onert::backend::ggml::ops::FullyConnectedLayer Class Reference

#include <FullyConnectedLayer.h>

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

Public Member Functions
	FullyConnectedLayer ()
	~FullyConnectedLayer ()
void	fullyConnectedGGMLWeight ()
void	configure (const IPortableTensor *input, const IPortableTensor *weights, const IPortableTensor *bias, ir::Activation activation, IPortableTensor *output, const std::shared_ptr< ExternalContext > &external_context)
void	run () override
void	prepare () override
Public Member Functions inherited from onert::exec::IFunction
virtual	~IFunction ()=default

Protected Attributes
const IPortableTensor *	_input
const IPortableTensor *	_weights
const IPortableTensor *	_bias
IPortableTensor *	_output
ir::Activation	_activation
std::shared_ptr< ExternalContext >	_external_context

Detailed Description

Definition at line 29 of file FullyConnectedLayer.h.

Constructor & Destructor Documentation

FullyConnectedLayer()

onert::backend::ggml::ops::FullyConnectedLayer::FullyConnectedLayer

(

)

Definition at line 76 of file FullyConnectedLayer.cc.

  : _input(nullptr), _weights(nullptr), _bias(nullptr), _output(nullptr),
    _activation(ir::Activation::NONE), _external_context(nullptr)
{
  // DO NOTHING
}

~FullyConnectedLayer()

onert::backend::ggml::ops::FullyConnectedLayer::~FullyConnectedLayer ( )

default

Member Function Documentation

configure()

void onert::backend::ggml::ops::FullyConnectedLayer::configure	(	const IPortableTensor *	input,
		const IPortableTensor *	weights,
		const IPortableTensor *	bias,
		ir::Activation	activation,
		IPortableTensor *	output,
		const std::shared_ptr< ExternalContext > &	external_context
	)

Definition at line 118 of file FullyConnectedLayer.cc.

{
  _input = input;
  _weights = weights;
  _bias = bias;
  _activation = activation;
  _output = output;
  _external_context = external_context;
}

References _activation, _bias, _external_context, _input, _output, and _weights.

fullyConnectedGGMLWeight()

void onert::backend::ggml::ops::FullyConnectedLayer::fullyConnectedGGMLWeight

(

)

Definition at line 85 of file FullyConnectedLayer.cc.

{
  if (_bias)
    throw std::runtime_error{"FullyConnected: GGML weights format does not support bias yet."};
 
  // convert tensor
  auto input = getGGMLTensor(_input);
  auto weights = getGGMLTensor(_weights);
  auto output = getGGMLTensor(_output);
  {
    output.op = GGML_OP_MUL_MAT;
    output.src[0] = &weights;
    output.src[1] = &input;
  }
  auto *nodes = &output;
 
  // create graph
  struct ggml_cgraph graph;
  {
    memset(&graph, 0, sizeof(graph));
    graph.n_nodes = 1;
    graph.nodes = &nodes;
  }
 
  // get cplan
  auto cplan = ggml_graph_plan(&graph, _external_context->maxNumThreads());
  std::vector<uint8_t> buf(cplan.work_size);
  cplan.work_data = buf.data();
 
  // compute
  ggml_graph_compute(&graph, &cplan);
}

References _bias, _external_context, _input, _output, _weights, and onert::backend::ggml::ops::getGGMLTensor().

Referenced by run().

prepare()

void onert::backend::ggml::ops::FullyConnectedLayer::prepare ( )

overridevirtual

Reimplemented from onert::exec::IFunction.

Definition at line 144 of file FullyConnectedLayer.cc.

{
  // DO NOTHING
}

run()

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

overridevirtual

Implements onert::exec::IFunction.

Definition at line 131 of file FullyConnectedLayer.cc.

{
  if (_weights->data_type() == ir::DataType::QUANT_GGML_Q4_0 ||
      _weights->data_type() == ir::DataType::QUANT_GGML_Q8_0)
  {
    fullyConnectedGGMLWeight();
  }
  else
  {
    throw std::runtime_error{"FullyConnected: unsupported data type"};
  }
}

References _weights, onert::backend::IPortableTensor::data_type(), and fullyConnectedGGMLWeight().

Field Documentation

_activation

ir::Activation onert::backend::ggml::ops::FullyConnectedLayer::_activation

protected

Definition at line 51 of file FullyConnectedLayer.h.

Referenced by configure().

_bias

const IPortableTensor* onert::backend::ggml::ops::FullyConnectedLayer::_bias

protected

Definition at line 49 of file FullyConnectedLayer.h.

Referenced by configure(), and fullyConnectedGGMLWeight().

_external_context

std::shared_ptr<ExternalContext> onert::backend::ggml::ops::FullyConnectedLayer::_external_context

protected

Definition at line 53 of file FullyConnectedLayer.h.

Referenced by configure(), and fullyConnectedGGMLWeight().

_input

const IPortableTensor* onert::backend::ggml::ops::FullyConnectedLayer::_input

protected

Definition at line 47 of file FullyConnectedLayer.h.

Referenced by configure(), and fullyConnectedGGMLWeight().

_output

IPortableTensor* onert::backend::ggml::ops::FullyConnectedLayer::_output

protected

Definition at line 50 of file FullyConnectedLayer.h.

Referenced by configure(), and fullyConnectedGGMLWeight().

_weights

const IPortableTensor* onert::backend::ggml::ops::FullyConnectedLayer::_weights

protected

Definition at line 48 of file FullyConnectedLayer.h.

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

---

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

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