tvm::tirx::transform Namespace Reference

Functions
TVM_TIRX_DLL Pass	VerifyMemory ()
	Pass variant of VerifyMemory.
TVM_TIRX_DLL Pass	DecorateDeviceScope ()
	Wrap a function body in a device scope region.
TVM_TIRX_DLL Pass	CreateFunctionPass (std::function< ffi::Optional< Function >(Function, IRModule, PassContext)> pass_func, int opt_level, ffi::String name)
TVM_TIRX_DLL Pass	VectorizeLoop (bool enable_vectorize=true)
	Lower vectorization loops.
TVM_TIRX_DLL Pass	StorageRewrite ()
	Rewrite storage allocation pattern. Moves the allocation to outer most possible scope. Trying to share space between allocations to make a static allocation plan when possible.
TVM_TIRX_DLL Pass	UnrollLoop ()
	unroll the constant loop marked by unroll. This pass also automatically attach pragma unroll tag to loops which meets the standard.
TVM_TIRX_DLL Pass	RemoveNoOp ()
	Remove No Op from the Stmt.
TVM_TIRX_DLL Pass	HoistIf ()
	Hoist loop-invariant If statements outside eligible loops.
TVM_TIRX_DLL Pass	RenormalizeSplitPattern ()
	Renormalize the split pattern from floordiv(floormod()) to floormod(floordiv()).
TVM_TIRX_DLL Pass	StmtSimplify ()
	Run statement-level arithmetic simplifications on the TIR Function.
TVM_TIRX_DLL Pass	ConvertSSA ()
	Convert an IRModule to be SSA form.
TVM_TIRX_DLL Pass	LowerThreadAllreduce ()
	Lower cross-thread reductions to target-specific communication.
TVM_TIRX_DLL Pass	MakePackedAPI ()
	Transform the high-level Function to a low-level version that can be used as an API function.
TVM_TIRX_DLL Pass	SplitHostDevice ()
	Annotate, split, and lower host/device functions.
TVM_TIRX_DLL Pass	SkipAssert ()
	skip assert stmt.
TVM_TIRX_DLL Pass	LowerTVMBuiltin ()
	Lower builtin intrinsics.
TVM_TIRX_DLL Pass	LowerIntrin ()
	Lower the target specific function intrinsics in each of the function.
TVM_TIRX_DLL Pass	LowerWarpMemory ()
	Lower warp memory access to low-level device related function calls.
TVM_TIRX_DLL Pass	NarrowDataType (int target_bits)
	Narrow down PrimExpr datatype in stmt to target_bits.
TVM_TIRX_DLL Pass	ForceNarrowIndexToInt32 ()
	Force to narrow down indexing expressions and integer tensors to int32 dtype.
TVM_TIRX_DLL Pass	BF16ComputeLegalize ()
	Legalize bf16 compute Ops. Add a cast to fp32 before Ops, then add a cast back to bf16.
TVM_TIRX_DLL Pass	FP8ComputeLegalize (ffi::String promote_dtype="float16")
	Legalize fp8 compute Ops. Add a cast to fp16/fp32 before Ops, then add a cast back to fp8.
TVM_TIRX_DLL Pass	BF16StorageLegalize ()
	Legalize bf16 storage types to u16.
TVM_TIRX_DLL Pass	FP8StorageLegalize ()
	Legalize fp8 storage types to u8.
TVM_TIRX_DLL Pass	PointerValueTypeRewrite ()
	Rewrite the pointer content type of arguments, as well as Alloc internal to the function to use the most frequently accessed type for load/store to avoid pointer casting in backend when possible.
TVM_TIRX_DLL Pass	FlattenTensor ()
	Flatten the multi-dimensional TensorLoad and TensorStore to single dimensional TensorLoad/TensorStore for the TIR not contains opaque block.
TVM_TIRX_DLL Pass	CommonSubexprElim ()
	Implements Common Subexpression Elimination (CSE) for TIR which introduces Bind statements for duplicated sub-expressions.
TVM_TIRX_DLL Pass	BindTarget (Target target)
	Annotate a Function with a given target.
TVM_TIRX_DLL Pass	AnnotateEntryFunc ()
	Set a Function as the entry point if it is only function in IRModule.
TVM_TIRX_DLL Pass	Filter (ffi::TypedFunction< bool(Function)> fcond)
	Filter Functions with a given condition.
TVM_TIRX_DLL Pass	TileDispatch ()
	Lower TIRx op calls using registered op dispatchers for the given target.
TVM_TIRX_DLL Pass	LowerTIRxCleanup ()
	Finalize TIRx lowering by applying layout rewriters and cleanup passes.
TVM_TIRX_DLL Pass	LowerTIRx ()
	Lower the TIR to a lower level IR for the given target.

