Managed reference to ScheduleRuleNode. More...

#include <schedule_rule.h>

Inheritance diagram for tvm::meta_schedule::ScheduleRule:

Collaboration diagram for tvm::meta_schedule::ScheduleRule:

Public Member Functions
	TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS (ScheduleRule, ObjectRef, ScheduleRuleNode)

Public Member Functions inherited from tvm::runtime::ObjectRef
	ObjectRef ()=default
	default constructor More...

	ObjectRef (ObjectPtr< Object > data)
	Constructor from existing object ptr. More...

bool	same_as (const ObjectRef &other) const
	Comparator. More...

bool	operator== (const ObjectRef &other) const
	Comparator. More...

bool	operator!= (const ObjectRef &other) const
	Comparator. More...

bool	operator< (const ObjectRef &other) const
	Comparator. More...

bool	defined () const

const Object *	get () const

const Object *	operator-> () const

bool	unique () const

int	use_count () const

template<typename ObjectType >
const ObjectType *	as () const
	Try to downcast the internal Object to a raw pointer of a corresponding type. More...

Static Public Member Functions
static ScheduleRule	AutoInline (bool into_producer, bool into_consumer, bool inline_const_tensor, bool disallow_if_then_else, bool require_injective, bool require_ordered, Optional< Array< String >> disallow_op)
	Create an auto-inline rule that inlines spatial blocks if it satisfies some conditions. More...

static ScheduleRule	MultiLevelTiling (String structure, Optional< Array< String >> tile_binds, Optional< Integer > max_innermost_factor, Optional< Array< Integer >> vector_load_lens, Optional< Map< String, ObjectRef >> reuse_read, Optional< Map< String, ObjectRef >> reuse_write)
	Create a mega rule: multi-level tiling with data reuse. More...

static ScheduleRule	MultiLevelTilingWithIntrin (String intrin_name, String structure, Optional< Array< String >> tile_binds, Optional< Integer > max_innermost_factor, Optional< Array< Integer >> vector_load_lens, Optional< Map< String, ObjectRef >> reuse_read, Optional< Map< String, ObjectRef >> reuse_write)
	Extension of MultiLevelTiling for auto-tensorizing with a single intrinsic. More...

static ScheduleRule	AddRFactor (int max_jobs_per_core, Optional< Integer > max_innermost_factor)
	Create a rule: add-rfactor to some blocks if needed. More...

static ScheduleRule	CrossThreadReduction (Array< Integer > thread_extents)
	Create a schedule rule which applies cross-thread reduction to some reduction blocks correspondingly when needed. More...

static ScheduleRule	RandomComputeLocation ()
	A rule that randomly select a compute-at location for a free block. More...

static ScheduleRule	ParallelizeVectorizeUnroll (int max_jobs_per_core, int max_vectorize_extent, Array< Integer > unroll_max_steps, bool unroll_explicit)
	Mark parallelize, vectorize and unroll to the root block. The mark will be applied to each block in a follow-up post processor. More...

static ScheduleRule	AutoBind (int max_threadblocks, Array< Integer > thread_extents)
	Auto bind loops around the block to BlockIdx and ThreadIdx. More...

static ScheduleRule	PyScheduleRule (PyScheduleRuleNode::FInitializeWithTuneContext f_initialize_with_tune_context, PyScheduleRuleNode::FApply f_apply, PyScheduleRuleNode::FAsString f_as_string)
	Create a schedule rule with customized methods on the python-side. More...

Additional Inherited Members
Public Types inherited from tvm::runtime::ObjectRef
using	ContainerType = Object
	type indicate the container type. More...

