ClLstmFloatWorkload Class Reference

#include <ClLstmFloatWorkload.hpp>

Inheritance diagram for ClLstmFloatWorkload:

Public Member Functions
	ClLstmFloatWorkload (const LstmQueueDescriptor &descriptor, const WorkloadInfo &info, const arm_compute::CLCompileContext &clCompileContext)
void	Execute () const override
void	ReplaceInputTensorHandle (ITensorHandle *tensorHandle, unsigned int slot) override
void	ReplaceOutputTensorHandle (ITensorHandle *tensorHandle, unsigned int slot) override
Public Member Functions inherited from TypedWorkload< QueueDescriptor, DataTypes >
	TypedWorkload (const QueueDescriptor &descriptor, const WorkloadInfo &info)
Public Member Functions inherited from BaseWorkload< QueueDescriptor >
	BaseWorkload (const QueueDescriptor &descriptor, const WorkloadInfo &info)
void	ExecuteAsync (WorkingMemDescriptor &workingMemDescriptor) override
void	PostAllocationConfigure () override
const QueueDescriptor &	GetData () const
profiling::ProfilingGuid	GetGuid () const final
virtual bool	SupportsTensorHandleReplacement () const override
Public Member Functions inherited from IWorkload
virtual	~IWorkload ()
virtual void	RegisterDebugCallback (const DebugCallbackFunction &)

Additional Inherited Members
Protected Attributes inherited from BaseWorkload< QueueDescriptor >
QueueDescriptor	m_Data
const profiling::ProfilingGuid	m_Guid

Detailed Description

Definition at line 18 of file ClLstmFloatWorkload.hpp.

Constructor & Destructor Documentation

ClLstmFloatWorkload()

ClLstmFloatWorkload	(	const LstmQueueDescriptor &	descriptor,
		const WorkloadInfo &	info,
		const arm_compute::CLCompileContext &	clCompileContext
	)

Definition at line 22 of file ClLstmFloatWorkload.cpp.

References ARMNN_REPORT_PROFILING_WORKLOAD_DESC, and QueueDescriptorWithParameters< LayerDescriptor >::m_Parameters.

        : FloatWorkload<LstmQueueDescriptor>(descriptor, info)
{
    // Report Profiling Details
    ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClLstmFloatWorkload_Construct",
                                         descriptor.m_Parameters,
                                         info,
                                         this->GetGuid());

    arm_compute::LSTMParams<arm_compute::ICLTensor> lstm_param;

    // Basic parameters
    m_InputToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>();
    BuildArmComputeTensor(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights->GetTensorInfo());

    m_InputToCellWeightsTensor = std::make_unique<arm_compute::CLTensor>();
    BuildArmComputeTensor(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights->GetTensorInfo());

    m_InputToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
    BuildArmComputeTensor(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights->GetTensorInfo());

    m_RecurrentToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>();
    BuildArmComputeTensor(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights->GetTensorInfo());

    m_RecurrentToCellWeightsTensor = std::make_unique<arm_compute::CLTensor>();
    BuildArmComputeTensor(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights->GetTensorInfo());

    m_RecurrentToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
    BuildArmComputeTensor(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights->GetTensorInfo());

    m_ForgetGateBiasTensor = std::make_unique<arm_compute::CLTensor>();
    BuildArmComputeTensor(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias->GetTensorInfo());

    m_CellBiasTensor = std::make_unique<arm_compute::CLTensor>();
    BuildArmComputeTensor(*m_CellBiasTensor, m_Data.m_CellBias->GetTensorInfo());

    m_OutputGateBiasTensor = std::make_unique<arm_compute::CLTensor>();
    BuildArmComputeTensor(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias->GetTensorInfo());

