Tensor expression language DSL. More...

Classes
class	BaseComputeOpNode
	A Compute op that compute a tensor on certain domain. This is the base class for ComputeOp (operating on a scalar at a time) and TensorComputeOp (operating on a TensorSlice at a time) More...

class	ComputeOp
	Managed reference to ComputeOpNode. More...

class	ComputeOpNode
	A Compute op that compute a tensor on certain domain. More...

class	ExternOp
	Managed reference to ExternOpNode. More...

class	ExternOpNode
	External computation that cannot be splitted. More...

class	Fuse
	Managed reference to FuseNode. More...

class	FuseNode
	Fuse two domains into one domain. More...

class	HybridOp
	Managed reference to HybridOpNode. More...

class	HybridOpNode
	A computation operator that generated by hybrid script. More...

class	IterVarAttr
	Additional scheduable attributes about IterVar. More...

class	IterVarAttrNode
	node container for IterVar attr More...

class	IterVarRelation
	The schedule relation between IterVars can be Split, Fuse. More...

class	IterVarRelationNode
	base node of iteration var More...

class	Operation
	Operation that produces tensors. More...

class	OperationNode
	Base class of all operation nodes. More...

class	PlaceholderOp
	Managed reference to PlaceholderOpNode. More...

class	PlaceholderOpNode
	A placeholder op represents an input placeholder. More...

class	Rebase
	Managed reference to RebaseNode. More...

class	RebaseNode
	Rebase the iteration to make min to be 0. This is useful to normalize the Schedule to make every leaf variable's min to be 0. More...

class	ScanOp
	Managed reference to ScanOpNode. More...

class	ScanOpNode
	Symbolic scan. More...

class	Schedule
	Global schedule container For operations and all the operations they depend on. The schedule per Operation is named as stage. More...

class	ScheduleContext
	Context helper to collect debug information of Schedule. More...

class	ScheduleNode
	node container for schedule More...

class	Singleton
	Managed reference to SingletonNode. More...

class	SingletonNode
	Singleton iterator [0, 1) More...

class	SpecializedCondition
	Specialized condition to enable op specialization. More...

class	SpecializedConditionNode
	Container for specialization conditions. More...

class	Split
	Managed reference to SplitNode. More...

class	SplitNode
	Split the parent domain into product of outer and iter. More...

class	Stage
	Stage, contains scheduling for a stage of computation. More...

class	StageNode
	represents a stage. More...

class	Tensor
	Tensor structure representing a possible input, or intermediate computation result. More...

class	TensorComputeOp
	Managed reference to TensorComputeOpNode. More...

class	TensorComputeOpNode
	A TenorCompute op that compute a tensor with an tensor intrinsic. More...

struct	TensorDom
	Temporary data structure to store union of bounds of each axis of Tensor. More...

class	TensorIntrin
	Managed reference to TensorIntrinNode. More...

class	TensorIntrinCall
	Managed reference to TensorIntrinCallNode. More...

class	TensorIntrinCallNode

class	TensorIntrinNode
	Node to represent a Tensor intrinsic operator. More...

class	TensorNode
	Node to represent a tensor. More...

class	Transform

class	TransformNode
	Transform iterator according to some arbitrary expression. More...

Typedefs
using	FCompute = std::function< PrimExpr(const Array< Var > &i)>
	The compute function to specify the input source of a Tensor. More...

using	FBatchCompute = std::function< Array< PrimExpr >(const Array< Var > &i)>
	The compute function to specify the inputs source of Tensors. More...

Enumerations
enum	AttachType : int { kGroupRoot = 1, kInline = 2, kInlinedAlready = 3, kScope = 4, kScanUpdate = 5 }
	the attachment type More...

Functions
PrimExpr	Derivative (const PrimExpr &expr, const Var &var)
	Take the derivative of the expression with respect to the given variable. More...

Tensor	Jacobian (const Tensor &output, const Tensor &input)
	Get the tensor representing the Jacobian of the output with respect to the input. More...

Tensor	VectorJacobianProduct (const Tensor &output, const Tensor &input, const Tensor &head)
	The building block for reverse-mode AD. More...

