Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into port_python3_syntax

Author: velconia

.travis.yml

@@ -27,15 +27,6 @@ script:
     # 43min timeout
     paddle/scripts/paddle_docker_build.sh ${JOB}
     if [ $? -eq 0 ] || [ $? -eq 142 ]; then true; else exit 1; fi;
-  - |
-    if [[ "$JOB" != "doc" ]]; then exit 0; fi;
-    # For document only
-    if [[ "$TRAVIS_PULL_REQUEST" != "false" ]]; then exit 0; fi;
-    if [[ "$TRAVIS_BRANCH" != "develop"  && ! "$TRAVIS_BRANCH" =~ ^v|release/[[:digit:]]+\.[[:digit:]]+(\.[[:digit:]]+)?(-\S*)?$ ]]; then exit 0; fi;
-    export DEPLOY_DOCS_SH=https://raw.githubusercontent.com/PaddlePaddle/PaddlePaddle.org/master/scripts/deploy/deploy_docs.sh
-    export DOCS_DIR=`pwd`
-    cd ..
-    curl $DEPLOY_DOCS_SH | bash -s $CONTENT_DEC_PASSWD $TRAVIS_BRANCH $DOCS_DIR $DOCS_DIR/build/doc/
 notifications:
   email:
     on_success: change

CMakeLists.txt

@@ -65,6 +65,7 @@ option(REPLACE_ENFORCE_GLOG "Replace PADDLE_ENFORCE with glog/CHECK for better d
 option(WITH_ANAKIN      "Compile with Anakin library"                   OFF)
 option(WITH_GRPC     "Use grpc as the default rpc framework"            ${WITH_DISTRIBUTE})
 option(WITH_BRPC_RDMA     "Use brpc rdma as the rpc protocal"           OFF)
+option(WITH_INFERENCE    "Compile fluid inference library"              ON)
 option(WITH_SYSTEM_BLAS   "Use system blas library"           OFF)
 option(PY_VERSION       "Compile PaddlePaddle with python3 support"     ${PY_VERSION})
 
@@ -175,6 +176,7 @@ include(external/any)       # download libn::any
 include(external/eigen)     # download eigen3
 include(external/pybind11)  # download pybind11
 include(external/cares)
+include(external/cub)
 
 if(WITH_DISTRIBUTE)
     if(WITH_GRPC)

cmake/external/cub.cmake

@@ -0,0 +1,35 @@
+if(NOT WITH_GPU)
+  return()
+endif()
+
+include(ExternalProject)
+
+set(CUB_SOURCE_DIR ${THIRD_PARTY_PATH}/cub)
+set(CUB_INCLUDE_DIR ${CUB_SOURCE_DIR}/src/extern_cub)
+
+include_directories(${CUB_INCLUDE_DIR})
+
+ExternalProject_Add(
+  extern_cub
+  ${EXTERNAL_PROJECT_LOG_ARGS}
+  GIT_REPOSITORY "https://github.com/NVlabs/cub.git"
+  GIT_TAG        "v1.8.0"
+  PREFIX         ${CUB_SOURCE_DIR}
+  UPDATE_COMMAND ""
+  CONFIGURE_COMMAND ""
+  BUILD_COMMAND     ""
+  INSTALL_COMMAND   ""
+  TEST_COMMAND      ""
+)
+
+if(${CMAKE_VERSION} VERSION_LESS "3.3.0")
+  set(dummyfile ${CMAKE_CURRENT_BINARY_DIR}/cub_dummy.c)
+  file(WRITE ${dummyfile} "const char *dummy = \"${dummyfile}\";")
+  add_library(cub STATIC ${dummyfile})
+else()
+  add_library(cub INTERFACE)
+endif()
+
+add_dependencies(cub extern_cub)
+
+LIST(APPEND externl_project_dependencies cub)

cmake/generic.cmake

@@ -264,7 +264,10 @@ function(cc_test TARGET_NAME)
              WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
     if (${cc_test_SERIAL})
         set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
+
+    set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
+    set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
     endif()
   endif()
 endfunction(cc_test)
@@ -329,7 +332,10 @@ function(nv_test TARGET_NAME)
     add_test(${TARGET_NAME} ${TARGET_NAME})
     if (nv_test_SERIAL)
         set_property(TEST ${TARGET_NAME} PROPERTY RUN_SERIAL 1)
+
+    set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
+    set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
     endif()
   endif()
 endfunction(nv_test)
@@ -577,7 +583,9 @@ function(py_test TARGET_NAME)
     set(multiValueArgs SRCS DEPS ARGS ENVS)
     cmake_parse_arguments(py_test "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
     add_test(NAME ${TARGET_NAME}
-             COMMAND env FLAGS_init_allocated_mem=true PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
+             COMMAND env FLAGS_init_allocated_mem=true FLAGS_cudnn_deterministic=true
+             FLAGS_cpu_deterministic=true
+             PYTHONPATH=${PADDLE_BINARY_DIR}/python ${py_test_ENVS}
              ${PYTHON_EXECUTABLE} -u ${py_test_SRCS} ${py_test_ARGS}
              WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
   endif()

