tvm::s_tir Namespace Reference

Namespaces
namespace	attr
namespace	backend
namespace	meta_schedule
namespace	transform

Classes
class	Instruction
	Managed reference to InstructionNode. More...
class	InstructionKind
	Managed reference to InstructionKindObj. More...
class	InstructionKindDef
	Temporary builder for a canonical named instruction kind. More...
class	InstructionKindObj
	Kind of an instruction, e.g. Split, Reorder, etc. Besides the name, every kind of instruction has its own properties, including: 1) A boolean indicating if the instruction is pure, i.e. change nothing in the schedule state 2) A functor that applies the instruction to a TensorIR schedule 3) A functor that converts the instruction to a statement in python syntax 4) A functor that serialize its attributes to JSON 5) A functor that deserialize its attributes from JSON. More...
class	InstructionNode
	Schedule instructions each corresponds to a schedule primitive. More...
class	IterVar
	Iteration Variable, represents an iteration over an integer interval. More...
class	IterVarNode
	An iteration variable representing an iteration over a one dimensional interval. More...
class	LinearCongruentialEngine
	This linear congruential engine is a drop-in replacement for std::minstd_rand. It strictly corresponds to std::minstd_rand and is designed to be platform-independent. More...
class	LoopRV
	Managed reference to LoopRVNode. More...
class	LoopRVNode
	A random variable that evaluates to a TensorIR for loop. More...
class	MatchTensorRegion
	Managed reference to MatchTensorRegionNode. More...
class	MatchTensorRegionNode
	Match introduces a constraint that the source tensor region can be remapped to the data layout specified by the tensor field. The constraint can be checked in later part of lowering (or optionally during runtime). More...
struct	MemCpyDetails
	Helper struct for return value of IdentifyMemCpy. More...
class	SBlock
	Managed reference to SBlockNode. More...
struct	SBlockInfo
	The information about a TensorIR block, it contains two categories of information 1) Info on the block scope rooted at a specific block, including dependency tracking, flags indicating if the scope is a stage pipeline, etc. 2) Info on the block itself, including if the block has a quasi-affine binding, if the regions it reads are completely covered by their producers, etc. More...
class	SBlockNode
	A block is a basic schedule unit in TIR. More...
class	SBlockRealize
	Managed reference to BlockRealizeNode. More...
class	SBlockRealizeNode
	A block realization node represents execution of the block at the binding values. More...
class	SBlockRV
	Managed reference to SBlockRVNode. More...
class	SBlockRVNode
	A random variable that evaluates to a TensorIR block. More...
class	Schedule
	Managed reference to ScheduleObj. More...
class	ScheduleObj
	The user-facing schedule class. More...
class	ScheduleState
	Managed reference to ScheduleStateObj. More...
class	ScheduleStateObj
	The state of scheduling, which exposes a Replace method as the primary interface for all the scheduling primitives to manipulate the TensorIR. More...
class	StmtExprMutator
	Extend native TIRX mutation while preserving block iterator binders. More...
class	StmtExprMutatorWithAnalyzer
class	StmtExprVisitor
	Extend native TIRX traversal with schedulable block semantics. More...
class	StmtExprVisitorWithAnalyzer
class	StmtFunctor
	Extend TIRX statement dispatch with schedulable blocks. More...
class	StmtFunctor< R(const Stmt &, Args...)>
class	TensorIntrin
	Managed reference to TensorIntrinNode. More...
class	TensorIntrinNode
	Tensor intrinsics for tensorization. More...
class	Trace
	Managed reference to TraceNode. More...
class	TraceNode
	An execution trace of a scheduling program. More...

Typedefs
using	FInstructionApply = ffi::TypedFunction< ffi::Array< Any >(Schedule sch, const ffi::Array< Any > &inputs, const ffi::Array< Any > &attrs, const Any &decision)>
	Type of the functor that applies the instruction to a TensorIR schedule.
using	FInstructionAsPython = ffi::TypedFunction< ffi::String(const ffi::Array< Any > &inputs, const ffi::Array< Any > &attrs, const Any &decision, const ffi::Array< ffi::String > &outputs)>
	Type of the functor that converts the instruction to a statement in python syntax.
using	FInstructionAttrsAsJSON = ffi::TypedFunction< ffi::ObjectRef(ffi::Array< Any > attrs)>
	Type of the functor that serialize its attributes to JSON.
using	FInstructionAttrsFromJSON = ffi::TypedFunction< ffi::Array< Any >(ffi::ObjectRef json_attrs)>
	Type of the functor that deserialize its attributes from JSON.
using	ExprRV = PrimExpr
	An expr random variable.
using	ExprRVNode = ExprNode
using	FTraceDecisionProvider = ffi::TypedFunction< Any(const Instruction &inst, const ffi::Array< Any > &inputs, const ffi::Array< Any > &attrs, const Any &decision)>
	A callback that allows users to mutate decisions on the fly when applying instructions. The signature of the callback is:
using	TensorAllocatedAddresses = ffi::Array< ffi::Tuple< Var, ffi::Array< PrimExpr > > >
	Placement entries owned by a block, traversable with their tensor references.

