Classes
class	TuningOptionsNode
	Tuning and measurement options. More...

class	TuningOptions
	Managed reference to TuningOptionsNode. More...

class	AccessAnalyzerNode
	Static analyzer for a ComputeDAG. More...

class	AccessAnalyzer
	Managed reference to AccessAnalyzerNode. More...

class	ComputeDAGNode
	The auto-scheduler's computational graph and related program analyses. More...

class	ComputeDAG
	Managed reference to ComputeDAGNode. More...

class	CostModelNode
	The base class for cost model. More...

class	CostModel
	Managed reference to CostModelNode. More...

class	RandomModelNode
	The cost model returning random value for all predictions. More...

class	RandomModel
	Managed reference to RandomModelNode. More...

class	PythonBasedModelNode
	A wrapper for cost model defined by python code This class will call functions defined in the python. More...

class	PythonBasedModel
	Managed reference to PythonBasedModelNode. More...

struct	StageAttributes
	Stage-level attributes. More...

class	StageNode
	A op stage in the compute declaration. Similar to te::Stage in `include/tvm/te/schedule.h`. More...

class	Stage
	Managed reference to StageNode. More...

class	AttachMapNode
	stores the compute_at relation between stages This stores a bi-directional mapping from stages and iter: More...

class	AttachMap
	Managed reference to AttachMapNode. More...

class	StateNode
	A state in the search process. It consists of the current loop structure and a list of transformation steps used to construct it. Each State corresponds to a specific schedule for its ComputeDAG. More...

class	State
	Managed reference to StateNode. More...

class	MeasureInputNode
	Store the input of a measurement. More...

class	MeasureInput
	Managed reference to MeasureInputNode. More...

class	BuildResultNode
	Store the result of a build. More...

class	BuildResult
	Managed reference to BuildResultNode. More...

class	MeasureResultNode
	Store the results of a measurement. More...

class	MeasureResult
	Managed reference to MeasureResultNode. More...

class	MeasureCallbackNode
	Bass class of measurement callbacks. More...

class	MeasureCallback
	Managed reference to MeasureCallbackNode. More...

class	PythonBasedMeasureCallbackNode
	A wrapper for measure callback defined by python code This class will call functions defined in the python. More...

class	PythonBasedMeasureCallback
	Managed reference to PythonBasedMeasureCallbackNode. More...

class	ProgramBuilderNode
	ProgramBuilder that builds the programs. More...

class	ProgramBuilder
	Managed reference to ProgramBuilderNode. More...

class	ProgramRunnerNode
	ProgramRunner that runs the built programs and measure the time cost. More...

class	ProgramRunner
	Managed reference to ProgramRunnerNode. More...

class	LocalBuilderNode
	LocalBuilder use local CPU cores to build programs in parallel. More...

class	LocalBuilder
	Managed reference to LocalBuilderNode. More...

class	LocalRunnerNode
	LocalRunner that uses local CPU/GPU to measure the time cost of programs. More...

class	LocalRunner
	Managed reference to LocalRunnerNode. More...

class	RPCRunnerNode
	RPCRunner that uses RPC call to measures the time cost of programs on remote devices. Or sometime we may need to use RPC even in local running to insulate the thread environment. (e.g. running CUDA programs) More...

class	RPCRunner
	Managed reference to RPCRunnerNode. More...

class	ProgramMeasurerNode
	Measurer that measures the time costs of tvm programs This class combines ProgramBuilder and ProgramRunner, and provides a simpler API. More...

class	ProgramMeasurer
	Managed reference to ProgramMeasurerNode. More...

class	RecordToFileNode
	Callback for logging the input and results of measurements to file. More...

class	RecordToFile
	Managed reference to RecordToFileNode. More...

class	RecordReaderNode
	Log reader to load step logs from a file. More...

class	RecordReader
	Managed reference to RecordReaderNode. More...

class	SearchCallbackNode
	Callback function to be called by the search process. This interface allows to do extra initializations before schedule search or extra check during/after the schedule search. More...

class	SearchCallback
	Managed reference to SearchCallbackNode. More...

class	PreloadMeasuredStatesNode
	Preload measured states from a log file. This can resume the state of the search policy. More...