doc/fluid/howto/optimization/timeline_cn.md

@@ -1,21 +1,27 @@
 # 如何使用timeline工具做性能分析
 
-1. 在训练的主循环外加上`with profiler.profiler(...)`。运行之后，代码会在`/tmp/profile`目录下生成一个profile的记录文件。
+1. 在训练的主循环外加上`profiler.start_profiler(...)`和`profiler.stop_profiler(...)`。运行之后，代码会在`/tmp/profile`目录下生成一个profile的记录文件。
 
 	**提示：**
 	请不要在timeline记录信息时运行太多次迭代，因为timeline中的记录数量和迭代次数是成正比的。
 
 	```python
-	with profiler.profiler('All', 'total', '/tmp/profile') as prof:
-	    for pass_id in range(pass_num):
-	        for batch_id, data in enumerate(train_reader()):
-	            exe.run(fluid.default_main_program(),
-	                    feed=feeder.feed(data),
-	                    fetch_list=[])
+    for pass_id in range(pass_num):
+        for batch_id, data in enumerate(train_reader()):
+            if pass_id == 0 and batch_id == 5:
+                profiler.start_profiler("All")
+            elif pass_id == 0 and batch_id == 10:
+                profiler.stop_profiler("total", "/tmp/profile")
+            exe.run(fluid.default_main_program(),
+                    feed=feeder.feed(data),
+                    fetch_list=[])
 	            ...
 	```
 
 1. 运行`python paddle/tools/timeline.py`来处理`/tmp/profile`，这个程序默认会生成一个`/tmp/timeline`文件，你也可以用命令行参数来修改这个路径，请参考[timeline.py](https://github.com/PaddlePaddle/Paddle/blob/develop/tools/timeline.py)。
+```python
+python Paddle/tools/timeline.py --profile_path=/tmp/profile --timeline_path=timeline
+```
 
 1. 打开chrome浏览器，访问<chrome://tracing/>，用`load`按钮来加载生成的`timeline`文件。
 

doc/fluid/howto/optimization/timeline_en.md

@@ -1,22 +1,28 @@
 # how to use timeline tool to do profile
 
-1. Add `with profiler.profiler(...)` to the main training loop. After run, the code will generate a profile record file `/tmp/profile`. **Warning**: Please do not run too many batches when use profiler to record timeline information, for the profile record will grow with the batch number.
+1. Add `profiler.start_profiler(...)`和`profiler.stop_profiler(...)` to the main training loop. After run, the code will generate a profile record file `/tmp/profile`. **Warning**: Please do not run too many batches when use profiler to record timeline information, for the profile record will grow with the batch number.
 
 	```python
-	with profiler.profiler('All', 'total', '/tmp/profile') as prof:
-	    for pass_id in range(pass_num):
-	        for batch_id, data in enumerate(train_reader()):
-	            exe.run(fluid.default_main_program(),
-	                    feed=feeder.feed(data),
-	                    fetch_list=[],
-	                    use_program_cache=True)
+    for pass_id in range(pass_num):
+        for batch_id, data in enumerate(train_reader()):
+            if pass_id == 0 and batch_id == 5:
+                profiler.start_profiler("All")
+            elif pass_id == 0 and batch_id == 10:
+                profiler.stop_profiler("total", "/tmp/profile")
+            exe.run(fluid.default_main_program(),
+                    feed=feeder.feed(data),
+                    fetch_list=[])
 	            ...
 	```
 
 1. Run `python paddle/tools/timeline.py` to process `/tmp/profile`, it will generate another
 file `/tmp/timeline` by default. You can change the path by cmd parameter, please take a look at
 [timeline.py](https://github.com/PaddlePaddle/Paddle/blob/develop/tools/timeline.py) for details.
 
+```python
+python Paddle/tools/timeline.py --profile_path=/tmp/profile --timeline_path=timeline
+```
+
 1. Open chrome and visit <chrome://tracing/>, use `load` button to load the generated `timeline` file.
 