Enumerations
enum	IterVarType : int {   kDataPar = 0 , kThreadIndex = 1 , kCommReduce = 2 , kOrdered = 3 ,   kOpaque = 4 , kUnrolled = 5 , kVectorized = 6 , kParallelized = 7 ,   kTensorized = 8 }
	Type of iteration variable. Each IterVar have a specific type. More...
enum class	ScheduleErrorRenderLevel : int32_t { kDetail = 0 , kFast = 1 , kNone = 2 }
	The level of detailed error message rendering. More...
enum class	TensorIndexType : int32_t { kRead = 0 , kWrite = 1 }
	Type of tensor index. More...
enum	ScheduleDebugMask : uint32_t { kVerifySRefTree = 1 , kVerifyCachedFlags = 2 }
	The bitmask of the debug flag in the ScheduleStateObj. More...

Functions
bool	VerifyWellFormed (const tirx::Function &func, bool assert_mode=true)
	Verify variable/tensor definitions, load types and schedulable block boundaries.
bool	VerifyWellFormed (const IRModule &mod, bool assert_mode=true)
	Verify S-TIR or mixed modules, including definitions shared across functions.
double	EstimateTIRFlops (const Stmt &stmt)
	Estimate the FLOPs of a TIR fragment.
double	EstimateTIRFlops (const IRModule &mod)
	Estimate the FLOPs of TIRs in an IRModule.
bool	IsPureFunction (const Function &func, bool assert_on_error=false)
	Analyze the side effect of a function.
bool	VerifyGPUCode (const Function &func, ffi::Map< ffi::String, PrimExpr > constraints)
	Verify the correctness of a GPU code.
std::optional< MemCpyDetails >	IdentifyMemCpy (const For &loop, const sym::Analyzer &analyzer)
	Identify whether a For loop is semantically equivalent to MemCpy.
ffi::Array< ffi::Optional< Range > >	DomainTouched (const Stmt &body, const TensorVar &tensor, bool consider_loads, bool consider_stores)
	Infer the domain touched by tensor accesses within a statement.
ffi::Map< ffi::String, ffi::Map< ffi::String, int64_t > >	CalculateAllocatedBytes (const Function &func)
	Calculate the allocated memory per scope in bytes needed inside the TIR Function.
ffi::Map< ffi::String, ffi::Map< ffi::String, int64_t > >	CalculateAllocatedBytes (const IRModule &mod)
	Calculate the allocated memory per scope in bytes for each function inside the module.
ffi::Array< tvm::transform::Pass >	GetVTCMCompactionPasses ()
	Get the list of lowering passes to calculate the compacted VTCM allocation size.
bool	VerifyVTCMLimit (const IRModule &mod, int64_t limit)
	Verifies that the VTCM usage for all functions in the given IRModule.
bool	VerifyVTCMLimit (const Function &func, int64_t limit)
	Verifies that the VTCM usage of the given function is within the provided limit.
const char *	IterVarType2String (IterVarType t)
const Op &	async_copy_scope ()
	Region marking copies eligible for asynchronous lowering.
const Op &	async_commit ()
	Commit the preceding asynchronous copies to a queue.
const Op &	async_wait ()
	Wait until at most the given number of committed groups remain in flight.
const Op &	manual_sync ()
	Region whose author manages cross-thread synchronization explicitly.
ffi::Optional< ffi::String >	GetThreadBinding (const ForNode *loop)
	Interpret and validate S-TIR thread placement on a shared loop.
ffi::Optional< ffi::String >	GetThreadBinding (const For &loop)

Typedef Documentation

ExprRV

using tvm::s_tir::ExprRV = typedef PrimExpr

An expr random variable.

ExprRVNode

using tvm::s_tir::ExprRVNode = typedef ExprNode

FInstructionApply

using tvm::s_tir::FInstructionApply = typedef ffi::TypedFunction<ffi::Array<Any>(Schedule sch, const ffi::Array<Any>& inputs, const ffi::Array<Any>& attrs, const Any& decision)>

Type of the functor that applies the instruction to a TensorIR schedule.

