Typedefs
using	Pass = tvm::transform::Pass

using	PassNode = tvm::transform::PassNode

using	PassInfo = tvm::transform::PassInfo

using	PassInfoNode = tvm::transform::PassInfoNode

using	PassContext = tvm::transform::PassContext

using	PassContextNode = tvm::transform::PassContextNode

using	Sequential = tvm::transform::Sequential

using	FTVMRelayToTIR = tvm::transform::Pass
	RelayToTIR tvm::transform::Pass specific to a TargetKind. More...

using	FTVMTIRToRuntime = tvm::runtime::TypedPackedFunc< runtime::Module(IRModule, Target)>
	TIRToRuntime conversion specific to a TargetKind. More...

Functions
Pass	CreateFunctionPass (const runtime::TypedPackedFunc< Function(Function, IRModule, PassContext)> &pass_func, int opt_level, String name, tvm::Array< String > required, bool traceable=false)

Pass	DeadCodeElimination (bool inline_once=false, bool ignore_purity=false)
	Remove let-bound expressions which do not effect the program result. More...

Pass	LazyGradientInit ()
	Convert all expressions of TensorType into GradCell, an algebraic data type defined in gradient.rly. More...

Pass	FoldConstant (bool fold_qnn=false)
	Fold constant expressions. More...

Pass	SplitArgs (uint64_t max_function_args)
	Split function with huge number of arguments to smaller pieces. More...

Pass	FuseOps (int fuse_opt_level=-1)
	Fuse operations into expr into separate functions. More...

Pass	DefuseOps ()
	The inverse operation of FuseOps. It transforms a fused program returned by FuseOps into the program before FuseOps. (i.e. x == DefuseOps(FuseOps(x))) More...

Pass	RewriteAnnotatedOps (int fallback_device)
	Rewrite the annotated program. More...

Pass	ToBasicBlockNormalForm ()
	Turn an expression to Basic Block Normal Form. More...

Pass	ToANormalForm ()
	turn a dataflow graph into Administrative Normal Form, or A-Normal Form (ANF). More...

Expr	ToANormalForm (const Expr &expr)
	ToANormalForm but on incomplete graph. More...

Pass	ToCPS ()
	Turn an expression into continuation passing style(CPS). More...

Pass	ToGraphNormalForm ()
	Remove let binding and directly share via pointer instead. More...

Pass	PartialEval ()
	Aggressive constant propagation/constant folding/inlining. More...

Pass	SimplifyInference ()
	Simplify certain operators during inference. For example, the result of a batch norm which is indexed at tuple index 0 will be unpacked into a number of simplified operators. More...

Pass	FastMath ()
	Replaces non linear activation functions with their fast but approximate counterparts. More...

Pass	DynamicToStatic ()
	Find Dynamic ops and make them static. More...

Pass	InferType ()
	Infer the type of an expression. More...

Type	InferTypeLocal (const Expr &expr)
	Infer the type of an expression, reusing existing type information. More...

Pass	EliminateCommonSubexpr (runtime::PackedFunc fskip=nullptr)
	Search and eliminate common subexpression. For example, if there are two expressions evaluated to an identical value, a single variable is created and these two expressions are replaced by this variable. More...

Pass	CombineParallelConv2D (uint64_t min_num_branches=3)
	Combine parallel 2d convolutions into a single convolution if the number of branches of this conv2d operator is not less than `min_num_branch`. More...

Pass	CombineParallelDense (uint64_t min_num_branches=3, bool to_batch_matmul=true)
	Combine parallel dense ops into a single batch_matmul if the number of branches of this dense operator is not less than `min_num_branch`. More...

Pass	CombineParallelBatchMatmul (uint64_t min_num_branches=3)
	Combine parallel batch_matmul ops into a single batch_matmul if the number of branches of this dense operator is not less than `min_num_branch`. More...