 	![chrome tracing](./tracing.jpeg)

doc/survey/op_fusion_design.md

@@ -0,0 +1,20 @@
+# Operator fusion  
+Fusing multiple operators together is an important method to optimize the program execution, particularly for GPU or other specialized accelerators. An obvious benefit is to avoid the overhead of saving the intermediate result back into global memory.   
+
+There are generally two ways to fuse operators, fusing directly connected operators and fusing non directly connected operators. The first method is mainly used by [NNVM Compiler](https://github.com/dmlc/tvm/) and [XLA](https://www.tensorflow.org/performance/xla/). The second method is mainly used by Dynet and TensorFlow Fold to do auto-batching. The principle of fusing operator is according to some rules to combine multiple operations into one, for example, `Y = X * W` and `Z = Y + B` can be fused to `Z = X * W + B`, and `Y1 = X1 * W` and `Y2 = X2 * W` can be fused to `[Y1;Y2] = [X1;X2] * W`. In order to get a short-term profit, we decided to try to manually specify these rules.   
+
+## Challenge
+The challenge of fusing operators is:
+  - how to make the rules.
+  - how to implement these rules efficiently.
+
+### How to make the rules?
+
+The problem of determining the best single location for a fusion operator is an NP-hard combinatorial problem. After analysis the operators of the DL model, we found there are two group of operators can be fused explicitly, one is the simple and adjacent operations, for example, `tmp = x + y` and `z = Relu(tmp)`, and the other is the operators that have the same function, for example, a serials of `SGD` or `Momentum`. They usually appear in the model in a large number. So we should think about how to fuse them separately first.
+
+### How to implement these rules efficiently?
+#### How to fuse the adjacent operations efficiently?
+Here we use a template function to represent the fused operations. The pros of using a template function are that it is simple and efficient, and the cons are that it is not easy to expand, and it can only be used to express some simple operations. So taking into account our current needs, the template function is more appropriate.
+
+#### How to fuse the operators that have the same function efficiently?
+We take SGD operator as an example, the training model may have hundreds of parameters and correspondingly have the same number of SGD operators. The expression(`w = w - lr*w_g`) of those operators is the same, so during of training, the executor will execute this expression hundreds time in CPU or other specialized accelerators. If we can fuse them and make the address of all `w` and all `w_g` continuous respectively, we only need execute one time. For some accelerators, the time of launching kernel is not neglected, so the time of hundreds of times of launching and executing kernel may be larger than launching and executing only once. There usually are many operators that similar to `SGD` in the DL model, such as `AllReduce` and `FC`.

paddle/fluid/API.spec

@@ -336,6 +336,7 @@ paddle.fluid.contrib.BeamSearchDecoder.decode ArgSpec(args=['self'], varargs=Non
 paddle.fluid.contrib.BeamSearchDecoder.early_stop ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.contrib.BeamSearchDecoder.read_array ArgSpec(args=['self', 'init', 'is_ids', 'is_scores'], varargs=None, keywords=None, defaults=(False, False))
 paddle.fluid.contrib.BeamSearchDecoder.update_array ArgSpec(args=['self', 'array', 'value'], varargs=None, keywords=None, defaults=None)
+paddle.fluid.contrib.memory_usage ArgSpec(args=['program', 'batch_size'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.DistributeTranspiler.__init__ ArgSpec(args=['self', 'config'], varargs=None, keywords=None, defaults=(None,))
 paddle.fluid.transpiler.DistributeTranspiler.create_splited_vars ArgSpec(args=['self', 'source_var', 'block', 'tag'], varargs=None, keywords=None, defaults=None)
 paddle.fluid.transpiler.DistributeTranspiler.get_pserver_program ArgSpec(args=['self', 'endpoint'], varargs=None, keywords=None, defaults=None)

paddle/fluid/CMakeLists.txt

@@ -5,5 +5,7 @@ add_subdirectory(operators)
 add_subdirectory(pybind)
 add_subdirectory(string)
 add_subdirectory(recordio)
-# NOTE: please add subdirectory inference at last.
-add_subdirectory(inference)
+if(WITH_INFERENCE)
+  # NOTE: please add subdirectory inference at last.
+  add_subdirectory(inference)
+endif()

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -17,6 +17,7 @@
 #include "paddle/fluid/framework/details/container_cast.h"
 #include "paddle/fluid/framework/details/reduce_and_gather.h"
 #include "paddle/fluid/framework/details/variable_visitor.h"
+#include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
 namespace framework {
@@ -45,6 +46,7 @@ AllReduceOpHandle::AllReduceOpHandle(ir::Node *node,
 #endif
 
 void AllReduceOpHandle::RunImpl() {
+  platform::RecordEvent r("all_reduce", nullptr);
   if (NoDummyInputSize() == 1) {
     return;  // No need to all reduce when GPU count = 1;
   } else {