Merge commit '33b8b3d22034423455a493712955e419aac7b19b' into develop

* commit '33b8b3d22034423455a493712955e419aac7b19b': (251 commits)
  Remove redundant commands in build.sh and build_doc.sh
  Add dependencies
  Move v2/api/fluid to fluid/api and Adjust doc build commands
  Plain LRN op throws an exception when is_test is set in backward pass
  fix compiler error of profiler_test in ONLY_CPU mode
  fix server shutdown
  Translation for Model Configuration (PaddlePaddle#9513)
  Fix data transform when inplace (PaddlePaddle#9450)
  refine parallel
  add FAQ (PaddlePaddle#9494)
  Fix dist error with lr decay layer (PaddlePaddle#9489)
  add prefetch_op (PaddlePaddle#9495)
  Fix some errors (PaddlePaddle#9403)
  hookup WITH_FLUID_ONLY in TeamCity build.sh (PaddlePaddle#9509)
  Fix the order of reads and write from buffered channel  (PaddlePaddle#9423)
  change WITH_FLUID to WITH_FLUID_ONLY (PaddlePaddle#9427)
  fix block num
  Revert "make append activation in place by default (PaddlePaddle#9417)"
  Speed/sequence op1 (PaddlePaddle#9217)
  fix a compile error
  ...

CMakeLists.txt

@@ -53,8 +53,7 @@ option(WITH_COVERAGE    "Compile PaddlePaddle with code coverage"       OFF)
 option(COVERALLS_UPLOAD "Package code coverage data to coveralls"       OFF)
 option(ON_TRAVIS        "Exclude special unit test on Travis CI"        OFF)
 option(WITH_C_API       "Compile PaddlePaddle with C-API(Prediction)"   OFF)
-# TODO: Only compile PaddlePaddle fluid version by WITH_FLUID option. 
-option(WITH_FLUID       "Compile PaddlePaddle fluid only(TODO)"         OFF)
+option(WITH_FLUID_ONLY  "Compile PaddlePaddle fluid only"               OFF)
 option(WITH_GOLANG      "Compile PaddlePaddle with GOLANG"              OFF)
 option(GLIDE_INSTALL    "Download and install go dependencies "         ON)
 option(USE_NNPACK       "Compile PaddlePaddle with NNPACK library"      OFF)
@@ -109,7 +108,7 @@ if (WITH_C_API AND WITH_PYTHON)
 endif()
 
 if (WITH_C_API)
-  set(WITH_FLUID OFF CACHE STRING "Disable install fluid when compile the C_API" FORCE)
+  set(WITH_FLUID_ONLY OFF CACHE STRING "Disable install fluid when compile the C_API" FORCE)
 endif()
 
 if(MOBILE_INFERENCE)
@@ -147,6 +146,7 @@ include(external/cares)
 include(external/grpc)
 include(external/snappy)    # download snappy
 include(external/snappystream)
+include(external/threadpool)
 
 include(cudnn)              # set cudnn libraries, must before configure
 include(cupti)

cmake/external/mklml.cmake

@@ -34,7 +34,7 @@ SET(MKLML_DOWNLOAD_DIR  "${MKLML_SOURCE_DIR}/src/${MKLML_PROJECT}")
 SET(MKLML_DST_DIR       "mklml")
 SET(MKLML_INSTALL_ROOT  "${THIRD_PARTY_PATH}/install")
 SET(MKLML_INSTALL_DIR   ${MKLML_INSTALL_ROOT}/${MKLML_DST_DIR})
-SET(MKLML_ROOT          ${MKLML_INSTALL_DIR}/${MKLML_VER})
+SET(MKLML_ROOT          ${MKLML_INSTALL_DIR})
 SET(MKLML_INC_DIR       ${MKLML_ROOT}/include)
 SET(MKLML_LIB_DIR       ${MKLML_ROOT}/lib)
 SET(MKLML_LIB           ${MKLML_LIB_DIR}/libmklml_intel.so)
@@ -46,7 +46,7 @@ INCLUDE_DIRECTORIES(${MKLML_INC_DIR})
 FILE(WRITE ${MKLML_DOWNLOAD_DIR}/CMakeLists.txt
   "PROJECT(MKLML)\n"
   "cmake_minimum_required(VERSION 3.0)\n"
-  "install(DIRECTORY ${MKLML_VER}\n"
+  "install(DIRECTORY ${MKLML_VER}/include ${MKLML_VER}/lib \n"
   "        DESTINATION ${MKLML_DST_DIR})\n")
 
