onert::backend::xnnpack::ops::FullyConnectedLayer Class Reference

#include <FullyConnectedLayer.h>

Collaboration diagram for onert::backend::xnnpack::ops::FullyConnectedLayer:

Public Member Functions
	FullyConnectedLayer (const std::shared_ptr< ExternalContext > external_context)
void	configure (const IPortableTensor *input, const IPortableTensor *_kernel, const IPortableTensor *bias, ir::Activation activation, IPortableTensor *output)
void	run () override
bool	create () override
bool	setup () override
Public Member Functions inherited from onert::backend::xnnpack::ops::Layer
	Layer (const std::shared_ptr< ExternalContext > external_context)
	~Layer ()
void	prepare () override
Public Member Functions inherited from onert::exec::IFunction
virtual	~IFunction ()=default

Additional Inherited Members
Protected Attributes inherited from onert::backend::xnnpack::ops::Layer
xnn_operator_t	_kernel_op
bool	_create
bool	_setup
const std::shared_ptr< ExternalContext >	_external_context

Detailed Description

Definition at line 33 of file FullyConnectedLayer.h.

Constructor & Destructor Documentation

FullyConnectedLayer()

onert::backend::xnnpack::ops::FullyConnectedLayer::FullyConnectedLayer

(

const std::shared_ptr< ExternalContext >

external_context

)

Definition at line 30 of file FullyConnectedLayer.cc.

  : Layer(external_context), _input(nullptr), _kernel(nullptr), _bias(nullptr), _output(nullptr),
    _activation(ir::Activation::NONE)
{
  // DO NOTHING
}

Member Function Documentation

configure()

void onert::backend::xnnpack::ops::FullyConnectedLayer::configure	(	const IPortableTensor *	input,
		const IPortableTensor *	_kernel,
		const IPortableTensor *	bias,
		ir::Activation	activation,
		IPortableTensor *	output
	)

Definition at line 37 of file FullyConnectedLayer.cc.

{
  _input = input;
  _kernel = weights;
  _bias = bias;
  _activation = activation;
  _output = output;
 
  assert(_activation == ir::Activation::NONE || _activation == ir::Activation::RELU ||
         _activation == ir::Activation::RELU1 || _activation == ir::Activation::RELU6);
}

References onert::ir::NONE, onert::ir::RELU, onert::ir::RELU1, and onert::ir::RELU6.

create()

bool onert::backend::xnnpack::ops::FullyConnectedLayer::create ( )

overridevirtual

Implements onert::backend::xnnpack::ops::Layer.

Definition at line 74 of file FullyConnectedLayer.cc.

{
  float output_activation_min = 0.f, output_activation_max = 0.f;
  CalculateActivationRange<float>(_activation, &output_activation_min, &output_activation_max);
 
  const auto &kernel_shape = _kernel->getShape();
  assert(kernel_shape.rank() == 2);
  uint32_t output_channels = kernel_shape.dim(0);
  uint32_t input_channels = kernel_shape.dim(1);
 
  const auto &input_shape = _input->getShape();
  const auto &output_shape = _output->getShape();
  uint32_t flag = 0;
  if (input_shape.rank() != output_shape.rank())
  {
    flag |= XNN_FLAG_TENSORFLOW_RESHAPE_2D;
    assert(input_shape.num_elements() % input_channels == 0);
  }
  else
  {
    assert(static_cast<uint32_t>(input_shape.dim(input_shape.rank() - 1)) == input_channels);
  }
 
  assert(_kernel && _kernel->buffer());
  const float *kernel_buffer = reinterpret_cast<const float *>(_kernel->buffer());
  const float *bias_buffer = (_bias) ? reinterpret_cast<const float *>(_bias->buffer()) : nullptr;
 
  enum xnn_status status = xnn_create_fully_connected_nc_f32(
    input_channels, output_channels, input_channels /* input stride */,
    output_channels /* output stride */, kernel_buffer, bias_buffer, output_activation_min,
    output_activation_max, flag, nullptr, nullptr, &_kernel_op);
  if (status != xnn_status_success)
  {
    throw std::runtime_error{"failed to create FP32 FullyConnected operator"};
  }
  assert(_kernel_op != nullptr);
  return true;
}

References onert::backend::xnnpack::ops::Layer::_kernel_op, onert::backend::ITensor::buffer(), onert::backend::IPortableTensor::getShape(), and output_shape.

run()

void onert::backend::xnnpack::ops::FullyConnectedLayer::run ( )

overridevirtual

Implements onert::exec::IFunction.

Definition at line 51 of file FullyConnectedLayer.cc.

{
  assert(_external_context && _external_context->getThreadPool());
  if (!_setup)
  {
    _setup = setup();
    assert(_setup);
  }
 
  if (_input->data_type() == OperandType::FLOAT32)
  {
    enum xnn_status status = xnn_run_operator(_kernel_op, _external_context->getThreadPool());
    if (status != xnn_status_success)
    {
      throw std::runtime_error{"failed to run FP32 FullyConnected operator"};
    }
  }
  else
  {
    throw std::runtime_error{"XNNPACK FC: unsupported data type"};
  }
}

References onert::backend::xnnpack::ops::Layer::_external_context, onert::backend::xnnpack::ops::Layer::_kernel_op, onert::backend::xnnpack::ops::Layer::_setup, onert::backend::IPortableTensor::data_type(), and setup().

Referenced by package.infer.session::inference().

setup()

bool onert::backend::xnnpack::ops::FullyConnectedLayer::setup ( )

overridevirtual

Implements onert::backend::xnnpack::ops::Layer.

Definition at line 113 of file FullyConnectedLayer.cc.

{
  if (_input->buffer() == nullptr || _output->buffer() == nullptr)
  {
    // it could be models's input or output
    return false;
  }
 
  uint32_t batch_size = _input->getShape().num_elements() / _kernel->getShape().dim(1);
  enum xnn_status status =
    xnn_reshape_fully_connected_nc_f32(_kernel_op, batch_size, _external_context->getThreadPool());
  if (status != xnn_status_success)
  {
    throw std::runtime_error{"failed to create FP32 FullyConnected operator"};
  }
 
  status =
    xnn_setup_fully_connected_nc_f32(_kernel_op, reinterpret_cast<const float *>(_input->buffer()),
                                     reinterpret_cast<float *>(_output->buffer()));
  if (status != xnn_status_success)
  {
    throw std::runtime_error{"failed to create FP32 FullyConnected operator"};
  }
  return true;
}

References onert::backend::xnnpack::ops::Layer::_external_context, onert::backend::xnnpack::ops::Layer::_kernel_op, onert::backend::ITensor::buffer(), and onert::backend::IPortableTensor::getShape().

Referenced by run().

---

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

• runtime/onert/backend/xnnpack/ops/FullyConnectedLayer.h
• runtime/onert/backend/xnnpack/ops/FullyConnectedLayer.cc