Class to represent Permute operation. More...

#include <Permute.h>

Collaboration diagram for onert::ir::operation::Permute:

[legend]

Public Member Functions
void	accept (OperationVisitor &v) const override

OpCode	opcode () const final

	Permute (const OperandIndex &input, const OperandIndex &output)

Public Member Functions inherited from onert::ir::Operation
	Operation (OperandConstraint input_constr, const OperandIndexSequence &inputs, const OperandIndexSequence &outputs, OperandConstraint output_constr=OperandConstraint::createAny())

	Operation (OperandConstraint input_constr, OperandConstraint output_constr=OperandConstraint::createAny())

	Operation (const Operation &)=default

	Operation (Operation &&)=default

Operation &	operator= (const Operation &)=default

Operation &	operator= (Operation &&)=default

virtual	~Operation ()

void	replaceInputs (const OperandIndex &from, const OperandIndex &to) override

void	replaceOutputs (const OperandIndex &from, const OperandIndex &to) override

OperandIndexSequence &	getInputs ()

const OperandIndexSequence &	getInputs () const override

const OperandIndexSequence &	getOutputs () const override

void	setInputs (const OperandIndexSequence &indexes)

void	setOutputs (const OperandIndexSequence &indexes)

Public Member Functions inherited from onert::ir::IOperation
virtual	~IOperation ()=default

virtual std::string	name () const

Detailed Description

Class to represent Permute operation.

Note

Permute operation reorders the dimensions of a tensor.

   This operation is virtual operation, which is not used on real model, but used internally.
   It was introduced to support various model layout (NHWC, NCHW, etc) and backend layout.
   But currently, model layout and backend layout are always same as NHWC.
   So this operation is used for below cases.
   1) Handle model output buffer's special case
     1-1) Model output is comes from model constant
     1-2) Model output is comes from model input
     1-3) Model output shares tensor with other model output(s)
   2) Handle shared tensor between different backend

   Q) Why name is still 'Permute'?
   A) It is handled as copy operation on compile phase,
      but it can be permute operation if output buffer layout is changed by API call

Definition at line 55 of file Permute.h.

Constructor & Destructor Documentation

◆ Permute()

onert::ir::operation::Permute::Permute	(	const OperandIndex &	input,
		const OperandIndex &	output
	)

Definition at line 29 of file Permute.cc.

  : Operation{OperandConstraint::createExact(1u)}
{
  setInputs({input});
  setOutputs({output});
}

References onert::ir::Operation::setInputs(), and onert::ir::Operation::setOutputs().

Member Function Documentation

◆ accept()

void onert::ir::operation::Permute::accept ( OperationVisitor & v ) const

overridevirtual

Implements onert::ir::IOperation.

Reimplemented in onert::ir::train::operation::Permute.

Definition at line 27 of file Permute.cc.

27{ v.visit(*this); }

◆ opcode()

OpCode onert::ir::operation::Permute::opcode ( ) const

inlinefinalvirtual

Implements onert::ir::IOperation.

Definition at line 59 of file Permute.h.

59{ return OpCode::Permute; }

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