Static Public Attributes inherited from tvm::runtime::ObjectRef
static constexpr bool	_type_is_nullable = true

Protected Member Functions inherited from tvm::runtime::ObjectRef
Object *	get_mutable () const

Static Protected Member Functions inherited from tvm::runtime::ObjectRef
template<typename T >
static T	DowncastNoCheck (ObjectRef ref)
	Internal helper function downcast a ref without check. More...

static void	FFIClearAfterMove (ObjectRef *ref)
	Clear the object ref data field without DecRef after we successfully moved the field. More...

template<typename ObjectType >
static ObjectPtr< ObjectType >	GetDataPtr (const ObjectRef &ref)
	Internal helper function get data_ as ObjectPtr of ObjectType. More...

Protected Attributes inherited from tvm::runtime::ObjectRef
ObjectPtr< Object >	data_
	Internal pointer that backs the reference. More...

Detailed Description

Managed reference to ScheduleRuleNode.

See also: ScheduleRuleNode

Member Function Documentation

◆ AddRFactor()

static ScheduleRule tvm::meta_schedule::ScheduleRule::AddRFactor	(	int	max_jobs_per_core,
		Optional< Integer >	max_innermost_factor
	)

static

Create a rule: add-rfactor to some blocks if needed.

Parameters

max_jobs_per_core	The maximum number of jobs to be launched per CPU core. It sets the uplimit of CPU parallelism, i.e. `num_cores * max_jobs_per_core`. Use -1 to disable parallelism.
max_innermost_factor	The maximum size of the innermost factor. NullOpt means no limit

Returns: The schedule rule created

◆ AutoBind()

static ScheduleRule tvm::meta_schedule::ScheduleRule::AutoBind	(	int	max_threadblocks,
		Array< Integer >	thread_extents
	)

static

Auto bind loops around the block to BlockIdx and ThreadIdx.

Parameters

max_threadblocks	The maximum number of threadblock on GPU
thread_extents	Candidates of thread axis extent.

Returns: The schedule rule created

◆ AutoInline()

static ScheduleRule tvm::meta_schedule::ScheduleRule::AutoInline	(	bool	into_producer,
		bool	into_consumer,
		bool	inline_const_tensor,
		bool	disallow_if_then_else,
		bool	require_injective,
		bool	require_ordered,
		Optional< Array< String >>	disallow_op
	)

static

Create an auto-inline rule that inlines spatial blocks if it satisfies some conditions.

Parameters

into_producer	If allows to inline a block into its producer
into_consumer	If allows to inline a block into its consumer
inline_const_tensor	Always inline constant tensors
disallow_if_then_else	Always disallow if-then-else-like constructs
require_ordered	Always require the read-to-write mapping to be ordered
require_injective	Always require the read-to-write mapping to be injective
disallow_op	The operators that are disallowed in auto inline

Returns: The schedule rule created

◆ CrossThreadReduction()

static ScheduleRule tvm::meta_schedule::ScheduleRule::CrossThreadReduction ( Array< Integer > thread_extents )

static

Create a schedule rule which applies cross-thread reduction to some reduction blocks correspondingly when needed.

Parameters

thread_extents Candidates of thread axis extent (values are required to be positive).

Returns: The schedule rule created

◆ MultiLevelTiling()

static ScheduleRule tvm::meta_schedule::ScheduleRule::MultiLevelTiling	(	String	structure,
		Optional< Array< String >>	tile_binds,
		Optional< Integer >	max_innermost_factor,
		Optional< Array< Integer >>	vector_load_lens,
		Optional< Map< String, ObjectRef >>	reuse_read,
		Optional< Map< String, ObjectRef >>	reuse_write
	)

static

Create a mega rule: multi-level tiling with data reuse.

Parameters

structure	The tiling structure. Recommended: 'SSRSRS' on CPU 'SSSRRSRS' on GPU
tile_binds	For each level of tiles, which thread axis it is bound to. Recommended: NullOpt on CPU [blockIdx.x, vthread.x, threadIdx.x] on GPU
max_innermost_factor	The maximum size of the innermost factor. NullOpt means no limit
vector_load_lens	The length of vector lane in vectorized cooperative fetching. NullOpt means disable vectorization
reuse_read	Data reuse configuration for reading. NullOpt means no reuse.
reuse_write	Data reuse configuration for writing. NullOpt means no reuse.