Array< Tensor >	Gradient (const Tensor &output, const Array< Tensor > &inputs, const Tensor &head=Tensor())
	Perform reverse mode automatic differentiation. More...

Var	var (std::string name_hint, DataType t=DataType::Int(32))
	Construct a new Var expression. More...

IterVar	thread_axis (Range dom, std::string tag)
	Create a new IterVar that represents an axis in thread. More...

IterVar	reduce_axis (Range dom, std::string name="rv")
	Create a new IterVar for reduction operations. More...

Tensor	placeholder (Array< PrimExpr > shape, DataType dtype=DataType::Float(32), std::string name="placeholder")
	create a place holder tensor. More...

Tensor	compute (Array< PrimExpr > shape, FCompute fcompute, std::string name="tensor", std::string tag="", Map< String, ObjectRef > attrs={})
	Construct a new tensor by computing over shape, using the computation rule: result_tensor[axis] = fcompute(axis) More...

Array< Tensor >	compute (Array< PrimExpr > shape, FBatchCompute fcompute, std::string name="tensor", std::string tag="", Map< String, ObjectRef > attrs={})
	Construct a new tensor by computing over shape, using the computation rule: result_tensor[axis] = fcompute(axis) More...

Array< Tensor >	scan (Array< Tensor > init, Array< Tensor > update, Array< Tensor > state_placeholder, Array< Tensor > inputs=Array< Tensor >(), std::string name="scan", std::string tag="", Map< String, ObjectRef > attrs={})
	Construct new tensors by scan. More...

Tensor	compute (Array< PrimExpr > shape, std::function< PrimExpr(Var)> f, std::string name="tensor", std::string tag="", Map< String, ObjectRef > attrs={})

Tensor	compute (Array< PrimExpr > shape, std::function< PrimExpr(Var, Var)> f, std::string name="tensor", std::string tag="", Map< String, ObjectRef > attrs={})

Tensor	compute (Array< PrimExpr > shape, std::function< PrimExpr(Var, Var, Var)> f, std::string name="tensor", std::string tag="", Map< String, ObjectRef > attrs={})

Tensor	compute (Array< PrimExpr > shape, std::function< PrimExpr(Var, Var, Var, Var)> f, std::string name="tensor", std::string tag="", Map< String, ObjectRef > attrs={})

Schedule	create_schedule (Array< Operation > ops)
	Create a schedule for array of ops(and their dependencies). More...

void	AutoInlineElemWise (Schedule sch)
	To automatically inline the element-wise operations. More...

void	AutoInlineBroarcast (Schedule sch)
	To automatically inline the broadcast operations. More...

void	AutoInlineInjective (Schedule sch)
	To automatically inline operations with injective writes (i.e. writes without reduction or sequential loops). Note that in this case, guarantees about contiguity, transpose, stride, alignemnt and memory footprint in general do not hold. More...

Map< IterVar, Range >	InferBound (const Schedule &sch)
	Infer the bound of all iteration variables relates to the schedule. More...

bool	VerifyCompactBuffer (const Stmt &stmt)
	Verify if there is any argument bound to compact buffer. More...

Stmt	ScheduleOps (Schedule s, Map< IterVar, Range > dom_map, bool debug_keep_trivial_loop)
	Schedule s' dependent operations. More...

PrimFunc	SchedulePostProcToPrimFunc (Array< ObjectRef > arg_list, Stmt body, Optional< Map< Tensor, Buffer >> bindings)
	Postprocessing the Stmt generated by ScheduleOps to create a PrimFunc that can then be used for further TIR optimizations. More...

