Merge pull request #138 from gangliao/master

Add Mac OS X port

.gitignore

 *.DS_Store
 build/
+*.user
+
+.vscode
+.idea
\ No newline at end of file

cmake/cblas.cmake

@@ -65,12 +65,14 @@ set(OPENBLAS_ROOT $ENV{OPENBLAS_ROOT} CACHE PATH "Folder contains Openblas")
 set(OPENBLAS_INCLUDE_SEARCH_PATHS
         ${OPENBLAS_ROOT}/include
         /usr/include
-        /usr/include/openblas)
+        /usr/include/openblas
+        /usr/local/opt/openblas/include)
 set(OPENBLAS_LIB_SEARCH_PATHS
         ${OPENBLAS_ROOT}/lib
         /usr/lib
         /usr/lib/blas/openblas
-        /usr/lib/openblas)
+        /usr/lib/openblas
+        /usr/local/opt/openblas/lib)
 
 find_path(OPENBLAS_INC_DIR NAMES cblas.h
   PATHS ${OPENBLAS_INCLUDE_SEARCH_PATHS})

cmake/cudnn.cmake

@@ -15,7 +15,7 @@ list(APPEND CUDNN_CHECK_LIBRARY_DIRS
     $ENV{CUDNN_ROOT}/lib64
     $ENV{CUDNN_ROOT}/lib
     /usr/lib)
-find_library(CUDNN_LIBRARY NAMES libcudnn.so # libcudnn_static.a
+find_library(CUDNN_LIBRARY NAMES libcudnn.so libcudnn.dylib # libcudnn_static.a
     PATHS ${CUDNN_CHECK_LIBRARY_DIRS} ${CUDNN_INCLUDE_DIR} ${__libpath_hist}
           NO_DEFAULT_PATH
     DOC "Path to cuDNN library.")

cmake/util.cmake

 # Some common routine for paddle compile.
 
-
 # target_circle_link_libraries
 # Link libraries to target which has circle dependencies.
 #
 # First Argument: target name want to be linked with libraries
 # Rest Arguments: libraries which link together.
 function(target_circle_link_libraries TARGET_NAME)
-    target_link_libraries(${TARGET_NAME}
-        -Wl,--start-group
-        ${ARGN}
-        -Wl,--end-group)
+    if(APPLE)
+        set(LIBS)
+        set(inArchive OFF)
+        set(libsInArgn)
+
+        foreach(arg ${ARGN})
+            if(${arg} STREQUAL "ARCHIVE_START")
+                set(inArchive ON)
+            elseif(${arg} STREQUAL "ARCHIVE_END")
+                set(inArchive OFF)
+            else()
+                if(inArchive)
+                    list(APPEND LIBS "-Wl,-force_load")
+                endif()
+                list(APPEND LIBS ${arg})
+                list(APPEND libsInArgn ${arg})
+            endif()
+        endforeach()
+        if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
+            list(APPEND LIBS "-undefined dynamic_lookup")
+        endif()
+        list(REVERSE libsInArgn)
+        target_link_libraries(${TARGET_NAME}
+            ${LIBS}
+            ${libsInArgn})
+
+    else()  # LINUX
+        set(LIBS)
+
+        foreach(arg ${ARGN})
+            if(${arg} STREQUAL "ARCHIVE_START")
+                list(APPEND LIBS "-Wl,--whole-archive")
+            elseif(${arg} STREQUAL "ARCHIVE_END")
+                list(APPEND LIBS "-Wl,--no-whole-archive")
+            else()
+                list(APPEND LIBS ${arg})
+            endif()
+        endforeach()
+
+        target_link_libraries(${TARGET_NAME}
+                "-Wl,--start-group"
+                ${LIBS}
+                "-Wl,--end-group")
+    endif()
 endfunction()
 
 # compile_cu_as_cpp
@@ -41,20 +80,20 @@ function(link_paddle_exe TARGET_NAME)
     if(PADDLE_WITH_INTERNAL)
         set(INTERAL_LIBS paddle_internal_gserver paddle_internal_parameter)
         target_circle_link_libraries(${TARGET_NAME}
-            -Wl,--whole-archive
+            ARCHIVE_START
             paddle_internal_gserver
             paddle_internal_owlqn
-            -Wl,--no-whole-archive
+            ARCHIVE_END
             paddle_internal_parameter)
     else()
         set(INTERAL_LIBS "")
     endif()
 
     target_circle_link_libraries(${TARGET_NAME}
-        -Wl,--whole-archive
+        ARCHIVE_START
         paddle_gserver
         ${METRIC_LIBS}
-        -Wl,--no-whole-archive
+        ARCHIVE_END
         paddle_pserver
         paddle_trainer_lib
         paddle_network

doc/build/build_from_source.md

-Build and Install
+Installing from Sources
 =================
 
-## Requirement
+* [1. Download and Setup](#download)
+* [2. Requirements](#requirements)
+* [3. Build on Ubuntu](#ubuntu)
+* [4. Build on Mac OS X](#mac)
 
-### Dependents
+## <span id="download">Download and Setup</span> 
+You can download PaddlePaddle from the [github source](https://github.com/gangliao/Paddle).
 
-- **CMake**: required for 2.8+ version
-- **g++**: a recent c++ compiler supporting c++11, >= 4.6, < 5
-- **BLAS library**: such as openBLAS, MKL, ATLAS
-- **protobuf**: required for 2.4+ version, 3.x is not supported
-- **python**: currently only 2.7 version is supported
+```bash
+git clone https://github.com/baidu/Paddle paddle
+```
+
+## <span id="requirements">Requirements</span>
+
+To compile the source code, your computer must be equipped with GCC >=4.6 or Clang Compiler.
+### Dependencies
+
+- **CMake**: version >= 2.8
+- **BLAS**: MKL, OpenBlas or ATLAS
+- **protobuf**: version >= 2.4, **Note: 3.x is not supported**
+- **python**: only python 2.7 is supported currently
+
+### Options
+
+PaddlePaddle supports some build options. To enable it, first you need to install the related libraries. 
+
+    Optional            | Description
+    ------------        | :-----------
+    **WITH_GPU**        | Compile with GPU mode.
+    **WITH_DOUBLE**     | Compile with double precision floating-point, default: single precision. |
+    **WITH_GLOG**       | Compile with glog. If not found, default: an internal log implementation.
+    **WITH_GFLAGS**     | Compile with gflags. If not found, default: an internal flag implementation.
+    **WITH_TESTING**    | Compile with gtest for PaddlePaddle's unit testing. 
+    **WITH_DOC**        | Compile to generate PaddlePaddle's docs, default: disabled (OFF).
+    **WITH_SWIG_PY**    | Compile with python predict API, default: disabled (OFF).
+    **WITH_STYLE_CHECK**| Compile with code style check, default: enabled (ON).
+|
 
-### Optional
+**Note:**
+  - The GPU version works best with Cuda Toolkit 7.5 and cuDNN v5.
+  - Other versions like Cuda Toolkit 6.5, 7.0, 8.0 and cuDNN v2, v3, v4 are also supported.
+  - **To utilize cuDNN v5, Cuda Toolkit 7.5 is prerequisite and vice versa.**
 
-PaddlePaddle also support some build options, you have to install related libraries. 
+As a simple example, consider the following:  
 
-- **WITH_GPU**: Compile with gpu mode
-  - The GPU version works best with Cuda Toolkit 7.5 and cuDNN v5
-  - Other versions Cuda Toolkit 6.5, 7.0 and cuDNN v2, v3, v4 are also supported
-  - Note: to utilize cuDNN v5, Cuda Toolkit 7.5 is prerequisite and vice versa
-- **WITH_DOUBLE**: Compile with double precision, otherwise use single precision 
-- **WITH_GLOG**: Compile with glog, otherwise use a log implement internally
-- **WITH_GFLAGS**: Compile with gflags, otherwise use a flag implement internally
-- **WITH_TESTING**: Compile with gtest and run unittest for PaddlePaddle 
-- **WITH_DOC**: Compile with documentation
-- **WITH_SWIG_PY**: Compile with python predict api
-- **WITH_STYLE_CHECK**: Style check for source code
+1. **Python Dependencies(optional)**
+  
+    To compile PaddlePaddle with python predict API, make sure swig installed and set `-DWITH_SWIG_PY=ON` as follows:
+
+    ```bash
+    # install swig on ubuntu
+    sudo apt-get install swig
+    # install swig on Mac OS X
+    brew install swig
+
+    # active swig in cmake
+    cmake .. -DWITH_SWIG_PY=ON
+    ```
+
+2. **Doc Dependencies(optional)**
+
+    To generate PaddlePaddle's documentation, install dependencies and set `-DWITH_DOC=ON` as follows:
 
+    ```bash
+    pip install 'sphinx>=1.4.0'
+    pip install sphinx_rtd_theme breathe recommonmark
 
-## Building on Ubuntu14.04
+    # install doxygen on Ubuntu
+    sudo apt-get install doxygen 
+    # install doxygen on Mac OS X
+    brew install doxygen
+
+    # active docs in cmake
+    cmake .. -DWITH_DOC=ON`
+    ```
+
+## <span id="ubuntu">Build on Ubuntu 14.04</span>
 
