#include <ConvolutionLayer.h>

Collaboration diagram for onert::backend::cpu::ops::ConvolutionLayer:

[legend]

Public Member Functions
	ConvolutionLayer ()

	~ConvolutionLayer ()

void	configure (const IPortableTensor input, const IPortableTensor kernel, const IPortableTensor bias, ir::PaddingType _paddingType, const uint32_t paddingLeft, const uint32_t paddingRight, const uint32_t paddingTop, const uint32_t paddingBottom, const uint32_t strideWidth, const uint32_t strideHeight, const uint32_t dilationWidthFactor, const uint32_t dilationHeightFactor, const ir::Activation activation, IPortableTensor output, bool is_cachable_weights)

void	prepare () override

void	run () override

Public Member Functions inherited from onert::exec::IFunction
virtual	~IFunction ()=default

Protected Attributes
const IPortableTensor *	_input

const IPortableTensor *	_kernel

const IPortableTensor *	_bias

IPortableTensor *	_output

ir::PaddingType	_paddingType

uint32_t	_paddingLeft

uint32_t	_paddingTop

uint32_t	_paddingRight

uint32_t	_paddingBottom

uint32_t	_strideWidth

uint32_t	_strideHeight

uint32_t	_dilationWidthFactor

uint32_t	_dilationHeightFactor

ir::Activation	_activation

std::unique_ptr< nnfw::cker::Conv >	_conv_kernel

std::unique_ptr< nnfw::cker::ConvHybridTempArena >	_hybrid_arena

bool	_prepare

bool	_is_cachable_weights

bool	_is_hybrid

Detailed Description

Definition at line 37 of file ConvolutionLayer.h.

Constructor & Destructor Documentation

◆ ConvolutionLayer()

onert::backend::cpu::ops::ConvolutionLayer::ConvolutionLayer ( )

Definition at line 27 of file ConvolutionLayer.cc.

  : _input(nullptr), _kernel(nullptr), _bias(nullptr), _output(nullptr),
    _paddingType(ir::PaddingType::EXPLICIT), _paddingLeft(0), _paddingTop(0), _paddingRight(0),
    _paddingBottom(0), _strideWidth(0), _strideHeight(0), _dilationWidthFactor(1),
    _dilationHeightFactor(1), _activation(ir::Activation::NONE),
    _conv_kernel(new nnfw::cker::Conv()), _prepare(false), _is_cachable_weights(false),
    _is_hybrid(false)
{
  // DO NOTHING
}

◆ ~ConvolutionLayer()

onert::backend::cpu::ops::ConvolutionLayer::~ConvolutionLayer ( )

default

Member Function Documentation

◆ configure()

void onert::backend::cpu::ops::ConvolutionLayer::configure	(	const IPortableTensor *	input,
		const IPortableTensor *	kernel,
		const IPortableTensor *	bias,
		ir::PaddingType	_paddingType,
		const uint32_t	paddingLeft,
		const uint32_t	paddingRight,
		const uint32_t	paddingTop,
		const uint32_t	paddingBottom,
		const uint32_t	strideWidth,
		const uint32_t	strideHeight,
		const uint32_t	dilationWidthFactor,
		const uint32_t	dilationHeightFactor,
		const ir::Activation	activation,
		IPortableTensor *	output,
		bool	is_cachable_weights
	)

Definition at line 191 of file ConvolutionLayer.cc.

{
  _input = input;
  _kernel = kernel;
  _bias = bias;
  _paddingType = paddingType;
  _paddingLeft = paddingLeft;
  _paddingRight = paddingRight;
  _paddingTop = paddingTop;
  _paddingBottom = paddingBottom;
  _strideWidth = strideWidth;
  _strideHeight = strideHeight;
  _dilationWidthFactor = dilationWidthFactor;
  _dilationHeightFactor = dilationHeightFactor;
  _activation = activation;
  _output = output;
  _is_cachable_weights = is_cachable_weights;
  _is_hybrid = _input->data_type() == OperandType::FLOAT32 &&
               _kernel->data_type() == OperandType::QUANT_INT8_SYMM;
}

References _activation, _bias, _dilationHeightFactor, _dilationWidthFactor, _input, _is_cachable_weights, _is_hybrid, _kernel, _output, _paddingBottom, _paddingLeft, _paddingRight, _paddingTop, _paddingType, _strideHeight, _strideWidth, and onert::backend::IPortableTensor::data_type().

