tvm
Public Types | Public Member Functions | Static Public Member Functions | List of all members
tvm::meta_schedule::Postproc Class Reference

Managed reference to PostprocNode. More...

#include <postproc.h>

Inheritance diagram for tvm::meta_schedule::Postproc:
Collaboration diagram for tvm::meta_schedule::Postproc:

Public Types

using FInitializeWithTuneContext = ffi::TypedFunction< void(const TuneContext &)>
 The function type of InitializeWithTuneContext method. More...
 
using FApply = ffi::TypedFunction< bool(const tir::Schedule &)>
 Apply a postprocessor to the given schedule. More...
 
using FClone = ffi::TypedFunction< Postproc()>
 Clone the postprocessor. More...
 
using FAsString = ffi::TypedFunction< String()>
 Get the postprocessor function as string with name. More...
 

Public Member Functions

 TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS (Postproc, ObjectRef, PostprocNode)
 

Static Public Member Functions

static Postproc PyPostproc (FInitializeWithTuneContext f_initialize_with_tune_context, FApply f_apply, FClone f_clone, FAsString f_as_string)
 Create a postprocessor with customized methods on the python-side. More...
 
static Postproc DisallowDynamicLoop ()
 Create a postprocessor that checks if all loops are static. More...
 
static Postproc DisallowAsyncStridedMemCopy ()
 Create a postprocessor that checks if all async mem copies are not strided. More...
 
static Postproc RewriteCooperativeFetch ()
 Create a postprocessor that rewrites the cooperative fetch annotation to actual vectorized cooperative fetching in loop bindings. More...
 
static Postproc RewriteParallelVectorizeUnroll ()
 Creates a postprocessor that applies parallelization, vectorization and auto unrolling according to the annotation of each block. More...
 
static Postproc RewriteReductionBlock ()
 Create a postprocessor that rewrites reduction block by moving the init block out. More...
 
static Postproc RewriteUnboundBlock (int max_threadblocks)
 Create a postprocessor that adds thread binding to unbound blocks. More...
 
static Postproc RewriteTensorize (bool vectorize_init_loop=false)
 Create a postprocessor that applies tensorization to annotated blocks. More...
 
static Postproc VerifyGPUCode ()
 Creates a postprocessor that verifies if the GPU code is correct. More...
 
static Postproc VerifyVTCMLimit ()
 Verifies that the VTCM usage of a given schedule is within the provided limit. More...
 
static Postproc RewriteLayout ()
 Creates a postprocessor that rewrites the layout of input tensor. More...
 
static Array< Postproc, void > DefaultLLVM ()
 Create default postprocessors for LLVM. More...
 
static Array< Postproc, void > DefaultCPUTensorization ()
 Create default postprocessors for x86 (AVX512 and VNNI) More...
 
static Array< Postproc, void > DefaultCUDA ()
 Create default postprocessors for CUDA. More...
 
static Array< Postproc, void > DefaultCUDATensorCore ()
 Create default postprocessors for CUDA with TensorCore. More...
 
static Array< Postproc, void > DefaultHexagon ()
 Create default postprocessors for Hexagon. More...
 

Detailed Description

Managed reference to PostprocNode.

See also
PostprocNode

Member Typedef Documentation

◆ FApply

using tvm::meta_schedule::Postproc::FApply = ffi::TypedFunction<bool(const tir::Schedule&)>

Apply a postprocessor to the given schedule.

Parameters
schThe schedule to be post processed.
Returns
Whether the postprocessor was successfully applied.

◆ FAsString

using tvm::meta_schedule::Postproc::FAsString = ffi::TypedFunction<String()>

Get the postprocessor function as string with name.

Returns
The string of the postprocessor function.

◆ FClone

using tvm::meta_schedule::Postproc::FClone = ffi::TypedFunction<Postproc()>

Clone the postprocessor.

Returns
The cloned postprocessor.

◆ FInitializeWithTuneContext

The function type of InitializeWithTuneContext method.

Parameters
contextThe tuning context for initialization.