 ### Install Dependencies
 
 - **CPU Dependencies**
 
-```bash
-# necessary
-sudo apt-get update
-sudo apt-get install -y g++ make cmake build-essential libatlas-base-dev python python-pip libpython-dev m4 libprotobuf-dev protobuf-compiler python-protobuf python-numpy git
-# optional
-sudo apt-get install libgoogle-glog-dev
-sudo apt-get install libgflags-dev
-sudo apt-get install libgtest-dev
-sudo pip install wheel
-pushd /usr/src/gtest
-cmake .
-make
-sudo cp *.a /usr/lib
-popd
-```
-    
+    ```bash
+    # necessary
+    sudo apt-get update
+    sudo apt-get install -y g++ make cmake build-essential libatlas-base-dev python python-pip libpython-dev m4 libprotobuf-dev protobuf-compiler python-protobuf python-numpy git
+    # optional
+    sudo apt-get install libgoogle-glog-dev
+    sudo apt-get install libgflags-dev
+    sudo apt-get install libgtest-dev
+    sudo pip install wheel
+    pushd /usr/src/gtest
+    cmake .
+    make
+    sudo cp *.a /usr/lib
+    popd
+    ```
   
-- **GPU Dependencies(optional)**
+- **GPU Dependencies (optional)**
 
-If you need to build GPU version, the first thing you need is a machine that has GPU and CUDA installed.
-And you also need to install cuDNN.
+    To build GPU version, you will need the following installed:
 
-You can download CUDA toolkit and cuDNN from nvidia website:
-    
-```bash
-https://developer.nvidia.com/cuda-downloads
-https://developer.nvidia.com/cudnn
-```
-You can copy cuDNN files into the CUDA toolkit directory, such as:
+        1. a CUDA-capable GPU
+        2. A supported version of Linux with a gcc compiler and toolchain
+        3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
+        4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn)
+
+    The CUDA development environment relies on tight integration with the host development environment,
+    including the host compiler and C runtime libraries, and is therefore only supported on
+    distribution versions that have been qualified for this CUDA Toolkit release.
+        
+    After downloading cuDNN library, issue the following commands:
+
+    ```bash
+    sudo tar -xzf cudnn-7.5-linux-x64-v5.1.tgz -C /usr/local
+    sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*
+    ```
+    Then you need to set LD\_LIBRARY\_PATH, CUDA\_HOME and PATH environment variables in ~/.bashrc.
+
+    ```bash
+    export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
+    export CUDA_HOME=/usr/local/cuda
+    export PATH=/usr/local/cuda/bin:$PATH
+    ```
+
+### Build and Install
+
+As usual, the best option is to create build folder under paddle project directory.
 
 ```bash
-sudo tar -xzf cudnn-7.5-linux-x64-v5.1.tgz -C /usr/local
-sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*
+mkdir build && cd build
+cmake ..
 ```
-Then you need to set LD\_LIBRARY\_PATH, CUDA\_HOME and PATH environment variables in ~/.bashrc.
+
+CMake first check PaddlePaddle's dependencies in system default path. After installing some optional
+libraries, corresponding build option will be set automatically (for instance, glog, gtest and gflags).
+If still not found, you can manually set it based on CMake error information from your screen.
+
+As a simple example, consider the following:
+
+- **Only CPU**
+
+  ```bash
+  cmake  .. -DWITH_GPU=OFF -DWITH_DOC=OFF
+  ```
+- **GPU**
+
+  ```bash
+  cmake .. -DWITH_GPU=ON -DWITH_DOC=OFF
+  ```
+
+- **GPU with doc and swig**
+
+  ```bash
+  cmake .. -DWITH_GPU=ON -DWITH_DOC=ON -DWITH_SWIG_PY=ON
+  ``` 
+
+Finally, you can download source code and build:
 
 ```bash
-export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
-export CUDA_HOME=/usr/local/cuda
-export PATH=/usr/local/cuda/bin:$PATH
+# you can add build option here, such as:    
+cmake .. -DWITH_GPU=ON -DWITH_DOC=OFF -DCMAKE_INSTALL_PREFIX=<path to install>
+# please use sudo make install, if you want
+# to install PaddlePaddle into the system
+make -j `nproc` && make install
+# set PaddlePaddle installation path in ~/.bashrc
+export PATH=<path to install>/bin:$PATH
 ```
-- **Python Dependencies(optional)**
 
-If you want to compile PaddlePaddle with python predict api, you need to add -DWITH_SWIG_PY=ON in cmake command and install these first:
+**Note:**
+
+If you set `WITH_SWIG_PY=ON`, related python dependencies also need to be installed.
+Otherwise, PaddlePaddle will automatically install python dependencies
+at first time when user run paddle commands, such as `paddle version`, `paddle train`.
+It may require sudo privileges:
 
 ```bash
-sudo apt-get install swig
+# you can run
+sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
+# or just run 
+sudo paddle version
 ```
 
-- **Doc Dependencies(optional)**
+## <span id="mac">Building on Mac OS X</span>
 
-If you want to compile PaddlePaddle with doc, you need to add -DWITH_DOC=ON in cmake command and install these first:
+### Prerequisites
+This guide is based on Mac OS X 10.11 (El Capitan). Note that if you are running an up to date version of OS X, 
+you will already have Python 2.7.10 and Numpy 1.8 installed.
+
+The best option is to use the package manager homebrew to handle installations and upgrades for you.
+To install [homebrew](http://brew.sh/), first open a terminal window (you can find Terminal in the Utilities folder in Applications), and issue the command:
 
 ```bash
-pip install 'sphinx>=1.4.0'
-pip install sphinx_rtd_theme breathe recommonmark
-sudo apt-get install doxygen 
+# install brew
+/usr/bin/ruby -e "$(curl -fsSL https://raw.githubusercontent.com/Homebrew/install/master/install)"
+# install pip
+easy_install pip
 ```
 
-### Build and Install
+### Install Dependencies
 
-CMake will find dependent libraries in system default paths first. After installing some optional libraries, corresponding build option will automatically be on(such as glog, gtest and gflags). And if libraries are not found, you have to set following variables manually in cmake command(CUDNN_ROOT, ATLAS_ROOT, MKL_ROOT, OPENBLAS_ROOT).
+- **CPU Dependencies**
 
-Here are some examples of cmake command with different options:
+  ```bash
+  # Install fundamental dependents 
+  brew install glog gflags cmake protobuf openblas
+
+  # Install google test on Mac OS X
+  # Download gtest 1.7.0
+  wget https://github.com/google/googletest/archive/release-1.7.0.tar.gz
+  tar -xvf googletest-release-1.7.0.tar.gz && cd googletest-release-1.7.0
+  # Build gtest
+  mkdir build && cmake ..
+  make
+  # Install gtest library
+  sudo cp -r ../include/gtest /usr/local/include/
+  sudo cp lib*.a /usr/local/lib
+  ```
 
-**only cpu**
+- **GPU Dependencies(optional)**
 
-```bash
-cmake -DWITH_GPU=OFF -DWITH_DOC=OFF
-```
+    To build GPU version, you will need the following installed:
+
+        1. a CUDA-capable GPU
+        2. Mac OS X 10.11 or later
+        2. the Clang compiler and toolchain installed using Xcode
+        3. NVIDIA CUDA Toolkit (available at http://developer.nvidia.com/cuda-downloads)
+        4. NVIDIA cuDNN Library (availabel at https://developer.nvidia.com/cudnn)
+
+    The CUDA development environment relies on tight integration with the host development environment,
+    including the host compiler and C runtime libraries, and is therefore only supported on
+    distribution versions that have been qualified for this CUDA Toolkit release.
+        
+    1. After downloading cuDNN library, issue the following commands:
+
+        ```bash
+        sudo tar -xzf cudnn-7.5-osx-x64-v5.0-ga.tgz -C /usr/local
+        sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*
+        ```
+    2. Then you need to set DYLD\_LIBRARY\_PATH, CUDA\_HOME and PATH environment variables in ~/.bashrc.
 
-**gpu**
+        ```bash
+        export DYLD_LIBRARY_PATH=/usr/local/cuda/lib:$DYLD_LIBRARY_PATH
+        export PATH=/usr/local/cuda/bin:$PATH
+        ```
+
+### Build and Install
+
+As usual, the best option is to create build folder under paddle project directory.
 
 ```bash
-cmake -DWITH_GPU=ON -DWITH_DOC=OFF
+mkdir build && cd build
+cmake ..
 ```
 
-**gpu with doc and swig**
+CMake first check PaddlePaddle's dependencies in system default path. After installing some optional
+libraries, corresponding build option will be set automatically (for instance, glog, gtest and gflags).
+If still not found, you can manually set it based on CMake error information from your screen.
 
-```bash
-cmake -DWITH_GPU=ON -DWITH_DOC=ON -DWITH_SWIG_PY=ON
-``` 
+As a simple example, consider the following:
 
-Finally, you can download source code and build:
+- **Only CPU**
+
+  ```bash
+  cmake  .. -DWITH_GPU=OFF -DWITH_DOC=OFF
+  ```
+- **GPU**
+
+  ```bash
+  cmake .. -DWITH_GPU=ON -DWITH_DOC=OFF
+  ```
+
+- **GPU with doc and swig**
+
+  ```bash
+  cmake .. -DWITH_GPU=ON -DWITH_DOC=ON -DWITH_SWIG_PY=ON
+  ``` 
+
+Finally, you can build PaddlePaddle:
 
 ```bash
-git clone https://github.com/baidu/Paddle paddle
-cd paddle
-mkdir build
-cd build
 # you can add build option here, such as:    
-cmake -DWITH_GPU=ON -DWITH_DOC=OFF -DCMAKE_INSTALL_PREFIX=<path to install> ..
-# please use sudo make install, if you want
-# to install PaddlePaddle into the system
+cmake .. -DWITH_GPU=ON -DWITH_DOC=OFF -DCMAKE_INSTALL_PREFIX=<installation path>
+# please use sudo make install, if you want to install PaddlePaddle into the system
 make -j `nproc` && make install
-# PaddlePaddle installation path
-export PATH=<path to install>/bin:$PATH
+# set PaddlePaddle installation path in ~/.bashrc
+export PATH=<installation path>/bin:$PATH
 ```
-**Note**
 