Returns: The schedule rule created

◆ MultiLevelTilingWithIntrin()

static ScheduleRule tvm::meta_schedule::ScheduleRule::MultiLevelTilingWithIntrin	(	String	intrin_name,
		String	structure,
		Optional< Array< String >>	tile_binds,
		Optional< Integer >	max_innermost_factor,
		Optional< Array< Integer >>	vector_load_lens,
		Optional< Map< String, ObjectRef >>	reuse_read,
		Optional< Map< String, ObjectRef >>	reuse_write
	)

static

Extension of MultiLevelTiling for auto-tensorizing with a single intrinsic.

Parameters

intrin_name	The name of a tensor intrinsic, must be registerd via TensorIntrin.register(...) beforehand
structure	The tiling structure. Recommended: 'SSRSRS' on CPU 'SSSRRSRS' on GPU
tile_binds	For each level of tiles, which thread axis it is bound to. Recommended: NullOpt on CPU [blockIdx.x, vthread.x, threadIdx.x] on GPU
max_innermost_factor	The maximum size of the innermost factor. NullOpt means no limit
vector_load_lens	The length of vector lane in vectorized cooperative fetching. NullOpt means disable vectorization
reuse_read	Data reuse configuration for reading. NullOpt means no reuse.
reuse_write	Data reuse configuration for writing. NullOpt means no reuse.

Returns: The schedule rule created

◆ ParallelizeVectorizeUnroll()

static ScheduleRule tvm::meta_schedule::ScheduleRule::ParallelizeVectorizeUnroll	(	int	max_jobs_per_core,
		int	max_vectorize_extent,
		Array< Integer >	unroll_max_steps,
		bool	unroll_explicit
	)

static

Mark parallelize, vectorize and unroll to the root block. The mark will be applied to each block in a follow-up post processor.

Parameters

max_jobs_per_core	The maximum number of jobs to be launched per CPU core. It sets the upper limit of CPU parallelism, i.e. `num_cores * max_jobs_per_core`. Use -1 to disable parallelism.
max_vectorize_extent	The maximum extent to be vectorized. It sets the upper limit of the hardware target vectorization. Use -1 to disable vectorization.
unroll_max_steps	The options of the maximum number of unroll steps to be done. Use an empty array to disable unroll.
unroll_explicit	Whether to explicitly unroll the loop, or just add an "unroll" pragma.

Returns: The schedule rule created

◆ PyScheduleRule()

static ScheduleRule tvm::meta_schedule::ScheduleRule::PyScheduleRule	(	PyScheduleRuleNode::FInitializeWithTuneContext	f_initialize_with_tune_context,
		PyScheduleRuleNode::FApply	f_apply,
		PyScheduleRuleNode::FAsString	f_as_string
	)

static

Create a schedule rule with customized methods on the python-side.

Parameters

f_initialize_with_tune_context	The packed function of `InitializeWithTuneContext`.
f_apply	The packed function of `Apply`.
f_as_string	The packed function of `AsString`.

Returns: The schedule rule created.

◆ RandomComputeLocation()

static ScheduleRule tvm::meta_schedule::ScheduleRule::RandomComputeLocation ( )

static

A rule that randomly select a compute-at location for a free block.

Returns: The schedule rule created

◆ TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS()

tvm::meta_schedule::ScheduleRule::TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS	(	ScheduleRule	,
		ObjectRef	,
		ScheduleRuleNode
	)

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

include/tvm/meta_schedule/schedule_rule.h

Public Member Functions

Static Public Member Functions

Additional Inherited Members

Detailed Description

Member Function Documentation

◆ AddRFactor()

◆ AutoBind()

◆ AutoInline()

◆ CrossThreadReduction()

◆ MultiLevelTiling()

◆ MultiLevelTilingWithIntrin()

◆ ParallelizeVectorizeUnroll()

◆ PyScheduleRule()

◆ RandomComputeLocation()

◆ TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS()