diff --git a/cmake/flags.cmake b/cmake/flags.cmake
index b76852fc6c50e80633c8294fb2724b83f15293a7..0a13a02df3cae58232df95ffd8dd7a73238ce10b 100644
--- a/cmake/flags.cmake
+++ b/cmake/flags.cmake
@@ -2,6 +2,7 @@
 include(CheckCXXCompilerFlag)
 include(CheckCCompilerFlag)
 include(CheckCXXSymbolExists)
+include(CheckTypeSize)
 
 function(CheckCompilerCXX11Flag)
     if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
@@ -83,6 +84,17 @@ if(NOT UINT64_MAX_EXISTS)
   endif()
 endif()
 
+SET(CMAKE_EXTRA_INCLUDE_FILES "pthread.h")
+CHECK_TYPE_SIZE(pthread_spinlock_t SPINLOCK_FOUND)
+CHECK_TYPE_SIZE(pthread_barrier_t BARRIER_FOUND)
+if(SPINLOCK_FOUND)
+  add_definitions(-DPADDLE_USE_PTHREAD_SPINLOCK)
+endif(SPINLOCK_FOUND)
+if(BARRIER_FOUND)
+  add_definitions(-DPADDLE_USE_PTHREAD_BARRIER)
+endif(BARRIER_FOUND)
+SET(CMAKE_EXTRA_INCLUDE_FILES "")
+
 # Common flags. the compiler flag used for C/C++ sources whenever release or debug
 # Do not care if this flag is support for gcc.
 set(COMMON_FLAGS
diff --git a/paddle/utils/arch/linux/Locks.cpp b/paddle/utils/arch/linux/Locks.cpp
index c189229cfccf7d907118bd87a15889c8fab00724..310c9a6542563891d4ba5888e58406ea28d6a2ce 100644
--- a/paddle/utils/arch/linux/Locks.cpp
+++ b/paddle/utils/arch/linux/Locks.cpp
@@ -15,6 +15,7 @@ limitations under the License. */
 #include "paddle/utils/Locks.h"
 #include <semaphore.h>
 #include <unistd.h>
+#include "paddle/utils/Logging.h"
 
 namespace paddle {
 class SemaphorePrivate {
@@ -26,7 +27,10 @@ Semaphore::Semaphore(int initValue) : m(new SemaphorePrivate()) {
   sem_init(&m->sem, 0, initValue);
 }
 
-Semaphore::~Semaphore() { sem_destroy(&m->sem); }
+Semaphore::~Semaphore() {
+  sem_destroy(&m->sem);
+  delete m;
+}
 
 bool Semaphore::timeWait(struct timespec* ts) {
   return (0 == sem_timedwait(&m->sem, ts));
@@ -36,70 +40,101 @@ void Semaphore::wait() { sem_wait(&m->sem); }
 
 void Semaphore::post() { sem_post(&m->sem); }
 
+#ifdef PADDLE_USE_PTHREAD_SPINLOCK
+
 class SpinLockPrivate {
 public:
-  inline SpinLockPrivate() {
-#ifndef __ANDROID__
-    pthread_spin_init(&lock_, 0);
-#else
-    lock_ = 0;
-#endif
-  }
-  inline ~SpinLockPrivate() {
-#ifndef __ANDROID__
-    pthread_spin_destroy(&lock_);
-#endif
-  }
-#ifndef __ANDROID__
+  inline SpinLockPrivate() { pthread_spin_init(&lock_, 0); }
+  inline ~SpinLockPrivate() { pthread_spin_destroy(&lock_); }
+
+  inline void lock() { pthread_spin_lock(&lock_); }
+  inline void unlock() { pthread_spin_unlock(&lock_); }
+
   pthread_spinlock_t lock_;
-#else
-  unsigned long lock_;
-#endif
-  char padding_[64 - sizeof(lock_)];
+  char padding_[64 - sizeof(pthread_spinlock_t)];
 };
 
-SpinLock::SpinLock() : m(new SpinLockPrivate()) {}
+#else
 
-SpinLock::~SpinLock() { delete m; }
+#include <atomic>
+class SpinLockPrivate {
+public:
+  inline void lock() {
+    while (lock_.test_and_set(std::memory_order_acquire)) {
+    }
+  }
+  inline void unlock() { lock_.clear(std::memory_order_release); }
 
-void SpinLock::lock() {
-#ifndef __ANDROID__
-  pthread_spin_lock(&m->lock_);
-#endif
-}
+  std::atomic_flag lock_ = ATOMIC_FLAG_INIT;
+  char padding_[64 - sizeof(lock_)];  // Padding to cache line size
+};
 
-void SpinLock::unlock() {
-#ifndef __ANDROID__
-  pthread_spin_unlock(&m->lock_);
 #endif
-}
+
+SpinLock::SpinLock() : m(new SpinLockPrivate()) {}
+SpinLock::~SpinLock() { delete m; }
+void SpinLock::lock() { m->lock(); }
+void SpinLock::unlock() { m->unlock(); }
+
+#ifdef PADDLE_USE_PTHREAD_BARRIER
 
 class ThreadBarrierPrivate {
 public:
-#ifndef __ANDROID__
   pthread_barrier_t barrier_;
-#else
-  unsigned long barrier_;
-#endif
+
+  inline explicit ThreadBarrierPrivate(int count) {
+    pthread_barrier_init(&barrier_, nullptr, count);
+  }
+
+  inline ~ThreadBarrierPrivate() { pthread_barrier_destroy(&barrier_); }
+
+  inline void wait() { pthread_barrier_wait(&barrier_); }
 };
 
-ThreadBarrier::ThreadBarrier(int count) : m(new ThreadBarrierPrivate()) {
-#ifndef __ANDROID__
-  pthread_barrier_init(&m->barrier_, nullptr, count);
-#endif
-}
+#else
 
-ThreadBarrier::~ThreadBarrier() {
-#ifndef __ANDROID__
-  pthread_barrier_destroy(&m->barrier_);
-#endif
-  delete m;
-}
+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;
+    }
+  }
+};
 
-void ThreadBarrier::wait() {
-#ifndef __ANDROID__
-  pthread_barrier_wait(&m->barrier_);
 #endif
-}
+
+ThreadBarrier::ThreadBarrier(int count) : m(new ThreadBarrierPrivate(count)) {}
+ThreadBarrier::~ThreadBarrier() { delete m; }
+void ThreadBarrier::wait() { m->wait(); }
 
 }  // namespace paddle