    // for future reference: check the AndroidNN API for the logic here
    if (!m_Data.m_Parameters.m_CifgEnabled)
    {
        m_InputToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
        BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo());

        m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
        BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo());

        m_CellToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
        if (m_Data.m_CellToInputWeights != nullptr)
        {
            BuildArmComputeTensor(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights->GetTensorInfo());
        }

        m_InputGateBiasTensor = std::make_unique<arm_compute::CLTensor>();
        BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo());

        lstm_param.set_cifg_params(m_InputToInputWeightsTensor.get(),
                                   m_RecurrentToInputWeightsTensor.get(),
                                   m_Data.m_CellToInputWeights != nullptr ? m_CellToInputWeightsTensor.get() : nullptr,
                                   m_InputGateBiasTensor.get());
    }

    if (m_Data.m_Parameters.m_ProjectionEnabled)
    {
        m_ProjectionWeightsTensor = std::make_unique<arm_compute::CLTensor>();
        BuildArmComputeTensor(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights->GetTensorInfo());

        m_ProjectionBiasTensor = std::make_unique<arm_compute::CLTensor>();
        if (m_Data.m_ProjectionBias != nullptr)
        {
            BuildArmComputeTensor(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias->GetTensorInfo());
        }

        lstm_param.set_projection_params(m_ProjectionWeightsTensor.get(),
                                         m_Data.m_ProjectionBias != nullptr ? m_ProjectionBiasTensor.get() : nullptr);
    }

    if (m_Data.m_Parameters.m_PeepholeEnabled)
    {
        m_CellToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>();
        BuildArmComputeTensor(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights->GetTensorInfo());

        m_CellToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>();
        BuildArmComputeTensor(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights->GetTensorInfo());

        lstm_param.set_peephole_params(m_CellToForgetWeightsTensor.get(), m_CellToOutputWeightsTensor.get());
    }

    if (m_Data.m_Parameters.m_LayerNormEnabled)
    {
        m_InputLayerNormWeightsTensor  = std::make_unique<arm_compute::CLTensor>();
        m_ForgetLayerNormWeightsTensor = std::make_unique<arm_compute::CLTensor>();
        m_CellLayerNormWeightsTensor   = std::make_unique<arm_compute::CLTensor>();
        m_OutputLayerNormWeightsTensor = std::make_unique<arm_compute::CLTensor>();

        if (!m_Data.m_Parameters.m_CifgEnabled)
        {
            BuildArmComputeTensor(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights->GetTensorInfo());
        }
        BuildArmComputeTensor(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights->GetTensorInfo());
        BuildArmComputeTensor(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights->GetTensorInfo());
        BuildArmComputeTensor(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights->GetTensorInfo());

        lstm_param.set_layer_normalization_params(m_Data.m_Parameters.m_CifgEnabled ? nullptr :
                                                  m_InputLayerNormWeightsTensor.get(),
                                                  m_ForgetLayerNormWeightsTensor.get(),
                                                  m_CellLayerNormWeightsTensor.get(),
                                                  m_OutputLayerNormWeightsTensor.get());
    }

    const arm_compute::ICLTensor& input           = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
    const arm_compute::ICLTensor& output_state_in = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
    arm_compute::ICLTensor& cell_state_in         = static_cast<IClTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();

    arm_compute::ICLTensor& output_state_out      = static_cast<IClTensorHandle*>(m_Data.m_Outputs[1])->GetTensor();
    arm_compute::ICLTensor& cell_state_out        = static_cast<IClTensorHandle*>(m_Data.m_Outputs[2])->GetTensor();
    arm_compute::ICLTensor& output                = static_cast<IClTensorHandle*>(m_Data.m_Outputs[3])->GetTensor();