-And if you set WITH_SWIG_PY=ON, you have to install related python predict api at the same time:
+
+**Note:**
+
+If you set `WITH_SWIG_PY=ON`, related python dependencies also need to be installed.
+Otherwise, PaddlePaddle will automatically install python dependencies
+at first time when user run paddle commands, such as `paddle version`, `paddle train`.
+It may require sudo privileges:
 
 ```bash
-pip install <path to install>/opt/paddle/share/wheels/*.whl
-```
+# you can run
+sudo pip install <path to install>/opt/paddle/share/wheels/*.whl
+# or just run 
+sudo paddle version
+```
\ No newline at end of file

paddle/api/PaddleAPI.h

@@ -20,6 +20,7 @@ limitations under the License. */
 #include <string>
 #include <vector>
 #include "paddle/utils/GlobalConstants.h"
+#include "paddle/utils/TypeDefs.h"
 
 /// Import PaddlePaddle's enumeration into global namespace.
 using namespace paddle::enumeration_wrapper;  // NOLINT

paddle/api/Util.cpp

@@ -17,6 +17,7 @@ limitations under the License. */
 #include "paddle/utils/Util.h"
 #include "paddle/utils/PythonUtil.h"
 #include "paddle/utils/Flags.h"
+#include "paddle/utils/Excepts.h"
 #include "paddle/parameter/Parameter.h"
 
 #include <fenv.h>

paddle/api/paddle_ld_flags.py

@@ -15,6 +15,19 @@
 try:
     from paddle_api_config import *
     import os.path
+    import platform
+
+    system = platform.system().lower()
+    is_osx = (system == 'darwin')
+    is_win = (system == 'windows')
+    is_lin = (system == 'linux')
+
+    if is_lin:
+        whole_start = "-Wl,--whole-archive"
+        whole_end = "-Wl,--no-whole-archive"
+    elif is_osx:
+        whole_start = ""
+        whole_end = ""
 
     LIB_DIRS = ["math", 'utils', 'parameter', "gserver", "api", "cuda", "pserver", "trainer"]
     PARENT_LIB_DIRS = ['proto']
@@ -56,9 +69,9 @@ try:
 
         def libs_str(self):
             libs = [
-                "-Wl,--whole-archive",
+                whole_start,
                 "-lpaddle_gserver",
-                "-Wl,--no-whole-archive",
+                whole_end,
                 "-lpaddle_pserver",
                 "-lpaddle_trainer_lib",
                 "-lpaddle_network",

paddle/cuda/include/hl_device_functions.cuh

@@ -16,28 +16,37 @@ limitations under the License. */
 #ifndef HL_DEVICE_FUNCTIONS_CUH_
 #define HL_DEVICE_FUNCTIONS_CUH_
 
-namespace hppl {
-
-static __inline__ __device__ double atomicAdd(double* address, double val) {
-    // NOLINTNEXTLINE
-    unsigned long long int* address_as_ull = (unsigned long long int*)address;
-    unsigned long long int old = *address_as_ull, assumed; // NOLINT
-
-    do {
-        assumed = old;
-        old = atomicCAS(address_as_ull,
-                        assumed,
-                        __double_as_longlong(val +
-                        __longlong_as_double(assumed)));
-    } while (assumed != old);
-
-    return __longlong_as_double(old);
-}
+namespace paddle {
+
+template <class T>
+inline __device__ T paddleAtomicAdd(T* address, T val);
 
-}  // namespace hppl
+template <>
+inline __device__ float paddleAtomicAdd(float* address, float val) {
+  return atomicAdd(address, val);
+}
 
-#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ < 600
-using hppl::atomicAdd;
+template <>
+inline __device__ double paddleAtomicAdd(double* address, double val) {
+#if defined(__CUDA_ARCH__) && __CUDA_ARCH__ >= 600
+  return atomicAdd(address, val);
+#else
+  // NOLINTNEXTLINE
+  unsigned long long int* address_as_ull = (unsigned long long int*)address;
+  unsigned long long int old = *address_as_ull, assumed; // NOLINT
+
+  do {
+    assumed = old;
+    old = atomicCAS(address_as_ull,
+                    assumed,
+                    __double_as_longlong(val +
+                    __longlong_as_double(assumed)));
+  } while (assumed != old);
+
+  return __longlong_as_double(old);
 #endif
+}
+}  // namespace paddle
+
 
 #endif /* HL_DEVICE_FUNCTIONS_CUH_ */

paddle/cuda/include/hl_gpu_lstm.cuh

@@ -192,10 +192,10 @@ __global__ void KeLstmBackward(Op op,
 
   if (isBatch) {
     if (value.prevStateValue) {
-      if (grad.checkIgGrad) atomicAdd(grad.checkIgGrad+frameIdx, rCheckIGrad);
-      if (grad.checkFgGrad) atomicAdd(grad.checkFgGrad+frameIdx, rCheckFGrad);
+      if (grad.checkIgGrad) paddle::paddleAtomicAdd(grad.checkIgGrad+frameIdx, rCheckIGrad);
+      if (grad.checkFgGrad) paddle::paddleAtomicAdd(grad.checkFgGrad+frameIdx, rCheckFGrad);
     }
-    if (grad.checkOgGrad) atomicAdd(grad.checkOgGrad+frameIdx, rCheckOGrad);
+    if (grad.checkOgGrad) paddle::paddleAtomicAdd(grad.checkOgGrad+frameIdx, rCheckOGrad);
   } else {
     if (value.prevStateValue) {
       if (grad.checkIgGrad) grad.checkIgGrad[frameIdx] += rCheckIGrad;

paddle/cuda/include/hl_matrix_type.cuh

@@ -27,6 +27,8 @@ typedef float4 vecType;
 typedef double2 vecType;
 #endif
 #else
+#include <mmintrin.h>
+#include <xmmintrin.h>
 #include <emmintrin.h>
 #ifndef HPPL_TYPE_DOUBLE
 typedef __m128  vecType;

paddle/cuda/include/hl_sse_matrix_kernel.cuh

@@ -25,6 +25,9 @@ limitations under the License. */
 #define     VECTOR_LEN      4
 #define     VECTOR_SET      _mm_set_ps1
 #else
+#if   defined(__APPLE__) || defined(__OSX__)
+#define     _mm_set_pd1     _mm_set1_pd
+#endif
 /* number of double in vector */
 #define     VECTOR_LEN      2
 #define     VECTOR_SET      _mm_set_pd1

paddle/cuda/src/hl_cuda_device.cc

@@ -209,7 +209,18 @@ __thread cudaStream_t default_stream = 0;
 __thread bool g_sync_flag = true;
 bool hl_start_flag = false;
 
-#define gettid() syscall(SYS_gettid)
+inline pid_t gettid() {
+#if defined(__APPLE__) || defined(__OSX__)
+  pid_t tid = syscall(SYS_thread_selfid);
+#else
+  #ifndef __NR_gettid
+  #define __NR_gettid 224
+  #endif
+  pid_t tid = syscall(__NR_gettid);
+#endif
+  CHECK_NE(tid, -1);
+  return tid;    
+}
 
 void hl_init(int device) {
   CHECK(hl_start_flag)

paddle/cuda/src/hl_cuda_lstm.cu

@@ -564,11 +564,11 @@ __global__ void KeLstmBackward(real *gateValue,
 
   /* TODO: Temporary save & merger in another kernel */
   if (frameIdy == 1) {
-    if (checkIgGrad) atomicAdd(checkIgGrad+frameIdx, rCheckGrad);
+    if (checkIgGrad) paddle::paddleAtomicAdd(checkIgGrad+frameIdx, rCheckGrad);
   } else if (frameIdy == 2) {
-    if (checkFgGrad) atomicAdd(checkFgGrad+frameIdx, rCheckGrad);
+    if (checkFgGrad) paddle::paddleAtomicAdd(checkFgGrad+frameIdx, rCheckGrad);
   } else if (frameIdy == 3) {
-    if (checkOgGrad) atomicAdd(checkOgGrad+frameIdx, rCheckGrad);
+    if (checkOgGrad) paddle::paddleAtomicAdd(checkOgGrad+frameIdx, rCheckGrad);
   }
 }
 

paddle/cuda/src/hl_cuda_matrix.cu

@@ -623,7 +623,7 @@ __global__ void KeCosSimDerivative(real* grad,
         prevGradY[index] +=
           scale * grad[ty] * prevOutX[index] * reciprocal;
       } else {
-        atomicAdd(prevGradY + index,
+        paddle::paddleAtomicAdd(prevGradY + index,
           scale * grad[ty] * prevOutX[index] * reciprocal);
       }
     }
@@ -640,7 +640,7 @@ __global__ void KeCosSimDerivative(real* grad,
           (prevOutX[index] * reciprocalXY -
            prevOutY[index] * reciprocalSquareSumY);
       } else {
-        atomicAdd(prevGradY + index, output[ty] * grad[ty] *
+        paddle::paddleAtomicAdd(prevGradY + index, output[ty] * grad[ty] *
           (prevOutX[index] * reciprocalXY -
            prevOutY[index] * reciprocalSquareSumY));
       }

paddle/cuda/src/hl_cuda_sequence.cu

@@ -362,7 +362,7 @@ __global__ void KeMatrixAddRows(real* output,
         if (AddRow == 0) {
           outputData[i] += tableData[i];
         } else {
-          atomicAdd(&tableData[i], outputData[i]);
+          paddle::paddleAtomicAdd(&tableData[i], outputData[i]);