◆ prepare()

void onert::backend::cpu::ops::ConvolutionLayer::prepare ( )

overridevirtual

Reimplemented from onert::exec::IFunction.

Definition at line 278 of file ConvolutionLayer.cc.

{
  if (_prepare)
    return;
 
  if (_is_hybrid)
  {
    // ensure weight is per-channel quantized.
    int32_t kernel_output_channel = getShape(_kernel).Dims(0);
    // zero_points comes from flatbuffer vector. Its size is within uint32_t range.
    size_t kernel_zerop_cnt = _kernel->data_scales().size();
    // promote to int64_t to compare int32_t and uint32_t
    if ((int64_t)kernel_output_channel != (int64_t)kernel_zerop_cnt)
      throw std::runtime_error{"Conv2D hybrid supports only per-channel quantized weight."};
 
    // allocate memory for activation quantization.
    // - quantized values (int8_t type and same shape of original input)
    // - quantization params (= scale/zeropoint for each input)
    auto input_shape = getShape(_input);
    const int batch_size = input_shape.Dims(0);
    const int input_size = input_shape.FlatSize() / batch_size;
    _hybrid_arena = std::make_unique<nnfw::cker::ConvHybridTempArena>(batch_size, input_size);
    _prepare = true;
    return;
  }
 
  nnfw::cker::Conv &kernel = *_conv_kernel;
  if (_input->data_type() == OperandType::FLOAT32 && _is_cachable_weights)
  {
    bool is_transposed = false;
    kernel.prepareF32(getShape(_kernel), getBuffer<float>(_kernel), getPaddingType(_paddingType),
                      is_transposed, _dilationWidthFactor, _dilationHeightFactor);
 
    // Decrease reference of _kernel(weights) only when _kernel is constant
    if (is_transposed)
    {
      auto kernel_tensor = dynamic_cast<const Tensor *>(_kernel);
      if (kernel_tensor)
        // TODO Remove const_cast
        const_cast<Tensor *>(kernel_tensor)->decrease_ref();
    }
  }
  else if (_input->data_type() == OperandType::QUANT_UINT8_ASYMM && _is_cachable_weights &&
           !_input->is_dynamic() && !_output->is_dynamic())
  {
    const bool per_channel_quantized = _kernel->data_scales().size() > 1;
    if (per_channel_quantized)
    {
      GetQuantizedConvolutionMultipliersAndShifts(
        _input->data_scale(), _output->data_scale(), _kernel->data_scales().data(),
        _kernel->data_scales().size(), getShape(_kernel).Dims(0),
        kernel.per_channel_output_multiplier(), kernel.per_channel_output_shift());
    }
    else
    {
      kernel.prepareQ8uPerTensor(getShape(_input), getShape(_kernel), getShape(_output),
                                 _strideWidth, _strideHeight, _dilationWidthFactor,
                                 _dilationHeightFactor);
    }
  }
  else if (_input->data_type() == OperandType::QUANT_INT8_ASYMM)
  {
    if (_is_cachable_weights && !_input->is_dynamic() && !_output->is_dynamic())
    {
      GetQuantizedConvolutionMultipliersAndShifts(
        _input->data_scale(), _output->data_scale(), _kernel->data_scales().data(),
        _kernel->data_scales().size(), getShape(_kernel).Dims(0),
        kernel.per_channel_output_multiplier(), kernel.per_channel_output_shift());
    }
    else
    {
      throw std::runtime_error{"Conv2D: Int8 dynamic weight is not supported"};
    }
  }
  _prepare = true;
}

References _conv_kernel, _dilationHeightFactor, _dilationWidthFactor, _hybrid_arena, _input, _is_cachable_weights, _is_hybrid, _kernel, _output, _paddingType, _prepare, _strideHeight, _strideWidth, onert::backend::IPortableTensor::data_scale(), onert::backend::IPortableTensor::data_scales(), onert::backend::IPortableTensor::data_type(), nnfw::cker::Shape::Dims(), onert::backend::cpu::ops::getPaddingType(), onert::backend::cpu::ops::GetQuantizedConvolutionMultipliersAndShifts(), onert::backend::cpu::ops::getShape(), onert::backend::IPortableTensor::is_dynamic(), nnfw::cker::Conv::per_channel_output_multiplier(), nnfw::cker::Conv::per_channel_output_shift(), nnfw::cker::Conv::prepareF32(), and nnfw::cker::Conv::prepareQ8uPerTensor().

