tvm::tir::TraceNode Class Reference

An execution trace of a scheduling program. More...

#include <trace.h>

Inheritance diagram for tvm::tir::TraceNode:

Collaboration diagram for tvm::tir::TraceNode:

Public Member Functions
	TVM_DECLARE_FINAL_OBJECT_INFO (TraceNode, runtime::Object)
Any	GetDecision (const Instruction &inst) const
	Retrieve the decision made on a specific instruction. More...
void	Append (Instruction inst)
	Append a new instruction to the trace. More...
void	Append (Instruction inst, Any decision)
	Append a new instruction with a random decision to the trace. More...
Optional< Instruction >	Pop ()
	Remove the last instruction, along with the decision made on that instruction, if any. More...
void	ApplyToSchedule (Schedule sch, bool remove_postproc, FTraceDecisionProvider decision_provider=nullptr) const
	Apply the trace to a TensorIR schedule. More...
ObjectRef	AsJSON (bool remove_postproc) const
	Serialize the trace as a JSON-style object. More...
Array< String >	AsPython (bool remove_postproc) const
	Serialize the trace as a sequence of python statements. More...
Trace	WithDecision (Instruction inst, Any decision, bool remove_postproc) const
	Create a new trace with an instruction whose decision is changed, assuming this instruction exists in the resulting trace. More...
Trace	Simplified (bool remove_postproc) const
	Simplify the trace with dead-code elimination. More...

Static Public Member Functions
static void	RegisterReflection ()

Public Attributes
Array< Instruction >	insts
	The instructions invoked so far in the program execution. More...
Map< Instruction, Any >	decisions
	The random decisions made upon those instructions. More...

Static Public Attributes
static constexpr const char *	_type_key = "tir.Trace"

Detailed Description

An execution trace of a scheduling program.

A trace has two parts: 1) The instructions invoked so far in the program execution 2) The random decisions made upon those instructions, if any

A trace can be serialized to: 1) Roundtrippable JSON format: can be saved to file and loaded back 2) Python syntax: allows users to copy-paste the trace to reproduce the scheduling process

A trace can be applied to a TensorIR schedule by re-applying all its instructions possibly with their decisions accordingly. Re-sampling is invoked if a sampling instruction doesn't have its corresponding decision; Otherwise the existing decision will be reused accordingly.

Member Function Documentation

Append() [1/2]

void tvm::tir::TraceNode::Append

(

Instruction

inst

)

Append a new instruction to the trace.

Parameters

inst	The new instruction to be appended

Append() [2/2]

void tvm::tir::TraceNode::Append	(	Instruction	inst,
		Any	decision
	)

Append a new instruction with a random decision to the trace.

Parameters

inst	The new instruction to be appended
decision	The random decision made on this instruction The type of decision depends on the instruction, e.g. the decision of SamplePerfectTile has type Array<IntImm>

ApplyToSchedule()

void tvm::tir::TraceNode::ApplyToSchedule	(	Schedule	sch,
		bool	remove_postproc,
		FTraceDecisionProvider	decision_provider = nullptr
	)		const

Apply the trace to a TensorIR schedule.

Parameters

sch	The schedule to be applied onto
remove_postproc	If postprocessing instructions are removed
decision_provider	A callback that allows users to mutate decisions on the fly when applying instructions.

See also

FTraceDecisionProvider

AsJSON()

ObjectRef tvm::tir::TraceNode::AsJSON

(

bool

remove_postproc

)

const

Serialize the trace as a JSON-style object.

Parameters

remove_postproc

If postprocessing instructions are removed

Returns

The JSON-style object

AsPython()

Array<String> tvm::tir::TraceNode::AsPython

(

bool

remove_postproc

)

const

Serialize the trace as a sequence of python statements.

Parameters

remove_postproc

If postprocessing instructions are removed

Returns

A sequence of python statements

GetDecision()

Any tvm::tir::TraceNode::GetDecision

(

const Instruction &

inst

)

const

Retrieve the decision made on a specific instruction.

Parameters

inst	The instruction whose decision is to be retrieved

Returns

The corresponding decision; nullptr if there is no decision made on the instruction

Pop()

Optional<Instruction> tvm::tir::TraceNode::Pop

(

)

Remove the last instruction, along with the decision made on that instruction, if any.

Returns

The instruction removed; std::nullopt if the trace is empty

RegisterReflection()

static void tvm::tir::TraceNode::RegisterReflection ( )

inlinestatic

Simplified()

Trace tvm::tir::TraceNode::Simplified

(

bool

remove_postproc

)

const

Simplify the trace with dead-code elimination.

Parameters

remove_postproc

If postprocessing instructions are removed

Returns

A simplified trace

TVM_DECLARE_FINAL_OBJECT_INFO()

tvm::tir::TraceNode::TVM_DECLARE_FINAL_OBJECT_INFO	(	TraceNode	,
		runtime::Object
	)

WithDecision()

Trace tvm::tir::TraceNode::WithDecision	(	Instruction	inst,
		Any	decision,
		bool	remove_postproc
	)		const

Create a new trace with an instruction whose decision is changed, assuming this instruction exists in the resulting trace.

Parameters

inst	The instruction whose decision is to be changed
decision	The decision to be changed to
remove_postproc	If postprocessing instructions are removed

Returns

The new trace with the decision changed

Member Data Documentation

_type_key

constexpr const char* tvm::tir::TraceNode::_type_key = "tir.Trace"

staticconstexpr

decisions

Map<Instruction, Any> tvm::tir::TraceNode::decisions

The random decisions made upon those instructions.

insts

Array<Instruction> tvm::tir::TraceNode::insts

The instructions invoked so far in the program execution.

---

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

• include/tvm/tir/schedule/trace.h