PrimExpr	operator! (const Tensor::Slice &a)

PrimExpr	operator- (const Tensor::Slice &a)

template<typename T >
PrimExpr	operator+ (const Tensor::Slice &a, const T &b)

template<typename T >
PrimExpr	operator+ (const T &a, const Tensor::Slice &b)

PrimExpr	operator+ (const Tensor::Slice &a, const Tensor::Slice &b)

template<typename T >
PrimExpr	operator- (const Tensor::Slice &a, const T &b)

template<typename T >
PrimExpr	operator- (const T &a, const Tensor::Slice &b)

PrimExpr	operator- (const Tensor::Slice &a, const Tensor::Slice &b)

template<typename T >
PrimExpr	operator* (const Tensor::Slice &a, const T &b)

template<typename T >
PrimExpr	operator* (const T &a, const Tensor::Slice &b)

PrimExpr	operator* (const Tensor::Slice &a, const Tensor::Slice &b)

template<typename T >
PrimExpr	operator== (const Tensor::Slice &a, const T &b)

template<typename T >
PrimExpr	operator== (const T &a, const Tensor::Slice &b)

PrimExpr	operator== (const Tensor::Slice &a, const Tensor::Slice &b)

template<typename T >
PrimExpr	operator<= (const Tensor::Slice &a, const T &b)

template<typename T >
PrimExpr	operator<= (const T &a, const Tensor::Slice &b)

PrimExpr	operator<= (const Tensor::Slice &a, const Tensor::Slice &b)

template<typename T >
PrimExpr	operator>= (const Tensor::Slice &a, const T &b)

template<typename T >
PrimExpr	operator>= (const T &a, const Tensor::Slice &b)

PrimExpr	operator>= (const Tensor::Slice &a, const Tensor::Slice &b)

template<typename T >
PrimExpr	operator!= (const Tensor::Slice &a, const T &b)

template<typename T >
PrimExpr	operator!= (const T &a, const Tensor::Slice &b)

PrimExpr	operator!= (const Tensor::Slice &a, const Tensor::Slice &b)

template<typename T >
PrimExpr	operator&& (const Tensor::Slice &a, const T &b)

template<typename T >
PrimExpr	operator&& (const T &a, const Tensor::Slice &b)

PrimExpr	operator&& (const Tensor::Slice &a, const Tensor::Slice &b)

template<typename T >
PrimExpr	operator\|\| (const Tensor::Slice &a, const T &b)

template<typename T >
PrimExpr	operator\|\| (const T &a, const Tensor::Slice &b)

PrimExpr	operator\|\| (const Tensor::Slice &a, const Tensor::Slice &b)

template<typename T >
PrimExpr	operator>> (const Tensor::Slice &a, const T &b)

template<typename T >
PrimExpr	operator>> (const T &a, const Tensor::Slice &b)

PrimExpr	operator>> (const Tensor::Slice &a, const Tensor::Slice &b)

template<typename T >
PrimExpr	operator<< (const Tensor::Slice &a, const T &b)

template<typename T >
PrimExpr	operator<< (const T &a, const Tensor::Slice &b)

PrimExpr	operator<< (const Tensor::Slice &a, const Tensor::Slice &b)

template<typename T >
PrimExpr	operator> (const Tensor::Slice &a, const T &b)

template<typename T >
PrimExpr	operator> (const T &a, const Tensor::Slice &b)

PrimExpr	operator> (const Tensor::Slice &a, const Tensor::Slice &b)

template<typename T >
PrimExpr	operator< (const Tensor::Slice &a, const T &b)

template<typename T >
PrimExpr	operator< (const T &a, const Tensor::Slice &b)

PrimExpr	operator< (const Tensor::Slice &a, const Tensor::Slice &b)

Detailed Description

Tensor expression language DSL.

Typedef Documentation

◆ FBatchCompute

using tvm::te::FBatchCompute = typedef std::function<Array<PrimExpr>(const Array<Var>& i)>

The compute function to specify the inputs source of Tensors.

◆ FCompute

using tvm::te::FCompute = typedef std::function<PrimExpr(const Array<Var>& i)>

The compute function to specify the input source of a Tensor.

Enumeration Type Documentation

◆ AttachType

enum tvm::te::AttachType : int

the attachment type

Enumerator
kGroupRoot
kInline
kInlinedAlready
kScope
kScanUpdate

Function Documentation

◆ AutoInlineBroarcast()

void tvm::te::AutoInlineBroarcast ( Schedule sch )

To automatically inline the broadcast operations.

Parameters

sch	The schedule to be inlined.

◆ AutoInlineElemWise()

void tvm::te::AutoInlineElemWise ( Schedule sch )

To automatically inline the element-wise operations.