    // Get the batch_size and the num_units from the cellStateIn dimensions
    const TensorInfo& inputTensorInfo = info.m_InputTensorInfos[2];
    const unsigned int batch_size = armnn::numeric_cast<unsigned int>(inputTensorInfo.GetShape()[0]);
    const unsigned int num_units  = armnn::numeric_cast<unsigned int>(inputTensorInfo.GetShape()[1]);

    m_ScratchBuffer = std::make_unique<arm_compute::CLTensor>();
    if (m_Data.m_Parameters.m_CifgEnabled)
    {
        // 2D tensor with dimensions [num_units * 3, batch_size] with CIFG
        armnn::TensorInfo scratchBuffer1({ batch_size, num_units * 3 }, DataType::Float32);
        BuildArmComputeTensor(*m_ScratchBuffer, scratchBuffer1);
    }
    else
    {
        // scratch_buffer [num_units * 4, batch_size] without CIFG
        armnn::TensorInfo scratchBuffer2({ batch_size, num_units * 4 }, DataType::Float32);
        BuildArmComputeTensor(*m_ScratchBuffer, scratchBuffer2);
    }

    float cell_threshold = m_Data.m_Parameters.m_ClippingThresCell;
    float projection_threshold = m_Data.m_Parameters.m_ClippingThresProj;

    // for preparing the object for the class ActivationLayerInfo, we need to consider 5 situations
    arm_compute::ActivationLayerInfo activationLayerInfo;
    if (m_Data.m_Parameters.m_ActivationFunc == 0)
    {
        // no activation, do nothing
    }
    else if (m_Data.m_Parameters.m_ActivationFunc == 1)
    {
        activationLayerInfo = arm_compute::ActivationLayerInfo(
                arm_compute::ActivationLayerInfo::ActivationFunction::RELU);
    }
    else if (m_Data.m_Parameters.m_ActivationFunc == 3)
    {
        activationLayerInfo = arm_compute::ActivationLayerInfo(
                arm_compute::ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.0);
    }
    else if (m_Data.m_Parameters.m_ActivationFunc == 4)
    {
        activationLayerInfo =  arm_compute::ActivationLayerInfo(
                arm_compute::ActivationLayerInfo::ActivationFunction::TANH, 1.0, 1.0);
    }
    else if (m_Data.m_Parameters.m_ActivationFunc == 6)
    {
        activationLayerInfo =  arm_compute::ActivationLayerInfo(
                arm_compute::ActivationLayerInfo::ActivationFunction::LOGISTIC);
    }
    else
    {
        throw armnn::Exception("Wrong Type of Activation Function!");
    }

    {
        ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClLstmFloatWorkload_configure");
        m_LstmLayer.configure(clCompileContext, &input, m_InputToForgetWeightsTensor.get(),
                              m_InputToCellWeightsTensor.get(), m_InputToOutputWeightsTensor.get(),
                              m_RecurrentToForgetWeightsTensor.get(), m_RecurrentToCellWeightsTensor.get(),
                              m_RecurrentToOutputWeightsTensor.get(), m_ForgetGateBiasTensor.get(),
                              m_CellBiasTensor.get(), m_OutputGateBiasTensor.get(), &output_state_in,
                              &cell_state_in, m_ScratchBuffer.get(), &output_state_out,
                              &cell_state_out, &output, lstm_param, activationLayerInfo,
                              cell_threshold, projection_threshold);
    }

    armcomputetensorutils::InitialiseArmComputeTensorEmpty(*m_ScratchBuffer);

    InitializeArmComputeClTensorData(*m_InputToForgetWeightsTensor,     m_Data.m_InputToForgetWeights);
    InitializeArmComputeClTensorData(*m_InputToCellWeightsTensor,       m_Data.m_InputToCellWeights);
    InitializeArmComputeClTensorData(*m_InputToOutputWeightsTensor,     m_Data.m_InputToOutputWeights);
    InitializeArmComputeClTensorData(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights);
    InitializeArmComputeClTensorData(*m_RecurrentToCellWeightsTensor,   m_Data.m_RecurrentToCellWeights);
    InitializeArmComputeClTensorData(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights);
    InitializeArmComputeClTensorData(*m_ForgetGateBiasTensor,           m_Data.m_ForgetGateBias);
    InitializeArmComputeClTensorData(*m_CellBiasTensor,                 m_Data.m_CellBias);
    InitializeArmComputeClTensorData(*m_OutputGateBiasTensor,           m_Data.m_OutputGateBias);