Referenced by run().

◆ run()

void onert::backend::cpu::ops::ConvolutionLayer::run ( )

overridevirtual

Implements onert::exec::IFunction.

Definition at line 220 of file ConvolutionLayer.cc.

{
  prepare();
  if (_input->is_dynamic() || _kernel->is_dynamic())
  {
    const auto ifm_shape = _input->getShape().asFeature();
    const auto ofm_shape = _output->getShape().asFeature();
    // Kernel format is [depth_out, kernel_height, kernel_width, depth_in].
    const auto ker_shape = _kernel->getShape();
    const auto ker_height = ker_shape.dim(1);
    const auto ker_width = ker_shape.dim(2);
 
    ir::Stride stride;
    stride.vertical = _strideWidth;
    stride.horizontal = _strideWidth;
 
    ir::Padding param_padding;
    param_padding.type = _paddingType;
    param_padding.param.left = _paddingLeft;
    param_padding.param.right = _paddingRight;
    param_padding.param.top = _paddingTop;
    param_padding.param.bottom = _paddingBottom;
 
    const auto padding =
      ir::calculatePadding(param_padding, ifm_shape, ofm_shape, stride, ker_width, ker_height,
                           _dilationWidthFactor, _dilationHeightFactor);
 
    _paddingLeft = padding.left;
    _paddingRight = padding.right;
    _paddingTop = padding.top;
    _paddingBottom = padding.bottom;
  }
  if (_is_hybrid)
  {
    convQ8iHybridPerChannel();
  }
  else if (_input->data_type() == OperandType::FLOAT32)
  {
    convFloat32();
  }
  else if (_input->data_type() == OperandType::QUANT_UINT8_ASYMM)
  {
    const bool per_channel_quantized = _kernel->data_scales().size() > 1;
    if (per_channel_quantized)
      convQ8uPerChannel();
    else
      convQ8uPerTensor();
  }
  else if (_input->data_type() == OperandType::QUANT_INT8_ASYMM)
  {
    convQ8i();
  }
  else
  {
    throw std::runtime_error{"Conv: unsupported data type"};
  }
}

References _dilationHeightFactor, _dilationWidthFactor, _input, _is_hybrid, _kernel, _output, _paddingBottom, _paddingLeft, _paddingRight, _paddingTop, _paddingType, _strideWidth, onert::ir::ExplicitPadding::bottom, onert::ir::calculatePadding(), onert::backend::IPortableTensor::data_scales(), onert::backend::IPortableTensor::data_type(), onert::backend::IPortableTensor::getShape(), onert::ir::Stride::horizontal, onert::backend::IPortableTensor::is_dynamic(), onert::ir::ExplicitPadding::left, onert::ir::Padding::param, prepare(), onert::ir::ExplicitPadding::right, onert::ir::ExplicitPadding::top, onert::ir::Padding::type, and onert::ir::Stride::vertical.

Referenced by onert::backend::train::ops::ConvolutionLayer::forward().

Field Documentation

◆ _activation

ir::Activation onert::backend::cpu::ops::ConvolutionLayer::_activation

protected

Definition at line 78 of file ConvolutionLayer.h.

Referenced by configure().

◆ _bias

const IPortableTensor* onert::backend::cpu::ops::ConvolutionLayer::_bias

protected

Definition at line 64 of file ConvolutionLayer.h.

Referenced by configure().

◆ _conv_kernel

std::unique_ptr<nnfw::cker::Conv> onert::backend::cpu::ops::ConvolutionLayer::_conv_kernel

protected

Definition at line 80 of file ConvolutionLayer.h.

Referenced by prepare().

◆ _dilationHeightFactor

uint32_t onert::backend::cpu::ops::ConvolutionLayer::_dilationHeightFactor

protected

Definition at line 76 of file ConvolutionLayer.h.

Referenced by configure(), onert::backend::train::ops::ConvolutionLayer::configureBackward(), prepare(), and run().

