mlir::quantizer::TargetConfiguration Class Reference

#include <Configuration.h>

Inheritance diagram for mlir::quantizer::TargetConfiguration:

Public Types
using	OpHandlerFn = std::function< void(Operation *op, CAGSlice &cag)>

Public Member Functions
	TargetConfiguration (SolverContext &context)
virtual	~TargetConfiguration ()=default
unsigned	addCandidateType (quant::AnyQuantizedType quantizedType, CandidateQuantizedType::Scheme scheme)
	Adds a candidate type, returning its ordinal. More...
const CandidateQuantizedType &	getCandidateType (unsigned index) const
	Gets a prototype scheme by index. More...
ArrayRef< CandidateQuantizedType >	getCandidateTypes () const
llvm::SmallBitVector	getAllCandidateTypesMask () const
	Gets a mask of all enabled candidate types by ordinal. More...
llvm::SmallBitVector	getCandidateTypeDisabledExceptMask (ArrayRef< unsigned > exceptOrdinals) const
	Gets a mask with every candidate type except those in the given mask. More...
template<typename OpTy >
void	addOpHandler (OpHandlerFn fn)
	Adds an op handler. More...
template<typename OpTy >
void	addRequireStatsOp ()
bool	isRequireStatsOp (Operation *op) const
	Returns whether opName is a RequireStatsOp. More...
template<typename OpTy >
void	addValueIdentityOp ()
void	handleOp (Operation *op, CAGSlice &cag) const
	Handles the operation if a handler is defined for it. More...
virtual void	finalizeAnchors (CAGSlice &cag) const
	Finalizes the CAG after all anchors have been added. More...
virtual bool	isHandledType (Type t) const =0
	Whether an operand or result type is subject to analysis by this config. More...

Static Public Attributes
static constexpr size_t	MaxSchemeIndex = 31

Protected Member Functions
virtual void	addValueIdentityOpByName (StringRef opName)=0
void	addOpHandlerByName (StringRef name, OpHandlerFn fn)

Detailed Description

Defines quantization configuration for the target. The settings here depend on a variety of details about the deployment environment, although, where we have control over such things, we do try to standardize as possible.

Non-const methods are used to setup the configuration. It is expected that const instances/references are used post-build.

Member Typedef Documentation

OpHandlerFn

using mlir::quantizer::TargetConfiguration::OpHandlerFn = std::function<void(Operation *op, CAGSlice &cag)>

Constructor & Destructor Documentation

TargetConfiguration()

TargetConfiguration::TargetConfiguration

(

SolverContext &

context

)

~TargetConfiguration()

virtual mlir::quantizer::TargetConfiguration::~TargetConfiguration ( )

virtualdefault

Member Function Documentation

addCandidateType()

unsigned mlir::quantizer::TargetConfiguration::addCandidateType ( quant::AnyQuantizedType  quantizedType, CandidateQuantizedType::Scheme  scheme  )

inline

Adds a candidate type, returning its ordinal.

addOpHandler()

template<typename OpTy >

void mlir::quantizer::TargetConfiguration::addOpHandler ( OpHandlerFn  fn )

inline

Adds an op handler.

addOpHandlerByName()

void TargetConfiguration::addOpHandlerByName ( StringRef  name, OpHandlerFn  fn  )

protected

addRequireStatsOp()

template<typename OpTy >

void mlir::quantizer::TargetConfiguration::addRequireStatsOp ( )

inline

Adds an operation which requires statistics at its result nodes for best quantization performance. Note that the opName StringRef is expected to come from getOperationName() and be static.

addValueIdentityOp()

template<typename OpTy >

void mlir::quantizer::TargetConfiguration::addValueIdentityOp ( )

inline

Adds an op which does not mutate its values but may mutate its shape or combine its operands in an arbitrary way. Such ops are expected to have the same types for operands and results and must be capable of operating on storage types.

addValueIdentityOpByName()

virtual void mlir::quantizer::TargetConfiguration::addValueIdentityOpByName ( StringRef  opName )

protectedpure virtual

finalizeAnchors()

virtual void mlir::quantizer::TargetConfiguration::finalizeAnchors ( CAGSlice &  cag ) const

inlinevirtual

Finalizes the CAG after all anchors have been added.

getAllCandidateTypesMask()

llvm::SmallBitVector mlir::quantizer::TargetConfiguration::getAllCandidateTypesMask ( ) const

inline

Gets a mask of all enabled candidate types by ordinal.

getCandidateType()

const CandidateQuantizedType& mlir::quantizer::TargetConfiguration::getCandidateType ( unsigned  index ) const

inline

Gets a prototype scheme by index.

getCandidateTypeDisabledExceptMask()

llvm::SmallBitVector mlir::quantizer::TargetConfiguration::getCandidateTypeDisabledExceptMask ( ArrayRef< unsigned >  exceptOrdinals ) const

inline

Gets a mask with every candidate type except those in the given mask.

getCandidateTypes()

ArrayRef<CandidateQuantizedType> mlir::quantizer::TargetConfiguration::getCandidateTypes ( ) const

inline

handleOp()

void TargetConfiguration::handleOp	(	Operation *	op,
		CAGSlice &	cag
	)		const

Handles the operation if a handler is defined for it.

isHandledType()

virtual bool mlir::quantizer::TargetConfiguration::isHandledType ( Type  t ) const

pure virtual

Whether an operand or result type is subject to analysis by this config.

isRequireStatsOp()

bool TargetConfiguration::isRequireStatsOp

(

Operation *

op

)

const

Returns whether opName is a RequireStatsOp.

Member Data Documentation

MaxSchemeIndex

constexpr size_t mlir::quantizer::TargetConfiguration::MaxSchemeIndex = 31

static

---

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

• include/mlir/Quantizer/Support/Configuration.h
• lib/Quantizer/Support/Configuration.cpp