NeonFusedWorkload Class Reference

#include <NeonFusedWorkload.hpp>

Inheritance diagram for NeonFusedWorkload:

Collaboration diagram for NeonFusedWorkload:

Public Member Functions
	NeonFusedWorkload (const FusedQueueDescriptor &descriptor, const WorkloadInfo &info)
virtual void	Execute () const override
Public Member Functions inherited from NeonBaseWorkload< FusedQueueDescriptor >
	NeonBaseWorkload (const FusedQueueDescriptor &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< FusedQueueDescriptor >
	BaseWorkload (const FusedQueueDescriptor &descriptor, const WorkloadInfo &info)
virtual const std::string &	GetName () const override
void	ExecuteAsync (ExecutionData &executionData) override
void	PostAllocationConfigure () override
const FusedQueueDescriptor &	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 const std::string &	GetName () const =0
virtual void	RegisterDebugCallback (const DebugCallbackFunction &)
virtual armnn::Optional< armnn::MemoryRequirements >	GetMemoryRequirements ()

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

Detailed Description

Definition at line 22 of file NeonFusedWorkload.hpp.

Constructor & Destructor Documentation

NeonFusedWorkload()

NeonFusedWorkload	(	const FusedQueueDescriptor &	descriptor,
		const WorkloadInfo &	info
	)

Definition at line 63 of file NeonFusedWorkload.cpp.

    : NeonBaseWorkload<FusedQueueDescriptor>(descriptor, info)
{
    m_Data.ValidateInputsOutputs("NeonFusedWorkload",
                                 static_cast<unsigned int>(info.m_InputTensorInfos.size()),
                                 static_cast<unsigned int>(info.m_OutputTensorInfos.size()));
 
    std::vector<arm_compute::ITensor*> inputs;
    inputs.reserve(info.m_InputTensorInfos.size());
    for (auto input : m_Data.m_Inputs)
    {
        inputs.emplace_back(&PolymorphicDowncast<IAclTensorHandle*>(input)->GetTensor());
    }
 
    std::vector<arm_compute::ITensor*> outputs;
    outputs.reserve(info.m_OutputTensorInfos.size());
    for (auto output : m_Data.m_Outputs)
    {
        outputs.emplace_back(&PolymorphicDowncast<IAclTensorHandle*>(output)->GetTensor());
    }
 
    const arm_compute::ActivationLayerInfo activationInfo =
            ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
 
    switch (descriptor.m_Parameters.m_FusedKernelType)
    {
        case FusedKernelType::AddMulAdd:
        {
            auto layer = std::make_unique<arm_compute::NEAddMulAdd>();
            layer->configure(inputs[0],
                             inputs[1],
                             inputs[2],  // bn_mul
                             inputs[3],  // bn_add
                             outputs.size() == 1 ? nullptr : outputs[0], // add_output
                             outputs.size() == 1 ? outputs[0] : outputs[1], // final_output
                             arm_compute::ConvertPolicy::SATURATE,
                             activationInfo);
            m_FusedLayer.reset(layer.release());
            break;
        }
        default:
            throw Exception("NeonFusedWorkload: no valid kernel type.");
    }
}

References armnn::AddMulAdd, armnn::ConvertAdditionalInfoToAclActivationLayerInfo(), armnn::info, BaseWorkload< FusedQueueDescriptor >::m_Data, FusedDescriptor::m_FusedKernelType, QueueDescriptor::m_Inputs, QueueDescriptor::m_Outputs, QueueDescriptorWithParameters< LayerDescriptor >::m_Parameters, and QueueDescriptor::ValidateInputsOutputs().

Member Function Documentation

Execute()

void Execute ( ) const

overridevirtual

Implements IWorkload.

Definition at line 108 of file NeonFusedWorkload.cpp.

{
    ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonFusedWorkload_Execute", this->GetGuid());
    m_FusedLayer->run();
}

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

---

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

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