Pass	BackwardFoldScaleAxis ()
	Backward fold axis scaling into weights of conv/dense operators. More...

Pass	ForwardFoldScaleAxis ()
	Forward fold axis scaling into weights of conv/dense operators. More...

Pass	FoldScaleAxis ()
	A sequential pass that executes ForwardFoldScaleAxis and BackwardFoldScaleAxis passes. More...

Pass	CanonicalizeOps ()
	Canonicalize some operators to the simplified operators. For example, bias_add can be canonicalized to expand_dims and broadcast_add. More...

Pass	AlterOpLayout ()
	Alternate the layouts of operators or replace primitive operators with other expressions. More...

Pass	AutoSchedulerLayoutRewrite ()
	Do layout rewrite according to the tile structure created by auto-scheduler. More...

Pass	MetaScheduleLayoutRewrite ()
	Do layout rewrite according to the tile structure created by meta-schedule. More...

Pass	ConvertLayout (const Map< String, Array< String >> &desired_layouts)
	Given a dest layout, this pass transforms the expr such that most of the ops input data layout is changed to the dest layout. In ideal situation, there are only 2 layout transforms, one at the start and one at the end. More...

Pass	Legalize (const String &legalize_map_attr_name="FTVMLegalize")
	Legalizes an expr with another expression. More...

Pass	CanonicalizeCast ()
	Canonicalize cast expressions to make operator fusion more efficient. More...

Pass	EtaExpand (bool expand_constructor, bool expand_global_var)
	Add abstraction over a constructor or global variable bound to a function. More...

Pass	PartitionGraph ()
	Partition a Relay program into regions that can be executed on different backends. More...

Pass	Inline ()
	Inline the global functions marked as `inline` in a given Relay IRModule. More...

Pass	RemoveUnusedFunctions (Array< runtime::String > entry_functions)
	Remove the unused functions in the Relay IRModule. More...

Pass	SimplifyExpr ()
	Simplify the Relay expression. More...

Pass	SimplifyExprPostAlterOp ()
	Stripped down version of SimplifyExpr which is run after AlterOpLayout. More...

Pass	RelayToTIRTargetHook (CompilationConfig config)
	Run any custom passes registered under "RelayToTIR" attributes on TargetKinds. More...

Pass	ManifestAlloc (VirtualDevice cpu_virtual_device)
	A pass for manifesting explicit memory allocations and rewriting specific dialects. More...

Pass	ManifestLifetimes ()
	A pass for manifesting variable lifetimes by inserting kill operations when variables become dead. This pass should be run after ManifestAlloc, and should not be run more than once. More...

Pass	PlanDevices (CompilationConfig config)
	Uses existing "on_device" and "device_copy" CallNodes to infer the `VirtualDevice` on which every Relay sub-expression should run and the result stored. Captures the result of that analysis using new "on_device" and "device_copy" CallNodes. More...

Pass	FlattenAtrousConv ()
	This transform flattens atrous convolution, which corresponds to the sequence of operations: "space_to_batch_nd"->"conv2d"->"batch_to_space_nd" and convert them into subgraphs with a convolution with the modified "dilation" and recalculated "padding" parameters. More...

Pass	AnnotateUsedMemory ()
	Annotates the minimum required memory of each primitive function callsite by analyzing the liveness of the input/output tensors at each function callsite and calculating the total amount of memory these tensors require. This is added as a "used_memory" annotation to the function in question as a list of the number of bytes for each callsite. In addition, the containing function is annotated with an "io_used_memory" annotation which refers to the total memory required for the IO tensors. More...

Pass	CapturePostDfsIndexInSpans ()
	Captures the post-dfs index and dominator post-dfs index of (most) expression nodes in their span, in the form "index:<post-dfs index>:<dominator post-dfs index>". This is useful for debugging since a) it helps identify pretty-printed sub-expressions within the overall model and b) the indexes are heavily used by Collage for its compact representation of sub-graphs. More...

