ClBatchMatMulWorkload Class Reference

#include <ClBatchMatMulWorkload.hpp>

Inheritance 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< 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 ClBaseWorkload< BatchMatMulQueueDescriptor >
virtual void	Reconfigure ()
Protected Attributes inherited from BaseWorkload< BatchMatMulQueueDescriptor >
BatchMatMulQueueDescriptor	m_Data
const arm::pipe::ProfilingGuid	m_Guid

Detailed Description

Definition at line 21 of file ClBatchMatMulWorkload.hpp.

Constructor & Destructor Documentation

ClBatchMatMulWorkload()

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

Definition at line 115 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);
 
    const arm_compute::ICLTensor& inputX = PolymorphicDowncast<ClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
    const arm_compute::ICLTensor& inputY = PolymorphicDowncast<ClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
    arm_compute::ICLTensor& output = PolymorphicDowncast<ClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
 
    inputX.info()->set_data_layout(armcomputetensorutils::ConvertDataLayout(m_Data.m_Parameters.m_DataLayoutX));
    arm_compute::TensorShape inputXTensorInfo = armcomputetensorutils::BuildArmComputeTensorShape(
            info.m_InputTensorInfos[0].GetShape(), 3);
    inputX.info()->set_tensor_shape(inputXTensorInfo);
    inputY.info()->set_data_layout(armcomputetensorutils::ConvertDataLayout(m_Data.m_Parameters.m_DataLayoutY));
    arm_compute::TensorShape inputYTensorInfo = armcomputetensorutils::BuildArmComputeTensorShape(
            info.m_InputTensorInfos[1].GetShape(), 3);
    inputY.info()->set_tensor_shape(inputYTensorInfo);
 
    arm_compute::TensorInfo aclPermutedXInfo = arm_compute::TensorInfo();
    arm_compute::TensorInfo aclPermutedYInfo = arm_compute::TensorInfo();
 
    if (descriptor.m_Parameters.m_TransposeX == true)
    {
        armnn::TensorInfo strippedInfo = armnnUtils::ReduceDims(info.m_InputTensorInfos[0], 3);
 
        armnn::PermutationVector permutationXVector
                = GeneratePermutationVectorOnLastTwoDimensions(strippedInfo.GetNumDimensions());
        const TensorInfo permutedXInfo = armnnUtils::Permuted(strippedInfo, permutationXVector);
        const auto aclPermutationXVector = armcomputetensorutils::BuildArmComputePermutationVector(permutationXVector);
        armcomputetensorutils::BuildArmComputeTensor(m_PermutedTensorX, permutedXInfo);
        armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_PermutedTensorX);
 
        auto permuteLayerX = std::make_unique<arm_compute::CLPermute>();
        permuteLayerX->configure(clCompileContext,
                                 &inputX,
                                 &m_PermutedTensorX,
                                 aclPermutationXVector);
        m_PermuteLayerX.reset(permuteLayerX.release());
    }
 
    if (descriptor.m_Parameters.m_TransposeY == true)
    {
        armnn::TensorInfo strippedInfo = armnnUtils::ReduceDims(info.m_InputTensorInfos[1], 3);
 
        armnn::PermutationVector permutationYVector
                = GeneratePermutationVectorOnLastTwoDimensions(strippedInfo.GetNumDimensions());
        const TensorInfo permutedYInfo = armnnUtils::Permuted(strippedInfo, permutationYVector);
        const auto aclPermutationYVector = armcomputetensorutils::BuildArmComputePermutationVector(permutationYVector);
        armcomputetensorutils::BuildArmComputeTensor(m_PermutedTensorY, permutedYInfo);
        armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_PermutedTensorY);
 
        auto permuteLayerY = std::make_unique<arm_compute::CLPermute>();
        permuteLayerY->configure(clCompileContext,
                                 &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::CLGEMM>();
    gemmLayer->configure(clCompileContext,
                         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());
}

References ARMNN_REPORT_PROFILING_WORKLOAD_DESC, armnn::info, 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().

Member Function Documentation

Execute()

void Execute ( ) const

overridevirtual

Implements IWorkload.

Definition at line 207 of file ClBatchMatMulWorkload.cpp.

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

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

---

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

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