         }
       }
     }

paddle/cuda/src/hl_cuda_sparse.cuh

@@ -280,7 +280,7 @@ __global__ void KeSMatrixCscMulDense(real *C_d,
         if (index_n_t < dimN) {
           real tmp;
           tmp = alpha*a_r*b_r[n];
-          atomicAdd(C_d_r, tmp);
+          paddle::paddleAtomicAdd(C_d_r, tmp);
           C_d_r += CU_CSC_MUL_DENSE_THREAD_X;
           index_n_t += CU_CSC_MUL_DENSE_THREAD_X;
         }
@@ -328,7 +328,7 @@ __global__ void KeSMatrixCscMulDense(real *C_d,
         if (index_n_t < dimN) {
           real tmp;
           tmp = alpha*a_r*b_r[n];
-          atomicAdd(C_d_r, tmp);
+          paddle::paddleAtomicAdd(C_d_r, tmp);
           C_d_r += CU_CSC_MUL_DENSE_THREAD_X;
           index_n_t += CU_CSC_MUL_DENSE_THREAD_X;
         }
@@ -629,7 +629,7 @@ __global__ void KeSMatrixDenseMulCsr(real *C_d,
       for (int n=0; n < CU_DM_CSR_N; n++) {
         if (index_m_t++ < dimM) {
           tmp = alpha * b_r * a_r[n];
-          atomicAdd(C_d_r, tmp);
+          paddle::paddleAtomicAdd(C_d_r, tmp);
           C_d_r += dimN;
         }
       }
@@ -660,7 +660,7 @@ __global__ void KeSMatrixDenseMulCsr(real *C_d,
       for (int n=0; n < CU_DM_CSR_N; n++) {
         if (index_m_t++ < dimM) {
           tmp = alpha * b_r * a_r[n];
-          atomicAdd(C_d_r, tmp);
+          paddle::paddleAtomicAdd(C_d_r, tmp);
           C_d_r += dimN;
         }
       }
@@ -912,7 +912,7 @@ __global__ void KeSMatrixCsrColumnSum(real* a_val, real* csr_val,
   for (int idx = gid; idx < dimNNZ; idx += gridDim.x * blockDim.x) {
     int colIdx = csr_col[idx];
     real val = csr_val[idx];
-    atomicAdd(a_val + colIdx, val);
+    paddle::paddleAtomicAdd(a_val + colIdx, val);
   }
 }
 

paddle/cuda/src/hl_dso_loader.cc

@@ -69,23 +69,40 @@ static inline void GetDsoHandleWithSearchPath(
 
     CHECK(nullptr != *dso_handle)
       << "For Gpu version of PaddlePaddle, it couldn't find CUDA library: "
-      << dlPath.c_str() << " Please make sure you already specify its path."
-      << "Note: for training data on Cpu using Gpu version of PaddlePaddle,"
-      << "you must specify libcudart.so via LD_LIBRARY_PATH.";
+      << dlPath.c_str() << ". Please make sure you already specify its path. "
+      << "Note: for training data on Cpu using Gpu version of PaddlePaddle, "
+      << "you must specify libcudart via export LD_LIBRARY_PATH for Linux or "
+      << "export DYLD_LIBRARY_PATH for MAC OS.";
 }
 
 void GetCublasDsoHandle(void** dso_handle) {
+#if defined(__APPLE__) || defined(__OSX__)
+    GetDsoHandleWithSearchPath(FLAGS_cuda_dir, "libcublas.dylib", dso_handle);
+#else
     GetDsoHandleWithSearchPath(FLAGS_cuda_dir, "libcublas.so", dso_handle);
+#endif
 }
 
 void GetCudnnDsoHandle(void** dso_handle) {
+#if defined(__APPLE__) || defined(__OSX__)
+    GetDsoHandleWithSearchPath(FLAGS_cudnn_dir, "libcudnn.dylib", dso_handle);
+#else
     GetDsoHandleWithSearchPath(FLAGS_cudnn_dir, "libcudnn.so", dso_handle);
+#endif
 }
 
 void GetCudartDsoHandle(void** dso_handle) {
+#if defined(__APPLE__) || defined(__OSX__)
+    GetDsoHandleWithSearchPath("", "libcudart.dylib", dso_handle);
+#else
     GetDsoHandleWithSearchPath("", "libcudart.so", dso_handle);
+#endif
 }
 
 void GetCurandDsoHandle(void** dso_handle) {
+#if defined(__APPLE__) || defined(__OSX__)
+    GetDsoHandleWithSearchPath(FLAGS_cuda_dir, "libcurand.dylib", dso_handle);
+#else
     GetDsoHandleWithSearchPath(FLAGS_cuda_dir, "libcurand.so", dso_handle);
+#endif
 }

paddle/cuda/src/hl_table_apply.cu

@@ -35,7 +35,7 @@ __global__ void KeMatrixAddRows(real* output, int ldo,
       real *tab = table + tableId * ldt;
       for (int i = idx; i < dim; i += blockDimX) {
         if (AddRow) {
-          atomicAdd(&tab[i], out[i]);
+          paddle::paddleAtomicAdd(&tab[i], out[i]);
         } else {
           out[i] += tab[i];
         }

paddle/gserver/dataproviders/DataProviderGroup.h

@@ -65,7 +65,8 @@ void DataProviderGroup<T>::reset() {
   provider_ = nullptr;
 
   // shuffle file list
-  std::random_shuffle(fileList_.begin(), fileList_.end());
+  std::shuffle(fileList_.begin(), fileList_.end(),
+      ThreadLocalRandomEngine::get());
 
   startLoader();
   DataProvider::reset();

paddle/gserver/dataproviders/ProtoDataProvider.cpp

@@ -374,7 +374,8 @@ void ProtoDataProvider::reset() {
 }
 
 void ProtoDataProvider::shuffle() {
-  std::random_shuffle(shuffledSequenceIds_.begin(), shuffledSequenceIds_.end());
+  std::shuffle(shuffledSequenceIds_.begin(), shuffledSequenceIds_.end(),
+      ThreadLocalRandomEngine::get());
 }
 
 /*

paddle/gserver/dataproviders/PyDataProvider.cpp

@@ -17,6 +17,8 @@ limitations under the License. */
 #include "paddle/utils/PythonUtil.h"
 #include <fenv.h>
 #include "paddle/utils/Util.h"
+#include "paddle/utils/Excepts.h"
+
 
 namespace paddle {
 
@@ -44,7 +46,6 @@ PyDataProvider::PyDataProvider(const DataConfig& config, bool useGpu,
 }
 
 void PyDataProvider::loadData(const std::vector<std::string>& fileList) {
-  int feFlag = fegetexcept();
   VLOG(1) << "module:" << pyModuleName_ << " class:" << pyClassName_;
   classInstance_ =
       createPythonClass(pyModuleName_, pyClassName_, fileList, pyUserArgs_);
@@ -55,7 +56,7 @@ void PyDataProvider::loadData(const std::vector<std::string>& fileList) {
   std::string headerInfo =
       std::string(PyString_AsString(obj.get()), PyString_Size(obj.get()));
   parseHeaderData(headerInfo);
-  feenableexcept(feFlag);
+  feenableexcept(FE_INVALID | FE_DIVBYZERO | FE_OVERFLOW);
 }
 
 void PyDataProvider::parseHeaderData(const std::string& headerData) {

paddle/gserver/gradientmachines/NeuralNetwork.cpp

@@ -385,17 +385,17 @@ void NeuralNetwork::setOutputGrad(const std::vector<Argument>& args) {
   }
 }
 
-extern NeuralNetwork* newCustomNeuralNetwork(
-    const std::string& name, NeuralNetwork* network) __attribute__((weak));
+extern NeuralNetwork* newCustomNerualNetwork(
+  const std::string& name, NeuralNetwork* network) __attribute__((weak));
 
 NeuralNetwork* NeuralNetwork::newNeuralNetwork(
     const std::string& name,
     NeuralNetwork* rootNetwork) {
-  if (newCustomNeuralNetwork) {
-    return newCustomNeuralNetwork(name, rootNetwork);
-  } else {
-    return new NeuralNetwork(name, rootNetwork);
-  }
+    if (newCustomNerualNetwork) {
+      return newCustomNerualNetwork(name, rootNetwork);
+    } else {
+      return new NeuralNetwork(name, rootNetwork);
+    }
 }
 
 }  // namespace paddle

paddle/gserver/tests/concat_table_a.conf

@@ -16,9 +16,9 @@
 
 from paddle.trainer_config_helpers import *
 
-settings(batch_size=1000)
+settings(batch_size=300)
 
-data = data_layer(name ="input", size=100000)
+data = data_layer(name ="input", size=10000)
 
 # emb1 is equal to emb2, note that bias_attr=false 
 # and act=LinearActivation() in default.

paddle/gserver/tests/concat_table_b.conf

@@ -16,9 +16,9 @@
 
 from paddle.trainer_config_helpers import *
 
-settings(batch_size=1000)
+settings(batch_size=300)
 
-data = data_layer(name ="input", size=100000)
+data = data_layer(name ="input", size=10000)
 
 proj1 = table_projection(input=data, size=128)
 

paddle/gserver/tests/test_LayerGrad.cpp

@@ -50,7 +50,7 @@ TEST(Operator, dot_mul) {
 TEST(Projection, context) {
   for (auto contextStart : {-5, -3, -1, 0, 3}) {
     for (auto contextLength : {1, 2, 5, 7}) {
-      for (auto batchSize : {1, 2, 5, 20, 100}) {
+      for (auto batchSize : {1, 2, 5, 20, 50}) {
         for (auto trainablePadding : {false, true}) {
           LOG(INFO) << " contextStart=" << contextStart
                     << " contextLength=" << contextLength

paddle/gserver/tests/test_PyDataProvider2.cpp