◆ _dilationWidthFactor

uint32_t onert::backend::cpu::ops::ConvolutionLayer::_dilationWidthFactor

protected

Definition at line 75 of file ConvolutionLayer.h.

Referenced by configure(), onert::backend::train::ops::ConvolutionLayer::configureBackward(), prepare(), and run().

◆ _hybrid_arena

std::unique_ptr<nnfw::cker::ConvHybridTempArena> onert::backend::cpu::ops::ConvolutionLayer::_hybrid_arena

protected

Definition at line 81 of file ConvolutionLayer.h.

Referenced by prepare().

◆ _input

const IPortableTensor* onert::backend::cpu::ops::ConvolutionLayer::_input

protected

Definition at line 62 of file ConvolutionLayer.h.

Referenced by onert::backend::train::ops::ConvolutionLayer::backward(), configure(), prepare(), and run().

◆ _is_cachable_weights

bool onert::backend::cpu::ops::ConvolutionLayer::_is_cachable_weights

protected

Definition at line 84 of file ConvolutionLayer.h.

Referenced by configure(), and prepare().

◆ _is_hybrid

bool onert::backend::cpu::ops::ConvolutionLayer::_is_hybrid

protected

Definition at line 85 of file ConvolutionLayer.h.

Referenced by configure(), prepare(), and run().

◆ _kernel

const IPortableTensor* onert::backend::cpu::ops::ConvolutionLayer::_kernel

protected

Definition at line 63 of file ConvolutionLayer.h.

Referenced by configure(), prepare(), and run().

◆ _output

IPortableTensor* onert::backend::cpu::ops::ConvolutionLayer::_output

protected

Definition at line 65 of file ConvolutionLayer.h.

Referenced by configure(), prepare(), and run().

◆ _paddingBottom

uint32_t onert::backend::cpu::ops::ConvolutionLayer::_paddingBottom

protected

Definition at line 71 of file ConvolutionLayer.h.

Referenced by configure(), and run().

◆ _paddingLeft

uint32_t onert::backend::cpu::ops::ConvolutionLayer::_paddingLeft

protected

Definition at line 68 of file ConvolutionLayer.h.

Referenced by configure(), and run().

◆ _paddingRight

uint32_t onert::backend::cpu::ops::ConvolutionLayer::_paddingRight

protected

Definition at line 70 of file ConvolutionLayer.h.

Referenced by configure(), and run().

◆ _paddingTop

uint32_t onert::backend::cpu::ops::ConvolutionLayer::_paddingTop

protected

Definition at line 69 of file ConvolutionLayer.h.

Referenced by configure(), and run().

◆ _paddingType

ir::PaddingType onert::backend::cpu::ops::ConvolutionLayer::_paddingType

protected

Definition at line 67 of file ConvolutionLayer.h.

Referenced by configure(), prepare(), and run().

◆ _prepare

bool onert::backend::cpu::ops::ConvolutionLayer::_prepare

protected

Definition at line 83 of file ConvolutionLayer.h.

Referenced by prepare().

◆ _strideHeight

uint32_t onert::backend::cpu::ops::ConvolutionLayer::_strideHeight

protected

Definition at line 74 of file ConvolutionLayer.h.

Referenced by configure(), and prepare().

◆ _strideWidth

uint32_t onert::backend::cpu::ops::ConvolutionLayer::_strideWidth

protected

Definition at line 73 of file ConvolutionLayer.h.

Referenced by configure(), prepare(), and run().

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

runtime/onert/backend/cpu/ops/ConvolutionLayer.h
runtime/onert/backend/cpu/ops/ConvolutionLayer.cc

Public Member Functions

Protected Attributes

Detailed Description

Constructor & Destructor Documentation

◆ ConvolutionLayer()

◆ ~ConvolutionLayer()

Member Function Documentation

◆ configure()

◆ prepare()

◆ run()

Field Documentation

◆ _activation

◆ _bias

◆ _conv_kernel

◆ _dilationHeightFactor

◆ _dilationWidthFactor

◆ _hybrid_arena

◆ _input

◆ _is_cachable_weights

◆ _is_hybrid

◆ _kernel

◆ _output

◆ _paddingBottom

◆ _paddingLeft

◆ _paddingRight

◆ _paddingTop

◆ _paddingType

◆ _prepare

◆ _strideHeight

◆ _strideWidth