Pass	AnnotateMemoryScope ()
	Calls device dependent memory scope analysis pass, collects mapping of desirable expr->memory_scope and annotates expressions by VirtualDevice with required memory_scope. More...

Pass	RemoveStandaloneReshapes ()
	Removes non-fused reshapes after lowering the graph. InferType() cannot be invoked after calling this pass as it removes reshapes from the call graph. Many targets only need buffer addresses irrespective of the shapes of them. This makes reshapes symbolic once the graph has been lowered. Reshape removal results into smaller code size and reduced buffer allocations. It opens up opportunities of operator fusion in the target backend. Thus, consequently, it improves the performance of the inference. More...

Typedef Documentation

◆ FTVMRelayToTIR

typedef tvm::transform::Pass tvm::relay::transform::FTVMRelayToTIR

RelayToTIR tvm::transform::Pass specific to a TargetKind.

Called before the default lowering passes.

Parameters

mod	The module that an optimization pass runs on.
pass_ctx	The pass context that can provide information for the optimization.

Returns: The transformed module.

◆ FTVMTIRToRuntime

using tvm::relay::transform::FTVMTIRToRuntime = typedef tvm::runtime::TypedPackedFunc<runtime::Module(IRModule, Target)>

TIRToRuntime conversion specific to a TargetKind.

This function is responsible for scanning an IRModule for appropriate Target-specific functions and generating a Runtime module representing the compiled output

Parameters

ir_module	Unified IRModule
target	Target to filter on or retrieve arguments from

Returns: Runtime Module containing compiled functions

◆ Pass

using tvm::relay::transform::Pass = typedef tvm::transform::Pass

◆ PassContext

using tvm::relay::transform::PassContext = typedef tvm::transform::PassContext

◆ PassContextNode

using tvm::relay::transform::PassContextNode = typedef tvm::transform::PassContextNode

◆ PassInfo

using tvm::relay::transform::PassInfo = typedef tvm::transform::PassInfo

◆ PassInfoNode

using tvm::relay::transform::PassInfoNode = typedef tvm::transform::PassInfoNode

◆ PassNode

using tvm::relay::transform::PassNode = typedef tvm::transform::PassNode

◆ Sequential

using tvm::relay::transform::Sequential = typedef tvm::transform::Sequential

Function Documentation

◆ AlterOpLayout()

Pass tvm::relay::transform::AlterOpLayout ( )

Alternate the layouts of operators or replace primitive operators with other expressions.

Returns: The pass.

◆ AnnotateMemoryScope()

Pass tvm::relay::transform::AnnotateMemoryScope ( )

Calls device dependent memory scope analysis pass, collects mapping of desirable expr->memory_scope and annotates expressions by VirtualDevice with required memory_scope.

◆ AnnotateUsedMemory()

Pass tvm::relay::transform::AnnotateUsedMemory ( )

Annotates the minimum required memory of each primitive function callsite by analyzing the liveness of the input/output tensors at each function callsite and calculating the total amount of memory these tensors require. This is added as a "used_memory" annotation to the function in question as a list of the number of bytes for each callsite. In addition, the containing function is annotated with an "io_used_memory" annotation which refers to the total memory required for the IO tensors.

Note: This pass does not support dynamic shapes, it is the users responsibility to check this pass isn't applied where dynamic shapes may be input.

◆ AutoSchedulerLayoutRewrite()

Pass tvm::relay::transform::AutoSchedulerLayoutRewrite ( )

Do layout rewrite according to the tile structure created by auto-scheduler.

Returns: The pass

◆ BackwardFoldScaleAxis()

Pass tvm::relay::transform::BackwardFoldScaleAxis ( )

Backward fold axis scaling into weights of conv/dense operators.

Returns: The pass.

◆ CanonicalizeCast()

Pass tvm::relay::transform::CanonicalizeCast ( )

Canonicalize cast expressions to make operator fusion more efficient.

