merge develop

Author: NHZlX

.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

@@ -200,6 +200,14 @@ include(external/snappy)    # download snappy
 include(external/snappystream)
 include(external/threadpool)
 
+if(WITH_GPU)
+    include(cuda)
+    include(tensorrt)
+    include(external/anakin)
+else()
+  set(WITH_ANAKIN OFF CACHE STRING "Anakin is valid only when GPU is set." FORCE)
+endif()
+
 include(cudnn)              # set cudnn libraries, must before configure
 include(cupti)
 include(configure)          # add paddle env configuration
@@ -228,14 +236,6 @@ set(EXTERNAL_LIBS
     ${PYTHON_LIBRARIES}
 )
 
-if(WITH_GPU)
-    include(cuda)
-    include(tensorrt)
-    include(external/anakin)
-else()
-  set(WITH_ANAKIN OFF CACHE STRING "Anakin is valid only when GPU is set." FORCE)
-endif()
-
 if(WITH_AMD_GPU)
     find_package(HIP)
     include(hip)

cmake/cudnn.cmake

@@ -21,6 +21,7 @@ list(APPEND CUDNN_CHECK_LIBRARY_DIRS
     ${CUDNN_ROOT}/lib64
     ${CUDNN_ROOT}/lib
     ${CUDNN_ROOT}/lib/${TARGET_ARCH}-linux-gnu
+    ${CUDNN_ROOT}/local/cuda-${CUDA_VERSION}/targets/${TARGET_ARCH}-linux/lib/
     $ENV{CUDNN_ROOT}
     $ENV{CUDNN_ROOT}/lib64
     $ENV{CUDNN_ROOT}/lib

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)

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 {

paddle/fluid/framework/details/build_strategy.h

@@ -21,6 +21,26 @@ namespace framework {
 namespace details {
 
 struct BuildStrategy {
+  // ParallelExecutor supports two modes of ReduceStrategy, kAllReduce and
+  // kReduce, for CPU and GPU. If you use kAllReduce, different threads
+  // optimize their parameters separately. If you use kReduce, the optimizations
+  // of parameters are distributed to different threads.
+  // For example, a model has 100 parameters and is running with four threads,
+  // if you choose kAllReduce, every thread is to optimize 100 parameters
+  // separately, if you choose kReduce, every thread is to optimize 25
+  // parameters.
+  // Of particular note is, if you use kReduce when using CPU training,
+  // all the parameters are shared between different threads. This feature will
+  // save memory.
+  // FIXME(zcd): The result of the two modes(kAllReduce and kReduce) maybe not
+  // equal for GPU. Because, the result of the different order of summing maybe
+  // different, for example, the result of `a+b+c+d` may be different with the
+  // result of `c+a+b+d`.
+  // For GPU, the implementation of kAllReduce and kReduce is adopted NCCL,
+  // so the result of kAllReduce and kReduce maybe not equal.
+  // For CPU, if you want to fix the order of summing to make the result
+  // of kAllReduce and kReduce no diff, you can add
+  // `FLAGS_cpu_deterministic=true` to env.
   enum class ReduceStrategy { kAllReduce = 0, kReduce = 1 };
 
   enum class GradientScaleStrategy {

paddle/fluid/framework/details/reduce_op_handle.cc

@@ -16,12 +16,18 @@
 #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"
+
+DEFINE_bool(
+    cpu_deterministic, false,
+    "Whether to make the result of computation deterministic in CPU side.");
 
 namespace paddle {
 namespace framework {
 namespace details {
 
 void ReduceOpHandle::RunImpl() {
+  platform::RecordEvent r("reduce", nullptr);
   if (places_.size() == 1) return;
   // the input and output may have dummy var.
   auto in_var_handles = DynamicCast<VarHandle>(inputs_);
@@ -89,11 +95,33 @@ void ReduceOpHandle::RunImpl() {
   } else {
     std::vector<const LoDTensor *> lod_tensors =
         GetInputValues<LoDTensor>(in_var_handles, var_scopes);
+
     if (paddle::platform::is_cpu_place(lod_tensors[0]->place())) {
       this->RunAndRecordEvent([&] {
-        ReduceLoDTensor func(lod_tensors,
-                             out_var->GetMutable<framework::LoDTensor>());
-        VisitDataType(ToDataType(lod_tensors[0]->type()), func);
+        // FIXME(zcd): The order of summing is important,
+        // especially when the type of data is float or double.
+        // For example, the result of `a+b+c+d` may be different
+        // with the result of `c+a+b+d`, so the summing order should be fixed.
+        if (!FLAGS_cpu_deterministic) {
+          ReduceLoDTensor func(lod_tensors,
+                               out_var->GetMutable<framework::LoDTensor>());
+          VisitDataType(ToDataType(lod_tensors[0]->type()), func);
+        } else {
+          // We sum lod_tensors to reduce_sum_trg which is in local_scopes_0
+          // here, but it doesn't mean reduce_sum_trg must be in local_scopes_0.
+          auto &reduce_sum_trg = *this->local_scopes_[0]
+                                      ->FindVar(kLocalExecScopeName)
+                                      ->Get<Scope *>()
+                                      ->FindVar(out_var_handle->name_)
+                                      ->GetMutable<framework::LoDTensor>();
+          ReduceLoDTensor func(lod_tensors, &reduce_sum_trg);
+          VisitDataType(ToDataType(lod_tensors[0]->type()), func);
+
+          auto trg = out_var->GetMutable<framework::LoDTensor>();
+          if (reduce_sum_trg.data<void>() != trg->data<void>()) {
+            TensorCopy(reduce_sum_trg, platform::CPUPlace(), trg);
+          }
+        }
       });
     } else if (paddle::platform::is_gpu_place(lod_tensors[0]->place())) {
 #ifdef PADDLE_WITH_CUDA

paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.cc

@@ -17,6 +17,7 @@
 #include <string>
 #include <vector>
 #include "paddle/fluid/framework/executor.h"
+#include "paddle/fluid/platform/profiler.h"
 
 namespace paddle {
 namespace framework {
@@ -62,6 +63,7 @@ FeedFetchList ScopeBufferedSSAGraphExecutor::Run(
     eptr = std::current_exception();
   }
 
+  platform::RecordEvent e("ScopeBufferedSSAGraphExecutorAfterRun", nullptr);
   drop_scope_counter_ += 1;
   if (!fetch_tensors.empty() ||
       drop_scope_counter_ == strategy_.num_iteration_per_drop_scope_) {