@@ -321,7 +321,7 @@ TEST(PyDataProvider2, input_order) {
     if (!realBatchSize) {
       break;
     }
-    ASSERT_EQ(batch.getStreams().size(), 2);
+    ASSERT_EQ(batch.getStreams().size(), (size_t)2);
     for (size_t i = 0; i < realBatchSize; ++i) {
       ASSERT_EQ(batch.getStream(0).ids->getData()[i], 0);
       ASSERT_EQ(batch.getStream(1).ids->getData()[i], 1);

paddle/math/Allocator.h

@@ -16,7 +16,7 @@ limitations under the License. */
 #pragma once
 
 #include <mutex>
-#include <malloc.h>
+#include <stdlib.h>
 #include "hl_gpu.h"
 #include "paddle/utils/Logging.h"
 
@@ -48,9 +48,10 @@ public:
    * @return Pointer to the allocated memory
    */
   virtual void* alloc(size_t size) {
-    void* ptr = memalign(32ul, size);
-    CHECK(ptr) << "Fail to allocate CPU memory: size=" << size;
-    return ptr;
+      void* ptr;
+      CHECK_EQ(posix_memalign(&ptr, 32ul, size), 0);
+      CHECK(ptr) << "Fail to allocate CPU memory: size=" << size;
+      return ptr;
   }
 
   /**

paddle/math/MathFunctions.h

@@ -23,6 +23,8 @@ extern "C" {
 }
 #endif
 
+#include <cmath>
+
 namespace paddle {
 
 template<class T>

paddle/math/Matrix.cpp

@@ -2514,7 +2514,8 @@ void SharedCpuMatrix::mul(CpuSparseMatrix* a, CpuMatrix* b, real scaleAB,
     for (int k = 0; k < blockNum_; ++k) {
       blockSeq.push_back(k);
     }
-    std::random_shuffle(blockSeq.begin(), blockSeq.end());
+    std::shuffle(blockSeq.begin(), blockSeq.end(),
+        ThreadLocalRandomEngine::get());
   }
   std::vector<int>& localBufRows = *localBufRows_;
   int* cols = a->getCols();

paddle/math/PoolAllocator.h

@@ -19,6 +19,7 @@ limitations under the License. */
 #include <mutex>
 #include <vector>
 #include <unordered_map>
+#include <map>
 #include "Allocator.h"
 
 namespace paddle {

paddle/math/Storage.cpp

@@ -25,8 +25,8 @@ namespace paddle {
 // Initialization StorageEngine singleton.
 // Other modules may rely on storage management,
 // so StorageEngine need to be initialized before other modules.
-static InitFunction __init_storage_engine(
-  StorageEngine::singleton, std::numeric_limits<int>::max());
+static InitFunction __init_storage_engine([](){StorageEngine::singleton();},
+                                          std::numeric_limits<int>::max());
 
 StorageEngine::StorageEngine() : cpuAllocator_(nullptr) {
 }

paddle/math/tests/test_SIMDFunctions.cpp

@@ -24,7 +24,7 @@ limitations under the License. */
 #include <algorithm>
 #include <memory>
 
-#include <malloc.h>
+#include <stdlib.h>
 #include <time.h>
 
 static constexpr size_t VECTOR_LEN = 3072;
@@ -37,7 +37,9 @@ static std::mt19937 RandomEngine(time(0));
 
 inline static std::unique_ptr<float[]> NewVector(size_t len = VECTOR_LEN,
                                                  size_t align = ALIGN) {
-  return std::unique_ptr<float[]>((float*)memalign(align, len * sizeof(float)));
+  float* ptr;
+  CHECK_EQ(posix_memalign((void**)&ptr, align, len * sizeof(float)), 0);
+  return std::unique_ptr<float[]>(ptr);
 }
 
 inline static std::unique_ptr<float[]> NewRandomVector(size_t len = VECTOR_LEN,

paddle/math/tests/test_matrixUtil.h

@@ -124,8 +124,8 @@ void checkSMatrixErr(const CpuSparseMatrixPtr& a,
       if (a->getValueType() == FLOAT_VALUE) {
         real aVal = a->getValue()[r];
         real bVal = b->getValue()[r];
-        if (fabs(aVal - bVal) > err) {
-          if ((fabsf(aVal - bVal) / fabsf(aVal)) > (err / 10.0f)) {
+        if (std::abs(aVal - bVal) > err) {
+          if ((std::abs(aVal - bVal) / std::abs(aVal)) > (err / 10.0f)) {
             LOG(INFO) << "a=" << aVal << "\t" << "b=" << bVal;
             count++;
           }
@@ -141,8 +141,8 @@ void checkSMatrixErr(const CpuSparseMatrixPtr& a,
       if (a->getValueType() == FLOAT_VALUE) {
         real aVal = a->getValue()[r];
         real bVal = b->getValue()[r];
-        if (fabs(aVal - bVal) > err) {
-          if ((fabsf(aVal - bVal) / fabsf(aVal)) > (err / 10.0f)) {
+        if (std::abs(aVal - bVal) > err) {
+          if ((std::abs(aVal - bVal) / std::abs(aVal)) > (err / 10.0f)) {
             count++;
           }
         }
@@ -173,8 +173,8 @@ void checkMatrixErr(const Matrix& matrix1, const Matrix& matrix2) {
     for (int j = 0; j < width; j++) {
       real a = data1[i * width + j];
       real b = data2[i * width + j];
-      if (fabs(a - b) > err) {
-        if ((fabsf(a - b) / fabsf(a)) > (err / 10.0f)) {
+      if (std::abs(a - b) > err) {
+        if ((std::abs(a - b) / std::abs(a)) > (err / 10.0f)) {
           count++;
         }
       }

paddle/parameter/tests/test_common.cpp

@@ -13,7 +13,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 
-#include <malloc.h>
+#include <stdlib.h>
 #include <paddle/utils/Util.h>
 
 #include <gtest/gtest.h>
@@ -124,9 +124,13 @@ void CommonTest::test_sgdUpadate(real* gradientBuffer, real* valueBuffer,
 TEST_F(CommonTest, sgdUpdate) {
   const size_t alignHeader[] = {0, 2, 3, 5, 7, 8};
   for (auto& size : sizeVec_) {
-    real* gradientBuffer = (real*)memalign(32, sizeof(real) * size);
-    real* valueBuffer = (real*)memalign(32, sizeof(real) * size);
-    real* momentumBuffer = (real*)memalign(32, sizeof(real) * size);
+    real *gradientBuffer, *valueBuffer, *momentumBuffer;
+    CHECK_EQ(posix_memalign((void**)&gradientBuffer, 32, sizeof(real) * size),
+        0);
+    CHECK_EQ(posix_memalign((void**)&valueBuffer, 32, sizeof(real) * size), 0);
+    CHECK_EQ(posix_memalign((void**)&momentumBuffer, 32, sizeof(real) * size),
+        0);
+
     for (size_t i = 0; i < size; i++) {
       gradientBuffer[i] = 1.0;
       valueBuffer[i] = 2.0;

paddle/pserver/LightNetwork.cpp

@@ -17,6 +17,7 @@ limitations under the License. */
 #include <sys/socket.h>
 #include <netdb.h>
 #include <netinet/in.h>
+#include <netinet/tcp.h>
 #include <fcntl.h>
 
 #include <arpa/inet.h>
@@ -24,7 +25,6 @@ limitations under the License. */
 #include <net/if.h>
 #include <net/if_arp.h>
 #include <sstream>
-#include <linux/tcp.h>
 
 #include "LightNetwork.h"
 #include "paddle/utils/Util.h"
@@ -79,6 +79,7 @@ std::string getIpAddr(std::string &device) {
  * @note adjust some default sock option for better performance
  */
 void setOption(int sockfd) {
+#if !defined(__APPLE__) && !defined(__OSX__)
   int sendSize = FLAGS_sock_send_buf_size;
   int recvSize = FLAGS_sock_recv_buf_size;
   CHECK_GE(
@@ -87,15 +88,19 @@ void setOption(int sockfd) {
   CHECK_GE(
       setsockopt(sockfd, SOL_SOCKET, SO_SNDBUF, &sendSize, sizeof(sendSize)),
       0);
+#endif
+
   if (FLAGS_small_messages) {
     int optval = 1;
     CHECK_GE(
         setsockopt(sockfd, IPPROTO_TCP, TCP_NODELAY, &optval, sizeof(optval)),
         0);
+#ifdef TCP_QUICKACK
     optval = 1;
     CHECK_GE(
         setsockopt(sockfd, IPPROTO_TCP, TCP_QUICKACK, &optval, sizeof(optval)),
         0);
+#endif
   }
   int reuse = 1;
   CHECK_GE(setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, &reuse, sizeof(reuse)),
@@ -340,17 +345,27 @@ void SocketWorker::run() {
  */
 void SocketClient::TcpClient(const std::string &serverAddr, int serverPort) {
   struct sockaddr_in serv_addr;
-  struct hostent hostinfo, *server;
-  char buf[1024];  // temp for gethostbyname_r
+  struct hostent *server;
+
   int errRet;      // temp for gethostbyname_r
 
   /// Create a socket point
   int sockfd = socket(AF_INET, SOCK_STREAM, 0);
   PCHECK(sockfd >= 0) << "ERROR opening socket";
-  CHECK_EQ(0, gethostbyname_r(serverAddr.c_str(), &hostinfo, buf, sizeof(buf),
-                              &server, &errRet))
-      << "ERROR, no such host: " << serverAddr << " ret = " << errRet;
-  CHECK(server) << "gethostbyname_r err";
+