Returns: The pass.

◆ CanonicalizeOps()

Pass tvm::relay::transform::CanonicalizeOps ( )

Canonicalize some operators to the simplified operators. For example, bias_add can be canonicalized to expand_dims and broadcast_add.

Returns: The pass.

◆ CapturePostDfsIndexInSpans()

Pass tvm::relay::transform::CapturePostDfsIndexInSpans ( )

Captures the post-dfs index and dominator post-dfs index of (most) expression nodes in their span, in the form "index:<post-dfs index>:<dominator post-dfs index>". This is useful for debugging since a) it helps identify pretty-printed sub-expressions within the overall model and b) the indexes are heavily used by Collage for its compact representation of sub-graphs.

Note that Op and Constructor nodes are not changed even though they are assigned an post-dfs index.

◆ CombineParallelBatchMatmul()

Pass tvm::relay::transform::CombineParallelBatchMatmul ( uint64_t min_num_branches = 3 )

Combine parallel batch_matmul ops into a single batch_matmul if the number of branches of this dense operator is not less than min_num_branch.

Parameters

min_num_branches The minimun number of branches.

Returns: The pass.

◆ CombineParallelConv2D()

Pass tvm::relay::transform::CombineParallelConv2D ( uint64_t min_num_branches = 3 )

Combine parallel 2d convolutions into a single convolution if the number of branches of this conv2d operator is not less than min_num_branch.

Parameters

min_num_branches The minimun number of branches.

Returns: The pass.

◆ CombineParallelDense()

Pass tvm::relay::transform::CombineParallelDense	(	uint64_t	min_num_branches = `3`,
		bool	to_batch_matmul = `true`
	)

Combine parallel dense ops into a single batch_matmul if the number of branches of this dense operator is not less than min_num_branch.

Parameters

min_num_branches	The minimun number of branches.
to_batch_matmul	Whether to combine parallel dense ops to batch matmul. If set false, combine dense ops to single dense op.

Returns: The pass.

◆ ConvertLayout()

Pass tvm::relay::transform::ConvertLayout ( const Map< String, Array< String >> & desired_layouts )

Given a dest layout, this pass transforms the expr such that most of the ops input data layout is changed to the dest layout. In ideal situation, there are only 2 layout transforms, one at the start and one at the end.

This pass is not a part of relay.build and is expected to be called between framework-relay parser and relay.build call. This is very helpful for hardware backends that support/prefer only type of data layout.

RFC - https://discuss.tvm.ai/t/layout-conversion-pass/4009

This pass uses most of the AlterOpLayout and InferCorrectLayout infrastructure. We can define new layouts for conv2d ops for now. Most of the other operators try to adapt to their input layout using the InferCorrectLayout infrastructure.

Parameters

desired_layouts Specify mapping of op_name to array of desired layouts for each input. For example: Map("nn.conv2d", Array("NHWC", "OHWI")), this specifies the desired layout for data then kernel for nn.conv2d.

Returns: The pass.

◆ CreateFunctionPass()

Pass tvm::relay::transform::CreateFunctionPass	(	const runtime::TypedPackedFunc< Function(Function, IRModule, PassContext)> &	pass_func,
		int	opt_level,
		String	name,
		tvm::Array< String >	required,
		bool	traceable = `false`
	)

◆ DeadCodeElimination()

Pass tvm::relay::transform::DeadCodeElimination	(	bool	inline_once = `false`,
		bool	ignore_purity = `false`
	)

Remove let-bound expressions which do not effect the program result.

This pass will remove let bindings which are not referenced. If inline_once is True, let bindings which are only referenced once will also be inlined.

For example, this pass should turn let a = 1; 2 into 2, as the value of the expression does not depend on a.

As another example, let a = 1; a will be optimized into 1 if inline_once is True.

