ClBatchMatMulWorkload Class Reference

#include <ClBatchMatMulWorkload.hpp>

Inheritance diagram for ClBatchMatMulWorkload:

Collaboration diagram for ClBatchMatMulWorkload:

Public Member Functions
	ClBatchMatMulWorkload (const BatchMatMulQueueDescriptor &descriptor, const WorkloadInfo &info, const arm_compute::CLCompileContext &clCompileContext)
virtual void	Execute () const override
Public Member Functions inherited from ClBaseWorkload< BatchMatMulQueueDescriptor >
	ClBaseWorkload (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< 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 ClBaseWorkload< BatchMatMulQueueDescriptor >
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 20 of file ClBatchMatMulWorkload.hpp.

Constructor & Destructor Documentation

ClBatchMatMulWorkload()

ClBatchMatMulWorkload	(	const BatchMatMulQueueDescriptor &	descriptor,
		const WorkloadInfo &	info,
		const arm_compute::CLCompileContext &	clCompileContext )

Definition at line 57 of file ClBatchMatMulWorkload.cpp.

    : ClBaseWorkload<BatchMatMulQueueDescriptor>(descriptor, info)
{
    // Report Profiling Details
    ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClBatchMatMulWorkload_Construct",
                                         descriptor.m_Parameters,
                                         info,
                                         this->GetGuid());
 
    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("ClBatchMatMulWorkload", 2, 1);
 
    arm_compute::ICLTensor& inputX = PolymorphicDowncast<ClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
    arm_compute::ICLTensor& inputY = PolymorphicDowncast<ClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
    auto outputHandle = PolymorphicDowncast<ClTensorHandle*>(m_Data.m_Outputs[0]);
    arm_compute::ICLTensor& output = outputHandle->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);
 
    arm_compute::GpuMatMulSettings settings;
 
    m_MatMulLayer.configure(clCompileContext, &inputX, &inputY, &output, matMulInfo, settings, activationInfo);
 
    // Report Profiling Details
    WorkloadInfo detailsInfo;
    detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
    detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
    ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClBatchMatMulWorkload_Construct",
                                         descriptor.m_Parameters,
                                         detailsInfo,
                                         GetGuid());
}

References ARMNN_REPORT_PROFILING_WORKLOAD_DESC, armnn::ConvertAdditionalInfoToAclActivationLayerInfo(), BaseWorkload< QueueDescriptor >::GetGuid(), armnn::info, BatchMatMulDescriptor::m_AdjointX, BatchMatMulDescriptor::m_AdjointY, BaseWorkload< QueueDescriptor >::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 109 of file ClBatchMatMulWorkload.cpp.

{
    ARMNN_SCOPED_PROFILING_EVENT_CL_NAME_GUID("ClBatchMatMulWorkload_Execute");
    RunClFunction(m_MatMulLayer, CHECK_LOCATION());
}

References ARMNN_SCOPED_PROFILING_EVENT_CL_NAME_GUID, CHECK_LOCATION, and armnn::RunClFunction().

---

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

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