Parameters

sch	The schedule to be inlined.

◆ AutoInlineInjective()

void tvm::te::AutoInlineInjective ( Schedule sch )

To automatically inline operations with injective writes (i.e. writes without reduction or sequential loops). Note that in this case, guarantees about contiguity, transpose, stride, alignemnt and memory footprint in general do not hold.

Parameters

sch	The schedule to be inlined.

◆ compute() [1/6]

Tensor tvm::te::compute	(	Array< PrimExpr >	shape,
		FCompute	fcompute,
		std::string	name = `"tensor"`,
		std::string	tag = `""`,
		Map< String, ObjectRef >	attrs = `{}`
	)

Construct a new tensor by computing over shape, using the computation rule: result_tensor[axis] = fcompute(axis)

Parameters

shape	Shape of the tensor.
fcompute	The compute function to create the tensor.
name	The optional name of the tensor.
tag	The optional tag of the tensor.
attrs	Optional additional attributes of the compute.

◆ compute() [2/6]

Array<Tensor> tvm::te::compute	(	Array< PrimExpr >	shape,
		FBatchCompute	fcompute,
		std::string	name = `"tensor"`,
		std::string	tag = `""`,
		Map< String, ObjectRef >	attrs = `{}`
	)

Construct a new tensor by computing over shape, using the computation rule: result_tensor[axis] = fcompute(axis)

Parameters

shape	Shape of the tensor.
fcompute	The compute function to create the tensors.
name	The optional name of the tensor.
tag	The optional tag of the tensor.
attrs	Optional additional attributes of the compute.

◆ compute() [3/6]

Tensor tvm::te::compute	(	Array< PrimExpr >	shape,
		std::function< PrimExpr(Var)>	f,
		std::string	name = `"tensor"`,
		std::string	tag = `""`,
		Map< String, ObjectRef >	attrs = `{}`
	)

inline

◆ compute() [4/6]

Tensor tvm::te::compute	(	Array< PrimExpr >	shape,
		std::function< PrimExpr(Var, Var)>	f,
		std::string	name = `"tensor"`,
		std::string	tag = `""`,
		Map< String, ObjectRef >	attrs = `{}`
	)

inline

◆ compute() [5/6]

Tensor tvm::te::compute	(	Array< PrimExpr >	shape,
		std::function< PrimExpr(Var, Var, Var)>	f,
		std::string	name = `"tensor"`,
		std::string	tag = `""`,
		Map< String, ObjectRef >	attrs = `{}`
	)

inline

◆ compute() [6/6]

Tensor tvm::te::compute	(	Array< PrimExpr >	shape,
		std::function< PrimExpr(Var, Var, Var, Var)>	f,
		std::string	name = `"tensor"`,
		std::string	tag = `""`,
		Map< String, ObjectRef >	attrs = `{}`
	)

inline

◆ create_schedule()

Schedule tvm::te::create_schedule ( Array< Operation > ops )

inline

Create a schedule for array of ops(and their dependencies).

Parameters

ops	The ops to be scheduled.

Returns: sch The created Schedule.

◆ Derivative()

PrimExpr tvm::te::Derivative	(	const PrimExpr &	expr,
		const Var &	var
	)

Take the derivative of the expression with respect to the given variable.

Parameters

expr	The expression to differentiate.
var	The variable to differentiate with respect to.

Returns: The expression for the derivative.

◆ Gradient()

Array<Tensor> tvm::te::Gradient	(	const Tensor &	output,
		const Array< Tensor > &	inputs,
		const Tensor &	head = `Tensor()`
	)

Perform reverse mode automatic differentiation.

Each item of the result field of the result is an adjoint for the corresponding item of inputs, i.e. head multiplied by the Jacobian of output with respect to the corresponding item of inputs.

Parameters

output	The tensor to differentiate.
inputs	The array of input tensors. When the array is empty, will perform differentiation wrt all tensors the output depends on.
head	The adjoint of the output, in other words, some tensor, by which the Jacobians will be multiplied (using tensordot axes=`output.shape`). Its shape must be of the form `prefix + output.shape`. If the null pointer is provided, the identity tensor of shape `output.shape + output.shape` will be used.