If ignore_purity is False, possibly side-effecting expressions (such as memory allocation, random number generation, reading/writing references, or calls to primitive or external functions) are never elided or inlined. This is sound, but ignore_purity can be set to True to suppress this check.

The analysis is fairly conservative, for example it assumes all local functions may be called more than once, any functions passed as arguments have side effects, and so on.

Parameters

inline_once	whether or not to inline bindings used exactly once.
ignore_purity	whether to ignore whether expressions have side-effects

Returns: the pass.

◆ DefuseOps()

Pass tvm::relay::transform::DefuseOps ( )

The inverse operation of FuseOps. It transforms a fused program returned by FuseOps into the program before FuseOps. (i.e. x == DefuseOps(FuseOps(x)))

Returns: The pass.

◆ DynamicToStatic()

Pass tvm::relay::transform::DynamicToStatic ( )

Find Dynamic ops and make them static.

Searches the graph for dynamic ops. If the dynamic inputs to those ops are constants, it replaces them with static ops and re-performs type inference and constant folding. The pass repeats itself until the graph stops changing or we run too many iterations.

Returns: The pass.

◆ EliminateCommonSubexpr()

Pass tvm::relay::transform::EliminateCommonSubexpr ( runtime::PackedFunc fskip = nullptr )

Search and eliminate common subexpression. For example, if there are two expressions evaluated to an identical value, a single variable is created and these two expressions are replaced by this variable.

Parameters

fskip The callback argument that allows to skip certain expressions.

Returns: The pass.

◆ EtaExpand()

Pass tvm::relay::transform::EtaExpand	(	bool	expand_constructor,
		bool	expand_global_var
	)

Add abstraction over a constructor or global variable bound to a function.

For example: square is transformed to fn (x: int32) -> int32 { square(x) }.

See https://en.wikipedia.org/wiki/Lambda_calculus#%CE%B7-conversion for more details.

Parameters

expand_constructor	Whether to expand constructors.
expand_global_var	Whether to expand global variables.

Returns: The pass.

◆ FastMath()

Pass tvm::relay::transform::FastMath ( )

Replaces non linear activation functions with their fast but approximate counterparts.

Returns: The Pass.

◆ FlattenAtrousConv()

Pass tvm::relay::transform::FlattenAtrousConv ( )

This transform flattens atrous convolution, which corresponds to the sequence of operations: "space_to_batch_nd"->"conv2d"->"batch_to_space_nd" and convert them into subgraphs with a convolution with the modified "dilation" and recalculated "padding" parameters.

Returns: The pass.

◆ FoldConstant()

Pass tvm::relay::transform::FoldConstant ( bool fold_qnn = false )

Fold constant expressions.

Because of backward compatibility reason it skips QNN primitives from folding by default. There are some transformation passes like FakeQuantizationToInteger, which requires to keep QNN primitives for constant subgraphs. Uncontrolled constant folding of QNN primitives may break applicability of FakeQuantizationToInteger. We suggest to use FoldConstant pass with none default fold_qnn=True value only when all other QNN sensitive passes were already applied.

Parameters

fold_qnn Whether to fold constants for QNN operations.

Returns: The pass.

◆ FoldScaleAxis()

Pass tvm::relay::transform::FoldScaleAxis ( )

A sequential pass that executes ForwardFoldScaleAxis and BackwardFoldScaleAxis passes.

Returns: The pass.

◆ ForwardFoldScaleAxis()

Pass tvm::relay::transform::ForwardFoldScaleAxis ( )

Forward fold axis scaling into weights of conv/dense operators.

Returns: The pass.

◆ FuseOps()

Pass tvm::relay::transform::FuseOps ( int fuse_opt_level = -1 )

Fuse operations into expr into separate functions.

Parameters

fuse_opt_level Optimization level. If it is -1 it will be inferred from pass context.

Returns: The pass.

◆ InferType()

Pass tvm::relay::transform::InferType ( )

Infer the type of an expression.

