对caffe2的一些初步体会（草稿）

Caffe2的一些关键设计思想：

• 所有运算都抽象为Operator。
• Blob和Tensor的概念。
• Blob和Net都存放在Workspace中，一个Workspace中可以有多个Net，这些Net中使用到的相同名称的Blob实际对应于这个Workspace中的同一个Blob。
• 将Net转化为DAG运算图，从而多线程实现并行运算。
• 每个Operator绑定在一个Context上，这个Operator创建的Tensor分配的内存都在这个Context上。
• Net，Blob，Tensor不具体绑定Context，与Context的绑定只在Operator上实现。
• Memonger分析Blob重用，节约内存。

具体代码分析

caffe2/core

typeid.h, typeid.cc, types.h, types.cc

CaffeTypeId

TypeNameRegisterer, UninitializedTypeNameRegisterer

TypeMeta: PlacementNew, TypedCopy, TypedDestructor

CAFFE_KNOWN_TYPE

StorageOrder

allocator.h，allocator.cc

Allocator：CPUAllocator，MemoryDeleter，MemoryAllocationReporter，DefaultCPUAllocator，

blob.h, blob_serializer_base.h, blob_serialization.h, blob_serialization.cc, blob_serialization_gpu.cc

Blob, 

BlobSerializerBase, TensorSerializer

BlobDeserializerBase, TensorDeserializer,

BlobStatGetter, BlobStatRegistry

common.h, common.cc, common_gpu.h, common_gpu.cc, common_omp.h

TIndex, CaffeMap, 

DeviceGuard

common_cudnn.h, common_cudnn.cc, cudnn_wrappers.h

cudnnTypeWrapper, 

cudnnTensorDescWrapper, cudnnFilterDescWrapper, 

CuDNNWrapper, CuDNNWorkspace, CuDNNState, 

context.h, context_gpu.h, context_gpu.cu

CPUContext, 

CudaMemoryPoolType, ThreadLocalCUDAObjects, CUDAContext

PinnedCPUAllocator

TensorCUDA

Caffe2CudaInitializerHelper

db.h, db.cc

Cursor，Transaction，DB，DBReader，

event.h, event.cc, event_gpu.cc

Event

EventCreateFunction, EventRecordFunction, EventWaitFunction, EventFinishFunction

EventCreateFunctionRegisterer, EventRecordFunctionRegisterer, EventWaitFunctionRegisterer, EventFinishFunctionRegisterer

REGISTER_EVENT_CREATE_FUNCTION, REGISTER_EVENT_RECORD_FUNCTION, REGISTER_EVENT_WAIT_FUNCTION, REGISTER_EVENT_FINISH_FUNCTION

CudaEventWrapper

flags.h, flags.cc

graph.h, graph.cc, transform.h, transform.cc

Node, Graph, 

Transform

init.h, init.cc, init_omp.cc

Caffe2InitializeRegistry, InitRegisterer

REGISTER_CAFFE2_INIT_FUNCTION, REGISTER_CAFFE2_EARLY_INIT_FUNCTION

logging.h, logging.cc, logging_is_google_glog.h, logging_is_not_google_glog.h

macros.h, macros.h.in

memonger.h, memonger.cc

optimize_inference_net

compute_blob_recycling_for_dag

module.h, module.cc

ModuleSchema

net.h, net.cc, 

net_simple.h, net_simple.cc, net_simple_async.h, net_simple_async.cc, 

net_dag.h, net_dag.cc, net_async_dag_gpu.h, net_async_dag_gpu.cc

NetBase, SimpleNet, AsyncSimpleNet, 

OperatorNode, OpGraphNode, 

DAGNetBase, DAGNet, AsyncDAGNet

observer.h

ObserverBase, Observable

operator.h, operator.cc, operator_schema.h, operator_schema.cc, operator_gradient.h

OperatorBase, Operator, 

CPUOperatorRegistry, CUDAOperatorRegistry, GradientRegistry, 

GradientWrapper, GradientOpsMeta, GradientMakerBase, NoGradient, 

OpSchema, OpSchemaRegistry

plan_executor.h, plan_executor.cc

predictor.h, predictor.cc

tensor.h, tensor.cc, qtensor.h, qtensor.cc, qtensor_serialization.h, qtensor_serialization.cc

Tensor, TensorCPU, TensorPrinter, TensorCPUStatGetter

QTensor

registry.h

Registry, Registerer

scope_guard.h

static_tracepoint.h, static_tracepoint.elfx86.h

stats.h, stats.cc

timer.h

Timer

workspace.h, workspace.cc

StopOnSignal, Workspace

cudnnTypeWrapper, 

cudnnTensorDescWrapper, cudnnFilterDescWrapper,