From 5e2656449c5c193a9b62c13056fccb23b4306a8c Mon Sep 17 00:00:00 2001
From: Yan Chunwei <yanchunwei@outlook.com>
Date: Fri, 29 Jun 2018 17:09:28 +0800
Subject: [PATCH] add inference-analysis doc (#11813)

---
 paddle/fluid/inference/analysis/README.md | 57 +++++++++++++++++++++++
 1 file changed, 57 insertions(+)
 create mode 100644 paddle/fluid/inference/analysis/README.md

diff --git a/paddle/fluid/inference/analysis/README.md b/paddle/fluid/inference/analysis/README.md
new file mode 100644
index 00000000000..4c5de189cd1
--- /dev/null
+++ b/paddle/fluid/inference/analysis/README.md
@@ -0,0 +1,57 @@
+# Inference Analysis
+
+The `inference/analysis` module is used to analyze and optimize the inference program,
+it references some philosophy from `LLVM/analysis`, 
+and make the various optimization features be pluggable and co-exist in a pipeline.
+
+We borrowed some concepts from LLVM, such as
+
+- [Pass](./pass.h)es to implement optimization that traverse the inference program,
+- [DataFlowGraph](./data_flow_graph.h) to represent the data flow graph built from a program,
+- [PassManager](./pass_manager.h) to manage a sequence of `Pass`es over a graph.
+
+There are some other basic concepts here
+
+- [Node](./node.h), the node in a `DataFlowGraph`,
+  - `Function`, the Operator in Fluid,
+  - `Value`, the Variable in Fluid;
+- [Argument](./argument.h), the argument that treat as the input and output of all `Pass`es in the pipeline,
+
+## How it works
+
+The `inference/analysis` module make all the passes in a pipeline, and works in such way:
+
+1. Build a `DataFlowGraph` from a Fluid inference ProgramDesc,
+2. Call the middle passes one by one, the same `DataFlowGraph` is passed across all the passes,
+3. Transform a new ProgramDesc from the modified `DataFlowGraph`.
+
+The new optimization features can be added as an independent `Pass` and controlled by gflags,
+each pass will generate unified debug information or visualization for better debugging.
+
+## Supported Passes
+
+### `FluidToDataFlowGraphPass`
+Transform the fluid `ProgramDesc` to a `DataFlowGraph` to give an abstract representation for all the middle passes, 
+this should be the first pass of the pipeline.
+
+### `DataFlowGraphToFluidPass`
+Generate a final `ProgramDesc` from a data flow graph, this should be the last pass of the pipeline.
+
+### `TensorRTSubgraphNodeMarkPass`
+Mark the `Node` that are supported by TensorRT, 
+this pass will generate a visualization file which can be used for debugging.
+
+### `TensorRTSubGraphPass`
+Split the sub-graph that are can be accelerated by TensorRT.
+
+### `DFG_GraphvizDrawPass`
+This pass is just for debug, it will visualize the `DataFlowGraph` using the [graphviz](http://www.graphviz.org) tool.
+
+It can be used as a helper class that draws the modified graph after each pass.
+
+## Utilities
+
+There is some helper function/class for analysis.
+
+- [dot.h](./dot.h) give a easy to use interface for generating `DOT` codes,
+- [graph_traits.h](./graph_traits.h) contains the graph traversal algorithms, it uses `iterator` to make the algorithms easy to share across different passes.
-- 
GitLab