-
Executor: 执行Computation Sub Graph
- SubGraph预处理:Node/NodeItem/TaggedNode
- Flow control op: switch/merge/enter/exit/nextIteration
- Frame: ControlFlowInfo/FrameInfo/FrameState/IterationState
- Entry: 保存node之间的输入输出tensor
- Pending Count: node执行状态
- Execute Graph Computation: PrepareInputs/Process/ProcessOutputs/PropateOutputs
-
DirectSession: 单机执行computation graph
- SessionFactory
- Rewrite Graph
- CallFrame: Feed and Fetch
- Place node to device
- Graph Partition
- Create executors
- Tensor Store
-
- Rendezvous主要模块以及继承关系
- LocalRendezvous
- IntraProcessRendezvous
- CopyTensor::ViaDMA: GPU,CPU之间的通信
- BaseRemoteRendezvous
- RpcRemoteRenddezvous
- RendezvousMgr
-
- Device继承关系
- Device Factory
- Device Thread pool
- Device Context
- Device Allocator
-
OpKernal Context
-
Opkernal
- Variable
- Const
- PlaceHolder
- Matmul
- Conv
-
GrpSession: Grp分布式执行computation graph
-
Estimator
-
Dataset
-
tfslim
ahangchen / blog Goto Github PK
View Code? Open in Web Editor NEWThis project forked from xieyu/blog