class	PreloadMeasuredStates
	Managed reference to PreloadMeasuredStatesNode. More...

struct	SearchPolicyKey
	Attribute keys of ops used for SearchPolicy. More...

class	SearchPolicyNode
	The base class of search policies. More...

class	SearchPolicy
	Managed reference to SearchPolicyNode. More...

class	HardwareParamsNode
	The parameters of target hardware used to guide the SearchPolicy. More...

class	HardwareParams
	Managed reference to HardwareParamsNode. More...

class	SearchTaskNode
	The computation information and hardware parameters for a specific schedule search task. More...

class	SearchTask
	Managed reference to SearchTaskNode. More...

class	IteratorNode
	An iterator of a for-loop Similar to tvm::IterVar in `include/tvm/tir/expr.h` More...

class	Iterator
	Managed reference to IteratorNode. More...

class	StepNode
	The base class of transformation steps. Each step has its corresponding tvm.te schedule primitives. More...

class	Step
	Managed reference to StepNode. More...

class	AnnotationStepNode
	Annotation step that corresponds to vectorize, parallel, unroll and thread binding. (i.e. te::Stage::vectorize, te::Stage::parallel, te::Stage::vectorize, te::Stage::bind) More...

class	AnnotationStep
	Managed reference to AnnotationStepNode. More...

class	FuseStepNode
	Fuse step that corresponds to te::Stage::fuse. More...

class	FuseStep
	Managed reference to FuseStepNode. More...

class	PragmaStepNode
	Pragma step that corresponds to te::Stage::pragma. More...

class	PragmaStep
	Managed reference to PragmaStepNode. More...

class	ReorderStepNode
	Reorder step that corresponds to te::Stage::reorder. More...

class	ReorderStep
	Managed reference to ReorderStepNode. More...

class	SplitStepNode
	Split step that corresponds to te::Stage::split with additional support of multiple-level of factors. More...

class	SplitStep
	Managed reference to SplitStepNode. More...

class	FollowSplitStepNode
	Similar to SplitStepNode, but uses split factors from another step (i.e. Follow another split step) More...

class	FollowSplitStep
	Managed reference to FollowSplitStepNode. More...

class	FollowFusedSplitStepNode
	Similar to FollowSplitStep, but uses split factors from multiple steps. More...

class	FollowFusedSplitStep
	Managed reference to FollowFusedSplitStepNode. More...

class	StorageAlignStepNode
	Storage align step that corresponds to te::Stage::storage_align. More...

class	StorageAlignStep
	Managed reference to StorageAlignStepNode. More...

class	ComputeAtStepNode
	Compute at step that corresponds to te::Stage::compute_at. More...

class	ComputeAtStep
	Managed reference to ComputeAtStepNode. More...

class	ComputeInlineStepNode
	Compute inline step that corresponds to te::Stage::compute_inline. More...

class	ComputeInlineStep
	Managed reference to ComputeInlineStepNode. More...

class	ComputeRootStepNode
	Compute root step that corresponds to te::Stage::compute_root. More...

class	ComputeRootStep
	Managed reference to ComputeRootStepNode. More...

class	CacheReadStepNode
	Cache read step that corresponds to te::Schedule::cache_read. More...

class	CacheReadStep
	Managed reference to CacheReadStepNode. More...

class	CacheWriteStepNode
	Cache write step that corresponds to te::Schedule::cache_write. More...

class	CacheWriteStep
	Managed reference to CacheWriteStepNode. More...

class	RfactorStepNode
	Reduction factor step that corresponds to te::Schedule::rfactor. More...

class	RfactorStep
	Managed reference to RfactorStepNode. More...

Typedefs
using	StageKey = int
	Use stage_id to represent a stage. More...

using	IterKey = std::pair< int, int >
	Use stage_id and iter_id to represent a iterator. More...

typedef Map< tvm::te::Stage, Array< tir::IterVar >, ObjectHash, ObjectEqual >	StageToAxesMap

Enumerations
enum class	LayoutRewriteOption : int { NoRewrite = 0 , InsertTransformStage = 1 , RewriteForPreTransformed = 2 }
	Options for applying layout rewrite. This is an optimization to rewrite the layout of input tensors according to the schedule we get. More...

enum class	StageKind : int { kPlaceholder = 0 , kCompute = 1 }
	The type of a stage. More...