The result of type checking is a new expression with unambiguous type information filled in, as well as it's checked type field populated with the result type.

Returns: The pass.

◆ InferTypeLocal()

Type tvm::relay::transform::InferTypeLocal ( const Expr & expr )

Infer the type of an expression, reusing existing type information.

The result of type checking is a new expression with unambiguous type information filled in for the given node only. The local version can use existing type information populated throughout the expression and assumes this information is correct. The local version also avoids examining large amounts of the graph assuming type information is filled in properly which makes it much faster if we iteratively call type inference.

Returns: The type of the expression.

◆ Inline()

Pass tvm::relay::transform::Inline ( )

Inline the global functions marked as inline in a given Relay IRModule.

Returns: The pass.

◆ LazyGradientInit()

Pass tvm::relay::transform::LazyGradientInit ( )

Convert all expressions of TensorType into GradCell, an algebraic data type defined in gradient.rly.

This will delay or decrease memory usage. All calls to ones, ones_like, zeros, zeros_like will not immediately instantiate a tensor in memory, rather only instantiate if needed. It also defines + and * operation between GradCell types which can increase performance when using zero-filled or one-filled tensors, which is the case in reverse mode ad.

Returns: the pass

◆ Legalize()

Pass tvm::relay::transform::Legalize ( const String & legalize_map_attr_name = "FTVMLegalize" )

Legalizes an expr with another expression.

Parameters

legalize_map_attr_name The Op's attr name which corresponds to the legalize rule function. One can collect and isolate similar type of legalize transformations using this param. For example, transformations that only apply to Dialects can be isolated into a FTVMDialectLegalize string. This pass calls only those transformations that have been registered using the supplied legalize_map_attr_name.

Returns: The pass.

◆ ManifestAlloc()

Pass tvm::relay::transform::ManifestAlloc ( VirtualDevice cpu_virtual_device )

A pass for manifesting explicit memory allocations and rewriting specific dialects.

Parameters

cpu_virtual_device VirtualDevice for computations and data which must reside on a CPU, such as shapes and shape functions.

Returns: The pass.

◆ ManifestLifetimes()

Pass tvm::relay::transform::ManifestLifetimes ( )

A pass for manifesting variable lifetimes by inserting kill operations when variables become dead. This pass should be run after ManifestAlloc, and should not be run more than once.

Returns: The pass.

◆ MetaScheduleLayoutRewrite()

Pass tvm::relay::transform::MetaScheduleLayoutRewrite ( )

Do layout rewrite according to the tile structure created by meta-schedule.

Returns: The pass

◆ PartialEval()

Pass tvm::relay::transform::PartialEval ( )

Aggressive constant propagation/constant folding/inlining.

It will do as much computation in compile time as possible. It has two benefit: remove runtime overhead, and allow more optimization (typically fusion). As a side effect, code size will explode.

Returns: the optimized expression.

◆ PartitionGraph()

Pass tvm::relay::transform::PartitionGraph ( )

Partition a Relay program into regions that can be executed on different backends.

Returns: The pass.

◆ PlanDevices()

Pass tvm::relay::transform::PlanDevices ( CompilationConfig config )

Uses existing "on_device" and "device_copy" CallNodes to infer the VirtualDevice on which every Relay sub-expression should run and the result stored. Captures the result of that analysis using new "on_device" and "device_copy" CallNodes.

See tvm::relay::transform::{LexicalOnDeviceMixin,DeviceAwareExprVisitor,DeviceAwareExprMutator} for help recovering the device for an arbitrary sub-expression in downstream transformations.

Parameters

config Describes the targets and default VirtualDevice for all primitive operators and host sub-expressions.

Returns: The pass.

◆ RelayToTIRTargetHook()

Pass tvm::relay::transform::RelayToTIRTargetHook ( CompilationConfig config )

Run any custom passes registered under "RelayToTIR" attributes on TargetKinds.

