tvm::meta_schedule::SearchStrategyNode Class Reference

The search strategy for measure candidates generation. More...

#include <search_strategy.h>

Inheritance diagram for tvm::meta_schedule::SearchStrategyNode:

Collaboration diagram for tvm::meta_schedule::SearchStrategyNode:

Public Member Functions
virtual	~SearchStrategyNode ()=default
	Virtual destructor. More...
virtual void	InitializeWithTuneContext (const TuneContext &context)=0
	Initialize the search strategy with tuning context. More...
virtual void	PreTuning (int max_trials, int num_trials_per_iter, const Array< tir::Schedule > &design_spaces, const Optional< Database > &database, const Optional< CostModel > &cost_model)=0
	Pre-tuning for the search strategy. More...
virtual void	PostTuning ()=0
	Post-tuning for the search strategy. More...
virtual Optional< Array< MeasureCandidate > >	GenerateMeasureCandidates ()=0
	Generate measure candidates from design spaces for measurement. More...
virtual void	NotifyRunnerResults (const Array< MeasureCandidate > &measure_candidates, const Array< RunnerResult > &results)=0
	Update the search strategy with measurement results. More...
virtual SearchStrategy	Clone () const =0
	Clone the search strategy. More...
	TVM_DECLARE_BASE_OBJECT_INFO (SearchStrategyNode, Object)

Static Public Attributes
static constexpr const char *	_type_key = "meta_schedule.SearchStrategy"

Detailed Description

The search strategy for measure candidates generation.

Note

The relationship between SearchStrategy and other classes are as follows: +-----------------------------------------------------------—+ +–+--------------------------------------------------------—+ | +–+---------------— Tune Context --------------------------—+ | | | +------------------—+ | | | | | | Generate | | | | | Space Generator +-----------—+ | | | | | | | | | | | +------------------—+ v | | | | Design Space | | | | +------------------—+ | | | | | Generate | | Pretuning | | | | | +--------—+ Search Strategy |<----------—+ | | | | | | | | +–+ | | +------------------—+ +–+ +-—+-------------------------------------------------------—+ | | +-—+-------------— Managed By Task Scheduler ------------------—+ | | +--------—+ | | | Send to | | Send to | | v +----------—+| Builder +-------—+ | | Measure Candidate | Builder | | Runner | | | | | +--------—+ | | | | +---------—+---------—+ | | | | | | +--------—+ | | | +-—>| Task Scheduler | | | | | | | | | Runner |<--—+ | | +----------------------—+ | | | | ^ +--—+--—+ | | | | | | +-— Runner Future <----—+ | +------------------------------------------------------------------—+

Constructor & Destructor Documentation

~SearchStrategyNode()

virtual tvm::meta_schedule::SearchStrategyNode::~SearchStrategyNode ( )

virtualdefault

Virtual destructor.

Member Function Documentation

Clone()

virtual SearchStrategy tvm::meta_schedule::SearchStrategyNode::Clone ( ) const

pure virtual

Clone the search strategy.

Returns

The cloned search strategy.

Implemented in tvm::meta_schedule::PySearchStrategyNode.

GenerateMeasureCandidates()

virtual Optional<Array<MeasureCandidate> > tvm::meta_schedule::SearchStrategyNode::GenerateMeasureCandidates ( )

pure virtual

Generate measure candidates from design spaces for measurement.

Returns

The measure candidates generated, nullptr if finished.

Implemented in tvm::meta_schedule::PySearchStrategyNode.

InitializeWithTuneContext()

virtual void tvm::meta_schedule::SearchStrategyNode::InitializeWithTuneContext ( const TuneContext &  context )

pure virtual

Initialize the search strategy with tuning context.

Parameters

context

The tuning context for initialization.

Note

This method is supposed to be called only once before every other method.

Implemented in tvm::meta_schedule::PySearchStrategyNode.

NotifyRunnerResults()

virtual void tvm::meta_schedule::SearchStrategyNode::NotifyRunnerResults ( const Array< MeasureCandidate > &  measure_candidates, const Array< RunnerResult > &  results  )

pure virtual

Update the search strategy with measurement results.

Parameters

measure_candidates	The candidates to be measured.
results	The measurement results from the runner.

Implemented in tvm::meta_schedule::PySearchStrategyNode.

PostTuning()

virtual void tvm::meta_schedule::SearchStrategyNode::PostTuning ( )

pure virtual

Post-tuning for the search strategy.

Note

Post-tuning is supposed to be called after the tuning process and before we reset the search strategy with another pre-tuning. Post-tuning can be empty.

Implemented in tvm::meta_schedule::PySearchStrategyNode.

PreTuning()

virtual void tvm::meta_schedule::SearchStrategyNode::PreTuning ( int  max_trials, int  num_trials_per_iter, const Array< tir::Schedule > &  design_spaces, const Optional< Database > &  database, const Optional< CostModel > &  cost_model  )

pure virtual

Pre-tuning for the search strategy.

Parameters

max_trials	The maximum number of trials.
num_trials_per_iter	The number of trials per iteration.
design_spaces	The design spaces used during tuning process.
database	The database used during tuning process.
cost_model	The cost model used during tuning process.

Note

Pre-tuning is supposed to be called before the tuning process and after the initialization. Because the search strategy is stateful, we can always call pretuning and reset the search strategy.

Implemented in tvm::meta_schedule::PySearchStrategyNode.

TVM_DECLARE_BASE_OBJECT_INFO()

tvm::meta_schedule::SearchStrategyNode::TVM_DECLARE_BASE_OBJECT_INFO	(	SearchStrategyNode	,
		Object
	)

Member Data Documentation

_type_key

constexpr const char* tvm::meta_schedule::SearchStrategyNode::_type_key = "meta_schedule.SearchStrategy"

staticconstexpr

---

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

• include/tvm/meta_schedule/search_strategy.h