enum class	ComputeAtKind : int { kRoot = 0 , kInlined = 1 , kIter = 2 }
	The type of compute location. More...

enum class	MeasureErrorNO : int { kNoError = 0 , kInstantiationError = 1 , kCompileHostError = 2 , kCompileDeviceError = 3 , kRuntimeDeviceError = 4 , kWrongAnswerError = 5 , kBuildTimeoutError = 6 , kRunTimeoutError = 7 , kUnknownError = 8 }
	The error code of one measurement. More...

enum class	IteratorKind : int { kSpatial = 0 , kReduction = 1 , kMixed = 2 , kSpecial = 3 }
	The type of an iterator. More...

enum class	IteratorAnnotation : int { kNone = 0 , kUnroll = 1 , kVectorize = 2 , kParallel = 3 , kVThread = 4 , kBlockX = 5 , kThreadX = 6 , kBlockY = 7 , kThreadY = 8 , kBlockZ = 9 , kThreadZ = 10 , kTensorize = 11 }
	The type of an iterator's annotation. More...

Functions
std::pair< te::Schedule, Array< te::Tensor > >	AutoSchedule (SearchPolicy search_policy, TuningOptions tuning_options)
	Run schedule search for a given compute declaration. More...

Array< PrimExpr >	GetShapeFromRewrittenLayout (String rewritten_layout, Array< String > axis_names)
	Get the orginal shape from a rewritten layout string. More...

void	GetPerStoreFeature (const PrimFunc &func, int cache_line_size, int max_n_bufs, std::vector< float > *ret, bool log_scale=true)
	Get per-store features from a TIR PrimFunc. More...

void	GetPerStoreFeatureName (int max_n_bufs, std::vector< std::string > *ret)

void	GetPerStoreFeaturesFromStates (const Array< State > &states, const SearchTask &task, int skip_first_n_feature_extraction, int max_n_bufs, std::vector< std::vector< float >> *features)
	Get per-store feature from states of the same task. More...

void	GetPerStoreFeaturesFromStates (const Array< State > &states, const std::vector< SearchTask > &tasks, int skip_first_n_feature_extraction, int max_n_bufs, std::vector< std::vector< float >> *features)
	Get per-store feature from states of different tasks. More...

void	GetPerStoreFeaturesFromFile (const std::string &filename, int max_lines, int max_n_bufs, std::vector< std::vector< float >> features, std::vector< float > normalized_throughputs, std::vector< int > *task_ids)
	Get per-store features from a log file. More...

void	GetPerStoreFeaturesFromMeasurePairs (const Array< MeasureInput > &inputs, const Array< MeasureResult > &results, int skip_first_n_feature_extraction, int max_n_bufs, std::vector< std::vector< float >> features, std::vector< float > normalized_throughputs, std::vector< int > *task_ids)
	Get per-store features from measurement input/result pairs. More...

void	WriteMeasureRecords (std::ostream *os, const Array< MeasureInput > &inputs, const Array< MeasureResult > &results, const std::string log_version=AUTO_SCHEDULER_LOG_VERSION)
	Append measure records to an output stream. More...

void	ReadMeasureRecord (const std::string &str, MeasureInputNode inp, MeasureResultNode res, std::string *log_version)
	Read one measure record from a string. More...

void	UpdateStageToAxesMap (const te::Stage &stage, StageToAxesMap *stage_to_axes)
	Update the current stage IterVar information to StageToAxesMap. More...

Step	StepReadFromRecord (dmlc::JSONReader *reader)
	Read a step record from JSONReader and create the corresponding step. More...

void	StepApplyToState (const Step &step, State *state, const ComputeDAG &dag)
	Apply a general step to a State with runtime dynamic dispatching. More...

void	StepApplyToSchedule (const Step &step, Array< te::Stage > stages, StageToAxesMap stage_to_axes, te::Schedule *schedule, const Array< Step > &transform_steps)
	Apply a general step to tvm.schedule with runtime dynamic dispatching. More...

String	StepPrintAsPythonAPI (const Step &step, Array< te::Stage > stages, StageToAxesMap stage_to_axes, te::Schedule *schedule, const Array< Step > &transform_steps)
	Print a general step as equivalent python schedule API with runtime dynamic dispatching. More...

