feature/inference analysis

[Superjomn]

Inference Analysis

Borrow some ideas from LLVM/analysis, the analysis phase optimize the codes(similar concept as ProgramDesc) there is some analysis passes, the original ProgramDesc will be translated to a lower computation graph representation for simple operation, and the graph will pass through all the passes one by one just like a pipe.

For example

ProgramDesc -> DFG(DataFlowGraph) ->
# following are the analyasis/optimization passes
# just a demo, some passes havn't been implemented
InferenceOpPrunePass -> TensorRTSubgrahpSplitterPass -> InferenceMemoryReusePass ->

DFG -> ProgramDesc -> Executor

All the passes can be controled by GFLAGS, and turned on/off by some configurations.

This is an Analysis framework designed to contain all the transpiler like analysis/optimization, which rewrites a ProgramDesc to an optimized one.

code structure

• graph representation related
  • data_flow_graph.cc/.h
  • node
  • graph_traits some graph retrival algorithms
• Pass related
  • pass.h define the basic Pass interfaces
  • pass_manager
  • fluid_to_data_flow_graph_pass transform a data flow graph from a fluid ProgramDesc
  • tensorrt_subgraph_pass split sub-graphs for TensorRT
  • data_flow_graph_to_fluid_pass translate from a data flow graph to a fluid ProgramDesc that can be executed by a Executor
• Helpers
  • dot, help to visualize a graph for better debug
  • subgraph_splitter split a graph according to some conditions
  • helper some helper methods

status

• ready to review
• working on some changes

[luotao1]

Since this PR is large, could it split into several small PRs? Maybe one concept with its test one PR?

node_tester.cc
dot_tester.cc
data_flow_graph_tester.cc
subgraph_splitter_tester.cc
fluid_to_data_flow_graph_pass_tester.cc
data_flow_graph_to_fluid_pass_tester.cc