Function Documentation

AnnotateEntryFunc()

TVM_TIRX_DLL Pass tvm::tirx::transform::AnnotateEntryFunc

(

)

Set a Function as the entry point if it is only function in IRModule.

Returns

The pass.

BF16ComputeLegalize()

TVM_TIRX_DLL Pass tvm::tirx::transform::BF16ComputeLegalize

(

)

Legalize bf16 compute Ops. Add a cast to fp32 before Ops, then add a cast back to bf16.

Returns

The pass.

BF16StorageLegalize()

TVM_TIRX_DLL Pass tvm::tirx::transform::BF16StorageLegalize

(

)

Legalize bf16 storage types to u16.

Returns

The pass.

BindTarget()

TVM_TIRX_DLL Pass tvm::tirx::transform::BindTarget

(

Target

target

)

Annotate a Function with a given target.

Note

Device calls must be enclosed in explicit launch_thread or device_entry regions. Lower dialect-specific thread bindings before classifying host/device calls.

Returns

The pass.

CommonSubexprElim()

TVM_TIRX_DLL Pass tvm::tirx::transform::CommonSubexprElim

(

)

Implements Common Subexpression Elimination (CSE) for TIR which introduces Bind statements for duplicated sub-expressions.

Returns

The pass.

ConvertSSA()

TVM_TIRX_DLL Pass tvm::tirx::transform::ConvertSSA

(

)

Convert an IRModule to be SSA form.

This pass handles cases where the same tvm::Var appears in multiple functions within the same module. For example, after extracting a fragment from one function into another, where the same tvm::Var may be defined both as within the body of the original function, and as a parameter within the hoisted function.

Returns

The pass.

CreateFunctionPass()

TVM_TIRX_DLL Pass tvm::tirx::transform::CreateFunctionPass	(	std::function< ffi::Optional< Function >(Function, IRModule, PassContext)>	pass_func,
		int	opt_level,
		ffi::String	name
	)

DecorateDeviceScope()

TVM_TIRX_DLL Pass tvm::tirx::transform::DecorateDeviceScope

(

)

Wrap a function body in a device scope region.

Filter()

TVM_TIRX_DLL Pass tvm::tirx::transform::Filter

(

ffi::TypedFunction< bool(Function)>

fcond

)

Filter Functions with a given condition.

Returns

The pass.

FlattenTensor()

TVM_TIRX_DLL Pass tvm::tirx::transform::FlattenTensor

(

)

Flatten the multi-dimensional TensorLoad and TensorStore to single dimensional TensorLoad/TensorStore for the TIR not contains opaque block.

Returns

The pass.

ForceNarrowIndexToInt32()

TVM_TIRX_DLL Pass tvm::tirx::transform::ForceNarrowIndexToInt32

(

)

Force to narrow down indexing expressions and integer tensors to int32 dtype.

The function must not contain S-TIR blocks.

Returns

The pass.

Note

This pass should not be used in default cases.

FP8ComputeLegalize()

TVM_TIRX_DLL Pass tvm::tirx::transform::FP8ComputeLegalize

(

ffi::String

promote_dtype = "float16"

)

Legalize fp8 compute Ops. Add a cast to fp16/fp32 before Ops, then add a cast back to fp8.

Parameters

promote_dtype

The data type used for type promotion, defaults to float16

Note

Must be run after BindTarget, as it relies on target attributes for Functions

Returns

The pass.

FP8StorageLegalize()

TVM_TIRX_DLL Pass tvm::tirx::transform::FP8StorageLegalize

(

)

Legalize fp8 storage types to u8.

Note

Must be run after BindTarget, as it relies on target attributes for Functions

Returns

The pass.

HoistIf()

TVM_TIRX_DLL Pass tvm::tirx::transform::HoistIf

(

)

Hoist loop-invariant If statements outside eligible loops.

Returns

The pass.

LowerIntrin()

TVM_TIRX_DLL Pass tvm::tirx::transform::LowerIntrin

(

)

Lower the target specific function intrinsics in each of the function.

Returns

The pass.

LowerThreadAllreduce()

TVM_TIRX_DLL Pass tvm::tirx::transform::LowerThreadAllreduce

(

)

Lower cross-thread reductions to target-specific communication.

LowerTIRx()

TVM_TIRX_DLL Pass tvm::tirx::transform::LowerTIRx

(

)

Lower the TIR to a lower level IR for the given target.

Returns

The pass.

LowerTIRxCleanup()

TVM_TIRX_DLL Pass tvm::tirx::transform::LowerTIRxCleanup

(

)

Finalize TIRx lowering by applying layout rewriters and cleanup passes.

