onert::backend::builtin::IOTensor Class Reference

Tensor object that indirects to the tensor it is pointing to. More...

#include <IOTensor.h>

Collaboration diagram for onert::backend::builtin::IOTensor:

Public Member Functions
	IOTensor (const ir::OperandInfo &info)
	~IOTensor ()
void	setTensor (IPortableTensor *tensor)
uint8_t *	buffer () const override
ir::Layout	layout () const
void	set_dynamic () override
	set this tensor dynamic
void	setShape (const ir::Shape &shape) override
	Set the shape of tenser to new_shape.
bool	applyShape (const ir::Shape &shape) override
	Set the shape to shape and possibly re-allocate the buffer.
Public Member Functions inherited from onert::backend::IPortableTensor
	IPortableTensor (const ir::OperandInfo &info)
virtual	~IPortableTensor ()
const ir::OperandInfo &	get_info () const
const ir::Sparsity *	sparsity () const
size_t	total_size () const override final
size_t	calcOffset (const ir::Coordinates &coords) const override final
ir::DataType	data_type () const override final
float	data_scale () const override final
int32_t	data_zero_point () const override final
const std::vector< float > &	data_scales () const override final
const std::vector< int32_t > &	data_zero_points () const override
bool	is_constant () const override final
	Return true if the tensor is constant.
bool	is_dynamic () const override final
	Return true if the tensor needs dynamic allocation, meaning that during compile-time the outpus shape cannot be known and the output shape is calculated during kernel execution-time.
ir::Shape	getShape () const override final
	Get ir::Shape of tensor.
bool	has_padding () const final
void	access (const std::function< void(ITensor &tensor)> &fn) final
Public Member Functions inherited from onert::backend::ITensor
virtual	~ITensor ()
virtual void	deallocBuffer ()
	Dealloc the buffer (only for dynamic tensors)
virtual bool	is_subtensor () const
virtual bool	needMemoryMap () const
virtual void	enqueueWriteBuffer (const void *, bool)
virtual void	enqueueReadBuffer (void *, bool)

Additional Inherited Members
Protected Attributes inherited from onert::backend::IPortableTensor
ir::OperandInfo	_info

Detailed Description

Tensor object that indirects to the tensor it is pointing to.

A executor's I/O tensor could be two types.

1. UserTensor, if it is the primary graph (package's input/output)
2. Any other derivative of IPortableTensor from another executor, otherwise

To support these, this object indirects everything to the actual tensor pointer.

IOTensor is derived from IPortableTensor, and it also have "_info" field. "_info" field is accessed by IPortableTensor's getter method.

It assumes that IOTensor's info is always same with actual tensor's info except shape. setTensor() updates IOTensor's info's shape to actual tensor shape. Actual tensor's info should not be updated directly after setTensor() call until executor's execution is finished, instead it is allowed to update actual tensor's info indirectly by IOTensor's setter methods.

Definition at line 49 of file IOTensor.h.

Constructor & Destructor Documentation

IOTensor()

onert::backend::builtin::IOTensor::IOTensor

(

const ir::OperandInfo &

info

)

Definition at line 32 of file IOTensor.cc.

  : IPortableTensor{info}, _tensor{nullptr},
    _orig{std::make_unique<UserTensor>(info, ir::Layout::NHWC, (uint8_t *)nullptr, 0)}
{
  _tensor = _orig.get();
}

~IOTensor()

onert::backend::builtin::IOTensor::~IOTensor

(

)

Definition at line 30 of file IOTensor.cc.

{}

Member Function Documentation

applyShape()

bool onert::backend::builtin::IOTensor::applyShape ( const ir::Shape &  )

inlineoverridevirtual

Set the shape to shape and possibly re-allocate the buffer.

If a tensor is dynamic tensor and previously allocated memory exists, it will be deallocated. If a tensor is static tensor (with previously allocated memory by StaticTensorManager), buffer() will be overwriten

Parameters

shape

tensor's new shape. While allocating memory for this new_shape, tensor's shape is set to new_shape

Returns

true If applying shape is successful

false If not applying shape is not supported (it throws for other errors)

Reimplemented from onert::backend::ITensor.

Definition at line 91 of file IOTensor.h.

  {
    auto return_val = _tensor->applyShape(shape);
    if (return_val)
    {
      _info.shape(shape);
      _info.setDynamic();
    }
    return return_val;
  }

References onert::backend::IPortableTensor::_info, onert::backend::ITensor::applyShape(), onert::ir::OperandInfo::setDynamic(), and onert::ir::OperandInfo::shape().

buffer()

uint8_t * onert::backend::builtin::IOTensor::buffer ( ) const

inlineoverridevirtual

Implements onert::backend::ITensor.

Definition at line 59 of file IOTensor.h.

{ return _tensor->buffer(); }

References onert::backend::ITensor::buffer().

layout()

ir::Layout onert::backend::builtin::IOTensor::layout ( ) const

inline

Definition at line 60 of file IOTensor.h.

{ return _orig->layout(); }

set_dynamic()

void onert::backend::builtin::IOTensor::set_dynamic ( )

inlineoverridevirtual

set this tensor dynamic

Reimplemented from onert::backend::ITensor.

Definition at line 61 of file IOTensor.h.

  {
    _info.setDynamic();
    _tensor->set_dynamic();
  }

References onert::backend::IPortableTensor::_info, onert::backend::ITensor::set_dynamic(), and onert::ir::OperandInfo::setDynamic().

setShape()

void onert::backend::builtin::IOTensor::setShape ( const ir::Shape &  )

inlineoverridevirtual

Set the shape of tenser to new_shape.

Note

Higer dimension will be placed on front.

Reimplemented from onert::backend::ITensor.

Definition at line 66 of file IOTensor.h.

  {
    _info.shape(shape);
    _tensor->setShape(shape);
  }

References onert::backend::IPortableTensor::_info, onert::backend::ITensor::setShape(), and onert::ir::OperandInfo::shape().

setTensor()

void onert::backend::builtin::IOTensor::setTensor

(

IPortableTensor *

tensor

)

Definition at line 39 of file IOTensor.cc.

{
  assert(tensor);
  assert(tensor != this);
  _tensor = tensor;
  if (_info.shape() != tensor->getShape())
  {
    _info.shape(tensor->getShape());
 
    // If input tensor shape is updated, other effective buffers use dynamic memory manager.
    // Dynamic memory manager deallocate allcoated memory after each execution.
    // So we should remain input tensor as dynamic if we mark it dynamic at least once.
    // If dynamic memory manager maintains allocated memory after execution is finished,
    // we may need to reset it as static for each setTensor call.
    _info.setDynamic();
  }
}

References onert::backend::IPortableTensor::_info, onert::ir::OperandInfo::setDynamic(), and onert::ir::OperandInfo::shape().

---

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

• runtime/onert/core/src/backend/builtin/IOTensor.h
• runtime/onert/core/src/backend/builtin/IOTensor.cc