DepthwiseConvolution2dQueueDescriptor Struct Reference

Depthwise Convolution 2D layer workload data. More...

#include <WorkloadData.hpp>

Inheritance diagram for DepthwiseConvolution2dQueueDescriptor:

Collaboration diagram for DepthwiseConvolution2dQueueDescriptor:

Public Member Functions
void	Validate (const WorkloadInfo &workloadInfo) const
Public Member Functions inherited from QueueDescriptorWithParameters< DepthwiseConvolution2dDescriptor >
virtual	~QueueDescriptorWithParameters ()=default
Public Member Functions inherited from QueueDescriptor
virtual	~QueueDescriptor ()=default
void	ValidateTensorNumDimensions (const TensorInfo &tensor, std::string const &descName, unsigned int numDimensions, std::string const &tensorName) const
void	ValidateTensorNumDimNumElem (const TensorInfo &tensorInfo, unsigned int numDimension, unsigned int numElements, std::string const &tensorName) const
void	ValidateInputsOutputs (const std::string &descName, unsigned int numExpectedIn, unsigned int numExpectedOut) const
template<typename T>
const T *	GetAdditionalInformation () const

Additional Inherited Members
Public Attributes inherited from QueueDescriptorWithParameters< DepthwiseConvolution2dDescriptor >
DepthwiseConvolution2dDescriptor	m_Parameters
Public Attributes inherited from QueueDescriptor
std::vector< ITensorHandle * >	m_Inputs
std::vector< ITensorHandle * >	m_Outputs
void *	m_AdditionalInfoObject
bool	m_AllowExpandedDims = false
Protected Member Functions inherited from QueueDescriptorWithParameters< DepthwiseConvolution2dDescriptor >
	QueueDescriptorWithParameters ()=default
QueueDescriptorWithParameters &	operator= (QueueDescriptorWithParameters const &)=default
Protected Member Functions inherited from QueueDescriptor
	QueueDescriptor ()
	QueueDescriptor (QueueDescriptor const &)=default
QueueDescriptor &	operator= (QueueDescriptor const &)=default

Detailed Description

Depthwise Convolution 2D layer workload data.

Note

The weights are in the format [1, H, W, I*M]. Where I is the input channel size, M the depthwise mutliplier and H, W is the height and width of the filter kernel. If per channel quantization is applied the weights will be quantized along the last dimension/axis (I*M) which corresponds to the output channel size. If per channel quantization is applied the weights tensor will have I*M scales, one for each dimension of the quantization axis. You have to be aware of this when reshaping the weights tensor. Splitting the I*M axis, e.g. [1, H, W, I*M] --> [H, W, I, M], won't work without taking care of the corresponding quantization scales. If there is no per channel quantization applied reshaping the weights tensor won't cause any issues. There are preconfigured permutation functions available here.

Definition at line 234 of file WorkloadData.hpp.

Member Function Documentation

Validate()

void Validate

(

const WorkloadInfo &

workloadInfo

)

const

Definition at line 1392 of file WorkloadData.cpp.

{
    const std::string descriptorName{"DepthwiseConvolution2dQueueDescriptor"};
 
    uint32_t numInputs = 2;
    if (m_Parameters.m_BiasEnabled)
    {
        numInputs = 3;
    }
 
    ValidateNumInputs(workloadInfo,  descriptorName, numInputs);
    ValidateNumOutputs(workloadInfo, descriptorName, 1);
 
    const TensorInfo& inputTensorInfo  = workloadInfo.m_InputTensorInfos[0];
    const TensorInfo& outputTensorInfo = workloadInfo.m_OutputTensorInfos[0];
 
    ValidateTensorNumDimensions(inputTensorInfo,  descriptorName, 4, "input");
    ValidateTensorNumDimensions(outputTensorInfo, descriptorName, 4, "output");
 
    const TensorInfo& weightTensorInfo = workloadInfo.m_InputTensorInfos[1];
    ValidateTensorNumDimensions(weightTensorInfo, descriptorName, 4, "weight");
 
    if (m_Parameters.m_DilationX < 1 || m_Parameters.m_DilationY < 1 )
    {
        throw InvalidArgumentException(
            fmt::format("{}: dilationX (provided {}) and dilationY (provided {}) "
                        "cannot be smaller than 1.",
                        descriptorName, m_Parameters.m_DilationX, m_Parameters.m_DilationX));
    }
 