This pass looks for inline, let-bound or global functions which have a "Compiler" attribute. If the attribute value corresponds to a TargetKind with a "RelayToTIR" attribute, then the 'custom' pass bound to that attribute is run (at most once) on the IRModule as a whole.

If, in addition, the config has a Target with a matching TargetKind, that Target is set as the 'current' target before the custom pass is executed. In this way it is possible for custom passes to pick up target options which may guide how they transform the IRModule. (Those targets are referred to as 'extern codegen targets' elsewhere).

A typical custom pass will:

Find calls to "Compiler" attributes functions with matching compiler name.
Lower those function to TIR PrimFuncs.
Bind those functions into the IRModule under the functions' "global_symbol" attribute.
Replace all calls to those functions with 'call_lowered' to the matching global. Care should be taken to handle multiple calls to the same function. See src/relay/backend/contrib/example_target_hooks/relay_to_tir.cc for an example custom pass.

It is also possible (despite the pass and attribute names!) for the custom pass to proceed directly to a runtime::Module, which can be attached to the output IRModules "external_mods" attribute (taking care not to clobber any existing modules). In this case the flow is as above, except:

The runtime::Module must contain a binding for each compiled function under their "global_symbol" (ie runtime::Module::ImplementsFunction should return true).
A Relay Function must be bound (or re-bound) into the result IRModule, again with the same "global_symbol", but with only the "Extern" attribute set to Integer(1). The function body should be the original function body. In this way we always have a TVM definition matching every global function name.

There are many existing runtime::Modules, ranging from source to object to dynamic libaries to entirely custom implementations. Some of those may require additional compilation using 'export_library' on the final build artifact.

The OutlineCompilerFunctionsWithExistingGlobalSymbols and MarkCompilerFunctionsAsExtern utility passes can be used by custom passes to take care of some of the boilerplate.

TODO(mbs): Rename PreLoweringTargetHooks?

Parameters

config All available targets.

Returns: The pass.

◆ RemoveStandaloneReshapes()

Pass tvm::relay::transform::RemoveStandaloneReshapes ( )

Removes non-fused reshapes after lowering the graph. InferType() cannot be invoked after calling this pass as it removes reshapes from the call graph. Many targets only need buffer addresses irrespective of the shapes of them. This makes reshapes symbolic once the graph has been lowered. Reshape removal results into smaller code size and reduced buffer allocations. It opens up opportunities of operator fusion in the target backend. Thus, consequently, it improves the performance of the inference.

◆ RemoveUnusedFunctions()

Pass tvm::relay::transform::RemoveUnusedFunctions ( Array< runtime::String > entry_functions )

Remove the unused functions in the Relay IRModule.

Parameters

entry_functions The entry functions used to search the functions that are being used.

Returns: The pass.

◆ RewriteAnnotatedOps()

Pass tvm::relay::transform::RewriteAnnotatedOps ( int fallback_device )

Rewrite the annotated program.

Parameters

fallback_device The fallback device which is the default device for operators without annotation.

Returns: The pass.

◆ SimplifyExpr()

Pass tvm::relay::transform::SimplifyExpr ( )

Simplify the Relay expression.

Returns: The pass.

◆ SimplifyExprPostAlterOp()

Pass tvm::relay::transform::SimplifyExprPostAlterOp ( )

Stripped down version of SimplifyExpr which is run after AlterOpLayout.

Returns: The pass.

◆ SimplifyInference()

Pass tvm::relay::transform::SimplifyInference ( )

Simplify certain operators during inference. For example, the result of a batch norm which is indexed at tuple index 0 will be unpacked into a number of simplified operators.

Returns: The Pass.

◆ SplitArgs()

Pass tvm::relay::transform::SplitArgs ( uint64_t max_function_args )

Split function with huge number of arguments to smaller pieces.

Parameters

max_function_args Maximum number of function arguments. If it equals 0 then SplitArgs shouldn't split the function.

