NeonGatherNdWorkload Class Reference

#include <NeonGatherNdWorkload.hpp>

Inheritance diagram for NeonGatherNdWorkload:

Collaboration diagram for NeonGatherNdWorkload:

Public Member Functions
	NeonGatherNdWorkload (const GatherNdQueueDescriptor &descriptor, const WorkloadInfo &info)
virtual void	Execute () const override
Public Member Functions inherited from NeonBaseWorkload< GatherNdQueueDescriptor >
	NeonBaseWorkload (const GatherNdQueueDescriptor &descriptor, const WorkloadInfo &info)
void	ReplaceInputTensorHandle (ITensorHandle *tensorHandle, unsigned int slot) override
void	ReplaceOutputTensorHandle (ITensorHandle *tensorHandle, unsigned int slot) override
Public Member Functions inherited from BaseWorkload< QueueDescriptor >
	BaseWorkload (const QueueDescriptor &descriptor, const WorkloadInfo &info)
virtual const std::string &	GetName () const override
void	PostAllocationConfigure () override
const QueueDescriptor &	GetData () const
arm::pipe::ProfilingGuid	GetGuid () const final
virtual bool	SupportsTensorHandleReplacement () const override
Public Member Functions inherited from IWorkload
virtual	~IWorkload ()
virtual void	RegisterDebugCallback (const DebugCallbackFunction &)
virtual armnn::Optional< armnn::MemoryRequirements >	GetMemoryRequirements ()

Additional Inherited Members
Protected Member Functions inherited from NeonBaseWorkload< GatherNdQueueDescriptor >
virtual void	Reconfigure ()
Protected Attributes inherited from BaseWorkload< QueueDescriptor >
QueueDescriptor	m_Data
const arm::pipe::ProfilingGuid	m_Guid
const std::string	m_Name

Detailed Description

Definition at line 22 of file NeonGatherNdWorkload.hpp.

Constructor & Destructor Documentation

NeonGatherNdWorkload()

NeonGatherNdWorkload	(	const GatherNdQueueDescriptor &	descriptor,
		const WorkloadInfo &	info
	)

Calculate flattened indices: m_FlattenedIndices = indices * m_FlattenedCoeff. This could be done using MatMul instead of multiplication followed by reduce sum operation, but GeMM does not support s32 at the moment.

Call Gather with adequate shapes

Definition at line 107 of file NeonGatherNdWorkload.cpp.

        : NeonBaseWorkload<GatherNdQueueDescriptor>(descriptor, info)
{
    m_Data.ValidateInputsOutputs("NeonGatherNdWorkload", 2, 1);
 
    TensorInfo paramsInfo  = info.m_InputTensorInfos[0];
    TensorInfo indicesInfo = info.m_InputTensorInfos[1];
    TensorInfo outputInfo  = info.m_OutputTensorInfos[0];
 
    arm_compute::ITensor& input   = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
    arm_compute::ITensor& indices = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
    arm_compute::ITensor& output  = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
 
    // Calculate ND, K, W, C.
    std::map<std::string, unsigned int> keyIndices = CalculateGatherNdKeyIndices(paramsInfo, indicesInfo);
 
    /// Calculate flattened indices: m_FlattenedIndices = indices * m_FlattenedCoeff.
    /// This could be done using MatMul instead of multiplication followed by reduce sum operation,
    /// but GeMM does not support s32 at the moment.
 
    // Prepare the tensor to store the output of the reduce_sum operation
    armnn::TensorInfo flattenedIndices_Info = indicesInfo;
    flattenedIndices_Info.SetShape({ keyIndices["W"] });
    BuildArmComputeTensor(m_FlattenedIndices, flattenedIndices_Info);
    armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_FlattenedIndices);
 
    // Reshape indices into { W, ND }
    armnn::TensorInfo indicesInfoReshape = indicesInfo;
    indicesInfoReshape.SetShape({ keyIndices["W"], keyIndices["ND"] });
    BuildArmComputeTensor(m_IndicesReshaped, indicesInfoReshape);
    armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_IndicesReshaped);
 