    if (m_Parameters.m_StrideX <= 0 || m_Parameters.m_StrideY <= 0  )
    {
        throw InvalidArgumentException(
            fmt::format("{}: strideX (provided {}) and strideY (provided {}) "
                        "cannot be either negative or 0.",
                        descriptorName, m_Parameters.m_StrideX, m_Parameters.m_StrideY));
    }
 
    if (weightTensorInfo.GetShape()[0] != 1)
    {
        throw InvalidArgumentException(fmt::format(
                "{0}: The weight format in armnn is expected to be [1, H, W, Cout]."
                "But first dimension is not equal to 1. Provided weight shape: [{1}, {2}, {3}, {4}]",
                descriptorName,
                weightTensorInfo.GetShape()[0],
                weightTensorInfo.GetShape()[1],
                weightTensorInfo.GetShape()[2],
                weightTensorInfo.GetShape()[3]));
    }
 
    const unsigned int channelIndex = (m_Parameters.m_DataLayout == DataLayout::NCHW) ? 1 : 3;
    const unsigned int numWeightOutputChannelsRefFormat = weightTensorInfo.GetShape()[3];
    const unsigned int numWeightOutputChannelsAclFormat = weightTensorInfo.GetShape()[1];
    const unsigned int numOutputChannels = outputTensorInfo.GetShape()[channelIndex];
 
    // Weights format has two valid options: [1, H, W, Cout] (CpuRef) or [1, Cout, H, W] (CpuAcc/GpuAcc).
    bool validRefFormat = (numWeightOutputChannelsRefFormat == numOutputChannels);
    bool validAclFormat = (numWeightOutputChannelsAclFormat == numOutputChannels);
 
    if (!(validRefFormat || validAclFormat))
    {
        throw InvalidArgumentException(fmt::format(
            "{0}: The weight format in armnn is expected to be [1, H, W, Cout] (CpuRef) or [1, Cout, H, W] "
            "(CpuAcc/GpuAcc). But neither the 4th (CpuRef) or 2nd (CpuAcc/GpuAcc) dimension is equal to Cout."
            "Cout = {1} Provided weight shape: [{2}, {3}, {4}, {5}]",
            descriptorName,
            numOutputChannels,
            weightTensorInfo.GetShape()[0],
            weightTensorInfo.GetShape()[1],
            weightTensorInfo.GetShape()[2],
            weightTensorInfo.GetShape()[3]));
    }
 
    ValidateWeightDataType(inputTensorInfo, weightTensorInfo, descriptorName);
 
    Optional<TensorInfo> optionalBiasTensorInfo;
    if (m_Parameters.m_BiasEnabled)
    {
        optionalBiasTensorInfo = MakeOptional<TensorInfo>(workloadInfo.m_InputTensorInfos[2]);
        const TensorInfo& biasTensorInfo = optionalBiasTensorInfo.value();
 
        ValidateBiasTensorQuantization(biasTensorInfo, weightTensorInfo, descriptorName);
        ValidateTensorDataType(biasTensorInfo, GetBiasDataType(inputTensorInfo.GetDataType()), descriptorName, "bias");
    }
    ValidatePerAxisQuantization(inputTensorInfo,
                                outputTensorInfo,
                                weightTensorInfo,
                                optionalBiasTensorInfo,
                                descriptorName);
 
    std::vector<DataType> supportedTypes =
    {
        DataType::BFloat16,
        DataType::Float16,
        DataType::Float32,
        DataType::QAsymmS8,
        DataType::QAsymmU8,
        DataType::QSymmS16
    };
 
    ValidateDataTypes(inputTensorInfo, supportedTypes, descriptorName);
    ValidateTensorDataTypesMatch(inputTensorInfo, outputTensorInfo, descriptorName, "input", "output");
}

References armnn::BFloat16, armnn::Float16, armnn::Float32, armnn::GetBiasDataType(), TensorInfo::GetDataType(), TensorInfo::GetShape(), WorkloadInfo::m_InputTensorInfos, WorkloadInfo::m_OutputTensorInfos, QueueDescriptorWithParameters< DepthwiseConvolution2dDescriptor >::m_Parameters, armnn::MakeOptional(), armnn::NCHW, armnn::QAsymmS8, armnn::QAsymmU8, armnn::QSymmS16, QueueDescriptor::ValidateTensorNumDimensions(), and OptionalReferenceSwitch< IsReference, T >::value().

---

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

• include/armnn/backends/WorkloadData.hpp
• src/backends/backendsCommon/WorkloadData.cpp