NeonBatchMatMulWorkload Class Reference

#include <NeonBatchMatMulWorkload.hpp>

Inheritance diagram for NeonBatchMatMulWorkload:

Public Member Functions
	NeonBatchMatMulWorkload (const BatchMatMulQueueDescriptor &descriptor, const WorkloadInfo &info)
virtual void	Execute () const override
Public Member Functions inherited from NeonBaseWorkload< BatchMatMulQueueDescriptor >
	NeonBaseWorkload (const BatchMatMulQueueDescriptor &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< BatchMatMulQueueDescriptor >
	BaseWorkload (const BatchMatMulQueueDescriptor &descriptor, const WorkloadInfo &info)
void	ExecuteAsync (ExecutionData &executionData) override
void	PostAllocationConfigure () override
const BatchMatMulQueueDescriptor &	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< BatchMatMulQueueDescriptor >
virtual void	Reconfigure ()
Protected Attributes inherited from BaseWorkload< BatchMatMulQueueDescriptor >
BatchMatMulQueueDescriptor	m_Data
const arm::pipe::ProfilingGuid	m_Guid

Detailed Description

Definition at line 22 of file NeonBatchMatMulWorkload.hpp.

Constructor & Destructor Documentation

NeonBatchMatMulWorkload()

NeonBatchMatMulWorkload	(	const BatchMatMulQueueDescriptor &	descriptor,
		const WorkloadInfo &	info
	)

Definition at line 102 of file NeonBatchMatMulWorkload.cpp.

References ARMNN_REPORT_PROFILING_WORKLOAD_DESC, BatchMatMulDescriptor::m_AdjointX, BatchMatMulDescriptor::m_AdjointY, BaseWorkload< BatchMatMulQueueDescriptor >::m_Data, BatchMatMulDescriptor::m_DataLayoutX, BatchMatMulDescriptor::m_DataLayoutY, QueueDescriptor::m_Inputs, QueueDescriptor::m_Outputs, QueueDescriptorWithParameters< LayerDescriptor >::m_Parameters, armnn::NCHW, and QueueDescriptor::ValidateInputsOutputs().

    : NeonBaseWorkload<BatchMatMulQueueDescriptor>(descriptor, info)
{
    if (descriptor.m_Parameters.m_AdjointX || descriptor.m_Parameters.m_AdjointY )
    {
        throw Exception("Support for adjoint not implemented.");
    }
    if (descriptor.m_Parameters.m_DataLayoutX != armnn::DataLayout::NCHW ||
        descriptor.m_Parameters.m_DataLayoutY != armnn::DataLayout::NCHW )
    {
        throw Exception("Only supported the MatMul in the last 2 dimensions");
    }

    // Report Profiling Details
    ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonBatchMatMulWorkload_Construct",
                                         descriptor.m_Parameters,
                                         info,
                                         this->GetGuid());

    m_Data.ValidateInputsOutputs("NeonBatchMatMulWorkload", 2, 1);

    arm_compute::ITensor& inputX = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
    arm_compute::ITensor& inputY = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
    auto outputHandle = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0]);
    arm_compute::ITensor& output = outputHandle->GetTensor();

    arm_compute::DataLayout aclDataLayoutX = ConvertDataLayout(m_Data.m_Parameters.m_DataLayoutX);
    arm_compute::DataLayout aclDataLayoutY = ConvertDataLayout(m_Data.m_Parameters.m_DataLayoutY);

    inputX.info()->set_data_layout(aclDataLayoutX);
    inputY.info()->set_data_layout(aclDataLayoutY);

    if (descriptor.m_Parameters.m_TransposeX == true)
    {
        armnn::PermutationVector permutationXVector
                = GeneratePermutationVectorOnLastTwoDimensions(info.m_InputTensorInfos[0].GetNumDimensions());
        const TensorInfo permutedXInfo = armnnUtils::Permuted(info.m_InputTensorInfos[0], permutationXVector);
        const auto aclPermutationXVector = armcomputetensorutils::BuildArmComputePermutationVector(permutationXVector);

        auto permuteLayerX = std::make_unique<arm_compute::NEPermute>();
        BuildArmComputeTensor(m_PermutedTensorX, permutedXInfo);
        InitialiseArmComputeTensorEmpty(m_PermutedTensorX);
        permuteLayerX->configure(&inputX, &m_PermutedTensorX, aclPermutationXVector);
        m_PermuteLayerX.reset(permuteLayerX.release());
    }

    if (descriptor.m_Parameters.m_TransposeY == true)
    {
        armnn::PermutationVector permutationYVector
                = GeneratePermutationVectorOnLastTwoDimensions(info.m_InputTensorInfos[1].GetNumDimensions());
        const TensorInfo permutedYInfo = armnnUtils::Permuted(info.m_InputTensorInfos[1], permutationYVector);
        const auto aclPermutationYVector = armcomputetensorutils::BuildArmComputePermutationVector(permutationYVector);

        auto permuteLayerY = std::make_unique<arm_compute::NEPermute>();
        BuildArmComputeTensor(m_PermutedTensorY, permutedYInfo);
        InitialiseArmComputeTensorEmpty(m_PermutedTensorY);
        permuteLayerY->configure(&inputY, &m_PermutedTensorY, aclPermutationYVector);
        m_PermuteLayerY.reset(permuteLayerY.release());
    }

    const arm_compute::GEMMInfo& gemm_info = arm_compute::GEMMInfo(false,  // is inputX reshaped
                                                                   false,  // is inputY reshaped
                                                                   false); // is inputY reshaped only 1st run
    auto gemmLayer = std::make_unique<arm_compute::NEGEMM>();
    gemmLayer->configure(descriptor.m_Parameters.m_TransposeX ? &m_PermutedTensorX : &inputX,
                         descriptor.m_Parameters.m_TransposeY ? &m_PermutedTensorY : &inputY,
                         nullptr,
                         &output,
                         1.0,
                         0,
                         gemm_info);
    m_GEMMLayer.reset(gemmLayer.release());
}

Member Function Documentation

Execute()

void Execute ( ) const

overridevirtual

Implements IWorkload.

Definition at line 177 of file NeonBatchMatMulWorkload.cpp.

References ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID, and BaseWorkload< BatchMatMulQueueDescriptor >::GetGuid().

{
    ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonBatchMatMulWorkload_Execute", this->GetGuid());
    if (m_PermuteLayerX)
    {
        m_PermuteLayerX->run();
    }
    if (m_PermuteLayerY)
    {
        m_PermuteLayerY->run();
    }
    m_GEMMLayer->run();
}

---

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

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