Returns: An array of adjoints corresponding to inputs.

◆ InferBound()

Map<IterVar, Range> tvm::te::InferBound ( const Schedule & sch )

Infer the bound of all iteration variables relates to the schedule.

Parameters

sch	The root schedule to infer all the bounds.

Returns: the result bound of the iteration Variable

◆ Jacobian()

Tensor tvm::te::Jacobian	(	const Tensor &	output,
		const Tensor &	input
	)

Get the tensor representing the Jacobian of the output with respect to the input.

Note that if output depends on input indirectly (by using some other tensor depending on input), this dependency won't contribute to the resulting Jacobian. For such cases use the function Gradient.

Parameters

output	The tensor to differentiate.
input	The input tensor, which `output` should directly use.

Returns: The tensor representing the Jacobian of shape output.shape + input.shape.

◆ operator!()

PrimExpr tvm::te::operator! ( const Tensor::Slice & a )

inline

◆ operator!=() [1/3]

template<typename T >

PrimExpr tvm::te::operator!=	(	const Tensor::Slice &	a,
		const T &	b
	)

inline

◆ operator!=() [2/3]

template<typename T >

PrimExpr tvm::te::operator!=	(	const T &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator!=() [3/3]

PrimExpr tvm::te::operator!=	(	const Tensor::Slice &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator&&() [1/3]

template<typename T >

PrimExpr tvm::te::operator &&	(	const Tensor::Slice &	a,
		const T &	b
	)

inline

◆ operator&&() [2/3]

template<typename T >

PrimExpr tvm::te::operator &&	(	const T &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator&&() [3/3]

PrimExpr tvm::te::operator &&	(	const Tensor::Slice &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator*() [1/3]

template<typename T >

PrimExpr tvm::te::operator*	(	const Tensor::Slice &	a,
		const T &	b
	)

inline

◆ operator*() [2/3]

template<typename T >

PrimExpr tvm::te::operator*	(	const T &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator*() [3/3]

PrimExpr tvm::te::operator*	(	const Tensor::Slice &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator+() [1/3]

template<typename T >

PrimExpr tvm::te::operator+	(	const Tensor::Slice &	a,
		const T &	b
	)

inline

◆ operator+() [2/3]

template<typename T >

PrimExpr tvm::te::operator+	(	const T &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator+() [3/3]

PrimExpr tvm::te::operator+	(	const Tensor::Slice &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator-() [1/4]

PrimExpr tvm::te::operator- ( const Tensor::Slice & a )

inline

◆ operator-() [2/4]

template<typename T >

PrimExpr tvm::te::operator-	(	const Tensor::Slice &	a,
		const T &	b
	)

inline

◆ operator-() [3/4]

template<typename T >

PrimExpr tvm::te::operator-	(	const T &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator-() [4/4]

PrimExpr tvm::te::operator-	(	const Tensor::Slice &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator<() [1/3]

PrimExpr tvm::te::operator<	(	const Tensor::Slice &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator<() [2/3]

template<typename T >

PrimExpr tvm::te::operator<	(	const Tensor::Slice &	a,
		const T &	b
	)

inline

◆ operator<() [3/3]

template<typename T >

PrimExpr tvm::te::operator<	(	const T &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator<<() [1/3]

template<typename T >

PrimExpr tvm::te::operator<<	(	const T &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator<<() [2/3]

template<typename T >

PrimExpr tvm::te::operator<<	(	const Tensor::Slice &	a,
		const T &	b
	)

inline

◆ operator<<() [3/3]

PrimExpr tvm::te::operator<<	(	const Tensor::Slice &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator<=() [1/3]

template<typename T >

PrimExpr tvm::te::operator<=	(	const Tensor::Slice &	a,
		const T &	b
	)

inline

◆ operator<=() [2/3]

PrimExpr tvm::te::operator<=	(	const Tensor::Slice &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator<=() [3/3]

template<typename T >

PrimExpr tvm::te::operator<=	(	const T &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator==() [1/3]

template<typename T >

PrimExpr tvm::te::operator==	(	const Tensor::Slice &	a,
		const T &	b
	)

inline

◆ operator==() [2/3]

template<typename T >

PrimExpr tvm::te::operator==	(	const T &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator==() [3/3]