Variables
const std::string	AUTO_SCHEDULER_LOG_VERSION = "v0.6"

const char *	IteratorAnnotationString []

Typedef Documentation

◆ IterKey

using tvm::auto_scheduler::IterKey = typedef std::pair<int, int>

Use stage_id and iter_id to represent a iterator.

◆ StageKey

using tvm::auto_scheduler::StageKey = typedef int

Use stage_id to represent a stage.

◆ StageToAxesMap

typedef Map<tvm::te::Stage, Array<tir::IterVar>, ObjectHash, ObjectEqual> tvm::auto_scheduler::StageToAxesMap

Enumeration Type Documentation

◆ ComputeAtKind

enum tvm::auto_scheduler::ComputeAtKind : int

strong

The type of compute location.

Enumerator
kRoot	Compute at root.
kInlined	Compute inlined.
kIter	Compute at some iterator.

◆ IteratorAnnotation

enum tvm::auto_scheduler::IteratorAnnotation : int

strong

The type of an iterator's annotation.

Enumerator
kNone	This iterator has no annotation.
kUnroll	This iterator has been unrolled.
kVectorize	This iterator has been vectorized.
kParallel	This iterator has been paralleld.
kVThread	This iterator has been bind to vthread.
kBlockX	This iterator has been bind to blockIdx.x.
kThreadX	This iterator has been bind to threadIdx.x.
kBlockY	This iterator has been bind to blockIdx.y.
kThreadY	This iterator has been bind to threadIdx.y.
kBlockZ	This iterator has been bind to blockIdx.y.
kThreadZ	This iterator has been bind to threadIdx.y.
kTensorize	This iterator has been mapped with a tensorize intrinsic.

◆ IteratorKind

enum tvm::auto_scheduler::IteratorKind : int

strong

The type of an iterator.

Enumerator
kSpatial	Spatial iterator.
kReduction	Reduction iterator.
kMixed	Fused spatial and reduction iterator.
kSpecial	Special iterator. (e.g. virtual root iterator)

◆ LayoutRewriteOption

enum tvm::auto_scheduler::LayoutRewriteOption : int

strong

Options for applying layout rewrite. This is an optimization to rewrite the layout of input tensors according to the schedule we get.

Enumerator

NoRewrite

Do not perform layout rewrite.

InsertTransformStage

Insert layout transformation stages for input placeholders in the compute DAG.

RewriteForPreTransformed

Do not insert layout transformation stages and assume the input placeholders are pre-transformed.

Note: The lowered function with this option does not accept the origial input shapes, so this option must be used along with AutoSchedulerLayoutRewrite pass in Relay.

◆ MeasureErrorNO

enum tvm::auto_scheduler::MeasureErrorNO : int

strong

The error code of one measurement.

Enumerator
kNoError	No error.
kInstantiationError	Errors happen when apply transform steps from init state.
kCompileHostError	Errors happen when compiling code on host. (when build module)
kCompileDeviceError	Errors happen when compiling code on device. (when load module)
kRuntimeDeviceError	Errors happen when run program on device.
kWrongAnswerError	Answer is wrong when compared to a reference output.
kBuildTimeoutError	Timeout during compilation.
kRunTimeoutError	Timeout during run.
kUnknownError	Unknown error.

◆ StageKind

enum tvm::auto_scheduler::StageKind : int

strong

The type of a stage.

Enumerator
kPlaceholder	A placeholder stage.
kCompute	A compute stage.

Function Documentation

◆ AutoSchedule()

std::pair<te::Schedule, Array<te::Tensor> > tvm::auto_scheduler::AutoSchedule	(	SearchPolicy	search_policy,
		TuningOptions	tuning_options
	)

Run schedule search for a given compute declaration.

Parameters

search_policy	The search policy.
tuning_options	Tuning and measurement options.

Returns: A te::schedule and an Array of te::Tensor to be used in tvm.lower or tvm.build.

◆ GetPerStoreFeature()

void tvm::auto_scheduler::GetPerStoreFeature	(	const PrimFunc &	func,
		int	cache_line_size,
		int	max_n_bufs,
		std::vector< float > *	ret,
		bool	log_scale = `true`
	)

Get per-store features from a TIR PrimFunc.