Member Function Documentation

◆ DefaultCPUTensorization()

static Array<Postproc, void> tvm::meta_schedule::Postproc::DefaultCPUTensorization ( )
static

Create default postprocessors for x86 (AVX512 and VNNI)

◆ DefaultCUDA()

static Array<Postproc, void> tvm::meta_schedule::Postproc::DefaultCUDA ( )
static

Create default postprocessors for CUDA.

◆ DefaultCUDATensorCore()

static Array<Postproc, void> tvm::meta_schedule::Postproc::DefaultCUDATensorCore ( )
static

Create default postprocessors for CUDA with TensorCore.

◆ DefaultHexagon()

static Array<Postproc, void> tvm::meta_schedule::Postproc::DefaultHexagon ( )
static

Create default postprocessors for Hexagon.

◆ DefaultLLVM()

static Array<Postproc, void> tvm::meta_schedule::Postproc::DefaultLLVM ( )
static

Create default postprocessors for LLVM.

◆ DisallowAsyncStridedMemCopy()

static Postproc tvm::meta_schedule::Postproc::DisallowAsyncStridedMemCopy ( )
static

Create a postprocessor that checks if all async mem copies are not strided.

Returns
The postprocessor created

◆ DisallowDynamicLoop()

static Postproc tvm::meta_schedule::Postproc::DisallowDynamicLoop ( )
static

Create a postprocessor that checks if all loops are static.

Returns
The postprocessor created

◆ PyPostproc()

static Postproc tvm::meta_schedule::Postproc::PyPostproc ( FInitializeWithTuneContext  f_initialize_with_tune_context,
FApply  f_apply,
FClone  f_clone,
FAsString  f_as_string 
)
static

Create a postprocessor with customized methods on the python-side.

Parameters
f_initialize_with_tune_contextThe packed function of InitializeWithTuneContext.
f_applyThe packed function of Apply.
f_cloneThe packed function of Clone.
f_as_stringThe packed function of AsString.
Returns
The postprocessor created.

◆ RewriteCooperativeFetch()

static Postproc tvm::meta_schedule::Postproc::RewriteCooperativeFetch ( )
static

Create a postprocessor that rewrites the cooperative fetch annotation to actual vectorized cooperative fetching in loop bindings.

Returns
The postprocessor created.

◆ RewriteLayout()

static Postproc tvm::meta_schedule::Postproc::RewriteLayout ( )
static

Creates a postprocessor that rewrites the layout of input tensor.

Note
Weight layout rewrite is supported so far, activation layout rewrite will be added.
Returns
The postprocessor created

◆ RewriteParallelVectorizeUnroll()

static Postproc tvm::meta_schedule::Postproc::RewriteParallelVectorizeUnroll ( )
static

Creates a postprocessor that applies parallelization, vectorization and auto unrolling according to the annotation of each block.

Returns
The postprocessor created

◆ RewriteReductionBlock()

static Postproc tvm::meta_schedule::Postproc::RewriteReductionBlock ( )
static

Create a postprocessor that rewrites reduction block by moving the init block out.

Returns
The postprocessor created.

◆ RewriteTensorize()

static Postproc tvm::meta_schedule::Postproc::RewriteTensorize ( bool  vectorize_init_loop = false)
static

Create a postprocessor that applies tensorization to annotated blocks.

Parameters
vectorize_init_loopWhether or not vectorize the initialization loop produced by DecomposeReduction
Returns
The postprocessor created.

◆ RewriteUnboundBlock()

static Postproc tvm::meta_schedule::Postproc::RewriteUnboundBlock ( int  max_threadblocks)
static

Create a postprocessor that adds thread binding to unbound blocks.

Parameters
max_threadblocksThe max number of threadblocks in the cuda device.
Returns
The postprocessor created.

◆ TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS()

tvm::meta_schedule::Postproc::TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS ( Postproc  ,
ObjectRef  ,
PostprocNode   
)

◆ VerifyGPUCode()

static Postproc tvm::meta_schedule::Postproc::VerifyGPUCode ( )
static

Creates a postprocessor that verifies if the GPU code is correct.

Returns
The postprocessor created

◆ VerifyVTCMLimit()

static Postproc tvm::meta_schedule::Postproc::VerifyVTCMLimit ( )
static

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

Returns
The postprocessor created

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