PrimExpr tvm::te::operator==	(	const Tensor::Slice &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator>() [1/3]

PrimExpr tvm::te::operator>	(	const Tensor::Slice &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator>() [2/3]

template<typename T >

PrimExpr tvm::te::operator>	(	const T &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator>() [3/3]

template<typename T >

PrimExpr tvm::te::operator>	(	const Tensor::Slice &	a,
		const T &	b
	)

inline

◆ operator>=() [1/3]

PrimExpr tvm::te::operator>=	(	const Tensor::Slice &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator>=() [2/3]

template<typename T >

PrimExpr tvm::te::operator>=	(	const Tensor::Slice &	a,
		const T &	b
	)

inline

◆ operator>=() [3/3]

template<typename T >

PrimExpr tvm::te::operator>=	(	const T &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator>>() [1/3]

template<typename T >

PrimExpr tvm::te::operator>>	(	const T &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator>>() [2/3]

template<typename T >

PrimExpr tvm::te::operator>>	(	const Tensor::Slice &	a,
		const T &	b
	)

inline

◆ operator>>() [3/3]

PrimExpr tvm::te::operator>>	(	const Tensor::Slice &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator||() [1/3]

PrimExpr tvm::te::operator\|\|	(	const Tensor::Slice &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator||() [2/3]

template<typename T >

PrimExpr tvm::te::operator\|\|	(	const T &	a,
		const Tensor::Slice &	b
	)

inline

◆ operator||() [3/3]

template<typename T >

PrimExpr tvm::te::operator\|\|	(	const Tensor::Slice &	a,
		const T &	b
	)

inline

◆ placeholder()

Tensor tvm::te::placeholder	(	Array< PrimExpr >	shape,
		DataType	dtype = `DataType::Float(32)`,
		std::string	name = `"placeholder"`
	)

create a place holder tensor.

Parameters

shape	The shape of the tensor.
dtype	the data type of the tensor.
name	The name of the Tensor.

◆ reduce_axis()

IterVar tvm::te::reduce_axis	(	Range	dom,
		std::string	name = `"rv"`
	)

Create a new IterVar for reduction operations.

Parameters

dom	The domain of the reduction axis.
name	The name of the reduction axis.

◆ scan()

Array<Tensor> tvm::te::scan	(	Array< Tensor >	init,
		Array< Tensor >	update,
		Array< Tensor >	state_placeholder,
		Array< Tensor >	inputs = `Array< Tensor >()`,
		std::string	name = `"scan"`,
		std::string	tag = `""`,
		Map< String, ObjectRef >	attrs = `{}`
	)

Construct new tensors by scan.

Parameters

init	The intialize tensor of first K steps.
update	The update tensor indicated the updated result after each timestamp.
state_placeholder	The placeholder for the states.
inputs	The inputs to the scan body, this is optional, but recommended to provide concrete information about scan body.
name	The optional name of the tensor.
tag	The optional tag of the tensor.
attrs	Optional additional attributes of the compute.

◆ ScheduleOps()

Stmt tvm::te::ScheduleOps	(	Schedule	s,
		Map< IterVar, Range >	dom_map,
		bool	debug_keep_trivial_loop
	)

Schedule s' dependent operations.

Parameters

s	The schedule to be realized
dom_map	The domain of each iter vars.
debug_keep_trivial_loop	Whether keep trivial loops with extent of 1 during lowering. This is a debug feature for dataflow/axis analysis. Note: If this is true, The lowered IR may be incorrect, because we will also delete the init part of reduction

Returns: the result Stmt

◆ SchedulePostProcToPrimFunc()

PrimFunc tvm::te::SchedulePostProcToPrimFunc	(	Array< ObjectRef >	arg_list,
		Stmt	body,
		Optional< Map< Tensor, Buffer >>	bindings
	)

Postprocessing the Stmt generated by ScheduleOps to create a PrimFunc that can then be used for further TIR optimizations.

Perform this translation before running any TIR optimizations.