Parameters

sch	The schedule to be applied on
inputs	The input random variables
attrs	Instruction attributes
decision	Decisions made on the instruction

Returns

The functor returns an array of output random variables

FInstructionAsPython

using tvm::s_tir::FInstructionAsPython = typedef ffi::TypedFunction<ffi::String(const ffi::Array<Any>& inputs, const ffi::Array<Any>& attrs, const Any& decision, const ffi::Array<ffi::String>& outputs)>

Type of the functor that converts the instruction to a statement in python syntax.

Parameters

inputs	Names of the input random variables
attrs	Instruction attributes
decisions	Decisions made on the instruction
outputs	Names of the output random variables

Returns

A string representing the python api call

FInstructionAttrsAsJSON

using tvm::s_tir::FInstructionAttrsAsJSON = typedef ffi::TypedFunction<ffi::ObjectRef(ffi::Array<Any> attrs)>

Type of the functor that serialize its attributes to JSON.

Parameters

attrs

The attributes to be serialized

Returns

An array, serialized attributes

Note

This functor is nullable

FInstructionAttrsFromJSON

using tvm::s_tir::FInstructionAttrsFromJSON = typedef ffi::TypedFunction<ffi::Array<Any>(ffi::ObjectRef json_attrs)>

Type of the functor that deserialize its attributes from JSON.

Parameters

json_attrs

The attributes to be serialized

Returns

An array, deserialized attributes

Note

This functor is nullable

FTraceDecisionProvider

using tvm::s_tir::FTraceDecisionProvider = typedef ffi::TypedFunction<Any(const Instruction& inst, const ffi::Array<Any>& inputs, const ffi::Array<Any>& attrs, const Any& decision)>

A callback that allows users to mutate decisions on the fly when applying instructions. The signature of the callback is:

Parameters

inst	The instruction
inputs	The input random variables
attrs	The attributes
decision	The original decision

Returns

A new decision

TensorAllocatedAddresses

using tvm::s_tir::TensorAllocatedAddresses = typedef ffi::Array<ffi::Tuple<Var, ffi::Array<PrimExpr> >>

Placement entries owned by a block, traversable with their tensor references.

Enumeration Type Documentation

IterVarType

enum tvm::s_tir::IterVarType : int

Type of iteration variable. Each IterVar have a specific type.

The type of iter var can be overriden via stage.iter_var_attrs given they are compatible.

Enumerator
kDataPar	Data parallel iteration. This normally corresponds to axis of Tensor. Allow all IterVar manipulations. Note This does not mean the loop have to be executed in parallel fashion.
kThreadIndex	The IterVar itself is a thread-index of a fixed thread launching group. Note that this is already assumed to be parallelized. Disallow: split/fuse/vectorize/parallel
kCommReduce	Communicative reduction. Cannot be directly parallelized. Disallow: parallel/vectorize
kOrdered	Serial loops with loop carry dependency, the iteration must execute in order. Cannot be re-ordered. Disallow: reorder/parallel/vectorize
kOpaque	IterVar is opaque,. May not corresponds to any generated loop Disallow all IterVar manipulations and compute_at Note This is usually used to implement composite op or external op, where the
kUnrolled	The execution is unrolled.
kVectorized	The loop is vectorized.
kParallelized	The loop is parallelized.
kTensorized	Marks boundary of tensorization intrinsic.

ScheduleDebugMask

enum tvm::s_tir::ScheduleDebugMask : uint32_t

The bitmask of the debug flag in the ScheduleStateObj.

See also

ScheduleStateObj

Enumerator
kVerifySRefTree	Verify the correctness of the sref tree.
kVerifyCachedFlags	Verify the correctness of affine_binding, region_cover and stage_pipeline.

ScheduleErrorRenderLevel

enum class tvm::s_tir::ScheduleErrorRenderLevel : int32_t

strong

The level of detailed error message rendering.

Enumerator
kDetail	Render a detailed error message.
kFast	Render the error in fast mode.
kNone	No error message at all.

TensorIndexType

enum class tvm::s_tir::TensorIndexType : int32_t

strong

Type of tensor index.

Enumerator
kRead	Index of a read tensor.
kWrite	Index of a written tensor.

Function Documentation

async_commit()

const Op & tvm::s_tir::async_commit

(

)

Commit the preceding asynchronous copies to a queue.

async_copy_scope()

const Op & tvm::s_tir::async_copy_scope

(

)

Region marking copies eligible for asynchronous lowering.

async_wait()

const Op & tvm::s_tir::async_wait

(

)

Wait until at most the given number of committed groups remain in flight.

CalculateAllocatedBytes() [1/2]

ffi::Map< ffi::String, ffi::Map< ffi::String, int64_t > > tvm::s_tir::CalculateAllocatedBytes

(

const Function &

func

)

Calculate the allocated memory per scope in bytes needed inside the TIR Function.