 ExternalProject_Add(

cmake/external/threadpool.cmake

@@ -0,0 +1,30 @@
+INCLUDE(ExternalProject)
+
+SET(THREADPOOL_SOURCE_DIR ${THIRD_PARTY_PATH}/threadpool)
+SET(THREADPOOL_INCLUDE_DIR ${THREADPOOL_SOURCE_DIR}/src/extern_threadpool)
+INCLUDE_DIRECTORIES(${THREADPOOL_INCLUDE_DIR})
+
+ExternalProject_Add(
+    extern_threadpool
+    ${EXTERNAL_PROJECT_LOG_ARGS}
+    GIT_REPOSITORY  "https://github.com/progschj/ThreadPool.git"
+    GIT_TAG         9a42ec1329f259a5f4881a291db1dcb8f2ad9040
+    PREFIX          ${THREADPOOL_SOURCE_DIR}
+    UPDATE_COMMAND  ""
+    CONFIGURE_COMMAND ""
+    BUILD_COMMAND     ""
+    INSTALL_COMMAND   ""
+    TEST_COMMAND      ""
+)
+
+if (${CMAKE_VERSION} VERSION_LESS "3.3.0")
+    set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/threadpool_dummy.c)
+    file(WRITE ${dummyfile} "const char *dummy_threadpool = \"${dummyfile}\";")
+    add_library(simple_threadpool STATIC ${dummyfile})
+else()
+    add_library(simple_threadpool INTERFACE)
+endif()
+
+add_dependencies(simple_threadpool extern_threadpool)
+
+LIST(APPEND external_project_dependencies simple_threadpool)

cmake/generic.cmake

@@ -587,6 +587,9 @@ function(grpc_library TARGET_NAME)
   get_filename_component(PROTO_WE ${grpc_library_PROTO} NAME_WE)
   get_filename_component(PROTO_PATH ${ABS_PROTO} PATH)
 
+  #FIXME(putcn): the follwoing line is supposed to generate *.pb.h and cc, but
+  # somehow it didn't. line 602 to 604 is to patching this. Leaving this here 
+  # for now to enable dist CI.
   protobuf_generate_cpp(grpc_proto_srcs grpc_proto_hdrs "${ABS_PROTO}")
   set(grpc_grpc_srcs "${CMAKE_CURRENT_BINARY_DIR}/${PROTO_WE}.grpc.pb.cc")
   set(grpc_grpc_hdrs "${CMAKE_CURRENT_BINARY_DIR}/${PROTO_WE}.grpc.pb.h")
@@ -597,6 +600,9 @@ function(grpc_library TARGET_NAME)
           COMMAND ${PROTOBUF_PROTOC_EXECUTABLE}
           ARGS --grpc_out "${CMAKE_CURRENT_BINARY_DIR}" -I "${PROTO_PATH}"
           --plugin=protoc-gen-grpc="${GRPC_CPP_PLUGIN}" "${ABS_PROTO}"
+          COMMAND ${PROTOBUF_PROTOC_EXECUTABLE}
+          ARGS --cpp_out "${CMAKE_CURRENT_BINARY_DIR}" -I "${PROTO_PATH}"
+          "${ABS_PROTO}"
           DEPENDS "${ABS_PROTO}" ${PROTOBUF_PROTOC_EXECUTABLE} extern_grpc)
 
   # FIXME(typhoonzero): grpc generated code do not generate virtual-dtor, mark it

cmake/inference_lib.cmake

@@ -69,6 +69,12 @@ if(NOT CBLAS_FOUND)
       SRCS ${CBLAS_INSTALL_DIR}/lib ${CBLAS_INSTALL_DIR}/include
       DSTS ${dst_dir} ${dst_dir}
     )
+elseif (WITH_MKLML)
+    set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/mklml")
+    copy(mklml_lib
+      SRCS ${MKLML_LIB} ${MKLML_IOMP_LIB} ${MKLML_INC_DIR}
+      DSTS ${dst_dir}/lib ${dst_dir}/lib ${dst_dir}
+    )
 endif()
 