List of actions taken by the function:

Remove occurrences of te::Tensor, te::Operation in the IR and replace them by corresponding IR nodes via tir::Buffer.
Add annotation of extern buffers using the buffer_map field in the PrimFunc type.

Parameters

arg_list	Array of Tensor/Var/Buffer arguments to the function.
body	The body of the function.
bindings	potential Tensor to Buffer bindings for the Tensors in the body.

◆ thread_axis()

IterVar tvm::te::thread_axis	(	Range	dom,
		std::string	tag
	)

Create a new IterVar that represents an axis in thread.

Parameters

dom	Optional, domain of the thread axis.
tag	The thread tag of the axis.

◆ var()

Var tvm::te::var	(	std::string	name_hint,
		DataType	t = `DataType::Int(32)`
	)

Construct a new Var expression.

Parameters

name_hint	The name hint for the expression
t	The type of the expression

◆ VectorJacobianProduct()

Tensor tvm::te::VectorJacobianProduct	(	const Tensor &	output,
		const Tensor &	input,
		const Tensor &	head
	)

The building block for reverse-mode AD.

Differentiate output wrt input and multiply the result by head on the left using tensor dot product. input must be an immediate dependency of output (must be called from within the body of output). That is, the function will compute one summand of the adjoint for input given the adjoint for output (which is called head here).

Parameters

output	The tensor to differentiate.
input	The input tensor, which `output` should directly use.
head	The adjoint of `output`. Must be of shape `prefix + output.shape`

Returns: The tensor of shape prefix + input.shape representing the partial adjoint of input wrt one of its consumers (output)

◆ VerifyCompactBuffer()

bool tvm::te::VerifyCompactBuffer ( const Stmt & stmt )

Verify if there is any argument bound to compact buffer.

Parameters

stmt	The stmt to be verified.

Returns: true if there is any buffer_bind_scope attribute found, otherwise, false.

Classes

Typedefs

Enumerations

Functions

Detailed Description

Typedef Documentation

◆ FBatchCompute

◆ FCompute

Enumeration Type Documentation

◆ AttachType

Function Documentation

◆ AutoInlineBroarcast()

◆ AutoInlineElemWise()

◆ AutoInlineInjective()

◆ compute() [1/6]

◆ compute() [2/6]

◆ compute() [3/6]

◆ compute() [4/6]

◆ compute() [5/6]

◆ compute() [6/6]

◆ create_schedule()

◆ Derivative()

◆ Gradient()

◆ InferBound()

◆ Jacobian()

◆ operator!()

◆ operator!=() [1/3]

◆ operator!=() [2/3]

◆ operator!=() [3/3]

◆ operator&&() [1/3]

◆ operator&&() [2/3]

◆ operator&&() [3/3]

◆ operator*() [1/3]

◆ operator*() [2/3]

◆ operator*() [3/3]

◆ operator+() [1/3]

◆ operator+() [2/3]

◆ operator+() [3/3]

◆ operator-() [1/4]

◆ operator-() [2/4]

◆ operator-() [3/4]

◆ operator-() [4/4]

◆ operator<() [1/3]

◆ operator<() [2/3]

◆ operator<() [3/3]

◆ operator<<() [1/3]

◆ operator<<() [2/3]

◆ operator<<() [3/3]

◆ operator<=() [1/3]

◆ operator<=() [2/3]

◆ operator<=() [3/3]

◆ operator==() [1/3]

◆ operator==() [2/3]

◆ operator==() [3/3]

◆ operator>() [1/3]

◆ operator>() [2/3]

◆ operator>() [3/3]

◆ operator>=() [1/3]

◆ operator>=() [2/3]

◆ operator>=() [3/3]

◆ operator>>() [1/3]

◆ operator>>() [2/3]

◆ operator>>() [3/3]

◆ operator||() [1/3]

◆ operator||() [2/3]

◆ operator||() [3/3]

◆ placeholder()

◆ reduce_axis()

◆ scan()

◆ ScheduleOps()

◆ SchedulePostProcToPrimFunc()

◆ thread_axis()

◆ var()

◆ VectorJacobianProduct()

◆ VerifyCompactBuffer()