tvm
Public Member Functions | Public Attributes | Static Public Attributes | List of all members
tvm::runtime::cuda_ipc::CUDAIPCMemoryObj Class Reference

The CUDA IPC (interprocess communication) memory object, which internally contains data pointers to CUDA IPC memory. It is be useful for efficient all-reduce implementation. More...

#include <cuda_ipc_memory.h>

Inheritance diagram for tvm::runtime::cuda_ipc::CUDAIPCMemoryObj:
Collaboration diagram for tvm::runtime::cuda_ipc::CUDAIPCMemoryObj:

Public Member Functions

 TVM_DECLARE_BASE_OBJECT_INFO (CUDAIPCMemoryObj, Object)
 
- Public Member Functions inherited from tvm::runtime::Object
uint32_t type_index () const
 
std::string GetTypeKey () const
 
size_t GetTypeKeyHash () const
 
template<typename TargetType >
bool IsInstance () const
 
bool unique () const
 
 Object ()
 
 Object (const Object &other)
 
 Object (Object &&other)
 
Objectoperator= (const Object &other)
 
Objectoperator= (Object &&other)
 

Public Attributes

int num_workers
 The number of GPU workers. More...
 
int worker_id
 The worker id corresponding to this IPC memory object. More...
 
std::vector< void * > remote_data
 The data pointers of all all-reduce inputs. It has "num_workers" pointers. The i-th pointer is the data pointer on worker i. If "i != worker_id", the pointer is an IPC data pointer. Otherwise, the pointer is a local CUDA data pointer. More...
 
std::vector< void * > barrier_in
 The pointers to input barrier signals of all workers for all-reduce. It has "num_workers" pointers, and the pointer arrangement is the same as "remote_data". More...
 
std::vector< void * > barrier_out
 The pointers to output barrier signals of all workers for all-reduce. It has "num_workers" pointers, and the pointer arrangement is the same as "remote_data". More...
 
int barrier_flag
 The integer buffer flag for all-reduce. More...
 

Static Public Attributes

static constexpr const char * _type_key = "tvm.runtime.disco.cuda_ipc_memory"
 
static constexpr const bool _type_has_method_sequal_reduce = false
 
static constexpr const bool _type_has_method_shash_reduce = false
 
- Static Public Attributes inherited from tvm::runtime::Object
static constexpr const char * _type_key = "runtime.Object"
 
static constexpr bool _type_final = false
 
static constexpr uint32_t _type_child_slots = 0
 
static constexpr bool _type_child_slots_can_overflow = true
 
static constexpr bool _type_has_method_visit_attrs = true
 
static constexpr bool _type_has_method_sequal_reduce = false
 
static constexpr bool _type_has_method_shash_reduce = false
 
static constexpr uint32_t _type_index = TypeIndex::kDynamic
 

Additional Inherited Members

- Public Types inherited from tvm::runtime::Object
typedef void(* FDeleter) (Object *self)
 Object deleter. More...
 
using RefCounterType = std::atomic< int32_t >
 
- Static Public Member Functions inherited from tvm::runtime::Object
static std::string TypeIndex2Key (uint32_t tindex)
 Get the type key of the corresponding index from runtime. More...
 
static size_t TypeIndex2KeyHash (uint32_t tindex)
 Get the type key hash of the corresponding index from runtime. More...
 
static uint32_t TypeKey2Index (const std::string &key)
 Get the type index of the corresponding key from runtime. More...
 
static uint32_t _GetOrAllocRuntimeTypeIndex ()
 
static uint32_t RuntimeTypeIndex ()
 
- Protected Member Functions inherited from tvm::runtime::Object
void IncRef ()
 developer function, increases reference counter. More...
 
void DecRef ()
 developer function, decrease reference counter. More...
 
- Static Protected Member Functions inherited from tvm::runtime::Object
static uint32_t GetOrAllocRuntimeTypeIndex (const std::string &key, uint32_t static_tindex, uint32_t parent_tindex, uint32_t type_child_slots, bool type_child_slots_can_overflow)
 Get the type index using type key. More...
 
- Protected Attributes inherited from tvm::runtime::Object
uint32_t type_index_ {0}
 Type index(tag) that indicates the type of the object. More...
 
RefCounterType ref_counter_ {0}
 The internal reference counter. More...
 
FDeleter deleter_ = nullptr
 deleter of this object to enable customized allocation. If the deleter is nullptr, no deletion will be performed. The creator of the object must always set the deleter field properly. More...
 

Detailed Description

The CUDA IPC (interprocess communication) memory object, which internally contains data pointers to CUDA IPC memory. It is be useful for efficient all-reduce implementation.

Note
Right now the class members are closely tied with customized all-reduce kernel. They may also be extended for other uses in the future.

Member Function Documentation

◆ TVM_DECLARE_BASE_OBJECT_INFO()

tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::TVM_DECLARE_BASE_OBJECT_INFO ( CUDAIPCMemoryObj  ,
Object   
)

Member Data Documentation

◆ _type_has_method_sequal_reduce

constexpr const bool tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::_type_has_method_sequal_reduce = false
staticconstexpr

◆ _type_has_method_shash_reduce

constexpr const bool tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::_type_has_method_shash_reduce = false
staticconstexpr

◆ _type_key

constexpr const char* tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::_type_key = "tvm.runtime.disco.cuda_ipc_memory"
staticconstexpr

◆ barrier_flag

int tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::barrier_flag

The integer buffer flag for all-reduce.

◆ barrier_in

std::vector<void*> tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::barrier_in

The pointers to input barrier signals of all workers for all-reduce. It has "num_workers" pointers, and the pointer arrangement is the same as "remote_data".

◆ barrier_out

std::vector<void*> tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::barrier_out

The pointers to output barrier signals of all workers for all-reduce. It has "num_workers" pointers, and the pointer arrangement is the same as "remote_data".

◆ num_workers

int tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::num_workers

The number of GPU workers.

◆ remote_data

std::vector<void*> tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::remote_data

The data pointers of all all-reduce inputs. It has "num_workers" pointers. The i-th pointer is the data pointer on worker i. If "i != worker_id", the pointer is an IPC data pointer. Otherwise, the pointer is a local CUDA data pointer.

◆ worker_id

int tvm::runtime::cuda_ipc::CUDAIPCMemoryObj::worker_id

The worker id corresponding to this IPC memory object.


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