src/armnn/layers/Convolution2dLayer.cpp

//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "Convolution2dLayer.hpp"

#include "LayerCloneBase.hpp"

#include <armnn/TypesUtils.hpp>
#include <backendsCommon/CpuTensorHandle.hpp>
#include <backendsCommon/WorkloadFactory.hpp>
#include <string>
#include <DataLayoutIndexed.hpp>

using namespace armnnUtils;

namespace armnn
{

Convolution2dLayer::Convolution2dLayer(const Convolution2dDescriptor& param, const char* name)
    : LayerWithParameters(1, 1, LayerType::Convolution2d, param, name)
{

}

void Convolution2dLayer::SerializeLayerParameters(ParameterStringifyFunction& fn) const
{
    //using DescriptorType = Parameters;
    const std::vector<TensorShape>& inputShapes =
    {
        GetInputSlot(0).GetConnection()->GetTensorInfo().GetShape(),
        m_Weight->GetTensorInfo().GetShape()
    };
    const TensorShape filterShape = inputShapes[1];
    DataLayoutIndexed dataLayoutIndex(m_Param.m_DataLayout);
    unsigned int filterWidth = filterShape[dataLayoutIndex.GetWidthIndex()];
    unsigned int filterHeight = filterShape[dataLayoutIndex.GetHeightIndex()];
    unsigned int outChannels = filterShape[0];

    fn("OutputChannels",std::to_string(outChannels));
    fn("FilterWidth",std::to_string(filterWidth));
    fn("FilterHeight",std::to_string(filterHeight));
    LayerWithParameters<Convolution2dDescriptor>::SerializeLayerParameters(fn);
}

std::unique_ptr<IWorkload> Convolution2dLayer::CreateWorkload(const Graph& graph, const IWorkloadFactory& factory) const
{
    // on this level constant data should not be released..
    BOOST_ASSERT_MSG(m_Weight != nullptr, "Convolution2dLayer: Weights data should not be null.");

    Convolution2dQueueDescriptor descriptor;

    descriptor.m_Weight = m_Weight.get();

    if (m_Param.m_BiasEnabled)
    {
        BOOST_ASSERT_MSG(m_Bias != nullptr, "Convolution2dLayer: Bias data should not be null.");
        descriptor.m_Bias = m_Bias.get();
    }
    return factory.CreateConvolution2d(descriptor, PrepInfoAndDesc(descriptor, graph));
}

Convolution2dLayer* Convolution2dLayer::Clone(Graph& graph) const
{
    auto layer = CloneBase<Convolution2dLayer>(graph, m_Param, GetName());

    layer->m_Weight = m_Weight ? std::make_unique<ScopedCpuTensorHandle>(*m_Weight) : nullptr;

    if (layer->m_Param.m_BiasEnabled)
    {
        layer->m_Bias = m_Bias ? std::make_unique<ScopedCpuTensorHandle>(*m_Bias) : nullptr;
    }

    return std::move(layer);
}

std::vector<TensorShape> Convolution2dLayer::InferOutputShapes(const std::vector<TensorShape>& inputShapes) const
{
    BOOST_ASSERT(inputShapes.size() == 2);
    const TensorShape& inputShape = inputShapes[0];
    const TensorShape filterShape = inputShapes[1];

    // If we support multiple batch dimensions in the future, then this assert will need to change.
    BOOST_ASSERT_MSG(inputShape.GetNumDimensions() == 4, "Convolutions will always have 4D input.");

    DataLayoutIndexed dataLayoutIndex(m_Param.m_DataLayout);

    unsigned int inWidth = inputShape[dataLayoutIndex.GetWidthIndex()];
    unsigned int inHeight = inputShape[dataLayoutIndex.GetHeightIndex()];
    unsigned int inBatchSize = inputShape[0];

    unsigned int filterWidth = filterShape[dataLayoutIndex.GetWidthIndex()];
    unsigned int dilatedFilterWidth = filterWidth + (m_Param.m_DilationX - 1) * (filterWidth - 1);
    unsigned int readWidth = (inWidth + m_Param.m_PadLeft + m_Param.m_PadRight) - dilatedFilterWidth;
    unsigned int outWidth = 1 + (readWidth / m_Param.m_StrideX);

    unsigned int filterHeight = filterShape[dataLayoutIndex.GetHeightIndex()];
    unsigned int dilatedFilterHeight = filterHeight + (m_Param.m_DilationY - 1) * (filterHeight - 1);
    unsigned int readHeight = (inHeight + m_Param.m_PadTop + m_Param.m_PadBottom) - dilatedFilterHeight;
    unsigned int outHeight = 1 + (readHeight / m_Param.m_StrideY);

    unsigned int outChannels = filterShape[0];
    unsigned int outBatchSize = inBatchSize;

    TensorShape tensorShape = m_Param.m_DataLayout == armnn::DataLayout::NHWC ?
        TensorShape( { outBatchSize, outHeight, outWidth, outChannels } ) :
        TensorShape( { outBatchSize, outChannels, outHeight, outWidth });

    return std::vector<TensorShape>({ tensorShape });
}

void Convolution2dLayer::ValidateTensorShapesFromInputs()
{
    VerifyLayerConnections(1, CHECK_LOCATION());

    // check if we m_Weight data is not nullptr
    BOOST_ASSERT_MSG(m_Weight != nullptr, "Convolution2dLayer: Weights data should not be null.");

    auto inferredShapes = InferOutputShapes({
        GetInputSlot(0).GetConnection()->GetTensorInfo().GetShape(),
        m_Weight->GetTensorInfo().GetShape() });

    BOOST_ASSERT(inferredShapes.size() == 1);

    ConditionalThrowIfNotEqual<LayerValidationException>(
        "Convolution2dLayer: TensorShape set on OutputSlot[0] does not match the inferred shape.",
        GetOutputSlot(0).GetTensorInfo().GetShape(),
        inferredShapes[0]);
}

Layer::ConstantTensors Convolution2dLayer::GetConstantTensorsByRef()
{
    return {m_Weight, m_Bias};
}

void Convolution2dLayer::Accept(ILayerVisitor& visitor) const
{
    ConstTensor weightsTensor(m_Weight->GetTensorInfo(), m_Weight->Map(true)) ;
    Optional<ConstTensor> optionalBiasTensor = EmptyOptional();

    if (GetParameters().m_BiasEnabled)
    {
        ConstTensor biasTensor(m_Bias->GetTensorInfo(), m_Bias->Map(true));
        optionalBiasTensor = Optional<ConstTensor>(biasTensor);
    }

    visitor.VisitConvolution2dLayer(this, GetParameters(), weightsTensor, optionalBiasTensor, GetName());
}

} // namespace armnn