Parameters

func	The input lowered TIR PrimFunc
cache_line_size	The size of cache line in bytes
max_n_bufs	The maximum number of extracted buffers for one statement
ret	The returned feature vector
log_scale	Should the outputs be scaled by log2(1+x).

◆ GetPerStoreFeatureName()

void tvm::auto_scheduler::GetPerStoreFeatureName	(	int	max_n_bufs,
		std::vector< std::string > *	ret
	)

◆ GetPerStoreFeaturesFromFile()

void tvm::auto_scheduler::GetPerStoreFeaturesFromFile	(	const std::string &	filename,
		int	max_lines,
		int	max_n_bufs,
		std::vector< std::vector< float >> *	features,
		std::vector< float > *	normalized_throughputs,
		std::vector< int > *	task_ids
	)

Get per-store features from a log file.

Parameters

filename	The name of log file
max_lines	Only read the first n lines of the file
max_n_bufs	The maximum number of extracted buffers for one statement
features	The returned feature vector. The innermost vector contains the feature vectors for all BufferStoreNode statements
normalized_throughputs	The normalized throughputs for all states
task_ids	The task ids for all states

◆ GetPerStoreFeaturesFromMeasurePairs()

void tvm::auto_scheduler::GetPerStoreFeaturesFromMeasurePairs	(	const Array< MeasureInput > &	inputs,
		const Array< MeasureResult > &	results,
		int	skip_first_n_feature_extraction,
		int	max_n_bufs,
		std::vector< std::vector< float >> *	features,
		std::vector< float > *	normalized_throughputs,
		std::vector< int > *	task_ids
	)

Get per-store features from measurement input/result pairs.

Parameters

inputs	The measurement inputs
results	The measurement results
skip_first_n_feature_extraction	Skip feature extraction for the first n measurement pairs
max_n_bufs	The maximum number of extracted buffers for one statement
features	The returned feature vector. The innermost vector contains the feature vectors for all BufferStoreNode statements
normalized_throughputs	The normalized throughputs for all states
task_ids	The task ids for all states

◆ GetPerStoreFeaturesFromStates() [1/2]

void tvm::auto_scheduler::GetPerStoreFeaturesFromStates	(	const Array< State > &	states,
		const SearchTask &	task,
		int	skip_first_n_feature_extraction,
		int	max_n_bufs,
		std::vector< std::vector< float >> *	features
	)

Get per-store feature from states of the same task.

Parameters

states	The input states
task	The same search task for all states
skip_first_n_feature_extraction	Skip feature extraction for the first n states
max_n_bufs	The maximum number of extracted buffers for one statement
features	The returned feature vector. The innermost vector contains the feature vectors for all BufferStoreNode statements

◆ GetPerStoreFeaturesFromStates() [2/2]

void tvm::auto_scheduler::GetPerStoreFeaturesFromStates	(	const Array< State > &	states,
		const std::vector< SearchTask > &	tasks,
		int	skip_first_n_feature_extraction,
		int	max_n_bufs,
		std::vector< std::vector< float >> *	features
	)

Get per-store feature from states of different tasks.

Parameters

states	The input states
tasks	The search tasks corresponding to the input states
skip_first_n_feature_extraction	Skip feature extraction for the first n states
max_n_bufs	The maximum number of extracted buffers for one statement
features	The returned feature vector. The innermost vector contains the feature vectors for all BufferStoreNode statements

◆ GetShapeFromRewrittenLayout()

Array<PrimExpr> tvm::auto_scheduler::GetShapeFromRewrittenLayout	(	String	rewritten_layout,
		Array< String >	axis_names
	)

Get the orginal shape from a rewritten layout string.

Parameters

rewritten_layout	The layout after auto-scheduler's layout rewrite.
axis_names	Specifiy the names of axes.

Returns: shape The original shape.

◆ ReadMeasureRecord()

void tvm::auto_scheduler::ReadMeasureRecord	(	const std::string &	str,
		MeasureInputNode *	inp,
		MeasureResultNode *	res,
		std::string *	log_version
	)

Read one measure record from a string.

Parameters

str	The record string to be parsed.
inp	A pointer to a MeasureInputNode used to store the return value.
res	A pointer to a MeasureResultNode used to store the return value.
log_version	A pointer to a string used to store the log version.