+#if defined(__OSX__) || defined(__APPLE__)
+   server = getipnodebyname(serverAddr.c_str(), AF_INET, AI_DEFAULT, &errRet);
+   CHECK_NE(HOST_NOT_FOUND, errRet)
+     << "ERROR, no such host: " << serverAddr << " ret = " << errRet;
+   CHECK(server) << "getipnodebyname error!";
+#else
+   struct hostent hostinfo;
+   char buf[1024];  // temp for gethostbyname_r
+   CHECK_EQ(0, gethostbyname_r(serverAddr.c_str(), &hostinfo, buf, sizeof(buf),
+                               &server, &errRet))
+       << "ERROR, no such host: " << serverAddr << " ret = " << errRet;
+   CHECK(server) << "gethostbyname_r error!";
+#endif
 
   bzero((char *)&serv_addr, sizeof(serv_addr));
   serv_addr.sin_family = AF_INET;

paddle/pserver/SocketChannel.cpp

@@ -27,6 +27,15 @@ limitations under the License. */
 
 namespace paddle {
 
+/**
+ * UIO_MAXIOV is documented in writev(2), but <sys/uio.h> only
+ * declares it on osx/ios if defined(KERNEL)
+ */
+#ifndef UIO_MAXIOV
+#define UIO_MAXIOV 512
+#endif
+
+
 SocketChannel::~SocketChannel() {
   if (tcpRdma_ == F_TCP)
     close(tcpSocket_);
@@ -148,8 +157,7 @@ void SocketChannel::writeMessage(const std::vector<struct iovec>& userIovs) {
   std::vector<iovec> iovs;
   iovs.reserve(userIovs.size() + 2);
   iovs.push_back({&header, sizeof(header)});
-  iovs.push_back({&iovLengths[0],
-       sizeof(iovLengths[0]) * (size_t) header.numIovs});
+  iovs.push_back({&iovLengths[0], sizeof(iovLengths[0]) * header.numIovs});
   iovs.insert(iovs.end(), userIovs.begin(), userIovs.end());
 
   header.totalLength = 0;

paddle/setup.py.in

@@ -17,6 +17,14 @@
 from setuptools import setup, Extension
 import numpy as np
 import api.paddle_ld_flags
+import platform
+
+system = platform.system().lower()
+
+is_osx = (system == 'darwin')
+is_win = (system == 'windows')
+is_lin = (system == 'linux')
+
 
 # The extra links will passed from COMAKE
 #   because generate paddle LDFLAGS is too complicated to do in setup.py
@@ -34,17 +42,24 @@ try:
 except:
   pass
 
+if is_lin == True:
+    extra_links = ["-Xlinker", '-start-group'] + extra_links + ["-Xlinker", "-end-group"]
+elif is_osx == True:
+    extra_links = ["-Wl,-all_load"] + extra_links
+
+include_dirs = [np.get_include(), "../"]    # include numpy and paddle.
+
 setup(name="py_paddle",
   version="@PADDLE_VERSION@",
   ext_modules=[
     Extension('py_paddle._swig_paddle',      # Build SWIG Extension.
-      ['Paddle_wrap.cxx'],
-      extra_link_args=["-Xlinker", '-start-group'] +
-                        extra_links + ["-Xlinker", "-end-group"]
+       ['Paddle_wrap.cxx'],
+       include_dirs = include_dirs,
+       extra_link_args = extra_links
     )
   ],
   packages=['py_paddle'],
-  include_dirs = [np.get_include(), "../"],   # include numpy and paddle.
+  include_dirs = include_dirs,
   install_requires = [
     'numpy>=1.8.0',      # The numpy is required.
     'protobuf>=2.4.1' # The paddle protobuf version

paddle/trainer/Trainer.cpp

@@ -28,6 +28,7 @@ limitations under the License. */
 #include "paddle/utils/PythonUtil.h"
 #include "paddle/utils/Stat.h"
 #include "paddle/utils/Util.h"
+#include "paddle/utils/Excepts.h"
 #include "paddle/utils/GlobalConstants.h"
 
 #include "paddle/gserver/gradientmachines/NeuralNetwork.h"

paddle/trainer/TrainerMain.cpp

@@ -16,6 +16,7 @@ limitations under the License. */
 #include <fenv.h>
 #include "paddle/utils/PythonUtil.h"
 #include "paddle/utils/StringUtil.h"
+#include "paddle/utils/Excepts.h"
 #include "paddle/pserver/ParameterServer2.h"
 
 #include "ParamUtil.h"

paddle/trainer/tests/test_CompareSparse.cpp

@@ -146,12 +146,12 @@ TEST(compareSparse, remote_cpu) {
 TEST(compareSparse, cpu10_local_vs_remote) {
   FLAGS_local = 1;  // disable remote sparse update in parameter config
   std::vector<ParameterPtr> localParameters =
-      trainerOnePassTest(configFile1, true, 10);
+      trainerOnePassTest(configFile1, true, 2);
 
   FLAGS_local = 0;  // will enable remote sparse update
   FLAGS_ports_num_for_sparse = 5;
   std::vector<ParameterPtr> remoteParameters =
-      trainerOnePassTest(configFile1, true, 10);
+      trainerOnePassTest(configFile1, true, 2);
 
   compareValue(localParameters, remoteParameters);
 }
@@ -174,7 +174,7 @@ TEST(compareSparse, multiGradientMachine) {
       FLAGS_parallel_nn = useGpu;
       LOG(INFO) << " local=" << local
                 << " useGpu=" << useGpu;
-      int trainerCount = useGpu ? numGpu : 10;
+      int trainerCount = useGpu ? numGpu : 2;
       std::vector<ParameterPtr> parameters =
           trainerOnePassTest(configFile1, true, trainerCount, useGpu);
       compareValue(getDenseParameters(), parameters, eps);

paddle/trainer/tests/test_Trainer.cpp

@@ -62,7 +62,11 @@ TEST(checkGradient, multiGpu) {
   }
 }
 
-TEST(checkGradient, parallel) { checkGradientTest(configFile4, true, true); }
+TEST(checkGradient, parallel) {
+  if (hl_get_device_count() >= 2) {
+    checkGradientTest(configFile4, true, true);
+  }
+}
 
 TEST(checkGradient, multiParallel) {
   FLAGS_allow_only_one_model_on_one_gpu = false;
@@ -90,7 +94,11 @@ TEST(checkGradient, multi) {
 TEST(checkGradient, hsigmoid) { checkGradientTest(configFile2, false, false); }
 
 TEST(checkGradient, chunk) {
+#if defined(__APPLE__) || defined (__OSX__)
+  EXPECT_EQ(0, system("python trainer/tests/gen_proto_data.py"));
+#else
   EXPECT_EQ(0, system("python2 trainer/tests/gen_proto_data.py"));
+#endif
   checkGradientTest(configFile3, false, false);
 #ifndef PADDLE_ONLY_CPU
   checkGradientTest(configFile3, true, true);

paddle/trainer/tests/test_TrainerOnePass.cpp

@@ -82,7 +82,11 @@ TEST(trainerOnePass, gpu2) { trainerOnePassTest(configFile1, true, false, 2); }
 
 TEST(trainerOnePass, gpu4) { trainerOnePassTest(configFile1, true, false, 4); }
 
-TEST(trainerOnePass, parallel) { trainerOnePassTest(configFile2, true, true); }
+TEST(trainerOnePass, parallel) {
+  if (hl_get_device_count() >= 2) {
+    trainerOnePassTest(configFile2, true, true);
+  }
+}
 #endif
 
 // 2. test average_window.

paddle/utils/CMakeLists.txt

@@ -2,12 +2,18 @@
 
 file(GLOB UTIL_HEADERS . *.h)
 file(GLOB UTIL_SOURCES . *.cpp)
-
+if(APPLE)
+    file(GLOB UTIL_ARCH_SOURCES . arch/osx/*.cpp)
+else()
+    file(GLOB UTIL_ARCH_SOURCES . arch/linux/*.cpp)
+endif()
 add_library(paddle_utils STATIC
-        ${UTIL_SOURCES})
+        ${UTIL_SOURCES}
+        ${UTIL_ARCH_SOURCES})
 add_style_check_target(paddle_utils ${UTIL_HEADERS})
-add_style_check_target(paddle_utils ${UTIL_SOURCES})
+add_style_check_target(paddle_utils ${UTIL_SOURCES}
+    ${UTIL_ARCH_SOURCES})
 add_dependencies(paddle_utils gen_proto_cpp)
 if(WITH_TESTING)
     add_subdirectory(tests)
-endif()
\ No newline at end of file
+endif()

paddle/utils/Excepts.cpp

+/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "Excepts.h"
+
+#if defined(__APPLE__) || defined(__OSX__)
+
+#include <fenv.h>
+
+int fegetexcept(void) {
+  static fenv_t fenv;
+  return fegetenv(&fenv) ? -1 : (fenv.__control & FE_ALL_EXCEPT);
+}
+
+int feenableexcept(unsigned int excepts) {
+  static fenv_t fenv;
+  unsigned int new_excepts = excepts & FE_ALL_EXCEPT, old_excepts;
+
+  if ( fegetenv (&fenv) ) return -1;
+  old_excepts = fenv.__control & FE_ALL_EXCEPT;
+
+  // unmask
+  fenv.__control &= ~new_excepts;
+  fenv.__mxcsr   &= ~(new_excepts << 7);
+
+  return ( fesetenv (&fenv) ? -1 : old_excepts );
+}
+
+int fedisableexcept(unsigned int excepts) {
+  static fenv_t fenv;
+  unsigned int new_excepts = excepts & FE_ALL_EXCEPT, old_excepts;
+
+  if ( fegetenv (&fenv) ) return -1;
+  old_excepts = fenv.__control & FE_ALL_EXCEPT;
+
+  // mask
+  fenv.__control |= new_excepts;
+  fenv.__mxcsr   |= new_excepts << 7;
+
