DepthToSpace, armnn::DepthwiseConvolution2d, armnn::Dequantize, armnn::DetectionPostProcess, armnn::Div, armnn::Division, armnn::ElementwiseBinary, armnn::ElementwiseUnary, armnn::Exp, armnn::Fill, armnn::Floor, armnn::FloorDiv, armnn::FullyConnected, armnn::Gather, armnn::GatherNd, armnn::info, armnn::Input, armnn::InstanceNormalization, ClBackendModelContext::IsFastMathEnabled(), armnn::L2Normalization, armnn::Log, armnn::LogicalAnd, armnn::LogicalBinary, armnn::LogicalNot, armnn::LogicalOr, armnn::LogSoftmax, armnn::Lstm, QueueDescriptor::m_AdditionalInfoObject, QueueDescriptor::m_Inputs, QueueDescriptor::m_Outputs, armnn::Maximum, armnn::Mean, armnn::MemCopy, armnn::MemImport, armnn::Minimum, armnn::Mul, armnn::Multiplication, armnn::Neg, armnn::Normalization, armnn::Output, armnn::Pad, armnn::Permute, armnn::Pooling2d, armnn::Pooling3d, armnn::Power, armnn::PreCompiled, armnn::Prelu, armnn::QLstm, armnn::Quantize, armnn::QuantizedLstm, armnn::Rank, armnn::Reduce, armnn::Reshape, armnn::Resize, armnn::ReverseV2, armnn::Rsqrt, armnn::ScatterNd, armnn::Sin, armnn::Slice, armnn::Softmax, armnn::SpaceToBatchNd, armnn::SpaceToDepth, armnn::Splitter, armnn::SqDiff, armnn::Sqrt, armnn::Stack, armnn::StridedSlice, armnn::Sub, armnn::Subtraction, armnn::Tile, armnn::Transpose, armnn::TransposeConvolution2d, and armnn::UnidirectionalSequenceLstm.

GetBackendId()

const BackendId & GetBackendId ( ) const

overridevirtual

Implements IWorkloadFactory.

Definition at line 61 of file ClWorkloadFactory.cpp.

{
    return s_Id;
}

IsLayerSupported() [1/2]

bool IsLayerSupported ( const IConnectableLayer & layer, Optional< DataType > dataType, std::string & outReasonIfUnsupported, const ModelOptions & modelOptions )

static

Definition at line 53 of file ClWorkloadFactory.cpp.

{
    return IWorkloadFactory::IsLayerSupported(s_Id, layer, dataType, outReasonIfUnsupported, modelOptions);
}

References IWorkloadFactory::IsLayerSupported().

IsLayerSupported() [2/2]

bool IsLayerSupported ( const Layer & layer, Optional< DataType > dataType, std::string & outReasonIfUnsupported )

static

Definition at line 46 of file ClWorkloadFactory.cpp.

{
    return IWorkloadFactory::IsLayerSupported(s_Id, layer, dataType, outReasonIfUnsupported);
}

References IWorkloadFactory::IsLayerSupported().

SupportsSubTensors()

bool SupportsSubTensors ( ) const

inlineoverridevirtual

Implements IWorkloadFactory.

Definition at line 42 of file ClWorkloadFactory.hpp.

{ return true; }

---

The documentation for this class was generated from the following files:

• src/backends/cl/ClWorkloadFactory.hpp
• src/backends/cl/ClWorkloadFactory.cpp