 # paddle fluid module

doc/CMakeLists.txt

@@ -1,2 +1,9 @@
+add_custom_target(paddle_apis ALL
+                  DEPENDS paddle_v2_apis paddle_fluid_apis)
+
+add_custom_target(paddle_docs ALL
+                  DEPENDS paddle_v2_docs paddle_v2_docs_cn
+                  paddle_fluid_docs paddle_fluid_docs_cn)
+
 add_subdirectory(v2)
 add_subdirectory(fluid)

doc/design/images/parallel_executor_overview.dot

@@ -0,0 +1,83 @@
+digraph G {
+  subgraph cluster_init {
+    label="Initialization"
+    startup_program [label="startup", shape=box]
+    node_w_g0 [label="W\nGPU0"]
+    startup_program -> node_w_g0 [label="Initialize"]
+    node_w_g1 [label="W\nGPU1"]
+    node_w_g0 -> node_w_g1 [label="broadcast"]
+  }
+
+  subgraph cluster_train {
+    label="forward_backward"
+
+    subgraph cluster_gpu0 {
+      label="GPU0"
+      fc_0 [label="fc\nGPU0", shape=box]
+      hidden_0 [label="hidden\nGPU0"]
+      node_w_g0 -> fc_0
+      fc_0 -> hidden_0
+      loss0 [label="loss\nGPU0"]
+      hidden_0 -> loss0 [label="many ops omitted"]
+      scale_loss_0 [label="scale_loss_gradient\nGPU0", shape=box]
+      loss_g0 [label="loss_grad\nGPU0"]
+      scale_loss_0->loss_g0
+
+      fc_g_0 [label="w_grad\nGPU0", shape=box]
+      loss0 -> fc_g_0
+      loss_g0 -> fc_g_0
+      hidden_0 -> fc_g_0
+    }
+
+    subgraph cluster_gpu1 {
+      label="GPU1"
+      fc_1 [label="fc\nGPU1", shape=box]
+      hidden_1 [label="hidden\nGPU1"]
+      node_w_g1 -> fc_1
+      fc_1 -> hidden_1
+      loss1 [label="loss\nGPU1"]
+      hidden_1 -> loss1 [label="many ops omitted"]
+      scale_loss_1 [label="scale_loss_gradient\nGPU1", shape=box]
+      loss_g1 [label="loss_grad\nGPU1"]
+      scale_loss_1->loss_g1
+
+      fc_g_1 [label="w_grad\nGPU1", shape=box]
+      loss1 -> fc_g_1
+      loss_g1 -> fc_g_1
+      hidden_1 -> fc_g_1
+    }
+  }
+
+  all_reduce_w [label="Merge Gradients(AllReduce)", shape=box]
+  fc_g_0 -> all_reduce_w
+  fc_g_1 -> all_reduce_w
+
+  fc_g_0_merged [label="w_grad\nMerged\nGPU0"]
+  fc_g_1_merged [label="w_grad\nMerged\nGPU1"]
+  all_reduce_w -> fc_g_0_merged
+  all_reduce_w -> fc_g_1_merged
+
+  subgraph cluster_optimization {
+    label="Optimization"
+    subgraph cluster_opt_gpu0 {
+      label="GPU0"
+      sgd_0 [label="SGD Op\nGPU0", shape=box]
+
+      fc_g_0_merged -> sgd_0
+      node_w_g0 -> sgd_0
+      optimized_w_0 [label="Optimized W\nGPU0"]
+      sgd_0 -> optimized_w_0
+    }
+    subgraph cluster_opt_gpu1 {
+      label="GPU1"
+      sgd_1 [label="SGD Op\nGPU1", shape=box]
+
+      fc_g_1_merged -> sgd_1
+      node_w_g1 -> sgd_1
+      optimized_w_1 [label="Optimized W\nGPU0"]
+      sgd_1 -> optimized_w_1
+    }
+  }
+
+
+}

