NeonBatchMatMulWorkload Class Reference

#include <NeonBatchMatMulWorkload.hpp>

Inheritance diagram for NeonBatchMatMulWorkload:

Public Member Functions
	NeonBatchMatMulWorkload (const BatchMatMulQueueDescriptor &descriptor, const WorkloadInfo &info, const bool isFastMathEnabled)
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 arm::pipe::ProfilingGuid	GetGuid () const =0
virtual bool	SupportsTensorHandleReplacement () const =0
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,
		const bool	isFastMathEnabled
	)

Definition at line 55 of file NeonBatchMatMulWorkload.cpp.

    : 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");
    }
 
    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();
    arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
 
    // GeMM dispatches kernel handles dynamic inputs differently to static so this flag needs to be set
    inputX.info()->set_are_values_constant(false);
    inputY.info()->set_are_values_constant(false);
 
    const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
 
    arm_compute::MatMulInfo matMulInfo;
    matMulInfo.adj_lhs(descriptor.m_Parameters.m_TransposeX);
    matMulInfo.adj_rhs(descriptor.m_Parameters.m_TransposeY);
    matMulInfo.fused_activation(activationInfo);
 
    arm_compute::CpuMatMulSettings settings;
    settings.fast_math(isFastMathEnabled);
 
    m_MatMulLayer.configure(&inputX, &inputY, &output, matMulInfo, settings);
 
    // Report Profiling Details
    WorkloadInfo detailsInfo;
    detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
    detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
    ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonBatchMatMulWorkload_Construct",
                                         descriptor.m_Parameters,
                                         detailsInfo,
                                         GetGuid());
}

References ARMNN_REPORT_PROFILING_WORKLOAD_DESC, armnn::ConvertAdditionalInfoToAclActivationLayerInfo(), BaseWorkload< BatchMatMulQueueDescriptor >::GetGuid(), armnn::info, BatchMatMulDescriptor::m_AdjointX, BatchMatMulDescriptor::m_AdjointY, BaseWorkload< BatchMatMulQueueDescriptor >::m_Data, BatchMatMulDescriptor::m_DataLayoutX, BatchMatMulDescriptor::m_DataLayoutY, QueueDescriptor::m_Inputs, WorkloadInfo::m_InputTensorInfos, QueueDescriptor::m_Outputs, WorkloadInfo::m_OutputTensorInfos, QueueDescriptorWithParameters< LayerDescriptor >::m_Parameters, BatchMatMulDescriptor::m_TransposeX, BatchMatMulDescriptor::m_TransposeY, armnn::NCHW, and QueueDescriptor::ValidateInputsOutputs().

Member Function Documentation

Execute()

void Execute ( ) const

overridevirtual

Implements IWorkload.

Definition at line 102 of file NeonBatchMatMulWorkload.cpp.

{
    ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonBatchMatMulWorkload_Execute", this->GetGuid());
    m_MatMulLayer.run();
}

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

---

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

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