Returns

The pass.

LowerTVMBuiltin()

TVM_TIRX_DLL Pass tvm::tirx::transform::LowerTVMBuiltin

(

)

Lower builtin intrinsics.

Returns

The pass.

LowerWarpMemory()

TVM_TIRX_DLL Pass tvm::tirx::transform::LowerWarpMemory

(

)

Lower warp memory access to low-level device related function calls.

Returns

The pass.

MakePackedAPI()

TVM_TIRX_DLL Pass tvm::tirx::transform::MakePackedAPI

(

)

Transform the high-level Function to a low-level version that can be used as an API function.

The main task of this function is to create code to :

• Map the values in the api_args to Var that is required by body.
• Insert assertions to check type/value of the passed arguments.

Note

The function signature have two cases

let num_packed_args = len(api_args);

if num_packed_args is zero: f()

if num_packed_args is not zero: f(void *, TVMFFIAny* packed_args, int num_packed_args, api_arg_k, api_arg_k+1, ... api_arg_n, TVMFFIAny* out_ret_val)

where n == len(api_args), k == num_packed_args

Returns

The pass.

NarrowDataType()

TVM_TIRX_DLL Pass tvm::tirx::transform::NarrowDataType

(

int

target_bits

)

Narrow down PrimExpr datatype in stmt to target_bits.

Parameters

target_bits

The target bits

Note

Run this pass after storage flatten.

Returns

The pass.

PointerValueTypeRewrite()

TVM_TIRX_DLL Pass tvm::tirx::transform::PointerValueTypeRewrite

(

)

Rewrite the pointer content type of arguments, as well as Alloc internal to the function to use the most frequently accessed type for load/store to avoid pointer casting in backend when possible.

Returns

The pass.

RemoveNoOp()

TVM_TIRX_DLL Pass tvm::tirx::transform::RemoveNoOp

(

)

Remove No Op from the Stmt.

Returns

The pass.

RenormalizeSplitPattern()

TVM_TIRX_DLL Pass tvm::tirx::transform::RenormalizeSplitPattern

(

)

Renormalize the split pattern from floordiv(floormod()) to floormod(floordiv()).

Returns

The pass.

SkipAssert()

TVM_TIRX_DLL Pass tvm::tirx::transform::SkipAssert

(

)

skip assert stmt.

Returns

The pass.

SplitHostDevice()

TVM_TIRX_DLL Pass tvm::tirx::transform::SplitHostDevice

(

)

Annotate, split, and lower host/device functions.

This pass first annotates device regions within host functions, then splits them into host and device-side Functions, and finally lowers host-to-device calls into the device kernel launch ABI.

The resulting host-side function will keep the same tvm::attr::kTarget attribute (e.g. T.target("cuda", host=T.target("llvm"))). This ensures that MakePackedAPI knows which device type should be used for the input tensors.

The resulting device-side function will have the host stripped from its target attribute (e.g. T.target("cuda")).

Returns

The pass.

StmtSimplify()

TVM_TIRX_DLL Pass tvm::tirx::transform::StmtSimplify

(

)

Run statement-level arithmetic simplifications on the TIR Function.

Returns

The pass.

StorageRewrite()

TVM_TIRX_DLL Pass tvm::tirx::transform::StorageRewrite

(

)

Rewrite storage allocation pattern. Moves the allocation to outer most possible scope. Trying to share space between allocations to make a static allocation plan when possible.

Returns

The pass.

TileDispatch()

TVM_TIRX_DLL Pass tvm::tirx::transform::TileDispatch

(

)

Lower TIRx op calls using registered op dispatchers for the given target.

Lowers CUDA index calls to ordinary bindings using the independent device-entry launch configuration, preserving execution-context information during dispatch.

Returns

The pass.

UnrollLoop()

TVM_TIRX_DLL Pass tvm::tirx::transform::UnrollLoop

(

)

unroll the constant loop marked by unroll. This pass also automatically attach pragma unroll tag to loops which meets the standard.

Returns

The pass.

VectorizeLoop()

TVM_TIRX_DLL Pass tvm::tirx::transform::VectorizeLoop

(

bool

enable_vectorize = true

)

Lower vectorization loops.

Target-dependent vectorization uses the Function's target attribute, not an ambient target context or nested target attributes. Target-independent fixed-width loops do not require a target. Code needing different vectorization targets must be separated into functions before this pass.

Parameters

enable_vectorize

Whether vectorization is enabled.

Returns

The pass.

VerifyMemory()

TVM_TIRX_DLL Pass tvm::tirx::transform::VerifyMemory

(

)

Pass variant of VerifyMemory.

Returns

The pass.

See also

tvm::tirx::VerifyMemory