Returns: The pass.

◆ ToANormalForm() [1/2]

Pass tvm::relay::transform::ToANormalForm ( )

turn a dataflow graph into Administrative Normal Form, or A-Normal Form (ANF).

It will turn an expression that is in a graph form (with sharing implicit), to an expression with explicit sharing (A-Normal Form).

The scope of the root expression is the global scope.

The scope of any non root expression is the least common ancestor of all it's scope.

Values are ordered by post-DFS order in each scope.

Returns: The pass.

◆ ToANormalForm() [2/2]

Expr tvm::relay::transform::ToANormalForm ( const Expr & expr )

ToANormalForm but on incomplete graph.

Parameters

expr	the graph.

Returns: The transformed program.

◆ ToBasicBlockNormalForm()

Pass tvm::relay::transform::ToBasicBlockNormalForm ( )

Turn an expression to Basic Block Normal Form.

We define a block as a group of expressions implied by the scope structure.

Each graph node can only belong to a single block.

For any value that is being used in multiple blocks, it has to be referred by a Var which is defined in a block, whose scope is the least common ancestor of blocks this value is used.

Returns: The pass.

◆ ToCPS()

Pass tvm::relay::transform::ToCPS ( )

Turn an expression into continuation passing style(CPS).

CPS mean that every function will, instead of returning the result directly, be passed down an extra function (called the continuation) as argument, and pass the result to the continuation instead.

Thus, every function call has to be passed an extra argument that represent the rest of the computation (Hence the name of continuation).

Similarly, all other compute will be wrapped and call the continuation as well.

Returns: the pass.

◆ ToGraphNormalForm()

Pass tvm::relay::transform::ToGraphNormalForm ( )

Remove let binding and directly share via pointer instead.

It will remove all let binding, and turn all of the variable bound by let into direct pointer reference.

Returns: the expression in graph normal form.

Typedefs

Functions

Typedef Documentation

◆ FTVMRelayToTIR

◆ FTVMTIRToRuntime

◆ Pass

◆ PassContext

◆ PassContextNode

◆ PassInfo

◆ PassInfoNode

◆ PassNode

◆ Sequential

Function Documentation

◆ AlterOpLayout()

◆ AnnotateMemoryScope()

◆ AnnotateUsedMemory()

◆ AutoSchedulerLayoutRewrite()

◆ BackwardFoldScaleAxis()

◆ CanonicalizeCast()

◆ CanonicalizeOps()

◆ CapturePostDfsIndexInSpans()

◆ CombineParallelBatchMatmul()

◆ CombineParallelConv2D()

◆ CombineParallelDense()

◆ ConvertLayout()

◆ CreateFunctionPass()

◆ DeadCodeElimination()

◆ DefuseOps()

◆ DynamicToStatic()

◆ EliminateCommonSubexpr()

◆ EtaExpand()

◆ FastMath()

◆ FlattenAtrousConv()

◆ FoldConstant()

◆ FoldScaleAxis()

◆ ForwardFoldScaleAxis()

◆ FuseOps()

◆ InferType()

◆ InferTypeLocal()

◆ Inline()

◆ LazyGradientInit()

◆ Legalize()

◆ ManifestAlloc()

◆ ManifestLifetimes()

◆ MetaScheduleLayoutRewrite()

◆ PartialEval()

◆ PartitionGraph()

◆ PlanDevices()

◆ RelayToTIRTargetHook()

◆ RemoveStandaloneReshapes()

◆ RemoveUnusedFunctions()

◆ RewriteAnnotatedOps()

◆ SimplifyExpr()

◆ SimplifyExprPostAlterOp()

◆ SimplifyInference()

◆ SplitArgs()

◆ ToANormalForm() [1/2]

◆ ToANormalForm() [2/2]

◆ ToBasicBlockNormalForm()

◆ ToCPS()

◆ ToGraphNormalForm()