+  return ( fesetenv (&fenv) ? -1 : old_excepts );
+}
+
+#endif

paddle/utils/Excepts.h

+/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#ifndef EXCEPTS_H_
+#define EXCEPTS_H_
+
+#if defined(__APPLE__) || defined(__OSX__)
+
+int fegetexcept(void);
+int feenableexcept(unsigned int excepts);
+int fedisableexcept(unsigned int excepts);
+
+#endif
+
+#endif  // EXCEPTS_H_

paddle/utils/Locks.h

@@ -16,13 +16,12 @@ limitations under the License. */
 #pragma once
 
 #include <pthread.h>
-#include <semaphore.h>
 #include <sys/time.h>
-#include <unistd.h>
-
 #include <condition_variable>
 #include <mutex>
 
+#include "DisableCopy.h"
+
 namespace paddle {
 
 /**
@@ -98,35 +97,44 @@ protected:
  * which means it will keep trying to lock until lock on successfully.
  * The SpinLock disable copy.
  */
+class SpinLockPrivate;
 class SpinLock {
 public:
-  SpinLock() { pthread_spin_init(&lock_, 0); }
-  ~SpinLock() { pthread_spin_destroy(&lock_); }
-  SpinLock(const SpinLock&) = delete;
-  SpinLock& operator=(const SpinLock&) = delete;
+  DISABLE_COPY(SpinLock);
+  SpinLock();
+  ~SpinLock();
 
   // std::mutext interface
-  void lock() { pthread_spin_lock(&lock_); }
-  void unlock() { pthread_spin_unlock(&lock_); }
+  void lock();
+  void unlock();
 
-protected:
-  pthread_spinlock_t lock_;
-  char padding_[64 - sizeof(pthread_spinlock_t)];
+private:
+  SpinLockPrivate* m;
 };
 
 /**
  * A simple wapper of semaphore which can only be shared in the same process.
  */
+class SemaphorePrivate;
 class Semaphore {
+public:
+  //! Disable copy & assign
+  Semaphore(const Semaphore& other) = delete;
+  Semaphore& operator= (const Semaphore&& other) = delete;
+
+  //! Enable move.
+  Semaphore(Semaphore&& other): m(std::move(other.m)) {
+  }
+
 public:
   /**
    * @brief Construct Function. 
    * @param[in] initValue the initial value of the 
    * semaphore, default 0.
    */
-  explicit Semaphore(int initValue = 0) { sem_init(&sem_, 0, initValue); }
+  explicit Semaphore(int initValue = 0);
 
-  ~Semaphore() { sem_destroy(&sem_); }
+  ~Semaphore();
 
   /**
    * @brief The same as wait(), except if the decrement can not 
@@ -136,41 +144,38 @@ public:
    * @return ture if the decrement proceeds before ts, 
    * else return false.
    */
-  bool timeWait(struct timespec* ts) { return (0 == sem_timedwait(&sem_, ts)); }
+  bool timeWait(struct timespec* ts);
 
   /**
    * @brief decrement the semaphore. If the semaphore's value is 0, then call blocks.
    */
-  void wait() { sem_wait(&sem_); }
+  void wait();
 
   /**
    * @brief increment the semaphore. If the semaphore's value 
    * greater than 0, wake up a thread blocked in wait().
    */
-  void post() { sem_post(&sem_); }
+  void post();
 
-protected:
-  sem_t sem_;
+private:
+  SemaphorePrivate* m;
 };
 
-static_assert(sizeof(SpinLock) == 64, "Wrong padding");
-
 /**
  * A simple wrapper of thread barrier.
  * The ThreadBarrier disable copy.
  */
+class ThreadBarrierPrivate;
 class ThreadBarrier {
 public:
+  DISABLE_COPY(ThreadBarrier);
+
   /**
    * @brief Construct Function. Initialize the barrier should
    * wait for count threads in wait().
    */
-  explicit ThreadBarrier(int count) {
-    pthread_barrier_init(&barrier_, NULL, count);
-  }
-  ~ThreadBarrier() { pthread_barrier_destroy(&barrier_); }
-  ThreadBarrier(const ThreadBarrier&) = delete;
-  ThreadBarrier& operator=(const ThreadBarrier&) = delete;
+  explicit ThreadBarrier(int count);
+  ~ThreadBarrier();
 
   /**
    * @brief . 
@@ -178,10 +183,10 @@ public:
    * then wake up all the count - 1 threads and continue run together. 
    * Else block the thread until waked by other thread .
    */
-  void wait() { pthread_barrier_wait(&barrier_); }
+  void wait();
 
-protected:
-  pthread_barrier_t barrier_;
+private:
+  ThreadBarrierPrivate* m;
 };
 
 /**

paddle/utils/PythonUtil.h

@@ -18,6 +18,12 @@ limitations under the License. */
 #ifndef PADDLE_NO_PYTHON
 // must include the following two blocks, otherwise,
 // gcc compiler may produce warning
+#ifdef __APPLE__
+#define _POSIX_SOURCE
+#define _POSIX_C_SOURCE 200809L
+#define _XOPEN_SOURCE 700
+#endif
+
 #ifdef _POSIX_C_SOURCE
 #define __TEMP_POSIX_C_SOURCE _POSIX_C_SOURCE
 #undef _POSIX_C_SOURCE
@@ -28,12 +34,7 @@ limitations under the License. */
 #endif
 #include <Python.h>
 #include <frameobject.h>
-#ifndef _POSIX_C_SOURCE
-#warning "no _POSIX_C_SOURCE defined in Python.h"
-#endif
-#ifndef _XOPEN_SOURCE
-#warning "no _XOPEN_SOURCE defined in Python.h"
-#endif
+
 #endif
 
 #include "paddle/utils/Util.h"

paddle/utils/Stat.cpp

@@ -13,24 +13,12 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "Stat.h"
-
-#include <sys/syscall.h>  // for syscall()
-#include <sys/types.h>
+#include "Util.h"
 #include <iomanip>
 #include <algorithm>
 
 namespace paddle {
 
-// return the thread id used by glog
-pid_t getTID() {
-#ifndef __NR_gettid
-#define __NR_gettid 224
-#endif
-  pid_t tid = syscall(__NR_gettid);
-  CHECK_NE(tid, -1);
-  return tid;
-}
-
 StatSet globalStat("GlobalStatInfo");
 
 void Stat::addSample(uint64_t value) {

paddle/utils/Thread.h

@@ -13,13 +13,10 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include "Util.h"
 #include "Logging.h"
 #include <thread>
 
-#include <sys/syscall.h>
-#include <unistd.h>
-inline pid_t gettid() { return syscall(SYS_gettid); }
-
 #include "Queue.h"
 #include "ThreadLocal.h"
 
@@ -175,7 +172,7 @@ public:
         jobFinishBarrier_(numWorkers + 1),
         jobFunc_(nullptr),
         checkOwner_(checkOwner) {
-    ownerThreadId_ = ::gettid();
+    ownerThreadId_ = getTID();
     workers_.resize(numWorkers);
     start();
   }
@@ -199,7 +196,7 @@ public:
    */
   void exec(JobFunc jobFunc, JobFunc ownerFunc = nullptr) {
     if (checkOwner_) {
-      CHECK_EQ(ownerThreadId_, ::gettid())
+      CHECK_EQ(ownerThreadId_, getTID())
           << "this sync thread pool should be used in one thread";
     }
 

paddle/utils/ThreadLocal.cpp

@@ -12,10 +12,8 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
+#include "Util.h"
 #include "ThreadLocal.h"
-
-#include "Thread.h"
-
 #include "CommandLineParser.h"
 
 P_DEFINE_bool(thread_local_rand_use_global_seed, false,
@@ -31,11 +29,11 @@ unsigned int* ThreadLocalRand::getSeed() {
   if (!p) {  // init seed
     if (FLAGS_thread_local_rand_use_global_seed) {
       p = new unsigned int(defaultSeed_);
-    } else if (getpid() == gettid()) {  // main thread
+    } else if (getpid() == getTID()) {  // main thread
       // deterministic, but differs from global srand()
       p = new unsigned int(defaultSeed_ - 1);
     } else {
-      p = new unsigned int(defaultSeed_ + gettid());
+      p = new unsigned int(defaultSeed_ + getTID());
       LOG(INFO) << "thread use undeterministic rand seed:" << *p;
     }
     seed_.set(p);
@@ -51,7 +49,7 @@ std::default_random_engine& ThreadLocalRandomEngine::get() {
     int defaultSeed = ThreadLocalRand::getDefaultSeed();
     engine->seed(FLAGS_thread_local_rand_use_global_seed
                      ? defaultSeed
-                     : defaultSeed + gettid());
+                     : defaultSeed + getTID());
     engine_.set(engine);
   }
   return *engine;

paddle/utils/ThreadLocal.h

@@ -156,7 +156,15 @@ private:
   static void dataDestructor(void* p) { delete (T*)p; }
 
   void updateMap(T* p) {
-    pid_t tid = syscall(SYS_gettid);
+#if defined(__APPLE__) || defined(__OSX__)
+    pid_t tid = syscall(SYS_thread_selfid);
+#else
+    #ifndef __NR_gettid
+    #define __NR_gettid 224
+    #endif
+    pid_t tid = syscall(__NR_gettid);
+#endif
+    CHECK_NE(tid, -1);
     std::lock_guard<std::mutex> guard(mutex_);