doc/design/parallel_executor.md

@@ -0,0 +1,104 @@
+# ParallelExecutor
+
+## Background
+
+Neural network models are defined as a `ProgramDesc` in Fluid. The `ProgramDesc` can be executed by an interpreter(i.e. the `executor` concept in Fluid). The instructions or operators in a `Program` will be executed, and the results will be fetched in Python side.
+
+The executor is a very naive interpreter. It runs operators one by one. We can use `Parallel.Do` to support data parallelism, however, lacking device information in `ProgramDesc`; it is not possible to optimize the performance of `Parallel.Do`.
+
+We want a `ProgramDesc` can be run on different nodes. It is better not to contain device information in `ProgramDesc`. However, we can write a high-performance interpreter, which can hold an alternative intermediate representation of `ProgramDesc`, to take full usage of Multi-GPUs. 
+
+ParallelExecutor is an interpreter of `ProgramDesc` which will [out-of-order execute](https://en.wikipedia.org/wiki/Out-of-order_execution) `Program` in data parallelism mode and maximise the utility of Multi-GPUs.
+
+
+## Overview of MultiGPUs logic
+
+The ParallelExecutor takes the startup program and main program as inputs. The parameters will be initialised on `GPU0` by startup program and will broadcast to multi-GPUs. The main program will be duplicated into multi-GPUs. The gradient will be merged during each iteration, and each device will optimize parameters independently. Since the gradients on each device will be merged before parameter optimization, the parameters will be the same on each device and it does not need to be broadcast the parameters.
+
+![alt](images/parallel_executor_overview.png)
+
+There are several optimizations for this logic.
+
+1. We use an alternate representation in ParallelExecutor. It because the device information is critical for performance optimization.
+2. The execution is out-of-order, i.e., an operator will be executed whenever the inputs of the operator are ready. 
+   * GPU is a high-performance device; only one CPU thread cannot fulfil one GPU. So there is a thread pool to execute operators.
+   * Out-of-order also helps transpilers to generate `ProgramDesc`. It is no need to concern about the best order of performance when implementing a transpiler.
+3. The streams of computation, merge gradients and fetch data are different.
+
+The performance of `ResNeXt152` on `TitanX` which `batch_size=12` is shown below.
+
+| Number of GPUs | 1 | 2 | 3 | 4|
+| --- | --- | --- | --- | --- |
+| Image/Sec | 17.9906 | 25.771 | 36.911 | 48.8428 |
+| Speed Up | N/A | 1.43247029 | 2.05168255 | 2.71490667 |
+
+
+## Static single assignment Graph
+
+[Static single assignment form](https://en.wikipedia.org/wiki/Static_single_assignment_form)(`SSA` for short) is a common form for compiler optimization. To implement concurrent execution, we uses an `SSA` graph as an intermedia representation of `ProgramDesc`.
+
+The `Program` is a directed acyclic graph, since a variable can be assigned multiple times. We enforce a variable will be assigned once, by adding version number to varaibles. We parsing the `Program` into a `SSA` graph. Also, ProgramExecutor duplicate `Program` into multi-devices. We also add a device number to varaibles and insert `NCCLAllReduce` into Graph.
+
+The data structure of `SSA` graph is:
+
+```c++
+struct VarHandleBase {
+  OpHandleBase* generated_op_;
+  vector<OpHandleBase*> pending_ops_;
+
+  string name;
+  Place place;
+  size_t version;
+};
+
+struct OpHandleBase {
+  vector<OpHandleBase*> inputs_;
+  vector<OpHnadleBase*> outputs_;
+};
+
+struct SSAGraph {
+  // vars on each devices. 
+  //   * the vars in each map in vector is on different device.
+  //   * the map is mapping a variable name to variable handles
+  //   with different versions
+  vector<std::unordered_map<string, vector<VarHandleBase>>> vars_;
+
+  // All ops
+  vector<OpHandleBase> ops_;
+};
+```
+The variable handles are the wrapper of `Variables`. The operator handles are the wrapper of `OperatorBase`. Some `OpHandle` is not an `OperatorBase`, such as `NCCLAllReduceOpHandle`, because `AllReduceOpHandle` will use new device contexts.
+
+When the `ProgramDesc` converted into an `SSA` Graph, the [data hazard](https://en.wikipedia.org/wiki/Hazard_(computer_architecture)) problem is also need to be taken care. The dummy variables, which represent the dependency between operators, will be manually inserted into SSA graph to resolve the [data hazard](https://en.wikipedia.org/wiki/Hazard_(computer_architecture)) problem.
+
+## Execute SSA Graph
+
+The SSA graph can be out-of-order executed by an approximate [topological sorting](https://en.wikipedia.org/wiki/Topological_sorting) algorithm. The algorithm is
+
+1. Maintaining a map of an operator and its needed input number.
+2. If a variable is not generated by an operator, i.e., `var.generated_op == nullptr`, decrease the needed input number of its pending operators.
+3. If there is an operator which needed input number is decreased to zero, just run this operator.
+4. After run this operator, just mark the variables are generated and repeat step 2 until all variables are generated.
+
+Running an operator can be asynchronized. There is a thread pool to execute an `SSA` graph.
+
+## Synchronize GPU Kernels
+
+The GPU is a non-blocking device. The different streams need be synchronized when switing streams. In current implementation, the synchronization based on the following algorithm:
+
+1. `OpHandle` will record `DeviceContext` that it is used.
+2. In `OpHandle::Run`, if the `DeviceContext` of current operator is different from `DeviceContext` of any input variable, just wait the generate operator of this input variable.
+
+The `wait` are implemented by two strategies:
+
+1. Invoke `DeviceContext->Wait()`, It will wait all operators on this device contexts complete.
+2. Uses `cudaStreamWaitEvent` to sending a event to the stream. It is a non-blocking call. The wait operators will be executed in GPU.
+
+Generally, the `cudaStreamWaitEvent` will have a better perforamnce. However, `DeviceContext->Wait()` strategy is easier to debug. The strategy can be changed in runtime.
+
+## What's next?
+
+* Merging gradient of dense parameters has been done. However, the merging of sparse parameters has not been done.
+* The CPU version of Parallel Executor has not been implemented. The out-of-order logic will make CPU compuatation faster, too.
+* A better strategy to merge gradients can be introduced. We can shrink the gradients from `float32` to `int8` or `int4` while merging. It will significantly speed up multi-GPUs training without much loss of precision.
+* Combine multi-Nodes implementation. By the benifit of out-of-order, sending and recving operator can be an blocking operator, and the transpiler does not need to concern about the best position of operator.