◆ StepApplyToSchedule()

void tvm::auto_scheduler::StepApplyToSchedule	(	const Step &	step,
		Array< te::Stage > *	stages,
		StageToAxesMap *	stage_to_axes,
		te::Schedule *	schedule,
		const Array< Step > &	transform_steps
	)

Apply a general step to tvm.schedule with runtime dynamic dispatching.

Parameters

step	The step to be applied to tvm.schedule.
stages	The list of current stages
stage_to_axes	A map that maps stage ot all its iterators.
schedule	A mutable point to the current schedule
transform_steps	An array of all history transform steps.

◆ StepApplyToState()

void tvm::auto_scheduler::StepApplyToState	(	const Step &	step,
		State *	state,
		const ComputeDAG &	dag
	)

Apply a general step to a State with runtime dynamic dispatching.

Parameters

step	The step to be applied to State.
state	A mutable pointer to state, which will be updated.
dag	The original ComputeDAG of this state.

◆ StepPrintAsPythonAPI()

String tvm::auto_scheduler::StepPrintAsPythonAPI	(	const Step &	step,
		Array< te::Stage > *	stages,
		StageToAxesMap *	stage_to_axes,
		te::Schedule *	schedule,
		const Array< Step > &	transform_steps
	)

Print a general step as equivalent python schedule API with runtime dynamic dispatching.

Parameters

step	The step to be printed as python API.
stages	The list of current stages
stage_to_axes	A map that maps stage ot all its iterators.
schedule	A mutable point to the current schedule
transform_steps	An array of all history transform steps.

Returns: Python schedule code.

◆ StepReadFromRecord()

Step tvm::auto_scheduler::StepReadFromRecord ( dmlc::JSONReader * reader )

Read a step record from JSONReader and create the corresponding step.

Parameters

reader The input JSONReader.

◆ UpdateStageToAxesMap()

void tvm::auto_scheduler::UpdateStageToAxesMap	(	const te::Stage &	stage,
		StageToAxesMap *	stage_to_axes
	)

Update the current stage IterVar information to StageToAxesMap.

Parameters

stage	The stage to be updated.
stage_to_axes	The map to be updated.

◆ WriteMeasureRecords()

void tvm::auto_scheduler::WriteMeasureRecords	(	std::ostream *	os,
		const Array< MeasureInput > &	inputs,
		const Array< MeasureResult > &	results,
		const std::string	log_version = `AUTO_SCHEDULER_LOG_VERSION`
	)

Append measure records to an output stream.

Parameters

os	A pointer to a output stream.
inputs	The MeasureInputs to be written.
results	The MeasureResults to be written.
log_version	The log version for the given record.

Variable Documentation

◆ AUTO_SCHEDULER_LOG_VERSION

const std::string tvm::auto_scheduler::AUTO_SCHEDULER_LOG_VERSION = "v0.6"

◆ IteratorAnnotationString

const char* tvm::auto_scheduler::IteratorAnnotationString[]

extern

Classes

Typedefs

Enumerations

Functions

Variables

Typedef Documentation

◆ IterKey

◆ StageKey

◆ StageToAxesMap

Enumeration Type Documentation

◆ ComputeAtKind

◆ IteratorAnnotation

◆ IteratorKind

◆ LayoutRewriteOption

◆ MeasureErrorNO

◆ StageKind

Function Documentation

◆ AutoSchedule()

◆ GetPerStoreFeature()

◆ GetPerStoreFeatureName()

◆ GetPerStoreFeaturesFromFile()

◆ GetPerStoreFeaturesFromMeasurePairs()

◆ GetPerStoreFeaturesFromStates() [1/2]

◆ GetPerStoreFeaturesFromStates() [2/2]

◆ GetShapeFromRewrittenLayout()

◆ ReadMeasureRecord()

◆ StepApplyToSchedule()

◆ StepApplyToState()

◆ StepPrintAsPythonAPI()

◆ StepReadFromRecord()

◆ UpdateStageToAxesMap()

◆ WriteMeasureRecords()

Variable Documentation

◆ AUTO_SCHEDULER_LOG_VERSION

◆ IteratorAnnotationString