     auto ret = threadMap_.insert(std::make_pair(tid, p));
     if (!ret.second) {

paddle/utils/Util.cpp

@@ -93,6 +93,19 @@ static void installProfilerSwitch() {}
 
 namespace paddle {
 
+pid_t getTID() {
+  #if defined(__APPLE__) || defined(__OSX__)
+      pid_t tid = syscall(SYS_thread_selfid);
+  #else
+      #ifndef __NR_gettid
+      #define __NR_gettid 224
+      #endif
+      pid_t tid = syscall(__NR_gettid);
+  #endif
+  CHECK_NE(tid, -1);
+  return tid;
+}
+
 static bool g_initialized = false;
 typedef std::pair<int, std::function<void()>> PriorityFuncPair;
 typedef std::vector<PriorityFuncPair> InitFuncList;

paddle/utils/Util.h

@@ -24,6 +24,8 @@ limitations under the License. */
 #include <unordered_map>
 #include <mutex>
 #include <functional>
+#include <sys/syscall.h>  // for syscall()
+#include <sys/types.h>
 
 #include "CommandLineParser.h"
 #include "Logging.h"
@@ -63,6 +65,9 @@ limitations under the License. */
 
 namespace paddle {
 
+// return the thread id used by glog
+pid_t getTID();
+
 /**
  * return the 1-based index of the highest bit set
  *

paddle/utils/arch/linux/Locks.cpp

+/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/utils/Locks.h"
+#include <semaphore.h>
+#include <unistd.h>
+
+namespace paddle {
+class SemaphorePrivate {
+public:
+  sem_t sem;
+};
+
+Semaphore::Semaphore(int initValue): m(new SemaphorePrivate()) {
+  sem_init(&m->sem, 0, initValue);
+}
+
+Semaphore::~Semaphore() {
+  sem_destroy(&m->sem);
+}
+
+bool Semaphore::timeWait(struct timespec* ts) {
+  return (0 == sem_timedwait(&m->sem, ts));
+}
+
+void Semaphore::wait() {
+  sem_wait(&m->sem);
+}
+
+void Semaphore::post() {
+  sem_post(&m->sem);
+}
+
+
+class SpinLockPrivate {
+public:
+  inline SpinLockPrivate() { pthread_spin_init(&lock_, 0); }
+  inline ~SpinLockPrivate() { pthread_spin_destroy(&lock_); }
+  pthread_spinlock_t lock_;
+  char padding_[64 - sizeof(pthread_spinlock_t)];
+};
+
+SpinLock::SpinLock():m(new SpinLockPrivate()) {}
+
+
+SpinLock::~SpinLock() { delete m; }
+
+void SpinLock::lock() {
+  pthread_spin_lock(&m->lock_);
+}
+
+void SpinLock::unlock() {
+  pthread_spin_unlock(&m->lock_);
+}
+
+class ThreadBarrierPrivate {
+public:
+  pthread_barrier_t barrier_;
+};
+
+ThreadBarrier::ThreadBarrier(int count): m(new ThreadBarrierPrivate()) {
+  pthread_barrier_init(&m->barrier_, nullptr, count);
+}
+
+ThreadBarrier::~ThreadBarrier() {
+  pthread_barrier_destroy(&m->barrier_);
+  delete m;
+}
+
+void ThreadBarrier::wait() {
+  pthread_barrier_wait(&m->barrier_);
+}
+
+}  // namespace paddle

paddle/utils/arch/osx/Locks.cpp

+/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include "paddle/utils/Locks.h"
+#include "paddle/utils/Logging.h"
+#include <dispatch/dispatch.h>
+#include <libkern/OSAtomic.h>
+namespace paddle {
+
+class SemaphorePrivate {
+public:
+  ~SemaphorePrivate() {
+    dispatch_release(sem);
+  }
+
+  dispatch_semaphore_t sem;
+};
+
+Semaphore::Semaphore(int initValue): m(new SemaphorePrivate()) {
+  m->sem = dispatch_semaphore_create(initValue);
+}
+
+Semaphore::~Semaphore() {
+  delete m;
+}
+
+bool Semaphore::timeWait(timespec *ts) {
+  dispatch_time_t tm = dispatch_walltime(ts, 0);
+  return (0 == dispatch_semaphore_wait(m->sem, tm));
+}
+
+void Semaphore::wait() {
+  dispatch_semaphore_wait(m->sem, DISPATCH_TIME_FOREVER);
+}
+
+void Semaphore::post() {
+  dispatch_semaphore_signal(m->sem);
+}
+
+class SpinLockPrivate {
+public:
+  SpinLockPrivate(): lock_(OS_SPINLOCK_INIT) {}
+
+  OSSpinLock lock_;
+  char padding_[64 - sizeof(OSSpinLock)];  // Padding to cache line size
+};
+
+SpinLock::SpinLock(): m(new SpinLockPrivate()) {}
+SpinLock::~SpinLock() { delete m; }
+
+void SpinLock::lock() {
+  OSSpinLockLock(&m->lock_);
+}
+
+void SpinLock::unlock() {
+  OSSpinLockUnlock(&m->lock_);
+}
+
+
+class ThreadBarrierPrivate {
+public:
+  pthread_mutex_t mutex_;
+  pthread_cond_t cond_;
+  int count_;
+  int tripCount_;
+
+  inline explicit ThreadBarrierPrivate(int cnt):count_(0), tripCount_(cnt) {
+    CHECK_NE(cnt, 0);
+    CHECK_GE(pthread_mutex_init(&mutex_, 0), 0);
+    CHECK_GE(pthread_cond_init(&cond_, 0), 0);
+  }
+
+  inline ~ThreadBarrierPrivate() {
+    pthread_cond_destroy(&cond_);
+    pthread_mutex_destroy(&mutex_);
+  }
+
+  /**
+   * @brief wait
+   * @return true if the last wait
+   */
+  inline bool wait() {
+    pthread_mutex_lock(&mutex_);
+    ++count_;
+    if (count_ >= tripCount_) {
+      count_ = 0;
+      pthread_cond_broadcast(&cond_);
+      pthread_mutex_unlock(&mutex_);
+      return true;
+    } else {
+      pthread_cond_wait(&cond_, &mutex_);
+      pthread_mutex_unlock(&mutex_);
+      return false;
+    }
+  }
+};
+
+ThreadBarrier::ThreadBarrier(int count): m(new ThreadBarrierPrivate(count)) {}
+ThreadBarrier::~ThreadBarrier() { delete m; }
+void ThreadBarrier::wait() { m->wait(); }
+
+}  // namespace paddle

paddle/utils/tests/CMakeLists.txt

@@ -3,12 +3,15 @@ add_simple_unittest(test_Logging)
 add_simple_unittest(test_Thread)
 add_simple_unittest(test_StringUtils)
 add_simple_unittest(test_CustomStackTrace)
+add_simple_unittest(test_ThreadBarrier)
 
 add_executable(
     test_CustomStackTracePrint
     test_CustomStackTracePrint.cpp
 )
 link_paddle_exe(test_CustomStackTracePrint)
-add_test(NAME test_CustomStackTracePrint
-    COMMAND ${PROJ_ROOT}/paddle/utils/tests/test_CustomStackTracePrint.sh
-    WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+if(NOT APPLE)
+    add_test(NAME test_CustomStackTracePrint
+        COMMAND ${PROJ_ROOT}/paddle/utils/tests/test_CustomStackTracePrint.sh
+        WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+endif()

paddle/utils/tests/test_CustomStackTrace.cpp

@@ -45,6 +45,7 @@ void testNormalImpl(const std::function<void(
   size_t cntDown = countDown;
   while (cntDown-- > 0) {
     startBarrier.wait();
+    sleep(1);
     doneBarrier.wait();
     ASSERT_TRUE(tracer.empty());
   }

paddle/utils/tests/test_ThreadBarrier.cpp

+/* Copyright (c) 2016 Baidu, Inc. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License. */
+
+#include <gtest/gtest.h>
+#include <set>
+#include <vector>
+#include "paddle/utils/Logging.h"
+#include "paddle/utils/CommandLineParser.h"
+#include "paddle/utils/Util.h"
+#include "paddle/utils/Locks.h"
+
+P_DEFINE_int32(test_thread_num, 100, "testing thread number");
+
+void testNormalImpl(size_t thread_num,
+                    const std::function<void(size_t,
+                    std::mutex&, std::set<std::thread::id>&,
+                    paddle::ThreadBarrier&)>& callback) {
+ std::mutex mutex;
+ std::set<std::thread::id> tids;
+ paddle::ThreadBarrier barrier(thread_num);
+
+ std::vector<std::thread> threads;
+ threads.reserve(thread_num);
+ for (size_t i = 0; i < thread_num; ++i) {
+    threads.emplace_back([&thread_num, &mutex,
+                         &tids, &barrier, &callback]{
+        callback(thread_num, mutex, tids, barrier);
+    });
+ }
+
+ for (auto& thread : threads) {
+   thread.join();
+ }
+}
+
+TEST(ThreadBarrier, normalTest) {
+  for (auto &thread_num : {10, 30, 50 , 100 , 300, 1000}) {
+    testNormalImpl(thread_num,
+                  [](size_t thread_num, std::mutex& mutex,
+                  std::set<std::thread::id>& tids,
+                  paddle::ThreadBarrier& barrier){
+      {
+        std::lock_guard<std::mutex> guard(mutex);
+        tids.insert(std::this_thread::get_id());
+      }
+      barrier.wait();
+      // Check whether all threads reach this point or not
+      CHECK_EQ(tids.size(), thread_num);
+    });
+  }
+}
+
+int main(int argc, char** argv) {
+  testing::InitGoogleTest(&argc, argv);
+  paddle::initMain(argc, argv);
+  return RUN_ALL_TESTS();
+}