doc/fluid/CMakeLists.txt

@@ -27,6 +27,8 @@ sphinx_add_target(paddle_fluid_docs
                   ${CMAKE_CURRENT_SOURCE_DIR}
                   ${SPHINX_HTML_DIR_EN})
 
+add_dependencies(paddle_fluid_docs gen_proto_py)
+
 # configured documentation tools and intermediate build results
 set(BINARY_BUILD_DIR_CN "${CMAKE_CURRENT_BINARY_DIR}/cn/_build")
 
@@ -47,3 +49,7 @@ sphinx_add_target(paddle_fluid_docs_cn
                   ${SPHINX_CACHE_DIR_CN}
                   ${CMAKE_CURRENT_SOURCE_DIR}
                   ${SPHINX_HTML_DIR_CN})
+
+add_dependencies(paddle_fluid_docs_cn gen_proto_py)
+
+add_subdirectory(api)

doc/fluid/api/CMakeLists.txt

@@ -0,0 +1,22 @@
+# configured documentation tools and intermediate build results
+set(BINARY_BUILD_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_build")
+
+# Sphinx cache with pickled ReST documents
+set(SPHINX_CACHE_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/_doctrees")
+
+# HTML output director
+set(SPHINX_HTML_DIR_EN "${CMAKE_CURRENT_BINARY_DIR}/en/html")
+
+configure_file(
+    "${CMAKE_CURRENT_SOURCE_DIR}/../../templates/conf.py.en.in"
+    "${BINARY_BUILD_DIR_EN}/conf.py"
+    @ONLY)
+
+sphinx_add_target(paddle_fluid_apis
+                  html
+                  ${BINARY_BUILD_DIR_EN}
+                  ${SPHINX_CACHE_DIR_EN}
+                  ${CMAKE_CURRENT_SOURCE_DIR}
+                  ${SPHINX_HTML_DIR_EN})
+
+add_dependencies(paddle_fluid_apis  gen_proto_py framework_py_proto copy_paddle_pybind)

doc/fluid/build_and_install/build_from_source_cn.rst

@@ -0,0 +1 @@
+../../v2/build_and_install/build_from_source_cn.rst

doc/fluid/build_and_install/build_from_source_en.rst

@@ -0,0 +1 @@
+../../v2/build_and_install/build_from_source_en.rst

doc/fluid/build_and_install/docker_install_cn.rst

@@ -0,0 +1 @@
+../../v2/build_and_install/docker_install_cn.rst

doc/fluid/build_and_install/docker_install_en.rst

@@ -0,0 +1 @@
+../../v2/build_and_install/docker_install_en.rst

doc/fluid/build_and_install/index_cn.rst

@@ -0,0 +1 @@
+../../v2/build_and_install/index_cn.rst

doc/fluid/build_and_install/index_en.rst

@@ -0,0 +1 @@
+../../v2/build_and_install/index_en.rst

doc/fluid/build_and_install/pip_install_cn.rst

@@ -0,0 +1 @@
+../../v2/build_and_install/pip_install_cn.rst

doc/fluid/build_and_install/pip_install_en.rst

@@ -0,0 +1 @@
+../../v2/build_and_install/pip_install_en.rst

doc/fluid/design/algorithm/index_cn.rst

@@ -0,0 +1,7 @@
+梯度更新算法
+------------
+
+.. toctree::
+  :maxdepth: 1
+
+  parameter_average.md

doc/fluid/design/algorithm/index_en.rst

@@ -0,0 +1,7 @@
+Gradient Update Algorithm
+--------------------------------------
+
+.. toctree::
+  :maxdepth: 1
+
+  parameter_average.md