    if (!m_Data.m_Parameters.m_CifgEnabled)
    {
        InitializeArmComputeClTensorData(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights);
        InitializeArmComputeClTensorData(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights);
        if (m_Data.m_CellToInputWeights != nullptr)
        {
            InitializeArmComputeClTensorData(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights);
        }
        InitializeArmComputeClTensorData(*m_InputGateBiasTensor, m_Data.m_InputGateBias);
    }

    if (m_Data.m_Parameters.m_ProjectionEnabled)
    {
        InitializeArmComputeClTensorData(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights);
        if (m_Data.m_ProjectionBias != nullptr)
        {
            InitializeArmComputeClTensorData(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias);
        }
    }

    if (m_Data.m_Parameters.m_PeepholeEnabled)
    {
        InitializeArmComputeClTensorData(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights);
        InitializeArmComputeClTensorData(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights);
    }

    if (m_Data.m_Parameters.m_LayerNormEnabled)
    {
        if (!m_Data.m_Parameters.m_CifgEnabled)
        {
            InitializeArmComputeClTensorData(*m_InputLayerNormWeightsTensor,  m_Data.m_InputLayerNormWeights);
        }

        InitializeArmComputeClTensorData(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights);
        InitializeArmComputeClTensorData(*m_CellLayerNormWeightsTensor,   m_Data.m_CellLayerNormWeights);
        InitializeArmComputeClTensorData(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights);
    }

    // Force Compute Library to perform the necessary copying and reshaping, after which
    // delete all the input tensors that will no longer be needed
    m_LstmLayer.prepare();
    FreeUnusedTensors();
}

Member Function Documentation

Execute()

void Execute ( ) const

overridevirtual

Implements IWorkload.

Definition at line 264 of file ClLstmFloatWorkload.cpp.

References ARMNN_SCOPED_PROFILING_EVENT_CL_GUID, CHECK_LOCATION, BaseWorkload< QueueDescriptor >::GetGuid(), and armnn::RunClFunction().

{
    ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClLstmFloatWorkload_Execute", this->GetGuid());
    RunClFunction(m_LstmLayer, CHECK_LOCATION());
}

ReplaceInputTensorHandle()

void ReplaceInputTensorHandle ( ITensorHandle *  tensorHandle, unsigned int  slot  )

overridevirtual

Reimplemented from BaseWorkload< QueueDescriptor >.

Definition at line 449 of file ClLstmFloatWorkload.cpp.

References BaseWorkload< QueueDescriptor >::m_Data, and QueueDescriptor::m_Inputs.

{
    ITensorHandle* backupHandle = this->m_Data.m_Inputs[slot];
    this->m_Data.m_Inputs[slot] = tensorHandle;
    try
    {
        Reconfigure();
    }
    catch(armnn::UnimplementedException& e)
    {
        // Cannot reconfigure, revert the slot back and throw the exception.
        this->m_Data.m_Inputs[slot] = backupHandle;
        throw e;
    }
}

ReplaceOutputTensorHandle()

void ReplaceOutputTensorHandle ( ITensorHandle *  tensorHandle, unsigned int  slot  )

overridevirtual

Reimplemented from BaseWorkload< QueueDescriptor >.

Definition at line 466 of file ClLstmFloatWorkload.cpp.

References BaseWorkload< QueueDescriptor >::m_Data, and QueueDescriptor::m_Inputs.

{
    ITensorHandle* backupHandle = this->m_Data.m_Inputs[slot];
    this->m_Data.m_Inputs[slot] = tensorHandle;
    try
    {
        Reconfigure();
    }
    catch(armnn::UnimplementedException& e)
    {
        // Cannot reconfigure, revert the slot back and throw the exception.
        this->m_Data.m_Inputs[slot] = backupHandle;
        throw e;
    }
}

---

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

• src/backends/cl/workloads/ClLstmFloatWorkload.hpp
• src/backends/cl/workloads/ClLstmFloatWorkload.cpp