    // Calculate the m_FlattenedCoeff
    TensorShape paramsShape = paramsInfo.GetShape();
    std::vector<int32_t> flattenedCoeff(keyIndices["ND"], 1);
    for (unsigned int i = 1; i < keyIndices["ND"]; ++i)
    {
        flattenedCoeff[i - 1] = static_cast<int32_t>(paramsShape[i]);
    }
    for (unsigned int i = keyIndices["ND"] - 1; i > 0; --i)
    {
        flattenedCoeff[i - 1] *= flattenedCoeff[i];
    }
    armnn::TensorInfo flattenedCoeff_Info = indicesInfo;
    flattenedCoeff_Info.SetShape({ keyIndices["ND"] });
    BuildArmComputeTensor(m_FlattenedCoeff, flattenedCoeff_Info);
    armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_FlattenedCoeff);
    CopyArmComputeITensorData<int32_t>(flattenedCoeff.data(), m_FlattenedCoeff);
 
    // Prepare the tensor to store the output of the multiplication
    armnn::TensorInfo outputMul_Info = indicesInfo;
    outputMul_Info.SetShape({ keyIndices["W"], keyIndices["ND"] });
    BuildArmComputeTensor(m_OutputMul, outputMul_Info);
    armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_OutputMul);
 
    // Reshape indices to the mul layer input shape
    m_ReshapeIndicesLayer.configure(&indices, &m_IndicesReshaped);
 
    // Multiply
    m_MulLayer.configure(&m_IndicesReshaped,
                         &m_FlattenedCoeff,
                         &m_OutputMul,
                         1.0f,
                         arm_compute::ConvertPolicy::WRAP,
                         arm_compute::RoundingPolicy::TO_ZERO,
                         arm_compute::ActivationLayerInfo());
 
    // Reduce Sum
    const std::vector<unsigned int> armnnReduceAxes(1, 1);
    arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(m_OutputMul.info()->num_dimensions(),
                                                                          outputMul_Info.GetNumDimensions(),
                                                                          armnnReduceAxes);
    m_ReduceSumLayer.configure(&m_OutputMul,
                               &m_FlattenedIndices,
                               static_cast<unsigned int>(coords[0]),
                               arm_compute::ReductionOperation::SUM,
                               false);
 
    /// Call Gather with adequate shapes
    // Reshape params into { K, C }
    armnn::TensorInfo paramsInfoReshape = paramsInfo;
    paramsInfoReshape.SetShape({ keyIndices["K"], keyIndices["C"] });
    BuildArmComputeTensor(m_InputGather, paramsInfoReshape);
    armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_InputGather);
 
    // Reshape input to the gather params input shape
    m_ReshapeInputLayer.configure(&input, &m_InputGather);
 
    // Reshape output to have the shape given by gather { W, C }
    // (the original outputInfo has the shape given by gatherNd)
    armnn::TensorInfo outputGather_Info = outputInfo;
    outputGather_Info.SetShape({ keyIndices["W"], keyIndices["C"] });
    BuildArmComputeTensor(m_OutputGather, outputGather_Info);
    armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_OutputGather);
 
    m_GatherLayer.configure(&m_InputGather,
                            &m_FlattenedIndices,
                            &m_OutputGather,
                            ComputeAclAxis(0, paramsInfoReshape));
 
    // Reshape output to the original output shape
    m_ReshapeOutputLayer.configure(&m_OutputGather, &output);
}

References armnn::CalculateGatherNdKeyIndices(), armnn::info, BaseWorkload< QueueDescriptor >::m_Data, QueueDescriptor::m_Inputs, QueueDescriptor::m_Outputs, TensorInfo::SetShape(), and QueueDescriptor::ValidateInputsOutputs().

Member Function Documentation

Execute()

void Execute ( ) const

overridevirtual

Implements IWorkload.

Definition at line 212 of file NeonGatherNdWorkload.cpp.

{
    ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID("NeonGatherNdWorkload_Execute");
    m_ReshapeInputLayer.run();
    m_ReshapeIndicesLayer.run();
    m_MulLayer.run();
    m_ReduceSumLayer.run();
    m_GatherLayer.run();
    m_ReshapeOutputLayer.run();
}

References ARMNN_SCOPED_PROFILING_EVENT_NEON_NAME_GUID.

---

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

• src/backends/neon/workloads/NeonGatherNdWorkload.hpp
• src/backends/neon/workloads/NeonGatherNdWorkload.cpp