Parameters

func	The TIR Function for which the allocated memory size to be calculated

Returns

Allocated memory size per scope in bytes.

CalculateAllocatedBytes() [2/2]

ffi::Map< ffi::String, ffi::Map< ffi::String, int64_t > > tvm::s_tir::CalculateAllocatedBytes

(

const IRModule &

mod

)

Calculate the allocated memory per scope in bytes for each function inside the module.

Parameters

mod	The IRModule for which the allocated memory size has to be calculated

Returns

Allocated memory size per scope in bytes for each function.

DomainTouched()

ffi::Array< ffi::Optional< Range > > tvm::s_tir::DomainTouched	(	const Stmt &	body,
		const TensorVar &	tensor,
		bool	consider_loads,
		bool	consider_stores
	)

Infer the domain touched by tensor accesses within a statement.

Parameters

body	The statement to analyze.
tensor	The tensor whose accesses are analyzed.
consider_loads	Whether to include loads.
consider_stores	Whether to include stores.

Returns

The domain covering the selected accesses.

EstimateTIRFlops() [1/2]

double tvm::s_tir::EstimateTIRFlops

(

const IRModule &

mod

)

Estimate the FLOPs of TIRs in an IRModule.

Parameters

mod	The IRModule to be estimated.

Returns

The estimated FLOPs.

EstimateTIRFlops() [2/2]

double tvm::s_tir::EstimateTIRFlops

(

const Stmt &

stmt

)

Estimate the FLOPs of a TIR fragment.

Parameters

stmt	The TIR fragment to be estimated.

Returns

The estimated FLOPs.

GetThreadBinding() [1/2]

ffi::Optional< ffi::String > tvm::s_tir::GetThreadBinding ( const For &  loop )

inline

GetThreadBinding() [2/2]

ffi::Optional< ffi::String > tvm::s_tir::GetThreadBinding ( const ForNode *  loop )

inline

Interpret and validate S-TIR thread placement on a shared loop.

GetVTCMCompactionPasses()

ffi::Array< tvm::transform::Pass > tvm::s_tir::GetVTCMCompactionPasses

(

)

Get the list of lowering passes to calculate the compacted VTCM allocation size.

Returns

The list of passes.

IdentifyMemCpy()

std::optional< MemCpyDetails > tvm::s_tir::IdentifyMemCpy	(	const For &	loop,
		const sym::Analyzer &	analyzer
	)

Identify whether a For loop is semantically equivalent to MemCpy.

Parameters

loop	The loop to be checked
analyzer	The analyzer with which to check any algebraic expressions

Returns

The source and destination regions being copied, if the loop is equivalent to memcpy.

IsPureFunction()

bool tvm::s_tir::IsPureFunction	(	const Function &	func,
		bool	assert_on_error = false
	)

Analyze the side effect of a function.

Parameters

func	The function to be checked.
assert_on_error	If true, an error will be thrown for an impure function.

Returns

The purity of the function.

IterVarType2String()

const char * tvm::s_tir::IterVarType2String ( IterVarType  t )

inline

manual_sync()

const Op & tvm::s_tir::manual_sync

(

)

Region whose author manages cross-thread synchronization explicitly.

VerifyGPUCode()

bool tvm::s_tir::VerifyGPUCode	(	const Function &	func,
		ffi::Map< ffi::String, PrimExpr >	constraints
	)

Verify the correctness of a GPU code.

Parameters

func	The function to be checked.
constraints	The dict to specify constraints to check.

Returns

valid Whether it is a valid GPU code.

VerifyVTCMLimit() [1/2]

bool tvm::s_tir::VerifyVTCMLimit	(	const Function &	func,
		int64_t	limit
	)

Verifies that the VTCM usage of the given function is within the provided limit.

Parameters

func	The function to be checked.
limit	The limit to check.

Returns

true if the VTCM usage is within the provided limit.

VerifyVTCMLimit() [2/2]

bool tvm::s_tir::VerifyVTCMLimit	(	const IRModule &	mod,
		int64_t	limit
	)

Verifies that the VTCM usage for all functions in the given IRModule.

Parameters

mod	The module to be checked.
limit	The limit to check.

Returns

true if the VTCM usage is within the provided limit.

VerifyWellFormed() [1/2]

bool tvm::s_tir::VerifyWellFormed	(	const IRModule &	mod,
		bool	assert_mode = true
	)

Verify S-TIR or mixed modules, including definitions shared across functions.

VerifyWellFormed() [2/2]

bool tvm::s_tir::VerifyWellFormed	(	const tirx::Function &	func,
		bool	assert_mode = true
	)

Verify variable/tensor definitions, load types and schedulable block boundaries.